diff --git a/hy3dpaint/DifferentiableRenderer/MeshRender.py b/hy3dpaint/DifferentiableRenderer/MeshRender.py deleted file mode 100644 index 83c8cf8d9a74d29e2a96756673c9987f9c38fd02..0000000000000000000000000000000000000000 --- a/hy3dpaint/DifferentiableRenderer/MeshRender.py +++ /dev/null @@ -1,1414 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import cv2 -import torch -import trimesh -import numpy as np -from PIL import Image -import torch.nn.functional as F -from typing import Union, Optional, Tuple, List, Any, Callable -from dataclasses import dataclass -from enum import Enum -from .camera_utils import ( - transform_pos, - get_mv_matrix, - get_orthographic_projection_matrix, - get_perspective_projection_matrix, -) - -try: - from .mesh_utils import load_mesh, save_mesh -except: - print("Bpy IO CAN NOT BE Imported!!!") - -try: - from .mesh_inpaint_processor import meshVerticeInpaint # , meshVerticeColor -except: - print("InPaint Function CAN NOT BE Imported!!!") - - -class RenderMode(Enum): - """Rendering mode enumeration.""" - NORMAL = "normal" - POSITION = "position" - ALPHA = "alpha" - UV_POS = "uvpos" - - -class ReturnType(Enum): - """Return type enumeration.""" - TENSOR = "th" - NUMPY = "np" - PIL = "pl" - - -class TextureType(Enum): - """Texture type enumeration.""" - DIFFUSE = "diffuse" - METALLIC_ROUGHNESS = "mr" - NORMAL = "normal" - - -@dataclass -class RenderConfig: - """Unified rendering configuration.""" - elev: float = 0 - azim: float = 0 - camera_distance: Optional[float] = None - center: Optional[List[float]] = None - resolution: Optional[Union[int, Tuple[int, int]]] = None - bg_color: List[float] = None - return_type: str = "th" - - def __post_init__(self): - if self.bg_color is None: - self.bg_color = [1, 1, 1] - - -@dataclass -class ViewState: - """Camera view state for rendering pipeline.""" - proj_mat: torch.Tensor - mv_mat: torch.Tensor - pos_camera: torch.Tensor - pos_clip: torch.Tensor - resolution: Tuple[int, int] - - -def stride_from_shape(shape): - """ - Calculate stride values from a given shape for multi-dimensional indexing. - - Args: - shape: Tuple or list representing tensor dimensions - - Returns: - List of stride values for each dimension - """ - stride = [1] - for x in reversed(shape[1:]): - stride.append(stride[-1] * x) - return list(reversed(stride)) - - -def scatter_add_nd_with_count(input, count, indices, values, weights=None): - """ - Perform scatter-add operation on N-dimensional tensors with counting. - - Args: - input: Input tensor [..., C] with D dimensions + C channels - count: Count tensor [..., 1] with D dimensions - indices: Index tensor [N, D] of type long - values: Value tensor [N, C] to scatter - weights: Optional weight tensor [N, C], defaults to ones if None - - Returns: - Tuple of (updated_input, updated_count) tensors - """ - # input: [..., C], D dimension + C channel - # count: [..., 1], D dimension - # indices: [N, D], long - # values: [N, C] - - D = indices.shape[-1] - C = input.shape[-1] - size = input.shape[:-1] - stride = stride_from_shape(size) - - assert len(size) == D - - input = input.view(-1, C) # [HW, C] - count = count.view(-1, 1) - - flatten_indices = (indices * torch.tensor(stride, dtype=torch.long, device=indices.device)).sum(-1) # [N] - - if weights is None: - weights = torch.ones_like(values[..., :1]) - - input.scatter_add_(0, flatten_indices.unsqueeze(1).repeat(1, C), values) - count.scatter_add_(0, flatten_indices.unsqueeze(1), weights) - - return input.view(*size, C), count.view(*size, 1) - - -def linear_grid_put_2d(H, W, coords, values, return_count=False): - """ - Place values on a 2D grid using bilinear interpolation. - - Args: - H: Grid height - W: Grid width - coords: Coordinate tensor [N, 2] with values in range [0, 1] - values: Value tensor [N, C] to place on grid - return_count: Whether to return count information - - Returns: - 2D grid tensor [H, W, C] with interpolated values, optionally with count tensor - """ - # coords: [N, 2], float in [0, 1] - # values: [N, C] - - C = values.shape[-1] - - indices = coords * torch.tensor([H - 1, W - 1], dtype=torch.float32, device=coords.device) - indices_00 = indices.floor().long() # [N, 2] - indices_00[:, 0].clamp_(0, H - 2) - indices_00[:, 1].clamp_(0, W - 2) - indices_01 = indices_00 + torch.tensor([0, 1], dtype=torch.long, device=indices.device) - indices_10 = indices_00 + torch.tensor([1, 0], dtype=torch.long, device=indices.device) - indices_11 = indices_00 + torch.tensor([1, 1], dtype=torch.long, device=indices.device) - - h = indices[..., 0] - indices_00[..., 0].float() - w = indices[..., 1] - indices_00[..., 1].float() - w_00 = (1 - h) * (1 - w) - w_01 = (1 - h) * w - w_10 = h * (1 - w) - w_11 = h * w - - result = torch.zeros(H, W, C, device=values.device, dtype=values.dtype) # [H, W, C] - count = torch.zeros(H, W, 1, device=values.device, dtype=values.dtype) # [H, W, 1] - weights = torch.ones_like(values[..., :1]) # [N, 1] - - result, count = scatter_add_nd_with_count( - result, count, indices_00, values * w_00.unsqueeze(1), weights * w_00.unsqueeze(1) - ) - result, count = scatter_add_nd_with_count( - result, count, indices_01, values * w_01.unsqueeze(1), weights * w_01.unsqueeze(1) - ) - result, count = scatter_add_nd_with_count( - result, count, indices_10, values * w_10.unsqueeze(1), weights * w_10.unsqueeze(1) - ) - result, count = scatter_add_nd_with_count( - result, count, indices_11, values * w_11.unsqueeze(1), weights * w_11.unsqueeze(1) - ) - - if return_count: - return result, count - - mask = count.squeeze(-1) > 0 - result[mask] = result[mask] / count[mask].repeat(1, C) - - return result - - -def mipmap_linear_grid_put_2d(H, W, coords, values, min_resolution=128, return_count=False): - """ - Place values on 2D grid using mipmap-based multiresolution interpolation to fill holes. - - Args: - H: Grid height - W: Grid width - coords: Coordinate tensor [N, 2] with values in range [0, 1] - values: Value tensor [N, C] to place on grid - min_resolution: Minimum resolution for mipmap levels - return_count: Whether to return count information - - Returns: - 2D grid tensor [H, W, C] with filled values, optionally with count tensor - """ - # coords: [N, 2], float in [0, 1] - # values: [N, C] - - C = values.shape[-1] - - result = torch.zeros(H, W, C, device=values.device, dtype=values.dtype) # [H, W, C] - count = torch.zeros(H, W, 1, device=values.device, dtype=values.dtype) # [H, W, 1] - - cur_H, cur_W = H, W - - while min(cur_H, cur_W) > min_resolution: - - # try to fill the holes - mask = count.squeeze(-1) == 0 - if not mask.any(): - break - - cur_result, cur_count = linear_grid_put_2d(cur_H, cur_W, coords, values, return_count=True) - result[mask] = ( - result[mask] - + F.interpolate( - cur_result.permute(2, 0, 1).unsqueeze(0).contiguous(), (H, W), mode="bilinear", align_corners=False - ) - .squeeze(0) - .permute(1, 2, 0) - .contiguous()[mask] - ) - count[mask] = ( - count[mask] - + F.interpolate(cur_count.view(1, 1, cur_H, cur_W), (H, W), mode="bilinear", align_corners=False).view( - H, W, 1 - )[mask] - ) - cur_H //= 2 - cur_W //= 2 - - if return_count: - return result, count - - mask = count.squeeze(-1) > 0 - result[mask] = result[mask] / count[mask].repeat(1, C) - - return result - - -# ============ Core utility functions for reducing duplication ============ - -def _normalize_image_input(image: Union[np.ndarray, torch.Tensor, Image.Image]) -> Union[np.ndarray, torch.Tensor]: - """Normalize image input to consistent format.""" - if isinstance(image, Image.Image): - return np.array(image) / 255.0 - elif isinstance(image, torch.Tensor): - return image.cpu().numpy() if image.is_cuda else image - return image - - -def _convert_texture_format(tex: Union[np.ndarray, torch.Tensor, Image.Image], - texture_size: Tuple[int, int], device: str, force_set: bool = False) -> torch.Tensor: - """Unified texture format conversion logic.""" - if not force_set: - if isinstance(tex, np.ndarray): - tex = Image.fromarray((tex * 255).astype(np.uint8)) - elif isinstance(tex, torch.Tensor): - tex_np = tex.cpu().numpy() - tex = Image.fromarray((tex_np * 255).astype(np.uint8)) - - tex = tex.resize(texture_size).convert("RGB") - tex = np.array(tex) / 255.0 - return torch.from_numpy(tex).to(device).float() - else: - if isinstance(tex, np.ndarray): - tex = torch.from_numpy(tex) - return tex.to(device).float() - - -def _format_output(image: torch.Tensor, return_type: str) -> Union[torch.Tensor, np.ndarray, Image.Image]: - """Convert output to requested format.""" - if return_type == ReturnType.NUMPY.value: - return image.cpu().numpy() - elif return_type == ReturnType.PIL.value: - img_np = image.cpu().numpy() * 255 - return Image.fromarray(img_np.astype(np.uint8)) - return image - - -def _ensure_resolution_format(resolution: Optional[Union[int, Tuple[int, int]]], - default: Tuple[int, int]) -> Tuple[int, int]: - """Ensure resolution is in (height, width) format.""" - if resolution is None: - return default - if isinstance(resolution, (int, float)): - return (int(resolution), int(resolution)) - return tuple(resolution) - - -def _apply_background_mask(content: torch.Tensor, visible_mask: torch.Tensor, - bg_color: List[float], device: str) -> torch.Tensor: - """Apply background color to masked regions.""" - bg_tensor = torch.tensor(bg_color, dtype=torch.float32, device=device) - return content * visible_mask + bg_tensor * (1 - visible_mask) - - -class MeshRender: - def __init__( - self, - camera_distance=1.45, - camera_type="orth", - default_resolution=1024, - texture_size=1024, - use_antialias=True, - max_mip_level=None, - filter_mode="linear-mipmap-linear", - bake_mode="back_sample", - raster_mode="cr", - shader_type="face", - use_opengl=False, - device="cuda", - ): - """ - Initialize mesh renderer with configurable parameters. - - Args: - camera_distance: Distance from camera to object center - camera_type: Type of camera projection ("orth" or "perspective") - default_resolution: Default rendering resolution - texture_size: Size of texture maps - use_antialias: Whether to use antialiasing - max_mip_level: Maximum mipmap level for texture filtering - filter_mode: Texture filtering mode - bake_mode: Texture baking method ("back_sample", "linear", "mip-map") - raster_mode: Rasterization backend ("cr" for custom rasterizer) - shader_type: Shading type ("face" or "vertex") - use_opengl: Whether to use OpenGL backend (deprecated) - device: Computing device ("cuda" or "cpu") - """ - - self.device = device - - self.set_default_render_resolution(default_resolution) - self.set_default_texture_resolution(texture_size) - - self.camera_distance = camera_distance - self.use_antialias = use_antialias - self.max_mip_level = max_mip_level - self.filter_mode = filter_mode - self.bake_angle_thres = 75 - self.set_boundary_unreliable_scale(2) - self.bake_mode = bake_mode - self.shader_type = shader_type - - self.raster_mode = raster_mode - if self.raster_mode == "cr": - import custom_rasterizer as cr - - self.raster = cr - else: - raise f"No raster named {self.raster_mode}" - - if camera_type == "orth": - self.set_orth_scale(1.2) - elif camera_type == "perspective": - self.camera_proj_mat = get_perspective_projection_matrix( - 49.13, self.default_resolution[1] / self.default_resolution[0], 0.01, 100.0 - ) - else: - raise f"No camera type {camera_type}" - - # Removed multiprocessing components for single-threaded version - - def _create_view_state(self, config: RenderConfig) -> ViewState: - """Create unified view state for rendering pipeline.""" - proj = self.camera_proj_mat - r_mv = get_mv_matrix( - elev=config.elev, - azim=config.azim, - camera_distance=self.camera_distance if config.camera_distance is None else config.camera_distance, - center=config.center, - ) - - pos_camera = transform_pos(r_mv, self.vtx_pos, keepdim=True) - pos_clip = transform_pos(proj, pos_camera) - resolution = _ensure_resolution_format(config.resolution, self.default_resolution) - - return ViewState(proj, r_mv, pos_camera, pos_clip, resolution) - - def _compute_face_normals(self, triangles: torch.Tensor) -> torch.Tensor: - """Compute face normals from triangle vertices.""" - return F.normalize( - torch.cross( - triangles[:, 1, :] - triangles[:, 0, :], - triangles[:, 2, :] - triangles[:, 0, :], - dim=-1, - ), - dim=-1, - ) - - def _get_normals_for_shading(self, view_state: ViewState, use_abs_coor: bool = False) -> torch.Tensor: - """Get normals based on shader type and coordinate system.""" - if use_abs_coor: - mesh_triangles = self.vtx_pos[self.pos_idx[:, :3], :] - else: - pos_camera = view_state.pos_camera[:, :3] / view_state.pos_camera[:, 3:4] - mesh_triangles = pos_camera[self.pos_idx[:, :3], :] - - face_normals = self._compute_face_normals(mesh_triangles) - - # Common rasterization - rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution) - - if self.shader_type == "vertex": - vertex_normals = trimesh.geometry.mean_vertex_normals( - vertex_count=self.vtx_pos.shape[0], - faces=self.pos_idx.cpu(), - face_normals=face_normals.cpu(), - ) - vertex_normals = torch.from_numpy(vertex_normals).float().to(self.device).contiguous() - normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx) - - elif self.shader_type == "face": - tri_ids = rast_out[..., 3] - tri_ids_mask = tri_ids > 0 - tri_ids = ((tri_ids - 1) * tri_ids_mask).long() - normal = torch.zeros(rast_out.shape[0], rast_out.shape[1], rast_out.shape[2], 3).to(rast_out) - normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[tri_ids[tri_ids_mask].view(-1)] - - return normal, rast_out - - def _unified_render_pipeline(self, config: RenderConfig, mode: RenderMode, **kwargs) -> torch.Tensor: - """Unified rendering pipeline for all render modes.""" - view_state = self._create_view_state(config) - - if mode == RenderMode.ALPHA: - rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution) - return rast_out[..., -1:].long() - - elif mode == RenderMode.UV_POS: - return self.uv_feature_map(self.vtx_pos * 0.5 + 0.5) - - elif mode == RenderMode.NORMAL: - use_abs_coor = kwargs.get('use_abs_coor', False) - normalize_rgb = kwargs.get('normalize_rgb', True) - - normal, rast_out = self._get_normals_for_shading(view_state, use_abs_coor) - visible_mask = torch.clamp(rast_out[..., -1:], 0, 1) - - result = _apply_background_mask(normal, visible_mask, config.bg_color, self.device) - - if normalize_rgb: - result = (result + 1) * 0.5 - - if self.use_antialias: - result = self.raster_antialias(result, rast_out, view_state.pos_clip, self.pos_idx) - - return result[0, ...] - - elif mode == RenderMode.POSITION: - rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution) - - tex_position = 0.5 - self.vtx_pos[:, :3] / self.scale_factor - tex_position = tex_position.contiguous() - - position, _ = self.raster_interpolate(tex_position[None, ...], rast_out, self.pos_idx) - visible_mask = torch.clamp(rast_out[..., -1:], 0, 1) - - result = _apply_background_mask(position, visible_mask, config.bg_color, self.device) - - if self.use_antialias: - result = self.raster_antialias(result, rast_out, view_state.pos_clip, self.pos_idx) - - return result[0, ...] - - def set_orth_scale(self, ortho_scale): - """ - Set the orthographic projection scale and update camera projection matrix. - - Args: - ortho_scale: Scale factor for orthographic projection - """ - self.ortho_scale = ortho_scale - self.camera_proj_mat = get_orthographic_projection_matrix( - left=-self.ortho_scale * 0.5, - right=self.ortho_scale * 0.5, - bottom=-self.ortho_scale * 0.5, - top=self.ortho_scale * 0.5, - near=0.1, - far=100, - ) - - def raster_rasterize(self, pos, tri, resolution, ranges=None, grad_db=True): - """ - Rasterize triangular mesh using the configured rasterization backend. - - Args: - pos: Vertex positions in clip space - tri: Triangle indices - resolution: Rendering resolution [height, width] - ranges: Optional rendering ranges (unused in current implementation) - grad_db: Whether to compute gradients (unused in current implementation) - - Returns: - Tuple of (rasterization_output, gradient_info) - """ - - if self.raster_mode == "cr": - rast_out_db = None - if pos.dim() == 2: - pos = pos.unsqueeze(0) - - # 确保pos是float32类型 - if pos.dtype == torch.float64: - pos = pos.to(torch.float32) - - # 确保tri是int32类型 - if tri.dtype == torch.int64: - tri = tri.to(torch.int32) - - findices, barycentric = self.raster.rasterize(pos, tri, resolution) - rast_out = torch.cat((barycentric, findices.unsqueeze(-1)), dim=-1) - rast_out = rast_out.unsqueeze(0) - else: - raise f"No raster named {self.raster_mode}" - - return rast_out, rast_out_db - - def raster_interpolate(self, uv, rast_out, uv_idx): - """ - Interpolate texture coordinates or vertex attributes across rasterized triangles. - - Args: - uv: UV coordinates or vertex attributes to interpolate - rast_out: Rasterization output containing barycentric coordinates - uv_idx: UV or vertex indices for triangles - - Returns: - Tuple of (interpolated_values, gradient_info) - """ - - if self.raster_mode == "cr": - textd = None - barycentric = rast_out[0, ..., :-1] - findices = rast_out[0, ..., -1] - if uv.dim() == 2: - uv = uv.unsqueeze(0) - textc = self.raster.interpolate(uv, findices, barycentric, uv_idx) - else: - raise f"No raster named {self.raster_mode}" - - return textc, textd - - def raster_antialias(self, color, rast, pos, tri, topology_hash=None, pos_gradient_boost=1.0): - """ - Apply antialiasing to rendered colors (currently returns input unchanged). - - Args: - color: Input color values - rast: Rasterization output - pos: Vertex positions - tri: Triangle indices - topology_hash: Optional topology hash for optimization - pos_gradient_boost: Gradient boosting factor - - Returns: - Antialiased color values - """ - - if self.raster_mode == "cr": - color = color - else: - raise f"No raster named {self.raster_mode}" - - return color - - def set_boundary_unreliable_scale(self, scale): - """ - Set the kernel size for boundary unreliable region detection during texture baking. - - Args: - scale: Scale factor relative to 512 resolution baseline - """ - self.bake_unreliable_kernel_size = int( - (scale / 512) * max(self.default_resolution[0], self.default_resolution[1]) - ) - - def load_mesh( - self, - mesh, - scale_factor=1.15, - auto_center=True, - ): - """ - Load mesh from file and set up rendering data structures. - - Args: - mesh: Path to mesh file or mesh object - scale_factor: Scaling factor for mesh normalization - auto_center: Whether to automatically center the mesh - """ - vtx_pos, pos_idx, vtx_uv, uv_idx, texture_data = load_mesh(mesh) - self.set_mesh( - vtx_pos, pos_idx, vtx_uv=vtx_uv, uv_idx=uv_idx, scale_factor=scale_factor, auto_center=auto_center - ) - if texture_data is not None: - self.set_texture(texture_data) - - def save_mesh(self, mesh_path, downsample=False): - """ - Save current mesh with textures to file. - - Args: - mesh_path: Output file path - downsample: Whether to downsample textures by half - """ - - vtx_pos, pos_idx, vtx_uv, uv_idx = self.get_mesh(normalize=False) - texture_data = self.get_texture() - texture_metallic, texture_roughness = self.get_texture_mr() - texture_normal = self.get_texture_normal() - if downsample: - texture_data = cv2.resize(texture_data, (texture_data.shape[1] // 2, texture_data.shape[0] // 2)) - if texture_metallic is not None: - texture_metallic = cv2.resize( - texture_metallic, (texture_metallic.shape[1] // 2, texture_metallic.shape[0] // 2) - ) - if texture_roughness is not None: - texture_roughness = cv2.resize( - texture_roughness, (texture_roughness.shape[1] // 2, texture_roughness.shape[0] // 2) - ) - if texture_normal is not None: - texture_normal = cv2.resize( - texture_normal, (texture_normal.shape[1] // 2, texture_normal.shape[0] // 2) - ) - - save_mesh( - mesh_path, - vtx_pos, - pos_idx, - vtx_uv, - uv_idx, - texture_data, - metallic=texture_metallic, - roughness=texture_roughness, - normal=texture_normal, - ) - - def set_mesh(self, vtx_pos, pos_idx, vtx_uv=None, uv_idx=None, scale_factor=1.15, auto_center=True): - """ - Set mesh geometry data and perform coordinate transformations. - - Args: - vtx_pos: Vertex positions [N, 3] - pos_idx: Triangle vertex indices [F, 3] - vtx_uv: UV coordinates [N, 2], optional - uv_idx: Triangle UV indices [F, 3], optional - scale_factor: Scaling factor for mesh normalization - auto_center: Whether to automatically center and scale the mesh - """ - self.vtx_pos = torch.from_numpy(vtx_pos).to(self.device) - self.pos_idx = torch.from_numpy(pos_idx).to(self.device) - - # 确保顶点位置是float32类型 - if self.vtx_pos.dtype == torch.float64: - self.vtx_pos = self.vtx_pos.to(torch.float32) - - # 确保索引类型为int32 - if self.pos_idx.dtype == torch.int64: - self.pos_idx = self.pos_idx.to(torch.int32) - - if (vtx_uv is not None) and (uv_idx is not None): - self.vtx_uv = torch.from_numpy(vtx_uv).to(self.device) - self.uv_idx = torch.from_numpy(uv_idx).to(self.device) - - # 确保UV坐标是float32类型 - if self.vtx_uv.dtype == torch.float64: - self.vtx_uv = self.vtx_uv.to(torch.float32) - - # 确保UV索引类型为int32 - if self.uv_idx.dtype == torch.int64: - self.uv_idx = self.uv_idx.to(torch.int32) - else: - self.vtx_uv = None - self.uv_idx = None - - self.vtx_pos[:, [0, 1]] = -self.vtx_pos[:, [0, 1]] - self.vtx_pos[:, [1, 2]] = self.vtx_pos[:, [2, 1]] - if (vtx_uv is not None) and (uv_idx is not None): - self.vtx_uv[:, 1] = 1.0 - self.vtx_uv[:, 1] - pass - - if auto_center: - max_bb = (self.vtx_pos - 0).max(0)[0] - min_bb = (self.vtx_pos - 0).min(0)[0] - center = (max_bb + min_bb) / 2 - scale = torch.norm(self.vtx_pos - center, dim=1).max() * 2.0 - self.vtx_pos = (self.vtx_pos - center) * (scale_factor / float(scale)) - self.scale_factor = scale_factor - self.mesh_normalize_scale_factor = scale_factor / float(scale) - self.mesh_normalize_scale_center = center.unsqueeze(0).cpu().numpy() - else: - self.scale_factor = 1.0 - self.mesh_normalize_scale_factor = 1.0 - self.mesh_normalize_scale_center = np.array([[0, 0, 0]]) - - if uv_idx is not None: - self.extract_textiles() - - def _set_texture_unified(self, tex: Union[np.ndarray, torch.Tensor, Image.Image], - texture_type: TextureType, force_set: bool = False): - """Unified texture setting method.""" - converted_tex = _convert_texture_format(tex, self.texture_size, self.device, force_set) - - if texture_type == TextureType.DIFFUSE: - self.tex = converted_tex - elif texture_type == TextureType.METALLIC_ROUGHNESS: - self.tex_mr = converted_tex - elif texture_type == TextureType.NORMAL: - self.tex_normalMap = converted_tex - - def set_texture(self, tex, force_set=False): - """Set the main diffuse texture for the mesh.""" - self._set_texture_unified(tex, TextureType.DIFFUSE, force_set) - - def set_texture_mr(self, mr, force_set=False): - """Set metallic-roughness texture for PBR rendering.""" - self._set_texture_unified(mr, TextureType.METALLIC_ROUGHNESS, force_set) - - def set_texture_normal(self, normal, force_set=False): - """Set normal map texture for surface detail.""" - self._set_texture_unified(normal, TextureType.NORMAL, force_set) - - def set_default_render_resolution(self, default_resolution): - """ - Set the default resolution for rendering operations. - - Args: - default_resolution: Resolution as int (square) or tuple (height, width) - """ - if isinstance(default_resolution, int): - default_resolution = (default_resolution, default_resolution) - self.default_resolution = default_resolution - - def set_default_texture_resolution(self, texture_size): - """ - Set the default texture resolution for UV mapping operations. - - Args: - texture_size: Texture size as int (square) or tuple (height, width) - """ - if isinstance(texture_size, int): - texture_size = (texture_size, texture_size) - self.texture_size = texture_size - - def get_face_num(self): - """ - Get the number of triangular faces in the mesh. - - Returns: - Number of faces as integer - """ - return self.pos_idx.shape[0] - - def get_vertex_num(self): - """ - Get the number of vertices in the mesh. - - Returns: - Number of vertices as integer - """ - return self.vtx_pos.shape[0] - - def get_face_areas(self, from_one_index=False): - """ - Calculate the area of each triangular face in the mesh. - - Args: - from_one_index: If True, insert zero at beginning for 1-indexed face IDs - - Returns: - Numpy array of face areas - """ - v0 = self.vtx_pos[self.pos_idx[:, 0], :] - v1 = self.vtx_pos[self.pos_idx[:, 1], :] - v2 = self.vtx_pos[self.pos_idx[:, 2], :] - - # 计算两个边向量 - edge1 = v1 - v0 - edge2 = v2 - v0 - - # 计算叉积的模长的一半即为面积 - areas = torch.norm(torch.cross(edge1, edge2, dim=-1), dim=-1) * 0.5 - - areas = areas.cpu().numpy() - - if from_one_index: - # 在数组前面插入一个0,因为三角片索引是从1开始的 - areas = np.insert(areas, 0, 0) - - return areas - - def get_mesh(self, normalize=True): - """ - Get mesh geometry with optional coordinate denormalization. - - Args: - normalize: Whether to keep normalized coordinates (True) or restore original scale (False) - - Returns: - Tuple of (vertex_positions, face_indices, uv_coordinates, uv_indices) - """ - vtx_pos = self.vtx_pos.cpu().numpy() - pos_idx = self.pos_idx.cpu().numpy() - vtx_uv = self.vtx_uv.cpu().numpy() - uv_idx = self.uv_idx.cpu().numpy() - - # 坐标变换的逆变换 - if not normalize: - vtx_pos = vtx_pos / self.mesh_normalize_scale_factor - vtx_pos = vtx_pos + self.mesh_normalize_scale_center - vtx_pos[:, [1, 2]] = vtx_pos[:, [2, 1]] - vtx_pos[:, [0, 1]] = -vtx_pos[:, [0, 1]] - - vtx_uv[:, 1] = 1.0 - vtx_uv[:, 1] - return vtx_pos, pos_idx, vtx_uv, uv_idx - - def get_texture(self): - """ - Get the current diffuse texture as numpy array. - - Returns: - Texture as numpy array in range [0, 1] - """ - return self.tex.cpu().numpy() - - def get_texture_mr(self): - """ - Get metallic and roughness textures as separate channels. - - Returns: - Tuple of (metallic_texture, roughness_texture) as numpy arrays, or (None, None) if not set - """ - metallic, roughness = None, None - if hasattr(self, "tex_mr"): - mr = self.tex_mr.cpu().numpy() - metallic = np.repeat(mr[:, :, 0:1], repeats=3, axis=2) - roughness = np.repeat(mr[:, :, 1:2], repeats=3, axis=2) - return metallic, roughness - - def get_texture_normal(self): - """ - Get the normal map texture as numpy array. - - Returns: - Normal map as numpy array, or None if not set - """ - normal = None - if hasattr(self, "tex_normalMap"): - normal = self.tex_normalMap.cpu().numpy() - return normal - - def to(self, device): - """ - Move all tensor attributes to the specified device. - - Args: - device: Target device ("cuda", "cpu", etc.) - """ - self.device = device - - for attr_name in dir(self): - attr_value = getattr(self, attr_name) - if isinstance(attr_value, torch.Tensor): - setattr(self, attr_name, attr_value.to(self.device)) - - def color_rgb_to_srgb(self, image): - """ - Convert RGB color values to sRGB color space using gamma correction. - - Args: - image: Input image as PIL Image, numpy array, or torch tensor - - Returns: - sRGB corrected image in same format as input - """ - if isinstance(image, Image.Image): - image_rgb = torch.tesnor(np.array(image) / 255.0).float().to(self.device) - elif isinstance(image, np.ndarray): - image_rgb = torch.tensor(image).float() - else: - image_rgb = image.to(self.device) - - image_srgb = torch.where( - image_rgb <= 0.0031308, 12.92 * image_rgb, 1.055 * torch.pow(image_rgb, 1 / 2.4) - 0.055 - ) - - if isinstance(image, Image.Image): - image_srgb = Image.fromarray((image_srgb.cpu().numpy() * 255).astype(np.uint8)) - elif isinstance(image, np.ndarray): - image_srgb = image_srgb.cpu().numpy() - else: - image_srgb = image_srgb.to(image.device) - - return image_srgb - - def extract_textiles(self): - """ - Extract texture-space position and normal information by rasterizing - the mesh in UV coordinate space. Creates texture-space geometry mappings. - """ - - vnum = self.vtx_uv.shape[0] - vtx_uv = torch.cat( - (self.vtx_uv, torch.zeros_like(self.vtx_uv[:, 0:1]), torch.ones_like(self.vtx_uv[:, 0:1])), axis=1 - ) - vtx_uv = vtx_uv.view(1, vnum, 4) * 2 - 1 - - rast_out, rast_out_db = self.raster_rasterize(vtx_uv, self.uv_idx, resolution=self.texture_size) - position, _ = self.raster_interpolate(self.vtx_pos, rast_out, self.pos_idx) - - v0 = self.vtx_pos[self.pos_idx[:, 0], :] - v1 = self.vtx_pos[self.pos_idx[:, 1], :] - v2 = self.vtx_pos[self.pos_idx[:, 2], :] - face_normals = F.normalize(torch.cross(v1 - v0, v2 - v0, dim=-1), dim=-1) - vertex_normals = trimesh.geometry.mean_vertex_normals( - vertex_count=self.vtx_pos.shape[0], - faces=self.pos_idx.cpu(), - face_normals=face_normals.cpu(), - ) - vertex_normals = torch.from_numpy(vertex_normals).to(self.vtx_pos).contiguous() - position_normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx) - visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ..., 0] - position = position[0] - position_normal = position_normal[0] - tri_ids = rast_out[0, ..., 3] - tri_ids_mask = tri_ids > 0 - tri_ids = ((tri_ids - 1) * tri_ids_mask).long() - position_normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[ - tri_ids[tri_ids_mask].view(-1) - ] - - row = torch.arange(position.shape[0]).to(visible_mask.device) - col = torch.arange(position.shape[1]).to(visible_mask.device) - grid_i, grid_j = torch.meshgrid(row, col, indexing="ij") - - mask = visible_mask.reshape(-1) > 0 - position = position.reshape(-1, 3)[mask] - position_normal = position_normal.reshape(-1, 3)[mask] - position = torch.cat((position, torch.ones_like(position[:, :1])), axis=-1) - grid = torch.stack((grid_i, grid_j), -1).reshape(-1, 2)[mask] - - texture_indices = ( - torch.ones(self.texture_size[0], self.texture_size[1], device=self.device, dtype=torch.long) * -1 - ) - texture_indices.view(-1)[grid[:, 0] * self.texture_size[1] + grid[:, 1]] = torch.arange(grid.shape[0]).to( - device=self.device, dtype=torch.long - ) - - self.tex_position = position - self.tex_normal = position_normal - self.tex_grid = grid - self.texture_indices = texture_indices - - def render_normal(self, elev, azim, camera_distance=None, center=None, resolution=None, - bg_color=[1, 1, 1], use_abs_coor=False, normalize_rgb=True, return_type="th"): - """Render surface normals of the mesh from specified viewpoint.""" - config = RenderConfig(elev, azim, camera_distance, center, resolution, bg_color, return_type) - image = self._unified_render_pipeline(config, RenderMode.NORMAL, - use_abs_coor=use_abs_coor, normalize_rgb=normalize_rgb) - return _format_output(image, return_type) - - def convert_normal_map(self, image): - """ - Convert normal map from standard format to renderer's coordinate system. - Applies coordinate transformations for proper normal interpretation. - - Args: - image: Input normal map as PIL Image or numpy array - - Returns: - Converted normal map as PIL Image - """ - # blue is front, red is left, green is top - if isinstance(image, Image.Image): - image = np.array(image) - mask = (image == [255, 255, 255]).all(axis=-1) - - image = (image / 255.0) * 2.0 - 1.0 - - image[..., [1]] = -image[..., [1]] - image[..., [1, 2]] = image[..., [2, 1]] - image[..., [0]] = -image[..., [0]] - - image = (image + 1.0) * 0.5 - - image = (image * 255).astype(np.uint8) - image[mask] = [127, 127, 255] - - return Image.fromarray(image) - - def render_position(self, elev, azim, camera_distance=None, center=None, resolution=None, - bg_color=[1, 1, 1], return_type="th"): - """Render world-space positions of visible mesh surface points.""" - config = RenderConfig(elev, azim, camera_distance, center, resolution, bg_color, return_type) - image = self._unified_render_pipeline(config, RenderMode.POSITION) - - if return_type == ReturnType.PIL.value: - image = image.squeeze(-1).cpu().numpy() * 255 - return Image.fromarray(image.astype(np.uint8)) - return _format_output(image, return_type) - - def render_uvpos(self, return_type="th"): - """Render vertex positions mapped to UV texture space.""" - config = RenderConfig(return_type=return_type) - image = self._unified_render_pipeline(config, RenderMode.UV_POS) - return _format_output(image, return_type) - - def render_alpha(self, elev, azim, camera_distance=None, center=None, resolution=None, return_type="th"): - """Render binary alpha mask indicating visible mesh regions.""" - config = RenderConfig(elev, azim, camera_distance, center, resolution, return_type=return_type) - image = self._unified_render_pipeline(config, RenderMode.ALPHA) - - if return_type == ReturnType.PIL.value: - raise Exception("PIL format not supported for alpha rendering") - return _format_output(image, return_type) - - def uv_feature_map(self, vert_feat, bg=None): - """ - Map per-vertex features to UV texture space using mesh topology. - - Args: - vert_feat: Per-vertex feature tensor [N, C] - bg: Background value for unmapped regions (optional) - - Returns: - Feature map in UV texture space [H, W, C] - """ - vtx_uv = self.vtx_uv * 2 - 1.0 - vtx_uv = torch.cat([vtx_uv, torch.zeros_like(self.vtx_uv)], dim=1).unsqueeze(0) - vtx_uv[..., -1] = 1 - uv_idx = self.uv_idx - rast_out, rast_out_db = self.raster_rasterize(vtx_uv, uv_idx, resolution=self.texture_size) - feat_map, _ = self.raster_interpolate(vert_feat[None, ...], rast_out, uv_idx) - feat_map = feat_map[0, ...] - if bg is not None: - visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ...] - feat_map[visible_mask == 0] = bg - return feat_map - - def render_sketch_from_geometry(self, normal_image, depth_image): - """ - Generate sketch-style edge image from rendered normal and depth maps. - - Args: - normal_image: Rendered normal map tensor - depth_image: Rendered depth map tensor - - Returns: - Binary edge sketch image as tensor - """ - normal_image_np = normal_image.cpu().numpy() - depth_image_np = depth_image.cpu().numpy() - - normal_image_np = (normal_image_np * 255).astype(np.uint8) - depth_image_np = (depth_image_np * 255).astype(np.uint8) - normal_image_np = cv2.cvtColor(normal_image_np, cv2.COLOR_RGB2GRAY) - - normal_edges = cv2.Canny(normal_image_np, 80, 150) - depth_edges = cv2.Canny(depth_image_np, 30, 80) - - combined_edges = np.maximum(normal_edges, depth_edges) - - sketch_image = torch.from_numpy(combined_edges).to(normal_image.device).float() / 255.0 - sketch_image = sketch_image.unsqueeze(-1) - - return sketch_image - - def render_sketch_from_depth(self, depth_image): - """ - Generate sketch-style edge image from depth map using edge detection. - - Args: - depth_image: Input depth map tensor - - Returns: - Binary edge sketch image as tensor - """ - depth_image_np = depth_image.cpu().numpy() - depth_image_np = (depth_image_np * 255).astype(np.uint8) - depth_edges = cv2.Canny(depth_image_np, 30, 80) - combined_edges = depth_edges - sketch_image = torch.from_numpy(combined_edges).to(depth_image.device).float() / 255.0 - sketch_image = sketch_image.unsqueeze(-1) - return sketch_image - - def back_project(self, image, elev, azim, camera_distance=None, center=None, method=None): - """ - Back-project a rendered image onto the mesh's UV texture space. - Handles visibility, viewing angle, and boundary detection for texture baking. - - Args: - image: Input image to back-project (PIL Image, numpy array, or tensor) - elev: Camera elevation angle in degrees used for rendering - azim: Camera azimuth angle in degrees used for rendering - camera_distance: Camera distance (uses default if None) - center: Camera focus center (uses origin if None) - method: Back-projection method ("linear", "mip-map", "back_sample", uses default if None) - - Returns: - Tuple of (texture, cosine_map, boundary_map) tensors in UV space - """ - - if isinstance(image, Image.Image): - image = torch.tensor(np.array(image) / 255.0) - elif isinstance(image, np.ndarray): - image = torch.tensor(image) - if image.dim() == 2: - image = image.unsqueeze(-1) - image = image.float().to(self.device) - resolution = image.shape[:2] - channel = image.shape[-1] - texture = torch.zeros(self.texture_size + (channel,)).to(self.device) - cos_map = torch.zeros(self.texture_size + (1,)).to(self.device) - - proj = self.camera_proj_mat - r_mv = get_mv_matrix( - elev=elev, - azim=azim, - camera_distance=self.camera_distance if camera_distance is None else camera_distance, - center=center, - ) - pos_camera = transform_pos(r_mv, self.vtx_pos, keepdim=True) - pos_clip = transform_pos(proj, pos_camera) - pos_camera = pos_camera[:, :3] / pos_camera[:, 3:4] - - v0 = pos_camera[self.pos_idx[:, 0], :] - v1 = pos_camera[self.pos_idx[:, 1], :] - v2 = pos_camera[self.pos_idx[:, 2], :] - face_normals = F.normalize(torch.cross(v1 - v0, v2 - v0, dim=-1), dim=-1) - - tex_depth = pos_camera[:, 2].reshape(1, -1, 1).contiguous() - rast_out, rast_out_db = self.raster_rasterize(pos_clip, self.pos_idx, resolution=resolution) - visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ...] - - if self.shader_type == "vertex": - vertex_normals = trimesh.geometry.mean_vertex_normals( - vertex_count=self.vtx_pos.shape[0], - faces=self.pos_idx.cpu(), - face_normals=face_normals.cpu(), - ) - vertex_normals = torch.from_numpy(vertex_normals).float().to(self.device).contiguous() - normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx) - elif self.shader_type == "face": - tri_ids = rast_out[..., 3] - tri_ids_mask = tri_ids > 0 - tri_ids = ((tri_ids - 1) * tri_ids_mask).long() - normal = torch.zeros(rast_out.shape[0], rast_out.shape[1], rast_out.shape[2], 3).to(rast_out) - normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[tri_ids[tri_ids_mask].view(-1)] - - normal = normal[0, ...] - uv, _ = self.raster_interpolate(self.vtx_uv[None, ...], rast_out, self.uv_idx) - depth, _ = self.raster_interpolate(tex_depth, rast_out, self.pos_idx) - depth = depth[0, ...] - - depth_max, depth_min = depth[visible_mask > 0].max(), depth[visible_mask > 0].min() - depth_normalized = (depth - depth_min) / (depth_max - depth_min) - depth_image = depth_normalized * visible_mask # Mask out background. - - sketch_image = self.render_sketch_from_depth(depth_image) - - lookat = torch.tensor([[0, 0, -1]], device=self.device) - cos_image = torch.nn.functional.cosine_similarity(lookat, normal.view(-1, 3)) - cos_image = cos_image.view(normal.shape[0], normal.shape[1], 1) - - cos_thres = np.cos(self.bake_angle_thres / 180 * np.pi) - cos_image[cos_image < cos_thres] = 0 - - # shrink - if self.bake_unreliable_kernel_size > 0: - kernel_size = self.bake_unreliable_kernel_size * 2 + 1 - kernel = torch.ones((1, 1, kernel_size, kernel_size), dtype=torch.float32).to(sketch_image.device) - - visible_mask = visible_mask.permute(2, 0, 1).unsqueeze(0).float() - visible_mask = F.conv2d(1.0 - visible_mask, kernel, padding=kernel_size // 2) - visible_mask = 1.0 - (visible_mask > 0).float() # 二值化 - visible_mask = visible_mask.squeeze(0).permute(1, 2, 0) - - sketch_image = sketch_image.permute(2, 0, 1).unsqueeze(0) - sketch_image = F.conv2d(sketch_image, kernel, padding=kernel_size // 2) - sketch_image = (sketch_image > 0).float() # 二值化 - sketch_image = sketch_image.squeeze(0).permute(1, 2, 0) - visible_mask = visible_mask * (sketch_image < 0.5) - - cos_image[visible_mask == 0] = 0 - - method = self.bake_mode if method is None else method - - if method == "linear": - proj_mask = (visible_mask != 0).view(-1) - uv = uv.squeeze(0).contiguous().view(-1, 2)[proj_mask] - image = image.squeeze(0).contiguous().view(-1, channel)[proj_mask] - cos_image = cos_image.contiguous().view(-1, 1)[proj_mask] - sketch_image = sketch_image.contiguous().view(-1, 1)[proj_mask] - - texture = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], image) - cos_map = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], cos_image) - boundary_map = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], sketch_image) - elif method == "mip-map": - proj_mask = (visible_mask != 0).view(-1) - uv = uv.squeeze(0).contiguous().view(-1, 2)[proj_mask] - image = image.squeeze(0).contiguous().view(-1, channel)[proj_mask] - cos_image = cos_image.contiguous().view(-1, 1)[proj_mask] - - texture = mipmap_linear_grid_put_2d( - self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], image, min_resolution=128 - ) - cos_map = mipmap_linear_grid_put_2d( - self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], cos_image, min_resolution=256 - ) - - if self.vtx_map is not None: - vertex_normals = vertex_normals[self.vtx_map, :] - normal_map = self.uv_feature_map(vertex_normals) - cos_map_uv = torch.nn.functional.cosine_similarity(lookat, normal_map.view(-1, 3)) # .abs() - cos_map_uv = cos_map_uv.view(1, 1, normal_map.shape[0], normal_map.shape[1]) - cos_map_uv = torch.nn.functional.max_pool2d(cos_map_uv, kernel_size=3, stride=1, padding=1) - cos_map_uv = cos_map_uv.reshape(self.texture_size[0], self.texture_size[1], 1) - cos_map_uv[cos_map_uv < cos_thres] = 0 - # cos_map = torch.min(cos_map, cos_map_uv) - cos_map[cos_map_uv < cos_thres] = 0 - elif method == "back_sample": - - img_proj = torch.from_numpy( - np.array(((proj[0, 0], 0, 0, 0), (0, proj[1, 1], 0, 0), (0, 0, 1, 0), (0, 0, 0, 1))) - ).to(self.tex_position) - w2c = torch.from_numpy(r_mv).to(self.tex_position) - v_proj = self.tex_position @ w2c.T @ img_proj - inner_mask = (v_proj[:, 0] <= 1.0) & (v_proj[:, 0] >= -1.0) & (v_proj[:, 1] <= 1.0) & (v_proj[:, 1] >= -1.0) - inner_valid_idx = torch.where(inner_mask)[0].long() - img_x = torch.clamp( - ((v_proj[:, 0].clamp(-1, 1) * 0.5 + 0.5) * (resolution[0])).long(), 0, resolution[0] - 1 - ) - img_y = torch.clamp( - ((v_proj[:, 1].clamp(-1, 1) * 0.5 + 0.5) * (resolution[1])).long(), 0, resolution[1] - 1 - ) - - indices = img_y * resolution[0] + img_x - sampled_z = depth.reshape(-1)[indices] - sampled_m = visible_mask.reshape(-1)[indices] - v_z = v_proj[:, 2] - - sampled_w = cos_image.reshape(-1)[indices] - depth_thres = 3e-3 - - # valid_idx = torch.where((torch.abs(v_z - sampled_z) < depth_thres) * (sampled_m*sampled_w>0))[0] - valid_idx = torch.where((torch.abs(v_z - sampled_z) < depth_thres) & (sampled_m * sampled_w > 0))[0] - - intersection_mask = torch.isin(valid_idx, inner_valid_idx) - valid_idx = valid_idx[intersection_mask].to(inner_valid_idx) - - indices = indices[valid_idx] - sampled_b = sketch_image.reshape(-1)[indices] - sampled_w = sampled_w[valid_idx] - - # bilinear sampling rgb - wx = ((v_proj[:, 0] * 0.5 + 0.5) * resolution[0] - img_x)[valid_idx].reshape(-1, 1) - wy = ((v_proj[:, 1] * 0.5 + 0.5) * resolution[1] - img_y)[valid_idx].reshape(-1, 1) - img_x = img_x[valid_idx] - img_y = img_y[valid_idx] - img_x_r = torch.clamp(img_x + 1, 0, resolution[0] - 1) - img_y_r = torch.clamp(img_y + 1, 0, resolution[1] - 1) - indices_lr = img_y * resolution[0] + img_x_r - indices_rl = img_y_r * resolution[0] + img_x - indices_rr = img_y_r * resolution[0] + img_x_r - rgb = image.reshape(-1, channel) - sampled_rgb = (rgb[indices] * (1 - wx) + rgb[indices_lr] * wx) * (1 - wy) + ( - rgb[indices_rl] * (1 - wx) + rgb[indices_rr] * wx - ) * wy - - # return sampled_rgb, sampled_w, sampled_b, valid_idx - texture = torch.zeros(self.texture_size[0], self.texture_size[1], channel, device=self.device).reshape( - -1, channel - ) - cos_map = torch.zeros(self.texture_size[0], self.texture_size[1], 1, device=self.device).reshape(-1) - boundary_map = torch.zeros(self.texture_size[0], self.texture_size[1], 1, device=self.device).reshape(-1) - - valid_tex_indices = self.tex_grid[valid_idx, 0] * self.texture_size[1] + self.tex_grid[valid_idx, 1] - texture[valid_tex_indices, :] = sampled_rgb - cos_map[valid_tex_indices] = sampled_w - boundary_map[valid_tex_indices] = sampled_b - - texture = texture.view(self.texture_size[0], self.texture_size[1], channel) - cos_map = cos_map.view(self.texture_size[0], self.texture_size[1], 1) - # texture = torch.clamp(texture,0,1) - - else: - raise f"No bake mode {method}" - return texture, cos_map, boundary_map - - def bake_texture(self, colors, elevs, azims, camera_distance=None, center=None, exp=6, weights=None): - """ - Bake multiple view images into a single UV texture using weighted blending. - - Args: - colors: List of input images (tensors, numpy arrays, or PIL Images) - elevs: List of elevation angles for each view - azims: List of azimuth angles for each view - camera_distance: Camera distance (uses default if None) - center: Camera focus center (uses origin if None) - exp: Exponent for cosine weighting (higher values favor front-facing views) - weights: Optional per-view weights (defaults to 1.0 for all views) - - Returns: - Tuple of (merged_texture, trust_map) tensors in UV space - """ - if isinstance(colors, torch.Tensor): - colors = [colors[i, ...].float().permute(1, 2, 0) for i in range(colors.shape[0])] - else: - for i in range(len(colors)): - if isinstance(colors[i], Image.Image): - colors[i] = torch.tensor(np.array(colors[i]) / 255.0, device=self.device).float() - if weights is None: - weights = [1.0 for _ in range(len(colors))] - textures = [] - cos_maps = [] - for color, elev, azim, weight in zip(colors, elevs, azims, weights): - texture, cos_map, _ = self.back_project(color, elev, azim, camera_distance, center) - cos_map = weight * (cos_map**exp) - textures.append(texture) - cos_maps.append(cos_map) - - texture_merge, trust_map_merge = self.fast_bake_texture(textures, cos_maps) - return texture_merge, trust_map_merge - - @torch.no_grad() - def fast_bake_texture(self, textures, cos_maps): - """ - Efficiently merge multiple textures using cosine-weighted blending. - Optimizes by skipping views that don't contribute new information. - - Args: - textures: List of texture tensors to merge - cos_maps: List of corresponding cosine weight maps - - Returns: - Tuple of (merged_texture, valid_mask) tensors - """ - - channel = textures[0].shape[-1] - texture_merge = torch.zeros(self.texture_size + (channel,)).to(self.device) - trust_map_merge = torch.zeros(self.texture_size + (1,)).to(self.device) - for texture, cos_map in zip(textures, cos_maps): - view_sum = (cos_map > 0).sum() - painted_sum = ((cos_map > 0) * (trust_map_merge > 0)).sum() - if painted_sum / view_sum > 0.99: - continue - texture_merge += texture * cos_map - trust_map_merge += cos_map - texture_merge = texture_merge / torch.clamp(trust_map_merge, min=1e-8) - - return texture_merge, trust_map_merge > 1e-8 - - @torch.no_grad() - def uv_inpaint(self, texture, mask, vertex_inpaint=True, method="NS", return_float=False): - """ - Inpaint missing regions in UV texture using mesh-aware and traditional methods. - - Args: - texture: Input texture as tensor, numpy array, or PIL Image - mask: Binary mask indicating regions to inpaint (1 = keep, 0 = inpaint) - vertex_inpaint: Whether to use mesh vertex connectivity for inpainting - method: Inpainting method ("NS" for Navier-Stokes) - return_float: Whether to return float values (False returns uint8) - - Returns: - Inpainted texture as numpy array - """ - - if isinstance(texture, torch.Tensor): - texture_np = texture.cpu().numpy() - elif isinstance(texture, np.ndarray): - texture_np = texture - elif isinstance(texture, Image.Image): - texture_np = np.array(texture) / 255.0 - - if isinstance(mask, torch.Tensor): - mask = (mask.squeeze(-1).cpu().numpy() * 255).astype(np.uint8) - - if vertex_inpaint: - vtx_pos, pos_idx, vtx_uv, uv_idx = self.get_mesh() - texture_np, mask = meshVerticeInpaint(texture_np, mask, vtx_pos, vtx_uv, pos_idx, uv_idx) - - if method == "NS": - texture_np = cv2.inpaint((texture_np * 255).astype(np.uint8), 255 - mask, 3, cv2.INPAINT_NS) - assert return_float == False - - return texture_np diff --git a/hy3dpaint/DifferentiableRenderer/__init__.py b/hy3dpaint/DifferentiableRenderer/__init__.py deleted file mode 100644 index e69de29bb2d1d6434b8b29ae775ad8c2e48c5391..0000000000000000000000000000000000000000 diff --git a/hy3dpaint/DifferentiableRenderer/camera_utils.py b/hy3dpaint/DifferentiableRenderer/camera_utils.py deleted file mode 100644 index 54369b557b23c7481ed0bbcff25840eb0802cfbf..0000000000000000000000000000000000000000 --- a/hy3dpaint/DifferentiableRenderer/camera_utils.py +++ /dev/null @@ -1,107 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import math - -import numpy as np -import torch - - -def transform_pos(mtx, pos, keepdim=False): - t_mtx = torch.from_numpy(mtx).to(pos.device) if isinstance(mtx, np.ndarray) else mtx - if pos.shape[-1] == 3: - posw = torch.cat([pos, torch.ones([pos.shape[0], 1]).to(pos.device)], axis=1) - else: - posw = pos - - if keepdim: - return torch.matmul(posw, t_mtx.t())[...] - else: - return torch.matmul(posw, t_mtx.t())[None, ...] - - -def get_mv_matrix(elev, azim, camera_distance, center=None): - elev = -elev - azim += 90 - - elev_rad = math.radians(elev) - azim_rad = math.radians(azim) - - camera_position = np.array( - [ - camera_distance * math.cos(elev_rad) * math.cos(azim_rad), - camera_distance * math.cos(elev_rad) * math.sin(azim_rad), - camera_distance * math.sin(elev_rad), - ] - ) - - if center is None: - center = np.array([0, 0, 0]) - else: - center = np.array(center) - - lookat = center - camera_position - lookat = lookat / np.linalg.norm(lookat) - - up = np.array([0, 0, 1.0]) - right = np.cross(lookat, up) - right = right / np.linalg.norm(right) - up = np.cross(right, lookat) - up = up / np.linalg.norm(up) - - c2w = np.concatenate([np.stack([right, up, -lookat], axis=-1), camera_position[:, None]], axis=-1) - - w2c = np.zeros((4, 4)) - w2c[:3, :3] = np.transpose(c2w[:3, :3], (1, 0)) - w2c[:3, 3:] = -np.matmul(np.transpose(c2w[:3, :3], (1, 0)), c2w[:3, 3:]) - w2c[3, 3] = 1.0 - - return w2c.astype(np.float32) - - -def get_orthographic_projection_matrix(left=-1, right=1, bottom=-1, top=1, near=0, far=2): - """ - 计算正交投影矩阵。 - - 参数: - left (float): 投影区域左侧边界。 - right (float): 投影区域右侧边界。 - bottom (float): 投影区域底部边界。 - top (float): 投影区域顶部边界。 - near (float): 投影区域近裁剪面距离。 - far (float): 投影区域远裁剪面距离。 - - 返回: - numpy.ndarray: 正交投影矩阵。 - """ - ortho_matrix = np.eye(4, dtype=np.float32) - ortho_matrix[0, 0] = 2 / (right - left) - ortho_matrix[1, 1] = 2 / (top - bottom) - ortho_matrix[2, 2] = -2 / (far - near) - ortho_matrix[0, 3] = -(right + left) / (right - left) - ortho_matrix[1, 3] = -(top + bottom) / (top - bottom) - ortho_matrix[2, 3] = -(far + near) / (far - near) - return ortho_matrix - - -def get_perspective_projection_matrix(fovy, aspect_wh, near, far): - fovy_rad = math.radians(fovy) - return np.array( - [ - [1.0 / (math.tan(fovy_rad / 2.0) * aspect_wh), 0, 0, 0], - [0, 1.0 / math.tan(fovy_rad / 2.0), 0, 0], - [0, 0, -(far + near) / (far - near), -2.0 * far * near / (far - near)], - [0, 0, -1, 0], - ] - ).astype(np.float32) diff --git a/hy3dpaint/DifferentiableRenderer/compile_mesh_painter.sh b/hy3dpaint/DifferentiableRenderer/compile_mesh_painter.sh deleted file mode 100644 index cf24fb32b51edfda7c9e19cb11c1f72ccb95b68c..0000000000000000000000000000000000000000 --- a/hy3dpaint/DifferentiableRenderer/compile_mesh_painter.sh +++ /dev/null @@ -1 +0,0 @@ -c++ -O3 -Wall -shared -std=c++11 -fPIC `python -m pybind11 --includes` mesh_inpaint_processor.cpp -o mesh_inpaint_processor`python3-config --extension-suffix` \ No newline at end of file diff --git a/hy3dpaint/DifferentiableRenderer/mesh_inpaint_processor.cpp b/hy3dpaint/DifferentiableRenderer/mesh_inpaint_processor.cpp deleted file mode 100644 index 25678713d95eb185b39a0fc6a3c88910e61f32bf..0000000000000000000000000000000000000000 --- a/hy3dpaint/DifferentiableRenderer/mesh_inpaint_processor.cpp +++ /dev/null @@ -1,395 +0,0 @@ -#include -#include -#include - -#include -#include -#include -#include -#include - -namespace py = pybind11; -using namespace std; - -namespace { -// 内部数据结构，避免重复的buffer获取和指针设置 -struct MeshData { - int texture_height, texture_width, texture_channel; - int vtx_num; - float* texture_ptr; - uint8_t* mask_ptr; - float* vtx_pos_ptr; - float* vtx_uv_ptr; - int* pos_idx_ptr; - int* uv_idx_ptr; - - // 存储buffer以防止被销毁 - py::buffer_info texture_buf, mask_buf, vtx_pos_buf, vtx_uv_buf, pos_idx_buf, uv_idx_buf; - - MeshData(py::array_t& texture, py::array_t& mask, - py::array_t& vtx_pos, py::array_t& vtx_uv, - py::array_t& pos_idx, py::array_t& uv_idx) { - - texture_buf = texture.request(); - mask_buf = mask.request(); - vtx_pos_buf = vtx_pos.request(); - vtx_uv_buf = vtx_uv.request(); - pos_idx_buf = pos_idx.request(); - uv_idx_buf = uv_idx.request(); - - texture_height = texture_buf.shape[0]; - texture_width = texture_buf.shape[1]; - texture_channel = texture_buf.shape[2]; - texture_ptr = static_cast(texture_buf.ptr); - mask_ptr = static_cast(mask_buf.ptr); - - vtx_num = vtx_pos_buf.shape[0]; - vtx_pos_ptr = static_cast(vtx_pos_buf.ptr); - vtx_uv_ptr = static_cast(vtx_uv_buf.ptr); - pos_idx_ptr = static_cast(pos_idx_buf.ptr); - uv_idx_ptr = static_cast(uv_idx_buf.ptr); - } -}; - -// 公共函数：计算UV坐标 -pair calculateUVCoordinates(int vtx_uv_idx, const MeshData& data) { - int uv_v = round(data.vtx_uv_ptr[vtx_uv_idx * 2] * (data.texture_width - 1)); - int uv_u = round((1.0 - data.vtx_uv_ptr[vtx_uv_idx * 2 + 1]) * (data.texture_height - 1)); - return make_pair(uv_u, uv_v); -} - -// 公共函数：计算距离权重 -float calculateDistanceWeight(const array& vtx_0, const array& vtx1) { - float dist_weight = 1.0f / max( - sqrt( - pow(vtx_0[0] - vtx1[0], 2) + - pow(vtx_0[1] - vtx1[1], 2) + - pow(vtx_0[2] - vtx1[2], 2) - ), 1E-4); - return dist_weight * dist_weight; -} - -// 公共函数：获取顶点位置 -array getVertexPosition(int vtx_idx, const MeshData& data) { - return {data.vtx_pos_ptr[vtx_idx * 3], - data.vtx_pos_ptr[vtx_idx * 3 + 1], - data.vtx_pos_ptr[vtx_idx * 3 + 2]}; -} - -// 公共函数：构建图结构 -void buildGraph(vector>& G, const MeshData& data) { - G.resize(data.vtx_num); - for(int i = 0; i < data.uv_idx_buf.shape[0]; ++i) { - for(int k = 0; k < 3; ++k) { - G[data.pos_idx_ptr[i * 3 + k]].push_back(data.pos_idx_ptr[i * 3 + (k + 1) % 3]); - } - } -} - -// 通用初始化函数：处理两种掩码类型（float和int） -template -void initializeVertexDataGeneric(const MeshData& data, vector& vtx_mask, - vector>& vtx_color, vector* uncolored_vtxs = nullptr, - MaskType mask_value = static_cast(1)) { - vtx_mask.assign(data.vtx_num, static_cast(0)); - vtx_color.assign(data.vtx_num, vector(data.texture_channel, 0.0f)); - - if(uncolored_vtxs) { - uncolored_vtxs->clear(); - } - - for(int i = 0; i < data.uv_idx_buf.shape[0]; ++i) { - for(int k = 0; k < 3; ++k) { - int vtx_uv_idx = data.uv_idx_ptr[i * 3 + k]; - int vtx_idx = data.pos_idx_ptr[i * 3 + k]; - auto uv_coords = calculateUVCoordinates(vtx_uv_idx, data); - - if(data.mask_ptr[uv_coords.first * data.texture_width + uv_coords.second] > 0) { - vtx_mask[vtx_idx] = mask_value; - for(int c = 0; c < data.texture_channel; ++c) { - vtx_color[vtx_idx][c] = data.texture_ptr[(uv_coords.first * data.texture_width + - uv_coords.second) * data.texture_channel + c]; - } - } else if(uncolored_vtxs) { - uncolored_vtxs->push_back(vtx_idx); - } - } - } -} - -// 通用平滑算法：支持不同的掩码类型和检查函数 -template -void performSmoothingAlgorithm(const MeshData& data, const vector>& G, - vector& vtx_mask, vector>& vtx_color, - const vector& uncolored_vtxs, - function is_colored_func, - function set_colored_func) { - int smooth_count = 2; - int last_uncolored_vtx_count = 0; - - while(smooth_count > 0) { - int uncolored_vtx_count = 0; - - for(int vtx_idx : uncolored_vtxs) { - vector sum_color(data.texture_channel, 0.0f); - float total_weight = 0.0f; - - array vtx_0 = getVertexPosition(vtx_idx, data); - - for(int connected_idx : G[vtx_idx]) { - if(is_colored_func(vtx_mask[connected_idx])) { - array vtx1 = getVertexPosition(connected_idx, data); - float dist_weight = calculateDistanceWeight(vtx_0, vtx1); - - for(int c = 0; c < data.texture_channel; ++c) { - sum_color[c] += vtx_color[connected_idx][c] * dist_weight; - } - total_weight += dist_weight; - } - } - - if(total_weight > 0.0f) { - for(int c = 0; c < data.texture_channel; ++c) { - vtx_color[vtx_idx][c] = sum_color[c] / total_weight; - } - set_colored_func(vtx_mask[vtx_idx]); - } else { - uncolored_vtx_count++; - } - } - - if(last_uncolored_vtx_count == uncolored_vtx_count) { - smooth_count--; - } else { - smooth_count++; - } - last_uncolored_vtx_count = uncolored_vtx_count; - } -} - -// 前向传播算法的通用实现 -void performForwardPropagation(const MeshData& data, const vector>& G, - vector& vtx_mask, vector>& vtx_color, - queue& active_vtxs) { - while(!active_vtxs.empty()) { - queue pending_active_vtxs; - - while(!active_vtxs.empty()) { - int vtx_idx = active_vtxs.front(); - active_vtxs.pop(); - array vtx_0 = getVertexPosition(vtx_idx, data); - - for(int connected_idx : G[vtx_idx]) { - if(vtx_mask[connected_idx] > 0) continue; - - array vtx1 = getVertexPosition(connected_idx, data); - float dist_weight = calculateDistanceWeight(vtx_0, vtx1); - - for(int c = 0; c < data.texture_channel; ++c) { - vtx_color[connected_idx][c] += vtx_color[vtx_idx][c] * dist_weight; - } - - if(vtx_mask[connected_idx] == 0) { - pending_active_vtxs.push(connected_idx); - } - vtx_mask[connected_idx] -= dist_weight; - } - } - - while(!pending_active_vtxs.empty()) { - int vtx_idx = pending_active_vtxs.front(); - pending_active_vtxs.pop(); - - for(int c = 0; c < data.texture_channel; ++c) { - vtx_color[vtx_idx][c] /= -vtx_mask[vtx_idx]; - } - vtx_mask[vtx_idx] = 1.0f; - active_vtxs.push(vtx_idx); - } - } -} - -// 公共函数：创建输出数组 -pair, py::array_t> createOutputArrays( - const MeshData& data, const vector& vtx_mask, - const vector>& vtx_color) { - - py::array_t new_texture(data.texture_buf.size); - py::array_t new_mask(data.mask_buf.size); - - auto new_texture_buf = new_texture.request(); - auto new_mask_buf = new_mask.request(); - - float* new_texture_ptr = static_cast(new_texture_buf.ptr); - uint8_t* new_mask_ptr = static_cast(new_mask_buf.ptr); - - // Copy original texture and mask to new arrays - copy(data.texture_ptr, data.texture_ptr + data.texture_buf.size, new_texture_ptr); - copy(data.mask_ptr, data.mask_ptr + data.mask_buf.size, new_mask_ptr); - - for(int face_idx = 0; face_idx < data.uv_idx_buf.shape[0]; ++face_idx) { - for(int k = 0; k < 3; ++k) { - int vtx_uv_idx = data.uv_idx_ptr[face_idx * 3 + k]; - int vtx_idx = data.pos_idx_ptr[face_idx * 3 + k]; - - if(vtx_mask[vtx_idx] == 1.0f) { - auto uv_coords = calculateUVCoordinates(vtx_uv_idx, data); - - for(int c = 0; c < data.texture_channel; ++c) { - new_texture_ptr[ - (uv_coords.first * data.texture_width + uv_coords.second) * - data.texture_channel + c - ] = vtx_color[vtx_idx][c]; - } - new_mask_ptr[uv_coords.first * data.texture_width + uv_coords.second] = 255; - } - } - } - - // Reshape the new arrays to match the original texture and mask shapes - new_texture.resize({data.texture_height, data.texture_width, 3}); - new_mask.resize({data.texture_height, data.texture_width}); - - return make_pair(new_texture, new_mask); -} - -// 创建顶点颜色输出数组的专用函数 -pair, py::array_t> createVertexColorOutput( - const MeshData& data, const vector& vtx_mask, - const vector>& vtx_color) { - - py::array_t py_vtx_color({data.vtx_num, data.texture_channel}); - py::array_t py_vtx_mask({data.vtx_num}); - - auto py_vtx_color_buf = py_vtx_color.request(); - auto py_vtx_mask_buf = py_vtx_mask.request(); - - float* py_vtx_color_ptr = static_cast(py_vtx_color_buf.ptr); - uint8_t* py_vtx_mask_ptr = static_cast(py_vtx_mask_buf.ptr); - - for(int i = 0; i < data.vtx_num; ++i) { - py_vtx_mask_ptr[i] = vtx_mask[i]; - for(int c = 0; c < data.texture_channel; ++c) { - py_vtx_color_ptr[i * data.texture_channel + c] = vtx_color[i][c]; - } - } - - return make_pair(py_vtx_color, py_vtx_mask); -} - -} // anonymous namespace - -// 重构后的 meshVerticeInpaint_smooth 函数 -pair, py::array_t> meshVerticeInpaint_smooth( - py::array_t texture, py::array_t mask, py::array_t vtx_pos, py::array_t vtx_uv, - py::array_t pos_idx, py::array_t uv_idx) { - - MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - - vector vtx_mask; - vector> vtx_color; - vector uncolored_vtxs; - vector> G; - - initializeVertexDataGeneric(data, vtx_mask, vtx_color, &uncolored_vtxs, 1.0f); - buildGraph(G, data); - - // 使用通用平滑算法 - performSmoothingAlgorithm(data, G, vtx_mask, vtx_color, uncolored_vtxs, - [](float mask_val) { return mask_val > 0; }, // 检查是否着色 - [](float& mask_val) { mask_val = 1.0f; } // 设置为已着色 - ); - - return createOutputArrays(data, vtx_mask, vtx_color); -} - -// 重构后的 meshVerticeInpaint_forward 函数 -pair, py::array_t> meshVerticeInpaint_forward( - py::array_t texture, py::array_t mask, py::array_t vtx_pos, py::array_t vtx_uv, - py::array_t pos_idx, py::array_t uv_idx) { - - MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - - vector vtx_mask; - vector> vtx_color; - vector> G; - queue active_vtxs; - - // 使用通用初始化（不需要 uncolored_vtxs） - initializeVertexDataGeneric(data, vtx_mask, vtx_color, nullptr, 1.0f); - buildGraph(G, data); - - // 收集活跃顶点 - for(int i = 0; i < data.vtx_num; ++i) { - if(vtx_mask[i] == 1.0f) { - active_vtxs.push(i); - } - } - - // 使用通用前向传播算法 - performForwardPropagation(data, G, vtx_mask, vtx_color, active_vtxs); - - return createOutputArrays(data, vtx_mask, vtx_color); -} - -// 主接口函数 -pair, py::array_t> meshVerticeInpaint( - py::array_t texture, py::array_t mask, py::array_t vtx_pos, py::array_t vtx_uv, - py::array_t pos_idx, py::array_t uv_idx, const string& method = "smooth") { - - if(method == "smooth") { - return meshVerticeInpaint_smooth(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - } else if(method == "forward") { - return meshVerticeInpaint_forward(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - } else { - throw invalid_argument("Invalid method. Use 'smooth' or 'forward'."); - } -} - -//============================ - -// 重构后的 meshVerticeColor_smooth 函数 -pair, py::array_t> meshVerticeColor_smooth( - py::array_t texture, py::array_t mask, py::array_t vtx_pos, py::array_t vtx_uv, - py::array_t pos_idx, py::array_t uv_idx) { - - MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - - vector vtx_mask; - vector> vtx_color; - vector uncolored_vtxs; - vector> G; - - initializeVertexDataGeneric(data, vtx_mask, vtx_color, &uncolored_vtxs, 1); - buildGraph(G, data); - - // 使用通用平滑算法 - performSmoothingAlgorithm(data, G, vtx_mask, vtx_color, uncolored_vtxs, - [](int mask_val) { return mask_val > 0; }, // 检查是否着色 - [](int& mask_val) { mask_val = 2; } // 设置为已着色（值为2） - ); - - return createVertexColorOutput(data, vtx_mask, vtx_color); -} - -// meshVerticeColor 主接口函数 -pair, py::array_t> meshVerticeColor( - py::array_t texture, py::array_t mask, py::array_t vtx_pos, py::array_t vtx_uv, - py::array_t pos_idx, py::array_t uv_idx, const string& method = "smooth") { - - if(method == "smooth") { - return meshVerticeColor_smooth(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx); - } else { - throw invalid_argument("Invalid method. Use 'smooth' or 'forward'."); - } -} - -// Python绑定 -PYBIND11_MODULE(mesh_inpaint_processor, m) { - m.def("meshVerticeInpaint", &meshVerticeInpaint, "A function to process mesh", - py::arg("texture"), py::arg("mask"), py::arg("vtx_pos"), py::arg("vtx_uv"), - py::arg("pos_idx"), py::arg("uv_idx"), py::arg("method") = "smooth"); - m.def("meshVerticeColor", &meshVerticeColor, "A function to process mesh", - py::arg("texture"), py::arg("mask"), py::arg("vtx_pos"), py::arg("vtx_uv"), - py::arg("pos_idx"), py::arg("uv_idx"), py::arg("method") = "smooth"); -} diff --git a/hy3dpaint/DifferentiableRenderer/mesh_utils.py b/hy3dpaint/DifferentiableRenderer/mesh_utils.py deleted file mode 100644 index 5b3e8317b7b8e8e3a293532cc278ede157d88e07..0000000000000000000000000000000000000000 --- a/hy3dpaint/DifferentiableRenderer/mesh_utils.py +++ /dev/null @@ -1,284 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import cv2 -import bpy -import math -import numpy as np -from io import StringIO -from typing import Optional, Tuple, Dict, Any - - -def _safe_extract_attribute(obj: Any, attr_path: str, default: Any = None) -> Any: - """Extract nested attribute safely from object.""" - try: - for attr in attr_path.split("."): - obj = getattr(obj, attr) - return obj - except AttributeError: - return default - - -def _convert_to_numpy(data: Any, dtype: np.dtype) -> Optional[np.ndarray]: - """Convert data to numpy array with specified dtype, handling None values.""" - if data is None: - return None - return np.asarray(data, dtype=dtype) - - -def load_mesh(mesh): - """Load mesh data including vertices, faces, UV coordinates and texture.""" - # Extract vertex positions and face indices - vtx_pos = _safe_extract_attribute(mesh, "vertices") - pos_idx = _safe_extract_attribute(mesh, "faces") - - # Extract UV coordinates (reusing face indices for UV indices) - vtx_uv = _safe_extract_attribute(mesh, "visual.uv") - uv_idx = pos_idx # Reuse face indices for UV mapping - - # Convert to numpy arrays with appropriate dtypes - vtx_pos = _convert_to_numpy(vtx_pos, np.float32) - pos_idx = _convert_to_numpy(pos_idx, np.int32) - vtx_uv = _convert_to_numpy(vtx_uv, np.float32) - uv_idx = _convert_to_numpy(uv_idx, np.int32) - - texture_data = None - return vtx_pos, pos_idx, vtx_uv, uv_idx, texture_data - - -def _get_base_path_and_name(mesh_path: str) -> Tuple[str, str]: - """Get base path without extension and mesh name.""" - base_path = os.path.splitext(mesh_path)[0] - name = os.path.basename(base_path) - return base_path, name - - -def _save_texture_map( - texture: np.ndarray, - base_path: str, - suffix: str = "", - image_format: str = ".jpg", - color_convert: Optional[int] = None, -) -> str: - """Save texture map with optional color conversion.""" - path = f"{base_path}{suffix}{image_format}" - processed_texture = (texture * 255).astype(np.uint8) - - if color_convert is not None: - processed_texture = cv2.cvtColor(processed_texture, color_convert) - cv2.imwrite(path, processed_texture) - else: - cv2.imwrite(path, processed_texture[..., ::-1]) # RGB to BGR - - return os.path.basename(path) - - -def _write_mtl_properties(f, properties: Dict[str, Any]): - """Write material properties to MTL file.""" - for key, value in properties.items(): - if isinstance(value, (list, tuple)): - f.write(f"{key} {' '.join(map(str, value))}\n") - else: - f.write(f"{key} {value}\n") - - -def _create_obj_content( - vtx_pos: np.ndarray, vtx_uv: np.ndarray, pos_idx: np.ndarray, uv_idx: np.ndarray, name: str -) -> str: - """Create OBJ file content.""" - buffer = StringIO() - - # Write header and vertices - buffer.write(f"mtllib {name}.mtl\no {name}\n") - np.savetxt(buffer, vtx_pos, fmt="v %.6f %.6f %.6f") - np.savetxt(buffer, vtx_uv, fmt="vt %.6f %.6f") - buffer.write("s 0\nusemtl Material\n") - - # Write faces - pos_idx_plus1 = pos_idx + 1 - uv_idx_plus1 = uv_idx + 1 - face_format = np.frompyfunc(lambda *x: f"{int(x[0])}/{int(x[1])}", 2, 1) - faces = face_format(pos_idx_plus1, uv_idx_plus1) - face_strings = [f"f {' '.join(face)}" for face in faces] - buffer.write("\n".join(face_strings) + "\n") - - return buffer.getvalue() - - -def save_obj_mesh(mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=None, roughness=None, normal=None): - """Save mesh as OBJ file with textures and material.""" - # Convert inputs to numpy arrays - vtx_pos = _convert_to_numpy(vtx_pos, np.float32) - vtx_uv = _convert_to_numpy(vtx_uv, np.float32) - pos_idx = _convert_to_numpy(pos_idx, np.int32) - uv_idx = _convert_to_numpy(uv_idx, np.int32) - - base_path, name = _get_base_path_and_name(mesh_path) - - # Create and save OBJ content - obj_content = _create_obj_content(vtx_pos, vtx_uv, pos_idx, uv_idx, name) - with open(mesh_path, "w") as obj_file: - obj_file.write(obj_content) - - # Save texture maps - texture_maps = {} - texture_maps["diffuse"] = _save_texture_map(texture, base_path) - - if metallic is not None: - texture_maps["metallic"] = _save_texture_map(metallic, base_path, "_metallic", color_convert=cv2.COLOR_RGB2GRAY) - if roughness is not None: - texture_maps["roughness"] = _save_texture_map( - roughness, base_path, "_roughness", color_convert=cv2.COLOR_RGB2GRAY - ) - if normal is not None: - texture_maps["normal"] = _save_texture_map(normal, base_path, "_normal") - - # Create MTL file - _create_mtl_file(base_path, texture_maps, metallic is not None) - - -def _create_mtl_file(base_path: str, texture_maps: Dict[str, str], is_pbr: bool): - """Create MTL material file.""" - mtl_path = f"{base_path}.mtl" - - with open(mtl_path, "w") as f: - f.write("newmtl Material\n") - - if is_pbr: - # PBR material properties - properties = { - "Kd": [0.800, 0.800, 0.800], - "Ke": [0.000, 0.000, 0.000], # 鐜鍏夐伄钄� - "Ni": 1.500, # 鎶樺皠绯绘暟 - "d": 1.0, # 閫忔槑搴� - "illum": 2, # 鍏夌収妯″瀷 - "map_Kd": texture_maps["diffuse"], - } - _write_mtl_properties(f, properties) - - # Additional PBR maps - map_configs = [("metallic", "map_Pm"), ("roughness", "map_Pr"), ("normal", "map_Bump -bm 1.0")] - - for texture_key, mtl_key in map_configs: - if texture_key in texture_maps: - f.write(f"{mtl_key} {texture_maps[texture_key]}\n") - else: - # Standard material properties - properties = { - "Ns": 250.000000, - "Ka": [0.200, 0.200, 0.200], - "Kd": [0.800, 0.800, 0.800], - "Ks": [0.500, 0.500, 0.500], - "Ke": [0.000, 0.000, 0.000], - "Ni": 1.500, - "d": 1.0, - "illum": 3, - "map_Kd": texture_maps["diffuse"], - } - _write_mtl_properties(f, properties) - - -def save_mesh(mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=None, roughness=None, normal=None): - """Save mesh using OBJ format.""" - save_obj_mesh( - mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=metallic, roughness=roughness, normal=normal - ) - - -def _setup_blender_scene(): - """Setup Blender scene for conversion.""" - if "convert" not in bpy.data.scenes: - bpy.data.scenes.new("convert") - bpy.context.window.scene = bpy.data.scenes["convert"] - - -def _clear_scene_objects(): - """Clear all objects from current Blender scene.""" - for obj in bpy.context.scene.objects: - obj.select_set(True) - bpy.data.objects.remove(obj, do_unlink=True) - - -def _select_mesh_objects(): - """Select all mesh objects in scene.""" - bpy.ops.object.select_all(action="DESELECT") - for obj in bpy.context.scene.objects: - if obj.type == "MESH": - obj.select_set(True) - - -def _merge_vertices_if_needed(merge_vertices: bool): - """Merge duplicate vertices if requested.""" - if not merge_vertices: - return - - for obj in bpy.context.selected_objects: - if obj.type == "MESH": - bpy.context.view_layer.objects.active = obj - bpy.ops.object.mode_set(mode="EDIT") - bpy.ops.mesh.select_all(action="SELECT") - bpy.ops.mesh.remove_doubles() - bpy.ops.object.mode_set(mode="OBJECT") - - -def _apply_shading(shade_type: str, auto_smooth_angle: float): - """Apply shading to selected objects.""" - shading_ops = { - "SMOOTH": lambda: bpy.ops.object.shade_smooth(), - "FLAT": lambda: bpy.ops.object.shade_flat(), - "AUTO_SMOOTH": lambda: _apply_auto_smooth(auto_smooth_angle), - } - - if shade_type in shading_ops: - shading_ops[shade_type]() - - -def _apply_auto_smooth(auto_smooth_angle: float): - """Apply auto smooth based on Blender version.""" - angle_rad = math.radians(auto_smooth_angle) - - if bpy.app.version < (4, 1, 0): - bpy.ops.object.shade_smooth(use_auto_smooth=True, auto_smooth_angle=angle_rad) - elif bpy.app.version < (4, 2, 0): - bpy.ops.object.shade_smooth_by_angle(angle=angle_rad) - else: - bpy.ops.object.shade_auto_smooth(angle=angle_rad) - - -def convert_obj_to_glb( - obj_path: str, - glb_path: str, - shade_type: str = "SMOOTH", - auto_smooth_angle: float = 60, - merge_vertices: bool = False, -) -> bool: - """Convert OBJ file to GLB format using Blender.""" - try: - _setup_blender_scene() - _clear_scene_objects() - - # Import OBJ file - bpy.ops.wm.obj_import(filepath=obj_path) - _select_mesh_objects() - - # Process meshes - _merge_vertices_if_needed(merge_vertices) - _apply_shading(shade_type, auto_smooth_angle) - - # Export to GLB - bpy.ops.export_scene.gltf(filepath=glb_path, use_active_scene=True) - return True - except Exception: - return False diff --git a/hy3dpaint/LICENSE b/hy3dpaint/LICENSE deleted file mode 100644 index 96313b341082fb9f72a00b7c77accd8196390af0..0000000000000000000000000000000000000000 --- a/hy3dpaint/LICENSE +++ /dev/null @@ -1,81 +0,0 @@ -TENCENT HUNYUAN 3D 2.1 COMMUNITY LICENSE AGREEMENT -Tencent Hunyuan 3D 2.1 Release Date: June 13, 2025 -THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW. -By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan 3D 2.1 Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately. -1. DEFINITIONS. -a. “Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A. -b. “Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of Tencent Hunyuan 3D 2.1 Works or any portion or element thereof set forth herein. -c. “Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan 3D 2.1 made publicly available by Tencent. -d. “Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means. -e. “Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan 3D 2.1 Works for any purpose and in any field of use. -f. “Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan 3D 2.1 and Documentation (and any portion thereof) as made available by Tencent under this Agreement. -g. “Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; (ii) works based on Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1, to that model in order to cause that model to perform similarly to Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1 for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives. -h. “Output” shall mean the information and/or content output of Tencent Hunyuan 3D 2.1 or a Model Derivative that results from operating or otherwise using Tencent Hunyuan 3D 2.1 or a Model Derivative, including via a Hosted Service. -i. “Tencent,” “We” or “Us” shall mean THL Q Limited. -j. “Tencent Hunyuan 3D 2.1” shall mean the 3D generation models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at [ https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1]. -k. “Tencent Hunyuan 3D 2.1 Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof. -l. “Territory” shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea. -m. “Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You. -n. “including” shall mean including but not limited to. -2. GRANT OF RIGHTS. -We grant You, for the Territory only, a non-exclusive, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy. -3. DISTRIBUTION. -You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan 3D 2.1 Works, exclusively in the Territory, provided that You meet all of the following conditions: -a. You must provide all such Third Party recipients of the Tencent Hunyuan 3D 2.1 Works or products or services using them a copy of this Agreement; -b. You must cause any modified files to carry prominent notices stating that You changed the files; -c. You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan 3D 2.1 Works; and (ii) mark the products or services developed by using the Tencent Hunyuan 3D 2.1 Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and -d. All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan 3D 2.1 is licensed under the Tencent Hunyuan 3D 2.1 Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.” -You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement (including as regards the Territory). If You receive Tencent Hunyuan 3D 2.1 Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You. -4. ADDITIONAL COMMERCIAL TERMS. -If, on the Tencent Hunyuan 3D 2.1 version release date, the monthly active users of all products or services made available by or for Licensee is greater than 1 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights. -Subject to Tencent's written approval, you may request a license for the use of Tencent Hunyuan 3D 2.1 by submitting the following information to hunyuan3d@tencent.com: -a. Your company’s name and associated business sector that plans to use Tencent Hunyuan 3D 2.1. -b. Your intended use case and the purpose of using Tencent Hunyuan 3D 2.1. -c. Your plans to modify Tencent Hunyuan 3D 2.1 or create Model Derivatives. -5. RULES OF USE. -a. Your use of the Tencent Hunyuan 3D 2.1 Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan 3D 2.1 Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan 3D 2.1 Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan 3D 2.1 Works are subject to the use restrictions in these Sections 5(a) and 5(b). -b. You must not use the Tencent Hunyuan 3D 2.1 Works or any Output or results of the Tencent Hunyuan 3D 2.1 Works to improve any other AI model (other than Tencent Hunyuan 3D 2.1 or Model Derivatives thereof). -c. You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan 3D 2.1 Works, Output or results of the Tencent Hunyuan 3D 2.1 Works outside the Territory. Any such use outside the Territory is unlicensed and unauthorized under this Agreement. -6. INTELLECTUAL PROPERTY. -a. Subject to Tencent’s ownership of Tencent Hunyuan 3D 2.1 Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You. -b. No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan 3D 2.1 Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan 3D 2.1 Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) in the Territory solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent. -c. If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan 3D 2.1 Works. -d. Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses. -7. DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY. -a. We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan 3D 2.1 Works or to grant any license thereto. -b. UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN 3D 2.1 WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT. -c. TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING. -8. SURVIVAL AND TERMINATION. -a. The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. -b. We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan 3D 2.1 Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement. -9. GOVERNING LAW AND JURISDICTION. -a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. -b. Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute. -  -EXHIBIT A -ACCEPTABLE USE POLICY - -Tencent reserves the right to update this Acceptable Use Policy from time to time. -Last modified: November 5, 2024 - -Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan 3D 2.1. You agree not to use Tencent Hunyuan 3D 2.1 or Model Derivatives: -1. Outside the Territory; -2. In any way that violates any applicable national, federal, state, local, international or any other law or regulation; -3. To harm Yourself or others; -4. To repurpose or distribute output from Tencent Hunyuan 3D 2.1 or any Model Derivatives to harm Yourself or others; -5. To override or circumvent the safety guardrails and safeguards We have put in place; -6. For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; -7. To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections; -8. To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement; -9. To intentionally defame, disparage or otherwise harass others; -10. To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems; -11. To generate or disseminate personal identifiable information with the purpose of harming others; -12. To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated; -13. To impersonate another individual without consent, authorization, or legal right; -14. To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance); -15. In a manner that violates or disrespects the social ethics and moral standards of other countries or regions; -16. To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism; -17. For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics; -18. To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; -19. For military purposes; -20. To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices. diff --git a/hy3dpaint/README.md b/hy3dpaint/README.md deleted file mode 100644 index 70d1bdec760bcd741802737d84a930cf150f53b8..0000000000000000000000000000000000000000 --- a/hy3dpaint/README.md +++ /dev/null @@ -1,96 +0,0 @@ -# Hunyuan3D-Paint 2.1 - -Hunyuan3D-Paint 2.1 is a high quality PBR texture generation model for 3D meshes, powered by [RomanTex](https://github.com/oakshy/RomanTex) and [MaterialMVP](https://github.com/ZebinHe/MaterialMVP/). - - -## Quick Inference -You need to manually download the RealESRGAN weight to the `ckpt` folder using the following command: -```bash -wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P ckpt -``` - -Given a 3D mesh `mesh.glb` and a reference image `image.png`, you can run inference using the following code. The result will be saved as `textured_mesh.glb`. - -```bash -python3 demo.py -``` -**Optional arguments in `demo.py`:** - -- `max_num_view` : Maximum number of views, adaptively selected by the model (integer between 6 to 9) - -- `resolution` : Resolution for generated PBR textures (512 or 768) - -**Memory Recommendation:** For `max_num_view=6` and `resolution=512`, we recommend using a GPU with at least **21GB VRAM**. - -## Training - -### Data Prepare -We provide a piece of data in `train_examples` for the overfitting training test. The data structure should be organized as follows: - -``` -train_examples/ -├── examples.json -└── 001/ - ├── render_tex/ # Rendered generated PBR images - │ ├── 000.png # Rendered views (RGB images) - │ ├── 000_albedo.png # Albedo maps for each view - │ ├── 000_mr.png # Metallic-Roughness maps for each view, R and G channels - │ ├── 000_normal.png # Normal maps - │ ├── 000_normal.png # Normal maps - │ ├── 000_pos.png # Position maps - │ ├── 000_pos.png # Position maps - │ ├── 001.png # Additional views... - │ ├── 001_albedo.png - │ ├── 001_mr.png - │ ├── 001_normal.png - │ ├── 001_pos.png - │ └── ... # More views (002, 003, 004, 005, ...) - └── render_cond/ # Rendered reference images (at least two light conditions should be rendered to facilitate consistency loss) - ├── 000_light_AL.png # Light condition 1 (Area Light) - ├── 000_light_ENVMAP.png # Light condition 2 (Environment map) - ├── 000_light_PL.png # Light condition 3 (Point lighting) - ├── 001_light_AL.png - ├── 001_light_ENVMAP.png - ├── 001_light_PL.png - └── ... # More lighting conditions (002-005, ...) -``` - -Each training example contains: -- **render_tex/**: Multi-view renderings with PBR material properties - - Main RGB images (`XXX.png`) - - Albedo maps (`XXX_albedo.png`) - - Metallic-Roughness maps (`XXX_mr.png`) - - Normal maps (`XXX_normal.png/jpg`) - - Position maps (`XXX_pos.png/jpg`) - - Camera transforms (`transforms.json`) -- **render_cond/**: Lighting condition maps for each view - - Ambient lighting (`XXX_light_AL.png`) - - Environment map lighting (`XXX_light_ENVMAP.png`) - - Point lighting (`XXX_light_PL.png`) - -### Launch Training - - -```bash -python3 train.py --base 'cfgs/hunyuan-paint-pbr.yaml' --name overfit --logdir logs/ -``` - -## BibTeX - -If you found Hunyuan3D-Paint 2.1 helpful, please cite our papers: - -```bibtex -@article{feng2025romantex, - title={RomanTex: Decoupling 3D-aware Rotary Positional Embedded Multi-Attention Network for Texture Synthesis}, - author={Feng, Yifei and Yang, Mingxin and Yang, Shuhui and Zhang, Sheng and Yu, Jiaao and Zhao, Zibo and Liu, Yuhong and Jiang, Jie and Guo, Chunchao}, - journal={arXiv preprint arXiv:2503.19011}, - year={2025} -} - -@article{he2025materialmvp, - title={MaterialMVP: Illumination-Invariant Material Generation via Multi-view PBR Diffusion}, - author={He, Zebin and Yang, Mingxin and Yang, Shuhui and Tang, Yixuan and Wang, Tao and Zhang, Kaihao and Chen, Guanying and Liu, Yuhong and Jiang, Jie and Guo, Chunchao and Luo, Wenhan}, - journal={arXiv preprint arXiv:2503.10289}, - year={2025} -} -``` diff --git a/hy3dpaint/cfgs/hunyuan-paint-pbr.yaml b/hy3dpaint/cfgs/hunyuan-paint-pbr.yaml deleted file mode 100644 index 27ed492e128832e2218d260265195b2a9b310d34..0000000000000000000000000000000000000000 --- a/hy3dpaint/cfgs/hunyuan-paint-pbr.yaml +++ /dev/null @@ -1,52 +0,0 @@ -model: - base_learning_rate: 5.0e-05 - target: hunyuanpaintpbr.model.HunyuanPaint - params: - num_view: 6 - view_size: 512 - drop_cond_prob: 0.1 - - noise_in_channels: 12 - - stable_diffusion_config: - pretrained_model_name_or_path: stabilityai/stable-diffusion-2-1 - custom_pipeline: ./hunyuanpaintpbr - - -data: - target: src.data.objaverse_hunyuan.DataModuleFromConfig - params: - batch_size: 1 - num_workers: 4 - train: - - - target: src.data.dataloader.objaverse_loader_forTexturePBR.TextureDataset - params: - num_view: 6 - json_path: train_examples/examples.json - validation: - - - target: src.data.dataloader.objaverse_loader_forTexturePBR.TextureDataset - params: - num_view: 6 - json_path: train_examples/examples.json - -lightning: - modelcheckpoint: - params: - every_n_train_steps: 10000 - save_top_k: -1 - save_last: true - callbacks: {} - - trainer: - benchmark: true - max_epochs: -1 - gradient_clip_val: 1.0 - val_check_interval: 1000 - num_sanity_val_steps: 0 - accumulate_grad_batches: 1 - check_val_every_n_epoch: null # if not set this, validation does not run - -init_control_from: -resume_from: diff --git a/hy3dpaint/convert_utils.py b/hy3dpaint/convert_utils.py deleted file mode 100644 index 7ca66c0b5bae337417e40a0669a780ef94602097..0000000000000000000000000000000000000000 --- a/hy3dpaint/convert_utils.py +++ /dev/null @@ -1,140 +0,0 @@ -import trimesh -import pygltflib -import numpy as np -from PIL import Image -import base64 -import io - - -def combine_metallic_roughness(metallic_path, roughness_path, output_path): - """ - 将metallic和roughness贴图合并为一张贴图 - GLB格式要求metallic在B通道，roughness在G通道 - """ - # 加载贴图 - metallic_img = Image.open(metallic_path).convert("L") # 转为灰度 - roughness_img = Image.open(roughness_path).convert("L") # 转为灰度 - - # 确保尺寸一致 - if metallic_img.size != roughness_img.size: - roughness_img = roughness_img.resize(metallic_img.size) - - # 创建RGB图像 - width, height = metallic_img.size - combined = Image.new("RGB", (width, height)) - - # 转为numpy数组便于操作 - metallic_array = np.array(metallic_img) - roughness_array = np.array(roughness_img) - - # 创建合并的数组 (R, G, B) = (AO, Roughness, Metallic) - combined_array = np.zeros((height, width, 3), dtype=np.uint8) - combined_array[:, :, 0] = 255 # R通道：AO (如果没有AO贴图，设为白色) - combined_array[:, :, 1] = roughness_array # G通道：Roughness - combined_array[:, :, 2] = metallic_array # B通道：Metallic - - # 转回PIL图像并保存 - combined = Image.fromarray(combined_array) - combined.save(output_path) - return output_path - - -def create_glb_with_pbr_materials(obj_path, textures_dict, output_path): - """ - 使用pygltflib创建包含完整PBR材质的GLB文件 - - textures_dict = { - 'albedo': 'path/to/albedo.png', - 'metallic': 'path/to/metallic.png', - 'roughness': 'path/to/roughness.png', - 'normal': 'path/to/normal.png', # 可选 - 'ao': 'path/to/ao.png' # 可选 - } - """ - # 1. 加载OBJ文件 - mesh = trimesh.load(obj_path) - - # 2. 先导出为临时GLB - temp_glb = "temp.glb" - mesh.export(temp_glb) - - # 3. 加载GLB文件进行材质编辑 - gltf = pygltflib.GLTF2().load(temp_glb) - - # 4. 准备纹理数据 - def image_to_data_uri(image_path): - """将图像转换为data URI""" - with open(image_path, "rb") as f: - image_data = f.read() - encoded = base64.b64encode(image_data).decode() - return f"data:image/png;base64,{encoded}" - - # 5. 合并metallic和roughness - if "metallic" in textures_dict and "roughness" in textures_dict: - mr_combined_path = "mr_combined.png" - combine_metallic_roughness(textures_dict["metallic"], textures_dict["roughness"], mr_combined_path) - textures_dict["metallicRoughness"] = mr_combined_path - - # 6. 添加图像到GLTF - images = [] - textures = [] - - texture_mapping = { - "albedo": "baseColorTexture", - "metallicRoughness": "metallicRoughnessTexture", - "normal": "normalTexture", - "ao": "occlusionTexture", - } - - for tex_type, tex_path in textures_dict.items(): - if tex_type in texture_mapping and tex_path: - # 添加图像 - image = pygltflib.Image(uri=image_to_data_uri(tex_path)) - images.append(image) - - # 添加纹理 - texture = pygltflib.Texture(source=len(images) - 1) - textures.append(texture) - - # 7. 创建PBR材质 - pbr_metallic_roughness = pygltflib.PbrMetallicRoughness( - baseColorFactor=[1.0, 1.0, 1.0, 1.0], metallicFactor=1.0, roughnessFactor=1.0 - ) - - # 设置纹理索引 - texture_index = 0 - if "albedo" in textures_dict: - pbr_metallic_roughness.baseColorTexture = pygltflib.TextureInfo(index=texture_index) - texture_index += 1 - - if "metallicRoughness" in textures_dict: - pbr_metallic_roughness.metallicRoughnessTexture = pygltflib.TextureInfo(index=texture_index) - texture_index += 1 - - # 创建材质 - material = pygltflib.Material(name="PBR_Material", pbrMetallicRoughness=pbr_metallic_roughness) - - # 添加法线贴图 - if "normal" in textures_dict: - material.normalTexture = pygltflib.NormalTextureInfo(index=texture_index) - texture_index += 1 - - # 添加AO贴图 - if "ao" in textures_dict: - material.occlusionTexture = pygltflib.OcclusionTextureInfo(index=texture_index) - - # 8. 更新GLTF - gltf.images = images - gltf.textures = textures - gltf.materials = [material] - - # 确保mesh使用材质 - if gltf.meshes: - for primitive in gltf.meshes[0].primitives: - primitive.material = 0 - - # 9. 保存最终GLB - gltf.save(output_path) - print(f"PBR GLB文件已保存: {output_path}") - - diff --git a/hy3dpaint/demo.py b/hy3dpaint/demo.py deleted file mode 100644 index 44fb296526dacb4d82b85ac1228c34c6e7f26e00..0000000000000000000000000000000000000000 --- a/hy3dpaint/demo.py +++ /dev/null @@ -1,35 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig - -try: - from utils.torchvision_fix import apply_fix - - apply_fix() -except ImportError: - print("Warning: torchvision_fix module not found, proceeding without compatibility fix") -except Exception as e: - print(f"Warning: Failed to apply torchvision fix: {e}") - - -if __name__ == "__main__": - - max_num_view = 6 # can be 6 to 9 - resolution = 512 # can be 768 or 512 - - conf = Hunyuan3DPaintConfig(max_num_view, resolution) - paint_pipeline = Hunyuan3DPaintPipeline(conf) - output_mesh_path = paint_pipeline(mesh_path="./assets/case_1/mesh.glb", image_path="./assets/case_1/image.png") - print(f"Output mesh path: {output_mesh_path}") diff --git a/hy3dpaint/hunyuanpaintpbr/__init__.py b/hy3dpaint/hunyuanpaintpbr/__init__.py deleted file mode 100644 index 930d2fc856340ae4d9739925087304b4957f2c7f..0000000000000000000000000000000000000000 --- a/hy3dpaint/hunyuanpaintpbr/__init__.py +++ /dev/null @@ -1,39 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from .pipeline import HunyuanPaintPipeline -from .unet.model import HunyuanPaint -from .unet.modules import ( - Dino_v2, - Basic2p5DTransformerBlock, - ImageProjModel, - UNet2p5DConditionModel, -) -from .unet.attn_processor import ( - PoseRoPEAttnProcessor2_0, - SelfAttnProcessor2_0, - RefAttnProcessor2_0, -) - -__all__ = [ - 'HunyuanPaintPipeline', - 'HunyuanPaint', - 'Dino_v2', - 'Basic2p5DTransformerBlock', - 'ImageProjModel', - 'UNet2p5DConditionModel', - 'PoseRoPEAttnProcessor2_0', - 'SelfAttnProcessor2_0', - 'RefAttnProcessor2_0', -] diff --git a/hy3dpaint/hunyuanpaintpbr/pipeline.py b/hy3dpaint/hunyuanpaintpbr/pipeline.py deleted file mode 100644 index 871dae58e968cf887529a25b8d6fb48da5a97b4b..0000000000000000000000000000000000000000 --- a/hy3dpaint/hunyuanpaintpbr/pipeline.py +++ /dev/null @@ -1,736 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from typing import Any, Dict, Optional -from diffusers.models import AutoencoderKL, UNet2DConditionModel -from diffusers.schedulers import KarrasDiffusionSchedulers - -import numpy -import torch -import torch.utils.checkpoint -import torch.distributed -import numpy as np -import transformers -from PIL import Image -from einops import rearrange -from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection -from typing import Any, Callable, Dict, List, Optional, Union, Tuple - -import diffusers -from diffusers import ( - AutoencoderKL, - DiffusionPipeline, - UNet2DConditionModel, -) -from diffusers.image_processor import VaeImageProcessor - -from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import ( - StableDiffusionPipeline, - retrieve_timesteps, - rescale_noise_cfg, -) - -from diffusers.utils import deprecate -from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback -from diffusers.image_processor import PipelineImageInput -from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput -from .unet.modules import UNet2p5DConditionModel -from .unet.attn_processor import SelfAttnProcessor2_0, RefAttnProcessor2_0, PoseRoPEAttnProcessor2_0 - -__all__ = [ - "HunyuanPaintPipeline", - "UNet2p5DConditionModel", - "SelfAttnProcessor2_0", - "RefAttnProcessor2_0", - "PoseRoPEAttnProcessor2_0", -] - - -def to_rgb_image(maybe_rgba: Image.Image): - if maybe_rgba.mode == "RGB": - return maybe_rgba - elif maybe_rgba.mode == "RGBA": - rgba = maybe_rgba - img = numpy.random.randint(127, 128, size=[rgba.size[1], rgba.size[0], 3], dtype=numpy.uint8) - img = Image.fromarray(img, "RGB") - img.paste(rgba, mask=rgba.getchannel("A")) - return img - else: - raise ValueError("Unsupported image type.", maybe_rgba.mode) - - -class HunyuanPaintPipeline(StableDiffusionPipeline): - - """Custom pipeline for multiview PBR texture generation. - - Extends Stable Diffusion with: - - Material-specific conditioning - - Multiview processing - - Position-aware attention - - 2.5D UNet integration - """ - - def __init__( - self, - vae: AutoencoderKL, - text_encoder: CLIPTextModel, - tokenizer: CLIPTokenizer, - unet: UNet2DConditionModel, - scheduler: KarrasDiffusionSchedulers, - feature_extractor: CLIPImageProcessor, - safety_checker=None, - use_torch_compile=False, - ): - DiffusionPipeline.__init__(self) - - safety_checker = None - self.register_modules( - vae=torch.compile(vae) if use_torch_compile else vae, - text_encoder=text_encoder, - tokenizer=tokenizer, - unet=unet, - scheduler=scheduler, - safety_checker=safety_checker, - feature_extractor=torch.compile(feature_extractor) if use_torch_compile else feature_extractor, - ) - - self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) - self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) - - if isinstance(self.unet, UNet2DConditionModel): - self.unet = UNet2p5DConditionModel(self.unet, None, self.scheduler) - - def eval(self): - self.unet.eval() - self.vae.eval() - - def set_pbr_settings(self, pbr_settings: List[str]): - self.pbr_settings = pbr_settings - - def set_learned_parameters(self): - - """Configures parameter freezing strategy. - - Freezes: - - Standard attention layers - - Dual-stream reference UNet - - Unfreezes: - - Material-specific parameters - - DINO integration components - """ - - freezed_names = ["attn1", "unet_dual"] - added_learned_names = ["albedo", "mr", "dino"] - - for name, params in self.unet.named_parameters(): - if any(freeze_name in name for freeze_name in freezed_names) and all( - learned_name not in name for learned_name in added_learned_names - ): - params.requires_grad = False - else: - params.requires_grad = True - - def prepare(self): - if isinstance(self.unet, UNet2DConditionModel): - self.unet = UNet2p5DConditionModel(self.unet, None, self.scheduler).eval() - - @torch.no_grad() - def encode_images(self, images): - - """Encodes multiview image batches into latent space. - - Args: - images: Input images [B, N_views, C, H, W] - - Returns: - torch.Tensor: Latent representations [B, N_views, C, H_latent, W_latent] - """ - - B = images.shape[0] - images = rearrange(images, "b n c h w -> (b n) c h w") - - dtype = next(self.vae.parameters()).dtype - images = (images - 0.5) * 2.0 - posterior = self.vae.encode(images.to(dtype)).latent_dist - latents = posterior.sample() * self.vae.config.scaling_factor - - latents = rearrange(latents, "(b n) c h w -> b n c h w", b=B) - return latents - - @torch.no_grad() - def __call__( - self, - images=None, - prompt=None, - negative_prompt="watermark, ugly, deformed, noisy, blurry, low contrast", - *args, - num_images_per_prompt: Optional[int] = 1, - guidance_scale=3.0, - output_type: Optional[str] = "pil", - width=512, - height=512, - num_inference_steps=15, - return_dict=True, - sync_condition=None, - **cached_condition, - ): - - """Main generation method for multiview PBR textures. - - Steps: - 1. Input validation and preparation - 2. Reference image encoding - 3. Condition processing (normal/position maps) - 4. Prompt embedding setup - 5. Classifier-free guidance preparation - 6. Diffusion sampling loop - - Args: - images: List of reference PIL images - prompt: Text prompt (overridden by learned embeddings) - cached_condition: Dictionary containing: - - images_normal: Normal maps (PIL or tensor) - - images_position: Position maps (PIL or tensor) - - Returns: - List[PIL.Image]: Generated multiview PBR textures - """ - - self.prepare() - if images is None: - raise ValueError("Inputting embeddings not supported for this pipeline. Please pass an image.") - assert not isinstance(images, torch.Tensor) - - if not isinstance(images, List): - images = [images] - - images = [to_rgb_image(image) for image in images] - images_vae = [torch.tensor(np.array(image) / 255.0) for image in images] - images_vae = [image_vae.unsqueeze(0).permute(0, 3, 1, 2).unsqueeze(0) for image_vae in images_vae] - images_vae = torch.cat(images_vae, dim=1) - images_vae = images_vae.to(device=self.vae.device, dtype=self.unet.dtype) - - batch_size = images_vae.shape[0] - N_ref = images_vae.shape[1] - - assert batch_size == 1 - assert num_images_per_prompt == 1 - - if self.unet.use_ra: - ref_latents = self.encode_images(images_vae) - cached_condition["ref_latents"] = ref_latents - - def convert_pil_list_to_tensor(images): - bg_c = [1.0, 1.0, 1.0] - images_tensor = [] - for batch_imgs in images: - view_imgs = [] - for pil_img in batch_imgs: - img = numpy.asarray(pil_img, dtype=numpy.float32) / 255.0 - if img.shape[2] > 3: - alpha = img[:, :, 3:] - img = img[:, :, :3] * alpha + bg_c * (1 - alpha) - img = torch.from_numpy(img).permute(2, 0, 1).unsqueeze(0).contiguous().half().to("cuda") - view_imgs.append(img) - view_imgs = torch.cat(view_imgs, dim=0) - images_tensor.append(view_imgs.unsqueeze(0)) - - images_tensor = torch.cat(images_tensor, dim=0) - return images_tensor - - if "images_normal" in cached_condition: - if isinstance(cached_condition["images_normal"], List): - cached_condition["images_normal"] = convert_pil_list_to_tensor(cached_condition["images_normal"]) - - cached_condition["embeds_normal"] = self.encode_images(cached_condition["images_normal"]) - - if "images_position" in cached_condition: - - if isinstance(cached_condition["images_position"], List): - cached_condition["images_position"] = convert_pil_list_to_tensor(cached_condition["images_position"]) - - cached_condition["position_maps"] = cached_condition["images_position"] - cached_condition["embeds_position"] = self.encode_images(cached_condition["images_position"]) - - if self.unet.use_learned_text_clip: - - all_shading_tokens = [] - for token in self.unet.pbr_setting: - all_shading_tokens.append( - getattr(self.unet, f"learned_text_clip_{token}").unsqueeze(dim=0).repeat(batch_size, 1, 1) - ) - prompt_embeds = torch.stack(all_shading_tokens, dim=1) - negative_prompt_embeds = torch.stack(all_shading_tokens, dim=1) - # negative_prompt_embeds = torch.zeros_like(prompt_embeds) - - else: - if prompt is None: - prompt = "high quality" - if isinstance(prompt, str): - prompt = [prompt for _ in range(batch_size)] - device = self._execution_device - prompt_embeds, _ = self.encode_prompt( - prompt, device=device, num_images_per_prompt=num_images_per_prompt, do_classifier_free_guidance=False - ) - - if isinstance(negative_prompt, str): - negative_prompt = [negative_prompt for _ in range(batch_size)] - if negative_prompt is not None: - negative_prompt_embeds, _ = self.encode_prompt( - negative_prompt, - device=device, - num_images_per_prompt=num_images_per_prompt, - do_classifier_free_guidance=False, - ) - else: - negative_prompt_embeds = torch.zeros_like(prompt_embeds) - - if guidance_scale > 1: - if self.unet.use_ra: - cached_condition["ref_latents"] = cached_condition["ref_latents"].repeat( - 3, *([1] * (cached_condition["ref_latents"].dim() - 1)) - ) - cached_condition["ref_scale"] = torch.as_tensor([0.0, 1.0, 1.0]).to(cached_condition["ref_latents"]) - - if self.unet.use_dino: - zero_states = torch.zeros_like(cached_condition["dino_hidden_states"]) - cached_condition["dino_hidden_states"] = torch.cat( - [zero_states, zero_states, cached_condition["dino_hidden_states"]] - ) - - del zero_states - if "embeds_normal" in cached_condition: - cached_condition["embeds_normal"] = cached_condition["embeds_normal"].repeat( - 3, *([1] * (cached_condition["embeds_normal"].dim() - 1)) - ) - - if "embeds_position" in cached_condition: - cached_condition["embeds_position"] = cached_condition["embeds_position"].repeat( - 3, *([1] * (cached_condition["embeds_position"].dim() - 1)) - ) - - if "position_maps" in cached_condition: - cached_condition["position_maps"] = cached_condition["position_maps"].repeat( - 3, *([1] * (cached_condition["position_maps"].dim() - 1)) - ) - - images = self.denoise( - None, - *args, - cross_attention_kwargs=None, - guidance_scale=guidance_scale, - num_images_per_prompt=num_images_per_prompt, - prompt_embeds=prompt_embeds, - negative_prompt_embeds=negative_prompt_embeds, - num_inference_steps=num_inference_steps, - output_type=output_type, - width=width, - height=height, - return_dict=return_dict, - **cached_condition, - ) - - return images - - def denoise( - self, - prompt: Union[str, List[str]] = None, - height: Optional[int] = None, - width: Optional[int] = None, - num_inference_steps: int = 50, - timesteps: List[int] = None, - sigmas: List[float] = None, - guidance_scale: float = 7.5, - negative_prompt: Optional[Union[str, List[str]]] = None, - num_images_per_prompt: Optional[int] = 1, - eta: float = 0.0, - generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, - latents: Optional[torch.Tensor] = None, - prompt_embeds: Optional[torch.Tensor] = None, - negative_prompt_embeds: Optional[torch.Tensor] = None, - ip_adapter_image: Optional[PipelineImageInput] = None, - ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, - output_type: Optional[str] = "pil", - return_dict: bool = True, - cross_attention_kwargs: Optional[Dict[str, Any]] = None, - guidance_rescale: float = 0.0, - clip_skip: Optional[int] = None, - callback_on_step_end: Optional[ - Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks] - ] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - **kwargs, - ): - r""" - The call function to the pipeline for generation. - - Args: - prompt (`str` or `List[str]`, *optional*): - The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. - height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): - The height in pixels of the generated image. - width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): - The width in pixels of the generated image. - num_inference_steps (`int`, *optional*, defaults to 50): - The number of denoising steps. More denoising steps usually lead to a higher quality image at the - expense of slower inference. - timesteps (`List[int]`, *optional*): - Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument - in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is - passed will be used. Must be in descending order. - sigmas (`List[float]`, *optional*): - Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in - their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed - will be used. - guidance_scale (`float`, *optional*, defaults to 7.5): - A higher guidance scale value encourages the model to generate images closely linked to the text - `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. - negative_prompt (`str` or `List[str]`, *optional*): - The prompt or prompts to guide what to not include in image generation. If not defined, you need to - pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). - num_images_per_prompt (`int`, *optional*, defaults to 1): - The number of images to generate per prompt. - eta (`float`, *optional*, defaults to 0.0): - Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies - to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. - generator (`torch.Generator` or `List[torch.Generator]`, *optional*): - A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make - generation deterministic. - latents (`torch.Tensor`, *optional*): - Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image - generation. Can be used to tweak the same generation with different prompts. If not provided, a latents - tensor is generated by sampling using the supplied random `generator`. - prompt_embeds (`torch.Tensor`, *optional*): - Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not - provided, text embeddings are generated from the `prompt` input argument. - negative_prompt_embeds (`torch.Tensor`, *optional*): - Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If - not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. - ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. - ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): - Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of - IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should - contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not - provided, embeddings are computed from the `ip_adapter_image` input argument. - output_type (`str`, *optional*, defaults to `"pil"`): - The output format of the generated image. Choose between `PIL.Image` or `np.array`. - return_dict (`bool`, *optional*, defaults to `True`): - Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a - plain tuple. - cross_attention_kwargs (`dict`, *optional*): - A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in - [`self.processor`] - (https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - guidance_rescale (`float`, *optional*, defaults to 0.0): - Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are - Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when - using zero terminal SNR. - clip_skip (`int`, *optional*): - Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that - the output of the pre-final layer will be used for computing the prompt embeddings. - callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): - A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of - each denoising step during the inference. with the following arguments: `callback_on_step_end(self: - DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a - list of all tensors as specified by `callback_on_step_end_tensor_inputs`. - callback_on_step_end_tensor_inputs (`List`, *optional*): - The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list - will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the - `._callback_tensor_inputs` attribute of your pipeline class. - - Examples: - - Returns: - [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: - If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, - otherwise a `tuple` is returned where the first element is a list with the generated images and the - second element is a list of `bool`s indicating whether the corresponding generated image contains - "not-safe-for-work" (nsfw) content. - - Core denoising procedure for multiview PBR texture generation. - - Handles the complete diffusion process including: - - Input validation and preparation - - Timestep scheduling - - Latent noise initialization - - Iterative denoising with specialized guidance - - Output decoding and post-processing - - Key innovations: - 1. Triple-batch classifier-free guidance: - - Negative (unconditional) - - Reference-conditioned - - Full-conditioned - 2. View-dependent guidance scaling: - - Adjusts influence based on camera azimuth - 3. PBR-aware latent organization: - - Maintains material/view separation throughout - 4. Optimized VRAM management: - - Selective tensor reshaping - - Processing Stages: - 1. Setup & Validation: Configures pipeline components and validates inputs - 2. Prompt Encoding: Processes text/material conditioning - 3. Latent Initialization: Prepares noise for denoising process - 4. Iterative Denoising: - a) Scales and organizes latent variables - b) Predicts noise at current timestep - c) Applies view-dependent guidance - d) Computes previous latent state - 5. Output Decoding: Converts latents to final images - 6. Cleanup: Releases resources and formats output - - """ - - callback = kwargs.pop("callback", None) - callback_steps = kwargs.pop("callback_steps", None) - - # open cache - kwargs["cache"] = {} - - if callback is not None: - deprecate( - "callback", - "1.0.0", - "Passing `callback` as an input argument to `__call__` is deprecated," - "consider using `callback_on_step_end`", - ) - if callback_steps is not None: - deprecate( - "callback_steps", - "1.0.0", - "Passing `callback` as an input argument to `__call__` is deprecated," - "consider using `callback_on_step_end`", - ) - - if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): - callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs - - # 0. Default height and width to unet - height = height or self.unet.config.sample_size * self.vae_scale_factor - width = width or self.unet.config.sample_size * self.vae_scale_factor - # to deal with lora scaling and other possible forward hooks - - # 1. Check inputs. Raise error if not correct - self.check_inputs( - prompt, - height, - width, - callback_steps, - negative_prompt, - prompt_embeds, - negative_prompt_embeds, - ip_adapter_image, - ip_adapter_image_embeds, - callback_on_step_end_tensor_inputs, - ) - - self._guidance_scale = guidance_scale - self._guidance_rescale = guidance_rescale - self._clip_skip = clip_skip - self._cross_attention_kwargs = cross_attention_kwargs - self._interrupt = False - - # 2. Define call parameters - if prompt is not None and isinstance(prompt, str): - batch_size = 1 - elif prompt is not None and isinstance(prompt, list): - batch_size = len(prompt) - else: - batch_size = prompt_embeds.shape[0] - - device = self._execution_device - - # 3. Encode input prompt - lora_scale = self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None - - """ - prompt_embeds, negative_prompt_embeds = self.encode_prompt( - prompt, - device, - num_images_per_prompt, - self.do_classifier_free_guidance, - negative_prompt, - prompt_embeds=prompt_embeds, - negative_prompt_embeds=negative_prompt_embeds, - lora_scale=lora_scale, - clip_skip=self.clip_skip, - )' - """ - - # For classifier free guidance, we need to do two forward passes. - # Here we concatenate the unconditional and text embeddings into a single batch - # to avoid doing two forward passes - if self.do_classifier_free_guidance: - # prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds]) - - if ip_adapter_image is not None or ip_adapter_image_embeds is not None: - image_embeds = self.prepare_ip_adapter_image_embeds( - ip_adapter_image, - ip_adapter_image_embeds, - device, - batch_size * num_images_per_prompt, - self.do_classifier_free_guidance, - ) - - # 4. Prepare timesteps - timesteps, num_inference_steps = retrieve_timesteps( - self.scheduler, num_inference_steps, device, timesteps, sigmas - ) - assert num_images_per_prompt == 1 - # 5. Prepare latent variables - n_pbr = len(self.unet.pbr_setting) - num_channels_latents = self.unet.config.in_channels - latents = self.prepare_latents( - batch_size * kwargs["num_in_batch"] * n_pbr, # num_images_per_prompt, - num_channels_latents, - height, - width, - prompt_embeds.dtype, - device, - generator, - latents, - ) - - # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline - extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) - - # 6.1 Add image embeds for IP-Adapter - added_cond_kwargs = ( - {"image_embeds": image_embeds} - if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) - else None - ) - - # 6.2 Optionally get Guidance Scale Embedding - timestep_cond = None - if self.unet.config.time_cond_proj_dim is not None: - guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) - timestep_cond = self.get_guidance_scale_embedding( - guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim - ).to(device=device, dtype=latents.dtype) - - # 7. Denoising loop - num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order - self._num_timesteps = len(timesteps) - with self.progress_bar(total=num_inference_steps) as progress_bar: - for i, t in enumerate(timesteps): - if self.interrupt: - continue - - # expand the latents if we are doing classifier free guidance - latents = rearrange( - latents, "(b n_pbr n) c h w -> b n_pbr n c h w", n=kwargs["num_in_batch"], n_pbr=n_pbr - ) - # latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents - latent_model_input = latents.repeat(3, 1, 1, 1, 1, 1) if self.do_classifier_free_guidance else latents - latent_model_input = rearrange(latent_model_input, "b n_pbr n c h w -> (b n_pbr n) c h w") - latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) - latent_model_input = rearrange( - latent_model_input, "(b n_pbr n) c h w ->b n_pbr n c h w", n=kwargs["num_in_batch"], n_pbr=n_pbr - ) - - # predict the noise residual - - noise_pred = self.unet( - latent_model_input, - t, - encoder_hidden_states=prompt_embeds, - timestep_cond=timestep_cond, - cross_attention_kwargs=self.cross_attention_kwargs, - added_cond_kwargs=added_cond_kwargs, - return_dict=False, - **kwargs, - )[0] - latents = rearrange(latents, "b n_pbr n c h w -> (b n_pbr n) c h w") - # perform guidance - if self.do_classifier_free_guidance: - # noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) - # noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) - noise_pred_uncond, noise_pred_ref, noise_pred_full = noise_pred.chunk(3) - - if "camera_azims" in kwargs.keys(): - camera_azims = kwargs["camera_azims"] - else: - camera_azims = [0] * kwargs["num_in_batch"] - - def cam_mapping(azim): - if azim < 90 and azim >= 0: - return float(azim) / 90.0 + 1 - elif azim >= 90 and azim < 330: - return 2.0 - else: - return -float(azim) / 90.0 + 5.0 - - view_scale_tensor = ( - torch.from_numpy(np.asarray([cam_mapping(azim) for azim in camera_azims])) - .unsqueeze(0) - .repeat(n_pbr, 1) - .view(-1) - .to(noise_pred_uncond)[:, None, None, None] - ) - noise_pred = noise_pred_uncond + self.guidance_scale * view_scale_tensor * ( - noise_pred_ref - noise_pred_uncond - ) - noise_pred += self.guidance_scale * view_scale_tensor * (noise_pred_full - noise_pred_ref) - - if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: - # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf - noise_pred = rescale_noise_cfg(noise_pred, noise_pred_ref, guidance_rescale=self.guidance_rescale) - - # compute the previous noisy sample x_t -> x_t-1 - latents = self.scheduler.step( - noise_pred, t, latents[:, :num_channels_latents, :, :], **extra_step_kwargs, return_dict=False - )[0] - - if callback_on_step_end is not None: - callback_kwargs = {} - for k in callback_on_step_end_tensor_inputs: - callback_kwargs[k] = locals()[k] - callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) - - latents = callback_outputs.pop("latents", latents) - prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) - negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) - - # call the callback, if provided - if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): - progress_bar.update() - if callback is not None and i % callback_steps == 0: - step_idx = i // getattr(self.scheduler, "order", 1) - callback(step_idx, t, latents) - - if not output_type == "latent": - image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0] - image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) - else: - image = latents - has_nsfw_concept = None - - if has_nsfw_concept is None: - do_denormalize = [True] * image.shape[0] - else: - do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] - - image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) - - # Offload all models - self.maybe_free_model_hooks() - - if not return_dict: - return (image, has_nsfw_concept) - - return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/hy3dpaint/hunyuanpaintpbr/unet/attn_processor.py b/hy3dpaint/hunyuanpaintpbr/unet/attn_processor.py deleted file mode 100644 index 37ebde0f6f30230351f757e050e3b86fd8a06402..0000000000000000000000000000000000000000 --- a/hy3dpaint/hunyuanpaintpbr/unet/attn_processor.py +++ /dev/null @@ -1,839 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import torch -import torch.nn as nn -import torch.nn.functional as F -from typing import Optional, Dict, Tuple, Union, Literal, List, Callable -from einops import rearrange -from diffusers.utils import deprecate -from diffusers.models.attention_processor import Attention, AttnProcessor - - -class AttnUtils: - """ - Shared utility functions for attention processing. - - This class provides common operations used across different attention processors - to eliminate code duplication and improve maintainability. - """ - - @staticmethod - def check_pytorch_compatibility(): - """ - Check PyTorch compatibility for scaled_dot_product_attention. - - Raises: - ImportError: If PyTorch version doesn't support scaled_dot_product_attention - """ - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - @staticmethod - def handle_deprecation_warning(args, kwargs): - """ - Handle deprecation warning for the 'scale' argument. - - Args: - args: Positional arguments passed to attention processor - kwargs: Keyword arguments passed to attention processor - """ - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = ( - "The `scale` argument is deprecated and will be ignored." - "Please remove it, as passing it will raise an error in the future." - "`scale` should directly be passed while calling the underlying pipeline component" - "i.e., via `cross_attention_kwargs`." - ) - deprecate("scale", "1.0.0", deprecation_message) - - @staticmethod - def prepare_hidden_states( - hidden_states, attn, temb, spatial_norm_attr="spatial_norm", group_norm_attr="group_norm" - ): - """ - Common preprocessing of hidden states for attention computation. - - Args: - hidden_states: Input hidden states tensor - attn: Attention module instance - temb: Optional temporal embedding tensor - spatial_norm_attr: Attribute name for spatial normalization - group_norm_attr: Attribute name for group normalization - - Returns: - Tuple of (processed_hidden_states, residual, input_ndim, shape_info) - """ - residual = hidden_states - - spatial_norm = getattr(attn, spatial_norm_attr, None) - if spatial_norm is not None: - hidden_states = spatial_norm(hidden_states, temb) - - input_ndim = hidden_states.ndim - - if input_ndim == 4: - batch_size, channel, height, width = hidden_states.shape - hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) - else: - batch_size, channel, height, width = None, None, None, None - - group_norm = getattr(attn, group_norm_attr, None) - if group_norm is not None: - hidden_states = group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) - - return hidden_states, residual, input_ndim, (batch_size, channel, height, width) - - @staticmethod - def prepare_attention_mask(attention_mask, attn, sequence_length, batch_size): - """ - Prepare attention mask for scaled_dot_product_attention. - - Args: - attention_mask: Input attention mask tensor or None - attn: Attention module instance - sequence_length: Length of the sequence - batch_size: Batch size - - Returns: - Prepared attention mask tensor reshaped for multi-head attention - """ - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - return attention_mask - - @staticmethod - def reshape_qkv_for_attention(tensor, batch_size, attn_heads, head_dim): - """ - Reshape Q/K/V tensors for multi-head attention computation. - - Args: - tensor: Input tensor to reshape - batch_size: Batch size - attn_heads: Number of attention heads - head_dim: Dimension per attention head - - Returns: - Reshaped tensor with shape [batch_size, attn_heads, seq_len, head_dim] - """ - return tensor.view(batch_size, -1, attn_heads, head_dim).transpose(1, 2) - - @staticmethod - def apply_norms(query, key, norm_q, norm_k): - """ - Apply Q/K normalization layers if available. - - Args: - query: Query tensor - key: Key tensor - norm_q: Query normalization layer (optional) - norm_k: Key normalization layer (optional) - - Returns: - Tuple of (normalized_query, normalized_key) - """ - if norm_q is not None: - query = norm_q(query) - if norm_k is not None: - key = norm_k(key) - return query, key - - @staticmethod - def finalize_output(hidden_states, input_ndim, shape_info, attn, residual, to_out): - """ - Common output processing including projection, dropout, reshaping, and residual connection. - - Args: - hidden_states: Processed hidden states from attention - input_ndim: Original input tensor dimensions - shape_info: Tuple containing original shape information - attn: Attention module instance - residual: Residual connection tensor - to_out: Output projection layers [linear, dropout] - - Returns: - Final output tensor after all processing steps - """ - batch_size, channel, height, width = shape_info - - # Apply output projection and dropout - hidden_states = to_out[0](hidden_states) - hidden_states = to_out[1](hidden_states) - - # Reshape back if needed - if input_ndim == 4: - hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # Apply residual connection - if attn.residual_connection: - hidden_states = hidden_states + residual - - # Apply rescaling - hidden_states = hidden_states / attn.rescale_output_factor - return hidden_states - - -# Base class for attention processors (eliminating initialization duplication) -class BaseAttnProcessor(nn.Module): - """ - Base class for attention processors with common initialization. - - This base class provides shared parameter initialization and module registration - functionality to reduce code duplication across different attention processor types. - """ - - def __init__( - self, - query_dim: int, - pbr_setting: List[str] = ["albedo", "mr"], - cross_attention_dim: Optional[int] = None, - heads: int = 8, - kv_heads: Optional[int] = None, - dim_head: int = 64, - dropout: float = 0.0, - bias: bool = False, - upcast_attention: bool = False, - upcast_softmax: bool = False, - cross_attention_norm: Optional[str] = None, - cross_attention_norm_num_groups: int = 32, - qk_norm: Optional[str] = None, - added_kv_proj_dim: Optional[int] = None, - added_proj_bias: Optional[bool] = True, - norm_num_groups: Optional[int] = None, - spatial_norm_dim: Optional[int] = None, - out_bias: bool = True, - scale_qk: bool = True, - only_cross_attention: bool = False, - eps: float = 1e-5, - rescale_output_factor: float = 1.0, - residual_connection: bool = False, - _from_deprecated_attn_block: bool = False, - processor: Optional["AttnProcessor"] = None, - out_dim: int = None, - out_context_dim: int = None, - context_pre_only=None, - pre_only=False, - elementwise_affine: bool = True, - is_causal: bool = False, - **kwargs, - ): - """ - Initialize base attention processor with common parameters. - - Args: - query_dim: Dimension of query features - pbr_setting: List of PBR material types to process (e.g., ["albedo", "mr"]) - cross_attention_dim: Dimension of cross-attention features (optional) - heads: Number of attention heads - kv_heads: Number of key-value heads for grouped query attention (optional) - dim_head: Dimension per attention head - dropout: Dropout rate - bias: Whether to use bias in linear projections - upcast_attention: Whether to upcast attention computation to float32 - upcast_softmax: Whether to upcast softmax computation to float32 - cross_attention_norm: Type of cross-attention normalization (optional) - cross_attention_norm_num_groups: Number of groups for cross-attention norm - qk_norm: Type of query-key normalization (optional) - added_kv_proj_dim: Dimension for additional key-value projections (optional) - added_proj_bias: Whether to use bias in additional projections - norm_num_groups: Number of groups for normalization (optional) - spatial_norm_dim: Dimension for spatial normalization (optional) - out_bias: Whether to use bias in output projection - scale_qk: Whether to scale query-key products - only_cross_attention: Whether to only perform cross-attention - eps: Small epsilon value for numerical stability - rescale_output_factor: Factor to rescale output values - residual_connection: Whether to use residual connections - _from_deprecated_attn_block: Flag for deprecated attention blocks - processor: Optional attention processor instance - out_dim: Output dimension (optional) - out_context_dim: Output context dimension (optional) - context_pre_only: Whether to only process context in pre-processing - pre_only: Whether to only perform pre-processing - elementwise_affine: Whether to use element-wise affine transformations - is_causal: Whether to use causal attention masking - **kwargs: Additional keyword arguments - """ - super().__init__() - AttnUtils.check_pytorch_compatibility() - - # Store common attributes - self.pbr_setting = pbr_setting - self.n_pbr_tokens = len(self.pbr_setting) - self.inner_dim = out_dim if out_dim is not None else dim_head * heads - self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads - self.query_dim = query_dim - self.use_bias = bias - self.is_cross_attention = cross_attention_dim is not None - self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim - self.upcast_attention = upcast_attention - self.upcast_softmax = upcast_softmax - self.rescale_output_factor = rescale_output_factor - self.residual_connection = residual_connection - self.dropout = dropout - self.fused_projections = False - self.out_dim = out_dim if out_dim is not None else query_dim - self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim - self.context_pre_only = context_pre_only - self.pre_only = pre_only - self.is_causal = is_causal - self._from_deprecated_attn_block = _from_deprecated_attn_block - self.scale_qk = scale_qk - self.scale = dim_head**-0.5 if self.scale_qk else 1.0 - self.heads = out_dim // dim_head if out_dim is not None else heads - self.sliceable_head_dim = heads - self.added_kv_proj_dim = added_kv_proj_dim - self.only_cross_attention = only_cross_attention - self.added_proj_bias = added_proj_bias - - # Validation - if self.added_kv_proj_dim is None and self.only_cross_attention: - raise ValueError( - "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None." - "Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." - ) - - def register_pbr_modules(self, module_types: List[str], **kwargs): - """ - Generic PBR module registration to eliminate code repetition. - - Dynamically registers PyTorch modules for different PBR material types - based on the specified module types and PBR settings. - - Args: - module_types: List of module types to register ("qkv", "v_only", "out", "add_kv") - **kwargs: Additional arguments for module configuration - """ - for pbr_token in self.pbr_setting: - if pbr_token == "albedo": - continue - - for module_type in module_types: - if module_type == "qkv": - self.register_module( - f"to_q_{pbr_token}", nn.Linear(self.query_dim, self.inner_dim, bias=self.use_bias) - ) - self.register_module( - f"to_k_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias) - ) - self.register_module( - f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias) - ) - elif module_type == "v_only": - self.register_module( - f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias) - ) - elif module_type == "out": - if not self.pre_only: - self.register_module( - f"to_out_{pbr_token}", - nn.ModuleList( - [ - nn.Linear(self.inner_dim, self.out_dim, bias=kwargs.get("out_bias", True)), - nn.Dropout(self.dropout), - ] - ), - ) - else: - self.register_module(f"to_out_{pbr_token}", None) - elif module_type == "add_kv": - if self.added_kv_proj_dim is not None: - self.register_module( - f"add_k_proj_{pbr_token}", - nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias), - ) - self.register_module( - f"add_v_proj_{pbr_token}", - nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias), - ) - else: - self.register_module(f"add_k_proj_{pbr_token}", None) - self.register_module(f"add_v_proj_{pbr_token}", None) - - -# Rotary Position Embedding utilities (specialized for PoseRoPE) -class RotaryEmbedding: - """ - Rotary position embedding utilities for 3D spatial attention. - - Provides functions to compute and apply rotary position embeddings (RoPE) - for 1D, 3D spatial coordinates used in 3D-aware attention mechanisms. - """ - - @staticmethod - def get_1d_rotary_pos_embed(dim: int, pos: torch.Tensor, theta: float = 10000.0, linear_factor=1.0, ntk_factor=1.0): - """ - Compute 1D rotary position embeddings. - - Args: - dim: Embedding dimension (must be even) - pos: Position tensor - theta: Base frequency for rotary embeddings - linear_factor: Linear scaling factor - ntk_factor: NTK (Neural Tangent Kernel) scaling factor - - Returns: - Tuple of (cos_embeddings, sin_embeddings) - """ - assert dim % 2 == 0 - theta = theta * ntk_factor - freqs = ( - 1.0 - / (theta ** (torch.arange(0, dim, 2, dtype=pos.dtype, device=pos.device)[: (dim // 2)] / dim)) - / linear_factor - ) - freqs = torch.outer(pos, freqs) - freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float() - freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float() - return freqs_cos, freqs_sin - - @staticmethod - def get_3d_rotary_pos_embed(position, embed_dim, voxel_resolution, theta: int = 10000): - """ - Compute 3D rotary position embeddings for spatial coordinates. - - Args: - position: 3D position tensor with shape [..., 3] - embed_dim: Embedding dimension - voxel_resolution: Resolution of the voxel grid - theta: Base frequency for rotary embeddings - - Returns: - Tuple of (cos_embeddings, sin_embeddings) for 3D positions - """ - assert position.shape[-1] == 3 - dim_xy = embed_dim // 8 * 3 - dim_z = embed_dim // 8 * 2 - - grid = torch.arange(voxel_resolution, dtype=torch.float32, device=position.device) - freqs_xy = RotaryEmbedding.get_1d_rotary_pos_embed(dim_xy, grid, theta=theta) - freqs_z = RotaryEmbedding.get_1d_rotary_pos_embed(dim_z, grid, theta=theta) - - xy_cos, xy_sin = freqs_xy - z_cos, z_sin = freqs_z - - embed_flattn = position.view(-1, position.shape[-1]) - x_cos = xy_cos[embed_flattn[:, 0], :] - x_sin = xy_sin[embed_flattn[:, 0], :] - y_cos = xy_cos[embed_flattn[:, 1], :] - y_sin = xy_sin[embed_flattn[:, 1], :] - z_cos = z_cos[embed_flattn[:, 2], :] - z_sin = z_sin[embed_flattn[:, 2], :] - - cos = torch.cat((x_cos, y_cos, z_cos), dim=-1) - sin = torch.cat((x_sin, y_sin, z_sin), dim=-1) - - cos = cos.view(*position.shape[:-1], embed_dim) - sin = sin.view(*position.shape[:-1], embed_dim) - return cos, sin - - @staticmethod - def apply_rotary_emb(x: torch.Tensor, freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]]): - """ - Apply rotary position embeddings to input tensor. - - Args: - x: Input tensor to apply rotary embeddings to - freqs_cis: Tuple of (cos_embeddings, sin_embeddings) or single tensor - - Returns: - Tensor with rotary position embeddings applied - """ - cos, sin = freqs_cis - cos, sin = cos.to(x.device), sin.to(x.device) - cos = cos.unsqueeze(1) - sin = sin.unsqueeze(1) - - x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) - x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3) - - out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype) - return out - - -# Core attention processing logic (eliminating major duplication) -class AttnCore: - """ - Core attention processing logic shared across processors. - - This class provides the fundamental attention computation pipeline - that can be reused across different attention processor implementations. - """ - - @staticmethod - def process_attention_base( - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - get_qkv_fn: Callable = None, - apply_rope_fn: Optional[Callable] = None, - **kwargs, - ): - """ - Generic attention processing core shared across different processors. - - This function implements the common attention computation pipeline including: - 1. Hidden state preprocessing - 2. Attention mask preparation - 3. Q/K/V computation via provided function - 4. Tensor reshaping for multi-head attention - 5. Optional normalization and RoPE application - 6. Scaled dot-product attention computation - - Args: - attn: Attention module instance - hidden_states: Input hidden states tensor - encoder_hidden_states: Optional encoder hidden states for cross-attention - attention_mask: Optional attention mask tensor - temb: Optional temporal embedding tensor - get_qkv_fn: Function to compute Q, K, V tensors - apply_rope_fn: Optional function to apply rotary position embeddings - **kwargs: Additional keyword arguments passed to subfunctions - - Returns: - Tuple containing (attention_output, residual, input_ndim, shape_info, - batch_size, num_heads, head_dim) - """ - # Prepare hidden states - hidden_states, residual, input_ndim, shape_info = AttnUtils.prepare_hidden_states(hidden_states, attn, temb) - - batch_size, sequence_length, _ = ( - hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape - ) - - # Prepare attention mask - attention_mask = AttnUtils.prepare_attention_mask(attention_mask, attn, sequence_length, batch_size) - - # Get Q, K, V - if encoder_hidden_states is None: - encoder_hidden_states = hidden_states - elif attn.norm_cross: - encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) - - query, key, value = get_qkv_fn(attn, hidden_states, encoder_hidden_states, **kwargs) - - # Reshape for attention - inner_dim = key.shape[-1] - head_dim = inner_dim // attn.heads - - query = AttnUtils.reshape_qkv_for_attention(query, batch_size, attn.heads, head_dim) - key = AttnUtils.reshape_qkv_for_attention(key, batch_size, attn.heads, head_dim) - value = AttnUtils.reshape_qkv_for_attention(value, batch_size, attn.heads, value.shape[-1] // attn.heads) - - # Apply normalization - query, key = AttnUtils.apply_norms(query, key, getattr(attn, "norm_q", None), getattr(attn, "norm_k", None)) - - # Apply RoPE if provided - if apply_rope_fn is not None: - query, key = apply_rope_fn(query, key, head_dim, **kwargs) - - # Compute attention - hidden_states = F.scaled_dot_product_attention( - query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False - ) - - return hidden_states, residual, input_ndim, shape_info, batch_size, attn.heads, head_dim - - -# Specific processor implementations (minimal unique code) -class PoseRoPEAttnProcessor2_0: - """ - Attention processor with Rotary Position Encoding (RoPE) for 3D spatial awareness. - - This processor extends standard attention with 3D rotary position embeddings - to provide spatial awareness for 3D scene understanding tasks. - """ - - def __init__(self): - """Initialize the RoPE attention processor.""" - AttnUtils.check_pytorch_compatibility() - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - position_indices: Dict = None, - temb: Optional[torch.Tensor] = None, - n_pbrs=1, - *args, - **kwargs, - ) -> torch.Tensor: - """ - Apply RoPE-enhanced attention computation. - - Args: - attn: Attention module instance - hidden_states: Input hidden states tensor - encoder_hidden_states: Optional encoder hidden states for cross-attention - attention_mask: Optional attention mask tensor - position_indices: Dictionary containing 3D position information for RoPE - temb: Optional temporal embedding tensor - n_pbrs: Number of PBR material types - *args: Additional positional arguments - **kwargs: Additional keyword arguments - - Returns: - Attention output tensor with applied rotary position encodings - """ - AttnUtils.handle_deprecation_warning(args, kwargs) - - def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs): - return attn.to_q(hidden_states), attn.to_k(encoder_hidden_states), attn.to_v(encoder_hidden_states) - - def apply_rope(query, key, head_dim, **kwargs): - if position_indices is not None: - if head_dim in position_indices: - image_rotary_emb = position_indices[head_dim] - else: - image_rotary_emb = RotaryEmbedding.get_3d_rotary_pos_embed( - rearrange( - position_indices["voxel_indices"].unsqueeze(1).repeat(1, n_pbrs, 1, 1), - "b n_pbrs l c -> (b n_pbrs) l c", - ), - head_dim, - voxel_resolution=position_indices["voxel_resolution"], - ) - position_indices[head_dim] = image_rotary_emb - - query = RotaryEmbedding.apply_rotary_emb(query, image_rotary_emb) - key = RotaryEmbedding.apply_rotary_emb(key, image_rotary_emb) - return query, key - - # Core attention processing - hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base( - attn, - hidden_states, - encoder_hidden_states, - attention_mask, - temb, - get_qkv_fn=get_qkv, - apply_rope_fn=apply_rope, - position_indices=position_indices, - n_pbrs=n_pbrs, - ) - - # Finalize output - hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim) - hidden_states = hidden_states.to(hidden_states.dtype) - - return AttnUtils.finalize_output(hidden_states, input_ndim, shape_info, attn, residual, attn.to_out) - - -class SelfAttnProcessor2_0(BaseAttnProcessor): - """ - Self-attention processor with PBR (Physically Based Rendering) material support. - - This processor handles multiple PBR material types (e.g., albedo, metallic-roughness) - with separate attention computation paths for each material type. - """ - - def __init__(self, **kwargs): - """ - Initialize self-attention processor with PBR support. - - Args: - **kwargs: Arguments passed to BaseAttnProcessor initialization - """ - super().__init__(**kwargs) - self.register_pbr_modules(["qkv", "out", "add_kv"], **kwargs) - - def process_single( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - token: Literal["albedo", "mr"] = "albedo", - multiple_devices=False, - *args, - **kwargs, - ): - """ - Process attention for a single PBR material type. - - Args: - attn: Attention module instance - hidden_states: Input hidden states tensor - encoder_hidden_states: Optional encoder hidden states for cross-attention - attention_mask: Optional attention mask tensor - temb: Optional temporal embedding tensor - token: PBR material type to process ("albedo", "mr", etc.) - multiple_devices: Whether to use multiple GPU devices - *args: Additional positional arguments - **kwargs: Additional keyword arguments - - Returns: - Processed attention output for the specified PBR material type - """ - target = attn if token == "albedo" else attn.processor - token_suffix = "" if token == "albedo" else "_" + token - - # Device management (if needed) - if multiple_devices: - device = torch.device("cuda:0") if token == "albedo" else torch.device("cuda:1") - for attr in [f"to_q{token_suffix}", f"to_k{token_suffix}", f"to_v{token_suffix}", f"to_out{token_suffix}"]: - getattr(target, attr).to(device) - - def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs): - return ( - getattr(target, f"to_q{token_suffix}")(hidden_states), - getattr(target, f"to_k{token_suffix}")(encoder_hidden_states), - getattr(target, f"to_v{token_suffix}")(encoder_hidden_states), - ) - - # Core processing using shared logic - hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base( - attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv - ) - - # Finalize - hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim) - hidden_states = hidden_states.to(hidden_states.dtype) - - return AttnUtils.finalize_output( - hidden_states, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}") - ) - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - """ - Apply self-attention with PBR material processing. - - Processes multiple PBR material types sequentially, applying attention - computation for each material type separately and combining results. - - Args: - attn: Attention module instance - hidden_states: Input hidden states tensor with PBR dimension - encoder_hidden_states: Optional encoder hidden states for cross-attention - attention_mask: Optional attention mask tensor - temb: Optional temporal embedding tensor - *args: Additional positional arguments - **kwargs: Additional keyword arguments - - Returns: - Combined attention output for all PBR material types - """ - AttnUtils.handle_deprecation_warning(args, kwargs) - - B = hidden_states.size(0) - pbr_hidden_states = torch.split(hidden_states, 1, dim=1) - - # Process each PBR setting - results = [] - for token, pbr_hs in zip(self.pbr_setting, pbr_hidden_states): - processed_hs = rearrange(pbr_hs, "b n_pbrs n l c -> (b n_pbrs n) l c").to("cuda:0") - result = self.process_single(attn, processed_hs, None, attention_mask, temb, token, False) - results.append(result) - - outputs = [rearrange(result, "(b n_pbrs n) l c -> b n_pbrs n l c", b=B, n_pbrs=1) for result in results] - return torch.cat(outputs, dim=1) - - -class RefAttnProcessor2_0(BaseAttnProcessor): - """ - Reference attention processor with shared value computation across PBR materials. - - This processor computes query and key once, but uses separate value projections - for different PBR material types, enabling efficient multi-material processing. - """ - - def __init__(self, **kwargs): - """ - Initialize reference attention processor. - - Args: - **kwargs: Arguments passed to BaseAttnProcessor initialization - """ - super().__init__(**kwargs) - self.pbr_settings = self.pbr_setting # Alias for compatibility - self.register_pbr_modules(["v_only", "out"], **kwargs) - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - """ - Apply reference attention with shared Q/K and separate V projections. - - This method computes query and key tensors once and reuses them across - all PBR material types, while using separate value projections for each - material type to maintain material-specific information. - - Args: - attn: Attention module instance - hidden_states: Input hidden states tensor - encoder_hidden_states: Optional encoder hidden states for cross-attention - attention_mask: Optional attention mask tensor - temb: Optional temporal embedding tensor - *args: Additional positional arguments - **kwargs: Additional keyword arguments - - Returns: - Stacked attention output for all PBR material types - """ - AttnUtils.handle_deprecation_warning(args, kwargs) - - def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs): - query = attn.to_q(hidden_states) - key = attn.to_k(encoder_hidden_states) - - # Concatenate values from all PBR settings - value_list = [attn.to_v(encoder_hidden_states)] - for token in ["_" + token for token in self.pbr_settings if token != "albedo"]: - value_list.append(getattr(attn.processor, f"to_v{token}")(encoder_hidden_states)) - value = torch.cat(value_list, dim=-1) - - return query, key, value - - # Core processing - hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base( - attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv - ) - - # Split and process each PBR setting output - hidden_states_list = torch.split(hidden_states, head_dim, dim=-1) - output_hidden_states_list = [] - - for i, hs in enumerate(hidden_states_list): - hs = hs.transpose(1, 2).reshape(batch_size, -1, heads * head_dim).to(hs.dtype) - token_suffix = "_" + self.pbr_settings[i] if self.pbr_settings[i] != "albedo" else "" - target = attn if self.pbr_settings[i] == "albedo" else attn.processor - - hs = AttnUtils.finalize_output( - hs, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}") - ) - output_hidden_states_list.append(hs) - - return torch.stack(output_hidden_states_list, dim=1) diff --git a/hy3dpaint/hunyuanpaintpbr/unet/model.py b/hy3dpaint/hunyuanpaintpbr/unet/model.py deleted file mode 100644 index f98ca7e73a5fdc59d6fc1afb0b2f1f6eb75be169..0000000000000000000000000000000000000000 --- a/hy3dpaint/hunyuanpaintpbr/unet/model.py +++ /dev/null @@ -1,622 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os - -# import ipdb -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F -import pytorch_lightning as pl -from tqdm import tqdm -from torchvision.transforms import v2 -from torchvision.utils import make_grid, save_image -from einops import rearrange - -from diffusers import ( - DiffusionPipeline, - EulerAncestralDiscreteScheduler, - DDPMScheduler, - UNet2DConditionModel, - ControlNetModel, -) - -from .modules import Dino_v2, UNet2p5DConditionModel -import math - - -def extract_into_tensor(a, t, x_shape): - b, *_ = t.shape - out = a.gather(-1, t) - return out.reshape(b, *((1,) * (len(x_shape) - 1))) - - -class HunyuanPaint(pl.LightningModule): - def __init__( - self, - stable_diffusion_config, - control_net_config=None, - num_view=6, - view_size=320, - drop_cond_prob=0.1, - with_normal_map=None, - with_position_map=None, - pbr_settings=["albedo", "mr"], - **kwargs, - ): - """Initializes the HunyuanPaint Lightning Module. - - Args: - stable_diffusion_config: Configuration for loading the Stable Diffusion pipeline - control_net_config: Configuration for ControlNet (optional) - num_view: Number of views to process - view_size: Size of input views (height/width) - drop_cond_prob: Probability of dropping conditioning input during training - with_normal_map: Flag indicating whether normal maps are used - with_position_map: Flag indicating whether position maps are used - pbr_settings: List of PBR materials to generate (e.g., albedo, metallic-roughness) - **kwargs: Additional keyword arguments - """ - super(HunyuanPaint, self).__init__() - - self.num_view = num_view - self.view_size = view_size - self.drop_cond_prob = drop_cond_prob - self.pbr_settings = pbr_settings - - # init modules - pipeline = DiffusionPipeline.from_pretrained(**stable_diffusion_config) - pipeline.set_pbr_settings(self.pbr_settings) - pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config( - pipeline.scheduler.config, timestep_spacing="trailing" - ) - - self.with_normal_map = with_normal_map - self.with_position_map = with_position_map - - self.pipeline = pipeline - - self.pipeline.vae.use_slicing = True - - train_sched = DDPMScheduler.from_config(self.pipeline.scheduler.config) - - if isinstance(self.pipeline.unet, UNet2DConditionModel): - self.pipeline.unet = UNet2p5DConditionModel( - self.pipeline.unet, train_sched, self.pipeline.scheduler, self.pbr_settings - ) - self.train_scheduler = train_sched # use ddpm scheduler during training - - self.register_schedule() - - pipeline.set_learned_parameters() - - if control_net_config is not None: - pipeline.unet = pipeline.unet.bfloat16().requires_grad_(control_net_config.train_unet) - self.pipeline.add_controlnet( - ControlNetModel.from_pretrained(control_net_config.pretrained_model_name_or_path), - conditioning_scale=0.75, - ) - - self.unet = pipeline.unet - - self.pipeline.set_progress_bar_config(disable=True) - self.pipeline.vae = self.pipeline.vae.bfloat16() - self.pipeline.text_encoder = self.pipeline.text_encoder.bfloat16() - - if self.unet.use_dino: - self.dino_v2 = Dino_v2("facebook/dinov2-giant") - self.dino_v2 = self.dino_v2.bfloat16() - - self.validation_step_outputs = [] - - def register_schedule(self): - - self.num_timesteps = self.train_scheduler.config.num_train_timesteps - - betas = self.train_scheduler.betas.detach().cpu() - - alphas = 1.0 - betas - alphas_cumprod = torch.cumprod(alphas, dim=0) - alphas_cumprod_prev = torch.cat([torch.ones(1, dtype=torch.float64), alphas_cumprod[:-1]], 0) - - self.register_buffer("betas", betas.float()) - self.register_buffer("alphas_cumprod", alphas_cumprod.float()) - self.register_buffer("alphas_cumprod_prev", alphas_cumprod_prev.float()) - - # calculations for diffusion q(x_t | x_{t-1}) and others - self.register_buffer("sqrt_alphas_cumprod", torch.sqrt(alphas_cumprod).float()) - self.register_buffer("sqrt_one_minus_alphas_cumprod", torch.sqrt(1 - alphas_cumprod).float()) - - self.register_buffer("sqrt_recip_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod).float()) - self.register_buffer("sqrt_recipm1_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod - 1).float()) - - def on_fit_start(self): - device = torch.device(f"cuda:{self.local_rank}") - self.pipeline.to(device) - if self.global_rank == 0: - os.makedirs(os.path.join(self.logdir, "images_val"), exist_ok=True) - - def prepare_batch_data(self, batch): - """Preprocesses a batch of input data for training/inference. - - Args: - batch: Raw input batch dictionary - - Returns: - tuple: Contains: - - cond_imgs: Primary conditioning images (B, 1, C, H, W) - - cond_imgs_another: Secondary conditioning images (B, 1, C, H, W) - - target_imgs: Dictionary of target PBR images resized and clamped - - images_normal: Preprocessed normal maps (if available) - - images_position: Preprocessed position maps (if available) - """ - - images_cond = batch["images_cond"].to(self.device) # (B, M, C, H, W), where M is the number of reference images - cond_imgs, cond_imgs_another = images_cond[:, 0:1, ...], images_cond[:, 1:2, ...] - - cond_size = self.view_size - cond_imgs = v2.functional.resize(cond_imgs, cond_size, interpolation=3, antialias=True).clamp(0, 1) - cond_imgs_another = v2.functional.resize(cond_imgs_another, cond_size, interpolation=3, antialias=True).clamp( - 0, 1 - ) - - target_imgs = {} - for pbr_token in self.pbr_settings: - target_imgs[pbr_token] = batch[f"images_{pbr_token}"].to(self.device) - target_imgs[pbr_token] = v2.functional.resize( - target_imgs[pbr_token], self.view_size, interpolation=3, antialias=True - ).clamp(0, 1) - - images_normal = None - if "images_normal" in batch: - images_normal = batch["images_normal"] # (B, N, C, H, W) - images_normal = v2.functional.resize(images_normal, self.view_size, interpolation=3, antialias=True).clamp( - 0, 1 - ) - images_normal = [images_normal] - - images_position = None - if "images_position" in batch: - images_position = batch["images_position"] # (B, N, C, H, W) - images_position = v2.functional.resize( - images_position, self.view_size, interpolation=3, antialias=True - ).clamp(0, 1) - images_position = [images_position] - - return cond_imgs, cond_imgs_another, target_imgs, images_normal, images_position - - @torch.no_grad() - def forward_text_encoder(self, prompts): - device = next(self.pipeline.vae.parameters()).device - text_embeds = self.pipeline.encode_prompt(prompts, device, 1, False)[0] - return text_embeds - - @torch.no_grad() - def encode_images(self, images): - """Encodes input images into latent representations using the VAE. - - Handles both standard input (B, N, C, H, W) and PBR input (B, N_pbrs, N, C, H, W) - Maintains original batch structure in output latents. - - Args: - images: Input images tensor - - Returns: - torch.Tensor: Latent representations with original batch dimensions preserved - """ - - B = images.shape[0] - image_ndims = images.ndim - if image_ndims != 5: - N_pbrs, N = images.shape[1:3] - images = ( - rearrange(images, "b n c h w -> (b n) c h w") - if image_ndims == 5 - else rearrange(images, "b n_pbrs n c h w -> (b n_pbrs n) c h w") - ) - dtype = next(self.pipeline.vae.parameters()).dtype - - images = (images - 0.5) * 2.0 - posterior = self.pipeline.vae.encode(images.to(dtype)).latent_dist - latents = posterior.sample() * self.pipeline.vae.config.scaling_factor - - latents = ( - rearrange(latents, "(b n) c h w -> b n c h w", b=B) - if image_ndims == 5 - else rearrange(latents, "(b n_pbrs n) c h w -> b n_pbrs n c h w", b=B, n_pbrs=N_pbrs) - ) - - return latents - - def forward_unet(self, latents, t, **cached_condition): - """Runs the UNet model to predict noise/latent residuals. - - Args: - latents: Noisy latent representations (B, C, H, W) - t: Timestep tensor (B,) - **cached_condition: Dictionary of conditioning inputs (text embeds, reference images, etc) - - Returns: - torch.Tensor: UNet output (predicted noise or velocity) - """ - - dtype = next(self.unet.parameters()).dtype - latents = latents.to(dtype) - shading_embeds = cached_condition["shading_embeds"] - pred_noise = self.pipeline.unet(latents, t, encoder_hidden_states=shading_embeds, **cached_condition) - return pred_noise[0] - - def predict_start_from_z_and_v(self, x_t, t, v): - """ - Predicts clean image (x0) from noisy latents (x_t) and - velocity prediction (v) using the v-prediction formula. - - Args: - x_t: Noisy latents at timestep t - t: Current timestep - v: Predicted velocity (v) from UNet - - Returns: - torch.Tensor: Predicted clean image (x0) - """ - - return ( - extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t - - extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v - ) - - def get_v(self, x, noise, t): - """Computes the target velocity (v) for v-prediction training. - - Args: - x: Clean latents (x0) - noise: Added noise - t: Current timestep - - Returns: - torch.Tensor: Target velocity - """ - - return ( - extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise - - extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x - ) - - def training_step(self, batch, batch_idx): - """Performs a single training step with both conditioning paths. - - Implements: - 1. Dual-conditioning path training (main ref + secondary ref) - 2. Velocity-prediction with consistency loss - 3. Conditional dropout for robust learning - 4. PBR-specific losses (albedo/metallic-roughness) - - Args: - batch: Input batch from dataloader - batch_idx: Index of current batch - - Returns: - torch.Tensor: Combined loss value - """ - - cond_imgs, cond_imgs_another, target_imgs, normal_imgs, position_imgs = self.prepare_batch_data(batch) - - B, N_ref = cond_imgs.shape[:2] - _, N_gen, _, H, W = target_imgs["albedo"].shape - N_pbrs = len(self.pbr_settings) - t = torch.randint(0, self.num_timesteps, size=(B,)).long().to(self.device) - t = t.unsqueeze(-1).repeat(1, N_pbrs, N_gen) - t = rearrange(t, "b n_pbrs n -> (b n_pbrs n)") - - all_target_pbrs = [] - for pbr_token in self.pbr_settings: - all_target_pbrs.append(target_imgs[pbr_token]) - all_target_pbrs = torch.stack(all_target_pbrs, dim=0).transpose(1, 0) - gen_latents = self.encode_images(all_target_pbrs) #! B, N_pbrs N C H W - ref_latents = self.encode_images(cond_imgs) #! B, M, C, H, W - ref_latents_another = self.encode_images(cond_imgs_another) #! B, M, C, H, W - - all_shading_tokens = [] - for token in self.pbr_settings: - if token in ["albedo", "mr"]: - all_shading_tokens.append( - getattr(self.unet, f"learned_text_clip_{token}").unsqueeze(dim=0).repeat(B, 1, 1) - ) - shading_embeds = torch.stack(all_shading_tokens, dim=1) - - if self.unet.use_dino: - dino_hidden_states = self.dino_v2(cond_imgs[:, :1, ...]) - dino_hidden_states_another = self.dino_v2(cond_imgs_another[:, :1, ...]) - - gen_latents = rearrange(gen_latents, "b n_pbrs n c h w -> (b n_pbrs n) c h w") - noise = torch.randn_like(gen_latents).to(self.device) - latents_noisy = self.train_scheduler.add_noise(gen_latents, noise, t).to(self.device) - latents_noisy = rearrange(latents_noisy, "(b n_pbrs n) c h w -> b n_pbrs n c h w", b=B, n_pbrs=N_pbrs) - - cached_condition = {} - - if normal_imgs is not None: - normal_embeds = self.encode_images(normal_imgs[0]) - cached_condition["embeds_normal"] = normal_embeds #! B, N, C, H, W - - if position_imgs is not None: - position_embeds = self.encode_images(position_imgs[0]) - cached_condition["embeds_position"] = position_embeds #! B, N, C, H, W - cached_condition["position_maps"] = position_imgs[0] #! B, N, C, H, W - - for b in range(B): - prob = np.random.rand() - if prob < self.drop_cond_prob: - if "normal_imgs" in cached_condition: - cached_condition["embeds_normal"][b, ...] = torch.zeros_like( - cached_condition["embeds_normal"][b, ...] - ) - if "position_imgs" in cached_condition: - cached_condition["embeds_position"][b, ...] = torch.zeros_like( - cached_condition["embeds_position"][b, ...] - ) - - prob = np.random.rand() - if prob < self.drop_cond_prob: - if "position_maps" in cached_condition: - cached_condition["position_maps"][b, ...] = torch.zeros_like( - cached_condition["position_maps"][b, ...] - ) - - prob = np.random.rand() - if prob < self.drop_cond_prob: - dino_hidden_states[b, ...] = torch.zeros_like(dino_hidden_states[b, ...]) - prob = np.random.rand() - if prob < self.drop_cond_prob: - dino_hidden_states_another[b, ...] = torch.zeros_like(dino_hidden_states_another[b, ...]) - - # MVA & Ref Attention - prob = np.random.rand() - cached_condition["mva_scale"] = 1.0 - cached_condition["ref_scale"] = 1.0 - if prob < self.drop_cond_prob: - cached_condition["mva_scale"] = 0.0 - cached_condition["ref_scale"] = 0.0 - elif prob > 1.0 - self.drop_cond_prob: - prob = np.random.rand() - if prob < 0.5: - cached_condition["mva_scale"] = 0.0 - else: - cached_condition["ref_scale"] = 0.0 - else: - pass - - if self.train_scheduler.config.prediction_type == "v_prediction": - - cached_condition["shading_embeds"] = shading_embeds - cached_condition["ref_latents"] = ref_latents - cached_condition["dino_hidden_states"] = dino_hidden_states - v_pred = self.forward_unet(latents_noisy, t, **cached_condition) - v_pred_albedo, v_pred_mr = torch.split( - rearrange( - v_pred, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view - ), - 1, - dim=1, - ) - v_target = self.get_v(gen_latents, noise, t) - v_target_albedo, v_target_mr = torch.split( - rearrange( - v_target, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view - ), - 1, - dim=1, - ) - - albedo_loss_1, _ = self.compute_loss(v_pred_albedo, v_target_albedo) - mr_loss_1, _ = self.compute_loss(v_pred_mr, v_target_mr) - - cached_condition["ref_latents"] = ref_latents_another - cached_condition["dino_hidden_states"] = dino_hidden_states_another - v_pred_another = self.forward_unet(latents_noisy, t, **cached_condition) - v_pred_another_albedo, v_pred_another_mr = torch.split( - rearrange( - v_pred_another, - "(b n_pbr n) c h w -> b n_pbr n c h w", - n_pbr=len(self.pbr_settings), - n=self.num_view, - ), - 1, - dim=1, - ) - - albedo_loss_2, _ = self.compute_loss(v_pred_another_albedo, v_target_albedo) - mr_loss_2, _ = self.compute_loss(v_pred_another_mr, v_target_mr) - - consistency_loss, _ = self.compute_loss(v_pred_another, v_pred) - - albedo_loss = (albedo_loss_1 + albedo_loss_2) * 0.5 - mr_loss = (mr_loss_1 + mr_loss_2) * 0.5 - - log_loss_dict = {} - log_loss_dict.update({f"train/albedo_loss": albedo_loss}) - log_loss_dict.update({f"train/mr_loss": mr_loss}) - log_loss_dict.update({f"train/cons_loss": consistency_loss}) - - loss_dict = log_loss_dict - - elif self.train_scheduler.config.prediction_type == "epsilon": - e_pred = self.forward_unet(latents_noisy, t, **cached_condition) - loss, loss_dict = self.compute_loss(e_pred, noise) - else: - raise f"No {self.train_scheduler.config.prediction_type}" - - # logging - self.log_dict(loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=True) - self.log("global_step", self.global_step, prog_bar=True, logger=True, on_step=True, on_epoch=False) - lr = self.optimizers().param_groups[0]["lr"] - self.log("lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False) - - return 0.85 * (albedo_loss + mr_loss) + 0.15 * consistency_loss - - def compute_loss(self, noise_pred, noise_gt): - loss = F.mse_loss(noise_pred, noise_gt) - prefix = "train" - loss_dict = {} - loss_dict.update({f"{prefix}/loss": loss}) - return loss, loss_dict - - @torch.no_grad() - def validation_step(self, batch, batch_idx): - """Performs validation on a single batch. - - Generates predicted images using: - 1. Reference conditioning images - 2. Optional normal/position maps - 3. Frozen DINO features (if enabled) - 4. Text prompt conditioning - - Compares predictions against ground truth targets and prepares visualization. - Stores results for epoch-level aggregation. - - Args: - batch: Input batch from validation dataloader - batch_idx: Index of current batch - """ - # [Validation image generation and comparison logic...] - # Key steps: - # 1. Preprocess conditioning images to PIL format - # 2. Set up conditioning inputs (normal maps, position maps, DINO features) - # 3. Run pipeline inference with fixed prompt ("high quality") - # 4. Decode latent outputs to image space - # 5. Arrange predictions and ground truths for visualization - - cond_imgs_tensor, _, target_imgs, normal_imgs, position_imgs = self.prepare_batch_data(batch) - resolution = self.view_size - image_pils = [] - for i in range(cond_imgs_tensor.shape[0]): - image_pils.append([]) - for j in range(cond_imgs_tensor.shape[1]): - image_pils[-1].append(v2.functional.to_pil_image(cond_imgs_tensor[i, j, ...])) - - outputs, gts = [], [] - for idx in range(len(image_pils)): - cond_imgs = image_pils[idx] - - cached_condition = dict(num_in_batch=self.num_view, N_pbrs=len(self.pbr_settings)) - if normal_imgs is not None: - cached_condition["images_normal"] = normal_imgs[0][idx, ...].unsqueeze(0) - if position_imgs is not None: - cached_condition["images_position"] = position_imgs[0][idx, ...].unsqueeze(0) - if self.pipeline.unet.use_dino: - dino_hidden_states = self.dino_v2([cond_imgs][0]) - cached_condition["dino_hidden_states"] = dino_hidden_states - - latent = self.pipeline( - cond_imgs, - prompt="high quality", - num_inference_steps=30, - output_type="latent", - height=resolution, - width=resolution, - **cached_condition, - ).images - - image = self.pipeline.vae.decode(latent / self.pipeline.vae.config.scaling_factor, return_dict=False)[ - 0 - ] # [-1, 1] - image = (image * 0.5 + 0.5).clamp(0, 1) - - image = rearrange( - image, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view - ) - image = torch.cat((torch.ones_like(image[:, :, :1, ...]) * 0.5, image), dim=2) - image = rearrange(image, "b n_pbr n c h w -> (b n_pbr n) c h w") - image = rearrange( - image, - "(b n_pbr n) c h w -> b c (n_pbr h) (n w)", - b=1, - n_pbr=len(self.pbr_settings), - n=self.num_view + 1, - ) - outputs.append(image) - - all_target_pbrs = [] - for pbr_token in self.pbr_settings: - all_target_pbrs.append(target_imgs[pbr_token]) - all_target_pbrs = torch.stack(all_target_pbrs, dim=0).transpose(1, 0) - all_target_pbrs = torch.cat( - (cond_imgs_tensor.unsqueeze(1).repeat(1, len(self.pbr_settings), 1, 1, 1, 1), all_target_pbrs), dim=2 - ) - all_target_pbrs = rearrange(all_target_pbrs, "b n_pbrs n c h w -> b c (n_pbrs h) (n w)") - gts = all_target_pbrs - outputs = torch.cat(outputs, dim=0).to(self.device) - images = torch.cat([gts, outputs], dim=-2) - self.validation_step_outputs.append(images) - - @torch.no_grad() - def on_validation_epoch_end(self): - """Aggregates validation results at epoch end. - - Gathers outputs from all GPUs (if distributed training), - creates a unified visualization grid, and saves to disk. - Only rank 0 process performs saving. - """ - # [Result aggregation and visualization...] - # Key steps: - # 1. Gather validation outputs from all processes - # 2. Create image grid combining ground truths and predictions - # 3. Save visualization with step-numbered filename - # 4. Clear memory for next validation cycle - - images = torch.cat(self.validation_step_outputs, dim=0) - all_images = self.all_gather(images) - all_images = rearrange(all_images, "r b c h w -> (r b) c h w") - - if self.global_rank == 0: - grid = make_grid(all_images, nrow=8, normalize=True, value_range=(0, 1)) - save_image(grid, os.path.join(self.logdir, "images_val", f"val_{self.global_step:07d}.png")) - - self.validation_step_outputs.clear() # free memory - - def configure_optimizers(self): - lr = self.learning_rate - optimizer = torch.optim.AdamW(self.unet.parameters(), lr=lr) - - def lr_lambda(step): - warm_up_step = 1000 - T_step = 9000 - gamma = 0.9 - min_lr = 0.1 if step >= warm_up_step else 0.0 - max_lr = 1.0 - normalized_step = step % (warm_up_step + T_step) - current_max_lr = max_lr * gamma ** (step // (warm_up_step + T_step)) - if current_max_lr < min_lr: - current_max_lr = min_lr - if normalized_step < warm_up_step: - lr_step = min_lr + (normalized_step / warm_up_step) * (current_max_lr - min_lr) - else: - step_wc_wp = normalized_step - warm_up_step - ratio = step_wc_wp / T_step - lr_step = min_lr + 0.5 * (current_max_lr - min_lr) * (1 + math.cos(math.pi * ratio)) - return lr_step - - lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda) - - lr_scheduler_config = { - "scheduler": lr_scheduler, - "interval": "step", - "frequency": 1, - "monitor": "val_loss", - "strict": False, - "name": None, - } - - return {"optimizer": optimizer, "lr_scheduler": lr_scheduler_config} diff --git a/hy3dpaint/hunyuanpaintpbr/unet/modules.py b/hy3dpaint/hunyuanpaintpbr/unet/modules.py deleted file mode 100644 index 55f0404a452ecd9d7bc20d84303c151c67d8bf8f..0000000000000000000000000000000000000000 --- a/hy3dpaint/hunyuanpaintpbr/unet/modules.py +++ /dev/null @@ -1,1102 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import json -import copy -import numpy as np -import torch -import torch.nn as nn -from einops import rearrange -from typing import Any, Callable, Dict, List, Optional, Union, Tuple, Literal -import diffusers -from diffusers.utils import deprecate -from diffusers import ( - DDPMScheduler, - EulerAncestralDiscreteScheduler, - UNet2DConditionModel, -) -from diffusers.models import UNet2DConditionModel -from diffusers.models.attention_processor import Attention, AttnProcessor -from diffusers.models.transformers.transformer_2d import BasicTransformerBlock -from .attn_processor import SelfAttnProcessor2_0, RefAttnProcessor2_0, PoseRoPEAttnProcessor2_0 - -from transformers import AutoImageProcessor, AutoModel - - -class Dino_v2(nn.Module): - - """Wrapper for DINOv2 vision transformer (frozen weights). - - Provides feature extraction for reference images. - - Args: - dino_v2_path: Custom path to DINOv2 model weights (uses default if None) - """ - - - def __init__(self, dino_v2_path): - super(Dino_v2, self).__init__() - self.dino_processor = AutoImageProcessor.from_pretrained(dino_v2_path) - self.dino_v2 = AutoModel.from_pretrained(dino_v2_path) - - for param in self.parameters(): - param.requires_grad = False - - self.dino_v2.eval() - - def forward(self, images): - - """Processes input images through DINOv2 ViT. - - Handles both tensor input (B, N, C, H, W) and PIL image lists. - Extracts patch embeddings and flattens spatial dimensions. - - Returns: - torch.Tensor: Feature vectors [B, N*(num_patches), feature_dim] - """ - - if isinstance(images, torch.Tensor): - batch_size = images.shape[0] - dino_proceesed_images = self.dino_processor( - images=rearrange(images, "b n c h w -> (b n) c h w"), return_tensors="pt", do_rescale=False - ).pixel_values - else: - batch_size = 1 - dino_proceesed_images = self.dino_processor(images=images, return_tensors="pt").pixel_values - dino_proceesed_images = torch.stack( - [torch.from_numpy(np.array(image)) for image in dino_proceesed_images], dim=0 - ) - dino_param = next(self.dino_v2.parameters()) - dino_proceesed_images = dino_proceesed_images.to(dino_param) - dino_hidden_states = self.dino_v2(dino_proceesed_images)[0] - dino_hidden_states = rearrange(dino_hidden_states.to(dino_param), "(b n) l c -> b (n l) c", b=batch_size) - - return dino_hidden_states - - -def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int): - # "feed_forward_chunk_size" can be used to save memory - - """Memory-efficient feedforward execution via chunking. - - Divides input along specified dimension for sequential processing. - - Args: - ff: Feedforward module to apply - hidden_states: Input tensor - chunk_dim: Dimension to split - chunk_size: Size of each chunk - - Returns: - torch.Tensor: Reassembled output tensor - """ - - if hidden_states.shape[chunk_dim] % chunk_size != 0: - raise ValueError( - f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]}" - f"has to be divisible by chunk size: {chunk_size}." - "Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." - ) - - num_chunks = hidden_states.shape[chunk_dim] // chunk_size - ff_output = torch.cat( - [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], - dim=chunk_dim, - ) - return ff_output - - -@torch.no_grad() -def compute_voxel_grid_mask(position, grid_resolution=8): - - """Generates view-to-view attention mask based on 3D position similarity. - - Uses voxel grid downsampling to determine spatially adjacent regions. - Mask indicates where features should interact across different views. - - Args: - position: Position maps [B, N, 3, H, W] (normalized 0-1) - grid_resolution: Spatial reduction factor - - Returns: - torch.Tensor: Attention mask [B, N*grid_res**2, N*grid_res**2] - """ - - position = position.half() - B, N, _, H, W = position.shape - assert H % grid_resolution == 0 and W % grid_resolution == 0 - - valid_mask = (position != 1).all(dim=2, keepdim=True) - valid_mask = valid_mask.expand_as(position) - position[valid_mask == False] = 0 - - position = rearrange( - position, - "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w", - num_h=grid_resolution, - num_w=grid_resolution, - ) - valid_mask = rearrange( - valid_mask, - "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w", - num_h=grid_resolution, - num_w=grid_resolution, - ) - - grid_position = position.sum(dim=(-2, -1)) - count_masked = valid_mask.sum(dim=(-2, -1)) - - grid_position = grid_position / count_masked.clamp(min=1) - grid_position[count_masked < 5] = 0 - - grid_position = grid_position.permute(0, 1, 4, 2, 3) - grid_position = rearrange(grid_position, "b n c h w -> b n (h w) c") - - grid_position_expanded_1 = grid_position.unsqueeze(2).unsqueeze(4) # 形状变为 B, N, 1, L, 1, 3 - grid_position_expanded_2 = grid_position.unsqueeze(1).unsqueeze(3) # 形状变为 B, 1, N, 1, L, 3 - - # 计算欧氏距离 - distances = torch.norm(grid_position_expanded_1 - grid_position_expanded_2, dim=-1) # 形状为 B, N, N, L, L - - weights = distances - grid_distance = 1.73 / grid_resolution - weights = weights < grid_distance - - return weights - - -def compute_multi_resolution_mask(position_maps, grid_resolutions=[32, 16, 8]): - - """Generates attention masks at multiple spatial resolutions. - - Creates pyramid of position-based masks for hierarchical attention. - - Args: - position_maps: Position maps [B, N, 3, H, W] - grid_resolutions: List of downsampling factors - - Returns: - dict: Resolution-specific masks keyed by flattened dimension size - """ - - position_attn_mask = {} - with torch.no_grad(): - for grid_resolution in grid_resolutions: - position_mask = compute_voxel_grid_mask(position_maps, grid_resolution) - position_mask = rearrange(position_mask, "b ni nj li lj -> b (ni li) (nj lj)") - position_attn_mask[position_mask.shape[1]] = position_mask - return position_attn_mask - - -@torch.no_grad() -def compute_discrete_voxel_indice(position, grid_resolution=8, voxel_resolution=128): - - """Quantizes position maps to discrete voxel indices. - - Creates sparse 3D coordinate representations for efficient hashing. - - Args: - position: Position maps [B, N, 3, H, W] - grid_resolution: Spatial downsampling factor - voxel_resolution: Quantization resolution - - Returns: - torch.Tensor: Voxel indices [B, N, grid_res, grid_res, 3] - """ - - position = position.half() - B, N, _, H, W = position.shape - assert H % grid_resolution == 0 and W % grid_resolution == 0 - - valid_mask = (position != 1).all(dim=2, keepdim=True) - valid_mask = valid_mask.expand_as(position) - position[valid_mask == False] = 0 - - position = rearrange( - position, - "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w", - num_h=grid_resolution, - num_w=grid_resolution, - ) - valid_mask = rearrange( - valid_mask, - "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w", - num_h=grid_resolution, - num_w=grid_resolution, - ) - - grid_position = position.sum(dim=(-2, -1)) - count_masked = valid_mask.sum(dim=(-2, -1)) - - grid_position = grid_position / count_masked.clamp(min=1) - voxel_mask_thres = (H // grid_resolution) * (W // grid_resolution) // (4 * 4) - grid_position[count_masked < voxel_mask_thres] = 0 - - grid_position = grid_position.permute(0, 1, 4, 2, 3).clamp(0, 1) # B N C H W - voxel_indices = grid_position * (voxel_resolution - 1) - voxel_indices = torch.round(voxel_indices).long() - return voxel_indices - - -def calc_multires_voxel_idxs(position_maps, grid_resolutions=[64, 32, 16, 8], voxel_resolutions=[512, 256, 128, 64]): - - """Generates multi-resolution voxel indices for position encoding. - - Creates pyramid of quantized position representations. - - Args: - position_maps: Input position maps - grid_resolutions: Spatial resolution levels - voxel_resolutions: Quantization levels - - Returns: - dict: Voxel indices keyed by flattened dimension size, with resolution metadata - """ - - voxel_indices = {} - with torch.no_grad(): - for grid_resolution, voxel_resolution in zip(grid_resolutions, voxel_resolutions): - voxel_indice = compute_discrete_voxel_indice(position_maps, grid_resolution, voxel_resolution) - voxel_indice = rearrange(voxel_indice, "b n c h w -> b (n h w) c") - voxel_indices[voxel_indice.shape[1]] = {"voxel_indices": voxel_indice, "voxel_resolution": voxel_resolution} - return voxel_indices - - -class Basic2p5DTransformerBlock(torch.nn.Module): - - - """Enhanced transformer block for multiview 2.5D image generation. - - Extends standard transformer blocks with: - - Material-specific attention (MDA) - - Multiview attention (MA) - - Reference attention (RA) - - DINO feature integration - - Args: - transformer: Base transformer block - layer_name: Identifier for layer - use_ma: Enable multiview attention - use_ra: Enable reference attention - use_mda: Enable material-aware attention - use_dino: Enable DINO feature integration - pbr_setting: List of PBR materials - """ - - def __init__( - self, - transformer: BasicTransformerBlock, - layer_name, - use_ma=True, - use_ra=True, - use_mda=True, - use_dino=True, - pbr_setting=None, - ) -> None: - - """ - Initialization: - 1. Material-Dimension Attention (MDA): - - Processes each PBR material with separate projection weights - - Uses custom SelfAttnProcessor2_0 with material awareness - - 2. Multiview Attention (MA): - - Adds cross-view attention with PoseRoPE - - Initialized as zero-initialized residual pathway - - 3. Reference Attention (RA): - - Conditions on reference view features - - Uses RefAttnProcessor2_0 for material-specific conditioning - - 4. DINO Attention: - - Incorporates DINO-ViT features - - Initialized as zero-initialized residual pathway - """ - - super().__init__() - self.transformer = transformer - self.layer_name = layer_name - self.use_ma = use_ma - self.use_ra = use_ra - self.use_mda = use_mda - self.use_dino = use_dino - self.pbr_setting = pbr_setting - - if self.use_mda: - self.attn1.set_processor( - SelfAttnProcessor2_0( - query_dim=self.dim, - heads=self.num_attention_heads, - dim_head=self.attention_head_dim, - dropout=self.dropout, - bias=self.attention_bias, - cross_attention_dim=None, - upcast_attention=self.attn1.upcast_attention, - out_bias=True, - pbr_setting=self.pbr_setting, - ) - ) - - # multiview attn - if self.use_ma: - self.attn_multiview = Attention( - query_dim=self.dim, - heads=self.num_attention_heads, - dim_head=self.attention_head_dim, - dropout=self.dropout, - bias=self.attention_bias, - cross_attention_dim=None, - upcast_attention=self.attn1.upcast_attention, - out_bias=True, - processor=PoseRoPEAttnProcessor2_0(), - ) - - # ref attn - if self.use_ra: - self.attn_refview = Attention( - query_dim=self.dim, - heads=self.num_attention_heads, - dim_head=self.attention_head_dim, - dropout=self.dropout, - bias=self.attention_bias, - cross_attention_dim=None, - upcast_attention=self.attn1.upcast_attention, - out_bias=True, - processor=RefAttnProcessor2_0( - query_dim=self.dim, - heads=self.num_attention_heads, - dim_head=self.attention_head_dim, - dropout=self.dropout, - bias=self.attention_bias, - cross_attention_dim=None, - upcast_attention=self.attn1.upcast_attention, - out_bias=True, - pbr_setting=self.pbr_setting, - ), - ) - - # dino attn - if self.use_dino: - self.attn_dino = Attention( - query_dim=self.dim, - heads=self.num_attention_heads, - dim_head=self.attention_head_dim, - dropout=self.dropout, - bias=self.attention_bias, - cross_attention_dim=self.cross_attention_dim, - upcast_attention=self.attn2.upcast_attention, - out_bias=True, - ) - - self._initialize_attn_weights() - - def _initialize_attn_weights(self): - - """Initializes specialized attention heads with base weights. - - Uses weight sharing strategy: - - Copies base transformer weights to specialized heads - - Initializes newly-added parameters to zero - """ - - if self.use_mda: - for token in self.pbr_setting: - if token == "albedo": - continue - getattr(self.attn1.processor, f"to_q_{token}").load_state_dict(self.attn1.to_q.state_dict()) - getattr(self.attn1.processor, f"to_k_{token}").load_state_dict(self.attn1.to_k.state_dict()) - getattr(self.attn1.processor, f"to_v_{token}").load_state_dict(self.attn1.to_v.state_dict()) - getattr(self.attn1.processor, f"to_out_{token}").load_state_dict(self.attn1.to_out.state_dict()) - - if self.use_ma: - self.attn_multiview.load_state_dict(self.attn1.state_dict(), strict=False) - with torch.no_grad(): - for layer in self.attn_multiview.to_out: - for param in layer.parameters(): - param.zero_() - - if self.use_ra: - self.attn_refview.load_state_dict(self.attn1.state_dict(), strict=False) - for token in self.pbr_setting: - if token == "albedo": - continue - getattr(self.attn_refview.processor, f"to_v_{token}").load_state_dict( - self.attn_refview.to_q.state_dict() - ) - getattr(self.attn_refview.processor, f"to_out_{token}").load_state_dict( - self.attn_refview.to_out.state_dict() - ) - with torch.no_grad(): - for layer in self.attn_refview.to_out: - for param in layer.parameters(): - param.zero_() - for token in self.pbr_setting: - if token == "albedo": - continue - for layer in getattr(self.attn_refview.processor, f"to_out_{token}"): - for param in layer.parameters(): - param.zero_() - - if self.use_dino: - self.attn_dino.load_state_dict(self.attn2.state_dict(), strict=False) - with torch.no_grad(): - for layer in self.attn_dino.to_out: - for param in layer.parameters(): - param.zero_() - - if self.use_dino: - self.attn_dino.load_state_dict(self.attn2.state_dict(), strict=False) - with torch.no_grad(): - for layer in self.attn_dino.to_out: - for param in layer.parameters(): - param.zero_() - - def __getattr__(self, name: str): - try: - return super().__getattr__(name) - except AttributeError: - return getattr(self.transformer, name) - - def forward( - self, - hidden_states: torch.Tensor, - attention_mask: Optional[torch.Tensor] = None, - encoder_hidden_states: Optional[torch.Tensor] = None, - encoder_attention_mask: Optional[torch.Tensor] = None, - timestep: Optional[torch.LongTensor] = None, - cross_attention_kwargs: Dict[str, Any] = None, - class_labels: Optional[torch.LongTensor] = None, - added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, - ) -> torch.Tensor: - - """Forward pass with multi-mechanism attention. - - Processing stages: - 1. Material-aware self-attention (MDA) - 2. Reference attention (RA) - 3. Multiview attention (MA) with position-aware attention - 4. Text conditioning (base attention) - 5. DINO feature conditioning (optional) - 6. Position-aware conditioning - 7. Feed-forward network - - Args: - hidden_states: Input features [B * N_materials * N_views, Seq_len, Feat_dim] - See base transformer for other parameters - - Returns: - torch.Tensor: Output features - """ - # [Full multi-mechanism processing pipeline...] - # Key processing stages: - # 1. Material-aware self-attention (handles albedo/mr separation) - # 2. Reference attention (conditioned on reference features) - # 3. View-to-view attention with geometric constraints - # 4. Text-to-image cross-attention - # 5. DINO feature fusion (when enabled) - # 6. Positional conditioning (RoPE-style) - # 7. Feed-forward network with conditional normalization - - # Notice that normalization is always applied before the real computation in the following blocks. - # 0. Self-Attention - batch_size = hidden_states.shape[0] - - cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} - num_in_batch = cross_attention_kwargs.pop("num_in_batch", 1) - mode = cross_attention_kwargs.pop("mode", None) - mva_scale = cross_attention_kwargs.pop("mva_scale", 1.0) - ref_scale = cross_attention_kwargs.pop("ref_scale", 1.0) - condition_embed_dict = cross_attention_kwargs.pop("condition_embed_dict", None) - dino_hidden_states = cross_attention_kwargs.pop("dino_hidden_states", None) - position_voxel_indices = cross_attention_kwargs.pop("position_voxel_indices", None) - N_pbr = len(self.pbr_setting) if self.pbr_setting is not None else 1 - - if self.norm_type == "ada_norm": - norm_hidden_states = self.norm1(hidden_states, timestep) - elif self.norm_type == "ada_norm_zero": - norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( - hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype - ) - elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: - norm_hidden_states = self.norm1(hidden_states) - elif self.norm_type == "ada_norm_continuous": - norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) - elif self.norm_type == "ada_norm_single": - shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( - self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) - ).chunk(6, dim=1) - norm_hidden_states = self.norm1(hidden_states) - norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa - else: - raise ValueError("Incorrect norm used") - - if self.pos_embed is not None: - norm_hidden_states = self.pos_embed(norm_hidden_states) - - # 1. Prepare GLIGEN inputs - cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} - gligen_kwargs = cross_attention_kwargs.pop("gligen", None) - - if self.use_mda: - mda_norm_hidden_states = rearrange( - norm_hidden_states, "(b n_pbr n) l c -> b n_pbr n l c", n=num_in_batch, n_pbr=N_pbr - ) - attn_output = self.attn1( - mda_norm_hidden_states, - encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, - attention_mask=attention_mask, - **cross_attention_kwargs, - ) - attn_output = rearrange(attn_output, "b n_pbr n l c -> (b n_pbr n) l c") - else: - attn_output = self.attn1( - norm_hidden_states, - encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, - attention_mask=attention_mask, - **cross_attention_kwargs, - ) - - if self.norm_type == "ada_norm_zero": - attn_output = gate_msa.unsqueeze(1) * attn_output - elif self.norm_type == "ada_norm_single": - attn_output = gate_msa * attn_output - - hidden_states = attn_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - # 1.2 Reference Attention - if "w" in mode: - condition_embed_dict[self.layer_name] = rearrange( - norm_hidden_states, "(b n) l c -> b (n l) c", n=num_in_batch - ) # B, (N L), C - - if "r" in mode and self.use_ra: - condition_embed = condition_embed_dict[self.layer_name] - - #! Only using albedo features for reference attention - ref_norm_hidden_states = rearrange( - norm_hidden_states, "(b n_pbr n) l c -> b n_pbr (n l) c", n=num_in_batch, n_pbr=N_pbr - )[:, 0, ...] - - attn_output = self.attn_refview( - ref_norm_hidden_states, - encoder_hidden_states=condition_embed, - attention_mask=None, - **cross_attention_kwargs, - ) # b (n l) c - attn_output = rearrange(attn_output, "b n_pbr (n l) c -> (b n_pbr n) l c", n=num_in_batch, n_pbr=N_pbr) - - ref_scale_timing = ref_scale - if isinstance(ref_scale, torch.Tensor): - ref_scale_timing = ref_scale.unsqueeze(1).repeat(1, num_in_batch * N_pbr).view(-1) - for _ in range(attn_output.ndim - 1): - ref_scale_timing = ref_scale_timing.unsqueeze(-1) - hidden_states = ref_scale_timing * attn_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - # 1.3 Multiview Attention - if num_in_batch > 1 and self.use_ma: - multivew_hidden_states = rearrange( - norm_hidden_states, "(b n_pbr n) l c -> (b n_pbr) (n l) c", n_pbr=N_pbr, n=num_in_batch - ) - position_indices = None - if position_voxel_indices is not None: - if multivew_hidden_states.shape[1] in position_voxel_indices: - position_indices = position_voxel_indices[multivew_hidden_states.shape[1]] - - attn_output = self.attn_multiview( - multivew_hidden_states, - encoder_hidden_states=multivew_hidden_states, - position_indices=position_indices, - n_pbrs=N_pbr, - **cross_attention_kwargs, - ) - - attn_output = rearrange(attn_output, "(b n_pbr) (n l) c -> (b n_pbr n) l c", n_pbr=N_pbr, n=num_in_batch) - - hidden_states = mva_scale * attn_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - # 1.2 GLIGEN Control - if gligen_kwargs is not None: - hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) - - # 3. Cross-Attention - if self.attn2 is not None: - if self.norm_type == "ada_norm": - norm_hidden_states = self.norm2(hidden_states, timestep) - elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: - norm_hidden_states = self.norm2(hidden_states) - elif self.norm_type == "ada_norm_single": - # For PixArt norm2 isn't applied here: - # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 - norm_hidden_states = hidden_states - elif self.norm_type == "ada_norm_continuous": - norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) - else: - raise ValueError("Incorrect norm") - - if self.pos_embed is not None and self.norm_type != "ada_norm_single": - norm_hidden_states = self.pos_embed(norm_hidden_states) - - attn_output = self.attn2( - norm_hidden_states, - encoder_hidden_states=encoder_hidden_states, - attention_mask=encoder_attention_mask, - **cross_attention_kwargs, - ) - hidden_states = attn_output + hidden_states - - # dino attn - if self.use_dino: - dino_hidden_states = dino_hidden_states.unsqueeze(1).repeat(1, N_pbr * num_in_batch, 1, 1) - dino_hidden_states = rearrange(dino_hidden_states, "b n l c -> (b n) l c") - attn_output = self.attn_dino( - norm_hidden_states, - encoder_hidden_states=dino_hidden_states, - attention_mask=None, - **cross_attention_kwargs, - ) - - hidden_states = attn_output + hidden_states - - # 4. Feed-forward - # i2vgen doesn't have this norm 🤷‍♂️ - if self.norm_type == "ada_norm_continuous": - norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) - elif not self.norm_type == "ada_norm_single": - norm_hidden_states = self.norm3(hidden_states) - - if self.norm_type == "ada_norm_zero": - norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] - - if self.norm_type == "ada_norm_single": - norm_hidden_states = self.norm2(hidden_states) - norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp - - if self._chunk_size is not None: - # "feed_forward_chunk_size" can be used to save memory - ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) - else: - ff_output = self.ff(norm_hidden_states) - - if self.norm_type == "ada_norm_zero": - ff_output = gate_mlp.unsqueeze(1) * ff_output - elif self.norm_type == "ada_norm_single": - ff_output = gate_mlp * ff_output - - hidden_states = ff_output + hidden_states - if hidden_states.ndim == 4: - hidden_states = hidden_states.squeeze(1) - - return hidden_states - - -class ImageProjModel(torch.nn.Module): - - """Projects image embeddings into cross-attention space. - - Transforms CLIP embeddings into additional context tokens for conditioning. - - Args: - cross_attention_dim: Dimension of attention space - clip_embeddings_dim: Dimension of input CLIP embeddings - clip_extra_context_tokens: Number of context tokens to generate - """ - - def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4): - super().__init__() - - self.generator = None - self.cross_attention_dim = cross_attention_dim - self.clip_extra_context_tokens = clip_extra_context_tokens - self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim) - self.norm = torch.nn.LayerNorm(cross_attention_dim) - - def forward(self, image_embeds): - - """Projects image embeddings to cross-attention context tokens. - - Args: - image_embeds: Input embeddings [B, N, C] or [B, C] - - Returns: - torch.Tensor: Context tokens [B, N*clip_extra_context_tokens, cross_attention_dim] - """ - - embeds = image_embeds - num_token = 1 - if embeds.dim() == 3: - num_token = embeds.shape[1] - embeds = rearrange(embeds, "b n c -> (b n) c") - - clip_extra_context_tokens = self.proj(embeds).reshape( - -1, self.clip_extra_context_tokens, self.cross_attention_dim - ) - clip_extra_context_tokens = self.norm(clip_extra_context_tokens) - - clip_extra_context_tokens = rearrange(clip_extra_context_tokens, "(b nt) n c -> b (nt n) c", nt=num_token) - - return clip_extra_context_tokens - - -class UNet2p5DConditionModel(torch.nn.Module): - - """2.5D UNet extension for multiview PBR generation. - - Enhances standard 2D UNet with: - - Multiview attention mechanisms - - Material-aware processing - - Position-aware conditioning - - Dual-stream reference processing - - Args: - unet: Base 2D UNet model - train_sched: Training scheduler (DDPM) - val_sched: Validation scheduler (EulerAncestral) - """ - - def __init__( - self, - unet: UNet2DConditionModel, - train_sched: DDPMScheduler = None, - val_sched: EulerAncestralDiscreteScheduler = None, - ) -> None: - super().__init__() - self.unet = unet - self.train_sched = train_sched - self.val_sched = val_sched - - self.use_ma = True - self.use_ra = True - self.use_mda = True - self.use_dino = True - self.use_position_rope = True - self.use_learned_text_clip = True - self.use_dual_stream = True - self.pbr_setting = ["albedo", "mr"] - self.pbr_token_channels = 77 - - if self.use_dual_stream and self.use_ra: - self.unet_dual = copy.deepcopy(unet) - self.init_attention(self.unet_dual) - - self.init_attention( - self.unet, - use_ma=self.use_ma, - use_ra=self.use_ra, - use_dino=self.use_dino, - use_mda=self.use_mda, - pbr_setting=self.pbr_setting, - ) - self.init_condition(use_dino=self.use_dino) - - @staticmethod - def from_pretrained(pretrained_model_name_or_path, **kwargs): - torch_dtype = kwargs.pop("torch_dtype", torch.float32) - config_path = os.path.join(pretrained_model_name_or_path, "config.json") - unet_ckpt_path = os.path.join(pretrained_model_name_or_path, "diffusion_pytorch_model.bin") - with open(config_path, "r", encoding="utf-8") as file: - config = json.load(file) - unet = UNet2DConditionModel(**config) - unet_2p5d = UNet2p5DConditionModel(unet) - unet_2p5d.unet.conv_in = torch.nn.Conv2d( - 12, - unet.conv_in.out_channels, - kernel_size=unet.conv_in.kernel_size, - stride=unet.conv_in.stride, - padding=unet.conv_in.padding, - dilation=unet.conv_in.dilation, - groups=unet.conv_in.groups, - bias=unet.conv_in.bias is not None, - ) - unet_ckpt = torch.load(unet_ckpt_path, map_location="cpu", weights_only=True) - unet_2p5d.load_state_dict(unet_ckpt, strict=True) - unet_2p5d = unet_2p5d.to(torch_dtype) - return unet_2p5d - - def init_condition(self, use_dino): - - """Initializes conditioning mechanisms for multiview PBR generation. - - Sets up: - 1. Learned text embeddings: Material-specific tokens (albedo, mr) initialized to zeros - 2. DINO projector: Model to process DINO-ViT features for cross-attention - - Args: - use_dino: Flag to enable DINO feature integration - """ - - if self.use_learned_text_clip: - for token in self.pbr_setting: - self.unet.register_parameter( - f"learned_text_clip_{token}", nn.Parameter(torch.zeros(self.pbr_token_channels, 1024)) - ) - self.unet.learned_text_clip_ref = nn.Parameter(torch.zeros(self.pbr_token_channels, 1024)) - - if use_dino: - self.unet.image_proj_model_dino = ImageProjModel( - cross_attention_dim=self.unet.config.cross_attention_dim, - clip_embeddings_dim=1536, - clip_extra_context_tokens=4, - ) - - def init_attention(self, unet, use_ma=False, use_ra=False, use_mda=False, use_dino=False, pbr_setting=None): - - """Recursively replaces standard transformers with enhanced 2.5D blocks. - - Processes UNet architecture: - 1. Downsampling blocks: Replaces transformers in attention layers - 2. Middle block: Upgrades central transformers - 3. Upsampling blocks: Modifies decoder transformers - - Args: - unet: UNet model to enhance - use_ma: Enable multiview attention - use_ra: Enable reference attention - use_mda: Enable material-specific attention - use_dino: Enable DINO feature integration - pbr_setting: List of PBR materials - """ - - for down_block_i, down_block in enumerate(unet.down_blocks): - if hasattr(down_block, "has_cross_attention") and down_block.has_cross_attention: - for attn_i, attn in enumerate(down_block.attentions): - for transformer_i, transformer in enumerate(attn.transformer_blocks): - if isinstance(transformer, BasicTransformerBlock): - attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock( - transformer, - f"down_{down_block_i}_{attn_i}_{transformer_i}", - use_ma, - use_ra, - use_mda, - use_dino, - pbr_setting, - ) - - if hasattr(unet.mid_block, "has_cross_attention") and unet.mid_block.has_cross_attention: - for attn_i, attn in enumerate(unet.mid_block.attentions): - for transformer_i, transformer in enumerate(attn.transformer_blocks): - if isinstance(transformer, BasicTransformerBlock): - attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock( - transformer, f"mid_{attn_i}_{transformer_i}", use_ma, use_ra, use_mda, use_dino, pbr_setting - ) - - for up_block_i, up_block in enumerate(unet.up_blocks): - if hasattr(up_block, "has_cross_attention") and up_block.has_cross_attention: - for attn_i, attn in enumerate(up_block.attentions): - for transformer_i, transformer in enumerate(attn.transformer_blocks): - if isinstance(transformer, BasicTransformerBlock): - attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock( - transformer, - f"up_{up_block_i}_{attn_i}_{transformer_i}", - use_ma, - use_ra, - use_mda, - use_dino, - pbr_setting, - ) - - def __getattr__(self, name: str): - try: - return super().__getattr__(name) - except AttributeError: - return getattr(self.unet, name) - - def forward( - self, - sample, - timestep, - encoder_hidden_states, - *args, - added_cond_kwargs=None, - cross_attention_kwargs=None, - down_intrablock_additional_residuals=None, - down_block_res_samples=None, - mid_block_res_sample=None, - **cached_condition, - ): - - """Forward pass with multiview/material conditioning. - - Key stages: - 1. Input preparation (concat normal/position maps) - 2. Reference feature extraction (dual-stream) - 3. Position encoding (voxel indices) - 4. DINO feature projection - 5. Main UNet processing with attention conditioning - - Args: - sample: Input latents [B, N_pbr, N_gen, C, H, W] - cached_condition: Dictionary containing: - - embeds_normal: Normal map embeddings - - embeds_position: Position map embeddings - - ref_latents: Reference image latents - - dino_hidden_states: DINO features - - position_maps: 3D position maps - - mva_scale: Multiview attention scale - - ref_scale: Reference attention scale - - Returns: - torch.Tensor: Output features - """ - - B, N_pbr, N_gen, _, H, W = sample.shape - assert H == W - - if "cache" not in cached_condition: - cached_condition["cache"] = {} - - sample = [sample] - if "embeds_normal" in cached_condition: - sample.append(cached_condition["embeds_normal"].unsqueeze(1).repeat(1, N_pbr, 1, 1, 1, 1)) - if "embeds_position" in cached_condition: - sample.append(cached_condition["embeds_position"].unsqueeze(1).repeat(1, N_pbr, 1, 1, 1, 1)) - sample = torch.cat(sample, dim=-3) - - sample = rearrange(sample, "b n_pbr n c h w -> (b n_pbr n) c h w") - - encoder_hidden_states_gen = encoder_hidden_states.unsqueeze(-3).repeat(1, 1, N_gen, 1, 1) - encoder_hidden_states_gen = rearrange(encoder_hidden_states_gen, "b n_pbr n l c -> (b n_pbr n) l c") - - if added_cond_kwargs is not None: - text_embeds_gen = added_cond_kwargs["text_embeds"].unsqueeze(1).repeat(1, N_gen, 1) - text_embeds_gen = rearrange(text_embeds_gen, "b n c -> (b n) c") - time_ids_gen = added_cond_kwargs["time_ids"].unsqueeze(1).repeat(1, N_gen, 1) - time_ids_gen = rearrange(time_ids_gen, "b n c -> (b n) c") - added_cond_kwargs_gen = {"text_embeds": text_embeds_gen, "time_ids": time_ids_gen} - else: - added_cond_kwargs_gen = None - - if self.use_position_rope: - if "position_voxel_indices" in cached_condition["cache"]: - position_voxel_indices = cached_condition["cache"]["position_voxel_indices"] - else: - if "position_maps" in cached_condition: - position_voxel_indices = calc_multires_voxel_idxs( - cached_condition["position_maps"], - grid_resolutions=[H, H // 2, H // 4, H // 8], - voxel_resolutions=[H * 8, H * 4, H * 2, H], - ) - cached_condition["cache"]["position_voxel_indices"] = position_voxel_indices - else: - position_voxel_indices = None - - if self.use_dino: - if "dino_hidden_states_proj" in cached_condition["cache"]: - dino_hidden_states = cached_condition["cache"]["dino_hidden_states_proj"] - else: - assert "dino_hidden_states" in cached_condition - dino_hidden_states = cached_condition["dino_hidden_states"] - dino_hidden_states = self.image_proj_model_dino(dino_hidden_states) - cached_condition["cache"]["dino_hidden_states_proj"] = dino_hidden_states - else: - dino_hidden_states = None - - if self.use_ra: - if "condition_embed_dict" in cached_condition["cache"]: - condition_embed_dict = cached_condition["cache"]["condition_embed_dict"] - else: - condition_embed_dict = {} - ref_latents = cached_condition["ref_latents"] - N_ref = ref_latents.shape[1] - - if not self.use_dual_stream: - ref_latents = [ref_latents] - if "embeds_normal" in cached_condition: - ref_latents.append(torch.zeros_like(ref_latents[0])) - if "embeds_position" in cached_condition: - ref_latents.append(torch.zeros_like(ref_latents[0])) - ref_latents = torch.cat(ref_latents, dim=2) - - ref_latents = rearrange(ref_latents, "b n c h w -> (b n) c h w") - - encoder_hidden_states_ref = self.unet.learned_text_clip_ref.repeat(B, N_ref, 1, 1) - - encoder_hidden_states_ref = rearrange(encoder_hidden_states_ref, "b n l c -> (b n) l c") - - if added_cond_kwargs is not None: - text_embeds_ref = added_cond_kwargs["text_embeds"].unsqueeze(1).repeat(1, N_ref, 1) - text_embeds_ref = rearrange(text_embeds_ref, "b n c -> (b n) c") - time_ids_ref = added_cond_kwargs["time_ids"].unsqueeze(1).repeat(1, N_ref, 1) - time_ids_ref = rearrange(time_ids_ref, "b n c -> (b n) c") - added_cond_kwargs_ref = { - "text_embeds": text_embeds_ref, - "time_ids": time_ids_ref, - } - else: - added_cond_kwargs_ref = None - - noisy_ref_latents = ref_latents - timestep_ref = 0 - if self.use_dual_stream: - unet_ref = self.unet_dual - else: - unet_ref = self.unet - unet_ref( - noisy_ref_latents, - timestep_ref, - encoder_hidden_states=encoder_hidden_states_ref, - class_labels=None, - added_cond_kwargs=added_cond_kwargs_ref, - # **kwargs - return_dict=False, - cross_attention_kwargs={ - "mode": "w", - "num_in_batch": N_ref, - "condition_embed_dict": condition_embed_dict, - }, - ) - cached_condition["cache"]["condition_embed_dict"] = condition_embed_dict - else: - condition_embed_dict = None - - mva_scale = cached_condition.get("mva_scale", 1.0) - ref_scale = cached_condition.get("ref_scale", 1.0) - - return self.unet( - sample, - timestep, - encoder_hidden_states_gen, - *args, - class_labels=None, - added_cond_kwargs=added_cond_kwargs_gen, - down_intrablock_additional_residuals=( - [sample.to(dtype=self.unet.dtype) for sample in down_intrablock_additional_residuals] - if down_intrablock_additional_residuals is not None - else None - ), - down_block_additional_residuals=( - [sample.to(dtype=self.unet.dtype) for sample in down_block_res_samples] - if down_block_res_samples is not None - else None - ), - mid_block_additional_residual=( - mid_block_res_sample.to(dtype=self.unet.dtype) if mid_block_res_sample is not None else None - ), - return_dict=False, - cross_attention_kwargs={ - "mode": "r", - "num_in_batch": N_gen, - "dino_hidden_states": dino_hidden_states, - "condition_embed_dict": condition_embed_dict, - "mva_scale": mva_scale, - "ref_scale": ref_scale, - "position_voxel_indices": position_voxel_indices, - }, - ) diff --git a/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/__init__.py b/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/__init__.py deleted file mode 100644 index 319995298d19a223710c1a608c5152c80a813ece..0000000000000000000000000000000000000000 --- a/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/__init__.py +++ /dev/null @@ -1,4 +0,0 @@ -""" -from .render import rasterize, interpolate -""" -from .render import * diff --git a/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/render.py b/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/render.py deleted file mode 100644 index 9d06b5195ed18d3cb581ace9efc1c2cebed232eb..0000000000000000000000000000000000000000 --- a/hy3dpaint/packages/custom_rasterizer/custom_rasterizer/render.py +++ /dev/null @@ -1,32 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import custom_rasterizer_kernel -import torch - - -def rasterize(pos, tri, resolution, clamp_depth=torch.zeros(0), use_depth_prior=0): - assert pos.device == tri.device - findices, barycentric = custom_rasterizer_kernel.rasterize_image( - pos[0], tri, clamp_depth, resolution[1], resolution[0], 1e-6, use_depth_prior - ) - return findices, barycentric - - -def interpolate(col, findices, barycentric, tri): - f = findices - 1 + (findices == 0) - vcol = col[0, tri.long()[f.long()]] - result = barycentric.view(*barycentric.shape, 1) * vcol - result = torch.sum(result, axis=-2) - return result.view(1, *result.shape) diff --git a/hy3dpaint/packages/custom_rasterizer/setup.py b/hy3dpaint/packages/custom_rasterizer/setup.py deleted file mode 100644 index 15192e9f5e1f4b0308ed6f1fdd3c93df8c35feef..0000000000000000000000000000000000000000 --- a/hy3dpaint/packages/custom_rasterizer/setup.py +++ /dev/null @@ -1,40 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from setuptools import setup, find_packages -import torch -from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CppExtension - -# build custom rasterizer - -custom_rasterizer_module = CUDAExtension( - "custom_rasterizer_kernel", - [ - "lib/custom_rasterizer_kernel/rasterizer.cpp", - "lib/custom_rasterizer_kernel/grid_neighbor.cpp", - "lib/custom_rasterizer_kernel/rasterizer_gpu.cu", - ], -) - -setup( - packages=find_packages(), - version="0.1", - name="custom_rasterizer", - include_package_data=True, - package_dir={"": "."}, - ext_modules=[ - custom_rasterizer_module, - ], - cmdclass={"build_ext": BuildExtension}, -) diff --git a/hy3dpaint/src/__init__.py b/hy3dpaint/src/__init__.py deleted file mode 100644 index 1614ff826f65b4720649d343c43fb09dbb6b9fa5..0000000000000000000000000000000000000000 --- a/hy3dpaint/src/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. diff --git a/hy3dpaint/src/utils/__init__.py b/hy3dpaint/src/utils/__init__.py deleted file mode 100644 index 1614ff826f65b4720649d343c43fb09dbb6b9fa5..0000000000000000000000000000000000000000 --- a/hy3dpaint/src/utils/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. diff --git a/hy3dpaint/src/utils/train_util.py b/hy3dpaint/src/utils/train_util.py deleted file mode 100644 index dff39da91cc24c812251728c786ad0509c544351..0000000000000000000000000000000000000000 --- a/hy3dpaint/src/utils/train_util.py +++ /dev/null @@ -1,40 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import importlib - - -def count_params(model, verbose=False): - total_params = sum(p.numel() for p in model.parameters()) - if verbose: - print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.") - return total_params - - -def instantiate_from_config(config): - if not "target" in config: - if config == "__is_first_stage__": - return None - elif config == "__is_unconditional__": - return None - raise KeyError("Expected key `target` to instantiate.") - return get_obj_from_str(config["target"])(**config.get("params", dict())) - - -def get_obj_from_str(string, reload=False): - module, cls = string.rsplit(".", 1) - if reload: - module_imp = importlib.import_module(module) - importlib.reload(module_imp) - return getattr(importlib.import_module(module, package=None), cls) diff --git a/hy3dpaint/textureGenPipeline.py b/hy3dpaint/textureGenPipeline.py deleted file mode 100644 index 9396eea492f88f5b10c3c44d8ef14a1540055fc3..0000000000000000000000000000000000000000 --- a/hy3dpaint/textureGenPipeline.py +++ /dev/null @@ -1,192 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import torch -import copy -import trimesh -import numpy as np -from PIL import Image -from typing import List -from DifferentiableRenderer.MeshRender import MeshRender -from utils.simplify_mesh_utils import remesh_mesh -from utils.multiview_utils import multiviewDiffusionNet -from utils.pipeline_utils import ViewProcessor -from utils.image_super_utils import imageSuperNet -from utils.uvwrap_utils import mesh_uv_wrap -from DifferentiableRenderer.mesh_utils import convert_obj_to_glb -import warnings - -warnings.filterwarnings("ignore") -from diffusers.utils import logging as diffusers_logging - -diffusers_logging.set_verbosity(50) - - -class Hunyuan3DPaintConfig: - def __init__(self, max_num_view, resolution): - self.device = "cuda" - - self.multiview_cfg_path = "cfgs/hunyuan-paint-pbr.yaml" - self.custom_pipeline = "hunyuanpaintpbr" - self.multiview_pretrained_path = "tencent/Hunyuan3D-2.1" - self.dino_ckpt_path = "facebook/dinov2-giant" - self.realesrgan_ckpt_path = "ckpt/RealESRGAN_x4plus.pth" - - self.raster_mode = "cr" - self.bake_mode = "back_sample" - self.render_size = 1024 * 2 - self.texture_size = 1024 * 4 - self.max_selected_view_num = max_num_view - self.resolution = resolution - self.bake_exp = 4 - self.merge_method = "fast" - - # view selection - self.candidate_camera_azims = [0, 90, 180, 270, 0, 180] - self.candidate_camera_elevs = [0, 0, 0, 0, 90, -90] - self.candidate_view_weights = [1, 0.1, 0.5, 0.1, 0.05, 0.05] - - for azim in range(0, 360, 30): - self.candidate_camera_azims.append(azim) - self.candidate_camera_elevs.append(20) - self.candidate_view_weights.append(0.01) - - self.candidate_camera_azims.append(azim) - self.candidate_camera_elevs.append(-20) - self.candidate_view_weights.append(0.01) - - -class Hunyuan3DPaintPipeline: - - def __init__(self, config=None) -> None: - self.config = config if config is not None else Hunyuan3DPaintConfig() - self.models = {} - self.stats_logs = {} - self.render = MeshRender( - default_resolution=self.config.render_size, - texture_size=self.config.texture_size, - bake_mode=self.config.bake_mode, - raster_mode=self.config.raster_mode, - ) - self.view_processor = ViewProcessor(self.config, self.render) - self.load_models() - - def load_models(self): - torch.cuda.empty_cache() - self.models["super_model"] = imageSuperNet(self.config) - self.models["multiview_model"] = multiviewDiffusionNet(self.config) - print("Models Loaded.") - - @torch.no_grad() - def __call__(self, mesh_path=None, image_path=None, output_mesh_path=None, use_remesh=True, save_glb=True): - """Generate texture for 3D mesh using multiview diffusion""" - # Ensure image_prompt is a list - if isinstance(image_path, str): - image_prompt = Image.open(image_path) - elif isinstance(image_path, Image.Image): - image_prompt = image_path - if not isinstance(image_prompt, List): - image_prompt = [image_prompt] - else: - image_prompt = image_path - - # Process mesh - path = os.path.dirname(mesh_path) - if use_remesh: - processed_mesh_path = os.path.join(path, "white_mesh_remesh.obj") - remesh_mesh(mesh_path, processed_mesh_path) - else: - processed_mesh_path = mesh_path - - # Output path - if output_mesh_path is None: - output_mesh_path = os.path.join(path, f"textured_mesh.obj") - - # Load mesh - mesh = trimesh.load(processed_mesh_path) - mesh = mesh_uv_wrap(mesh) - self.render.load_mesh(mesh=mesh) - - ########### View Selection ######### - selected_camera_elevs, selected_camera_azims, selected_view_weights = self.view_processor.bake_view_selection( - self.config.candidate_camera_elevs, - self.config.candidate_camera_azims, - self.config.candidate_view_weights, - self.config.max_selected_view_num, - ) - - normal_maps = self.view_processor.render_normal_multiview( - selected_camera_elevs, selected_camera_azims, use_abs_coor=True - ) - position_maps = self.view_processor.render_position_multiview(selected_camera_elevs, selected_camera_azims) - - ########## Style ########### - image_caption = "high quality" - image_style = [] - for image in image_prompt: - image = image.resize((512, 512)) - if image.mode == "RGBA": - white_bg = Image.new("RGB", image.size, (255, 255, 255)) - white_bg.paste(image, mask=image.getchannel("A")) - image = white_bg - image_style.append(image) - image_style = [image.convert("RGB") for image in image_style] - - ########### Multiview ########## - multiviews_pbr = self.models["multiview_model"]( - image_style, - normal_maps + position_maps, - prompt=image_caption, - custom_view_size=self.config.resolution, - resize_input=True, - ) - ########### Enhance ########## - enhance_images = {} - enhance_images["albedo"] = copy.deepcopy(multiviews_pbr["albedo"]) - enhance_images["mr"] = copy.deepcopy(multiviews_pbr["mr"]) - - for i in range(len(enhance_images["albedo"])): - enhance_images["albedo"][i] = self.models["super_model"](enhance_images["albedo"][i]) - enhance_images["mr"][i] = self.models["super_model"](enhance_images["mr"][i]) - - ########### Bake ########## - for i in range(len(enhance_images)): - enhance_images["albedo"][i] = enhance_images["albedo"][i].resize( - (self.config.render_size, self.config.render_size) - ) - enhance_images["mr"][i] = enhance_images["mr"][i].resize((self.config.render_size, self.config.render_size)) - texture, mask = self.view_processor.bake_from_multiview( - enhance_images["albedo"], selected_camera_elevs, selected_camera_azims, selected_view_weights - ) - mask_np = (mask.squeeze(-1).cpu().numpy() * 255).astype(np.uint8) - texture_mr, mask_mr = self.view_processor.bake_from_multiview( - enhance_images["mr"], selected_camera_elevs, selected_camera_azims, selected_view_weights - ) - mask_mr_np = (mask_mr.squeeze(-1).cpu().numpy() * 255).astype(np.uint8) - - ########## inpaint ########### - texture = self.view_processor.texture_inpaint(texture, mask_np) - self.render.set_texture(texture, force_set=True) - if "mr" in enhance_images: - texture_mr = self.view_processor.texture_inpaint(texture_mr, mask_mr_np) - self.render.set_texture_mr(texture_mr) - - self.render.save_mesh(output_mesh_path, downsample=True) - - if save_glb: - convert_obj_to_glb(output_mesh_path, output_mesh_path.replace(".obj", ".glb")) - output_glb_path = output_mesh_path.replace(".obj", ".glb") - - return output_mesh_path diff --git a/hy3dpaint/train.py b/hy3dpaint/train.py deleted file mode 100644 index 289f74b8e44b9c9931cc3dc749c9ca102811d62e..0000000000000000000000000000000000000000 --- a/hy3dpaint/train.py +++ /dev/null @@ -1,401 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import torch -import os, sys -import argparse -import shutil -import subprocess -from omegaconf import OmegaConf - -from pytorch_lightning import seed_everything -from pytorch_lightning.trainer import Trainer -from pytorch_lightning.strategies import DDPStrategy -from pytorch_lightning.callbacks import Callback -from pytorch_lightning.utilities import rank_zero_only, rank_zero_warn - -from src.utils.train_util import instantiate_from_config -import warnings - -warnings.filterwarnings("ignore") -from diffusers.utils import logging as diffusers_logging - -diffusers_logging.set_verbosity(50) - - -@rank_zero_only -def rank_zero_print(*args): - print(*args) - - -def get_parser(**parser_kwargs): - def str2bool(v): - if isinstance(v, bool): - return v - if v.lower() in ("yes", "true", "t", "y", "1"): - return True - elif v.lower() in ("no", "false", "f", "n", "0"): - return False - else: - raise argparse.ArgumentTypeError("Boolean value expected.") - - parser = argparse.ArgumentParser(**parser_kwargs) - parser.add_argument( - "-r", - "--resume", - type=str, - default=None, - help="resume from checkpoint", - ) - parser.add_argument( - "--resume_weights_only", - action="store_true", - help="only resume model weights", - ) - parser.add_argument( - "-b", - "--base", - type=str, - default="base_config.yaml", - help="path to base configs", - ) - parser.add_argument( - "-n", - "--name", - type=str, - default="", - help="experiment name", - ) - parser.add_argument( - "--num_nodes", - type=int, - default=1, - help="number of nodes to use", - ) - parser.add_argument( - "--gpus", - type=str, - default="0,", - help="gpu ids to use", - ) - parser.add_argument( - "-s", - "--seed", - type=int, - default=42, - help="seed for seed_everything", - ) - parser.add_argument( - "-l", - "--logdir", - type=str, - default="logs", - help="directory for logging data", - ) - return parser - - -class SetupCallback(Callback): - def __init__(self, resume, logdir, ckptdir, cfgdir, config): - super().__init__() - self.resume = resume - self.logdir = logdir - self.ckptdir = ckptdir - self.cfgdir = cfgdir - self.config = config - - def on_fit_start(self, trainer, pl_module): - if trainer.global_rank == 0: - # Create logdirs and save configs - os.makedirs(self.logdir, exist_ok=True) - os.makedirs(self.ckptdir, exist_ok=True) - os.makedirs(self.cfgdir, exist_ok=True) - - rank_zero_print("Project config") - rank_zero_print(OmegaConf.to_yaml(self.config)) - OmegaConf.save(self.config, os.path.join(self.cfgdir, "project.yaml")) - - -class CodeSnapshot(Callback): - """ - Modified from https://github.com/threestudio-project/threestudio/blob/main/threestudio/utils/callbacks.py#L60 - """ - - def __init__(self, savedir): - self.savedir = savedir - - def get_file_list(self): - return [ - b.decode() - for b in set(subprocess.check_output('git ls-files -- ":!:configs/*"', shell=True).splitlines()) - | set( # hard code, TODO: use config to exclude folders or files - subprocess.check_output("git ls-files --others --exclude-standard", shell=True).splitlines() - ) - ] - - @rank_zero_only - def save_code_snapshot(self): - os.makedirs(self.savedir, exist_ok=True) - - # for f in self.get_file_list(): - # if not os.path.exists(f) or os.path.isdir(f): - # continue - # os.makedirs(os.path.join(self.savedir, os.path.dirname(f)), exist_ok=True) - # shutil.copyfile(f, os.path.join(self.savedir, f)) - - def on_fit_start(self, trainer, pl_module): - try: - self.save_code_snapshot() - except: - rank_zero_warn( - "Code snapshot is not saved. Please make sure you have git installed and are in a git repository." - ) - - -if __name__ == "__main__": - # add cwd for convenience and to make classes in this file available when - # running as `python main.py` - sys.path.append(os.getcwd()) - torch.set_float32_matmul_precision("medium") - - parser = get_parser() - opt, unknown = parser.parse_known_args() - - cfg_fname = os.path.split(opt.base)[-1] - cfg_name = os.path.splitext(cfg_fname)[0] - exp_name = "-" + opt.name if opt.name != "" else "" - logdir = os.path.join(opt.logdir, cfg_name + exp_name) - - # assert not os.path.exists(logdir) or 'test' in logdir, logdir - if os.path.exists(logdir) and opt.resume is None: - auto_resume_path = os.path.join(logdir, "checkpoints", "last.ckpt") - if os.path.exists(auto_resume_path): - opt.resume = auto_resume_path - print(f"Auto set resume ckpt {opt.resume}") - - ckptdir = os.path.join(logdir, "checkpoints") - cfgdir = os.path.join(logdir, "configs") - codedir = os.path.join(logdir, "code") - - node_rank = int(os.environ.get("NODE_RANK", 0)) # 当前节点的编号 - local_rank = int(os.environ.get("LOCAL_RANK", 0)) # 当前节点上的 GPU 编号 - num_gpus_per_node = torch.cuda.device_count() # 每个节点上的 GPU 数量 - - global_rank = node_rank * num_gpus_per_node + local_rank - seed_everything(opt.seed + global_rank) - - # init configs - config = OmegaConf.load(opt.base) - lightning_config = config.lightning - trainer_config = lightning_config.trainer - - trainer_config["accelerator"] = "gpu" - rank_zero_print(f"Running on GPUs {opt.gpus}") - try: - ngpu = int(opt.gpus) - except: - ngpu = len(opt.gpus.strip(",").split(",")) - trainer_config["devices"] = ngpu - - trainer_opt = argparse.Namespace(**trainer_config) - lightning_config.trainer = trainer_config - - # model - model = instantiate_from_config(config.model) - - model_unet = model.unet.unet - model_unet_prefix = "unet.unet." - if hasattr(model_unet, "unet"): - model_unet = model_unet.unet - model_unet_prefix += "unet." - - if getattr(config, "init_unet_from", None): - unet_ckpt_path = config.init_unet_from - sd = torch.load(unet_ckpt_path, map_location="cpu") - model_unet.load_state_dict(sd, strict=True) - - if getattr(config, "init_vae_from", None): - vae_ckpt_path = config.init_vae_from - sd_vae = torch.load(vae_ckpt_path, map_location="cpu") - - def replace_key(key_str): - replace_pairs = [("key", "to_k"), ("query", "to_q"), ("value", "to_v"), ("proj_attn", "to_out.0")] - for replace_pair in replace_pairs: - key_str = key_str.replace(replace_pair[0], replace_pair[1]) - return key_str - - sd_vae = {replace_key(k): v for k, v in sd_vae.items()} - model.pipeline.vae.load_state_dict(sd_vae, strict=True) - - if hasattr(model.unet, "controlnet"): - if getattr(config, "init_control_from", None): - unet_ckpt_path = config.init_control_from - sd_control = torch.load(unet_ckpt_path, map_location="cpu") - model.unet.controlnet.load(sd_control, strict=True) - - noise_in_channels = config.model.params.get("noise_in_channels", None) - if noise_in_channels is not None: - with torch.no_grad(): - new_conv_in = torch.nn.Conv2d( - noise_in_channels, - model_unet.conv_in.out_channels, - model_unet.conv_in.kernel_size, - model_unet.conv_in.stride, - model_unet.conv_in.padding, - ) - new_conv_in.weight.zero_() - new_conv_in.weight[:, : model_unet.conv_in.in_channels, :, :].copy_(model_unet.conv_in.weight) - - new_conv_in.bias.zero_() - new_conv_in.bias[: model_unet.conv_in.bias.size(0)].copy_(model_unet.conv_in.bias) - - model_unet.conv_in = new_conv_in - - if hasattr(model.unet, "controlnet"): - if config.model.params.get("control_in_channels", None): - control_in_channels = config.model.params.control_in_channels - model.unet.controlnet.config["conditioning_channels"] = control_in_channels - condition_conv_in = model.unet.controlnet.controlnet_cond_embedding.conv_in - - new_condition_conv_in = torch.nn.Conv2d( - control_in_channels, - condition_conv_in.out_channels, - kernel_size=condition_conv_in.kernel_size, - stride=condition_conv_in.stride, - padding=condition_conv_in.padding, - ) - - with torch.no_grad(): - new_condition_conv_in.weight[:, : condition_conv_in.in_channels, :, :] = condition_conv_in.weight - if condition_conv_in.bias is not None: - new_condition_conv_in.bias = condition_conv_in.bias - - model.unet.controlnet.controlnet_cond_embedding.conv_in = new_condition_conv_in - - rank_zero_print(f"Loaded Init ...") - - if getattr(config, "resume_from", None): - cnet_ckpt_path = config.resume_from - sds = torch.load(cnet_ckpt_path, map_location="cpu")["state_dict"] - sd0 = {k[len(model_unet_prefix) :]: v for k, v in sds.items() if model_unet_prefix in k} - # model.unet.unet.unet.load_state_dict(sd0, strict=True) - model_unet.load_state_dict(sd0, strict=True) - if hasattr(model.unet, "controlnet"): - sd1 = {k[16:]: v for k, v in sds.items() if "unet.controlnet." in k} - model.unet.controlnet.load_state_dict(sd1, strict=True) - rank_zero_print(f"Loaded {cnet_ckpt_path} ...") - - if opt.resume and opt.resume_weights_only: - model = model.__class__.load_from_checkpoint(opt.resume, **config.model.params) - - model.logdir = logdir - - # trainer and callbacks - trainer_kwargs = dict() - - # logger - default_logger_cfg = { - "target": "pytorch_lightning.loggers.TensorBoardLogger", - "params": { - "name": "tensorboard", - "save_dir": logdir, - "version": "0", - }, - } - logger_cfg = OmegaConf.merge(default_logger_cfg) - trainer_kwargs["logger"] = instantiate_from_config(logger_cfg) - - # model checkpoint - default_modelckpt_cfg = { - "target": "pytorch_lightning.callbacks.ModelCheckpoint", - "params": { - "dirpath": ckptdir, - "filename": "{step:08}", - "verbose": True, - "save_last": True, - "every_n_train_steps": 5000, - "save_top_k": -1, # save all checkpoints - }, - } - - if "modelcheckpoint" in lightning_config: - modelckpt_cfg = lightning_config.modelcheckpoint - else: - modelckpt_cfg = OmegaConf.create() - modelckpt_cfg = OmegaConf.merge(default_modelckpt_cfg, modelckpt_cfg) - - # callbacks - default_callbacks_cfg = { - "setup_callback": { - "target": "train.SetupCallback", - "params": { - "resume": opt.resume, - "logdir": logdir, - "ckptdir": ckptdir, - "cfgdir": cfgdir, - "config": config, - }, - }, - "learning_rate_logger": { - "target": "pytorch_lightning.callbacks.LearningRateMonitor", - "params": { - "logging_interval": "step", - }, - }, - "code_snapshot": { - "target": "train.CodeSnapshot", - "params": { - "savedir": codedir, - }, - }, - } - default_callbacks_cfg["checkpoint_callback"] = modelckpt_cfg - - if "callbacks" in lightning_config: - callbacks_cfg = lightning_config.callbacks - else: - callbacks_cfg = OmegaConf.create() - callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg) - - trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg] - - trainer_kwargs["precision"] = "bf16" - trainer_kwargs["strategy"] = DDPStrategy(find_unused_parameters=False) - - # trainer - trainer = Trainer(**trainer_config, **trainer_kwargs, num_nodes=opt.num_nodes, inference_mode=False) - trainer.logdir = logdir - - # data - data = instantiate_from_config(config.data) - data.prepare_data() - data.setup("fit") - - # configure learning rate - base_lr = config.model.base_learning_rate - if "accumulate_grad_batches" in lightning_config.trainer: - accumulate_grad_batches = lightning_config.trainer.accumulate_grad_batches - else: - accumulate_grad_batches = 1 - rank_zero_print(f"accumulate_grad_batches = {accumulate_grad_batches}") - lightning_config.trainer.accumulate_grad_batches = accumulate_grad_batches - model.learning_rate = base_lr - rank_zero_print("++++ NOT USING LR SCALING ++++") - rank_zero_print(f"Setting learning rate to {model.learning_rate:.2e}") - - # run training loop - if opt.resume and not opt.resume_weights_only: - trainer.fit(model, data, ckpt_path=opt.resume) - else: - trainer.fit(model, data) diff --git a/hy3dpaint/train_examples/001/render_cond/mesh.ply b/hy3dpaint/train_examples/001/render_cond/mesh.ply deleted file mode 100644 index 652534d9d9e977934138ab66a365db55f89b443a..0000000000000000000000000000000000000000 Binary files a/hy3dpaint/train_examples/001/render_cond/mesh.ply and /dev/null differ diff --git a/hy3dpaint/train_examples/001/render_cond/transforms.json b/hy3dpaint/train_examples/001/render_cond/transforms.json deleted file mode 100644 index a9eae56d4335fca08a2de91bee8b97e87e7449df..0000000000000000000000000000000000000000 --- a/hy3dpaint/train_examples/001/render_cond/transforms.json +++ /dev/null @@ -1,838 +0,0 @@ -{ - "aabb": [ - [ - -0.5, - -0.5, - -0.5 - ], - [ - 0.5, - 0.5, - 0.5 - ] - ], - "scale": 0.5121457915021708, - "offset": [ - 0.00032189488410949707, - -0.03408946841955185, - -0.2757369577884674 - ], - "frames": [ - { - "file_path": "000.png", - "camera_angle_x": 0.47011032380678586, - "proj_type": 0, - "azimuth": 2.2539364673716054, - "elevation": -1.3001257577149983, - "cam_dis": 3.71849691615005, - "transform_matrix": [ - [ - -0.7755944728851318, - -0.6082496643066406, - -0.1687772274017334, - -0.6275975704193115 - ], - [ - -0.6312316656112671, - 0.7473564147949219, - 0.2073766440153122, - 0.7711292505264282 - ], - [ - -2.3015424233108206e-08, - 0.26737767457962036, - -0.9635918736457825, - -3.583113193511963 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "001.png", - "camera_angle_x": 0.5079176868006297, - "proj_type": 0, - "azimuth": 5.395529120961399, - "elevation": -0.9221878379342444, - "cam_dis": 3.4470348893999567, - "transform_matrix": [ - [ - 0.7755942940711975, - 0.5030443668365479, - 0.3813132345676422, - 1.3143998384475708 - ], - [ - 0.6312316656112671, - -0.6180905103683472, - -0.4685195982456207, - -1.615003228187561 - ], - [ - -4.003486608894491e-08, - 0.6040781140327454, - -0.7969250082969666, - -2.7470288276672363 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "002.png", - "camera_angle_x": 0.7090660480350758, - "proj_type": 0, - "azimuth": 3.824732794166502, - "elevation": -0.6818860998422269, - "cam_dis": 2.494654312477568, - "transform_matrix": [ - [ - 0.6312316060066223, - -0.4888249933719635, - -0.6021600961685181, - -1.5021814107894897 - ], - [ - -0.7755944132804871, - -0.39783918857574463, - -0.49007895588874817, - -1.222577691078186 - ], - [ - -2.0854262317016037e-08, - 0.7763853669166565, - -0.6302586197853088, - -1.5722770690917969 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "003.png", - "camera_angle_x": 0.7207915782104436, - "proj_type": 0, - "azimuth": 6.966325447756295, - "elevation": -0.48204466653151123, - "cam_dis": 2.4558020914544034, - "transform_matrix": [ - [ - -0.6312316060066223, - 0.35955917835235596, - 0.6872146725654602, - 1.687662959098816 - ], - [ - 0.7755944728851318, - 0.292633593082428, - 0.5593021512031555, - 1.37353515625 - ], - [ - 7.981980587601356e-08, - 0.8860490322113037, - -0.46359172463417053, - -1.1384897232055664 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "004.png", - "camera_angle_x": 1.1560563050396928, - "proj_type": 0, - "azimuth": 3.0393346307690536, - "elevation": -0.30147096310802857, - "cam_dis": 1.5850428518685533, - "transform_matrix": [ - [ - -0.10207992047071457, - -0.2953740656375885, - -0.9499125480651855, - -1.5056520700454712 - ], - [ - -0.99477618932724, - 0.03031015396118164, - 0.09747619181871414, - 0.15450391173362732 - ], - [ - 4.873770365065866e-08, - 0.9549007415771484, - -0.2969251275062561, - -0.47063907980918884 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "005.png", - "camera_angle_x": 1.1080063733527263, - "proj_type": 0, - "azimuth": 6.180927284358847, - "elevation": -0.13062968094200955, - "cam_dis": 1.6461361738590004, - "transform_matrix": [ - [ - 0.10207971930503845, - 0.1295779049396515, - 0.9863009452819824, - 1.6235853433609009 - ], - [ - 0.9947763085365295, - -0.013296757824718952, - -0.10121013969182968, - -0.16660575568675995 - ], - [ - -1.0836848218787054e-07, - 0.9914802312850952, - -0.1302584558725357, - -0.21442320942878723 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "006.png", - "camera_angle_x": 1.0631377508801954, - "proj_type": 0, - "azimuth": 4.610130957563951, - "elevation": 0.012136358647873546, - "cam_dis": 1.7085198579433762, - "transform_matrix": [ - [ - 0.99477618932724, - 0.0012388412142172456, - -0.10207240283489227, - -0.1743927001953125 - ], - [ - -0.10207992047071457, - 0.012072510085999966, - -0.9947029948234558, - -1.6994696855545044 - ], - [ - -2.03229033601815e-09, - 0.9999264478683472, - 0.012135906144976616, - 0.020734701305627823 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "007.png", - "camera_angle_x": 0.5894741433793819, - "proj_type": 0, - "azimuth": 7.751723611153743, - "elevation": 0.06774341874087009, - "cam_dis": 2.9812749570228485, - "transform_matrix": [ - [ - -0.99477618932724, - -0.006909956689924002, - 0.10184576362371445, - 0.30363020300865173 - ], - [ - 0.10207990556955338, - -0.06733796000480652, - 0.9924944639205933, - 2.9588990211486816 - ], - [ - -8.816746444040291e-09, - 0.9977062940597534, - 0.06769154965877533, - 0.20180732011795044 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "008.png", - "camera_angle_x": 1.1731446025768484, - "proj_type": 0, - "azimuth": 2.6466355490703295, - "elevation": 0.12356134208787495, - "cam_dis": 1.5646078921687379, - "transform_matrix": [ - [ - -0.47499409317970276, - 0.10845614224672318, - -0.8732801675796509, - -1.3663408756256104 - ], - [ - -0.8799892067909241, - -0.058541711419820786, - 0.4713726341724396, - 0.7375132441520691 - ], - [ - 5.2883926571212214e-08, - 0.9923761487007141, - 0.12324699014425278, - 0.19283349812030792 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "009.png", - "camera_angle_x": 0.9240898657044195, - "proj_type": 0, - "azimuth": 5.788228202660123, - "elevation": 0.17976943361645015, - "cam_dis": 1.9427212715365068, - "transform_matrix": [ - [ - 0.4749939441680908, - -0.15734440088272095, - 0.8658080101013184, - 1.6820236444473267 - ], - [ - 0.8799890875816345, - 0.08493020385503769, - -0.4673393964767456, - -0.9079101085662842 - ], - [ - -2.0080790363863343e-08, - 0.9838849902153015, - 0.17880268394947052, - 0.3473638594150543 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "010.png", - "camera_angle_x": 1.2064601116642644, - "proj_type": 0, - "azimuth": 4.217431875865226, - "elevation": 0.23655841484996287, - "cam_dis": 1.526559413456265, - "transform_matrix": [ - [ - 0.8799890875816345, - 0.11131872981786728, - -0.46176549792289734, - -0.7049124240875244 - ], - [ - -0.47499391436576843, - 0.20623266696929932, - -0.855481743812561, - -1.3059436082839966 - ], - [ - 2.947104427164504e-08, - 0.9721503853797913, - 0.23435820639133453, - 0.35776183009147644 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "011.png", - "camera_angle_x": 0.8106133703864192, - "proj_type": 0, - "azimuth": 7.359024529455018, - "elevation": 0.2941368085661682, - "cam_dis": 2.196358226678013, - "transform_matrix": [ - [ - -0.8799890875816345, - -0.13770732283592224, - 0.454594224691391, - 0.9984517693519592 - ], - [ - 0.47499391436576843, - -0.25512099266052246, - 0.8421959280967712, - 1.849764108657837 - ], - [ - -1.5272490827555885e-08, - 0.9570527672767639, - 0.2899138629436493, - 0.6367546319961548 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "012.png", - "camera_angle_x": 0.6777962932590744, - "proj_type": 0, - "azimuth": 3.4320337124677778, - "elevation": 0.35273892986621935, - "cam_dis": 2.604994196713999, - "transform_matrix": [ - [ - 0.2863747477531433, - 0.33100035786628723, - -0.8991264700889587, - -2.3422188758850098 - ], - [ - -0.9581177234649658, - 0.09893371164798737, - -0.2687426805496216, - -0.7000733017921448 - ], - [ - 5.095670019272802e-08, - 0.9384300708770752, - 0.3454693555831909, - 0.8999457955360413 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "013.png", - "camera_angle_x": 1.2201901414150222, - "proj_type": 0, - "azimuth": 6.57362636605757, - "elevation": 0.4126354310485474, - "cam_dis": 1.511532217466212, - "transform_matrix": [ - [ - -0.2863748371601105, - -0.38422903418540955, - 0.8777002096176147, - 1.3266719579696655 - ], - [ - 0.9581177234649658, - -0.11484345048666, - 0.2623385488986969, - 0.39653322100639343 - ], - [ - -1.850063036101801e-08, - 0.9160672426223755, - 0.40102484822273254, - 0.6061621308326721 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "014.png", - "camera_angle_x": 0.5705820549491503, - "proj_type": 0, - "azimuth": 5.002830039262674, - "elevation": 0.4741478890444091, - "cam_dis": 3.077158450915264, - "transform_matrix": [ - [ - 0.958117663860321, - -0.13075314462184906, - 0.2547825574874878, - 0.7840063571929932 - ], - [ - 0.2863748073577881, - 0.4374578595161438, - -0.8524201512336731, - -2.6230318546295166 - ], - [ - -1.4379679669218604e-08, - 0.8896821141242981, - 0.4565805196762085, - 1.4049705266952515 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "015.png", - "camera_angle_x": 1.1922232332059943, - "proj_type": 0, - "azimuth": 8.144422692852467, - "elevation": 0.5376699116109869, - "cam_dis": 1.5425377326560552, - "transform_matrix": [ - [ - -0.9581177234649658, - 0.1466628760099411, - -0.24596858024597168, - -0.3794158399105072 - ], - [ - -0.2863748371601105, - -0.4906865656375885, - 0.8229314684867859, - 1.2694027423858643 - ], - [ - 1.9456553701502344e-08, - 0.8589044213294983, - 0.5121359825134277, - 0.789989173412323 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "016.png", - "camera_angle_x": 0.4843348255851337, - "proj_type": 0, - "azimuth": 2.4502860082209676, - "elevation": 0.6036991169616068, - "cam_dis": 3.6113379310693747, - "transform_matrix": [ - [ - -0.6375443935394287, - 0.4373573660850525, - -0.6342363953590393, - -2.2904417514801025 - ], - [ - -0.7704136967658997, - -0.3619285523891449, - 0.5248528718948364, - 1.895421028137207 - ], - [ - 4.0876805940115446e-08, - 0.8232413530349731, - 0.5676915049552917, - 2.050126314163208 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "017.png", - "camera_angle_x": 1.0484863420149597, - "proj_type": 0, - "azimuth": 5.59187866181076, - "elevation": 0.6728881287032298, - "cam_dis": 1.7301201355596483, - "transform_matrix": [ - [ - 0.6375443935394287, - -0.4801580607891083, - 0.6024826765060425, - 1.0423673391342163 - ], - [ - 0.7704136371612549, - 0.39734768867492676, - -0.4985755980014801, - -0.8625956773757935 - ], - [ - 3.110944390982695e-09, - 0.7820249199867249, - 0.6232471466064453, - 1.0782924890518188 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "018.png", - "camera_angle_x": 0.8327046868500112, - "proj_type": 0, - "azimuth": 4.021082335015864, - "elevation": 0.7461309541186378, - "cam_dis": 2.1413633661978952, - "transform_matrix": [ - [ - 0.7704136371612549, - 0.432766854763031, - -0.46816205978393555, - -1.0025050640106201 - ], - [ - -0.6375443935394287, - 0.5229588747024536, - -0.5657307505607605, - -1.2114349603652954 - ], - [ - 2.405637644642411e-08, - 0.7343207597732544, - 0.6788026690483093, - 1.4535633325576782 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "019.png", - "camera_angle_x": 1.0378875875393918, - "proj_type": 0, - "azimuth": 7.162674988605656, - "elevation": 0.8247207741120075, - "cam_dis": 1.7461482839234206, - "transform_matrix": [ - [ - -0.7704136371612549, - -0.46818602085113525, - 0.4327409267425537, - 0.75562983751297 - ], - [ - 0.6375443935394287, - -0.5657596588134766, - 0.5229275226593018, - 0.9131090044975281 - ], - [ - 5.7037539136217674e-09, - 0.6787620186805725, - 0.7343583106994629, - 1.2822984457015991 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "020.png", - "camera_angle_x": 1.181698605034973, - "proj_type": 0, - "azimuth": 3.235684171618416, - "elevation": 0.9106684775229277, - "cam_dis": 1.5546175155141897, - "transform_matrix": [ - [ - 0.09395270049571991, - 0.7864198088645935, - -0.6105054020881653, - -0.9491023421287537 - ], - [ - -0.9955766797065735, - 0.07421454787254333, - -0.057613469660282135, - -0.08956695348024368 - ], - [ - 1.1830872281848315e-08, - 0.613217830657959, - 0.7899138331413269, - 1.2280138731002808 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "021.png", - "camera_angle_x": 1.197224930669693, - "proj_type": 0, - "azimuth": 6.377276825208208, - "elevation": 1.0074435322145288, - "cam_dis": 1.536877375373375, - "transform_matrix": [ - [ - -0.09395267069339752, - -0.8417297005653381, - 0.5316617488861084, - 0.81709885597229 - ], - [ - 0.9955767393112183, - -0.0794341042637825, - 0.05017295107245445, - 0.07710976153612137 - ], - [ - -3.2885168366192374e-08, - 0.5340239405632019, - 0.8454694151878357, - 1.2993828058242798 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "022.png", - "camera_angle_x": 0.7684080505969957, - "proj_type": 0, - "azimuth": 4.806480498413312, - "elevation": 1.122126648780772, - "cam_dis": 2.310502047946327, - "transform_matrix": [ - [ - 0.9955766797065735, - -0.08465376496315002, - 0.040753625333309174, - 0.09416133165359497 - ], - [ - 0.0939527377486229, - 0.8970394134521484, - -0.43184858560562134, - -0.9977869987487793 - ], - [ - -3.0048674659610697e-09, - 0.43376728892326355, - 0.9010249376296997, - 2.081820011138916 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "023.png", - "camera_angle_x": 1.2027037528892426, - "proj_type": 0, - "azimuth": 7.948073152003104, - "elevation": 1.2750349593459749, - "cam_dis": 1.5307354901339616, - "transform_matrix": [ - [ - -0.9955767393112183, - 0.08987335115671158, - -0.027384238317608833, - -0.04191803187131882 - ], - [ - -0.0939527302980423, - -0.9523493051528931, - 0.29017895460128784, - 0.4441872537136078 - ], - [ - -4.757865657012417e-09, - 0.29146820306777954, - 0.9565805196762085, - 1.4642717838287354 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - } - ] -} \ No newline at end of file diff --git a/hy3dpaint/train_examples/001/render_tex/transforms.json b/hy3dpaint/train_examples/001/render_tex/transforms.json deleted file mode 100644 index f7c46d458a00b2d44c1e4d0d03a8c4749846d979..0000000000000000000000000000000000000000 --- a/hy3dpaint/train_examples/001/render_tex/transforms.json +++ /dev/null @@ -1,226 +0,0 @@ -{ - "aabb": [ - [ - -0.5, - -0.5, - -0.5 - ], - [ - 0.5, - 0.5, - 0.5 - ] - ], - "scale": 0.5121457915021708, - "offset": [ - 0.00032189488410949707, - -0.03408946841955185, - -0.2757369577884674 - ], - "frames": [ - { - "file_path": "000.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": -1.5707963267948966, - "elevation": 0, - "cam_dis": 1.5, - "transform_matrix": [ - [ - 1.0, - 0.0, - 6.123233601181349e-17, - 9.184850732644269e-17 - ], - [ - 6.123233601181349e-17, - 3.422854177870249e-08, - -0.9999999403953552, - -1.5 - ], - [ - 0.0, - 0.9999999403953552, - 3.422854177870249e-08, - 0.0 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "001.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": 0, - "elevation": 0, - "cam_dis": 1.5, - "transform_matrix": [ - [ - -2.220446049250313e-16, - 0.0, - 1.0, - 1.5 - ], - [ - 1.0, - -2.220446049250313e-16, - 0.0, - 0.0 - ], - [ - 0.0, - 1.0, - -2.220446049250313e-16, - 0.0 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "002.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": 1.5707963267948966, - "elevation": 0, - "cam_dis": 1.5, - "transform_matrix": [ - [ - -0.9999999403953552, - 0.0, - 6.123233601181349e-17, - 9.184850732644269e-17 - ], - [ - 6.123233601181349e-17, - 3.422854177870249e-08, - 0.9999999403953552, - 1.5 - ], - [ - 0.0, - 0.9999999403953552, - 3.422854177870249e-08, - 0.0 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "003.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": 3.141592653589793, - "elevation": 0, - "cam_dis": 1.5, - "transform_matrix": [ - [ - -2.220446049250313e-16, - 0.0, - -1.0, - -1.5 - ], - [ - -1.0, - -2.220446049250313e-16, - 0.0, - 1.8369701465288538e-16 - ], - [ - 0.0, - 1.0, - -2.220446049250313e-16, - 0.0 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "004.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": -1.5707963267948966, - "elevation": 1.5707963267948966, - "cam_dis": 1.5, - "transform_matrix": [ - [ - 1.0, - -6.123233601181349e-17, - 2.812049544789739e-33, - 5.624099089579478e-33 - ], - [ - 6.123233601181349e-17, - 1.0, - -3.0616171314629196e-17, - -9.184850732644269e-17 - ], - [ - -9.373498482632464e-34, - 3.0616171314629196e-17, - 1.0, - 1.5 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - }, - { - "file_path": "005.png", - "camera_angle_x": 0.9232195630055808, - "proj_type": 1, - "azimuth": -1.5707963267948966, - "elevation": -1.5707963267948966, - "cam_dis": 1.5, - "transform_matrix": [ - [ - 1.0, - 6.123233601181349e-17, - 2.812049544789739e-33, - 5.624099089579478e-33 - ], - [ - 6.123233601181349e-17, - -1.0, - -3.0616171314629196e-17, - -9.184850732644269e-17 - ], - [ - 9.373498482632464e-34, - 3.0616171314629196e-17, - -1.0, - -1.5 - ], - [ - 0, - 0, - 0, - 1 - ] - ] - } - ] -} \ No newline at end of file diff --git a/hy3dpaint/train_examples/examples.json b/hy3dpaint/train_examples/examples.json deleted file mode 100644 index 11b519e210c18c687970a915ec98c27b02a454ab..0000000000000000000000000000000000000000 --- a/hy3dpaint/train_examples/examples.json +++ /dev/null @@ -1,3 +0,0 @@ -[ - "hy3dpaint/train_examples/001" -] \ No newline at end of file diff --git a/hy3dpaint/utils/__init__.py b/hy3dpaint/utils/__init__.py deleted file mode 100644 index 1614ff826f65b4720649d343c43fb09dbb6b9fa5..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/__init__.py +++ /dev/null @@ -1,13 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. diff --git a/hy3dpaint/utils/image_super_utils.py b/hy3dpaint/utils/image_super_utils.py deleted file mode 100644 index 8d56d4f2f27990392db9d531d77b220bfab61eb3..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/image_super_utils.py +++ /dev/null @@ -1,41 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import numpy as np -from PIL import Image - - -class imageSuperNet: - def __init__(self, config) -> None: - from realesrgan import RealESRGANer - from basicsr.archs.rrdbnet_arch import RRDBNet - - model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) - upsampler = RealESRGANer( - scale=4, - model_path=config.realesrgan_ckpt_path, - dni_weight=None, - model=model, - tile=0, - tile_pad=10, - pre_pad=0, - half=True, - gpu_id=None, - ) - self.upsampler = upsampler - - def __call__(self, image): - output, _ = self.upsampler.enhance(np.array(image)) - output = Image.fromarray(output) - return output diff --git a/hy3dpaint/utils/multiview_utils.py b/hy3dpaint/utils/multiview_utils.py deleted file mode 100644 index e27d6307b09ecdc7b5116b6fd2f2646705ef9cb6..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/multiview_utils.py +++ /dev/null @@ -1,128 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import torch -import random -import numpy as np -from PIL import Image -from typing import List -import huggingface_hub -from omegaconf import OmegaConf -from diffusers import DiffusionPipeline -from diffusers import EulerAncestralDiscreteScheduler, DDIMScheduler, UniPCMultistepScheduler - - -class multiviewDiffusionNet: - def __init__(self, config) -> None: - self.device = config.device - - cfg_path = config.multiview_cfg_path - custom_pipeline = config.custom_pipeline - cfg = OmegaConf.load(cfg_path) - self.cfg = cfg - self.mode = self.cfg.model.params.stable_diffusion_config.custom_pipeline[2:] - - model_path = huggingface_hub.snapshot_download( - repo_id=config.multiview_pretrained_path, - allow_patterns=["hunyuan3d-paintpbr-v2-1/*"], - ) - - model_path = os.path.join(model_path, "hunyuan3d-paintpbr-v2-1") - pipeline = DiffusionPipeline.from_pretrained( - model_path, - custom_pipeline=custom_pipeline, - torch_dtype=torch.float16 - ) - - pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing") - pipeline.set_progress_bar_config(disable=True) - pipeline.eval() - setattr(pipeline, "view_size", cfg.model.params.get("view_size", 320)) - self.pipeline = pipeline.to(self.device) - - if hasattr(self.pipeline.unet, "use_dino") and self.pipeline.unet.use_dino: - from hunyuanpaintpbr.unet.modules import Dino_v2 - self.dino_v2 = Dino_v2(config.dino_ckpt_path).to(torch.float16) - self.dino_v2 = self.dino_v2.to(self.device) - - def seed_everything(self, seed): - random.seed(seed) - np.random.seed(seed) - torch.manual_seed(seed) - os.environ["PL_GLOBAL_SEED"] = str(seed) - - @torch.no_grad() - def __call__(self, images, conditions, prompt=None, custom_view_size=None, resize_input=False): - pils = self.forward_one( - images, conditions, prompt=prompt, custom_view_size=custom_view_size, resize_input=resize_input - ) - return pils - - def forward_one(self, input_images, control_images, prompt=None, custom_view_size=None, resize_input=False): - self.seed_everything(0) - custom_view_size = custom_view_size if custom_view_size is not None else self.pipeline.view_size - if not isinstance(input_images, List): - input_images = [input_images] - if not resize_input: - input_images = [ - input_image.resize((self.pipeline.view_size, self.pipeline.view_size)) for input_image in input_images - ] - else: - input_images = [input_image.resize((custom_view_size, custom_view_size)) for input_image in input_images] - for i in range(len(control_images)): - control_images[i] = control_images[i].resize((custom_view_size, custom_view_size)) - if control_images[i].mode == "L": - control_images[i] = control_images[i].point(lambda x: 255 if x > 1 else 0, mode="1") - kwargs = dict(generator=torch.Generator(device=self.pipeline.device).manual_seed(0)) - - num_view = len(control_images) // 2 - normal_image = [[control_images[i] for i in range(num_view)]] - position_image = [[control_images[i + num_view] for i in range(num_view)]] - - kwargs["width"] = custom_view_size - kwargs["height"] = custom_view_size - kwargs["num_in_batch"] = num_view - kwargs["images_normal"] = normal_image - kwargs["images_position"] = position_image - - if hasattr(self.pipeline.unet, "use_dino") and self.pipeline.unet.use_dino: - dino_hidden_states = self.dino_v2(input_images[0]) - kwargs["dino_hidden_states"] = dino_hidden_states - - sync_condition = None - - infer_steps_dict = { - "EulerAncestralDiscreteScheduler": 30, - "UniPCMultistepScheduler": 15, - "DDIMScheduler": 50, - "ShiftSNRScheduler": 15, - } - - mvd_image = self.pipeline( - input_images[0:1], - num_inference_steps=infer_steps_dict[self.pipeline.scheduler.__class__.__name__], - prompt=prompt, - sync_condition=sync_condition, - guidance_scale=3.0, - **kwargs, - ).images - - if "pbr" in self.mode: - mvd_image = {"albedo": mvd_image[:num_view], "mr": mvd_image[num_view:]} - # mvd_image = {'albedo':mvd_image[:num_view]} - else: - mvd_image = {"hdr": mvd_image} - - return mvd_image diff --git a/hy3dpaint/utils/pipeline_utils.py b/hy3dpaint/utils/pipeline_utils.py deleted file mode 100644 index 6bc53826956211739250bc32914ce11a72809ce2..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/pipeline_utils.py +++ /dev/null @@ -1,135 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import torch -import numpy as np - - -class ViewProcessor: - def __init__(self, config, render): - self.config = config - self.render = render - - def render_normal_multiview(self, camera_elevs, camera_azims, use_abs_coor=True): - normal_maps = [] - for elev, azim in zip(camera_elevs, camera_azims): - normal_map = self.render.render_normal(elev, azim, use_abs_coor=use_abs_coor, return_type="pl") - normal_maps.append(normal_map) - - return normal_maps - - def render_position_multiview(self, camera_elevs, camera_azims): - position_maps = [] - for elev, azim in zip(camera_elevs, camera_azims): - position_map = self.render.render_position(elev, azim, return_type="pl") - position_maps.append(position_map) - - return position_maps - - def bake_view_selection( - self, candidate_camera_elevs, candidate_camera_azims, candidate_view_weights, max_selected_view_num - ): - - original_resolution = self.render.default_resolution - self.render.set_default_render_resolution(1024) - - selected_camera_elevs = [] - selected_camera_azims = [] - selected_view_weights = [] - selected_alpha_maps = [] - viewed_tri_idxs = [] - viewed_masks = [] - - # 计算每个三角片的面积 - face_areas = self.render.get_face_areas(from_one_index=True) - total_area = face_areas.sum() - face_area_ratios = face_areas / total_area - - candidate_view_num = len(candidate_camera_elevs) - self.render.set_boundary_unreliable_scale(2) - - for elev, azim in zip(candidate_camera_elevs, candidate_camera_azims): - viewed_tri_idx = self.render.render_alpha(elev, azim, return_type="np") - viewed_tri_idxs.append(set(np.unique(viewed_tri_idx.flatten()))) - viewed_masks.append(viewed_tri_idx[0, :, :, 0] > 0) - - is_selected = [False for _ in range(candidate_view_num)] - total_viewed_tri_idxs = set() - total_viewed_area = 0.0 - - for idx in range(6): - selected_camera_elevs.append(candidate_camera_elevs[idx]) - selected_camera_azims.append(candidate_camera_azims[idx]) - selected_view_weights.append(candidate_view_weights[idx]) - selected_alpha_maps.append(viewed_masks[idx]) - is_selected[idx] = True - total_viewed_tri_idxs.update(viewed_tri_idxs[idx]) - - total_viewed_area = face_area_ratios[list(total_viewed_tri_idxs)].sum() - for iter in range(max_selected_view_num - len(selected_view_weights)): - max_inc = 0 - max_idx = -1 - - for idx, (elev, azim, weight) in enumerate( - zip(candidate_camera_elevs, candidate_camera_azims, candidate_view_weights) - ): - if is_selected[idx]: - continue - new_tri_idxs = viewed_tri_idxs[idx] - total_viewed_tri_idxs - new_inc_area = face_area_ratios[list(new_tri_idxs)].sum() - - if new_inc_area > max_inc: - max_inc = new_inc_area - max_idx = idx - - if max_inc > 0.01: - is_selected[max_idx] = True - selected_camera_elevs.append(candidate_camera_elevs[max_idx]) - selected_camera_azims.append(candidate_camera_azims[max_idx]) - selected_view_weights.append(candidate_view_weights[max_idx]) - selected_alpha_maps.append(viewed_masks[max_idx]) - total_viewed_tri_idxs = total_viewed_tri_idxs.union(viewed_tri_idxs[max_idx]) - total_viewed_area += max_inc - else: - break - - self.render.set_default_render_resolution(original_resolution) - - return selected_camera_elevs, selected_camera_azims, selected_view_weights - - def bake_from_multiview(self, views, camera_elevs, camera_azims, view_weights): - project_textures, project_weighted_cos_maps = [], [] - project_boundary_maps = [] - - for view, camera_elev, camera_azim, weight in zip(views, camera_elevs, camera_azims, view_weights): - project_texture, project_cos_map, project_boundary_map = self.render.back_project( - view, camera_elev, camera_azim - ) - project_cos_map = weight * (project_cos_map**self.config.bake_exp) - project_textures.append(project_texture) - project_weighted_cos_maps.append(project_cos_map) - project_boundary_maps.append(project_boundary_map) - texture, ori_trust_map = self.render.fast_bake_texture(project_textures, project_weighted_cos_maps) - return texture, ori_trust_map > 1e-8 - - def texture_inpaint(self, texture, mask, defualt=None): - if defualt is not None: - mask = mask.astype(bool) - inpaint_value = torch.tensor(defualt, dtype=texture.dtype, device=texture.device) - texture[~mask] = inpaint_value - else: - texture_np = self.render.uv_inpaint(texture, mask) - texture = torch.tensor(texture_np / 255).float().to(texture.device) - - return texture diff --git a/hy3dpaint/utils/simplify_mesh_utils.py b/hy3dpaint/utils/simplify_mesh_utils.py deleted file mode 100644 index 7d3f10aa182c25ac4847f134a7189ab32b112054..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/simplify_mesh_utils.py +++ /dev/null @@ -1,37 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import trimesh -import pymeshlab - - -def remesh_mesh(mesh_path, remesh_path): - mesh = mesh_simplify_trimesh(mesh_path, remesh_path) - - -def mesh_simplify_trimesh(inputpath, outputpath, target_count=40000): - # 先去除离散面 - ms = pymeshlab.MeshSet() - if inputpath.endswith(".glb"): - ms.load_new_mesh(inputpath, load_in_a_single_layer=True) - else: - ms.load_new_mesh(inputpath) - ms.save_current_mesh(outputpath.replace(".glb", ".obj"), save_textures=False) - # 调用减面函数 - courent = trimesh.load(outputpath.replace(".glb", ".obj"), force="mesh") - face_num = courent.faces.shape[0] - - if face_num > target_count: - courent = courent.simplify_quadric_decimation(target_count) - courent.export(outputpath) diff --git a/hy3dpaint/utils/torchvision_fix.py b/hy3dpaint/utils/torchvision_fix.py deleted file mode 100644 index 40cef19f218704e0b92b8119b19d29a631e579e6..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/torchvision_fix.py +++ /dev/null @@ -1,111 +0,0 @@ -# Torchvision compatibility fix for functional_tensor module -# This file helps resolve compatibility issues between different torchvision versions - -import sys -import torch -import torchvision - -def fix_torchvision_functional_tensor(): - """ - Fix torchvision.transforms.functional_tensor import issue - """ - try: - # Check if the module exists in the expected location - import torchvision.transforms.functional_tensor - print("torchvision.transforms.functional_tensor is available") - return True - except ImportError: - print("torchvision.transforms.functional_tensor not found, applying compatibility fix...") - - try: - # Create a mock functional_tensor module with the required functions - import torchvision.transforms.functional as F - - class FunctionalTensorMock: - """Mock module to replace functional_tensor""" - - @staticmethod - def _get_grayscale_weights(img): - """Helper to create grayscale weights based on image dimensions""" - weights = torch.tensor([0.299, 0.587, 0.114], device=img.device, dtype=img.dtype) - return weights.view(1, 3, 1, 1) if len(img.shape) == 4 else weights.view(3, 1, 1) - - @staticmethod - def _try_import_fallback(module_names, attr_name): - """Helper to try importing from multiple modules""" - for module_name in module_names: - try: - module = __import__(module_name, fromlist=[attr_name]) - if hasattr(module, attr_name): - return getattr(module, attr_name) - except ImportError: - continue - return None - - @staticmethod - def rgb_to_grayscale(img, num_output_channels=1): - """Convert RGB image to grayscale""" - if hasattr(F, 'rgb_to_grayscale'): - return F.rgb_to_grayscale(img, num_output_channels) - - # Fallback implementation - weights = FunctionalTensorMock._get_grayscale_weights(img) - grayscale = torch.sum(img * weights, dim=-3, keepdim=True) - - if num_output_channels == 3: - repeat_dims = (1, 3, 1, 1) if len(img.shape) == 4 else (3, 1, 1) - grayscale = grayscale.repeat(*repeat_dims) - - return grayscale - - @staticmethod - def resize(img, size, interpolation=2, antialias=None): - """Resize function wrapper""" - # Try v2.functional first, then regular functional, then torch.nn.functional - resize_func = FunctionalTensorMock._try_import_fallback([ - 'torchvision.transforms.v2.functional', - 'torchvision.transforms.functional' - ], 'resize') - - if resize_func: - try: - return resize_func(img, size, interpolation=interpolation, antialias=antialias) - except TypeError: - # Fallback for older versions without antialias parameter - return resize_func(img, size, interpolation=interpolation) - - # Final fallback using torch.nn.functional - import torch.nn.functional as torch_F - size = (size, size) if isinstance(size, int) else size - img_input = img.unsqueeze(0) if len(img.shape) == 3 else img - return torch_F.interpolate(img_input, size=size, mode='bilinear', align_corners=False) - - def __getattr__(self, name): - """Fallback to regular functional module""" - func = self._try_import_fallback([ - 'torchvision.transforms.functional', - 'torchvision.transforms.v2.functional' - ], name) - - if func: - return func - - raise AttributeError(f"'{name}' not found in functional_tensor mock") - - # Create the mock module instance and monkey patch - sys.modules['torchvision.transforms.functional_tensor'] = FunctionalTensorMock() - print("Applied compatibility fix: created functional_tensor mock module") - return True - - except Exception as e: - print(f"Failed to create functional_tensor mock: {e}") - return False - -def apply_fix(): - """Apply the torchvision compatibility fix""" - print(f"Torchvision version: {torchvision.__version__}") - return fix_torchvision_functional_tensor() - -if __name__ == "__main__": - apply_fix() - \ No newline at end of file diff --git a/hy3dpaint/utils/uvwrap_utils.py b/hy3dpaint/utils/uvwrap_utils.py deleted file mode 100644 index f55a924f16aa6b2dce3347f39ff7e8ef190065f4..0000000000000000000000000000000000000000 --- a/hy3dpaint/utils/uvwrap_utils.py +++ /dev/null @@ -1,32 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import trimesh -import xatlas - - -def mesh_uv_wrap(mesh): - if isinstance(mesh, trimesh.Scene): - mesh = mesh.dump(concatenate=True) - - if len(mesh.faces) > 500000000: - raise ValueError("The mesh has more than 500,000,000 faces, which is not supported.") - - vmapping, indices, uvs = xatlas.parametrize(mesh.vertices, mesh.faces) - - mesh.vertices = mesh.vertices[vmapping] - mesh.faces = indices - mesh.visual.uv = uvs - - return mesh diff --git a/hy3dshape/.gitignore b/hy3dshape/.gitignore deleted file mode 100644 index 35000cb2173be759506038ac0372d2ba5177a3c8..0000000000000000000000000000000000000000 --- a/hy3dshape/.gitignore +++ /dev/null @@ -1,169 +0,0 @@ -# Byte-compiled / optimized / DLL files -__pycache__/ -*.py[cod] -*$py.class - -# C extensions -*.so - -# Distribution / packaging -.Python -build/ -develop-eggs/ -dist/ -downloads/ -eggs/ -.eggs/ -lib/ -!hy3dgen/texgen/custom_rasterizer/lib/ -lib64/ -parts/ -sdist/ -var/ -wheels/ -share/python-wheels/ -*.egg-info/ -.installed.cfg -*.egg -MANIFEST - -# PyInstaller -# Usually these files are written by a python script from a template -# before PyInstaller builds the exe, so as to inject date/other infos into it. -*.manifest -*.spec - -# Installer logs -pip-log.txt -pip-delete-this-directory.txt - -# Unit test / coverage reports -htmlcov/ -.tox/ -.nox/ -.coverage -.coverage.* -.cache -nosetests.xml -coverage.xml -*.cover -*.py,cover -.hypothesis/ -.pytest_cache/ -cover/ - -# Translations -*.mo -*.pot - -# Django stuff: -*.log -local_settings.py -db.sqlite3 -db.sqlite3-journal - -# Flask stuff: -instance/ -.webassets-cache - -# Scrapy stuff: -.scrapy - -# Sphinx documentation -docs/_build/ - -# PyBuilder -.pybuilder/ -target/ - -# Jupyter Notebook -.ipynb_checkpoints - -# IPython -profile_default/ -ipython_config.py - -# pyenv -# For a library or package, you might want to ignore these files since the code is -# intended to run in multiple environments; otherwise, check them in: -# .python-version - -# pipenv -# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. -# However, in case of collaboration, if having platform-specific dependencies or dependencies -# having no cross-platform support, pipenv may install dependencies that don't work, or not -# install all needed dependencies. -#Pipfile.lock - -# UV -# Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control. -# This is especially recommended for binary packages to ensure reproducibility, and is more -# commonly ignored for libraries. -#uv.lock - -# poetry -# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. -# This is especially recommended for binary packages to ensure reproducibility, and is more -# commonly ignored for libraries. -# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control -#poetry.lock - -# pdm -# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. -#pdm.lock -# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it -# in version control. -# https://pdm.fming.dev/latest/usage/project/#working-with-version-control -.pdm.toml -.pdm-python -.pdm-build/ - -# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm -__pypackages__/ - -# Celery stuff -celerybeat-schedule -celerybeat.pid - -# SageMath parsed files -*.sage.py - -# Environments -.env -.venv -env/ -venv/ -ENV/ -env.bak/ -venv.bak/ - -# Spyder project settings -.spyderproject -.spyproject - -# Rope project settings -.ropeproject - -# mkdocs documentation -/site - -# mypy -.mypy_cache/ -.dmypy.json -dmypy.json - -# Pyre type checker -.pyre/ - -# pytype static type analyzer -.pytype/ -.DS_Store -# Cython debug symbols -cython_debug/ -gradio_cache/ -# PyCharm -# JetBrains specific template is maintained in a separate JetBrains.gitignore that can -# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore -# and can be added to the global gitignore or merged into this file. For a more nuclear -# option (not recommended) you can uncomment the following to ignore the entire idea folder. -#.idea/ diff --git a/hy3dshape/LICENSE b/hy3dshape/LICENSE deleted file mode 100644 index 96313b341082fb9f72a00b7c77accd8196390af0..0000000000000000000000000000000000000000 --- a/hy3dshape/LICENSE +++ /dev/null @@ -1,81 +0,0 @@ -TENCENT HUNYUAN 3D 2.1 COMMUNITY LICENSE AGREEMENT -Tencent Hunyuan 3D 2.1 Release Date: June 13, 2025 -THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW. -By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan 3D 2.1 Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately. -1. DEFINITIONS. -a. “Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A. -b. “Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of Tencent Hunyuan 3D 2.1 Works or any portion or element thereof set forth herein. -c. “Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan 3D 2.1 made publicly available by Tencent. -d. “Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means. -e. “Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan 3D 2.1 Works for any purpose and in any field of use. -f. “Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan 3D 2.1 and Documentation (and any portion thereof) as made available by Tencent under this Agreement. -g. “Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; (ii) works based on Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1, to that model in order to cause that model to perform similarly to Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1 for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives. -h. “Output” shall mean the information and/or content output of Tencent Hunyuan 3D 2.1 or a Model Derivative that results from operating or otherwise using Tencent Hunyuan 3D 2.1 or a Model Derivative, including via a Hosted Service. -i. “Tencent,” “We” or “Us” shall mean THL Q Limited. -j. “Tencent Hunyuan 3D 2.1” shall mean the 3D generation models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at [ https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1]. -k. “Tencent Hunyuan 3D 2.1 Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof. -l. “Territory” shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea. -m. “Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You. -n. “including” shall mean including but not limited to. -2. GRANT OF RIGHTS. -We grant You, for the Territory only, a non-exclusive, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy. -3. DISTRIBUTION. -You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan 3D 2.1 Works, exclusively in the Territory, provided that You meet all of the following conditions: -a. You must provide all such Third Party recipients of the Tencent Hunyuan 3D 2.1 Works or products or services using them a copy of this Agreement; -b. You must cause any modified files to carry prominent notices stating that You changed the files; -c. You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan 3D 2.1 Works; and (ii) mark the products or services developed by using the Tencent Hunyuan 3D 2.1 Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and -d. All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan 3D 2.1 is licensed under the Tencent Hunyuan 3D 2.1 Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.” -You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement (including as regards the Territory). If You receive Tencent Hunyuan 3D 2.1 Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You. -4. ADDITIONAL COMMERCIAL TERMS. -If, on the Tencent Hunyuan 3D 2.1 version release date, the monthly active users of all products or services made available by or for Licensee is greater than 1 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights. -Subject to Tencent's written approval, you may request a license for the use of Tencent Hunyuan 3D 2.1 by submitting the following information to hunyuan3d@tencent.com: -a. Your company’s name and associated business sector that plans to use Tencent Hunyuan 3D 2.1. -b. Your intended use case and the purpose of using Tencent Hunyuan 3D 2.1. -c. Your plans to modify Tencent Hunyuan 3D 2.1 or create Model Derivatives. -5. RULES OF USE. -a. Your use of the Tencent Hunyuan 3D 2.1 Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan 3D 2.1 Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan 3D 2.1 Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan 3D 2.1 Works are subject to the use restrictions in these Sections 5(a) and 5(b). -b. You must not use the Tencent Hunyuan 3D 2.1 Works or any Output or results of the Tencent Hunyuan 3D 2.1 Works to improve any other AI model (other than Tencent Hunyuan 3D 2.1 or Model Derivatives thereof). -c. You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan 3D 2.1 Works, Output or results of the Tencent Hunyuan 3D 2.1 Works outside the Territory. Any such use outside the Territory is unlicensed and unauthorized under this Agreement. -6. INTELLECTUAL PROPERTY. -a. Subject to Tencent’s ownership of Tencent Hunyuan 3D 2.1 Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You. -b. No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan 3D 2.1 Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan 3D 2.1 Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) in the Territory solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent. -c. If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan 3D 2.1 Works. -d. Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses. -7. DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY. -a. We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan 3D 2.1 Works or to grant any license thereto. -b. UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN 3D 2.1 WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT. -c. TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING. -8. SURVIVAL AND TERMINATION. -a. The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. -b. We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan 3D 2.1 Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement. -9. GOVERNING LAW AND JURISDICTION. -a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. -b. Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute. -  -EXHIBIT A -ACCEPTABLE USE POLICY - -Tencent reserves the right to update this Acceptable Use Policy from time to time. -Last modified: November 5, 2024 - -Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan 3D 2.1. You agree not to use Tencent Hunyuan 3D 2.1 or Model Derivatives: -1. Outside the Territory; -2. In any way that violates any applicable national, federal, state, local, international or any other law or regulation; -3. To harm Yourself or others; -4. To repurpose or distribute output from Tencent Hunyuan 3D 2.1 or any Model Derivatives to harm Yourself or others; -5. To override or circumvent the safety guardrails and safeguards We have put in place; -6. For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; -7. To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections; -8. To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement; -9. To intentionally defame, disparage or otherwise harass others; -10. To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems; -11. To generate or disseminate personal identifiable information with the purpose of harming others; -12. To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated; -13. To impersonate another individual without consent, authorization, or legal right; -14. To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance); -15. In a manner that violates or disrespects the social ethics and moral standards of other countries or regions; -16. To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism; -17. For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics; -18. To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; -19. For military purposes; -20. To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices. diff --git a/hy3dshape/NOTICE b/hy3dshape/NOTICE deleted file mode 100644 index e02234d9704824b109ffe3625aa9b8b94dad2d55..0000000000000000000000000000000000000000 --- a/hy3dshape/NOTICE +++ /dev/null @@ -1,214 +0,0 @@ -Usage and Legal Notices: - -Tencent is pleased to support the open source community by making Hunyuan 3D 2.0 available. - -Copyright (C) 2025 THL A29 Limited, a Tencent company. All rights reserved. The below software and/or models in this distribution may have been modified by THL A29 Limited ("Tencent Modifications"). All Tencent Modifications are Copyright (C) THL A29 Limited. - -Hunyuan 3D 2.0 is licensed under the TENCENT HUNYUAN 3D 2.0 COMMUNITY LICENSE AGREEMENT except for the third-party components listed below, which is licensed under different terms. Hunyuan 3D 2.0 does not impose any additional limitations beyond what is outlined in the respective licenses of these third-party components. Users must comply with all terms and conditions of original licenses of these third-party components and must ensure that the usage of the third party components adheres to all relevant laws and regulations. - -For avoidance of doubts, Hunyuan 3D 2.0 means inference-enabling code, parameters, and weights of this Model only, which are made publicly available by Tencent in accordance with TENCENT HUNYUAN 3D 2.0 COMMUNITY LICENSE AGREEMENT. - - -Other dependencies and licenses: - - -Open Source Model Licensed under the MIT and CreativeML Open RAIL++-M License: --------------------------------------------------------------------- -1. Stable Diffusion -Copyright (c) 2022 Stability AI - - -Terms of the MIT and CreativeML Open RAIL++-M License: --------------------------------------------------------------------- -Permission is hereby granted, free of charge, to any person obtaining a copy -of this software and associated documentation files (the "Software"), to deal -in the Software without restriction, including without limitation the rights -to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -copies of the Software, and to permit persons to whom the Software is -furnished to do so, subject to the following conditions: - -The above copyright notice and this permission notice shall be included in all -copies or substantial portions of the Software. - -THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -SOFTWARE. - - -CreativeML Open RAIL++-M License -dated November 24, 2022 - -Section I: PREAMBLE - -Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation. - -Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations. - -In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation. - -Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI. - -This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model. - -NOW THEREFORE, You and Licensor agree as follows: - -1. Definitions - -- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document. -- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License. -- "Output" means the results of operating a Model as embodied in informational content resulting therefrom. -- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material. -- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model. -- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any. -- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access. -- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model. -- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator. -- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You. -- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." -- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model. - -Section II: INTELLECTUAL PROPERTY RIGHTS - -Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III. - -2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model. -3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed. - -Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION - -4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions: -Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material. -You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License; -You must cause any modified files to carry prominent notices stating that You changed the files; -You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model. -You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License. -5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5). -6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License. - -Section IV: OTHER PROVISIONS - -7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License. -8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors. -9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License. -10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. -11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. -12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein. - -END OF TERMS AND CONDITIONS - - - - -Attachment A - -Use Restrictions - -You agree not to use the Model or Derivatives of the Model: - -- In any way that violates any applicable national, federal, state, local or international law or regulation; -- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; -- To generate or disseminate verifiably false information and/or content with the purpose of harming others; -- To generate or disseminate personal identifiable information that can be used to harm an individual; -- To defame, disparage or otherwise harass others; -- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation; -- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics; -- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; -- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories; -- To provide medical advice and medical results interpretation; -- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use). - - - -Open Source Model Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT and Other Licenses of the Third-Party Components therein: --------------------------------------------------------------------- -1. HunyuanDiT -Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. - - -Terms of the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT: --------------------------------------------------------------------- -TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT -Tencent Hunyuan Release Date: 2024/5/14 -By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately. -1. DEFINITIONS. -a. “Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A. -b. “Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of the Hunyuan Works or any portion or element thereof set forth herein. -c. “Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan made publicly available by Tencent. -d. “Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means. -e. “Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan Works for any purpose and in any field of use. -f. “Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan and Documentation (and any portion thereof) as made available by Tencent under this Agreement. -g. “Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan or any Model Derivative of Tencent Hunyuan; (ii) works based on Tencent Hunyuan or any Model Derivative of Tencent Hunyuan; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan or any Model Derivative of Tencent Hunyuan, to that model in order to cause that model to perform similarly to Tencent Hunyuan or a Model Derivative of Tencent Hunyuan, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan or a Model Derivative of Tencent Hunyuan for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives. -h. “Output” shall mean the information and/or content output of Tencent Hunyuan or a Model Derivative that results from operating or otherwise using Tencent Hunyuan or a Model Derivative, including via a Hosted Service. -i. “Tencent,” “We” or “Us” shall mean THL A29 Limited. -j. “Tencent Hunyuan” shall mean the large language models, image/video/audio/3D generation models, and multimodal large language models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at https://huggingface.co/Tencent-Hunyuan/HunyuanDiT and https://github.com/Tencent/HunyuanDiT . -k. “Tencent Hunyuan Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof. -l. “Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You. -m. “including” shall mean including but not limited to. -2. GRANT OF RIGHTS. -We grant You a non-exclusive, worldwide, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy. -3. DISTRIBUTION. -You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan Works, provided that You meet all of the following conditions: -a. You must provide all such Third Party recipients of the Tencent Hunyuan Works or products or services using them a copy of this Agreement; -b. You must cause any modified files to carry prominent notices stating that You changed the files; -c. You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan Works; and (ii) mark the products or services developed by using the Tencent Hunyuan Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and -d. All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan is licensed under the Tencent Hunyuan Community License Agreement, Copyright © 2024 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.” -You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement. If You receive Tencent Hunyuan Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You. -4. ADDITIONAL COMMERCIAL TERMS. -If, on the Tencent Hunyuan version release date, the monthly active users of all products or services made available by or for Licensee is greater than 100 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights. -5. RULES OF USE. -a. Your use of the Tencent Hunyuan Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan Works are subject to the use restrictions in these Sections 5(a) and 5(b). -b. You must not use the Tencent Hunyuan Works or any Output or results of the Tencent Hunyuan Works to improve any other large language model (other than Tencent Hunyuan or Model Derivatives thereof). -6. INTELLECTUAL PROPERTY. -a. Subject to Tencent’s ownership of Tencent Hunyuan Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You. -b. No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent. -c. If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan Works. -d. Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses. -7. DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY. -a. We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan Works or to grant any license thereto. -b. UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT. -c. TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING. -8. SURVIVAL AND TERMINATION. -a. The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. -b. We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement. -9. GOVERNING LAW AND JURISDICTION. -a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. -b. Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute. -  - -EXHIBIT A -ACCEPTABLE USE POLICY - -Tencent reserves the right to update this Acceptable Use Policy from time to time. -Last modified: 2024/5/14 - -Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan. You agree not to use Tencent Hunyuan or Model Derivatives: -1. In any way that violates any applicable national, federal, state, local, international or any other law or regulation; -2. To harm Yourself or others; -3. To repurpose or distribute output from Tencent Hunyuan or any Model Derivatives to harm Yourself or others; -4. To override or circumvent the safety guardrails and safeguards We have put in place; -5. For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; -6. To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections; -7. To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement; -8. To intentionally defame, disparage or otherwise harass others; -9. To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems; -10. To generate or disseminate personal identifiable information with the purpose of harming others; -11. To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated; -12. To impersonate another individual without consent, authorization, or legal right; -13. To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance); -14. In a manner that violates or disrespects the social ethics and moral standards of other countries or regions; -15. To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism; -16. For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics; -17. To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; -18. For military purposes; -19. To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices. - -For the license of other third party components, please refer to the following URL: -https://huggingface.co/Tencent-Hunyuan/HunyuanDiT/blob/main/Notice - --------------------------------------------------------------------- - -This Model also incorporates insights from Flux's neural network architechtures (https://github.com/black-forest-labs/flux?tab=readme-ov-file). Credits are given to the orginal authors. \ No newline at end of file diff --git a/hy3dshape/README-zh.md b/hy3dshape/README-zh.md deleted file mode 100644 index b526a7e731fcb6fdbef14013d33f411e9af2d422..0000000000000000000000000000000000000000 --- a/hy3dshape/README-zh.md +++ /dev/null @@ -1,47 +0,0 @@ -# Hunyuan3D-2.1-Shape - - -# 训练 - -我们会展示小数据集上DiT的训练全流程 - -## 数据预处理 - -渲染和水密化参考[链接](tools/README.md)，最终得到如下结构 - -``` yaml -dataset/preprocessed/{uid} -├── geo_data -│ ├── {uid}_sdf.npz -│ ├── {uid}_surface.npz -│ └── {uid}_watertight.obj -└── render_cond - ├── 000.png - ├── ... - ├── 023.png - ├── mesh.ply - └── transforms.json -``` - -我们提供了一个8个case(均来自Objaverse-XL)预处理后的结果在 tools/mini_trainset，可以直接用于过拟合训练 - - - -## 启动训练 - -我们提供了可供参考的训练配置文件和启动脚本（默认单机8卡deepspeed训练），用户根据需要自行修改。 - -配置文件 -``` -configs/dit-from-scratch-overfitting-flowmatching-dinog518-bf16-lr1e4-1024.yaml -``` -启动脚本 - -``` -export node_num=1 -export node_rank=0 -export master_ip=0.0.0.0 # set your master_ip -export config='configs/dit-from-scratch-overfitting-flowmatching-dinog518-bf16-lr1e4-1024.yaml' -export output_dir='output_folder/dit/overfitting' -bash scripts/train_deepspeed.sh $node_num $node_rank $master_ip $config $output_dir -``` \ No newline at end of file diff --git a/hy3dshape/README.md b/hy3dshape/README.md deleted file mode 100644 index d7b864ed6f80c235423c5f515e153762ec8e0f08..0000000000000000000000000000000000000000 --- a/hy3dshape/README.md +++ /dev/null @@ -1,54 +0,0 @@ -# Hunyuan3D-2.1-Shape - -## Quick Inference - -Given a reference image `image.png`, you can run inference using the following code. The result will be saved as `demo.glb`. - -```bash -python3 minimal_demo.py -``` - -**Memory Recommendation:** For we recommend using a GPU with at least **10GB VRAM**. - -# Training - -Here we demonstrate the complete training workflow of DiT on a small dataset. - -## Data Preprocessing - -The rendering and watertight mesh generation process is described in detail in [this document](tools/README.md). After preprocessing, the dataset directory structure should look like the following: - -```yaml -dataset/preprocessed/{uid} -├── geo_data -│ ├── {uid}_sdf.npz -│ ├── {uid}_surface.npz -│ └── {uid}_watertight.obj -└── render_cond - ├── 000.png - ├── ... - ├── 023.png - ├── mesh.ply - └── transforms.json -``` - -We provide a preprocessed mini_dataset containing 8 cases (all sourced from Objaverse-XL) as `tools/mini_trainset`, which can be used directly for DiT overfitting training experiments. - -## Launching Training - -We provide example configuration files and launch scripts for reference. By default, the training runs on a single node with 8 GPUs using DeepSpeed. Users can modify the configurations and scripts as needed to suit their environment. - -Configuration File -``` -configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml -``` -Launch Script - -``` -export node_num=1 -export node_rank=0 -export master_ip=0.0.0.0 # set your master_ip -export config=configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml -export output_dir=output_folder/dit/overfitting -bash scripts/train_deepspeed.sh $node_num $node_rank $master_ip $config $output_dir -``` \ No newline at end of file diff --git a/hy3dshape/configs/hunyuan3ddit-full-params-finetuning-flowmatching-dinog518-bf16-lr1e5-512.yaml b/hy3dshape/configs/hunyuan3ddit-full-params-finetuning-flowmatching-dinog518-bf16-lr1e5-512.yaml deleted file mode 100644 index 3128e4447c0bd307be99bb1e11e9405bb8d43524..0000000000000000000000000000000000000000 --- a/hy3dshape/configs/hunyuan3ddit-full-params-finetuning-flowmatching-dinog518-bf16-lr1e5-512.yaml +++ /dev/null @@ -1,174 +0,0 @@ -name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518" - -training: - steps: 10_0000_0000 - use_amp: true - amp_type: "bf16" - base_lr: 1.e-5 - gradient_clip_val: 1.0 - gradient_clip_algorithm: "norm" - every_n_train_steps: 2000 # 5000 - val_check_interval: 50 # 4096 - limit_val_batches: 16 - -dataset: - target: hy3dshape.data.dit_asl.AlignedShapeLatentModule - params: - #! Base setting - batch_size: 4 - num_workers: 8 - val_num_workers: 4 - - # Data - train_data_list: tools/mini_trainset/preprocessed - val_data_list: tools/mini_trainset/preprocessed - - #! Image loading - cond_stage_key: "image" # image / text / image_text - image_size: 518 - mean: &mean [0.5, 0.5, 0.5] - std: &std [0.5, 0.5, 0.5] - - #! Point cloud sampling - pc_size: &pc_size 30720 - pc_sharpedge_size: &pc_sharpedge_size 30720 - sharpedge_label: &sharpedge_label true - return_normal: true - - #! Augmentation - padding: true - -model: - target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser - params: - first_stage_key: "surface" - cond_stage_key: "image" - scale_by_std: false - z_scale_factor: &z_scale_factor 0.9990943042622529 # 1 / 1.0009065167661184 - torch_compile: false - - # ema_config: - # ema_model: LitEma - # ema_decay: 0.999 - # ema_inference: false - - first_stage_config: - target: hy3dshape.models.autoencoders.ShapeVAE - from_pretrained: tencent/Hunyuan3D-2.1 - params: - num_latents: &num_latents 512 - embed_dim: 64 - num_freqs: 8 - include_pi: false - heads: 16 - width: 1024 - point_feats: 4 - num_decoder_layers: 16 - pc_size: *pc_size - pc_sharpedge_size: *pc_sharpedge_size - qkv_bias: false - qk_norm: true - scale_factor: *z_scale_factor - geo_decoder_mlp_expand_ratio: 4 - geo_decoder_downsample_ratio: 1 - geo_decoder_ln_post: true - - cond_stage_config: - target: hy3dshape.models.conditioner.SingleImageEncoder - params: - main_image_encoder: - type: DinoImageEncoder # dino giant - kwargs: - config: - attention_probs_dropout_prob: 0.0 - drop_path_rate: 0.0 - hidden_act: gelu - hidden_dropout_prob: 0.0 - hidden_size: 1536 - image_size: 518 - initializer_range: 0.02 - layer_norm_eps: 1.e-6 - layerscale_value: 1.0 - mlp_ratio: 4 - model_type: dinov2 - num_attention_heads: 24 - num_channels: 3 - num_hidden_layers: 40 - patch_size: 14 - qkv_bias: true - torch_dtype: float32 - use_swiglu_ffn: true - image_size: 518 - - denoiser_cfg: - target: hy3dshape.models.denoisers.hunyuan3ddit.Hunyuan3DDiT - params: - ckpt_path: ~/.cache/hy3dgen/tencent/Hunyuan3D-2-1-Shape/dit/model.fp16.ckpt - input_size: *num_latents - context_in_dim: 1536 - hidden_size: 1024 - mlp_ratio: 4.0 - num_heads: 16 - depth: 16 - depth_single_blocks: 32 - axes_dim: [64] - theta: 10000 - qkv_bias: true - use_pe: false - force_norm_fp32: true - - scheduler_cfg: - transport: - target: hy3dshape.models.diffusion.transport.create_transport - params: - path_type: Linear - prediction: velocity - sampler: - target: hy3dshape.models.diffusion.transport.Sampler - params: {} - ode_params: - sampling_method: euler # dopri5 ... - num_steps: &num_steps 50 - - optimizer_cfg: - optimizer: - target: torch.optim.AdamW - params: - betas: [0.9, 0.99] - eps: 1.e-6 - weight_decay: 1.e-2 - - scheduler: - target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler - params: - warm_up_steps: 50 # 5000 - f_start: 1.e-6 - f_min: 1.e-3 - f_max: 1.0 - - pipeline_cfg: - target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline - - image_processor_cfg: - target: hy3dshape.preprocessors.ImageProcessorV2 - params: {} - -callbacks: - logger: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger - params: - step_frequency: 100 # 10000 - num_samples: 1 - sample_times: 1 - mean: *mean - std: *std - bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01] - octree_depth: 8 - num_chunks: 50000 - mc_level: 0.0 - - file_loggers: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger - params: - step_frequency: 50 # 5000 - test_data_path: "tools/mini_testset/images.json" diff --git a/hy3dshape/configs/hunyuan3ddit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml b/hy3dshape/configs/hunyuan3ddit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml deleted file mode 100644 index 33dcbd0f0c71390b870e40f0e89d0bb7ddaefe70..0000000000000000000000000000000000000000 --- a/hy3dshape/configs/hunyuan3ddit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml +++ /dev/null @@ -1,173 +0,0 @@ -name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518" - -training: - steps: 10_0000_0000 - use_amp: true - amp_type: "bf16" - base_lr: 1e-4 - gradient_clip_val: 1.0 - gradient_clip_algorithm: "norm" - every_n_train_steps: 2000 # 5000 - val_check_interval: 50 # 4096 - limit_val_batches: 16 - -dataset: - target: hy3dshape.data.dit_asl.AlignedShapeLatentModule - params: - #! Base setting - batch_size: 2 - num_workers: 8 - val_num_workers: 4 - - # Data - train_data_list: tools/mini_trainset/preprocessed - val_data_list: tools/mini_trainset/preprocessed - - #! Image loading - cond_stage_key: "image" # image / text / image_text - image_size: 518 - mean: &mean [0.5, 0.5, 0.5] - std: &std [0.5, 0.5, 0.5] - - #! Point cloud sampling - pc_size: &pc_size 10240 - pc_sharpedge_size: &pc_sharpedge_size 10240 - sharpedge_label: &sharpedge_label true - return_normal: true - - #! Augmentation - padding: true - -model: - target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser - params: - first_stage_key: "surface" - cond_stage_key: "image" - scale_by_std: false - z_scale_factor: &z_scale_factor 0.9990943042622529 # 1 / 1.0009065167661184 - torch_compile: false - - # ema_config: - # ema_model: LitEma - # ema_decay: 0.999 - # ema_inference: false - - first_stage_config: - target: hy3dshape.models.autoencoders.ShapeVAE - from_pretrained: tencent/Hunyuan3D-2.1 - params: - num_latents: &num_latents 512 - embed_dim: 64 - num_freqs: 8 - include_pi: false - heads: 16 - width: 1024 - point_feats: 4 - num_decoder_layers: 16 - pc_size: *pc_size - pc_sharpedge_size: *pc_sharpedge_size - qkv_bias: false - qk_norm: true - scale_factor: *z_scale_factor - geo_decoder_mlp_expand_ratio: 4 - geo_decoder_downsample_ratio: 1 - geo_decoder_ln_post: true - - cond_stage_config: - target: hy3dshape.models.conditioner.SingleImageEncoder - params: - main_image_encoder: - type: DinoImageEncoder # dino giant - kwargs: - config: - attention_probs_dropout_prob: 0.0 - drop_path_rate: 0.0 - hidden_act: gelu - hidden_dropout_prob: 0.0 - hidden_size: 1536 - image_size: 518 - initializer_range: 0.02 - layer_norm_eps: 1.e-6 - layerscale_value: 1.0 - mlp_ratio: 4 - model_type: dinov2 - num_attention_heads: 24 - num_channels: 3 - num_hidden_layers: 40 - patch_size: 14 - qkv_bias: true - torch_dtype: float32 - use_swiglu_ffn: true - image_size: 518 - - denoiser_cfg: - target: hy3dshape.models.denoisers.hunyuan3ddit.Hunyuan3DDiT - params: - input_size: *num_latents - context_in_dim: 1536 - hidden_size: 1024 - mlp_ratio: 4.0 - num_heads: 16 - depth: 8 - depth_single_blocks: 16 - axes_dim: [64] - theta: 10000 - qkv_bias: true - use_pe: false - force_norm_fp32: true - - scheduler_cfg: - transport: - target: hy3dshape.models.diffusion.transport.create_transport - params: - path_type: Linear - prediction: velocity - sampler: - target: hy3dshape.models.diffusion.transport.Sampler - params: {} - ode_params: - sampling_method: euler # dopri5 ... - num_steps: &num_steps 50 - - optimizer_cfg: - optimizer: - target: torch.optim.AdamW - params: - betas: [0.9, 0.99] - eps: 1.e-6 - weight_decay: 1.e-2 - - scheduler: - target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler - params: - warm_up_steps: 50 # 5000 - f_start: 1.e-6 - f_min: 1.e-3 - f_max: 1.0 - - pipeline_cfg: - target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline - - image_processor_cfg: - target: hy3dshape.preprocessors.ImageProcessorV2 - params: {} - -callbacks: - logger: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger - params: - step_frequency: 100 # 10000 - num_samples: 1 - sample_times: 1 - mean: *mean - std: *std - bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01] - octree_depth: 8 - num_chunks: 50000 - mc_level: 0.0 - - file_loggers: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger - params: - step_frequency: 50 # 5000 - test_data_path: "tools/mini_testset/images.json" diff --git a/hy3dshape/configs/hunyuandit-finetuning-flowmatching-dinog518-bf16-lr1e5-4096.yaml b/hy3dshape/configs/hunyuandit-finetuning-flowmatching-dinog518-bf16-lr1e5-4096.yaml deleted file mode 100644 index e9fc84146294cc30d7bb546e8cbc7a36f342c0c5..0000000000000000000000000000000000000000 --- a/hy3dshape/configs/hunyuandit-finetuning-flowmatching-dinog518-bf16-lr1e5-4096.yaml +++ /dev/null @@ -1,180 +0,0 @@ -name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518" - -training: - steps: 10_0000_0000 - use_amp: true - amp_type: "bf16" - base_lr: 1e-5 - gradient_clip_val: 1.0 - gradient_clip_algorithm: "norm" - every_n_train_steps: 2000 # 5000 - val_check_interval: 50 # 4096 - limit_val_batches: 16 - -dataset: - target: hy3dshape.data.dit_asl.AlignedShapeLatentModule - params: - #! Base setting - batch_size: 4 - num_workers: 8 - val_num_workers: 4 - - # Data - train_data_list: tools/mini_trainset/preprocessed - val_data_list: tools/mini_trainset/preprocessed - - #! Image loading - cond_stage_key: "image" # image / text / image_text - image_size: 518 - mean: &mean [0.5, 0.5, 0.5] - std: &std [0.5, 0.5, 0.5] - - #! Point cloud sampling - pc_size: &pc_size 81920 - pc_sharpedge_size: &pc_sharpedge_size 0 - sharpedge_label: &sharpedge_label true - return_normal: true - - #! Augmentation - padding: true - -model: - target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser - params: - first_stage_key: "surface" - cond_stage_key: "image" - scale_by_std: false - z_scale_factor: &z_scale_factor 1.0039506158752403 - torch_compile: false - - # ema_config: - # ema_model: LitEma - # ema_decay: 0.999 - # ema_inference: false - - first_stage_config: - target: hy3dshape.models.autoencoders.ShapeVAE - from_pretrained: tencent/Hunyuan3D-2.1 - params: - num_latents: &num_latents 4096 - embed_dim: 64 - num_freqs: 8 - include_pi: false - heads: 16 - width: 1024 - num_encoder_layers: 8 - num_decoder_layers: 16 - qkv_bias: false - qk_norm: true - scale_factor: *z_scale_factor - geo_decoder_mlp_expand_ratio: 4 - geo_decoder_downsample_ratio: 1 - geo_decoder_ln_post: true - point_feats: 4 - pc_size: *pc_size - pc_sharpedge_size: *pc_sharpedge_size - - cond_stage_config: - target: hy3dshape.models.conditioner.SingleImageEncoder - params: - main_image_encoder: - type: DinoImageEncoder # dino large - kwargs: - config: - attention_probs_dropout_prob: 0.0 - drop_path_rate: 0.0 - hidden_act: gelu - hidden_dropout_prob: 0.0 - hidden_size: 1024 - image_size: 518 - initializer_range: 0.02 - layer_norm_eps: 1.e-6 - layerscale_value: 1.0 - mlp_ratio: 4 - model_type: dinov2 - num_attention_heads: 16 - num_channels: 3 - num_hidden_layers: 24 - patch_size: 14 - qkv_bias: true - torch_dtype: float32 - use_swiglu_ffn: false - image_size: 518 - use_cls_token: true - - - denoiser_cfg: - target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain - params: - input_size: *num_latents - in_channels: 64 - hidden_size: 2048 - context_dim: 1024 - depth: 21 - num_heads: 16 - qk_norm: true - text_len: 1370 - with_decoupled_ca: false - use_attention_pooling: false - qk_norm_type: 'rms' - qkv_bias: false - use_pos_emb: false - num_moe_layers: 6 - num_experts: 8 - moe_top_k: 2 - - scheduler_cfg: - transport: - target: hy3dshape.models.diffusion.transport.create_transport - params: - path_type: Linear - prediction: velocity - sampler: - target: hy3dshape.models.diffusion.transport.Sampler - params: {} - ode_params: - sampling_method: euler # dopri5 ... - num_steps: &num_steps 50 - - optimizer_cfg: - optimizer: - target: torch.optim.AdamW - params: - betas: [0.9, 0.99] - eps: 1.e-6 - weight_decay: 1.e-2 - - scheduler: - target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler - params: - warm_up_steps: 50 # 5000 - f_start: 1.e-6 - f_min: 1.e-3 - f_max: 1.0 - - pipeline_cfg: - target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline - - image_processor_cfg: - target: hy3dshape.preprocessors.ImageProcessorV2 - params: {} - -callbacks: - logger: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger - params: - step_frequency: 100 # 10000 - num_samples: 1 - sample_times: 1 - mean: *mean - std: *std - bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01] - octree_depth: 8 - num_chunks: 50000 - mc_level: 0.0 - - file_loggers: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger - params: - step_frequency: 50 # 5000 - test_data_path: "tools/mini_testset/images.json" diff --git a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-4096.yaml b/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-4096.yaml deleted file mode 100644 index 82e7ebb4b7475c6ea19c52bf63715b0fe40ecd20..0000000000000000000000000000000000000000 --- a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-4096.yaml +++ /dev/null @@ -1,180 +0,0 @@ -name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518" - -training: - steps: 10_0000_0000 - use_amp: true - amp_type: "bf16" - base_lr: 1e-4 - gradient_clip_val: 1.0 - gradient_clip_algorithm: "norm" - every_n_train_steps: 2000 # 5000 - val_check_interval: 50 # 4096 - limit_val_batches: 16 - -dataset: - target: hy3dshape.data.dit_asl.AlignedShapeLatentModule - params: - #! Base setting - batch_size: 2 - num_workers: 8 - val_num_workers: 4 - - # Data - train_data_list: tools/mini_trainset/preprocessed - val_data_list: tools/mini_trainset/preprocessed - - #! Image loading - cond_stage_key: "image" # image / text / image_text - image_size: 518 - mean: &mean [0.5, 0.5, 0.5] - std: &std [0.5, 0.5, 0.5] - - #! Point cloud sampling - pc_size: &pc_size 81920 - pc_sharpedge_size: &pc_sharpedge_size 0 - sharpedge_label: &sharpedge_label true - return_normal: true - - #! Augmentation - padding: true - -model: - target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser - params: - first_stage_key: "surface" - cond_stage_key: "image" - scale_by_std: false - z_scale_factor: &z_scale_factor 1.0039506158752403 - torch_compile: false - - # ema_config: - # ema_model: LitEma - # ema_decay: 0.999 - # ema_inference: false - - first_stage_config: - target: hy3dshape.models.autoencoders.ShapeVAE - from_pretrained: tencent/Hunyuan3D-2.1 - params: - num_latents: &num_latents 4096 - embed_dim: 64 - num_freqs: 8 - include_pi: false - heads: 16 - width: 1024 - num_encoder_layers: 8 - num_decoder_layers: 16 - qkv_bias: false - qk_norm: true - scale_factor: *z_scale_factor - geo_decoder_mlp_expand_ratio: 4 - geo_decoder_downsample_ratio: 1 - geo_decoder_ln_post: true - point_feats: 4 - pc_size: *pc_size - pc_sharpedge_size: *pc_sharpedge_size - - cond_stage_config: - target: hy3dshape.models.conditioner.SingleImageEncoder - params: - main_image_encoder: - type: DinoImageEncoder # dino large - kwargs: - config: - attention_probs_dropout_prob: 0.0 - drop_path_rate: 0.0 - hidden_act: gelu - hidden_dropout_prob: 0.0 - hidden_size: 1024 - image_size: 518 - initializer_range: 0.02 - layer_norm_eps: 1.e-6 - layerscale_value: 1.0 - mlp_ratio: 4 - model_type: dinov2 - num_attention_heads: 16 - num_channels: 3 - num_hidden_layers: 24 - patch_size: 14 - qkv_bias: true - torch_dtype: float32 - use_swiglu_ffn: false - image_size: 518 - use_cls_token: true - - - denoiser_cfg: - target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain - params: - input_size: *num_latents - in_channels: 64 - hidden_size: 2048 - context_dim: 1024 - depth: 11 - num_heads: 16 - qk_norm: true - text_len: 1370 - with_decoupled_ca: false - use_attention_pooling: false - qk_norm_type: 'rms' - qkv_bias: false - use_pos_emb: false - num_moe_layers: 6 - num_experts: 8 - moe_top_k: 2 - - scheduler_cfg: - transport: - target: hy3dshape.models.diffusion.transport.create_transport - params: - path_type: Linear - prediction: velocity - sampler: - target: hy3dshape.models.diffusion.transport.Sampler - params: {} - ode_params: - sampling_method: euler # dopri5 ... - num_steps: &num_steps 50 - - optimizer_cfg: - optimizer: - target: torch.optim.AdamW - params: - betas: [0.9, 0.99] - eps: 1.e-6 - weight_decay: 1.e-2 - - scheduler: - target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler - params: - warm_up_steps: 50 # 5000 - f_start: 1.e-6 - f_min: 1.e-3 - f_max: 1.0 - - pipeline_cfg: - target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline - - image_processor_cfg: - target: hy3dshape.preprocessors.ImageProcessorV2 - params: {} - -callbacks: - logger: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger - params: - step_frequency: 100 # 10000 - num_samples: 1 - sample_times: 1 - mean: *mean - std: *std - bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01] - octree_depth: 8 - num_chunks: 50000 - mc_level: 0.0 - - file_loggers: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger - params: - step_frequency: 50 # 5000 - test_data_path: "tools/mini_testset/images.json" diff --git a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml b/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml deleted file mode 100644 index ffed3aa9a9b721c23059e8f16f94d08e5b16cab4..0000000000000000000000000000000000000000 --- a/hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml +++ /dev/null @@ -1,180 +0,0 @@ -name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518" - -training: - steps: 10_0000_0000 - use_amp: true - amp_type: "bf16" - base_lr: 1e-4 - gradient_clip_val: 1.0 - gradient_clip_algorithm: "norm" - every_n_train_steps: 2000 # 5000 - val_check_interval: 50 # 4096 - limit_val_batches: 16 - -dataset: - target: hy3dshape.data.dit_asl.AlignedShapeLatentModule - params: - #! Base setting - batch_size: 2 - num_workers: 8 - val_num_workers: 4 - - # Data - train_data_list: tools/mini_trainset/preprocessed - val_data_list: tools/mini_trainset/preprocessed - - #! Image loading - cond_stage_key: "image" # image / text / image_text - image_size: 518 - mean: &mean [0.5, 0.5, 0.5] - std: &std [0.5, 0.5, 0.5] - - #! Point cloud sampling - pc_size: &pc_size 81920 - pc_sharpedge_size: &pc_sharpedge_size 0 - sharpedge_label: &sharpedge_label true - return_normal: true - - #! Augmentation - padding: true - -model: - target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser - params: - first_stage_key: "surface" - cond_stage_key: "image" - scale_by_std: false - z_scale_factor: &z_scale_factor 1.0039506158752403 - torch_compile: false - - # ema_config: - # ema_model: LitEma - # ema_decay: 0.999 - # ema_inference: false - - first_stage_config: - target: hy3dshape.models.autoencoders.ShapeVAE - from_pretrained: tencent/Hunyuan3D-2.1 - params: - num_latents: &num_latents 512 - embed_dim: 64 - num_freqs: 8 - include_pi: false - heads: 16 - width: 1024 - num_encoder_layers: 8 - num_decoder_layers: 16 - qkv_bias: false - qk_norm: true - scale_factor: *z_scale_factor - geo_decoder_mlp_expand_ratio: 4 - geo_decoder_downsample_ratio: 1 - geo_decoder_ln_post: true - point_feats: 4 - pc_size: *pc_size - pc_sharpedge_size: *pc_sharpedge_size - - cond_stage_config: - target: hy3dshape.models.conditioner.SingleImageEncoder - params: - main_image_encoder: - type: DinoImageEncoder # dino large - kwargs: - config: - attention_probs_dropout_prob: 0.0 - drop_path_rate: 0.0 - hidden_act: gelu - hidden_dropout_prob: 0.0 - hidden_size: 1024 - image_size: 518 - initializer_range: 0.02 - layer_norm_eps: 1.e-6 - layerscale_value: 1.0 - mlp_ratio: 4 - model_type: dinov2 - num_attention_heads: 16 - num_channels: 3 - num_hidden_layers: 24 - patch_size: 14 - qkv_bias: true - torch_dtype: float32 - use_swiglu_ffn: false - image_size: 518 - use_cls_token: true - - - denoiser_cfg: - target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain - params: - input_size: *num_latents - in_channels: 64 - hidden_size: 768 - context_dim: 1024 - depth: 6 - num_heads: 12 - qk_norm: true - text_len: 1370 - with_decoupled_ca: false - use_attention_pooling: false - qk_norm_type: 'rms' - qkv_bias: false - use_pos_emb: false - num_moe_layers: 3 - num_experts: 4 - moe_top_k: 2 - - scheduler_cfg: - transport: - target: hy3dshape.models.diffusion.transport.create_transport - params: - path_type: Linear - prediction: velocity - sampler: - target: hy3dshape.models.diffusion.transport.Sampler - params: {} - ode_params: - sampling_method: euler # dopri5 ... - num_steps: &num_steps 50 - - optimizer_cfg: - optimizer: - target: torch.optim.AdamW - params: - betas: [0.9, 0.99] - eps: 1.e-6 - weight_decay: 1.e-2 - - scheduler: - target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler - params: - warm_up_steps: 50 # 5000 - f_start: 1.e-6 - f_min: 1.e-3 - f_max: 1.0 - - pipeline_cfg: - target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline - - image_processor_cfg: - target: hy3dshape.preprocessors.ImageProcessorV2 - params: {} - -callbacks: - logger: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger - params: - step_frequency: 100 # 10000 - num_samples: 1 - sample_times: 1 - mean: *mean - std: *std - bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01] - octree_depth: 8 - num_chunks: 50000 - mc_level: 0.0 - - file_loggers: - target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger - params: - step_frequency: 50 # 5000 - test_data_path: "tools/mini_testset/images.json" diff --git a/hy3dshape/hy3dshape/__init__.py b/hy3dshape/hy3dshape/__init__.py deleted file mode 100644 index 1b1f9cce42c1d2bef94bbb2d7b088d24e76fe01e..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/__init__.py +++ /dev/null @@ -1,17 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from .pipelines import Hunyuan3DDiTPipeline, Hunyuan3DDiTFlowMatchingPipeline -from .postprocessors import FaceReducer, FloaterRemover, DegenerateFaceRemover, MeshSimplifier -from .preprocessors import ImageProcessorV2, IMAGE_PROCESSORS, DEFAULT_IMAGEPROCESSOR diff --git a/hy3dshape/hy3dshape/models/__init__.py b/hy3dshape/hy3dshape/models/__init__.py deleted file mode 100644 index 8179353ba7a5bdb8bcc30baa64e319fb8f884d57..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/__init__.py +++ /dev/null @@ -1,28 +0,0 @@ -# Open Source Model Licensed under the Apache License Version 2.0 -# and Other Licenses of the Third-Party Components therein: -# The below Model in this distribution may have been modified by THL A29 Limited -# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited. - -# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. -# The below software and/or models in this distribution may have been -# modified by THL A29 Limited ("Tencent Modifications"). -# All Tencent Modifications are Copyright (C) THL A29 Limited. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - - -from .autoencoders import ShapeVAE -from .conditioner import DualImageEncoder, SingleImageEncoder, DinoImageEncoder, CLIPImageEncoder -from .denoisers import Hunyuan3DDiT diff --git a/hy3dshape/hy3dshape/models/autoencoders/__init__.py b/hy3dshape/hy3dshape/models/autoencoders/__init__.py deleted file mode 100644 index 20bbf8d9559f8d5de7d7ae2d88bbb0348a197dc4..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/__init__.py +++ /dev/null @@ -1,20 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from .attention_blocks import CrossAttentionDecoder -from .attention_processors import FlashVDMCrossAttentionProcessor, CrossAttentionProcessor, \ - FlashVDMTopMCrossAttentionProcessor -from .model import ShapeVAE, VectsetVAE -from .surface_extractors import SurfaceExtractors, MCSurfaceExtractor, DMCSurfaceExtractor, Latent2MeshOutput -from .volume_decoders import HierarchicalVolumeDecoding, FlashVDMVolumeDecoding, VanillaVolumeDecoder diff --git a/hy3dshape/hy3dshape/models/autoencoders/attention_blocks.py b/hy3dshape/hy3dshape/models/autoencoders/attention_blocks.py deleted file mode 100644 index 918a695547a8733a30f039813a9a6a59bce2b388..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/attention_blocks.py +++ /dev/null @@ -1,716 +0,0 @@ -# Open Source Model Licensed under the Apache License Version 2.0 -# and Other Licenses of the Third-Party Components therein: -# The below Model in this distribution may have been modified by THL A29 Limited -# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited. - -# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. -# The below software and/or models in this distribution may have been -# modified by THL A29 Limited ("Tencent Modifications"). -# All Tencent Modifications are Copyright (C) THL A29 Limited. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - - -import os -from typing import Optional, Union, List - -import torch -import torch.nn as nn -from einops import rearrange -from torch import Tensor - -from .attention_processors import CrossAttentionProcessor -from ...utils import logger - -scaled_dot_product_attention = nn.functional.scaled_dot_product_attention - -if os.environ.get('USE_SAGEATTN', '0') == '1': - try: - from sageattention import sageattn - except ImportError: - raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.') - scaled_dot_product_attention = sageattn - - -class FourierEmbedder(nn.Module): - """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts - each feature dimension of `x[..., i]` into: - [ - sin(x[..., i]), - sin(f_1*x[..., i]), - sin(f_2*x[..., i]), - ... - sin(f_N * x[..., i]), - cos(x[..., i]), - cos(f_1*x[..., i]), - cos(f_2*x[..., i]), - ... - cos(f_N * x[..., i]), - x[..., i] # only present if include_input is True. - ], here f_i is the frequency. - - Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs]. - If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...]; - Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)]. - - Args: - num_freqs (int): the number of frequencies, default is 6; - logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...], - otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)]; - input_dim (int): the input dimension, default is 3; - include_input (bool): include the input tensor or not, default is True. - - Attributes: - frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...], - otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1); - - out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1), - otherwise, it is input_dim * num_freqs * 2. - - """ - - def __init__(self, - num_freqs: int = 6, - logspace: bool = True, - input_dim: int = 3, - include_input: bool = True, - include_pi: bool = True) -> None: - - """The initialization""" - - super().__init__() - - if logspace: - frequencies = 2.0 ** torch.arange( - num_freqs, - dtype=torch.float32 - ) - else: - frequencies = torch.linspace( - 1.0, - 2.0 ** (num_freqs - 1), - num_freqs, - dtype=torch.float32 - ) - - if include_pi: - frequencies *= torch.pi - - self.register_buffer("frequencies", frequencies, persistent=False) - self.include_input = include_input - self.num_freqs = num_freqs - - self.out_dim = self.get_dims(input_dim) - - def get_dims(self, input_dim): - temp = 1 if self.include_input or self.num_freqs == 0 else 0 - out_dim = input_dim * (self.num_freqs * 2 + temp) - - return out_dim - - def forward(self, x: torch.Tensor) -> torch.Tensor: - """ Forward process. - - Args: - x: tensor of shape [..., dim] - - Returns: - embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)] - where temp is 1 if include_input is True and 0 otherwise. - """ - - if self.num_freqs > 0: - embed = (x[..., None].contiguous() * self.frequencies).view(*x.shape[:-1], -1) - if self.include_input: - return torch.cat((x, embed.sin(), embed.cos()), dim=-1) - else: - return torch.cat((embed.sin(), embed.cos()), dim=-1) - else: - return x - - -class DropPath(nn.Module): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). - """ - - def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True): - super(DropPath, self).__init__() - self.drop_prob = drop_prob - self.scale_by_keep = scale_by_keep - - def forward(self, x): - """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). - - This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, - the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... - See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for - changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use - 'survival rate' as the argument. - - """ - if self.drop_prob == 0. or not self.training: - return x - keep_prob = 1 - self.drop_prob - shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets - random_tensor = x.new_empty(shape).bernoulli_(keep_prob) - if keep_prob > 0.0 and self.scale_by_keep: - random_tensor.div_(keep_prob) - return x * random_tensor - - def extra_repr(self): - return f'drop_prob={round(self.drop_prob, 3):0.3f}' - - -class MLP(nn.Module): - def __init__( - self, *, - width: int, - expand_ratio: int = 4, - output_width: int = None, - drop_path_rate: float = 0.0 - ): - super().__init__() - self.width = width - self.c_fc = nn.Linear(width, width * expand_ratio) - self.c_proj = nn.Linear(width * expand_ratio, output_width if output_width is not None else width) - self.gelu = nn.GELU() - self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity() - - def forward(self, x): - return self.drop_path(self.c_proj(self.gelu(self.c_fc(x)))) - - -class QKVMultiheadCrossAttention(nn.Module): - def __init__( - self, - *, - heads: int, - n_data: Optional[int] = None, - width=None, - qk_norm=False, - norm_layer=nn.LayerNorm - ): - super().__init__() - self.heads = heads - self.n_data = n_data - self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - - self.attn_processor = CrossAttentionProcessor() - - def forward(self, q, kv): - _, n_ctx, _ = q.shape - bs, n_data, width = kv.shape - attn_ch = width // self.heads // 2 - q = q.view(bs, n_ctx, self.heads, -1) - kv = kv.view(bs, n_data, self.heads, -1) - k, v = torch.split(kv, attn_ch, dim=-1) - - q = self.q_norm(q) - k = self.k_norm(k) - q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v)) - out = self.attn_processor(self, q, k, v) - out = out.transpose(1, 2).reshape(bs, n_ctx, -1) - return out - - -class MultiheadCrossAttention(nn.Module): - def __init__( - self, - *, - width: int, - heads: int, - qkv_bias: bool = True, - n_data: Optional[int] = None, - data_width: Optional[int] = None, - norm_layer=nn.LayerNorm, - qk_norm: bool = False, - kv_cache: bool = False, - ): - super().__init__() - self.n_data = n_data - self.width = width - self.heads = heads - self.data_width = width if data_width is None else data_width - self.c_q = nn.Linear(width, width, bias=qkv_bias) - self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias) - self.c_proj = nn.Linear(width, width) - self.attention = QKVMultiheadCrossAttention( - heads=heads, - n_data=n_data, - width=width, - norm_layer=norm_layer, - qk_norm=qk_norm - ) - self.kv_cache = kv_cache - self.data = None - - def forward(self, x, data): - x = self.c_q(x) - if self.kv_cache: - if self.data is None: - self.data = self.c_kv(data) - logger.info('Save kv cache,this should be called only once for one mesh') - data = self.data - else: - data = self.c_kv(data) - x = self.attention(x, data) - x = self.c_proj(x) - return x - - -class ResidualCrossAttentionBlock(nn.Module): - def __init__( - self, - *, - n_data: Optional[int] = None, - width: int, - heads: int, - mlp_expand_ratio: int = 4, - data_width: Optional[int] = None, - qkv_bias: bool = True, - norm_layer=nn.LayerNorm, - qk_norm: bool = False - ): - super().__init__() - - if data_width is None: - data_width = width - - self.attn = MultiheadCrossAttention( - n_data=n_data, - width=width, - heads=heads, - data_width=data_width, - qkv_bias=qkv_bias, - norm_layer=norm_layer, - qk_norm=qk_norm - ) - self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6) - self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6) - self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6) - self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio) - - def forward(self, x: torch.Tensor, data: torch.Tensor): - x = x + self.attn(self.ln_1(x), self.ln_2(data)) - x = x + self.mlp(self.ln_3(x)) - return x - - -class QKVMultiheadAttention(nn.Module): - def __init__( - self, - *, - heads: int, - n_ctx: int, - width=None, - qk_norm=False, - norm_layer=nn.LayerNorm - ): - super().__init__() - self.heads = heads - self.n_ctx = n_ctx - self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - - def forward(self, qkv): - bs, n_ctx, width = qkv.shape - attn_ch = width // self.heads // 3 - qkv = qkv.view(bs, n_ctx, self.heads, -1) - q, k, v = torch.split(qkv, attn_ch, dim=-1) - - q = self.q_norm(q) - k = self.k_norm(k) - - q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v)) - out = scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1) - return out - - -class MultiheadAttention(nn.Module): - def __init__( - self, - *, - n_ctx: int, - width: int, - heads: int, - qkv_bias: bool, - norm_layer=nn.LayerNorm, - qk_norm: bool = False, - drop_path_rate: float = 0.0 - ): - super().__init__() - self.n_ctx = n_ctx - self.width = width - self.heads = heads - self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias) - self.c_proj = nn.Linear(width, width) - self.attention = QKVMultiheadAttention( - heads=heads, - n_ctx=n_ctx, - width=width, - norm_layer=norm_layer, - qk_norm=qk_norm - ) - self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity() - - def forward(self, x): - x = self.c_qkv(x) - x = self.attention(x) - x = self.drop_path(self.c_proj(x)) - return x - - -class ResidualAttentionBlock(nn.Module): - def __init__( - self, - *, - n_ctx: int, - width: int, - heads: int, - qkv_bias: bool = True, - norm_layer=nn.LayerNorm, - qk_norm: bool = False, - drop_path_rate: float = 0.0, - ): - super().__init__() - self.attn = MultiheadAttention( - n_ctx=n_ctx, - width=width, - heads=heads, - qkv_bias=qkv_bias, - norm_layer=norm_layer, - qk_norm=qk_norm, - drop_path_rate=drop_path_rate - ) - self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6) - self.mlp = MLP(width=width, drop_path_rate=drop_path_rate) - self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6) - - def forward(self, x: torch.Tensor): - x = x + self.attn(self.ln_1(x)) - x = x + self.mlp(self.ln_2(x)) - return x - - -class Transformer(nn.Module): - def __init__( - self, - *, - n_ctx: int, - width: int, - layers: int, - heads: int, - qkv_bias: bool = True, - norm_layer=nn.LayerNorm, - qk_norm: bool = False, - drop_path_rate: float = 0.0 - ): - super().__init__() - self.n_ctx = n_ctx - self.width = width - self.layers = layers - self.resblocks = nn.ModuleList( - [ - ResidualAttentionBlock( - n_ctx=n_ctx, - width=width, - heads=heads, - qkv_bias=qkv_bias, - norm_layer=norm_layer, - qk_norm=qk_norm, - drop_path_rate=drop_path_rate - ) - for _ in range(layers) - ] - ) - - def forward(self, x: torch.Tensor): - for block in self.resblocks: - x = block(x) - return x - - -class CrossAttentionDecoder(nn.Module): - - def __init__( - self, - *, - num_latents: int, - out_channels: int, - fourier_embedder: FourierEmbedder, - width: int, - heads: int, - mlp_expand_ratio: int = 4, - downsample_ratio: int = 1, - enable_ln_post: bool = True, - qkv_bias: bool = True, - qk_norm: bool = False, - label_type: str = "binary" - ): - super().__init__() - - self.enable_ln_post = enable_ln_post - self.fourier_embedder = fourier_embedder - self.downsample_ratio = downsample_ratio - self.query_proj = nn.Linear(self.fourier_embedder.out_dim, width) - if self.downsample_ratio != 1: - self.latents_proj = nn.Linear(width * downsample_ratio, width) - if self.enable_ln_post == False: - qk_norm = False - self.cross_attn_decoder = ResidualCrossAttentionBlock( - n_data=num_latents, - width=width, - mlp_expand_ratio=mlp_expand_ratio, - heads=heads, - qkv_bias=qkv_bias, - qk_norm=qk_norm - ) - - if self.enable_ln_post: - self.ln_post = nn.LayerNorm(width) - self.output_proj = nn.Linear(width, out_channels) - self.label_type = label_type - self.count = 0 - - def set_cross_attention_processor(self, processor): - self.cross_attn_decoder.attn.attention.attn_processor = processor - - def set_default_cross_attention_processor(self): - self.cross_attn_decoder.attn.attention.attn_processor = CrossAttentionProcessor - - def forward(self, queries=None, query_embeddings=None, latents=None): - if query_embeddings is None: - query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype)) - self.count += query_embeddings.shape[1] - if self.downsample_ratio != 1: - latents = self.latents_proj(latents) - x = self.cross_attn_decoder(query_embeddings, latents) - if self.enable_ln_post: - x = self.ln_post(x) - occ = self.output_proj(x) - return occ - - -def fps( - src: torch.Tensor, - batch: Optional[Tensor] = None, - ratio: Optional[Union[Tensor, float]] = None, - random_start: bool = True, - batch_size: Optional[int] = None, - ptr: Optional[Union[Tensor, List[int]]] = None, -): - src = src.float() - from torch_cluster import fps as fps_fn - output = fps_fn(src, batch, ratio, random_start, batch_size, ptr) - return output - - -class PointCrossAttentionEncoder(nn.Module): - - def __init__( - self, *, - num_latents: int, - downsample_ratio: float, - pc_size: int, - pc_sharpedge_size: int, - fourier_embedder: FourierEmbedder, - point_feats: int, - width: int, - heads: int, - layers: int, - normal_pe: bool = False, - qkv_bias: bool = True, - use_ln_post: bool = False, - use_checkpoint: bool = False, - qk_norm: bool = False - ): - - super().__init__() - - self.use_checkpoint = use_checkpoint - self.num_latents = num_latents - self.downsample_ratio = downsample_ratio - self.point_feats = point_feats - self.normal_pe = normal_pe - - if pc_sharpedge_size == 0: - print( - f'PointCrossAttentionEncoder INFO: pc_sharpedge_size is not given, using pc_size as pc_sharpedge_size') - else: - print( - f'PointCrossAttentionEncoder INFO: pc_sharpedge_size is given, using pc_size={pc_size}, pc_sharpedge_size={pc_sharpedge_size}') - - self.pc_size = pc_size - self.pc_sharpedge_size = pc_sharpedge_size - - self.fourier_embedder = fourier_embedder - - self.input_proj = nn.Linear(self.fourier_embedder.out_dim + point_feats, width) - self.cross_attn = ResidualCrossAttentionBlock( - width=width, - heads=heads, - qkv_bias=qkv_bias, - qk_norm=qk_norm - ) - - self.self_attn = None - if layers > 0: - self.self_attn = Transformer( - n_ctx=num_latents, - width=width, - layers=layers, - heads=heads, - qkv_bias=qkv_bias, - qk_norm=qk_norm - ) - - if use_ln_post: - self.ln_post = nn.LayerNorm(width) - else: - self.ln_post = None - - def sample_points_and_latents(self, pc: torch.FloatTensor, feats: Optional[torch.FloatTensor] = None): - B, N, D = pc.shape - num_pts = self.num_latents * self.downsample_ratio - - # Compute number of latents - num_latents = int(num_pts / self.downsample_ratio) - - # Compute the number of random and sharpedge latents - num_random_query = self.pc_size / (self.pc_size + self.pc_sharpedge_size) * num_latents - num_sharpedge_query = num_latents - num_random_query - - # Split random and sharpedge surface points - random_pc, sharpedge_pc = torch.split(pc, [self.pc_size, self.pc_sharpedge_size], dim=1) - assert random_pc.shape[1] <= self.pc_size, "Random surface points size must be less than or equal to pc_size" - assert sharpedge_pc.shape[ - 1] <= self.pc_sharpedge_size, "Sharpedge surface points size must be less than or equal to pc_sharpedge_size" - - # Randomly select random surface points and random query points - input_random_pc_size = int(num_random_query * self.downsample_ratio) - random_query_ratio = num_random_query / input_random_pc_size - idx_random_pc = torch.randperm(random_pc.shape[1], device=random_pc.device)[:input_random_pc_size] - input_random_pc = random_pc[:, idx_random_pc, :] - flatten_input_random_pc = input_random_pc.view(B * input_random_pc_size, D) - N_down = int(flatten_input_random_pc.shape[0] / B) - batch_down = torch.arange(B).to(pc.device) - batch_down = torch.repeat_interleave(batch_down, N_down) - idx_query_random = fps(flatten_input_random_pc, batch_down, ratio=random_query_ratio) - query_random_pc = flatten_input_random_pc[idx_query_random].view(B, -1, D) - - # Randomly select sharpedge surface points and sharpedge query points - input_sharpedge_pc_size = int(num_sharpedge_query * self.downsample_ratio) - if input_sharpedge_pc_size == 0: - input_sharpedge_pc = torch.zeros(B, 0, D, dtype=input_random_pc.dtype).to(pc.device) - query_sharpedge_pc = torch.zeros(B, 0, D, dtype=query_random_pc.dtype).to(pc.device) - else: - sharpedge_query_ratio = num_sharpedge_query / input_sharpedge_pc_size - idx_sharpedge_pc = torch.randperm(sharpedge_pc.shape[1], device=sharpedge_pc.device)[ - :input_sharpedge_pc_size] - input_sharpedge_pc = sharpedge_pc[:, idx_sharpedge_pc, :] - flatten_input_sharpedge_surface_points = input_sharpedge_pc.view(B * input_sharpedge_pc_size, D) - N_down = int(flatten_input_sharpedge_surface_points.shape[0] / B) - batch_down = torch.arange(B).to(pc.device) - batch_down = torch.repeat_interleave(batch_down, N_down) - idx_query_sharpedge = fps(flatten_input_sharpedge_surface_points, batch_down, ratio=sharpedge_query_ratio) - query_sharpedge_pc = flatten_input_sharpedge_surface_points[idx_query_sharpedge].view(B, -1, D) - - # Concatenate random and sharpedge surface points and query points - query_pc = torch.cat([query_random_pc, query_sharpedge_pc], dim=1) - input_pc = torch.cat([input_random_pc, input_sharpedge_pc], dim=1) - - # PE - query = self.fourier_embedder(query_pc) - data = self.fourier_embedder(input_pc) - - # Concat normal if given - if self.point_feats != 0: - - random_surface_feats, sharpedge_surface_feats = torch.split(feats, [self.pc_size, self.pc_sharpedge_size], - dim=1) - input_random_surface_feats = random_surface_feats[:, idx_random_pc, :] - flatten_input_random_surface_feats = input_random_surface_feats.view(B * input_random_pc_size, -1) - query_random_feats = flatten_input_random_surface_feats[idx_query_random].view(B, -1, - flatten_input_random_surface_feats.shape[ - -1]) - - if input_sharpedge_pc_size == 0: - input_sharpedge_surface_feats = torch.zeros(B, 0, self.point_feats, - dtype=input_random_surface_feats.dtype).to(pc.device) - query_sharpedge_feats = torch.zeros(B, 0, self.point_feats, dtype=query_random_feats.dtype).to( - pc.device) - else: - input_sharpedge_surface_feats = sharpedge_surface_feats[:, idx_sharpedge_pc, :] - flatten_input_sharpedge_surface_feats = input_sharpedge_surface_feats.view(B * input_sharpedge_pc_size, - -1) - query_sharpedge_feats = flatten_input_sharpedge_surface_feats[idx_query_sharpedge].view(B, -1, - flatten_input_sharpedge_surface_feats.shape[ - -1]) - - query_feats = torch.cat([query_random_feats, query_sharpedge_feats], dim=1) - input_feats = torch.cat([input_random_surface_feats, input_sharpedge_surface_feats], dim=1) - - if self.normal_pe: - query_normal_pe = self.fourier_embedder(query_feats[..., :3]) - input_normal_pe = self.fourier_embedder(input_feats[..., :3]) - query_feats = torch.cat([query_normal_pe, query_feats[..., 3:]], dim=-1) - input_feats = torch.cat([input_normal_pe, input_feats[..., 3:]], dim=-1) - - query = torch.cat([query, query_feats], dim=-1) - data = torch.cat([data, input_feats], dim=-1) - - if input_sharpedge_pc_size == 0: - query_sharpedge_pc = torch.zeros(B, 1, D).to(pc.device) - input_sharpedge_pc = torch.zeros(B, 1, D).to(pc.device) - - # print(f'query_pc: {query_pc.shape}') - # print(f'input_pc: {input_pc.shape}') - # print(f'query_random_pc: {query_random_pc.shape}') - # print(f'input_random_pc: {input_random_pc.shape}') - # print(f'query_sharpedge_pc: {query_sharpedge_pc.shape}') - # print(f'input_sharpedge_pc: {input_sharpedge_pc.shape}') - - return query.view(B, -1, query.shape[-1]), data.view(B, -1, data.shape[-1]), [query_pc, input_pc, - query_random_pc, input_random_pc, - query_sharpedge_pc, - input_sharpedge_pc] - - def forward(self, pc, feats): - """ - - Args: - pc (torch.FloatTensor): [B, N, 3] - feats (torch.FloatTensor or None): [B, N, C] - - Returns: - - """ - - query, data, pc_infos = self.sample_points_and_latents(pc, feats) - - query = self.input_proj(query) - query = query - data = self.input_proj(data) - data = data - - latents = self.cross_attn(query, data) - if self.self_attn is not None: - latents = self.self_attn(latents) - - if self.ln_post is not None: - latents = self.ln_post(latents) - - return latents, pc_infos diff --git a/hy3dshape/hy3dshape/models/autoencoders/attention_processors.py b/hy3dshape/hy3dshape/models/autoencoders/attention_processors.py deleted file mode 100644 index f7b232eb0d16d67b9598a7e49f57d6616f2e5bed..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/attention_processors.py +++ /dev/null @@ -1,96 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os - -import torch -import torch.nn.functional as F - -scaled_dot_product_attention = F.scaled_dot_product_attention -if os.environ.get('CA_USE_SAGEATTN', '0') == '1': - try: - from sageattention import sageattn - except ImportError: - raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.') - scaled_dot_product_attention = sageattn - - -class CrossAttentionProcessor: - def __call__(self, attn, q, k, v): - out = scaled_dot_product_attention(q, k, v) - return out - - -class FlashVDMCrossAttentionProcessor: - def __init__(self, topk=None): - self.topk = topk - - def __call__(self, attn, q, k, v): - if k.shape[-2] == 3072: - topk = 1024 - elif k.shape[-2] == 512: - topk = 256 - else: - topk = k.shape[-2] // 3 - - if self.topk is True: - q1 = q[:, :, ::100, :] - sim = q1 @ k.transpose(-1, -2) - sim = torch.mean(sim, -2) - topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1) - topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1]) - v0 = torch.gather(v, dim=-2, index=topk_ind) - k0 = torch.gather(k, dim=-2, index=topk_ind) - out = scaled_dot_product_attention(q, k0, v0) - elif self.topk is False: - out = scaled_dot_product_attention(q, k, v) - else: - idx, counts = self.topk - start = 0 - outs = [] - for grid_coord, count in zip(idx, counts): - end = start + count - q_chunk = q[:, :, start:end, :] - k0, v0 = self.select_topkv(q_chunk, k, v, topk) - out = scaled_dot_product_attention(q_chunk, k0, v0) - outs.append(out) - start += count - out = torch.cat(outs, dim=-2) - self.topk = False - return out - - def select_topkv(self, q_chunk, k, v, topk): - q1 = q_chunk[:, :, ::50, :] - sim = q1 @ k.transpose(-1, -2) - sim = torch.mean(sim, -2) - topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1) - topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1]) - v0 = torch.gather(v, dim=-2, index=topk_ind) - k0 = torch.gather(k, dim=-2, index=topk_ind) - return k0, v0 - - -class FlashVDMTopMCrossAttentionProcessor(FlashVDMCrossAttentionProcessor): - def select_topkv(self, q_chunk, k, v, topk): - q1 = q_chunk[:, :, ::30, :] - sim = q1 @ k.transpose(-1, -2) - # sim = sim.to(torch.float32) - sim = sim.softmax(-1) - sim = torch.mean(sim, 1) - activated_token = torch.where(sim > 1e-6)[2] - index = torch.unique(activated_token, return_counts=True)[0].unsqueeze(0).unsqueeze(0).unsqueeze(-1) - index = index.expand(-1, v.shape[1], -1, v.shape[-1]) - v0 = torch.gather(v, dim=-2, index=index) - k0 = torch.gather(k, dim=-2, index=index) - return k0, v0 diff --git a/hy3dshape/hy3dshape/models/autoencoders/model.py b/hy3dshape/hy3dshape/models/autoencoders/model.py deleted file mode 100644 index 3d782f5804e6caf09ba72a47c8b3d2ddbedc1570..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/model.py +++ /dev/null @@ -1,339 +0,0 @@ -# Open Source Model Licensed under the Apache License Version 2.0 -# and Other Licenses of the Third-Party Components therein: -# The below Model in this distribution may have been modified by THL A29 Limited -# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited. - -# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. -# The below software and/or models in this distribution may have been -# modified by THL A29 Limited ("Tencent Modifications"). -# All Tencent Modifications are Copyright (C) THL A29 Limited. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - - -import os -from typing import Union, List - -import numpy as np -import torch -import torch.nn as nn -import yaml - -from .attention_blocks import FourierEmbedder, Transformer, CrossAttentionDecoder, PointCrossAttentionEncoder -from .surface_extractors import MCSurfaceExtractor, SurfaceExtractors -from .volume_decoders import VanillaVolumeDecoder, FlashVDMVolumeDecoding, HierarchicalVolumeDecoding -from ...utils import logger, synchronize_timer, smart_load_model - - -class DiagonalGaussianDistribution(object): - def __init__(self, parameters: Union[torch.Tensor, List[torch.Tensor]], deterministic=False, feat_dim=1): - """ - Initialize a diagonal Gaussian distribution with mean and log-variance parameters. - - Args: - parameters (Union[torch.Tensor, List[torch.Tensor]]): - Either a single tensor containing concatenated mean and log-variance along `feat_dim`, - or a list of two tensors [mean, logvar]. - deterministic (bool, optional): If True, the distribution is deterministic (zero variance). - Default is False. feat_dim (int, optional): Dimension along which mean and logvar are - concatenated if parameters is a single tensor. Default is 1. - """ - self.feat_dim = feat_dim - self.parameters = parameters - - if isinstance(parameters, list): - self.mean = parameters[0] - self.logvar = parameters[1] - else: - self.mean, self.logvar = torch.chunk(parameters, 2, dim=feat_dim) - - self.logvar = torch.clamp(self.logvar, -30.0, 20.0) - self.deterministic = deterministic - self.std = torch.exp(0.5 * self.logvar) - self.var = torch.exp(self.logvar) - if self.deterministic: - self.var = self.std = torch.zeros_like(self.mean) - - def sample(self): - """ - Sample from the diagonal Gaussian distribution. - - Returns: - torch.Tensor: A sample tensor with the same shape as the mean. - """ - x = self.mean + self.std * torch.randn_like(self.mean) - return x - - def kl(self, other=None, dims=(1, 2, 3)): - """ - Compute the Kullback-Leibler (KL) divergence between this distribution and another. - - If `other` is None, compute KL divergence to a standard normal distribution N(0, I). - - Args: - other (DiagonalGaussianDistribution, optional): Another diagonal Gaussian distribution. - dims (tuple, optional): Dimensions along which to compute the mean KL divergence. - Default is (1, 2, 3). - - Returns: - torch.Tensor: The mean KL divergence value. - """ - if self.deterministic: - return torch.Tensor([0.]) - else: - if other is None: - return 0.5 * torch.mean(torch.pow(self.mean, 2) - + self.var - 1.0 - self.logvar, - dim=dims) - else: - return 0.5 * torch.mean( - torch.pow(self.mean - other.mean, 2) / other.var - + self.var / other.var - 1.0 - self.logvar + other.logvar, - dim=dims) - - def nll(self, sample, dims=(1, 2, 3)): - if self.deterministic: - return torch.Tensor([0.]) - logtwopi = np.log(2.0 * np.pi) - return 0.5 * torch.sum( - logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, - dim=dims) - - def mode(self): - return self.mean - - -class VectsetVAE(nn.Module): - - @classmethod - @synchronize_timer('VectsetVAE Model Loading') - def from_single_file( - cls, - ckpt_path, - config_path, - device='cuda', - dtype=torch.float16, - use_safetensors=None, - **kwargs, - ): - # load config - with open(config_path, 'r') as f: - config = yaml.safe_load(f) - - # load ckpt - if use_safetensors: - ckpt_path = ckpt_path.replace('.ckpt', '.safetensors') - if not os.path.exists(ckpt_path): - raise FileNotFoundError(f"Model file {ckpt_path} not found") - - logger.info(f"Loading model from {ckpt_path}") - if use_safetensors: - import safetensors.torch - ckpt = safetensors.torch.load_file(ckpt_path, device='cpu') - else: - ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=True) - - model_kwargs = config['params'] - model_kwargs.update(kwargs) - - model = cls(**model_kwargs) - model.load_state_dict(ckpt) - model.to(device=device, dtype=dtype) - return model - - @classmethod - def from_pretrained( - cls, - model_path, - device='cuda', - dtype=torch.float16, - use_safetensors=False, - variant='fp16', - subfolder='hunyuan3d-vae-v2-1', - **kwargs, - ): - config_path, ckpt_path = smart_load_model( - model_path, - subfolder=subfolder, - use_safetensors=use_safetensors, - variant=variant - ) - - return cls.from_single_file( - ckpt_path, - config_path, - device=device, - dtype=dtype, - use_safetensors=use_safetensors, - **kwargs - ) - - def init_from_ckpt(self, path, ignore_keys=()): - state_dict = torch.load(path, map_location="cpu") - state_dict = state_dict.get("state_dict", state_dict) - keys = list(state_dict.keys()) - for k in keys: - for ik in ignore_keys: - if k.startswith(ik): - print("Deleting key {} from state_dict.".format(k)) - del state_dict[k] - missing, unexpected = self.load_state_dict(state_dict, strict=False) - print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") - if len(missing) > 0: - print(f"Missing Keys: {missing}") - print(f"Unexpected Keys: {unexpected}") - - def __init__( - self, - volume_decoder=None, - surface_extractor=None - ): - super().__init__() - if volume_decoder is None: - volume_decoder = VanillaVolumeDecoder() - if surface_extractor is None: - surface_extractor = MCSurfaceExtractor() - self.volume_decoder = volume_decoder - self.surface_extractor = surface_extractor - - def latents2mesh(self, latents: torch.FloatTensor, **kwargs): - with synchronize_timer('Volume decoding'): - grid_logits = self.volume_decoder(latents, self.geo_decoder, **kwargs) - with synchronize_timer('Surface extraction'): - outputs = self.surface_extractor(grid_logits, **kwargs) - return outputs - - def enable_flashvdm_decoder( - self, - enabled: bool = True, - adaptive_kv_selection=True, - topk_mode='mean', - mc_algo='dmc', - ): - if enabled: - if adaptive_kv_selection: - self.volume_decoder = FlashVDMVolumeDecoding(topk_mode) - else: - self.volume_decoder = HierarchicalVolumeDecoding() - if mc_algo not in SurfaceExtractors.keys(): - raise ValueError(f'Unsupported mc_algo {mc_algo}, available:{list(SurfaceExtractors.keys())}') - self.surface_extractor = SurfaceExtractors[mc_algo]() - else: - self.volume_decoder = VanillaVolumeDecoder() - self.surface_extractor = MCSurfaceExtractor() - - -class ShapeVAE(VectsetVAE): - def __init__( - self, - *, - num_latents: int, - embed_dim: int, - width: int, - heads: int, - num_decoder_layers: int, - num_encoder_layers: int = 8, - pc_size: int = 5120, - pc_sharpedge_size: int = 5120, - point_feats: int = 3, - downsample_ratio: int = 20, - geo_decoder_downsample_ratio: int = 1, - geo_decoder_mlp_expand_ratio: int = 4, - geo_decoder_ln_post: bool = True, - num_freqs: int = 8, - include_pi: bool = True, - qkv_bias: bool = True, - qk_norm: bool = False, - label_type: str = "binary", - drop_path_rate: float = 0.0, - scale_factor: float = 1.0, - use_ln_post: bool = True, - ckpt_path = None - ): - super().__init__() - self.geo_decoder_ln_post = geo_decoder_ln_post - self.downsample_ratio = downsample_ratio - - self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi) - - self.encoder = PointCrossAttentionEncoder( - fourier_embedder=self.fourier_embedder, - num_latents=num_latents, - downsample_ratio=self.downsample_ratio, - pc_size=pc_size, - pc_sharpedge_size=pc_sharpedge_size, - point_feats=point_feats, - width=width, - heads=heads, - layers=num_encoder_layers, - qkv_bias=qkv_bias, - use_ln_post=use_ln_post, - qk_norm=qk_norm - ) - - self.pre_kl = nn.Linear(width, embed_dim * 2) - self.post_kl = nn.Linear(embed_dim, width) - - self.transformer = Transformer( - n_ctx=num_latents, - width=width, - layers=num_decoder_layers, - heads=heads, - qkv_bias=qkv_bias, - qk_norm=qk_norm, - drop_path_rate=drop_path_rate - ) - - self.geo_decoder = CrossAttentionDecoder( - fourier_embedder=self.fourier_embedder, - out_channels=1, - num_latents=num_latents, - mlp_expand_ratio=geo_decoder_mlp_expand_ratio, - downsample_ratio=geo_decoder_downsample_ratio, - enable_ln_post=self.geo_decoder_ln_post, - width=width // geo_decoder_downsample_ratio, - heads=heads // geo_decoder_downsample_ratio, - qkv_bias=qkv_bias, - qk_norm=qk_norm, - label_type=label_type, - ) - - self.scale_factor = scale_factor - self.latent_shape = (num_latents, embed_dim) - - if ckpt_path is not None: - self.init_from_ckpt(ckpt_path) - - def forward(self, latents): - latents = self.post_kl(latents) - latents = self.transformer(latents) - return latents - - def encode(self, surface, sample_posterior=True): - pc, feats = surface[:, :, :3], surface[:, :, 3:] - latents, _ = self.encoder(pc, feats) - # print(latents.shape, self.pre_kl.weight.shape) - moments = self.pre_kl(latents) - posterior = DiagonalGaussianDistribution(moments, feat_dim=-1) - if sample_posterior: - latents = posterior.sample() - else: - latents = posterior.mode() - return latents - - def decode(self, latents): - latents = self.post_kl(latents) - latents = self.transformer(latents) - return latents diff --git a/hy3dshape/hy3dshape/models/autoencoders/surface_extractors.py b/hy3dshape/hy3dshape/models/autoencoders/surface_extractors.py deleted file mode 100644 index 8282aa2dc4e7b2b18b5605edf5dea7f32880e696..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/surface_extractors.py +++ /dev/null @@ -1,164 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from typing import Union, Tuple, List - -import numpy as np -import torch -from skimage import measure - - -class Latent2MeshOutput: - def __init__(self, mesh_v=None, mesh_f=None): - self.mesh_v = mesh_v - self.mesh_f = mesh_f - - -def center_vertices(vertices): - """Translate the vertices so that bounding box is centered at zero.""" - vert_min = vertices.min(dim=0)[0] - vert_max = vertices.max(dim=0)[0] - vert_center = 0.5 * (vert_min + vert_max) - return vertices - vert_center - - -class SurfaceExtractor: - def _compute_box_stat(self, bounds: Union[Tuple[float], List[float], float], octree_resolution: int): - """ - Compute grid size, bounding box minimum coordinates, and bounding box size based on input - bounds and resolution. - - Args: - bounds (Union[Tuple[float], List[float], float]): Bounding box coordinates or a single - float representing half side length. - If float, bounds are assumed symmetric around zero in all axes. - Expected format if list/tuple: [xmin, ymin, zmin, xmax, ymax, zmax]. - octree_resolution (int): Resolution of the octree grid. - - Returns: - grid_size (List[int]): Grid size along each axis (x, y, z), each equal to octree_resolution + 1. - bbox_min (np.ndarray): Minimum coordinates of the bounding box (xmin, ymin, zmin). - bbox_size (np.ndarray): Size of the bounding box along each axis (xmax - xmin, etc.). - """ - if isinstance(bounds, float): - bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds] - - bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6]) - bbox_size = bbox_max - bbox_min - grid_size = [int(octree_resolution) + 1, int(octree_resolution) + 1, int(octree_resolution) + 1] - return grid_size, bbox_min, bbox_size - - def run(self, *args, **kwargs): - """ - Abstract method to extract surface mesh from grid logits. - - This method should be implemented by subclasses. - - Raises: - NotImplementedError: Always, since this is an abstract method. - """ - return NotImplementedError - - def __call__(self, grid_logits, **kwargs): - """ - Process a batch of grid logits to extract surface meshes. - - Args: - grid_logits (torch.Tensor): Batch of grid logits with shape (batch_size, ...). - **kwargs: Additional keyword arguments passed to the `run` method. - - Returns: - List[Optional[Latent2MeshOutput]]: List of mesh outputs for each grid in the batch. - If extraction fails for a grid, None is appended at that position. - """ - outputs = [] - for i in range(grid_logits.shape[0]): - try: - vertices, faces = self.run(grid_logits[i], **kwargs) - vertices = vertices.astype(np.float32) - faces = np.ascontiguousarray(faces) - outputs.append(Latent2MeshOutput(mesh_v=vertices, mesh_f=faces)) - - except Exception: - import traceback - traceback.print_exc() - outputs.append(None) - - return outputs - - -class MCSurfaceExtractor(SurfaceExtractor): - def run(self, grid_logit, *, mc_level, bounds, octree_resolution, **kwargs): - """ - Extract surface mesh using the Marching Cubes algorithm. - - Args: - grid_logit (torch.Tensor): 3D grid logits tensor representing the scalar field. - mc_level (float): The level (iso-value) at which to extract the surface. - bounds (Union[Tuple[float], List[float], float]): Bounding box coordinates or half side length. - octree_resolution (int): Resolution of the octree grid. - **kwargs: Additional keyword arguments (ignored). - - Returns: - Tuple[np.ndarray, np.ndarray]: Tuple containing: - - vertices (np.ndarray): Extracted mesh vertices, scaled and translated to bounding - box coordinates. - - faces (np.ndarray): Extracted mesh faces (triangles). - """ - vertices, faces, normals, _ = measure.marching_cubes(grid_logit.cpu().numpy(), - mc_level, - method="lewiner") - grid_size, bbox_min, bbox_size = self._compute_box_stat(bounds, octree_resolution) - vertices = vertices / grid_size * bbox_size + bbox_min - return vertices, faces - - -class DMCSurfaceExtractor(SurfaceExtractor): - def run(self, grid_logit, *, octree_resolution, **kwargs): - """ - Extract surface mesh using Differentiable Marching Cubes (DMC) algorithm. - - Args: - grid_logit (torch.Tensor): 3D grid logits tensor representing the scalar field. - octree_resolution (int): Resolution of the octree grid. - **kwargs: Additional keyword arguments (ignored). - - Returns: - Tuple[np.ndarray, np.ndarray]: Tuple containing: - - vertices (np.ndarray): Extracted mesh vertices, centered and converted to numpy. - - faces (np.ndarray): Extracted mesh faces (triangles), with reversed vertex order. - - Raises: - ImportError: If the 'diso' package is not installed. - """ - device = grid_logit.device - if not hasattr(self, 'dmc'): - try: - from diso import DiffDMC - self.dmc = DiffDMC(dtype=torch.float32).to(device) - except: - raise ImportError("Please install diso via `pip install diso`, or set mc_algo to 'mc'") - sdf = -grid_logit / octree_resolution - sdf = sdf.to(torch.float32).contiguous() - verts, faces = self.dmc(sdf, deform=None, return_quads=False, normalize=True) - verts = center_vertices(verts) - vertices = verts.detach().cpu().numpy() - faces = faces.detach().cpu().numpy()[:, ::-1] - return vertices, faces - - -SurfaceExtractors = { - 'mc': MCSurfaceExtractor, - 'dmc': DMCSurfaceExtractor, -} diff --git a/hy3dshape/hy3dshape/models/autoencoders/volume_decoders.py b/hy3dshape/hy3dshape/models/autoencoders/volume_decoders.py deleted file mode 100644 index d7bfd84647786b43d7815c788931326a305c7dca..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/autoencoders/volume_decoders.py +++ /dev/null @@ -1,435 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from typing import Union, Tuple, List, Callable - -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F -from einops import repeat -from tqdm import tqdm - -from .attention_blocks import CrossAttentionDecoder -from .attention_processors import FlashVDMCrossAttentionProcessor, FlashVDMTopMCrossAttentionProcessor -from ...utils import logger - - -def extract_near_surface_volume_fn(input_tensor: torch.Tensor, alpha: float): - device = input_tensor.device - D = input_tensor.shape[0] - signed_val = 0.0 - - # 添加偏移并处理无效值 - val = input_tensor + alpha - valid_mask = val > -9000 # 假设-9000是无效值 - - # 改进的邻居获取函数（保持维度一致） - def get_neighbor(t, shift, axis): - """根据指定轴进行位移并保持维度一致""" - if shift == 0: - return t.clone() - - # 确定填充轴（输入为[D, D, D]对应z,y,x轴） - pad_dims = [0, 0, 0, 0, 0, 0] # 格式：[x前，x后，y前，y后，z前，z后] - - # 根据轴类型设置填充 - if axis == 0: # x轴（最后一个维度） - pad_idx = 0 if shift > 0 else 1 - pad_dims[pad_idx] = abs(shift) - elif axis == 1: # y轴（中间维度） - pad_idx = 2 if shift > 0 else 3 - pad_dims[pad_idx] = abs(shift) - elif axis == 2: # z轴（第一个维度） - pad_idx = 4 if shift > 0 else 5 - pad_dims[pad_idx] = abs(shift) - - # 执行填充（添加batch和channel维度适配F.pad） - padded = F.pad(t.unsqueeze(0).unsqueeze(0), pad_dims[::-1], mode='replicate') # 反转顺序适配F.pad - - # 构建动态切片索引 - slice_dims = [slice(None)] * 3 # 初始化为全切片 - if axis == 0: # x轴（dim=2） - if shift > 0: - slice_dims[0] = slice(shift, None) - else: - slice_dims[0] = slice(None, shift) - elif axis == 1: # y轴（dim=1） - if shift > 0: - slice_dims[1] = slice(shift, None) - else: - slice_dims[1] = slice(None, shift) - elif axis == 2: # z轴（dim=0） - if shift > 0: - slice_dims[2] = slice(shift, None) - else: - slice_dims[2] = slice(None, shift) - - # 应用切片并恢复维度 - padded = padded.squeeze(0).squeeze(0) - sliced = padded[slice_dims] - return sliced - - # 获取各方向邻居（确保维度一致） - left = get_neighbor(val, 1, axis=0) # x方向 - right = get_neighbor(val, -1, axis=0) - back = get_neighbor(val, 1, axis=1) # y方向 - front = get_neighbor(val, -1, axis=1) - down = get_neighbor(val, 1, axis=2) # z方向 - up = get_neighbor(val, -1, axis=2) - - # 处理边界无效值（使用where保持维度一致） - def safe_where(neighbor): - return torch.where(neighbor > -9000, neighbor, val) - - left = safe_where(left) - right = safe_where(right) - back = safe_where(back) - front = safe_where(front) - down = safe_where(down) - up = safe_where(up) - - # 计算符号一致性（转换为float32确保精度） - sign = torch.sign(val.to(torch.float32)) - neighbors_sign = torch.stack([ - torch.sign(left.to(torch.float32)), - torch.sign(right.to(torch.float32)), - torch.sign(back.to(torch.float32)), - torch.sign(front.to(torch.float32)), - torch.sign(down.to(torch.float32)), - torch.sign(up.to(torch.float32)) - ], dim=0) - - # 检查所有符号是否一致 - same_sign = torch.all(neighbors_sign == sign, dim=0) - - # 生成最终掩码 - mask = (~same_sign).to(torch.int32) - return mask * valid_mask.to(torch.int32) - - -def generate_dense_grid_points( - bbox_min: np.ndarray, - bbox_max: np.ndarray, - octree_resolution: int, - indexing: str = "ij", -): - length = bbox_max - bbox_min - num_cells = octree_resolution - - x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32) - y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32) - z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32) - [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing) - xyz = np.stack((xs, ys, zs), axis=-1) - grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1] - - return xyz, grid_size, length - - -class VanillaVolumeDecoder: - @torch.no_grad() - def __call__( - self, - latents: torch.FloatTensor, - geo_decoder: Callable, - bounds: Union[Tuple[float], List[float], float] = 1.01, - num_chunks: int = 10000, - octree_resolution: int = None, - enable_pbar: bool = True, - **kwargs, - ): - device = latents.device - dtype = latents.dtype - batch_size = latents.shape[0] - - # 1. generate query points - if isinstance(bounds, float): - bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds] - - bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6]) - xyz_samples, grid_size, length = generate_dense_grid_points( - bbox_min=bbox_min, - bbox_max=bbox_max, - octree_resolution=octree_resolution, - indexing="ij" - ) - xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3) - - # 2. latents to 3d volume - batch_logits = [] - for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc=f"Volume Decoding", - disable=not enable_pbar): - chunk_queries = xyz_samples[start: start + num_chunks, :] - chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size) - logits = geo_decoder(queries=chunk_queries, latents=latents) - batch_logits.append(logits) - - grid_logits = torch.cat(batch_logits, dim=1) - grid_logits = grid_logits.view((batch_size, *grid_size)).float() - - return grid_logits - - -class HierarchicalVolumeDecoding: - @torch.no_grad() - def __call__( - self, - latents: torch.FloatTensor, - geo_decoder: Callable, - bounds: Union[Tuple[float], List[float], float] = 1.01, - num_chunks: int = 10000, - mc_level: float = 0.0, - octree_resolution: int = None, - min_resolution: int = 63, - enable_pbar: bool = True, - **kwargs, - ): - device = latents.device - dtype = latents.dtype - - resolutions = [] - if octree_resolution < min_resolution: - resolutions.append(octree_resolution) - while octree_resolution >= min_resolution: - resolutions.append(octree_resolution) - octree_resolution = octree_resolution // 2 - resolutions.reverse() - - # 1. generate query points - if isinstance(bounds, float): - bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds] - bbox_min = np.array(bounds[0:3]) - bbox_max = np.array(bounds[3:6]) - bbox_size = bbox_max - bbox_min - - xyz_samples, grid_size, length = generate_dense_grid_points( - bbox_min=bbox_min, - bbox_max=bbox_max, - octree_resolution=resolutions[0], - indexing="ij" - ) - - dilate = nn.Conv3d(1, 1, 3, padding=1, bias=False, device=device, dtype=dtype) - dilate.weight = torch.nn.Parameter(torch.ones(dilate.weight.shape, dtype=dtype, device=device)) - - grid_size = np.array(grid_size) - xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3) - - # 2. latents to 3d volume - batch_logits = [] - batch_size = latents.shape[0] - for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), - desc=f"Hierarchical Volume Decoding [r{resolutions[0] + 1}]"): - queries = xyz_samples[start: start + num_chunks, :] - batch_queries = repeat(queries, "p c -> b p c", b=batch_size) - logits = geo_decoder(queries=batch_queries, latents=latents) - batch_logits.append(logits) - - grid_logits = torch.cat(batch_logits, dim=1).view((batch_size, grid_size[0], grid_size[1], grid_size[2])) - - for octree_depth_now in resolutions[1:]: - grid_size = np.array([octree_depth_now + 1] * 3) - resolution = bbox_size / octree_depth_now - next_index = torch.zeros(tuple(grid_size), dtype=dtype, device=device) - next_logits = torch.full(next_index.shape, -10000., dtype=dtype, device=device) - curr_points = extract_near_surface_volume_fn(grid_logits.squeeze(0), mc_level) - curr_points += grid_logits.squeeze(0).abs() < 0.95 - - if octree_depth_now == resolutions[-1]: - expand_num = 0 - else: - expand_num = 1 - for i in range(expand_num): - curr_points = dilate(curr_points.unsqueeze(0).to(dtype)).squeeze(0) - (cidx_x, cidx_y, cidx_z) = torch.where(curr_points > 0) - next_index[cidx_x * 2, cidx_y * 2, cidx_z * 2] = 1 - for i in range(2 - expand_num): - next_index = dilate(next_index.unsqueeze(0)).squeeze(0) - nidx = torch.where(next_index > 0) - - next_points = torch.stack(nidx, dim=1) - next_points = (next_points * torch.tensor(resolution, dtype=next_points.dtype, device=device) + - torch.tensor(bbox_min, dtype=next_points.dtype, device=device)) - batch_logits = [] - for start in tqdm(range(0, next_points.shape[0], num_chunks), - desc=f"Hierarchical Volume Decoding [r{octree_depth_now + 1}]"): - queries = next_points[start: start + num_chunks, :] - batch_queries = repeat(queries, "p c -> b p c", b=batch_size) - logits = geo_decoder(queries=batch_queries.to(latents.dtype), latents=latents) - batch_logits.append(logits) - grid_logits = torch.cat(batch_logits, dim=1) - next_logits[nidx] = grid_logits[0, ..., 0] - grid_logits = next_logits.unsqueeze(0) - grid_logits[grid_logits == -10000.] = float('nan') - - return grid_logits - - -class FlashVDMVolumeDecoding: - def __init__(self, topk_mode='mean'): - if topk_mode not in ['mean', 'merge']: - raise ValueError(f'Unsupported topk_mode {topk_mode}, available: {["mean", "merge"]}') - - if topk_mode == 'mean': - self.processor = FlashVDMCrossAttentionProcessor() - else: - self.processor = FlashVDMTopMCrossAttentionProcessor() - - @torch.no_grad() - def __call__( - self, - latents: torch.FloatTensor, - geo_decoder: CrossAttentionDecoder, - bounds: Union[Tuple[float], List[float], float] = 1.01, - num_chunks: int = 10000, - mc_level: float = 0.0, - octree_resolution: int = None, - min_resolution: int = 63, - mini_grid_num: int = 4, - enable_pbar: bool = True, - **kwargs, - ): - processor = self.processor - geo_decoder.set_cross_attention_processor(processor) - - device = latents.device - dtype = latents.dtype - - resolutions = [] - if octree_resolution < min_resolution: - resolutions.append(octree_resolution) - while octree_resolution >= min_resolution: - resolutions.append(octree_resolution) - octree_resolution = octree_resolution // 2 - resolutions.reverse() - resolutions[0] = round(resolutions[0] / mini_grid_num) * mini_grid_num - 1 - for i, resolution in enumerate(resolutions[1:]): - resolutions[i + 1] = resolutions[0] * 2 ** (i + 1) - - logger.info(f"FlashVDMVolumeDecoding Resolution: {resolutions}") - - # 1. generate query points - if isinstance(bounds, float): - bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds] - bbox_min = np.array(bounds[0:3]) - bbox_max = np.array(bounds[3:6]) - bbox_size = bbox_max - bbox_min - - xyz_samples, grid_size, length = generate_dense_grid_points( - bbox_min=bbox_min, - bbox_max=bbox_max, - octree_resolution=resolutions[0], - indexing="ij" - ) - - dilate = nn.Conv3d(1, 1, 3, padding=1, bias=False, device=device, dtype=dtype) - dilate.weight = torch.nn.Parameter(torch.ones(dilate.weight.shape, dtype=dtype, device=device)) - - grid_size = np.array(grid_size) - - # 2. latents to 3d volume - xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype) - batch_size = latents.shape[0] - mini_grid_size = xyz_samples.shape[0] // mini_grid_num - xyz_samples = xyz_samples.view( - mini_grid_num, mini_grid_size, - mini_grid_num, mini_grid_size, - mini_grid_num, mini_grid_size, 3 - ).permute( - 0, 2, 4, 1, 3, 5, 6 - ).reshape( - -1, mini_grid_size * mini_grid_size * mini_grid_size, 3 - ) - batch_logits = [] - num_batchs = max(num_chunks // xyz_samples.shape[1], 1) - for start in tqdm(range(0, xyz_samples.shape[0], num_batchs), - desc=f"FlashVDM Volume Decoding", disable=not enable_pbar): - queries = xyz_samples[start: start + num_batchs, :] - batch = queries.shape[0] - batch_latents = repeat(latents.squeeze(0), "p c -> b p c", b=batch) - processor.topk = True - logits = geo_decoder(queries=queries, latents=batch_latents) - batch_logits.append(logits) - grid_logits = torch.cat(batch_logits, dim=0).reshape( - mini_grid_num, mini_grid_num, mini_grid_num, - mini_grid_size, mini_grid_size, - mini_grid_size - ).permute(0, 3, 1, 4, 2, 5).contiguous().view( - (batch_size, grid_size[0], grid_size[1], grid_size[2]) - ) - - for octree_depth_now in resolutions[1:]: - grid_size = np.array([octree_depth_now + 1] * 3) - resolution = bbox_size / octree_depth_now - next_index = torch.zeros(tuple(grid_size), dtype=dtype, device=device) - next_logits = torch.full(next_index.shape, -10000., dtype=dtype, device=device) - curr_points = extract_near_surface_volume_fn(grid_logits.squeeze(0), mc_level) - curr_points += grid_logits.squeeze(0).abs() < 0.95 - - if octree_depth_now == resolutions[-1]: - expand_num = 0 - else: - expand_num = 1 - for i in range(expand_num): - curr_points = dilate(curr_points.unsqueeze(0).to(dtype)).squeeze(0) - (cidx_x, cidx_y, cidx_z) = torch.where(curr_points > 0) - - next_index[cidx_x * 2, cidx_y * 2, cidx_z * 2] = 1 - for i in range(2 - expand_num): - next_index = dilate(next_index.unsqueeze(0)).squeeze(0) - nidx = torch.where(next_index > 0) - - next_points = torch.stack(nidx, dim=1) - next_points = (next_points * torch.tensor(resolution, dtype=torch.float32, device=device) + - torch.tensor(bbox_min, dtype=torch.float32, device=device)) - - query_grid_num = 6 - min_val = next_points.min(axis=0).values - max_val = next_points.max(axis=0).values - vol_queries_index = (next_points - min_val) / (max_val - min_val) * (query_grid_num - 0.001) - index = torch.floor(vol_queries_index).long() - index = index[..., 0] * (query_grid_num ** 2) + index[..., 1] * query_grid_num + index[..., 2] - index = index.sort() - next_points = next_points[index.indices].unsqueeze(0).contiguous() - unique_values = torch.unique(index.values, return_counts=True) - grid_logits = torch.zeros((next_points.shape[1]), dtype=latents.dtype, device=latents.device) - input_grid = [[], []] - logits_grid_list = [] - start_num = 0 - sum_num = 0 - for grid_index, count in zip(unique_values[0].cpu().tolist(), unique_values[1].cpu().tolist()): - if sum_num + count < num_chunks or sum_num == 0: - sum_num += count - input_grid[0].append(grid_index) - input_grid[1].append(count) - else: - processor.topk = input_grid - logits_grid = geo_decoder(queries=next_points[:, start_num:start_num + sum_num], latents=latents) - start_num = start_num + sum_num - logits_grid_list.append(logits_grid) - input_grid = [[grid_index], [count]] - sum_num = count - if sum_num > 0: - processor.topk = input_grid - logits_grid = geo_decoder(queries=next_points[:, start_num:start_num + sum_num], latents=latents) - logits_grid_list.append(logits_grid) - logits_grid = torch.cat(logits_grid_list, dim=1) - grid_logits[index.indices] = logits_grid.squeeze(0).squeeze(-1) - next_logits[nidx] = grid_logits - grid_logits = next_logits.unsqueeze(0) - - grid_logits[grid_logits == -10000.] = float('nan') - - return grid_logits diff --git a/hy3dshape/hy3dshape/models/conditioner.py b/hy3dshape/hy3dshape/models/conditioner.py deleted file mode 100644 index d0d848c3b0d82eba4e4453d2c266f8fa7a1aeaaa..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/conditioner.py +++ /dev/null @@ -1,257 +0,0 @@ -# Open Source Model Licensed under the Apache License Version 2.0 -# and Other Licenses of the Third-Party Components therein: -# The below Model in this distribution may have been modified by THL A29 Limited -# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited. - -# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. -# The below software and/or models in this distribution may have been -# modified by THL A29 Limited ("Tencent Modifications"). -# All Tencent Modifications are Copyright (C) THL A29 Limited. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import numpy as np -import torch -import torch.nn as nn -from torchvision import transforms -from transformers import ( - CLIPVisionModelWithProjection, - CLIPVisionConfig, - Dinov2Model, - Dinov2Config, -) - - -def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): - """ - embed_dim: output dimension for each position - pos: a list of positions to be encoded: size (M,) - out: (M, D) - """ - assert embed_dim % 2 == 0 - omega = np.arange(embed_dim // 2, dtype=np.float64) - omega /= embed_dim / 2. - omega = 1. / 10000 ** omega # (D/2,) - - pos = pos.reshape(-1) # (M,) - out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product - - emb_sin = np.sin(out) # (M, D/2) - emb_cos = np.cos(out) # (M, D/2) - - return np.concatenate([emb_sin, emb_cos], axis=1) - - -class ImageEncoder(nn.Module): - def __init__( - self, - version=None, - config=None, - use_cls_token=True, - image_size=224, - **kwargs, - ): - super().__init__() - - if config is None: - self.model = self.MODEL_CLASS.from_pretrained(version) - else: - self.model = self.MODEL_CLASS(self.MODEL_CONFIG_CLASS.from_dict(config)) - self.model.eval() - self.model.requires_grad_(False) - self.use_cls_token = use_cls_token - self.size = image_size // 14 - self.num_patches = (image_size // 14) ** 2 - if self.use_cls_token: - self.num_patches += 1 - - self.transform = transforms.Compose( - [ - transforms.Resize(image_size, transforms.InterpolationMode.BILINEAR, antialias=True), - transforms.CenterCrop(image_size), - transforms.Normalize( - mean=self.mean, - std=self.std, - ), - ] - ) - - def forward(self, image, mask=None, value_range=(-1, 1), **kwargs): - if value_range is not None: - low, high = value_range - image = (image - low) / (high - low) - - image = image.to(self.model.device, dtype=self.model.dtype) - inputs = self.transform(image) - outputs = self.model(inputs) - - last_hidden_state = outputs.last_hidden_state - if not self.use_cls_token: - last_hidden_state = last_hidden_state[:, 1:, :] - - return last_hidden_state - - def unconditional_embedding(self, batch_size, **kwargs): - device = next(self.model.parameters()).device - dtype = next(self.model.parameters()).dtype - zero = torch.zeros( - batch_size, - self.num_patches, - self.model.config.hidden_size, - device=device, - dtype=dtype, - ) - - return zero - - -class CLIPImageEncoder(ImageEncoder): - MODEL_CLASS = CLIPVisionModelWithProjection - MODEL_CONFIG_CLASS = CLIPVisionConfig - mean = [0.48145466, 0.4578275, 0.40821073] - std = [0.26862954, 0.26130258, 0.27577711] - - -class DinoImageEncoder(ImageEncoder): - MODEL_CLASS = Dinov2Model - MODEL_CONFIG_CLASS = Dinov2Config - mean = [0.485, 0.456, 0.406] - std = [0.229, 0.224, 0.225] - - -class DinoImageEncoderMV(DinoImageEncoder): - def __init__( - self, - version=None, - config=None, - use_cls_token=True, - image_size=224, - view_num=4, - **kwargs, - ): - super().__init__(version, config, use_cls_token, image_size, **kwargs) - self.view_num = view_num - self.num_patches = self.num_patches - pos = np.arange(self.view_num, dtype=np.float32) - view_embedding = torch.from_numpy( - get_1d_sincos_pos_embed_from_grid(self.model.config.hidden_size, pos)).float() - - view_embedding = view_embedding.unsqueeze(1).repeat(1, self.num_patches, 1) - self.view_embed = view_embedding.unsqueeze(0) - - def forward(self, image, mask=None, value_range=(-1, 1), view_idxs=None): - if value_range is not None: - low, high = value_range - image = (image - low) / (high - low) - - image = image.to(self.model.device, dtype=self.model.dtype) - - bs, num_views, c, h, w = image.shape - image = image.view(bs * num_views, c, h, w) - - inputs = self.transform(image) - outputs = self.model(inputs) - - last_hidden_state = outputs.last_hidden_state - last_hidden_state = last_hidden_state.view( - bs, num_views, last_hidden_state.shape[-2], - last_hidden_state.shape[-1] - ) - - view_embedding = self.view_embed.to(last_hidden_state.dtype).to(last_hidden_state.device) - if view_idxs is not None: - assert len(view_idxs) == bs - view_embeddings = [] - for i in range(bs): - view_idx = view_idxs[i] - assert num_views == len(view_idx) - view_embeddings.append(self.view_embed[:, view_idx, ...]) - view_embedding = torch.cat(view_embeddings, 0).to(last_hidden_state.dtype).to(last_hidden_state.device) - - if num_views != self.view_num: - view_embedding = view_embedding[:, :num_views, ...] - last_hidden_state = last_hidden_state + view_embedding - last_hidden_state = last_hidden_state.view(bs, num_views * last_hidden_state.shape[-2], - last_hidden_state.shape[-1]) - return last_hidden_state - - def unconditional_embedding(self, batch_size, view_idxs=None, **kwargs): - device = next(self.model.parameters()).device - dtype = next(self.model.parameters()).dtype - zero = torch.zeros( - batch_size, - self.num_patches * len(view_idxs[0]), - self.model.config.hidden_size, - device=device, - dtype=dtype, - ) - return zero - - -def build_image_encoder(config): - if config['type'] == 'CLIPImageEncoder': - return CLIPImageEncoder(**config['kwargs']) - elif config['type'] == 'DinoImageEncoder': - return DinoImageEncoder(**config['kwargs']) - elif config['type'] == 'DinoImageEncoderMV': - return DinoImageEncoderMV(**config['kwargs']) - else: - raise ValueError(f'Unknown image encoder type: {config["type"]}') - - -class DualImageEncoder(nn.Module): - def __init__( - self, - main_image_encoder, - additional_image_encoder, - ): - super().__init__() - self.main_image_encoder = build_image_encoder(main_image_encoder) - self.additional_image_encoder = build_image_encoder(additional_image_encoder) - - def forward(self, image, mask=None, **kwargs): - outputs = { - 'main': self.main_image_encoder(image, mask=mask, **kwargs), - 'additional': self.additional_image_encoder(image, mask=mask, **kwargs), - } - return outputs - - def unconditional_embedding(self, batch_size, **kwargs): - outputs = { - 'main': self.main_image_encoder.unconditional_embedding(batch_size, **kwargs), - 'additional': self.additional_image_encoder.unconditional_embedding(batch_size, **kwargs), - } - return outputs - - -class SingleImageEncoder(nn.Module): - def __init__( - self, - main_image_encoder, - ): - super().__init__() - self.main_image_encoder = build_image_encoder(main_image_encoder) - - def forward(self, image, mask=None, **kwargs): - outputs = { - 'main': self.main_image_encoder(image, mask=mask, **kwargs), - } - return outputs - - def unconditional_embedding(self, batch_size, **kwargs): - outputs = { - 'main': self.main_image_encoder.unconditional_embedding(batch_size, **kwargs), - } - return outputs diff --git a/hy3dshape/hy3dshape/models/denoisers/__init__.py b/hy3dshape/hy3dshape/models/denoisers/__init__.py deleted file mode 100644 index 72609333e27d68d377905ba9e67655de7021c31b..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/denoisers/__init__.py +++ /dev/null @@ -1,15 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from .hunyuan3ddit import Hunyuan3DDiT diff --git a/hy3dshape/hy3dshape/models/denoisers/hunyuan3ddit.py b/hy3dshape/hy3dshape/models/denoisers/hunyuan3ddit.py deleted file mode 100644 index 5565cfaffa3500a60e56caa54f17ea46f4e27def..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/denoisers/hunyuan3ddit.py +++ /dev/null @@ -1,404 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import math -from dataclasses import dataclass -from typing import List, Tuple, Optional - -import torch -from torch import Tensor, nn -from einops import rearrange - -# set up attention backend -scaled_dot_product_attention = nn.functional.scaled_dot_product_attention -if os.environ.get('USE_SAGEATTN', '0') == '1': - try: - from sageattention import sageattn - except ImportError: - raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.') - scaled_dot_product_attention = sageattn - - -def attention(q: Tensor, k: Tensor, v: Tensor, **kwargs) -> Tensor: - x = scaled_dot_product_attention(q, k, v) - x = rearrange(x, "B H L D -> B L (H D)") - return x - - -def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0): - """ - Create sinusoidal timestep embeddings. - :param t: a 1-D Tensor of N indices, one per batch element. - These may be fractional. - :param dim: the dimension of the output. - :param max_period: controls the minimum frequency of the embeddings. - :return: an (N, D) Tensor of positional embeddings. - """ - t = time_factor * t - half = dim // 2 - freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half) - freqs = freqs.to(t.device) - - args = t[:, None].float() * freqs[None] - embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) - if dim % 2: - embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) - if torch.is_floating_point(t): - embedding = embedding.to(t) - return embedding - - -class GELU(nn.Module): - def __init__(self, approximate='tanh'): - super().__init__() - self.approximate = approximate - - def forward(self, x: Tensor) -> Tensor: - return nn.functional.gelu(x.contiguous(), approximate=self.approximate) - - -class MLPEmbedder(nn.Module): - def __init__(self, in_dim: int, hidden_dim: int): - super().__init__() - self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True) - self.silu = nn.SiLU() - self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True) - - def forward(self, x: Tensor) -> Tensor: - return self.out_layer(self.silu(self.in_layer(x))) - - -class RMSNorm(torch.nn.Module): - def __init__(self, dim: int): - super().__init__() - self.scale = nn.Parameter(torch.ones(dim)) - - def forward(self, x: Tensor): - x_dtype = x.dtype - x = x.float() - rrms = torch.rsqrt(torch.mean(x ** 2, dim=-1, keepdim=True) + 1e-6) - return (x * rrms).to(dtype=x_dtype) * self.scale - - -class QKNorm(torch.nn.Module): - def __init__(self, dim: int): - super().__init__() - self.query_norm = RMSNorm(dim) - self.key_norm = RMSNorm(dim) - - def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tuple[Tensor, Tensor]: - q = self.query_norm(q) - k = self.key_norm(k) - return q.to(v), k.to(v) - - -class SelfAttention(nn.Module): - def __init__( - self, - dim: int, - num_heads: int = 8, - qkv_bias: bool = False, - ): - super().__init__() - self.num_heads = num_heads - head_dim = dim // num_heads - - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.norm = QKNorm(head_dim) - self.proj = nn.Linear(dim, dim) - - def forward(self, x: Tensor, pe: Tensor) -> Tensor: - qkv = self.qkv(x) - q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) - q, k = self.norm(q, k, v) - x = attention(q, k, v, pe=pe) - x = self.proj(x) - return x - - -@dataclass -class ModulationOut: - shift: Tensor - scale: Tensor - gate: Tensor - - -class Modulation(nn.Module): - def __init__(self, dim: int, double: bool): - super().__init__() - self.is_double = double - self.multiplier = 6 if double else 3 - self.lin = nn.Linear(dim, self.multiplier * dim, bias=True) - - def forward(self, vec: Tensor) -> Tuple[ModulationOut, Optional[ModulationOut]]: - out = self.lin(nn.functional.silu(vec))[:, None, :] - out = out.chunk(self.multiplier, dim=-1) - - return ( - ModulationOut(*out[:3]), - ModulationOut(*out[3:]) if self.is_double else None, - ) - - -class DoubleStreamBlock(nn.Module): - def __init__( - self, - hidden_size: int, - num_heads: int, - mlp_ratio: float, - qkv_bias: bool = False, - ): - super().__init__() - mlp_hidden_dim = int(hidden_size * mlp_ratio) - self.num_heads = num_heads - self.hidden_size = hidden_size - self.img_mod = Modulation(hidden_size, double=True) - self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias) - - self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.img_mlp = nn.Sequential( - nn.Linear(hidden_size, mlp_hidden_dim, bias=True), - GELU(approximate="tanh"), - nn.Linear(mlp_hidden_dim, hidden_size, bias=True), - ) - - self.txt_mod = Modulation(hidden_size, double=True) - self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias) - - self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.txt_mlp = nn.Sequential( - nn.Linear(hidden_size, mlp_hidden_dim, bias=True), - GELU(approximate="tanh"), - nn.Linear(mlp_hidden_dim, hidden_size, bias=True), - ) - - def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor) -> Tuple[Tensor, Tensor]: - img_mod1, img_mod2 = self.img_mod(vec) - txt_mod1, txt_mod2 = self.txt_mod(vec) - - img_modulated = self.img_norm1(img) - img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift - img_qkv = self.img_attn.qkv(img_modulated) - img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) - img_q, img_k = self.img_attn.norm(img_q, img_k, img_v) - - txt_modulated = self.txt_norm1(txt) - txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift - txt_qkv = self.txt_attn.qkv(txt_modulated) - txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) - txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v) - - q = torch.cat((txt_q, img_q), dim=2) - k = torch.cat((txt_k, img_k), dim=2) - v = torch.cat((txt_v, img_v), dim=2) - - attn = attention(q, k, v, pe=pe) - txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:] - - img = img + img_mod1.gate * self.img_attn.proj(img_attn) - img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift) - - txt = txt + txt_mod1.gate * self.txt_attn.proj(txt_attn) - txt = txt + txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift) - return img, txt - - -class SingleStreamBlock(nn.Module): - """ - A DiT block with parallel linear layers as described in - https://arxiv.org/abs/2302.05442 and adapted modulation interface. - """ - - def __init__( - self, - hidden_size: int, - num_heads: int, - mlp_ratio: float = 4.0, - qk_scale: Optional[float] = None, - ): - super().__init__() - - self.hidden_dim = hidden_size - self.num_heads = num_heads - head_dim = hidden_size // num_heads - self.scale = qk_scale or head_dim ** -0.5 - - self.mlp_hidden_dim = int(hidden_size * mlp_ratio) - # qkv and mlp_in - self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim) - # proj and mlp_out - self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size) - - self.norm = QKNorm(head_dim) - - self.hidden_size = hidden_size - self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - - self.mlp_act = GELU(approximate="tanh") - self.modulation = Modulation(hidden_size, double=False) - - def forward(self, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor: - mod, _ = self.modulation(vec) - - x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift - qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1) - - q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads) - q, k = self.norm(q, k, v) - - # compute attention - attn = attention(q, k, v, pe=pe) - # compute activation in mlp stream, cat again and run second linear layer - output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2)) - return x + mod.gate * output - - -class LastLayer(nn.Module): - def __init__(self, hidden_size: int, patch_size: int, out_channels: int): - super().__init__() - self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) - self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True) - self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)) - - def forward(self, x: Tensor, vec: Tensor) -> Tensor: - shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1) - x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :] - x = self.linear(x) - return x - - -class Hunyuan3DDiT(nn.Module): - def __init__( - self, - in_channels: int = 64, - context_in_dim: int = 1536, - hidden_size: int = 1024, - mlp_ratio: float = 4.0, - num_heads: int = 16, - depth: int = 16, - depth_single_blocks: int = 32, - axes_dim: List[int] = [64], - theta: int = 10_000, - qkv_bias: bool = True, - time_factor: float = 1000, - guidance_embed: bool = False, - ckpt_path: Optional[str] = None, - **kwargs, - ): - super().__init__() - self.in_channels = in_channels - self.context_in_dim = context_in_dim - self.hidden_size = hidden_size - self.mlp_ratio = mlp_ratio - self.num_heads = num_heads - self.depth = depth - self.depth_single_blocks = depth_single_blocks - self.axes_dim = axes_dim - self.theta = theta - self.qkv_bias = qkv_bias - self.time_factor = time_factor - self.out_channels = self.in_channels - self.guidance_embed = guidance_embed - - if hidden_size % num_heads != 0: - raise ValueError( - f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}" - ) - pe_dim = hidden_size // num_heads - if sum(axes_dim) != pe_dim: - raise ValueError(f"Got {axes_dim} but expected positional dim {pe_dim}") - self.hidden_size = hidden_size - self.num_heads = num_heads - self.latent_in = nn.Linear(self.in_channels, self.hidden_size, bias=True) - self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) - self.cond_in = nn.Linear(context_in_dim, self.hidden_size) - self.guidance_in = ( - MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if guidance_embed else nn.Identity() - ) - - self.double_blocks = nn.ModuleList( - [ - DoubleStreamBlock( - self.hidden_size, - self.num_heads, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, - ) - for _ in range(depth) - ] - ) - - self.single_blocks = nn.ModuleList( - [ - SingleStreamBlock( - self.hidden_size, - self.num_heads, - mlp_ratio=mlp_ratio, - ) - for _ in range(depth_single_blocks) - ] - ) - - self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels) - - if ckpt_path is not None: - print('restored denoiser ckpt', ckpt_path) - - ckpt = torch.load(ckpt_path, map_location="cpu") - if 'state_dict' not in ckpt: - # deepspeed ckpt - state_dict = {} - for k in ckpt.keys(): - new_k = k.replace('_forward_module.', '') - state_dict[new_k] = ckpt[k] - else: - state_dict = ckpt["state_dict"] - - final_state_dict = {} - for k, v in state_dict.items(): - if k.startswith('model.'): - final_state_dict[k.replace('model.', '')] = v - else: - final_state_dict[k] = v - missing, unexpected = self.load_state_dict(final_state_dict, strict=False) - print('unexpected keys:', unexpected) - print('missing keys:', missing) - - def forward(self, x, t, contexts, **kwargs) -> Tensor: - cond = contexts['main'] - latent = self.latent_in(x) - - vec = self.time_in(timestep_embedding(t, 256, self.time_factor).to(dtype=latent.dtype)) - if self.guidance_embed: - guidance = kwargs.get('guidance', None) - if guidance is None: - raise ValueError("Didn't get guidance strength for guidance distilled model.") - vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.time_factor)) - - cond = self.cond_in(cond) - pe = None - - for block in self.double_blocks: - latent, cond = block(img=latent, txt=cond, vec=vec, pe=pe) - - latent = torch.cat((cond, latent), 1) - for block in self.single_blocks: - latent = block(latent, vec=vec, pe=pe) - - latent = latent[:, cond.shape[1]:, ...] - latent = self.final_layer(latent, vec) - return latent diff --git a/hy3dshape/hy3dshape/models/denoisers/hunyuandit.py b/hy3dshape/hy3dshape/models/denoisers/hunyuandit.py deleted file mode 100644 index b4b3b50baea7a544ee5cb89ca6dd4fabcdfad638..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/denoisers/hunyuandit.py +++ /dev/null @@ -1,596 +0,0 @@ -# Open Source Model Licensed under the Apache License Version 2.0 -# and Other Licenses of the Third-Party Components therein: -# The below Model in this distribution may have been modified by THL A29 Limited -# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited. - -# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved. -# The below software and/or models in this distribution may have been -# modified by THL A29 Limited ("Tencent Modifications"). -# All Tencent Modifications are Copyright (C) THL A29 Limited. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import math - -import numpy as np -import torch -import torch.nn as nn -import torch.nn.functional as F -from einops import rearrange - -from .moe_layers import MoEBlock - - -def modulate(x, shift, scale): - return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1) - - -def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): - """ - embed_dim: output dimension for each position - pos: a list of positions to be encoded: size (M,) - out: (M, D) - """ - assert embed_dim % 2 == 0 - omega = np.arange(embed_dim // 2, dtype=np.float64) - omega /= embed_dim / 2. - omega = 1. / 10000 ** omega # (D/2,) - - pos = pos.reshape(-1) # (M,) - out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product - - emb_sin = np.sin(out) # (M, D/2) - emb_cos = np.cos(out) # (M, D/2) - - return np.concatenate([emb_sin, emb_cos], axis=1) - - -class Timesteps(nn.Module): - def __init__(self, - num_channels: int, - downscale_freq_shift: float = 0.0, - scale: int = 1, - max_period: int = 10000 - ): - super().__init__() - self.num_channels = num_channels - self.downscale_freq_shift = downscale_freq_shift - self.scale = scale - self.max_period = max_period - - def forward(self, timesteps): - assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" - embedding_dim = self.num_channels - half_dim = embedding_dim // 2 - exponent = -math.log(self.max_period) * torch.arange( - start=0, end=half_dim, dtype=torch.float32, device=timesteps.device) - exponent = exponent / (half_dim - self.downscale_freq_shift) - emb = torch.exp(exponent) - emb = timesteps[:, None].float() * emb[None, :] - emb = self.scale * emb - emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) - if embedding_dim % 2 == 1: - emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) - return emb - - -class TimestepEmbedder(nn.Module): - """ - Embeds scalar timesteps into vector representations. - """ - - def __init__(self, hidden_size, frequency_embedding_size=256, cond_proj_dim=None, out_size=None): - super().__init__() - if out_size is None: - out_size = hidden_size - self.mlp = nn.Sequential( - nn.Linear(hidden_size, frequency_embedding_size, bias=True), - nn.GELU(), - nn.Linear(frequency_embedding_size, out_size, bias=True), - ) - self.frequency_embedding_size = frequency_embedding_size - - if cond_proj_dim is not None: - self.cond_proj = nn.Linear(cond_proj_dim, frequency_embedding_size, bias=False) - - self.time_embed = Timesteps(hidden_size) - - def forward(self, t, condition): - - t_freq = self.time_embed(t).type(self.mlp[0].weight.dtype) - - # t_freq = timestep_embedding(t, self.frequency_embedding_size).type(self.mlp[0].weight.dtype) - if condition is not None: - t_freq = t_freq + self.cond_proj(condition) - - t = self.mlp(t_freq) - t = t.unsqueeze(dim=1) - return t - - -class MLP(nn.Module): - def __init__(self, *, width: int): - super().__init__() - self.width = width - self.fc1 = nn.Linear(width, width * 4) - self.fc2 = nn.Linear(width * 4, width) - self.gelu = nn.GELU() - - def forward(self, x): - return self.fc2(self.gelu(self.fc1(x))) - - -class CrossAttention(nn.Module): - def __init__( - self, - qdim, - kdim, - num_heads, - qkv_bias=True, - qk_norm=False, - norm_layer=nn.LayerNorm, - with_decoupled_ca=False, - decoupled_ca_dim=16, - decoupled_ca_weight=1.0, - **kwargs, - ): - super().__init__() - self.qdim = qdim - self.kdim = kdim - self.num_heads = num_heads - assert self.qdim % num_heads == 0, "self.qdim must be divisible by num_heads" - self.head_dim = self.qdim // num_heads - assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" - self.scale = self.head_dim ** -0.5 - - self.to_q = nn.Linear(qdim, qdim, bias=qkv_bias) - self.to_k = nn.Linear(kdim, qdim, bias=qkv_bias) - self.to_v = nn.Linear(kdim, qdim, bias=qkv_bias) - - # TODO: eps should be 1 / 65530 if using fp16 - self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.out_proj = nn.Linear(qdim, qdim, bias=True) - - self.with_dca = with_decoupled_ca - if self.with_dca: - self.kv_proj_dca = nn.Linear(kdim, 2 * qdim, bias=qkv_bias) - self.k_norm_dca = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.dca_dim = decoupled_ca_dim - self.dca_weight = decoupled_ca_weight - - def forward(self, x, y): - """ - Parameters - ---------- - x: torch.Tensor - (batch, seqlen1, hidden_dim) (where hidden_dim = num heads * head dim) - y: torch.Tensor - (batch, seqlen2, hidden_dim2) - freqs_cis_img: torch.Tensor - (batch, hidden_dim // 2), RoPE for image - """ - b, s1, c = x.shape # [b, s1, D] - - if self.with_dca: - token_len = y.shape[1] - context_dca = y[:, -self.dca_dim:, :] - kv_dca = self.kv_proj_dca(context_dca).view(b, self.dca_dim, 2, self.num_heads, self.head_dim) - k_dca, v_dca = kv_dca.unbind(dim=2) # [b, s, h, d] - k_dca = self.k_norm_dca(k_dca) - y = y[:, :(token_len - self.dca_dim), :] - - _, s2, c = y.shape # [b, s2, 1024] - q = self.to_q(x) - k = self.to_k(y) - v = self.to_v(y) - - kv = torch.cat((k, v), dim=-1) - split_size = kv.shape[-1] // self.num_heads // 2 - kv = kv.view(1, -1, self.num_heads, split_size * 2) - k, v = torch.split(kv, split_size, dim=-1) - - q = q.view(b, s1, self.num_heads, self.head_dim) # [b, s1, h, d] - k = k.view(b, s2, self.num_heads, self.head_dim) # [b, s2, h, d] - v = v.view(b, s2, self.num_heads, self.head_dim) # [b, s2, h, d] - - q = self.q_norm(q) - k = self.k_norm(k) - - with torch.backends.cuda.sdp_kernel( - enable_flash=True, - enable_math=False, - enable_mem_efficient=True - ): - q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.num_heads), (q, k, v)) - context = F.scaled_dot_product_attention( - q, k, v - ).transpose(1, 2).reshape(b, s1, -1) - - if self.with_dca: - with torch.backends.cuda.sdp_kernel( - enable_flash=True, - enable_math=False, - enable_mem_efficient=True - ): - k_dca, v_dca = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.num_heads), - (k_dca, v_dca)) - context_dca = F.scaled_dot_product_attention( - q, k_dca, v_dca).transpose(1, 2).reshape(b, s1, -1) - - context = context + self.dca_weight * context_dca - - out = self.out_proj(context) # context.reshape - B, L1, -1 - - return out - - -class Attention(nn.Module): - """ - We rename some layer names to align with flash attention - """ - - def __init__( - self, - dim, - num_heads, - qkv_bias=True, - qk_norm=False, - norm_layer=nn.LayerNorm, - ): - super().__init__() - self.dim = dim - self.num_heads = num_heads - assert self.dim % num_heads == 0, 'dim should be divisible by num_heads' - self.head_dim = self.dim // num_heads - # This assertion is aligned with flash attention - assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" - self.scale = self.head_dim ** -0.5 - - self.to_q = nn.Linear(dim, dim, bias=qkv_bias) - self.to_k = nn.Linear(dim, dim, bias=qkv_bias) - self.to_v = nn.Linear(dim, dim, bias=qkv_bias) - # TODO: eps should be 1 / 65530 if using fp16 - self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() - self.out_proj = nn.Linear(dim, dim) - - def forward(self, x): - B, N, C = x.shape - - q = self.to_q(x) - k = self.to_k(x) - v = self.to_v(x) - - qkv = torch.cat((q, k, v), dim=-1) - split_size = qkv.shape[-1] // self.num_heads // 3 - qkv = qkv.view(1, -1, self.num_heads, split_size * 3) - q, k, v = torch.split(qkv, split_size, dim=-1) - - q = q.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2) # [b, h, s, d] - k = k.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2) # [b, h, s, d] - v = v.reshape(B, N, self.num_heads, self.head_dim).transpose(1, 2) - - q = self.q_norm(q) # [b, h, s, d] - k = self.k_norm(k) # [b, h, s, d] - - with torch.backends.cuda.sdp_kernel( - enable_flash=True, - enable_math=False, - enable_mem_efficient=True - ): - x = F.scaled_dot_product_attention(q, k, v) - x = x.transpose(1, 2).reshape(B, N, -1) - - x = self.out_proj(x) - return x - - -class HunYuanDiTBlock(nn.Module): - def __init__( - self, - hidden_size, - c_emb_size, - num_heads, - text_states_dim=1024, - use_flash_attn=False, - qk_norm=False, - norm_layer=nn.LayerNorm, - qk_norm_layer=nn.RMSNorm, - with_decoupled_ca=False, - decoupled_ca_dim=16, - decoupled_ca_weight=1.0, - init_scale=1.0, - qkv_bias=True, - skip_connection=True, - timested_modulate=False, - use_moe: bool = False, - num_experts: int = 8, - moe_top_k: int = 2, - **kwargs, - ): - super().__init__() - self.use_flash_attn = use_flash_attn - use_ele_affine = True - - # ========================= Self-Attention ========================= - self.norm1 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6) - self.attn1 = Attention(hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, qk_norm=qk_norm, - norm_layer=qk_norm_layer) - - # ========================= FFN ========================= - self.norm2 = norm_layer(hidden_size, elementwise_affine=use_ele_affine, eps=1e-6) - - # ========================= Add ========================= - # Simply use add like SDXL. - self.timested_modulate = timested_modulate - if self.timested_modulate: - self.default_modulation = nn.Sequential( - nn.SiLU(), - nn.Linear(c_emb_size, hidden_size, bias=True) - ) - - # ========================= Cross-Attention ========================= - self.attn2 = CrossAttention(hidden_size, text_states_dim, num_heads=num_heads, qkv_bias=qkv_bias, - qk_norm=qk_norm, norm_layer=qk_norm_layer, - with_decoupled_ca=with_decoupled_ca, decoupled_ca_dim=decoupled_ca_dim, - decoupled_ca_weight=decoupled_ca_weight, init_scale=init_scale, - ) - self.norm3 = norm_layer(hidden_size, elementwise_affine=True, eps=1e-6) - - if skip_connection: - self.skip_norm = norm_layer(hidden_size, elementwise_affine=True, eps=1e-6) - self.skip_linear = nn.Linear(2 * hidden_size, hidden_size) - else: - self.skip_linear = None - - self.use_moe = use_moe - if self.use_moe: - print("using moe") - self.moe = MoEBlock( - hidden_size, - num_experts=num_experts, - moe_top_k=moe_top_k, - dropout=0.0, - activation_fn="gelu", - final_dropout=False, - ff_inner_dim=int(hidden_size * 4.0), - ff_bias=True, - ) - else: - self.mlp = MLP(width=hidden_size) - - def forward(self, x, c=None, text_states=None, skip_value=None): - - if self.skip_linear is not None: - cat = torch.cat([skip_value, x], dim=-1) - x = self.skip_linear(cat) - x = self.skip_norm(x) - - # Self-Attention - if self.timested_modulate: - shift_msa = self.default_modulation(c).unsqueeze(dim=1) - x = x + shift_msa - - attn_out = self.attn1(self.norm1(x)) - - x = x + attn_out - - # Cross-Attention - x = x + self.attn2(self.norm2(x), text_states) - - # FFN Layer - mlp_inputs = self.norm3(x) - - if self.use_moe: - x = x + self.moe(mlp_inputs) - else: - x = x + self.mlp(mlp_inputs) - - return x - - -class AttentionPool(nn.Module): - def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None): - super().__init__() - self.positional_embedding = nn.Parameter(torch.randn(spacial_dim + 1, embed_dim) / embed_dim ** 0.5) - self.k_proj = nn.Linear(embed_dim, embed_dim) - self.q_proj = nn.Linear(embed_dim, embed_dim) - self.v_proj = nn.Linear(embed_dim, embed_dim) - self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim) - self.num_heads = num_heads - - def forward(self, x, attention_mask=None): - x = x.permute(1, 0, 2) # NLC -> LNC - if attention_mask is not None: - attention_mask = attention_mask.unsqueeze(-1).permute(1, 0, 2) - global_emb = (x * attention_mask).sum(dim=0) / attention_mask.sum(dim=0) - x = torch.cat([global_emb[None,], x], dim=0) - - else: - x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (L+1)NC - x = x + self.positional_embedding[:, None, :].to(x.dtype) # (L+1)NC - x, _ = F.multi_head_attention_forward( - query=x[:1], key=x, value=x, - embed_dim_to_check=x.shape[-1], - num_heads=self.num_heads, - q_proj_weight=self.q_proj.weight, - k_proj_weight=self.k_proj.weight, - v_proj_weight=self.v_proj.weight, - in_proj_weight=None, - in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]), - bias_k=None, - bias_v=None, - add_zero_attn=False, - dropout_p=0, - out_proj_weight=self.c_proj.weight, - out_proj_bias=self.c_proj.bias, - use_separate_proj_weight=True, - training=self.training, - need_weights=False - ) - return x.squeeze(0) - - -class FinalLayer(nn.Module): - """ - The final layer of HunYuanDiT. - """ - - def __init__(self, final_hidden_size, out_channels): - super().__init__() - self.final_hidden_size = final_hidden_size - self.norm_final = nn.LayerNorm(final_hidden_size, elementwise_affine=True, eps=1e-6) - self.linear = nn.Linear(final_hidden_size, out_channels, bias=True) - - def forward(self, x): - x = self.norm_final(x) - x = x[:, 1:] - x = self.linear(x) - return x - - -class HunYuanDiTPlain(nn.Module): - - def __init__( - self, - input_size=1024, - in_channels=4, - hidden_size=1024, - context_dim=1024, - depth=24, - num_heads=16, - mlp_ratio=4.0, - norm_type='layer', - qk_norm_type='rms', - qk_norm=False, - text_len=257, - with_decoupled_ca=False, - additional_cond_hidden_state=768, - decoupled_ca_dim=16, - decoupled_ca_weight=1.0, - use_pos_emb=False, - use_attention_pooling=True, - guidance_cond_proj_dim=None, - qkv_bias=True, - num_moe_layers: int = 6, - num_experts: int = 8, - moe_top_k: int = 2, - **kwargs - ): - super().__init__() - self.input_size = input_size - self.depth = depth - self.in_channels = in_channels - self.out_channels = in_channels - self.num_heads = num_heads - - self.hidden_size = hidden_size - self.norm = nn.LayerNorm if norm_type == 'layer' else nn.RMSNorm - self.qk_norm = nn.RMSNorm if qk_norm_type == 'rms' else nn.LayerNorm - self.context_dim = context_dim - - self.with_decoupled_ca = with_decoupled_ca - self.decoupled_ca_dim = decoupled_ca_dim - self.decoupled_ca_weight = decoupled_ca_weight - self.use_pos_emb = use_pos_emb - self.use_attention_pooling = use_attention_pooling - self.guidance_cond_proj_dim = guidance_cond_proj_dim - - self.text_len = text_len - - self.x_embedder = nn.Linear(in_channels, hidden_size, bias=True) - self.t_embedder = TimestepEmbedder(hidden_size, hidden_size * 4, cond_proj_dim=guidance_cond_proj_dim) - - # Will use fixed sin-cos embedding: - if self.use_pos_emb: - self.register_buffer("pos_embed", torch.zeros(1, input_size, hidden_size)) - pos = np.arange(self.input_size, dtype=np.float32) - pos_embed = get_1d_sincos_pos_embed_from_grid(self.pos_embed.shape[-1], pos) - self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0)) - - self.use_attention_pooling = use_attention_pooling - if use_attention_pooling: - self.pooler = AttentionPool(self.text_len, context_dim, num_heads=8, output_dim=1024) - self.extra_embedder = nn.Sequential( - nn.Linear(1024, hidden_size * 4), - nn.SiLU(), - nn.Linear(hidden_size * 4, hidden_size, bias=True), - ) - - if with_decoupled_ca: - self.additional_cond_hidden_state = additional_cond_hidden_state - self.additional_cond_proj = nn.Sequential( - nn.Linear(additional_cond_hidden_state, hidden_size * 4), - nn.SiLU(), - nn.Linear(hidden_size * 4, 1024, bias=True), - ) - - # HUnYuanDiT Blocks - self.blocks = nn.ModuleList([ - HunYuanDiTBlock(hidden_size=hidden_size, - c_emb_size=hidden_size, - num_heads=num_heads, - mlp_ratio=mlp_ratio, - text_states_dim=context_dim, - qk_norm=qk_norm, - norm_layer=self.norm, - qk_norm_layer=self.qk_norm, - skip_connection=layer > depth // 2, - with_decoupled_ca=with_decoupled_ca, - decoupled_ca_dim=decoupled_ca_dim, - decoupled_ca_weight=decoupled_ca_weight, - qkv_bias=qkv_bias, - use_moe=True if depth - layer <= num_moe_layers else False, - num_experts=num_experts, - moe_top_k=moe_top_k - ) - for layer in range(depth) - ]) - self.depth = depth - - self.final_layer = FinalLayer(hidden_size, self.out_channels) - - def forward(self, x, t, contexts, **kwargs): - cond = contexts['main'] - - t = self.t_embedder(t, condition=kwargs.get('guidance_cond')) - x = self.x_embedder(x) - - if self.use_pos_emb: - pos_embed = self.pos_embed.to(x.dtype) - x = x + pos_embed - - if self.use_attention_pooling: - extra_vec = self.pooler(cond, None) - c = t + self.extra_embedder(extra_vec) # [B, D] - else: - c = t - - if self.with_decoupled_ca: - additional_cond = self.additional_cond_proj(contexts['additional']) - cond = torch.cat([cond, additional_cond], dim=1) - - x = torch.cat([c, x], dim=1) - - skip_value_list = [] - for layer, block in enumerate(self.blocks): - skip_value = None if layer <= self.depth // 2 else skip_value_list.pop() - x = block(x, c, cond, skip_value=skip_value) - if layer < self.depth // 2: - skip_value_list.append(x) - - x = self.final_layer(x) - return x diff --git a/hy3dshape/hy3dshape/models/denoisers/moe_layers.py b/hy3dshape/hy3dshape/models/denoisers/moe_layers.py deleted file mode 100644 index dbcf210b0e43d6bf5ab957dd73c59e39561bab99..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/denoisers/moe_layers.py +++ /dev/null @@ -1,177 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import torch -import torch.nn as nn -import numpy as np -import math -from timm.models.vision_transformer import PatchEmbed, Attention, Mlp - -import torch.nn.functional as F -from diffusers.models.attention import FeedForward - -class AddAuxiliaryLoss(torch.autograd.Function): - """ - The trick function of adding auxiliary (aux) loss, - which includes the gradient of the aux loss during backpropagation. - """ - @staticmethod - def forward(ctx, x, loss): - assert loss.numel() == 1 - ctx.dtype = loss.dtype - ctx.required_aux_loss = loss.requires_grad - return x - - @staticmethod - def backward(ctx, grad_output): - grad_loss = None - if ctx.required_aux_loss: - grad_loss = torch.ones(1, dtype=ctx.dtype, device=grad_output.device) - return grad_output, grad_loss - -class MoEGate(nn.Module): - def __init__(self, embed_dim, num_experts=16, num_experts_per_tok=2, aux_loss_alpha=0.01): - super().__init__() - self.top_k = num_experts_per_tok - self.n_routed_experts = num_experts - - self.scoring_func = 'softmax' - self.alpha = aux_loss_alpha - self.seq_aux = False - - # topk selection algorithm - self.norm_topk_prob = False - self.gating_dim = embed_dim - self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim))) - self.reset_parameters() - - def reset_parameters(self) -> None: - import torch.nn.init as init - init.kaiming_uniform_(self.weight, a=math.sqrt(5)) - - def forward(self, hidden_states): - bsz, seq_len, h = hidden_states.shape - # print(bsz, seq_len, h) - ### compute gating score - hidden_states = hidden_states.view(-1, h) - logits = F.linear(hidden_states, self.weight, None) - if self.scoring_func == 'softmax': - scores = logits.softmax(dim=-1) - else: - raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}') - - ### select top-k experts - topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False) - - ### norm gate to sum 1 - if self.top_k > 1 and self.norm_topk_prob: - denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20 - topk_weight = topk_weight / denominator - - ### expert-level computation auxiliary loss - if self.training and self.alpha > 0.0: - scores_for_aux = scores - aux_topk = self.top_k - # always compute aux loss based on the naive greedy topk method - topk_idx_for_aux_loss = topk_idx.view(bsz, -1) - if self.seq_aux: - scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1) - ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device) - ce.scatter_add_( - 1, - topk_idx_for_aux_loss, - torch.ones( - bsz, seq_len * aux_topk, - device=hidden_states.device - ) - ).div_(seq_len * aux_topk / self.n_routed_experts) - aux_loss = (ce * scores_for_seq_aux.mean(dim = 1)).sum(dim = 1).mean() - aux_loss = aux_loss * self.alpha - else: - mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), - num_classes=self.n_routed_experts) - ce = mask_ce.float().mean(0) - Pi = scores_for_aux.mean(0) - fi = ce * self.n_routed_experts - aux_loss = (Pi * fi).sum() * self.alpha - else: - aux_loss = None - return topk_idx, topk_weight, aux_loss - -class MoEBlock(nn.Module): - def __init__(self, dim, num_experts=8, moe_top_k=2, - activation_fn = "gelu", dropout=0.0, final_dropout = False, - ff_inner_dim = None, ff_bias = True): - super().__init__() - self.moe_top_k = moe_top_k - self.experts = nn.ModuleList([ - FeedForward(dim,dropout=dropout, - activation_fn=activation_fn, - final_dropout=final_dropout, - inner_dim=ff_inner_dim, - bias=ff_bias) - for i in range(num_experts)]) - self.gate = MoEGate(embed_dim=dim, num_experts=num_experts, num_experts_per_tok=moe_top_k) - - self.shared_experts = FeedForward(dim,dropout=dropout, activation_fn=activation_fn, - final_dropout=final_dropout, inner_dim=ff_inner_dim, - bias=ff_bias) - - def initialize_weight(self): - pass - - def forward(self, hidden_states): - identity = hidden_states - orig_shape = hidden_states.shape - topk_idx, topk_weight, aux_loss = self.gate(hidden_states) - - hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) - flat_topk_idx = topk_idx.view(-1) - if self.training: - hidden_states = hidden_states.repeat_interleave(self.moe_top_k, dim=0) - y = torch.empty_like(hidden_states, dtype=hidden_states.dtype) - for i, expert in enumerate(self.experts): - tmp = expert(hidden_states[flat_topk_idx == i]) - y[flat_topk_idx == i] = tmp.to(hidden_states.dtype) - y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1) - y = y.view(*orig_shape) - y = AddAuxiliaryLoss.apply(y, aux_loss) - else: - y = self.moe_infer(hidden_states, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape) - y = y + self.shared_experts(identity) - return y - - - @torch.no_grad() - def moe_infer(self, x, flat_expert_indices, flat_expert_weights): - expert_cache = torch.zeros_like(x) - idxs = flat_expert_indices.argsort() - tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0) - token_idxs = idxs // self.moe_top_k - for i, end_idx in enumerate(tokens_per_expert): - start_idx = 0 if i == 0 else tokens_per_expert[i-1] - if start_idx == end_idx: - continue - expert = self.experts[i] - exp_token_idx = token_idxs[start_idx:end_idx] - expert_tokens = x[exp_token_idx] - expert_out = expert(expert_tokens) - expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]]) - - # for fp16 and other dtype - expert_cache = expert_cache.to(expert_out.dtype) - expert_cache.scatter_reduce_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), - expert_out, - reduce='sum') - return expert_cache diff --git a/hy3dshape/hy3dshape/models/diffusion/flow_matching_sit.py b/hy3dshape/hy3dshape/models/diffusion/flow_matching_sit.py deleted file mode 100644 index 9813b74235e8245f35f5479adc4d75abba5a80de..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/flow_matching_sit.py +++ /dev/null @@ -1,354 +0,0 @@ -import os -from contextlib import contextmanager -from typing import List, Tuple, Optional, Union - -import torch -import torch.nn as nn -from torch.optim import lr_scheduler -import pytorch_lightning as pl -from pytorch_lightning.utilities import rank_zero_info -from pytorch_lightning.utilities import rank_zero_only - -from ...utils.ema import LitEma -from ...utils.misc import instantiate_from_config, instantiate_non_trainable_model - - - -class Diffuser(pl.LightningModule): - def __init__( - self, - *, - first_stage_config, - cond_stage_config, - denoiser_cfg, - scheduler_cfg, - optimizer_cfg, - pipeline_cfg=None, - image_processor_cfg=None, - lora_config=None, - ema_config=None, - first_stage_key: str = "surface", - cond_stage_key: str = "image", - scale_by_std: bool = False, - z_scale_factor: float = 1.0, - ckpt_path: Optional[str] = None, - ignore_keys: Union[Tuple[str], List[str]] = (), - torch_compile: bool = False, - ): - super().__init__() - self.first_stage_key = first_stage_key - self.cond_stage_key = cond_stage_key - - # ========= init optimizer config ========= # - self.optimizer_cfg = optimizer_cfg - - # ========= init diffusion scheduler ========= # - self.scheduler_cfg = scheduler_cfg - self.sampler = None - if 'transport' in scheduler_cfg: - self.transport = instantiate_from_config(scheduler_cfg.transport) - self.sampler = instantiate_from_config(scheduler_cfg.sampler, transport=self.transport) - self.sample_fn = self.sampler.sample_ode(**scheduler_cfg.sampler.ode_params) - - # ========= init the model ========= # - self.denoiser_cfg = denoiser_cfg - self.model = instantiate_from_config(denoiser_cfg, device=None, dtype=None) - self.cond_stage_model = instantiate_from_config(cond_stage_config) - - self.ckpt_path = ckpt_path - if ckpt_path is not None: - self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) - - # ========= config lora model ========= # - if lora_config is not None: - from peft import LoraConfig, get_peft_model - loraconfig = LoraConfig( - r=lora_config.rank, - lora_alpha=lora_config.rank, - target_modules=lora_config.get('target_modules') - ) - self.model = get_peft_model(self.model, loraconfig) - - # ========= config ema model ========= # - self.ema_config = ema_config - if self.ema_config is not None: - if self.ema_config.ema_model == 'DSEma': - # from michelangelo.models.modules.ema_deepspeed import DSEma - from ..utils.ema_deepspeed import DSEma - self.model_ema = DSEma(self.model, decay=self.ema_config.ema_decay) - else: - self.model_ema = LitEma(self.model, decay=self.ema_config.ema_decay) - #do not initilize EMA weight from ckpt path, since I need to change moe layers - if ckpt_path is not None: - self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) - print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") - - # ========= init vae at last to prevent it is overridden by loaded ckpt ========= # - self.first_stage_model = instantiate_non_trainable_model(first_stage_config) - - self.scale_by_std = scale_by_std - if scale_by_std: - self.register_buffer("z_scale_factor", torch.tensor(z_scale_factor)) - else: - self.z_scale_factor = z_scale_factor - - # ========= init pipeline for inference ========= # - self.image_processor_cfg = image_processor_cfg - self.image_processor = None - if self.image_processor_cfg is not None: - self.image_processor = instantiate_from_config(self.image_processor_cfg) - self.pipeline_cfg = pipeline_cfg - from ...schedulers import FlowMatchEulerDiscreteScheduler - scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000) - self.pipeline = instantiate_from_config( - pipeline_cfg, - vae=self.first_stage_model, - model=self.model, - scheduler=scheduler, # self.sampler, - conditioner=self.cond_stage_model, - image_processor=self.image_processor, - ) - - # ========= torch compile to accelerate ========= # - self.torch_compile = torch_compile - if self.torch_compile: - torch.nn.Module.compile(self.model) - torch.nn.Module.compile(self.first_stage_model) - torch.nn.Module.compile(self.cond_stage_model) - print(f'*' * 100) - print(f'Compile model for acceleration') - print(f'*' * 100) - - @contextmanager - def ema_scope(self, context=None): - if self.ema_config is not None and self.ema_config.get('ema_inference', False): - self.model_ema.store(self.model) - self.model_ema.copy_to(self.model) - if context is not None: - print(f"{context}: Switched to EMA weights") - try: - yield None - finally: - if self.ema_config is not None and self.ema_config.get('ema_inference', False): - self.model_ema.restore(self.model) - if context is not None: - print(f"{context}: Restored training weights") - - def init_from_ckpt(self, path, ignore_keys=()): - ckpt = torch.load(path, map_location="cpu") - if 'state_dict' not in ckpt: - # deepspeed ckpt - state_dict = {} - for k in ckpt.keys(): - new_k = k.replace('_forward_module.', '') - state_dict[new_k] = ckpt[k] - else: - state_dict = ckpt["state_dict"] - - keys = list(state_dict.keys()) - for k in keys: - for ik in ignore_keys: - if ik in k: - print("Deleting key {} from state_dict.".format(k)) - del state_dict[k] - - missing, unexpected = self.load_state_dict(state_dict, strict=False) - print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") - if len(missing) > 0: - print(f"Missing Keys: {missing}") - print(f"Unexpected Keys: {unexpected}") - - def on_load_checkpoint(self, checkpoint): - """ - The pt_model is trained separately, so we already have access to its - checkpoint and load it separately with `self.set_pt_model`. - - However, the PL Trainer is strict about - checkpoint loading (not configurable), so it expects the loaded state_dict - to match exactly the keys in the model state_dict. - - So, when loading the checkpoint, before matching keys, we add all pt_model keys - from self.state_dict() to the checkpoint state dict, so that they match - """ - for key in self.state_dict().keys(): - if key.startswith("model_ema") and key not in checkpoint["state_dict"]: - checkpoint["state_dict"][key] = self.state_dict()[key] - - def configure_optimizers(self) -> Tuple[List, List]: - lr = self.learning_rate - - params_list = [] - trainable_parameters = list(self.model.parameters()) - params_list.append({'params': trainable_parameters, 'lr': lr}) - - no_decay = ['bias', 'norm.weight', 'norm.bias', 'norm1.weight', 'norm1.bias', 'norm2.weight', 'norm2.bias'] - - - if self.optimizer_cfg.get('train_image_encoder', False): - image_encoder_parameters = list(self.cond_stage_model.named_parameters()) - image_encoder_parameters_decay = [param for name, param in image_encoder_parameters if - not any((no_decay_name in name) for no_decay_name in no_decay)] - image_encoder_parameters_nodecay = [param for name, param in image_encoder_parameters if - any((no_decay_name in name) for no_decay_name in no_decay)] - # filter trainable params - image_encoder_parameters_decay = [param for param in image_encoder_parameters_decay if - param.requires_grad] - image_encoder_parameters_nodecay = [param for param in image_encoder_parameters_nodecay if - param.requires_grad] - - print(f"Image Encoder Params: {len(image_encoder_parameters_decay)} decay, ") - print(f"Image Encoder Params: {len(image_encoder_parameters_nodecay)} nodecay, ") - - image_encoder_lr = self.optimizer_cfg['image_encoder_lr'] - image_encoder_lr_multiply = self.optimizer_cfg.get('image_encoder_lr_multiply', 1.0) - image_encoder_lr = image_encoder_lr if image_encoder_lr is not None else lr * image_encoder_lr_multiply - params_list.append( - {'params': image_encoder_parameters_decay, 'lr': image_encoder_lr, - 'weight_decay': 0.05}) - params_list.append( - {'params': image_encoder_parameters_nodecay, 'lr': image_encoder_lr, - 'weight_decay': 0.}) - - optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=params_list, lr=lr) - if hasattr(self.optimizer_cfg, 'scheduler'): - scheduler_func = instantiate_from_config( - self.optimizer_cfg.scheduler, - max_decay_steps=self.trainer.max_steps, - lr_max=lr - ) - scheduler = { - "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule), - "interval": "step", - "frequency": 1 - } - schedulers = [scheduler] - else: - schedulers = [] - optimizers = [optimizer] - - return optimizers, schedulers - - @rank_zero_only - @torch.no_grad() - def on_train_batch_start(self, batch, batch_idx): - # only for very first batch - if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 \ - and batch_idx == 0 and self.ckpt_path is None: - # set rescale weight to 1./std of encodings - print("### USING STD-RESCALING ###") - - z_q = self.encode_first_stage(batch[self.first_stage_key]) - z = z_q.detach() - - del self.z_scale_factor - self.register_buffer("z_scale_factor", 1. / z.flatten().std()) - print(f"setting self.z_scale_factor to {self.z_scale_factor}") - - print("### USING STD-RESCALING ###") - - def on_train_batch_end(self, *args, **kwargs): - if self.ema_config is not None: - self.model_ema(self.model) - - def on_train_epoch_start(self) -> None: - pl.seed_everything(self.trainer.global_rank) - - def forward(self, batch): - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): #float32 for text - contexts = self.cond_stage_model(image=batch.get('image'), text=batch.get('text'), mask=batch.get('mask')) - # t5_text = contexts['t5_text']['prompt_embeds'] - # nan_count = torch.isnan(t5_text).sum() - # if nan_count > 0: - # print("t5_text has %d NaN values"%(nan_count)) - with torch.autocast(device_type="cuda", dtype=torch.float16): - with torch.no_grad(): - latents = self.first_stage_model.encode(batch[self.first_stage_key], sample_posterior=True) - latents = self.z_scale_factor * latents - # print(latents.shape) - - # check vae encode and decode is ok? answer is ok ! - # import time - # from hy3dshape.pipelines import export_to_trimesh - # latents = 1. / self.z_scale_factor * latents - # latents = self.first_stage_model(latents) - # outputs = self.first_stage_model.latents2mesh( - # latents, - # bounds=1.01, - # mc_level=0.0, - # num_chunks=20000, - # octree_resolution=256, - # mc_algo='mc', - # enable_pbar=True - # ) - # mesh = export_to_trimesh(outputs) - # if isinstance(mesh, list): - # for midx, m in enumerate(mesh): - # m.export(f"check_{midx}_{time.time()}.glb") - # else: - # mesh.export(f"check_{time.time()}.glb") - - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - loss = self.transport.training_losses(self.model, latents, dict(contexts=contexts))["loss"].mean() - return loss - - def training_step(self, batch, batch_idx, optimizer_idx=0): - loss = self.forward(batch) - split = 'train' - loss_dict = { - f"{split}/simple": loss.detach(), - f"{split}/total_loss": loss.detach(), - f"{split}/lr_abs": self.optimizers().param_groups[0]['lr'], - } - self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True) - - return loss - - def validation_step(self, batch, batch_idx, optimizer_idx=0): - loss = self.forward(batch) - split = 'val' - loss_dict = { - f"{split}/simple": loss.detach(), - f"{split}/total_loss": loss.detach(), - f"{split}/lr_abs": self.optimizers().param_groups[0]['lr'], - } - self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True) - - return loss - - @torch.no_grad() - def sample(self, batch, output_type='trimesh', **kwargs): - self.cond_stage_model.disable_drop = True - - generator = torch.Generator().manual_seed(0) - - with self.ema_scope("Sample"): - with torch.amp.autocast(device_type='cuda'): - try: - self.pipeline.device = self.device - self.pipeline.dtype = self.dtype - print("### USING PIPELINE ###") - print(f'device: {self.device} dtype : {self.dtype}') - additional_params = {'output_type':output_type} - - image = batch.get("image", None) - mask = batch.get('mask', None) - - # if not isinstance(image, torch.Tensor): print(image.shape) - # if isinstance(mask, torch.Tensor): print(mask.shape) - - outputs = self.pipeline(image=image, - mask=mask, - generator=generator, - **additional_params) - - except Exception as e: - import traceback - traceback.print_exc() - print(f"Unexpected {e=}, {type(e)=}") - with open("error.txt", "a") as f: - f.write(str(e)) - f.write(traceback.format_exc()) - f.write("\n") - outputs = [None] - self.cond_stage_model.disable_drop = False - return [outputs] diff --git a/hy3dshape/hy3dshape/models/diffusion/transport/__init__.py b/hy3dshape/hy3dshape/models/diffusion/transport/__init__.py deleted file mode 100644 index 71164eaf5c447f9fc7bf3bc6ab4e8d6e0c67a05c..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/transport/__init__.py +++ /dev/null @@ -1,97 +0,0 @@ -# This file includes code derived from the SiT project (https://github.com/willisma/SiT), -# which is licensed under the MIT License. -# -# MIT License -# -# Copyright (c) Meta Platforms, Inc. and affiliates. -# -# Permission is hereby granted, free of charge, to any person obtaining a copy -# of this software and associated documentation files (the "Software"), to deal -# in the Software without restriction, including without limitation the rights -# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -# copies of the Software, and to permit persons to whom the Software is -# furnished to do so, subject to the following conditions: -# -# The above copyright notice and this permission notice shall be included in all -# copies or substantial portions of the Software. -# -# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -# SOFTWARE. - -from .transport import Transport, ModelType, WeightType, PathType, Sampler - - -def create_transport( - path_type='Linear', - prediction="velocity", - loss_weight=None, - train_eps=None, - sample_eps=None, - train_sample_type="uniform", - mean = 0.0, - std = 1.0, - shift_scale = 1.0, -): - """function for creating Transport object - **Note**: model prediction defaults to velocity - Args: - - path_type: type of path to use; default to linear - - learn_score: set model prediction to score - - learn_noise: set model prediction to noise - - velocity_weighted: weight loss by velocity weight - - likelihood_weighted: weight loss by likelihood weight - - train_eps: small epsilon for avoiding instability during training - - sample_eps: small epsilon for avoiding instability during sampling - """ - - if prediction == "noise": - model_type = ModelType.NOISE - elif prediction == "score": - model_type = ModelType.SCORE - else: - model_type = ModelType.VELOCITY - - if loss_weight == "velocity": - loss_type = WeightType.VELOCITY - elif loss_weight == "likelihood": - loss_type = WeightType.LIKELIHOOD - else: - loss_type = WeightType.NONE - - path_choice = { - "Linear": PathType.LINEAR, - "GVP": PathType.GVP, - "VP": PathType.VP, - } - - path_type = path_choice[path_type] - - if (path_type in [PathType.VP]): - train_eps = 1e-5 if train_eps is None else train_eps - sample_eps = 1e-3 if train_eps is None else sample_eps - elif (path_type in [PathType.GVP, PathType.LINEAR] and model_type != ModelType.VELOCITY): - train_eps = 1e-3 if train_eps is None else train_eps - sample_eps = 1e-3 if train_eps is None else sample_eps - else: # velocity & [GVP, LINEAR] is stable everywhere - train_eps = 0 - sample_eps = 0 - - # create flow state - state = Transport( - model_type=model_type, - path_type=path_type, - loss_type=loss_type, - train_eps=train_eps, - sample_eps=sample_eps, - train_sample_type=train_sample_type, - mean=mean, - std=std, - shift_scale =shift_scale, - ) - - return state diff --git a/hy3dshape/hy3dshape/models/diffusion/transport/integrators.py b/hy3dshape/hy3dshape/models/diffusion/transport/integrators.py deleted file mode 100644 index e2e011492ce8d7acc296df9bfe57d45ae01893e7..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/transport/integrators.py +++ /dev/null @@ -1,142 +0,0 @@ -# This file includes code derived from the SiT project (https://github.com/willisma/SiT), -# which is licensed under the MIT License. -# -# MIT License -# -# Copyright (c) Meta Platforms, Inc. and affiliates. -# -# Permission is hereby granted, free of charge, to any person obtaining a copy -# of this software and associated documentation files (the "Software"), to deal -# in the Software without restriction, including without limitation the rights -# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -# copies of the Software, and to permit persons to whom the Software is -# furnished to do so, subject to the following conditions: -# -# The above copyright notice and this permission notice shall be included in all -# copies or substantial portions of the Software. -# -# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -# SOFTWARE. - -import numpy as np -import torch as th -import torch.nn as nn -from torchdiffeq import odeint -from functools import partial -from tqdm import tqdm - -class sde: - """SDE solver class""" - def __init__( - self, - drift, - diffusion, - *, - t0, - t1, - num_steps, - sampler_type, - ): - assert t0 < t1, "SDE sampler has to be in forward time" - - self.num_timesteps = num_steps - self.t = th.linspace(t0, t1, num_steps) - self.dt = self.t[1] - self.t[0] - self.drift = drift - self.diffusion = diffusion - self.sampler_type = sampler_type - - def __Euler_Maruyama_step(self, x, mean_x, t, model, **model_kwargs): - w_cur = th.randn(x.size()).to(x) - t = th.ones(x.size(0)).to(x) * t - dw = w_cur * th.sqrt(self.dt) - drift = self.drift(x, t, model, **model_kwargs) - diffusion = self.diffusion(x, t) - mean_x = x + drift * self.dt - x = mean_x + th.sqrt(2 * diffusion) * dw - return x, mean_x - - def __Heun_step(self, x, _, t, model, **model_kwargs): - w_cur = th.randn(x.size()).to(x) - dw = w_cur * th.sqrt(self.dt) - t_cur = th.ones(x.size(0)).to(x) * t - diffusion = self.diffusion(x, t_cur) - xhat = x + th.sqrt(2 * diffusion) * dw - K1 = self.drift(xhat, t_cur, model, **model_kwargs) - xp = xhat + self.dt * K1 - K2 = self.drift(xp, t_cur + self.dt, model, **model_kwargs) - return xhat + 0.5 * self.dt * (K1 + K2), xhat # at last time point we do not perform the heun step - - def __forward_fn(self): - """TODO: generalize here by adding all private functions ending with steps to it""" - sampler_dict = { - "Euler": self.__Euler_Maruyama_step, - "Heun": self.__Heun_step, - } - - try: - sampler = sampler_dict[self.sampler_type] - except: - raise NotImplementedError("Smapler type not implemented.") - - return sampler - - def sample(self, init, model, **model_kwargs): - """forward loop of sde""" - x = init - mean_x = init - samples = [] - sampler = self.__forward_fn() - for ti in self.t[:-1]: - with th.no_grad(): - x, mean_x = sampler(x, mean_x, ti, model, **model_kwargs) - samples.append(x) - - return samples - -class ode: - """ODE solver class""" - def __init__( - self, - drift, - *, - t0, - t1, - sampler_type, - num_steps, - atol, - rtol, - ): - assert t0 < t1, "ODE sampler has to be in forward time" - - self.drift = drift - self.t = th.linspace(t0, t1, num_steps) - self.atol = atol - self.rtol = rtol - self.sampler_type = sampler_type - - def sample(self, x, model, **model_kwargs): - - device = x[0].device if isinstance(x, tuple) else x.device - def _fn(t, x): - t = th.ones(x[0].size(0)).to(device) * t if isinstance(x, tuple) else th.ones(x.size(0)).to(device) * t - model_output = self.drift(x, t, model, **model_kwargs) - return model_output - - t = self.t.to(device) - atol = [self.atol] * len(x) if isinstance(x, tuple) else [self.atol] - rtol = [self.rtol] * len(x) if isinstance(x, tuple) else [self.rtol] - samples = odeint( - _fn, - x, - t, - method=self.sampler_type, - atol=atol, - rtol=rtol - ) - return samples diff --git a/hy3dshape/hy3dshape/models/diffusion/transport/path.py b/hy3dshape/hy3dshape/models/diffusion/transport/path.py deleted file mode 100644 index 9c867929e9c570a5465c6bfd337cce27b0447041..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/transport/path.py +++ /dev/null @@ -1,220 +0,0 @@ -# This file includes code derived from the SiT project (https://github.com/willisma/SiT), -# which is licensed under the MIT License. -# -# MIT License -# -# Copyright (c) Meta Platforms, Inc. and affiliates. -# -# Permission is hereby granted, free of charge, to any person obtaining a copy -# of this software and associated documentation files (the "Software"), to deal -# in the Software without restriction, including without limitation the rights -# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -# copies of the Software, and to permit persons to whom the Software is -# furnished to do so, subject to the following conditions: -# -# The above copyright notice and this permission notice shall be included in all -# copies or substantial portions of the Software. -# -# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -# SOFTWARE. - -import torch as th -import numpy as np -from functools import partial - -def expand_t_like_x(t, x): - """Function to reshape time t to broadcastable dimension of x - Args: - t: [batch_dim,], time vector - x: [batch_dim,...], data point - """ - dims = [1] * (len(x.size()) - 1) - t = t.view(t.size(0), *dims) - return t - - -#################### Coupling Plans #################### - -class ICPlan: - """Linear Coupling Plan""" - def __init__(self, sigma=0.0): - self.sigma = sigma - - def compute_alpha_t(self, t): - """Compute the data coefficient along the path""" - return t, 1 - - def compute_sigma_t(self, t): - """Compute the noise coefficient along the path""" - return 1 - t, -1 - - def compute_d_alpha_alpha_ratio_t(self, t): - """Compute the ratio between d_alpha and alpha""" - return 1 / t - - def compute_drift(self, x, t): - """We always output sde according to score parametrization; """ - t = expand_t_like_x(t, x) - alpha_ratio = self.compute_d_alpha_alpha_ratio_t(t) - sigma_t, d_sigma_t = self.compute_sigma_t(t) - drift = alpha_ratio * x - diffusion = alpha_ratio * (sigma_t ** 2) - sigma_t * d_sigma_t - - return -drift, diffusion - - def compute_diffusion(self, x, t, form="constant", norm=1.0): - """Compute the diffusion term of the SDE - Args: - x: [batch_dim, ...], data point - t: [batch_dim,], time vector - form: str, form of the diffusion term - norm: float, norm of the diffusion term - """ - t = expand_t_like_x(t, x) - choices = { - "constant": norm, - "SBDM": norm * self.compute_drift(x, t)[1], - "sigma": norm * self.compute_sigma_t(t)[0], - "linear": norm * (1 - t), - "decreasing": 0.25 * (norm * th.cos(np.pi * t) + 1) ** 2, - "inccreasing-decreasing": norm * th.sin(np.pi * t) ** 2, - } - - try: - diffusion = choices[form] - except KeyError: - raise NotImplementedError(f"Diffusion form {form} not implemented") - - return diffusion - - def get_score_from_velocity(self, velocity, x, t): - """Wrapper function: transfrom velocity prediction model to score - Args: - velocity: [batch_dim, ...] shaped tensor; velocity model output - x: [batch_dim, ...] shaped tensor; x_t data point - t: [batch_dim,] time tensor - """ - t = expand_t_like_x(t, x) - alpha_t, d_alpha_t = self.compute_alpha_t(t) - sigma_t, d_sigma_t = self.compute_sigma_t(t) - mean = x - reverse_alpha_ratio = alpha_t / d_alpha_t - var = sigma_t**2 - reverse_alpha_ratio * d_sigma_t * sigma_t - score = (reverse_alpha_ratio * velocity - mean) / var - return score - - def get_noise_from_velocity(self, velocity, x, t): - """Wrapper function: transfrom velocity prediction model to denoiser - Args: - velocity: [batch_dim, ...] shaped tensor; velocity model output - x: [batch_dim, ...] shaped tensor; x_t data point - t: [batch_dim,] time tensor - """ - t = expand_t_like_x(t, x) - alpha_t, d_alpha_t = self.compute_alpha_t(t) - sigma_t, d_sigma_t = self.compute_sigma_t(t) - mean = x - reverse_alpha_ratio = alpha_t / d_alpha_t - var = reverse_alpha_ratio * d_sigma_t - sigma_t - noise = (reverse_alpha_ratio * velocity - mean) / var - return noise - - def get_velocity_from_score(self, score, x, t): - """Wrapper function: transfrom score prediction model to velocity - Args: - score: [batch_dim, ...] shaped tensor; score model output - x: [batch_dim, ...] shaped tensor; x_t data point - t: [batch_dim,] time tensor - """ - t = expand_t_like_x(t, x) - drift, var = self.compute_drift(x, t) - velocity = var * score - drift - return velocity - - def compute_mu_t(self, t, x0, x1): - """Compute the mean of time-dependent density p_t""" - t = expand_t_like_x(t, x1) - alpha_t, _ = self.compute_alpha_t(t) - sigma_t, _ = self.compute_sigma_t(t) - # t*x1 + (1-t)*x0 ; t=0 x0; t=1 x1 - return alpha_t * x1 + sigma_t * x0 - - def compute_xt(self, t, x0, x1): - """Sample xt from time-dependent density p_t; rng is required""" - xt = self.compute_mu_t(t, x0, x1) - return xt - - def compute_ut(self, t, x0, x1, xt): - """Compute the vector field corresponding to p_t""" - t = expand_t_like_x(t, x1) - _, d_alpha_t = self.compute_alpha_t(t) - _, d_sigma_t = self.compute_sigma_t(t) - return d_alpha_t * x1 + d_sigma_t * x0 - - def plan(self, t, x0, x1): - xt = self.compute_xt(t, x0, x1) - ut = self.compute_ut(t, x0, x1, xt) - return t, xt, ut - - -class VPCPlan(ICPlan): - """class for VP path flow matching""" - - def __init__(self, sigma_min=0.1, sigma_max=20.0): - self.sigma_min = sigma_min - self.sigma_max = sigma_max - self.log_mean_coeff = lambda t: -0.25 * ((1 - t) ** 2) * \ - (self.sigma_max - self.sigma_min) - 0.5 * (1 - t) * self.sigma_min - self.d_log_mean_coeff = lambda t: 0.5 * (1 - t) * \ - (self.sigma_max - self.sigma_min) + 0.5 * self.sigma_min - - - def compute_alpha_t(self, t): - """Compute coefficient of x1""" - alpha_t = self.log_mean_coeff(t) - alpha_t = th.exp(alpha_t) - d_alpha_t = alpha_t * self.d_log_mean_coeff(t) - return alpha_t, d_alpha_t - - def compute_sigma_t(self, t): - """Compute coefficient of x0""" - p_sigma_t = 2 * self.log_mean_coeff(t) - sigma_t = th.sqrt(1 - th.exp(p_sigma_t)) - d_sigma_t = th.exp(p_sigma_t) * (2 * self.d_log_mean_coeff(t)) / (-2 * sigma_t) - return sigma_t, d_sigma_t - - def compute_d_alpha_alpha_ratio_t(self, t): - """Special purposed function for computing numerical stabled d_alpha_t / alpha_t""" - return self.d_log_mean_coeff(t) - - def compute_drift(self, x, t): - """Compute the drift term of the SDE""" - t = expand_t_like_x(t, x) - beta_t = self.sigma_min + (1 - t) * (self.sigma_max - self.sigma_min) - return -0.5 * beta_t * x, beta_t / 2 - - -class GVPCPlan(ICPlan): - def __init__(self, sigma=0.0): - super().__init__(sigma) - - def compute_alpha_t(self, t): - """Compute coefficient of x1""" - alpha_t = th.sin(t * np.pi / 2) - d_alpha_t = np.pi / 2 * th.cos(t * np.pi / 2) - return alpha_t, d_alpha_t - - def compute_sigma_t(self, t): - """Compute coefficient of x0""" - sigma_t = th.cos(t * np.pi / 2) - d_sigma_t = -np.pi / 2 * th.sin(t * np.pi / 2) - return sigma_t, d_sigma_t - - def compute_d_alpha_alpha_ratio_t(self, t): - """Special purposed function for computing numerical stabled d_alpha_t / alpha_t""" - return np.pi / (2 * th.tan(t * np.pi / 2)) diff --git a/hy3dshape/hy3dshape/models/diffusion/transport/transport.py b/hy3dshape/hy3dshape/models/diffusion/transport/transport.py deleted file mode 100644 index 6b51d028da744d1241345086be4a06af90eb3f10..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/transport/transport.py +++ /dev/null @@ -1,534 +0,0 @@ -# This file includes code derived from the SiT project (https://github.com/willisma/SiT), -# which is licensed under the MIT License. -# -# MIT License -# -# Copyright (c) Meta Platforms, Inc. and affiliates. -# -# Permission is hereby granted, free of charge, to any person obtaining a copy -# of this software and associated documentation files (the "Software"), to deal -# in the Software without restriction, including without limitation the rights -# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -# copies of the Software, and to permit persons to whom the Software is -# furnished to do so, subject to the following conditions: -# -# The above copyright notice and this permission notice shall be included in all -# copies or substantial portions of the Software. -# -# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -# SOFTWARE. - -import torch as th -import numpy as np -import logging - -import enum - -from . import path -from .utils import EasyDict, log_state, mean_flat -from .integrators import ode, sde - - -class ModelType(enum.Enum): - """ - Which type of output the model predicts. - """ - - NOISE = enum.auto() # the model predicts epsilon - SCORE = enum.auto() # the model predicts \nabla \log p(x) - VELOCITY = enum.auto() # the model predicts v(x) - - -class PathType(enum.Enum): - """ - Which type of path to use. - """ - - LINEAR = enum.auto() - GVP = enum.auto() - VP = enum.auto() - - -class WeightType(enum.Enum): - """ - Which type of weighting to use. - """ - - NONE = enum.auto() - VELOCITY = enum.auto() - LIKELIHOOD = enum.auto() - - -class Transport: - - def __init__( - self, - *, - model_type, - path_type, - loss_type, - train_eps, - sample_eps, - train_sample_type = "uniform", - **kwargs, - ): - path_options = { - PathType.LINEAR: path.ICPlan, - PathType.GVP: path.GVPCPlan, - PathType.VP: path.VPCPlan, - } - - self.loss_type = loss_type - self.model_type = model_type - self.path_sampler = path_options[path_type]() - self.train_eps = train_eps - self.sample_eps = sample_eps - self.train_sample_type = train_sample_type - if self.train_sample_type == "logit_normal": - self.mean = kwargs['mean'] - self.std = kwargs['std'] - self.shift_scale = kwargs['shift_scale'] - print(f"using logit normal sample, shift scale is {self.shift_scale}") - - def prior_logp(self, z): - ''' - Standard multivariate normal prior - Assume z is batched - ''' - shape = th.tensor(z.size()) - N = th.prod(shape[1:]) - _fn = lambda x: -N / 2. * np.log(2 * np.pi) - th.sum(x ** 2) / 2. - return th.vmap(_fn)(z) - - def check_interval( - self, - train_eps, - sample_eps, - *, - diffusion_form="SBDM", - sde=False, - reverse=False, - eval=False, - last_step_size=0.0, - ): - t0 = 0 - t1 = 1 - eps = train_eps if not eval else sample_eps - if (type(self.path_sampler) in [path.VPCPlan]): - - t1 = 1 - eps if (not sde or last_step_size == 0) else 1 - last_step_size - - elif (type(self.path_sampler) in [path.ICPlan, path.GVPCPlan]) \ - and ( - self.model_type != ModelType.VELOCITY or sde): # avoid numerical issue by taking a first semi-implicit step - - t0 = eps if (diffusion_form == "SBDM" and sde) or self.model_type != ModelType.VELOCITY else 0 - t1 = 1 - eps if (not sde or last_step_size == 0) else 1 - last_step_size - - if reverse: - t0, t1 = 1 - t0, 1 - t1 - - return t0, t1 - - def sample(self, x1): - """Sampling x0 & t based on shape of x1 (if needed) - Args: - x1 - data point; [batch, *dim] - """ - - x0 = th.randn_like(x1) - if self.train_sample_type=="uniform": - t0, t1 = self.check_interval(self.train_eps, self.sample_eps) - t = th.rand((x1.shape[0],)) * (t1 - t0) + t0 - t = t.to(x1) - elif self.train_sample_type=="logit_normal": - t = th.randn((x1.shape[0],)) * self.std + self.mean - t = t.to(x1) - t = 1/(1+th.exp(-t)) - - t = np.sqrt(self.shift_scale)*t/(1+(np.sqrt(self.shift_scale)-1)*t) - - return t, x0, x1 - - def training_losses( - self, - model, - x1, - model_kwargs=None - ): - """Loss for training the score model - Args: - - model: backbone model; could be score, noise, or velocity - - x1: datapoint - - model_kwargs: additional arguments for the model - """ - if model_kwargs == None: - model_kwargs = {} - - t, x0, x1 = self.sample(x1) - t, xt, ut = self.path_sampler.plan(t, x0, x1) - model_output = model(xt, t, **model_kwargs) - B, *_, C = xt.shape - assert model_output.size() == (B, *xt.size()[1:-1], C) - - terms = {} - terms['pred'] = model_output - if self.model_type == ModelType.VELOCITY: - terms['loss'] = mean_flat(((model_output - ut) ** 2)) - else: - _, drift_var = self.path_sampler.compute_drift(xt, t) - sigma_t, _ = self.path_sampler.compute_sigma_t(path.expand_t_like_x(t, xt)) - if self.loss_type in [WeightType.VELOCITY]: - weight = (drift_var / sigma_t) ** 2 - elif self.loss_type in [WeightType.LIKELIHOOD]: - weight = drift_var / (sigma_t ** 2) - elif self.loss_type in [WeightType.NONE]: - weight = 1 - else: - raise NotImplementedError() - - if self.model_type == ModelType.NOISE: - terms['loss'] = mean_flat(weight * ((model_output - x0) ** 2)) - else: - terms['loss'] = mean_flat(weight * ((model_output * sigma_t + x0) ** 2)) - - return terms - - def get_drift( - self - ): - """member function for obtaining the drift of the probability flow ODE""" - - def score_ode(x, t, model, **model_kwargs): - drift_mean, drift_var = self.path_sampler.compute_drift(x, t) - model_output = model(x, t, **model_kwargs) - return (-drift_mean + drift_var * model_output) # by change of variable - - def noise_ode(x, t, model, **model_kwargs): - drift_mean, drift_var = self.path_sampler.compute_drift(x, t) - sigma_t, _ = self.path_sampler.compute_sigma_t(path.expand_t_like_x(t, x)) - model_output = model(x, t, **model_kwargs) - score = model_output / -sigma_t - return (-drift_mean + drift_var * score) - - def velocity_ode(x, t, model, **model_kwargs): - model_output = model(x, t, **model_kwargs) - return model_output - - if self.model_type == ModelType.NOISE: - drift_fn = noise_ode - elif self.model_type == ModelType.SCORE: - drift_fn = score_ode - else: - drift_fn = velocity_ode - - def body_fn(x, t, model, **model_kwargs): - model_output = drift_fn(x, t, model, **model_kwargs) - assert model_output.shape == x.shape, "Output shape from ODE solver must match input shape" - return model_output - - return body_fn - - def get_score( - self, - ): - """member function for obtaining score of - x_t = alpha_t * x + sigma_t * eps""" - if self.model_type == ModelType.NOISE: - score_fn = lambda x, t, model, **kwargs: model(x, t, **kwargs) / - \ - self.path_sampler.compute_sigma_t(path.expand_t_like_x(t, x))[0] - elif self.model_type == ModelType.SCORE: - score_fn = lambda x, t, model, **kwagrs: model(x, t, **kwagrs) - elif self.model_type == ModelType.VELOCITY: - score_fn = lambda x, t, model, **kwargs: self.path_sampler.get_score_from_velocity(model(x, t, **kwargs), x, - t) - else: - raise NotImplementedError() - - return score_fn - - -class Sampler: - """Sampler class for the transport model""" - - def __init__( - self, - transport, - ): - """Constructor for a general sampler; supporting different sampling methods - Args: - - transport: an tranport object specify model prediction & interpolant type - """ - - self.transport = transport - self.drift = self.transport.get_drift() - self.score = self.transport.get_score() - - def __get_sde_diffusion_and_drift( - self, - *, - diffusion_form="SBDM", - diffusion_norm=1.0, - ): - - def diffusion_fn(x, t): - diffusion = self.transport.path_sampler.compute_diffusion(x, t, form=diffusion_form, norm=diffusion_norm) - return diffusion - - sde_drift = \ - lambda x, t, model, **kwargs: \ - self.drift(x, t, model, **kwargs) + diffusion_fn(x, t) * self.score(x, t, model, **kwargs) - - sde_diffusion = diffusion_fn - - return sde_drift, sde_diffusion - - def __get_last_step( - self, - sde_drift, - *, - last_step, - last_step_size, - ): - """Get the last step function of the SDE solver""" - - if last_step is None: - last_step_fn = \ - lambda x, t, model, **model_kwargs: \ - x - elif last_step == "Mean": - last_step_fn = \ - lambda x, t, model, **model_kwargs: \ - x + sde_drift(x, t, model, **model_kwargs) * last_step_size - elif last_step == "Tweedie": - alpha = self.transport.path_sampler.compute_alpha_t # simple aliasing; the original name was too long - sigma = self.transport.path_sampler.compute_sigma_t - last_step_fn = \ - lambda x, t, model, **model_kwargs: \ - x / alpha(t)[0][0] + (sigma(t)[0][0] ** 2) / alpha(t)[0][0] * self.score(x, t, model, - **model_kwargs) - elif last_step == "Euler": - last_step_fn = \ - lambda x, t, model, **model_kwargs: \ - x + self.drift(x, t, model, **model_kwargs) * last_step_size - else: - raise NotImplementedError() - - return last_step_fn - - def sample_sde( - self, - *, - sampling_method="Euler", - diffusion_form="SBDM", - diffusion_norm=1.0, - last_step="Mean", - last_step_size=0.04, - num_steps=250, - ): - """returns a sampling function with given SDE settings - Args: - - sampling_method: type of sampler used in solving the SDE; default to be Euler-Maruyama - - diffusion_form: function form of diffusion coefficient; default to be matching SBDM - - diffusion_norm: function magnitude of diffusion coefficient; default to 1 - - last_step: type of the last step; default to identity - - last_step_size: size of the last step; default to match the stride of 250 steps over [0,1] - - num_steps: total integration step of SDE - """ - - if last_step is None: - last_step_size = 0.0 - - sde_drift, sde_diffusion = self.__get_sde_diffusion_and_drift( - diffusion_form=diffusion_form, - diffusion_norm=diffusion_norm, - ) - - t0, t1 = self.transport.check_interval( - self.transport.train_eps, - self.transport.sample_eps, - diffusion_form=diffusion_form, - sde=True, - eval=True, - reverse=False, - last_step_size=last_step_size, - ) - - _sde = sde( - sde_drift, - sde_diffusion, - t0=t0, - t1=t1, - num_steps=num_steps, - sampler_type=sampling_method - ) - - last_step_fn = self.__get_last_step(sde_drift, last_step=last_step, last_step_size=last_step_size) - - def _sample(init, model, **model_kwargs): - xs = _sde.sample(init, model, **model_kwargs) - ts = th.ones(init.size(0), device=init.device) * t1 - x = last_step_fn(xs[-1], ts, model, **model_kwargs) - xs.append(x) - - assert len(xs) == num_steps, "Samples does not match the number of steps" - - return xs - - return _sample - - def sample_ode( - self, - *, - sampling_method="dopri5", - num_steps=50, - atol=1e-6, - rtol=1e-3, - reverse=False, - ): - """returns a sampling function with given ODE settings - Args: - - sampling_method: type of sampler used in solving the ODE; default to be Dopri5 - - num_steps: - - fixed solver (Euler, Heun): the actual number of integration steps performed - - adaptive solver (Dopri5): the number of datapoints saved during integration; produced by interpolation - - atol: absolute error tolerance for the solver - - rtol: relative error tolerance for the solver - - reverse: whether solving the ODE in reverse (data to noise); default to False - """ - if reverse: - drift = lambda x, t, model, **kwargs: self.drift(x, th.ones_like(t) * (1 - t), model, **kwargs) - else: - drift = self.drift - - t0, t1 = self.transport.check_interval( - self.transport.train_eps, - self.transport.sample_eps, - sde=False, - eval=True, - reverse=reverse, - last_step_size=0.0, - ) - - _ode = ode( - drift=drift, - t0=t0, - t1=t1, - sampler_type=sampling_method, - num_steps=num_steps, - atol=atol, - rtol=rtol, - ) - - return _ode.sample - - def sample_ode_intermediate( - self, - *, - sampling_method="dopri5", - num_steps=50, - atol=1e-6, - rtol=1e-3, - t=0.5, - reverse=False, - ): - """returns a sampling function with given ODE settings - Args: - - sampling_method: type of sampler used in solving the ODE; default to be Dopri5 - - num_steps: - - fixed solver (Euler, Heun): the actual number of integration steps performed - - adaptive solver (Dopri5): the number of datapoints saved during integration; produced by interpolation - - atol: absolute error tolerance for the solver - - rtol: relative error tolerance for the solver - - reverse: whether solving the ODE in reverse (data to noise); default to False - """ - if reverse: - drift = lambda x, t, model, **kwargs: self.drift(x, th.ones_like(t) * (1 - t), model, **kwargs) - else: - drift = self.drift - - t0, t1 = self.transport.check_interval( - self.transport.train_eps, - self.transport.sample_eps, - sde=False, - eval=True, - reverse=reverse, - last_step_size=0.0, - ) - - _ode = ode( - drift=drift, - t0=t, - t1=t1, - sampler_type=sampling_method, - num_steps=num_steps, - atol=atol, - rtol=rtol, - ) - - return _ode.sample - - def sample_ode_likelihood( - self, - *, - sampling_method="dopri5", - num_steps=50, - atol=1e-6, - rtol=1e-3, - ): - - """returns a sampling function for calculating likelihood with given ODE settings - Args: - - sampling_method: type of sampler used in solving the ODE; default to be Dopri5 - - num_steps: - - fixed solver (Euler, Heun): the actual number of integration steps performed - - adaptive solver (Dopri5): the number of datapoints saved during integration; produced by interpolation - - atol: absolute error tolerance for the solver - - rtol: relative error tolerance for the solver - """ - - def _likelihood_drift(x, t, model, **model_kwargs): - x, _ = x - eps = th.randint(2, x.size(), dtype=th.float, device=x.device) * 2 - 1 - t = th.ones_like(t) * (1 - t) - with th.enable_grad(): - x.requires_grad = True - grad = th.autograd.grad(th.sum(self.drift(x, t, model, **model_kwargs) * eps), x)[0] - logp_grad = th.sum(grad * eps, dim=tuple(range(1, len(x.size())))) - drift = self.drift(x, t, model, **model_kwargs) - return (-drift, logp_grad) - - t0, t1 = self.transport.check_interval( - self.transport.train_eps, - self.transport.sample_eps, - sde=False, - eval=True, - reverse=False, - last_step_size=0.0, - ) - - _ode = ode( - drift=_likelihood_drift, - t0=t0, - t1=t1, - sampler_type=sampling_method, - num_steps=num_steps, - atol=atol, - rtol=rtol, - ) - - def _sample_fn(x, model, **model_kwargs): - init_logp = th.zeros(x.size(0)).to(x) - input = (x, init_logp) - drift, delta_logp = _ode.sample(input, model, **model_kwargs) - drift, delta_logp = drift[-1], delta_logp[-1] - prior_logp = self.transport.prior_logp(drift) - logp = prior_logp - delta_logp - return logp, drift - - return _sample_fn diff --git a/hy3dshape/hy3dshape/models/diffusion/transport/utils.py b/hy3dshape/hy3dshape/models/diffusion/transport/utils.py deleted file mode 100644 index 5830aab015f9186c3e02d462af7580cfb223cfae..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/models/diffusion/transport/utils.py +++ /dev/null @@ -1,54 +0,0 @@ -# This file includes code derived from the SiT project (https://github.com/willisma/SiT), -# which is licensed under the MIT License. -# -# MIT License -# -# Copyright (c) Meta Platforms, Inc. and affiliates. -# -# Permission is hereby granted, free of charge, to any person obtaining a copy -# of this software and associated documentation files (the "Software"), to deal -# in the Software without restriction, including without limitation the rights -# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -# copies of the Software, and to permit persons to whom the Software is -# furnished to do so, subject to the following conditions: -# -# The above copyright notice and this permission notice shall be included in all -# copies or substantial portions of the Software. -# -# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -# SOFTWARE. - -import torch as th - -class EasyDict: - - def __init__(self, sub_dict): - for k, v in sub_dict.items(): - setattr(self, k, v) - - def __getitem__(self, key): - return getattr(self, key) - -def mean_flat(x): - """ - Take the mean over all non-batch dimensions. - """ - return th.mean(x, dim=list(range(1, len(x.size())))) - -def log_state(state): - result = [] - - sorted_state = dict(sorted(state.items())) - for key, value in sorted_state.items(): - # Check if the value is an instance of a class - if "=", "0.17.0.dev0"): - from accelerate import cpu_offload_with_hook - else: - raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") - - torch_device = torch.device(device) - device_index = torch_device.index - - if gpu_id is not None and device_index is not None: - raise ValueError( - f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" - f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of " - f"the device: `device`={torch_device.type}" - ) - - # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) - # or default to previously set id or default to 0 - self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) - - device_type = torch_device.type - device = torch.device(f"{device_type}:{self._offload_gpu_id}") - - if self.device.type != "cpu": - self.to("cpu") - device_mod = getattr(torch, self.device.type, None) - if hasattr(device_mod, "empty_cache") and device_mod.is_available(): - device_mod.empty_cache() - # otherwise we don't see the memory savings (but they probably exist) - - all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)} - - self._all_hooks = [] - hook = None - for model_str in self.model_cpu_offload_seq.split("->"): - model = all_model_components.pop(model_str, None) - if not isinstance(model, torch.nn.Module): - continue - - _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook) - self._all_hooks.append(hook) - - # CPU offload models that are not in the seq chain unless they are explicitly excluded - # these models will stay on CPU until maybe_free_model_hooks is called - # some models cannot be in the seq chain because they are iteratively called, - # such as controlnet - for name, model in all_model_components.items(): - if not isinstance(model, torch.nn.Module): - continue - - if name in self._exclude_from_cpu_offload: - model.to(device) - else: - _, hook = cpu_offload_with_hook(model, device) - self._all_hooks.append(hook) - - def maybe_free_model_hooks(self): - r""" - Function that offloads all components, removes all model hooks that were added when using - `enable_model_cpu_offload` and then applies them again. In case the model has not been offloaded this function - is a no-op. Make sure to add this function to the end of the `__call__` function of your pipeline so that it - functions correctly when applying enable_model_cpu_offload. - """ - if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0: - # `enable_model_cpu_offload` has not be called, so silently do nothing - return - - for hook in self._all_hooks: - # offload model and remove hook from model - hook.offload() - hook.remove() - - # make sure the model is in the same state as before calling it - self.enable_model_cpu_offload() - - @synchronize_timer('Encode cond') - def encode_cond(self, image, additional_cond_inputs, do_classifier_free_guidance, dual_guidance): - bsz = image.shape[0] - cond = self.conditioner(image=image, **additional_cond_inputs) - - if do_classifier_free_guidance: - un_cond = self.conditioner.unconditional_embedding(bsz, **additional_cond_inputs) - - if dual_guidance: - un_cond_drop_main = copy.deepcopy(un_cond) - un_cond_drop_main['additional'] = cond['additional'] - - def cat_recursive(a, b, c): - if isinstance(a, torch.Tensor): - return torch.cat([a, b, c], dim=0).to(self.dtype) - out = {} - for k in a.keys(): - out[k] = cat_recursive(a[k], b[k], c[k]) - return out - - cond = cat_recursive(cond, un_cond_drop_main, un_cond) - else: - def cat_recursive(a, b): - if isinstance(a, torch.Tensor): - return torch.cat([a, b], dim=0).to(self.dtype) - out = {} - for k in a.keys(): - out[k] = cat_recursive(a[k], b[k]) - return out - - cond = cat_recursive(cond, un_cond) - return cond - - def prepare_extra_step_kwargs(self, generator, eta): - # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature - # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. - # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 - # and should be between [0, 1] - - accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) - extra_step_kwargs = {} - if accepts_eta: - extra_step_kwargs["eta"] = eta - - # check if the scheduler accepts generator - accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) - if accepts_generator: - extra_step_kwargs["generator"] = generator - return extra_step_kwargs - - def prepare_latents(self, batch_size, dtype, device, generator, latents=None): - shape = (batch_size, *self.vae.latent_shape) - if isinstance(generator, list) and len(generator) != batch_size: - raise ValueError( - f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" - f" size of {batch_size}. Make sure the batch size matches the length of the generators." - ) - - if latents is None: - latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) - else: - latents = latents.to(device) - - # scale the initial noise by the standard deviation required by the scheduler - latents = latents * getattr(self.scheduler, 'init_noise_sigma', 1.0) - return latents - - def prepare_image(self, image, mask=None) -> dict: - if isinstance(image, torch.Tensor) and isinstance(mask, torch.Tensor): - outputs = { - 'image': image, - 'mask': mask - } - return outputs - - if isinstance(image, str) and not os.path.exists(image): - raise FileNotFoundError(f"Couldn't find image at path {image}") - - if not isinstance(image, list): - image = [image] - - outputs = [] - for img in image: - output = self.image_processor(img) - outputs.append(output) - - cond_input = {k: [] for k in outputs[0].keys()} - for output in outputs: - for key, value in output.items(): - cond_input[key].append(value) - for key, value in cond_input.items(): - if isinstance(value[0], torch.Tensor): - cond_input[key] = torch.cat(value, dim=0) - - return cond_input - - def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): - """ - See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 - - Args: - timesteps (`torch.Tensor`): - generate embedding vectors at these timesteps - embedding_dim (`int`, *optional*, defaults to 512): - dimension of the embeddings to generate - dtype: - data type of the generated embeddings - - Returns: - `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` - """ - assert len(w.shape) == 1 - w = w * 1000.0 - - half_dim = embedding_dim // 2 - emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) - emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) - emb = w.to(dtype)[:, None] * emb[None, :] - emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) - if embedding_dim % 2 == 1: # zero pad - emb = torch.nn.functional.pad(emb, (0, 1)) - assert emb.shape == (w.shape[0], embedding_dim) - return emb - - def set_surface_extractor(self, mc_algo): - if mc_algo is None: - return - logger.info('The parameters `mc_algo` is deprecated, and will be removed in future versions.\n' - 'Please use: \n' - 'from hy3dshape.models.autoencoders import SurfaceExtractors\n' - 'pipeline.vae.surface_extractor = SurfaceExtractors[mc_algo]() instead\n') - if mc_algo not in SurfaceExtractors.keys(): - raise ValueError(f"Unknown mc_algo {mc_algo}") - self.vae.surface_extractor = SurfaceExtractors[mc_algo]() - - @torch.no_grad() - def __call__( - self, - image: Union[str, List[str], Image.Image] = None, - num_inference_steps: int = 50, - timesteps: List[int] = None, - sigmas: List[float] = None, - eta: float = 0.0, - guidance_scale: float = 7.5, - dual_guidance_scale: float = 10.5, - dual_guidance: bool = True, - generator=None, - box_v=1.01, - octree_resolution=384, - mc_level=-1 / 512, - num_chunks=8000, - mc_algo=None, - output_type: Optional[str] = "trimesh", - enable_pbar=True, - **kwargs, - ) -> List[List[trimesh.Trimesh]]: - callback = kwargs.pop("callback", None) - callback_steps = kwargs.pop("callback_steps", None) - - self.set_surface_extractor(mc_algo) - - device = self.device - dtype = self.dtype - do_classifier_free_guidance = guidance_scale >= 0 and \ - getattr(self.model, 'guidance_cond_proj_dim', None) is None - dual_guidance = dual_guidance_scale >= 0 and dual_guidance - - if isinstance(image, torch.Tensor): - pass - else: - cond_inputs = self.prepare_image(image) - image = cond_inputs.pop('image') - - cond = self.encode_cond( - image=image, - additional_cond_inputs=cond_inputs, - do_classifier_free_guidance=do_classifier_free_guidance, - dual_guidance=False, - ) - batch_size = image.shape[0] - - t_dtype = torch.long - inner_scheduler = copy.deepcopy(self.scheduler) - timesteps, num_inference_steps = retrieve_timesteps( - inner_scheduler, num_inference_steps, device, timesteps, sigmas) - - latents = self.prepare_latents(batch_size, dtype, device, generator) - extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) - - guidance_cond = None - if getattr(self.model, 'guidance_cond_proj_dim', None) is not None: - logger.info('Using lcm guidance scale') - guidance_scale_tensor = torch.tensor(guidance_scale - 1).repeat(batch_size) - guidance_cond = self.get_guidance_scale_embedding( - guidance_scale_tensor, embedding_dim=self.model.guidance_cond_proj_dim - ).to(device=device, dtype=latents.dtype) - with synchronize_timer('Diffusion Sampling'): - for i, t in enumerate(tqdm(timesteps, disable=not enable_pbar, desc="Diffusion Sampling:", leave=False)): - # expand the latents if we are doing classifier free guidance - if do_classifier_free_guidance: - latent_model_input = torch.cat([latents] * (3 if dual_guidance else 2)) - else: - latent_model_input = latents - latent_model_input = inner_scheduler.scale_model_input(latent_model_input, t) - - # predict the noise residual - timestep_tensor = torch.tensor([t], dtype=t_dtype, device=device) - timestep_tensor = timestep_tensor.expand(latent_model_input.shape[0]) - noise_pred = self.model(latent_model_input, timestep_tensor, cond, guidance_cond=guidance_cond) - - # no drop, drop clip, all drop - if do_classifier_free_guidance: - if dual_guidance: - noise_pred_clip, noise_pred_dino, noise_pred_uncond = noise_pred.chunk(3) - noise_pred = ( - noise_pred_uncond - + guidance_scale * (noise_pred_clip - noise_pred_dino) - + dual_guidance_scale * (noise_pred_dino - noise_pred_uncond) - ) - else: - noise_pred_cond, noise_pred_uncond = noise_pred.chunk(2) - noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond) - - # compute the previous noisy sample x_t -> x_t-1 - outputs = inner_scheduler.step(noise_pred, t, latents, **extra_step_kwargs) - latents = outputs.prev_sample - - if callback is not None and i % callback_steps == 0: - step_idx = i // getattr(inner_scheduler, "order", 1) - callback(step_idx, t, outputs) - - return self._export( - latents, - output_type, - box_v, mc_level, num_chunks, octree_resolution, mc_algo, - ) - - def _export( - self, - latents, - output_type='trimesh', - box_v=1.01, - mc_level=0.0, - num_chunks=20000, - octree_resolution=256, - mc_algo='mc', - enable_pbar=True - ): - if not output_type == "latent": - latents = 1. / self.vae.scale_factor * latents - latents = self.vae(latents) - outputs = self.vae.latents2mesh( - latents, - bounds=box_v, - mc_level=mc_level, - num_chunks=num_chunks, - octree_resolution=octree_resolution, - mc_algo=mc_algo, - enable_pbar=enable_pbar, - ) - else: - outputs = latents - - if output_type == 'trimesh': - outputs = export_to_trimesh(outputs) - - return outputs - - -class Hunyuan3DDiTFlowMatchingPipeline(Hunyuan3DDiTPipeline): - - @torch.inference_mode() - def __call__( - self, - image: Union[str, List[str], Image.Image, dict, List[dict], torch.Tensor] = None, - num_inference_steps: int = 50, - timesteps: List[int] = None, - sigmas: List[float] = None, - eta: float = 0.0, - guidance_scale: float = 5.0, - generator=None, - box_v=1.01, - octree_resolution=384, - mc_level=0.0, - mc_algo=None, - num_chunks=8000, - output_type: Optional[str] = "trimesh", - enable_pbar=True, - mask = None, - **kwargs, - ) -> List[List[trimesh.Trimesh]]: - callback = kwargs.pop("callback", None) - callback_steps = kwargs.pop("callback_steps", None) - - self.set_surface_extractor(mc_algo) - - device = self.device - dtype = self.dtype - do_classifier_free_guidance = guidance_scale >= 0 and not ( - hasattr(self.model, 'guidance_embed') and - self.model.guidance_embed is True - ) - - # print('image', type(image), 'mask', type(mask)) - cond_inputs = self.prepare_image(image, mask) - image = cond_inputs.pop('image') - cond = self.encode_cond( - image=image, - additional_cond_inputs=cond_inputs, - do_classifier_free_guidance=do_classifier_free_guidance, - dual_guidance=False, - ) - - batch_size = image.shape[0] - - inner_scheduler = copy.deepcopy(self.scheduler) - - # 5. Prepare timesteps - # NOTE: this is slightly different from common usage, we start from 0. - sigmas = np.linspace(0, 1, num_inference_steps) if sigmas is None else sigmas - timesteps, num_inference_steps = retrieve_timesteps( - inner_scheduler, - num_inference_steps, - device, - sigmas=sigmas, - ) - latents = self.prepare_latents(batch_size, dtype, device, generator) - - guidance = None - if hasattr(self.model, 'guidance_embed') and \ - self.model.guidance_embed is True: - guidance = torch.tensor([guidance_scale] * batch_size, device=device, dtype=dtype) - # logger.info(f'Using guidance embed with scale {guidance_scale}') - - with synchronize_timer('Diffusion Sampling'): - for i, t in enumerate(tqdm(timesteps, disable=not enable_pbar, desc="Diffusion Sampling:")): - # expand the latents if we are doing classifier free guidance - if do_classifier_free_guidance: - latent_model_input = torch.cat([latents] * 2) - else: - latent_model_input = latents - - # NOTE: we assume model get timesteps ranged from 0 to 1 - timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype) - timestep = timestep / inner_scheduler.config.num_train_timesteps - noise_pred = self.model(latent_model_input, timestep, cond, guidance=guidance) - - if do_classifier_free_guidance: - noise_pred_cond, noise_pred_uncond = noise_pred.chunk(2) - noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond) - - # compute the previous noisy sample x_t -> x_t-1 - outputs = inner_scheduler.step(noise_pred, t, latents) - latents = outputs.prev_sample - - if callback is not None and i % callback_steps == 0: - step_idx = i // getattr(inner_scheduler, "order", 1) - callback(step_idx, t, outputs) - - del inner_scheduler - return self._export( - latents, - output_type, - box_v, mc_level, num_chunks, octree_resolution, mc_algo, - enable_pbar=enable_pbar, - ) diff --git a/hy3dshape/hy3dshape/postprocessors.py b/hy3dshape/hy3dshape/postprocessors.py deleted file mode 100644 index d258369e2b9116090ade3c955d57215d1481138e..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/postprocessors.py +++ /dev/null @@ -1,202 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import tempfile -from typing import Union - -import numpy as np -import pymeshlab -import torch -import trimesh - -from .models.autoencoders import Latent2MeshOutput -from .utils import synchronize_timer - - -def load_mesh(path): - if path.endswith(".glb"): - mesh = trimesh.load(path) - else: - mesh = pymeshlab.MeshSet() - mesh.load_new_mesh(path) - return mesh - - -def reduce_face(mesh: pymeshlab.MeshSet, max_facenum: int = 200000): - if max_facenum > mesh.current_mesh().face_number(): - return mesh - - mesh.apply_filter( - "meshing_decimation_quadric_edge_collapse", - targetfacenum=max_facenum, - qualitythr=1.0, - preserveboundary=True, - boundaryweight=3, - preservenormal=True, - preservetopology=True, - autoclean=True - ) - return mesh - - -def remove_floater(mesh: pymeshlab.MeshSet): - mesh.apply_filter("compute_selection_by_small_disconnected_components_per_face", - nbfaceratio=0.005) - mesh.apply_filter("compute_selection_transfer_face_to_vertex", inclusive=False) - mesh.apply_filter("meshing_remove_selected_vertices_and_faces") - return mesh - - -def pymeshlab2trimesh(mesh: pymeshlab.MeshSet): - with tempfile.NamedTemporaryFile(suffix='.ply', delete=False) as temp_file: - mesh.save_current_mesh(temp_file.name) - mesh = trimesh.load(temp_file.name) - # 检查加载的对象类型 - if isinstance(mesh, trimesh.Scene): - combined_mesh = trimesh.Trimesh() - # 如果是Scene，遍历所有的geometry并合并 - for geom in mesh.geometry.values(): - combined_mesh = trimesh.util.concatenate([combined_mesh, geom]) - mesh = combined_mesh - return mesh - - -def trimesh2pymeshlab(mesh: trimesh.Trimesh): - with tempfile.NamedTemporaryFile(suffix='.ply', delete=False) as temp_file: - if isinstance(mesh, trimesh.scene.Scene): - for idx, obj in enumerate(mesh.geometry.values()): - if idx == 0: - temp_mesh = obj - else: - temp_mesh = temp_mesh + obj - mesh = temp_mesh - mesh.export(temp_file.name) - mesh = pymeshlab.MeshSet() - mesh.load_new_mesh(temp_file.name) - return mesh - - -def export_mesh(input, output): - if isinstance(input, pymeshlab.MeshSet): - mesh = output - elif isinstance(input, Latent2MeshOutput): - output = Latent2MeshOutput() - output.mesh_v = output.current_mesh().vertex_matrix() - output.mesh_f = output.current_mesh().face_matrix() - mesh = output - else: - mesh = pymeshlab2trimesh(output) - return mesh - - -def import_mesh(mesh: Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput, str]) -> pymeshlab.MeshSet: - if isinstance(mesh, str): - mesh = load_mesh(mesh) - elif isinstance(mesh, Latent2MeshOutput): - mesh = pymeshlab.MeshSet() - mesh_pymeshlab = pymeshlab.Mesh(vertex_matrix=mesh.mesh_v, face_matrix=mesh.mesh_f) - mesh.add_mesh(mesh_pymeshlab, "converted_mesh") - - if isinstance(mesh, (trimesh.Trimesh, trimesh.scene.Scene)): - mesh = trimesh2pymeshlab(mesh) - - return mesh - - -class FaceReducer: - @synchronize_timer('FaceReducer') - def __call__( - self, - mesh: Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput, str], - max_facenum: int = 40000 - ) -> Union[pymeshlab.MeshSet, trimesh.Trimesh]: - ms = import_mesh(mesh) - ms = reduce_face(ms, max_facenum=max_facenum) - mesh = export_mesh(mesh, ms) - return mesh - - -class FloaterRemover: - @synchronize_timer('FloaterRemover') - def __call__( - self, - mesh: Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput, str], - ) -> Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput]: - ms = import_mesh(mesh) - ms = remove_floater(ms) - mesh = export_mesh(mesh, ms) - return mesh - - -class DegenerateFaceRemover: - @synchronize_timer('DegenerateFaceRemover') - def __call__( - self, - mesh: Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput, str], - ) -> Union[pymeshlab.MeshSet, trimesh.Trimesh, Latent2MeshOutput]: - ms = import_mesh(mesh) - - with tempfile.NamedTemporaryFile(suffix='.ply', delete=False) as temp_file: - ms.save_current_mesh(temp_file.name) - ms = pymeshlab.MeshSet() - ms.load_new_mesh(temp_file.name) - - mesh = export_mesh(mesh, ms) - return mesh - - -def mesh_normalize(mesh): - """ - Normalize mesh vertices to sphere - """ - scale_factor = 1.2 - vtx_pos = np.asarray(mesh.vertices) - max_bb = (vtx_pos - 0).max(0)[0] - min_bb = (vtx_pos - 0).min(0)[0] - - center = (max_bb + min_bb) / 2 - - scale = torch.norm(torch.tensor(vtx_pos - center, dtype=torch.float32), dim=1).max() * 2.0 - - vtx_pos = (vtx_pos - center) * (scale_factor / float(scale)) - mesh.vertices = vtx_pos - - return mesh - - -class MeshSimplifier: - def __init__(self, executable: str = None): - if executable is None: - CURRENT_DIR = os.path.dirname(os.path.abspath(__file__)) - executable = os.path.join(CURRENT_DIR, "mesh_simplifier.bin") - self.executable = executable - - @synchronize_timer('MeshSimplifier') - def __call__( - self, - mesh: Union[trimesh.Trimesh], - ) -> Union[trimesh.Trimesh]: - with tempfile.NamedTemporaryFile(suffix='.obj', delete=False) as temp_input: - with tempfile.NamedTemporaryFile(suffix='.obj', delete=False) as temp_output: - mesh.export(temp_input.name) - os.system(f'{self.executable} {temp_input.name} {temp_output.name}') - ms = trimesh.load(temp_output.name, process=False) - if isinstance(ms, trimesh.Scene): - combined_mesh = trimesh.Trimesh() - for geom in ms.geometry.values(): - combined_mesh = trimesh.util.concatenate([combined_mesh, geom]) - ms = combined_mesh - ms = mesh_normalize(ms) - return ms diff --git a/hy3dshape/hy3dshape/preprocessors.py b/hy3dshape/hy3dshape/preprocessors.py deleted file mode 100644 index 8a9cb9ea1591363fb77f4e02351a0f945e3bc1ab..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/preprocessors.py +++ /dev/null @@ -1,167 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import cv2 -import numpy as np -import torch -from PIL import Image -from einops import repeat, rearrange - - -def array_to_tensor(np_array): - image_pt = torch.tensor(np_array).float() - image_pt = image_pt / 255 * 2 - 1 - image_pt = rearrange(image_pt, "h w c -> c h w") - image_pts = repeat(image_pt, "c h w -> b c h w", b=1) - return image_pts - - -class ImageProcessorV2: - def __init__(self, size=512, border_ratio=None): - self.size = size - self.border_ratio = border_ratio - - @staticmethod - def recenter(image, border_ratio: float = 0.2): - """ recenter an image to leave some empty space at the image border. - - Args: - image (ndarray): input image, float/uint8 [H, W, 3/4] - mask (ndarray): alpha mask, bool [H, W] - border_ratio (float, optional): border ratio, image will be resized to (1 - border_ratio). Defaults to 0.2. - - Returns: - ndarray: output image, float/uint8 [H, W, 3/4] - """ - - if image.shape[-1] == 4: - mask = image[..., 3] - else: - mask = np.ones_like(image[..., 0:1]) * 255 - image = np.concatenate([image, mask], axis=-1) - mask = mask[..., 0] - - H, W, C = image.shape - - size = max(H, W) - result = np.zeros((size, size, C), dtype=np.uint8) - - coords = np.nonzero(mask) - x_min, x_max = coords[0].min(), coords[0].max() - y_min, y_max = coords[1].min(), coords[1].max() - h = x_max - x_min - w = y_max - y_min - if h == 0 or w == 0: - raise ValueError('input image is empty') - desired_size = int(size * (1 - border_ratio)) - scale = desired_size / max(h, w) - h2 = int(h * scale) - w2 = int(w * scale) - x2_min = (size - h2) // 2 - x2_max = x2_min + h2 - - y2_min = (size - w2) // 2 - y2_max = y2_min + w2 - - result[x2_min:x2_max, y2_min:y2_max] = cv2.resize(image[x_min:x_max, y_min:y_max], (w2, h2), - interpolation=cv2.INTER_AREA) - - bg = np.ones((result.shape[0], result.shape[1], 3), dtype=np.uint8) * 255 - - mask = result[..., 3:].astype(np.float32) / 255 - result = result[..., :3] * mask + bg * (1 - mask) - - mask = mask * 255 - result = result.clip(0, 255).astype(np.uint8) - mask = mask.clip(0, 255).astype(np.uint8) - return result, mask - - def load_image(self, image, border_ratio=0.15, to_tensor=True): - if isinstance(image, str): - image = cv2.imread(image, cv2.IMREAD_UNCHANGED) - image, mask = self.recenter(image, border_ratio=border_ratio) - image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) - elif isinstance(image, Image.Image): - image = image.convert("RGBA") - image = np.asarray(image) - image, mask = self.recenter(image, border_ratio=border_ratio) - - image = cv2.resize(image, (self.size, self.size), interpolation=cv2.INTER_CUBIC) - mask = cv2.resize(mask, (self.size, self.size), interpolation=cv2.INTER_NEAREST) - mask = mask[..., np.newaxis] - - if to_tensor: - image = array_to_tensor(image) - mask = array_to_tensor(mask) - return image, mask - - def __call__(self, image, border_ratio=0.15, to_tensor=True, **kwargs): - if self.border_ratio is not None: - border_ratio = self.border_ratio - image, mask = self.load_image(image, border_ratio=border_ratio, to_tensor=to_tensor) - outputs = { - 'image': image, - 'mask': mask - } - return outputs - - -class MVImageProcessorV2(ImageProcessorV2): - """ - view order: front, front clockwise 90, back, front clockwise 270 - """ - return_view_idx = True - - def __init__(self, size=512, border_ratio=None): - super().__init__(size, border_ratio) - self.view2idx = { - 'front': 0, - 'left': 1, - 'back': 2, - 'right': 3 - } - - def __call__(self, image_dict, border_ratio=0.15, to_tensor=True, **kwargs): - if self.border_ratio is not None: - border_ratio = self.border_ratio - - images = [] - masks = [] - view_idxs = [] - for idx, (view_tag, image) in enumerate(image_dict.items()): - view_idxs.append(self.view2idx[view_tag]) - image, mask = self.load_image(image, border_ratio=border_ratio, to_tensor=to_tensor) - images.append(image) - masks.append(mask) - - zipped_lists = zip(view_idxs, images, masks) - sorted_zipped_lists = sorted(zipped_lists) - view_idxs, images, masks = zip(*sorted_zipped_lists) - - image = torch.cat(images, 0).unsqueeze(0) - mask = torch.cat(masks, 0).unsqueeze(0) - outputs = { - 'image': image, - 'mask': mask, - 'view_idxs': view_idxs - } - return outputs - - -IMAGE_PROCESSORS = { - "v2": ImageProcessorV2, - 'mv_v2': MVImageProcessorV2, -} - -DEFAULT_IMAGEPROCESSOR = 'v2' diff --git a/hy3dshape/hy3dshape/rembg.py b/hy3dshape/hy3dshape/rembg.py deleted file mode 100644 index 6247f060c9f325b1e267668baf236ec8e4c2dae9..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/rembg.py +++ /dev/null @@ -1,25 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from PIL import Image -from rembg import remove, new_session - - -class BackgroundRemover(): - def __init__(self): - self.session = new_session() - - def __call__(self, image: Image.Image): - output = remove(image, session=self.session, bgcolor=[255, 255, 255, 0]) - return output diff --git a/hy3dshape/hy3dshape/schedulers.py b/hy3dshape/hy3dshape/schedulers.py deleted file mode 100644 index 13f0da86b278679d427498128392a7a1ad69b2e6..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/schedulers.py +++ /dev/null @@ -1,480 +0,0 @@ -# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import math -from dataclasses import dataclass -from typing import List, Optional, Tuple, Union - -import numpy as np -import torch -from diffusers.configuration_utils import ConfigMixin, register_to_config -from diffusers.schedulers.scheduling_utils import SchedulerMixin -from diffusers.utils import BaseOutput, logging - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - - -@dataclass -class FlowMatchEulerDiscreteSchedulerOutput(BaseOutput): - """ - Output class for the scheduler's `step` function output. - - Args: - prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): - Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the - denoising loop. - """ - - prev_sample: torch.FloatTensor - - -class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin): - """ - NOTE: this is very similar to diffusers.FlowMatchEulerDiscreteScheduler. Except our timesteps are reversed - - Euler scheduler. - - This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic - methods the library implements for all schedulers such as loading and saving. - - Args: - num_train_timesteps (`int`, defaults to 1000): - The number of diffusion steps to train the model. - timestep_spacing (`str`, defaults to `"linspace"`): - The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and - Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. - shift (`float`, defaults to 1.0): - The shift value for the timestep schedule. - """ - - _compatibles = [] - order = 1 - - @register_to_config - def __init__( - self, - num_train_timesteps: int = 1000, - shift: float = 1.0, - use_dynamic_shifting=False, - ): - timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32).copy() - timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) - - sigmas = timesteps / num_train_timesteps - if not use_dynamic_shifting: - # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution - sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) - - self.timesteps = sigmas * num_train_timesteps - - self._step_index = None - self._begin_index = None - - self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication - self.sigma_min = self.sigmas[-1].item() - self.sigma_max = self.sigmas[0].item() - - @property - def step_index(self): - """ - The index counter for current timestep. It will increase 1 after each scheduler step. - """ - return self._step_index - - @property - def begin_index(self): - """ - The index for the first timestep. It should be set from pipeline with `set_begin_index` method. - """ - return self._begin_index - - # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index - def set_begin_index(self, begin_index: int = 0): - """ - Sets the begin index for the scheduler. This function should be run from pipeline before the inference. - - Args: - begin_index (`int`): - The begin index for the scheduler. - """ - self._begin_index = begin_index - - def scale_noise( - self, - sample: torch.FloatTensor, - timestep: Union[float, torch.FloatTensor], - noise: Optional[torch.FloatTensor] = None, - ) -> torch.FloatTensor: - """ - Forward process in flow-matching - - Args: - sample (`torch.FloatTensor`): - The input sample. - timestep (`int`, *optional*): - The current timestep in the diffusion chain. - - Returns: - `torch.FloatTensor`: - A scaled input sample. - """ - # Make sure sigmas and timesteps have the same device and dtype as original_samples - sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype) - - if sample.device.type == "mps" and torch.is_floating_point(timestep): - # mps does not support float64 - schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) - timestep = timestep.to(sample.device, dtype=torch.float32) - else: - schedule_timesteps = self.timesteps.to(sample.device) - timestep = timestep.to(sample.device) - - # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index - if self.begin_index is None: - step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep] - elif self.step_index is not None: - # add_noise is called after first denoising step (for inpainting) - step_indices = [self.step_index] * timestep.shape[0] - else: - # add noise is called before first denoising step to create initial latent(img2img) - step_indices = [self.begin_index] * timestep.shape[0] - - sigma = sigmas[step_indices].flatten() - while len(sigma.shape) < len(sample.shape): - sigma = sigma.unsqueeze(-1) - - sample = sigma * noise + (1.0 - sigma) * sample - - return sample - - def _sigma_to_t(self, sigma): - return sigma * self.config.num_train_timesteps - - def time_shift(self, mu: float, sigma: float, t: torch.Tensor): - return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) - - def set_timesteps( - self, - num_inference_steps: int = None, - device: Union[str, torch.device] = None, - sigmas: Optional[List[float]] = None, - mu: Optional[float] = None, - ): - """ - Sets the discrete timesteps used for the diffusion chain (to be run before inference). - - Args: - num_inference_steps (`int`): - The number of diffusion steps used when generating samples with a pre-trained model. - device (`str` or `torch.device`, *optional*): - The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. - """ - - if self.config.use_dynamic_shifting and mu is None: - raise ValueError(" you have a pass a value for `mu` when `use_dynamic_shifting` is set to be `True`") - - if sigmas is None: - self.num_inference_steps = num_inference_steps - timesteps = np.linspace( - self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps - ) - - sigmas = timesteps / self.config.num_train_timesteps - - if self.config.use_dynamic_shifting: - sigmas = self.time_shift(mu, 1.0, sigmas) - else: - sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas) - - sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) - timesteps = sigmas * self.config.num_train_timesteps - - self.timesteps = timesteps.to(device=device) - self.sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)]) - - self._step_index = None - self._begin_index = None - - def index_for_timestep(self, timestep, schedule_timesteps=None): - if schedule_timesteps is None: - schedule_timesteps = self.timesteps - - indices = (schedule_timesteps == timestep).nonzero() - - # The sigma index that is taken for the **very** first `step` - # is always the second index (or the last index if there is only 1) - # This way we can ensure we don't accidentally skip a sigma in - # case we start in the middle of the denoising schedule (e.g. for image-to-image) - pos = 1 if len(indices) > 1 else 0 - - return indices[pos].item() - - def _init_step_index(self, timestep): - if self.begin_index is None: - if isinstance(timestep, torch.Tensor): - timestep = timestep.to(self.timesteps.device) - self._step_index = self.index_for_timestep(timestep) - else: - self._step_index = self._begin_index - - def step( - self, - model_output: torch.FloatTensor, - timestep: Union[float, torch.FloatTensor], - sample: torch.FloatTensor, - s_churn: float = 0.0, - s_tmin: float = 0.0, - s_tmax: float = float("inf"), - s_noise: float = 1.0, - generator: Optional[torch.Generator] = None, - return_dict: bool = True, - ) -> Union[FlowMatchEulerDiscreteSchedulerOutput, Tuple]: - """ - Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion - process from the learned model outputs (most often the predicted noise). - - Args: - model_output (`torch.FloatTensor`): - The direct output from learned diffusion model. - timestep (`float`): - The current discrete timestep in the diffusion chain. - sample (`torch.FloatTensor`): - A current instance of a sample created by the diffusion process. - s_churn (`float`): - s_tmin (`float`): - s_tmax (`float`): - s_noise (`float`, defaults to 1.0): - Scaling factor for noise added to the sample. - generator (`torch.Generator`, *optional*): - A random number generator. - return_dict (`bool`): - Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or - tuple. - - Returns: - [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`: - If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is - returned, otherwise a tuple is returned where the first element is the sample tensor. - """ - - if ( - isinstance(timestep, int) - or isinstance(timestep, torch.IntTensor) - or isinstance(timestep, torch.LongTensor) - ): - raise ValueError( - ( - "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" - " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" - " one of the `scheduler.timesteps` as a timestep." - ), - ) - - if self.step_index is None: - self._init_step_index(timestep) - - # Upcast to avoid precision issues when computing prev_sample - sample = sample.to(torch.float32) - - sigma = self.sigmas[self.step_index] - sigma_next = self.sigmas[self.step_index + 1] - - prev_sample = sample + (sigma_next - sigma) * model_output - - # Cast sample back to model compatible dtype - prev_sample = prev_sample.to(model_output.dtype) - - # upon completion increase step index by one - self._step_index += 1 - - if not return_dict: - return (prev_sample,) - - return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample) - - def __len__(self): - return self.config.num_train_timesteps - - -@dataclass -class ConsistencyFlowMatchEulerDiscreteSchedulerOutput(BaseOutput): - prev_sample: torch.FloatTensor - pred_original_sample: torch.FloatTensor - - -class ConsistencyFlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin): - _compatibles = [] - order = 1 - - @register_to_config - def __init__( - self, - num_train_timesteps: int = 1000, - pcm_timesteps: int = 50, - ): - sigmas = np.linspace(0, 1, num_train_timesteps) - step_ratio = num_train_timesteps // pcm_timesteps - - euler_timesteps = (np.arange(1, pcm_timesteps) * step_ratio).round().astype(np.int64) - 1 - euler_timesteps = np.asarray([0] + euler_timesteps.tolist()) - - self.euler_timesteps = euler_timesteps - self.sigmas = sigmas[self.euler_timesteps] - self.sigmas = torch.from_numpy((self.sigmas.copy())).to(dtype=torch.float32) - self.timesteps = self.sigmas * num_train_timesteps - self._step_index = None - self._begin_index = None - self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication - - @property - def step_index(self): - """ - The index counter for current timestep. It will increase 1 after each scheduler step. - """ - return self._step_index - - @property - def begin_index(self): - """ - The index for the first timestep. It should be set from pipeline with `set_begin_index` method. - """ - return self._begin_index - - # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index - def set_begin_index(self, begin_index: int = 0): - """ - Sets the begin index for the scheduler. This function should be run from pipeline before the inference. - - Args: - begin_index (`int`): - The begin index for the scheduler. - """ - self._begin_index = begin_index - - def _sigma_to_t(self, sigma): - return sigma * self.config.num_train_timesteps - - def set_timesteps( - self, - num_inference_steps: int = None, - device: Union[str, torch.device] = None, - sigmas: Optional[List[float]] = None, - ): - """ - Sets the discrete timesteps used for the diffusion chain (to be run before inference). - - Args: - num_inference_steps (`int`): - The number of diffusion steps used when generating samples with a pre-trained model. - device (`str` or `torch.device`, *optional*): - The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. - """ - self.num_inference_steps = num_inference_steps if num_inference_steps is not None else len(sigmas) - inference_indices = np.linspace( - 0, self.config.pcm_timesteps, num=self.num_inference_steps, endpoint=False - ) - inference_indices = np.floor(inference_indices).astype(np.int64) - inference_indices = torch.from_numpy(inference_indices).long() - - self.sigmas_ = self.sigmas[inference_indices] - timesteps = self.sigmas_ * self.config.num_train_timesteps - self.timesteps = timesteps.to(device=device) - self.sigmas_ = torch.cat( - [self.sigmas_, torch.ones(1, device=self.sigmas_.device)] - ) - - self._step_index = None - self._begin_index = None - - def index_for_timestep(self, timestep, schedule_timesteps=None): - if schedule_timesteps is None: - schedule_timesteps = self.timesteps - - indices = (schedule_timesteps == timestep).nonzero() - - # The sigma index that is taken for the **very** first `step` - # is always the second index (or the last index if there is only 1) - # This way we can ensure we don't accidentally skip a sigma in - # case we start in the middle of the denoising schedule (e.g. for image-to-image) - pos = 1 if len(indices) > 1 else 0 - - return indices[pos].item() - - def _init_step_index(self, timestep): - if self.begin_index is None: - if isinstance(timestep, torch.Tensor): - timestep = timestep.to(self.timesteps.device) - self._step_index = self.index_for_timestep(timestep) - else: - self._step_index = self._begin_index - - def step( - self, - model_output: torch.FloatTensor, - timestep: Union[float, torch.FloatTensor], - sample: torch.FloatTensor, - generator: Optional[torch.Generator] = None, - return_dict: bool = True, - ) -> Union[ConsistencyFlowMatchEulerDiscreteSchedulerOutput, Tuple]: - if ( - isinstance(timestep, int) - or isinstance(timestep, torch.IntTensor) - or isinstance(timestep, torch.LongTensor) - ): - raise ValueError( - ( - "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" - " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" - " one of the `scheduler.timesteps` as a timestep." - ), - ) - - if self.step_index is None: - self._init_step_index(timestep) - - sample = sample.to(torch.float32) - - sigma = self.sigmas_[self.step_index] - sigma_next = self.sigmas_[self.step_index + 1] - - prev_sample = sample + (sigma_next - sigma) * model_output - prev_sample = prev_sample.to(model_output.dtype) - - pred_original_sample = sample + (1.0 - sigma) * model_output - pred_original_sample = pred_original_sample.to(model_output.dtype) - - self._step_index += 1 - - if not return_dict: - return (prev_sample,) - - return ConsistencyFlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample, - pred_original_sample=pred_original_sample) - - def __len__(self): - return self.config.num_train_timesteps diff --git a/hy3dshape/hy3dshape/surface_loaders.py b/hy3dshape/hy3dshape/surface_loaders.py deleted file mode 100644 index cb638f367d4ca7a69078878aca5a36d5f60972dc..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/surface_loaders.py +++ /dev/null @@ -1,234 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - - - -import numpy as np - -import torch -import trimesh - - -def normalize_mesh(mesh, scale=0.9999): - """ - Normalize the mesh to fit inside a centered cube with a specified scale. - - The mesh is translated so that its bounding box center is at the origin, - then uniformly scaled so that the longest side of the bounding box fits within [-scale, scale]. - - Args: - mesh (trimesh.Trimesh): Input mesh to normalize. - scale (float, optional): Scaling factor to slightly shrink the mesh inside the unit cube. Default is 0.9999. - - Returns: - trimesh.Trimesh: The normalized mesh with applied translation and scaling. - """ - bbox = mesh.bounds - center = (bbox[1] + bbox[0]) / 2 - scale_ = (bbox[1] - bbox[0]).max() - - mesh.apply_translation(-center) - mesh.apply_scale(1 / scale_ * 2 * scale) - - return mesh - - -def sample_pointcloud(mesh, num=200000): - """ - Sample points uniformly from the surface of the mesh along with their corresponding face normals. - - Args: - mesh (trimesh.Trimesh): Input mesh to sample from. - num (int, optional): Number of points to sample. Default is 200000. - - Returns: - Tuple[torch.Tensor, torch.Tensor]: - - points: Sampled points as a float tensor of shape (num, 3). - - normals: Corresponding normals as a float tensor of shape (num, 3). - """ - points, face_idx = mesh.sample(num, return_index=True) - normals = mesh.face_normals[face_idx] - points = torch.from_numpy(points.astype(np.float32)) - normals = torch.from_numpy(normals.astype(np.float32)) - return points, normals - - -def load_surface(mesh, num_points=8192): - """ - Normalize the mesh, sample points and normals from its surface, and randomly select a subset. - - Args: - mesh (trimesh.Trimesh): Input mesh to process. - num_points (int, optional): Number of points to randomly select - from the sampled surface points. Default is 8192. - - Returns: - Tuple[torch.Tensor, trimesh.Trimesh]: - - surface: Tensor of shape (1, num_points, 6), concatenating points and normals. - - mesh: The normalized mesh. - """ - - mesh = normalize_mesh(mesh, scale=0.98) - surface, normal = sample_pointcloud(mesh) - - rng = np.random.default_rng(seed=0) - ind = rng.choice(surface.shape[0], num_points, replace=False) - surface = torch.FloatTensor(surface[ind]) - normal = torch.FloatTensor(normal[ind]) - - surface = torch.cat([surface, normal], dim=-1).unsqueeze(0) - - return surface, mesh - - -def sharp_sample_pointcloud(mesh, num=16384): - """ - Sample points and normals preferentially from sharp edges of the mesh. - - Sharp edges are detected based on the angle between vertex normals and face normals. - Points are sampled along these edges proportionally to edge length. - - Args: - mesh (trimesh.Trimesh): Input mesh to sample from. - num (int, optional): Number of points to sample from sharp edges. Default is 16384. - - Returns: - Tuple[np.ndarray, np.ndarray]: - - samples: Sampled points along sharp edges, shape (num, 3). - - normals: Corresponding interpolated normals, shape (num, 3). - """ - V = mesh.vertices - N = mesh.face_normals - VN = mesh.vertex_normals - F = mesh.faces - VN2 = np.ones(V.shape[0]) - for i in range(3): - dot = np.stack((VN2[F[:, i]], np.sum(VN[F[:, i]] * N, axis=-1)), axis=-1) - VN2[F[:, i]] = np.min(dot, axis=-1) - - sharp_mask = VN2 < 0.985 - # collect edge - edge_a = np.concatenate((F[:, 0], F[:, 1], F[:, 2])) - edge_b = np.concatenate((F[:, 1], F[:, 2], F[:, 0])) - sharp_edge = ((sharp_mask[edge_a] * sharp_mask[edge_b])) - edge_a = edge_a[sharp_edge > 0] - edge_b = edge_b[sharp_edge > 0] - - sharp_verts_a = V[edge_a] - sharp_verts_b = V[edge_b] - sharp_verts_an = VN[edge_a] - sharp_verts_bn = VN[edge_b] - - weights = np.linalg.norm(sharp_verts_b - sharp_verts_a, axis=-1) - weights /= np.sum(weights) - - random_number = np.random.rand(num) - w = np.random.rand(num, 1) - index = np.searchsorted(weights.cumsum(), random_number) - samples = w * sharp_verts_a[index] + (1 - w) * sharp_verts_b[index] - normals = w * sharp_verts_an[index] + (1 - w) * sharp_verts_bn[index] - return samples, normals - - -def load_surface_sharpegde(mesh, num_points=4096, num_sharp_points=4096, sharpedge_flag=True): - try: - mesh_full = trimesh.util.concatenate(mesh.dump()) - except Exception as err: - mesh_full = trimesh.util.concatenate(mesh) - mesh_full = normalize_mesh(mesh_full) - - origin_num = mesh_full.faces.shape[0] - original_vertices = mesh_full.vertices - original_faces = mesh_full.faces - - mesh = trimesh.Trimesh(vertices=original_vertices, faces=original_faces[:origin_num]) - mesh_fill = trimesh.Trimesh(vertices=original_vertices, faces=original_faces[origin_num:]) - area = mesh.area - area_fill = mesh_fill.area - sample_num = 499712 // 2 - num_fill = int(sample_num * (area_fill / (area + area_fill))) - num = sample_num - num_fill - - random_surface, random_normal = sample_pointcloud(mesh, num=num) - if num_fill == 0: - random_surface_fill, random_normal_fill = np.zeros((0, 3)), np.zeros((0, 3)) - else: - random_surface_fill, random_normal_fill = sample_pointcloud(mesh_fill, num=num_fill) - random_sharp_surface, sharp_normal = sharp_sample_pointcloud(mesh, num=sample_num) - - # save_surface - surface = np.concatenate((random_surface, random_normal), axis=1).astype(np.float16) - surface_fill = np.concatenate((random_surface_fill, random_normal_fill), axis=1).astype(np.float16) - sharp_surface = np.concatenate((random_sharp_surface, sharp_normal), axis=1).astype(np.float16) - surface = np.concatenate((surface, surface_fill), axis=0) - if sharpedge_flag: - sharpedge_label = np.zeros((surface.shape[0], 1)) - surface = np.concatenate((surface, sharpedge_label), axis=1) - sharpedge_label = np.ones((sharp_surface.shape[0], 1)) - sharp_surface = np.concatenate((sharp_surface, sharpedge_label), axis=1) - rng = np.random.default_rng() - ind = rng.choice(surface.shape[0], num_points, replace=False) - surface = torch.FloatTensor(surface[ind]) - ind = rng.choice(sharp_surface.shape[0], num_sharp_points, replace=False) - sharp_surface = torch.FloatTensor(sharp_surface[ind]) - - return torch.cat([surface, sharp_surface], dim=0).unsqueeze(0), mesh_full - - -class SurfaceLoader: - def __init__(self, num_points=8192): - self.num_points = num_points - - def __call__(self, mesh_or_mesh_path, num_points=None): - if num_points is None: - num_points = self.num_points - - mesh = mesh_or_mesh_path - if isinstance(mesh, str): - mesh = trimesh.load(mesh, force="mesh", merge_primitives=True) - if isinstance(mesh, trimesh.scene.Scene): - for idx, obj in enumerate(mesh.geometry.values()): - if idx == 0: - temp_mesh = obj - else: - temp_mesh = temp_mesh + obj - mesh = temp_mesh - surface, mesh = load_surface(mesh, num_points=num_points) - return surface - - -class SharpEdgeSurfaceLoader: - def __init__(self, num_uniform_points=8192, num_sharp_points=8192, **kwargs): - self.num_uniform_points = num_uniform_points - self.num_sharp_points = num_sharp_points - self.num_points = num_uniform_points + num_sharp_points - - def __call__(self, mesh_or_mesh_path, num_uniform_points=None, num_sharp_points=None): - if num_uniform_points is None: - num_uniform_points = self.num_uniform_points - if num_sharp_points is None: - num_sharp_points = self.num_sharp_points - - mesh = mesh_or_mesh_path - if isinstance(mesh, str): - mesh = trimesh.load(mesh, force="mesh", merge_primitives=True) - if isinstance(mesh, trimesh.scene.Scene): - for idx, obj in enumerate(mesh.geometry.values()): - if idx == 0: - temp_mesh = obj - else: - temp_mesh = temp_mesh + obj - mesh = temp_mesh - surface, mesh = load_surface_sharpegde(mesh, num_points=num_uniform_points, num_sharp_points=num_sharp_points) - return surface diff --git a/hy3dshape/hy3dshape/utils/__init__.py b/hy3dshape/hy3dshape/utils/__init__.py deleted file mode 100644 index 0f92daa391f814e1a2c2cad0601e10be49b71998..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/__init__.py +++ /dev/null @@ -1,5 +0,0 @@ -# -*- coding: utf-8 -*- - -from .misc import get_config_from_file -from .misc import instantiate_from_config -from .utils import get_logger, logger, synchronize_timer, smart_load_model diff --git a/hy3dshape/hy3dshape/utils/ema.py b/hy3dshape/hy3dshape/utils/ema.py deleted file mode 100644 index d90df922fd96ab831a2fa26cd16cf1b565fbeb2b..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/ema.py +++ /dev/null @@ -1,76 +0,0 @@ -import torch -from torch import nn - - -class LitEma(nn.Module): - def __init__(self, model, decay=0.9999, use_num_updates=True): - super().__init__() - if decay < 0.0 or decay > 1.0: - raise ValueError('Decay must be between 0 and 1') - - self.m_name2s_name = {} - self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) - self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_updates - else torch.tensor(-1, dtype=torch.int)) - - for name, p in model.named_parameters(): - if p.requires_grad: - # remove as '.'-character is not allowed in buffers - s_name = name.replace('.', '_____') - self.m_name2s_name.update({name: s_name}) - self.register_buffer(s_name, p.clone().detach().data) - - self.collected_params = [] - - def forward(self, model): - decay = self.decay - - if self.num_updates >= 0: - self.num_updates += 1 - decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) - - one_minus_decay = 1.0 - decay - - with torch.no_grad(): - m_param = dict(model.named_parameters()) - shadow_params = dict(self.named_buffers()) - - for key in m_param: - if m_param[key].requires_grad: - sname = self.m_name2s_name[key] - shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) - shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) - else: - assert not key in self.m_name2s_name - - def copy_to(self, model): - m_param = dict(model.named_parameters()) - shadow_params = dict(self.named_buffers()) - for key in m_param: - if m_param[key].requires_grad: - m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) - else: - assert not key in self.m_name2s_name - - def store(self, model): - """ - Save the current parameters for restoring later. - Args: - parameters: Iterable of `torch.nn.Parameter`; the parameters to be - temporarily stored. - """ - self.collected_params = [param.clone() for param in model.parameters()] - - def restore(self, model): - """ - Restore the parameters stored with the `store` method. - Useful to validate the model with EMA parameters without affecting the - original optimization process. Store the parameters before the - `copy_to` method. After validation (or model saving), use this to - restore the former parameters. - Args: - parameters: Iterable of `torch.nn.Parameter`; the parameters to be - updated with the stored parameters. - """ - for c_param, param in zip(self.collected_params, model.parameters()): - param.data.copy_(c_param.data) diff --git a/hy3dshape/hy3dshape/utils/misc.py b/hy3dshape/hy3dshape/utils/misc.py deleted file mode 100644 index 55e11369611ccc60bb7b4afdd8b2d9efb14bf0d2..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/misc.py +++ /dev/null @@ -1,125 +0,0 @@ -# -*- coding: utf-8 -*- - -import importlib -from omegaconf import OmegaConf, DictConfig, ListConfig - -import torch -import torch.distributed as dist -from typing import Union - - -def get_config_from_file(config_file: str) -> Union[DictConfig, ListConfig]: - config_file = OmegaConf.load(config_file) - - if 'base_config' in config_file.keys(): - if config_file['base_config'] == "default_base": - base_config = OmegaConf.create() - # base_config = get_default_config() - elif config_file['base_config'].endswith(".yaml"): - base_config = get_config_from_file(config_file['base_config']) - else: - raise ValueError(f"{config_file} must be `.yaml` file or it contains `base_config` key.") - - config_file = {key: value for key, value in config_file if key != "base_config"} - - return OmegaConf.merge(base_config, config_file) - - return config_file - - -def get_obj_from_str(string, reload=False): - module, cls = string.rsplit(".", 1) - if reload: - module_imp = importlib.import_module(module) - importlib.reload(module_imp) - return getattr(importlib.import_module(module, package=None), cls) - - -def get_obj_from_config(config): - if "target" not in config: - raise KeyError("Expected key `target` to instantiate.") - - return get_obj_from_str(config["target"]) - - -def instantiate_from_config(config, **kwargs): - if "target" not in config: - raise KeyError("Expected key `target` to instantiate.") - - cls = get_obj_from_str(config["target"]) - - if config.get("from_pretrained", None): - return cls.from_pretrained( - config["from_pretrained"], - use_safetensors=config.get('use_safetensors', False), - variant=config.get('variant', 'fp16')) - - params = config.get("params", dict()) - # params.update(kwargs) - # instance = cls(**params) - kwargs.update(params) - instance = cls(**kwargs) - - return instance - - -def disabled_train(self, mode=True): - """Overwrite model.train with this function to make sure train/eval mode - does not change anymore.""" - return self - - -def instantiate_non_trainable_model(config): - model = instantiate_from_config(config) - model = model.eval() - model.train = disabled_train - for param in model.parameters(): - param.requires_grad = False - - return model - - -def is_dist_avail_and_initialized(): - if not dist.is_available(): - return False - if not dist.is_initialized(): - return False - return True - - -def get_rank(): - if not is_dist_avail_and_initialized(): - return 0 - return dist.get_rank() - - -def get_world_size(): - if not is_dist_avail_and_initialized(): - return 1 - return dist.get_world_size() - - -def all_gather_batch(tensors): - """ - Performs all_gather operation on the provided tensors. - """ - # Queue the gathered tensors - world_size = get_world_size() - # There is no need for reduction in the single-proc case - if world_size == 1: - return tensors - tensor_list = [] - output_tensor = [] - for tensor in tensors: - tensor_all = [torch.ones_like(tensor) for _ in range(world_size)] - dist.all_gather( - tensor_all, - tensor, - async_op=False # performance opt - ) - - tensor_list.append(tensor_all) - - for tensor_all in tensor_list: - output_tensor.append(torch.cat(tensor_all, dim=0)) - return output_tensor diff --git a/hy3dshape/hy3dshape/utils/trainings/__init__.py b/hy3dshape/hy3dshape/utils/trainings/__init__.py deleted file mode 100644 index 40a96afc6ff09d58a702b76e3f7dd412fe975e26..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/__init__.py +++ /dev/null @@ -1 +0,0 @@ -# -*- coding: utf-8 -*- diff --git a/hy3dshape/hy3dshape/utils/trainings/callback.py b/hy3dshape/hy3dshape/utils/trainings/callback.py deleted file mode 100644 index e75ac9f99b89a397ae461a3df2f0c317c4216e06..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/callback.py +++ /dev/null @@ -1,213 +0,0 @@ -# ------------------------------------------------------------------------------------ -# Modified from Taming Transformers (https://github.com/CompVis/taming-transformers) -# Copyright (c) 2020 Patrick Esser and Robin Rombach and Björn Ommer. All Rights Reserved. -# ------------------------------------------------------------------------------------ - -import os -import time -import wandb -import numpy as np -from PIL import Image -from pathlib import Path -from omegaconf import OmegaConf, DictConfig -from typing import Tuple, Generic, Dict, Callable, Optional, Any -from pprint import pprint - -import torch -import torchvision -import pytorch_lightning as pl -import pytorch_lightning.loggers -from pytorch_lightning.loggers import WandbLogger -from pytorch_lightning.loggers.logger import DummyLogger -from pytorch_lightning.utilities import rank_zero_only, rank_zero_info -from pytorch_lightning.callbacks import Callback - -from functools import wraps - -def node_zero_only(fn: Callable) -> Callable: - @wraps(fn) - def wrapped_fn(*args, **kwargs) -> Optional[Any]: - if node_zero_only.node == 0: - return fn(*args, **kwargs) - return None - return wrapped_fn - -node_zero_only.node = getattr(node_zero_only, 'node', int(os.environ.get('NODE_RANK', 0))) - -def node_zero_experiment(fn: Callable) -> Callable: - """Returns the real experiment on rank 0 and otherwise the DummyExperiment.""" - @wraps(fn) - def experiment(self): - @node_zero_only - def get_experiment(): - return fn(self) - return get_experiment() or DummyLogger.experiment - return experiment - -# customize wandb for node 0 only -class MyWandbLogger(WandbLogger): - @WandbLogger.experiment.getter - @node_zero_experiment - def experiment(self): - return super().experiment - -class SetupCallback(Callback): - def __init__(self, config: DictConfig, exp_config: DictConfig, - basedir: Path, logdir: str = "log", ckptdir: str = "ckpt") -> None: - super().__init__() - self.logdir = basedir / logdir - self.ckptdir = basedir / ckptdir - self.config = config - self.exp_config = exp_config - - # def on_pretrain_routine_start(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule) -> None: - # if trainer.global_rank == 0: - # # Create logdirs and save configs - # os.makedirs(self.logdir, exist_ok=True) - # os.makedirs(self.ckptdir, exist_ok=True) - # - # print("Experiment config") - # print(self.exp_config.pretty()) - # - # print("Model config") - # print(self.config.pretty()) - - def on_fit_start(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule) -> None: - if trainer.global_rank == 0: - # Create logdirs and save configs - os.makedirs(self.logdir, exist_ok=True) - os.makedirs(self.ckptdir, exist_ok=True) - - # print("Experiment config") - # pprint(self.exp_config) - # - # print("Model config") - # pprint(self.config) - - -class ImageLogger(Callback): - def __init__(self, batch_frequency: int, max_images: int, clamp: bool = True, - increase_log_steps: bool = True) -> None: - - super().__init__() - self.batch_freq = batch_frequency - self.max_images = max_images - self.logger_log_images = { - pl.loggers.WandbLogger: self._wandb, - pl.loggers.TestTubeLogger: self._testtube, - } - self.log_steps = [2 ** n for n in range(int(np.log2(self.batch_freq)) + 1)] - if not increase_log_steps: - self.log_steps = [self.batch_freq] - self.clamp = clamp - - @rank_zero_only - def _wandb(self, pl_module, images, batch_idx, split): - # raise ValueError("No way wandb") - grids = dict() - for k in images: - grid = torchvision.utils.make_grid(images[k]) - grids[f"{split}/{k}"] = wandb.Image(grid) - pl_module.logger.experiment.log(grids) - - @rank_zero_only - def _testtube(self, pl_module, images, batch_idx, split): - for k in images: - grid = torchvision.utils.make_grid(images[k]) - grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w - - tag = f"{split}/{k}" - pl_module.logger.experiment.add_image( - tag, grid, - global_step=pl_module.global_step) - - @rank_zero_only - def log_local(self, save_dir: str, split: str, images: Dict, - global_step: int, current_epoch: int, batch_idx: int) -> None: - root = os.path.join(save_dir, "results", split) - os.makedirs(root, exist_ok=True) - for k in images: - grid = torchvision.utils.make_grid(images[k], nrow=4) - - grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1) - grid = grid.numpy() - grid = (grid * 255).astype(np.uint8) - filename = "{}_gs-{:06}_e-{:06}_b-{:06}.png".format( - k, - global_step, - current_epoch, - batch_idx) - path = os.path.join(root, filename) - os.makedirs(os.path.split(path)[0], exist_ok=True) - Image.fromarray(grid).save(path) - - def log_img(self, pl_module: pl.LightningModule, batch: Tuple[torch.LongTensor, torch.FloatTensor], batch_idx: int, - split: str = "train") -> None: - if (self.check_frequency(batch_idx) and # batch_idx % self.batch_freq == 0 - hasattr(pl_module, "log_images") and - callable(pl_module.log_images) and - self.max_images > 0): - logger = type(pl_module.logger) - - is_train = pl_module.training - if is_train: - pl_module.eval() - - with torch.no_grad(): - images = pl_module.log_images(batch, split=split, pl_module=pl_module) - - for k in images: - N = min(images[k].shape[0], self.max_images) - images[k] = images[k][:N].detach().cpu() - if self.clamp: - images[k] = images[k].clamp(0, 1) - - self.log_local(pl_module.logger.save_dir, split, images, - pl_module.global_step, pl_module.current_epoch, batch_idx) - - logger_log_images = self.logger_log_images.get(logger, lambda *args, **kwargs: None) - logger_log_images(pl_module, images, pl_module.global_step, split) - - if is_train: - pl_module.train() - - def check_frequency(self, batch_idx: int) -> bool: - if (batch_idx % self.batch_freq) == 0 or (batch_idx in self.log_steps): - try: - self.log_steps.pop(0) - except IndexError: - pass - return True - return False - - def on_train_batch_end(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule, - outputs: Generic, batch: Tuple[torch.LongTensor, torch.FloatTensor], batch_idx: int) -> None: - self.log_img(pl_module, batch, batch_idx, split="train") - - def on_validation_batch_end(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule, - outputs: Generic, batch: Tuple[torch.LongTensor, torch.FloatTensor], - dataloader_idx: int, batch_idx: int) -> None: - self.log_img(pl_module, batch, batch_idx, split="val") - - -class CUDACallback(Callback): - # see https://github.com/SeanNaren/minGPT/blob/master/mingpt/callback.py - def on_train_epoch_start(self, trainer, pl_module): - # Reset the memory use counter - torch.cuda.reset_peak_memory_stats(trainer.root_gpu) - torch.cuda.synchronize(trainer.root_gpu) - self.start_time = time.time() - - def on_train_epoch_end(self, trainer, pl_module, outputs): - torch.cuda.synchronize(trainer.root_gpu) - max_memory = torch.cuda.max_memory_allocated(trainer.root_gpu) / 2 ** 20 - epoch_time = time.time() - self.start_time - - try: - max_memory = trainer.training_type_plugin.reduce(max_memory) - epoch_time = trainer.training_type_plugin.reduce(epoch_time) - - rank_zero_info(f"Average Epoch time: {epoch_time:.2f} seconds") - rank_zero_info(f"Average Peak memory {max_memory:.2f}MiB") - except AttributeError: - pass diff --git a/hy3dshape/hy3dshape/utils/trainings/lr_scheduler.py b/hy3dshape/hy3dshape/utils/trainings/lr_scheduler.py deleted file mode 100644 index bb7224776457a4395d309628ab62360d85cec1dc..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/lr_scheduler.py +++ /dev/null @@ -1,53 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import numpy as np - - -class BaseScheduler(object): - - def schedule(self, n, **kwargs): - raise NotImplementedError - - -class LambdaWarmUpCosineFactorScheduler(BaseScheduler): - """ - note: use with a base_lr of 1.0 - """ - def __init__(self, warm_up_steps, f_min, f_max, f_start, max_decay_steps, verbosity_interval=0, **ignore_kwargs): - self.lr_warm_up_steps = warm_up_steps - self.f_start = f_start - self.f_min = f_min - self.f_max = f_max - self.lr_max_decay_steps = max_decay_steps - self.last_f = 0. - self.verbosity_interval = verbosity_interval - - def schedule(self, n, **kwargs): - if self.verbosity_interval > 0: - if n % self.verbosity_interval == 0: - print(f"current step: {n}, recent lr-multiplier: {self.f_start}") - if n < self.lr_warm_up_steps: - f = (self.f_max - self.f_start) / self.lr_warm_up_steps * n + self.f_start - self.last_f = f - return f - else: - t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps) - t = min(t, 1.0) - f = self.f_min + 0.5 * (self.f_max - self.f_min) * (1 + np.cos(t * np.pi)) - self.last_f = f - return f - - def __call__(self, n, **kwargs): - return self.schedule(n, **kwargs) diff --git a/hy3dshape/hy3dshape/utils/trainings/mesh.py b/hy3dshape/hy3dshape/utils/trainings/mesh.py deleted file mode 100644 index 0af7da6256ecde08c49e7a2bc88d0a22e8e67fc7..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/mesh.py +++ /dev/null @@ -1,128 +0,0 @@ -# -*- coding: utf-8 -*- - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -import cv2 -import numpy as np -import PIL.Image -from typing import Optional - -import trimesh - - -def save_obj(pointnp_px3, facenp_fx3, fname): - fid = open(fname, "w") - write_str = "" - for pidx, p in enumerate(pointnp_px3): - pp = p - write_str += "v %f %f %f\n" % (pp[0], pp[1], pp[2]) - - for i, f in enumerate(facenp_fx3): - f1 = f + 1 - write_str += "f %d %d %d\n" % (f1[0], f1[1], f1[2]) - fid.write(write_str) - fid.close() - return - - -def savemeshtes2(pointnp_px3, tcoords_px2, facenp_fx3, facetex_fx3, tex_map, fname): - fol, na = os.path.split(fname) - na, _ = os.path.splitext(na) - - matname = "%s/%s.mtl" % (fol, na) - fid = open(matname, "w") - fid.write("newmtl material_0\n") - fid.write("Kd 1 1 1\n") - fid.write("Ka 0 0 0\n") - fid.write("Ks 0.4 0.4 0.4\n") - fid.write("Ns 10\n") - fid.write("illum 2\n") - fid.write("map_Kd %s.png\n" % na) - fid.close() - #### - - fid = open(fname, "w") - fid.write("mtllib %s.mtl\n" % na) - - for pidx, p3 in enumerate(pointnp_px3): - pp = p3 - fid.write("v %f %f %f\n" % (pp[0], pp[1], pp[2])) - - for pidx, p2 in enumerate(tcoords_px2): - pp = p2 - fid.write("vt %f %f\n" % (pp[0], pp[1])) - - fid.write("usemtl material_0\n") - for i, f in enumerate(facenp_fx3): - f1 = f + 1 - f2 = facetex_fx3[i] + 1 - fid.write("f %d/%d %d/%d %d/%d\n" % (f1[0], f2[0], f1[1], f2[1], f1[2], f2[2])) - fid.close() - - PIL.Image.fromarray(np.ascontiguousarray(tex_map), "RGB").save( - os.path.join(fol, "%s.png" % na)) - - return - - -class MeshOutput(object): - - def __init__(self, - mesh_v: np.ndarray, - mesh_f: np.ndarray, - vertex_colors: Optional[np.ndarray] = None, - uvs: Optional[np.ndarray] = None, - mesh_tex_idx: Optional[np.ndarray] = None, - tex_map: Optional[np.ndarray] = None): - - self.mesh_v = mesh_v - self.mesh_f = mesh_f - self.vertex_colors = vertex_colors - self.uvs = uvs - self.mesh_tex_idx = mesh_tex_idx - self.tex_map = tex_map - - def contain_uv_texture(self): - return (self.uvs is not None) and (self.mesh_tex_idx is not None) and (self.tex_map is not None) - - def contain_vertex_colors(self): - return self.vertex_colors is not None - - def export(self, fname): - - if self.contain_uv_texture(): - savemeshtes2( - self.mesh_v, - self.uvs, - self.mesh_f, - self.mesh_tex_idx, - self.tex_map, - fname - ) - - elif self.contain_vertex_colors(): - mesh_obj = trimesh.Trimesh(vertices=self.mesh_v, faces=self.mesh_f, vertex_colors=self.vertex_colors) - mesh_obj.export(fname) - - else: - save_obj( - self.mesh_v, - self.mesh_f, - fname - ) - - - diff --git a/hy3dshape/hy3dshape/utils/trainings/mesh_log_callback.py b/hy3dshape/hy3dshape/utils/trainings/mesh_log_callback.py deleted file mode 100644 index f2466d581d75d4fbe923cdf91cbb8ddb919af456..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/mesh_log_callback.py +++ /dev/null @@ -1,336 +0,0 @@ -# -*- coding: utf-8 -*- - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import json -import math -import os -from typing import Tuple, Generic, Dict, List, Union, Optional - -import trimesh -import numpy as np -import pytorch_lightning as pl -import pytorch_lightning.loggers -import torch -import torchvision -from pytorch_lightning.callbacks import Callback -from pytorch_lightning.utilities import rank_zero_only - -from hy3dshape.pipelines import export_to_trimesh -from hy3dshape.utils.trainings.mesh import MeshOutput -from hy3dshape.utils.visualizers import html_util -from hy3dshape.utils.visualizers.pythreejs_viewer import PyThreeJSViewer - - -class ImageConditionalASLDiffuserLogger(Callback): - def __init__(self, - step_frequency: int, - num_samples: int = 1, - mean: Optional[Union[List[float], Tuple[float]]] = None, - std: Optional[Union[List[float], Tuple[float]]] = None, - bounds: Union[List[float], Tuple[float]] = (-1.1, -1.1, -1.1, 1.1, 1.1, 1.1), - **kwargs) -> None: - - super().__init__() - self.bbox_size = np.array(bounds[3:6]) - np.array(bounds[0:3]) - - if mean is not None: - mean = np.asarray(mean) - - if std is not None: - std = np.asarray(std) - - self.mean = mean - self.std = std - - self.step_freq = step_frequency - self.num_samples = num_samples - self.has_train_logged = False - self.logger_log_images = { - pl.loggers.WandbLogger: self._wandb, - } - - self.viewer = PyThreeJSViewer(settings={}, render_mode="WEBSITE") - - @rank_zero_only - def _wandb(self, pl_module, images, batch_idx, split): - # raise ValueError("No way wandb") - grids = dict() - for k in images: - grid = torchvision.utils.make_grid(images[k]) - grids[f"{split}/{k}"] = wandb.Image(grid) - pl_module.logger.experiment.log(grids) - - def log_local(self, - outputs: List[List['Latent2MeshOutput']], - images: Union[np.ndarray, List[np.ndarray]], - description: List[str], - keys: List[str], - save_dir: str, split: str, - global_step: int, current_epoch: int, batch_idx: int, - prog_bar: bool = False, - multi_views=None, # yf ... - ) -> None: - - folder = "gs-{:010}_e-{:06}_b-{:06}".format(global_step, current_epoch, batch_idx) - visual_dir = os.path.join(save_dir, "visuals", split, folder) - os.makedirs(visual_dir, exist_ok=True) - - num_samples = len(images) - - for i in range(num_samples): - key_i = keys[i] - image_i = self.denormalize_image(images[i]) - shape_tag_i = description[i] - - for j in range(1): - mesh = outputs[j][i] - if mesh is None: - continue - - mesh_v = mesh.mesh_v.copy() - mesh_v[:, 0] += j * np.max(self.bbox_size) - self.viewer.add_mesh(mesh_v, mesh.mesh_f) - - image_tag = html_util.to_image_embed_tag(image_i) - mesh_tag = self.viewer.to_html(html_frame=False) - - table_tag = f""" - - - - - - - -
{shape_tag_i} - {key_i} Input Image | Generated Mesh
{image_tag}{mesh_tag}
- """ - - if multi_views is not None: - multi_views_i = self.make_grid(multi_views[i]) - views_tag = html_util.to_image_embed_tag(self.denormalize_image(multi_views_i)) - table_tag = f""" - - - - - - - - -
{shape_tag_i} - {key_i} Input Image | Generated Mesh
{image_tag}{views_tag}{mesh_tag}
- """ - - html_frame = html_util.to_html_frame(table_tag) - if len(key_i) > 100: - key_i = key_i[:100] - with open(os.path.join(visual_dir, f"{key_i}.html"), "w") as writer: - writer.write(html_frame) - - self.viewer.reset() - - def log_sample(self, - pl_module: pl.LightningModule, - batch: Dict[str, torch.FloatTensor], - batch_idx: int, - split: str = "train") -> None: - """ - - Args: - pl_module: - batch (dict): the batch sample information, and it contains: - - surface (torch.FloatTensor): - - image (torch.FloatTensor): - batch_idx (int): - split (str): - - Returns: - - """ - - is_train = pl_module.training - if is_train: - pl_module.eval() - - batch_size = len(batch["surface"]) - replace = batch_size < self.num_samples - ids = np.random.choice(batch_size, self.num_samples, replace=replace) - - with torch.no_grad(): - # run text to mesh - # keys = [batch["__key__"][i] for i in ids] - keys = [f'key_{i}' for i in ids] - # texts = [batch["text"][i] for i in ids] - texts = [f'text_{i}'for i in ids] - # description = [batch["description"][i] for i in ids] - description = [f'desc_{i}' for i in ids] - images = batch["image"][ids] - mask_input = batch["mask"][ids] if 'mask' in batch else None - sample_batch = { - "__key__": keys, - "image": images, - 'text': texts, - 'mask': mask_input, - } - - # if 'cam_parm' in batch: - # sample_batch['cam_parm'] = batch['cam_parm'][ids] - - # if 'multi_views' in batch: # yf ... - # sample_batch['multi_views'] = batch['multi_views'][ids] - - outputs = pl_module.sample( - batch=sample_batch, - output_type='latents2mesh' - ) - - images = images.cpu().float().numpy() - # images = self.denormalize_image(images) - # images = np.transpose(images, (0, 2, 3, 1)) - # images = ((images + 1) / 2 * 255).astype(np.uint8) - - self.log_local(outputs, images, description, keys, pl_module.logger.save_dir, split, - pl_module.global_step, pl_module.current_epoch, batch_idx, prog_bar=False, - multi_views=sample_batch.get('multi_views')) - - if is_train: pl_module.train() - - def make_grid(self, images): # return (3,h,w) in (0,1) ... - images_resized = [] - for img in images: - img_resized = torchvision.transforms.functional.resize(img, (320, 320)) - images_resized.append(img_resized) - image = torchvision.utils.make_grid(images_resized, nrow=2, padding=5, pad_value=255) - - image = image.cpu().numpy() - # image = np.transpose(image, (1, 2, 0)) - # image = (image * 255).astype(np.uint8) - - return image - - def check_frequency(self, step: int) -> bool: - if step % self.step_freq == 0: - return True - return False - - def on_train_batch_end(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule, - outputs: Generic, batch: Dict[str, torch.FloatTensor], batch_idx: int) -> None: - - if (self.check_frequency(pl_module.global_step) and # batch_idx % self.batch_freq == 0 - hasattr(pl_module, "sample") and - callable(pl_module.sample) and - self.num_samples > 0): - self.log_sample(pl_module, batch, batch_idx, split="train") - self.has_train_logged = True - - def on_validation_batch_end(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule, - outputs: Generic, batch: Dict[str, torch.FloatTensor], - dataloader_idx: int, batch_idx: int) -> None: - - if self.has_train_logged: - self.log_sample(pl_module, batch, batch_idx, split="val") - self.has_train_logged = False - - def denormalize_image(self, image): - """ - - Args: - image (np.ndarray): [3, h, w] - - Returns: - image (np.ndarray): [h, w, 3], np.uint8, [0, 255]. - """ - # image = np.transpose(image, (0, 2, 3, 1)) - image = np.transpose(image, (1, 2, 0)) - - if self.std is not None: - image = image * self.std - - if self.mean is not None: - image = image + self.mean - - image = (image * 255).astype(np.uint8) - - return image - - -class ImageConditionalFixASLDiffuserLogger(Callback): - def __init__( - self, - step_frequency: int, - test_data_path: str, - max_size: int = None, - save_dir: str = 'infer', - **kwargs, - ) -> None: - super().__init__() - self.step_freq = step_frequency - self.viewer = PyThreeJSViewer(settings={}, render_mode="WEBSITE") - - self.test_data_path = test_data_path - with open(self.test_data_path, 'r') as f: - data = json.load(f) - self.file_list = data['file_list'] - self.file_folder = data['file_folder'] - if max_size is not None: - self.file_list = self.file_list[:max_size] - self.kwargs = kwargs - self.save_dir = save_dir - - def on_train_batch_end( - self, - trainer: pl.trainer.Trainer, - pl_module: pl.LightningModule, - outputs: Generic, - batch: Dict[str, torch.FloatTensor], - batch_idx: int, - ): - if pl_module.global_step % self.step_freq == 0: - is_train = pl_module.training - if is_train: - pl_module.eval() - - folder_path = self.file_folder - folder_name = os.path.basename(folder_path) - folder = "gs-{:010}_e-{:06}_b-{:06}".format(pl_module.global_step, pl_module.current_epoch, batch_idx) - visual_dir = os.path.join(pl_module.logger.save_dir, self.save_dir, folder, folder_name) - os.makedirs(visual_dir, exist_ok=True) - - image_paths = self.file_list - chunk_size = math.ceil(len(image_paths) / trainer.world_size) - if pl_module.global_rank == trainer.world_size - 1: - image_paths = image_paths[pl_module.global_rank * chunk_size:] - else: - image_paths = image_paths[pl_module.global_rank * chunk_size:(pl_module.global_rank + 1) * chunk_size] - - print(f'Rank{pl_module.global_rank}: processing {len(image_paths)}|{len(self.file_list)} images') - for image_path in image_paths: - if folder_path in image_path: - save_path = image_path.replace(folder_path, visual_dir) - else: - save_path = os.path.join(visual_dir, os.path.basename(image_path)) - save_path = os.path.splitext(save_path)[0] + '.glb' - - print(image_path) - with torch.no_grad(): - mesh = pl_module.sample(batch={"image": image_path}, **self.kwargs)[0][0] - if isinstance(mesh, tuple) and len(mesh)==2: - mesh = export_to_trimesh(mesh) - elif isinstance(mesh, trimesh.Trimesh): - os.makedirs(os.path.dirname(save_path), exist_ok=True) - mesh.export(save_path) - - if is_train: - pl_module.train() diff --git a/hy3dshape/hy3dshape/utils/trainings/peft.py b/hy3dshape/hy3dshape/utils/trainings/peft.py deleted file mode 100644 index 7b999defbb3b74a63ea86f9d6bf8e7851ff760b9..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/trainings/peft.py +++ /dev/null @@ -1,78 +0,0 @@ -# -*- coding: utf-8 -*- - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import os -from pytorch_lightning.callbacks import Callback -from omegaconf import OmegaConf, ListConfig - -class PeftSaveCallback(Callback): - def __init__(self, peft_model, save_dir: str, save_every_n_steps: int = None): - super().__init__() - self.peft_model = peft_model - self.save_dir = save_dir - self.save_every_n_steps = save_every_n_steps - os.makedirs(self.save_dir, exist_ok=True) - - def recursive_convert(self, obj): - from omegaconf import OmegaConf, ListConfig - if isinstance(obj, (OmegaConf, ListConfig)): - return OmegaConf.to_container(obj, resolve=True) - elif isinstance(obj, dict): - return {k: self.recursive_convert(v) for k, v in obj.items()} - elif isinstance(obj, list): - return [self.recursive_convert(i) for i in obj] - elif isinstance(obj, type): - # 避免修改类对象 - return obj - elif hasattr(obj, '__dict__'): - for attr_name, attr_value in vars(obj).items(): - setattr(obj, attr_name, self.recursive_convert(attr_value)) - return obj - else: - return obj - - # def recursive_convert(self, obj): - # if isinstance(obj, (OmegaConf, ListConfig)): - # return OmegaConf.to_container(obj, resolve=True) - # elif isinstance(obj, dict): - # return {k: self.recursive_convert(v) for k, v in obj.items()} - # elif isinstance(obj, list): - # return [self.recursive_convert(i) for i in obj] - # elif hasattr(obj, '__dict__'): - # for attr_name, attr_value in vars(obj).items(): - # setattr(obj, attr_name, self.recursive_convert(attr_value)) - # return obj - # else: - # return obj - - def _convert_peft_config(self): - pc = self.peft_model.peft_config - self.peft_model.peft_config = self.recursive_convert(pc) - - def on_train_epoch_end(self, trainer, pl_module): - self._convert_peft_config() - save_path = os.path.join(self.save_dir, f"epoch_{trainer.current_epoch}") - self.peft_model.save_pretrained(save_path) - print(f"[PeftSaveCallback] Saved LoRA weights to {save_path}") - - def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx): - if self.save_every_n_steps is not None: - global_step = trainer.global_step - if global_step % self.save_every_n_steps == 0 and global_step > 0: - self._convert_peft_config() - save_path = os.path.join(self.save_dir, f"step_{global_step}") - self.peft_model.save_pretrained(save_path) - print(f"[PeftSaveCallback] Saved LoRA weights to {save_path}") diff --git a/hy3dshape/hy3dshape/utils/utils.py b/hy3dshape/hy3dshape/utils/utils.py deleted file mode 100644 index c88f6bb9a9e3f7aaf0ce68753757cd7816489915..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/utils.py +++ /dev/null @@ -1,128 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import logging -import os -from functools import wraps - -import torch - - -def get_logger(name): - logger = logging.getLogger(name) - logger.setLevel(logging.INFO) - - console_handler = logging.StreamHandler() - console_handler.setLevel(logging.INFO) - - formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') - console_handler.setFormatter(formatter) - logger.addHandler(console_handler) - return logger - - -logger = get_logger('hy3dgen.shapgen') - - -class synchronize_timer: - """ Synchronized timer to count the inference time of `nn.Module.forward`. - - Supports both context manager and decorator usage. - - Example as context manager: - ```python - with synchronize_timer('name') as t: - run() - ``` - - Example as decorator: - ```python - @synchronize_timer('Export to trimesh') - def export_to_trimesh(mesh_output): - pass - ``` - """ - - def __init__(self, name=None): - self.name = name - - def __enter__(self): - """Context manager entry: start timing.""" - if os.environ.get('HY3DGEN_DEBUG', '0') == '1': - self.start = torch.cuda.Event(enable_timing=True) - self.end = torch.cuda.Event(enable_timing=True) - self.start.record() - return lambda: self.time - - def __exit__(self, exc_type, exc_value, exc_tb): - """Context manager exit: stop timing and log results.""" - if os.environ.get('HY3DGEN_DEBUG', '0') == '1': - self.end.record() - torch.cuda.synchronize() - self.time = self.start.elapsed_time(self.end) - if self.name is not None: - logger.info(f'{self.name} takes {self.time} ms') - - def __call__(self, func): - """Decorator: wrap the function to time its execution.""" - - @wraps(func) - def wrapper(*args, **kwargs): - with self: - result = func(*args, **kwargs) - return result - - return wrapper - - -def smart_load_model( - model_path, - subfolder, - use_safetensors, - variant, -): - original_model_path = model_path - # try local path - base_dir = os.environ.get('HY3DGEN_MODELS', '~/.cache/hy3dgen') - model_fld = os.path.expanduser(os.path.join(base_dir, model_path)) - model_path = os.path.expanduser(os.path.join(base_dir, model_path, subfolder)) - logger.info(f'Try to load model from local path: {model_path}') - if not os.path.exists(model_path): - logger.info('Model path not exists, try to download from huggingface') - try: - from huggingface_hub import snapshot_download - # 只下载指定子目录 - path = snapshot_download( - repo_id=original_model_path, - allow_patterns=[f"{subfolder}/*"], # 关键修改：模式匹配子文件夹 - local_dir=model_fld - ) - model_path = os.path.join(path, subfolder) # 保持路径拼接逻辑不变 - except ImportError: - logger.warning( - "You need to install HuggingFace Hub to load models from the hub." - ) - raise RuntimeError(f"Model path {model_path} not found") - except Exception as e: - raise e - - if not os.path.exists(model_path): - raise FileNotFoundError(f"Model path {original_model_path} not found") - - extension = 'ckpt' if not use_safetensors else 'safetensors' - variant = '' if variant is None else f'.{variant}' - ckpt_name = f'model{variant}.{extension}' - config_path = os.path.join(model_path, 'config.yaml') - ckpt_path = os.path.join(model_path, ckpt_name) - return config_path, ckpt_path diff --git a/hy3dshape/hy3dshape/utils/visualizers/__init__.py b/hy3dshape/hy3dshape/utils/visualizers/__init__.py deleted file mode 100644 index 40a96afc6ff09d58a702b76e3f7dd412fe975e26..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/visualizers/__init__.py +++ /dev/null @@ -1 +0,0 @@ -# -*- coding: utf-8 -*- diff --git a/hy3dshape/hy3dshape/utils/visualizers/color_util.py b/hy3dshape/hy3dshape/utils/visualizers/color_util.py deleted file mode 100644 index 64f5ed382d657bcab3420c57f12845ba74bc9df7..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/visualizers/color_util.py +++ /dev/null @@ -1,57 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import numpy as np -import matplotlib.pyplot as plt - - -# Helper functions -def get_colors(inp, colormap="viridis", normalize=True, vmin=None, vmax=None): - colormap = plt.cm.get_cmap(colormap) - if normalize: - vmin = np.min(inp) - vmax = np.max(inp) - - norm = plt.Normalize(vmin, vmax) - return colormap(norm(inp))[:, :3] - - -def gen_checkers(n_checkers_x, n_checkers_y, width=256, height=256): - # tex dims need to be power of two. - array = np.ones((width, height, 3), dtype='float32') - - # width in texels of each checker - checker_w = width / n_checkers_x - checker_h = height / n_checkers_y - - for y in range(height): - for x in range(width): - color_key = int(x / checker_w) + int(y / checker_h) - if color_key % 2 == 0: - array[x, y, :] = [1., 0.874, 0.0] - else: - array[x, y, :] = [0., 0., 0.] - return array - - -def gen_circle(width=256, height=256): - xx, yy = np.mgrid[:width, :height] - circle = (xx - width / 2 + 0.5) ** 2 + (yy - height / 2 + 0.5) ** 2 - array = np.ones((width, height, 4), dtype='float32') - array[:, :, 0] = (circle <= width) - array[:, :, 1] = (circle <= width) - array[:, :, 2] = (circle <= width) - array[:, :, 3] = circle <= width - return array - diff --git a/hy3dshape/hy3dshape/utils/visualizers/html_util.py b/hy3dshape/hy3dshape/utils/visualizers/html_util.py deleted file mode 100644 index b31c403777e44b948a088e6a60331061525e7a80..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/visualizers/html_util.py +++ /dev/null @@ -1,64 +0,0 @@ -# -*- coding: utf-8 -*- - -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import io -import base64 -import numpy as np -from PIL import Image - - -def to_html_frame(content): - - html_frame = f""" - - - {content} - - - """ - - return html_frame - - -def to_single_row_table(caption: str, content: str): - - table_html = f""" - - - - - -
{caption}
{content}
- """ - - return table_html - - -def to_image_embed_tag(image: np.ndarray): - - # Convert np.ndarray to bytes - img = Image.fromarray(image) - raw_bytes = io.BytesIO() - img.save(raw_bytes, "PNG") - - # Encode bytes to base64 - image_base64 = base64.b64encode(raw_bytes.getvalue()).decode("utf-8") - - image_tag = f""" - Embedded Image - """ - - return image_tag diff --git a/hy3dshape/hy3dshape/utils/visualizers/pythreejs_viewer.py b/hy3dshape/hy3dshape/utils/visualizers/pythreejs_viewer.py deleted file mode 100644 index 817d2efaf74f376a419eae893996540fa643a38d..0000000000000000000000000000000000000000 --- a/hy3dshape/hy3dshape/utils/visualizers/pythreejs_viewer.py +++ /dev/null @@ -1,549 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - - -import numpy as np -from ipywidgets import embed -import pythreejs as p3s -import uuid - -from .color_util import get_colors, gen_circle, gen_checkers - - -EMBED_URL = "https://cdn.jsdelivr.net/npm/@jupyter-widgets/html-manager@1.0.1/dist/embed-amd.js" - - -class PyThreeJSViewer(object): - - def __init__(self, settings, render_mode="WEBSITE"): - self.render_mode = render_mode - self.__update_settings(settings) - self._light = p3s.DirectionalLight(color='white', position=[0, 0, 1], intensity=0.6) - self._light2 = p3s.AmbientLight(intensity=0.5) - self._cam = p3s.PerspectiveCamera(position=[0, 0, 1], lookAt=[0, 0, 0], fov=self.__s["fov"], - aspect=self.__s["width"] / self.__s["height"], children=[self._light]) - self._orbit = p3s.OrbitControls(controlling=self._cam) - self._scene = p3s.Scene(children=[self._cam, self._light2], background=self.__s["background"]) # "#4c4c80" - self._renderer = p3s.Renderer(camera=self._cam, scene=self._scene, controls=[self._orbit], - width=self.__s["width"], height=self.__s["height"], - antialias=self.__s["antialias"]) - - self.__objects = {} - self.__cnt = 0 - - def jupyter_mode(self): - self.render_mode = "JUPYTER" - - def offline(self): - self.render_mode = "OFFLINE" - - def website(self): - self.render_mode = "WEBSITE" - - def __get_shading(self, shading): - shad = {"flat": True, "wireframe": False, "wire_width": 0.03, "wire_color": "black", - "side": 'DoubleSide', "colormap": "viridis", "normalize": [None, None], - "bbox": False, "roughness": 0.5, "metalness": 0.25, "reflectivity": 1.0, - "line_width": 1.0, "line_color": "black", - "point_color": "red", "point_size": 0.01, "point_shape": "circle", - "text_color": "red" - } - for k in shading: - shad[k] = shading[k] - return shad - - def __update_settings(self, settings={}): - sett = {"width": 1600, "height": 800, "antialias": True, "scale": 1.5, "background": "#ffffff", - "fov": 30} - for k in settings: - sett[k] = settings[k] - self.__s = sett - - def __add_object(self, obj, parent=None): - if not parent: # Object is added to global scene and objects dict - self.__objects[self.__cnt] = obj - self.__cnt += 1 - self._scene.add(obj["mesh"]) - else: # Object is added to parent object and NOT to objects dict - parent.add(obj["mesh"]) - - self.__update_view() - - if self.render_mode == "JUPYTER": - return self.__cnt - 1 - elif self.render_mode == "WEBSITE": - return self - - def __add_line_geometry(self, lines, shading, obj=None): - lines = lines.astype("float32", copy=False) - mi = np.min(lines, axis=0) - ma = np.max(lines, axis=0) - - geometry = p3s.LineSegmentsGeometry(positions=lines.reshape((-1, 2, 3))) - material = p3s.LineMaterial(linewidth=shading["line_width"], color=shading["line_color"]) - # , vertexColors='VertexColors'), - lines = p3s.LineSegments2(geometry=geometry, material=material) # type='LinePieces') - line_obj = {"geometry": geometry, "mesh": lines, "material": material, - "max": ma, "min": mi, "type": "Lines", "wireframe": None} - - if obj: - return self.__add_object(line_obj, obj), line_obj - else: - return self.__add_object(line_obj) - - def __update_view(self): - if len(self.__objects) == 0: - return - ma = np.zeros((len(self.__objects), 3)) - mi = np.zeros((len(self.__objects), 3)) - for r, obj in enumerate(self.__objects): - ma[r] = self.__objects[obj]["max"] - mi[r] = self.__objects[obj]["min"] - ma = np.max(ma, axis=0) - mi = np.min(mi, axis=0) - diag = np.linalg.norm(ma - mi) - mean = ((ma - mi) / 2 + mi).tolist() - scale = self.__s["scale"] * (diag) - self._orbit.target = mean - self._cam.lookAt(mean) - self._cam.position = [mean[0], mean[1], mean[2] + scale] - self._light.position = [mean[0], mean[1], mean[2] + scale] - - self._orbit.exec_three_obj_method('update') - self._cam.exec_three_obj_method('updateProjectionMatrix') - - def __get_bbox(self, v): - m = np.min(v, axis=0) - M = np.max(v, axis=0) - - # Corners of the bounding box - v_box = np.array([[m[0], m[1], m[2]], [M[0], m[1], m[2]], [M[0], M[1], m[2]], [m[0], M[1], m[2]], - [m[0], m[1], M[2]], [M[0], m[1], M[2]], [M[0], M[1], M[2]], [m[0], M[1], M[2]]]) - - f_box = np.array([[0, 1], [1, 2], [2, 3], [3, 0], [4, 5], [5, 6], [6, 7], [7, 4], - [0, 4], [1, 5], [2, 6], [7, 3]], dtype=np.uint32) - return v_box, f_box - - def __get_colors(self, v, f, c, sh): - coloring = "VertexColors" - if type(c) == np.ndarray and c.size == 3: # Single color - colors = np.ones_like(v) - colors[:, 0] = c[0] - colors[:, 1] = c[1] - colors[:, 2] = c[2] - # print("Single colors") - elif type(c) == np.ndarray and len(c.shape) == 2 and c.shape[1] == 3: # Color values for - if c.shape[0] == f.shape[0]: # faces - colors = np.hstack([c, c, c]).reshape((-1, 3)) - coloring = "FaceColors" - # print("Face color values") - elif c.shape[0] == v.shape[0]: # vertices - colors = c - # print("Vertex color values") - else: # Wrong size, fallback - print("Invalid color array given! Supported are numpy arrays.", type(c)) - colors = np.ones_like(v) - colors[:, 0] = 1.0 - colors[:, 1] = 0.874 - colors[:, 2] = 0.0 - elif type(c) == np.ndarray and c.size == f.shape[0]: # Function values for faces - normalize = sh["normalize"][0] != None and sh["normalize"][1] != None - cc = get_colors(c, sh["colormap"], normalize=normalize, - vmin=sh["normalize"][0], vmax=sh["normalize"][1]) - # print(cc.shape) - colors = np.hstack([cc, cc, cc]).reshape((-1, 3)) - coloring = "FaceColors" - # print("Face function values") - elif type(c) == np.ndarray and c.size == v.shape[0]: # Function values for vertices - normalize = sh["normalize"][0] != None and sh["normalize"][1] != None - colors = get_colors(c, sh["colormap"], normalize=normalize, - vmin=sh["normalize"][0], vmax=sh["normalize"][1]) - # print("Vertex function values") - - else: - colors = np.ones_like(v) - colors[:, 0] = 1.0 - colors[:, 1] = 0.874 - colors[:, 2] = 0.0 - - # No color - if c is not None: - print("Invalid color array given! Supported are numpy arrays.", type(c)) - - return colors, coloring - - def __get_point_colors(self, v, c, sh): - v_color = True - if c is None: # No color given, use global color - # conv = mpl.colors.ColorConverter() - colors = sh["point_color"] # np.array(conv.to_rgb(sh["point_color"])) - v_color = False - elif isinstance(c, str): # No color given, use global color - # conv = mpl.colors.ColorConverter() - colors = c # np.array(conv.to_rgb(c)) - v_color = False - elif type(c) == np.ndarray and len(c.shape) == 2 and c.shape[0] == v.shape[0] and c.shape[1] == 3: - # Point color - colors = c.astype("float32", copy=False) - - elif isinstance(c, np.ndarray) and len(c.shape) == 2 and c.shape[0] == v.shape[0] and c.shape[1] != 3: - # Function values for vertices, but the colors are features - c_norm = np.linalg.norm(c, ord=2, axis=-1) - normalize = sh["normalize"][0] != None and sh["normalize"][1] != None - colors = get_colors(c_norm, sh["colormap"], normalize=normalize, - vmin=sh["normalize"][0], vmax=sh["normalize"][1]) - colors = colors.astype("float32", copy=False) - - elif type(c) == np.ndarray and c.size == v.shape[0]: # Function color - normalize = sh["normalize"][0] != None and sh["normalize"][1] != None - colors = get_colors(c, sh["colormap"], normalize=normalize, - vmin=sh["normalize"][0], vmax=sh["normalize"][1]) - colors = colors.astype("float32", copy=False) - # print("Vertex function values") - - else: - print("Invalid color array given! Supported are numpy arrays.", type(c)) - colors = sh["point_color"] - v_color = False - - return colors, v_color - - def add_mesh(self, v, f, c=None, uv=None, n=None, shading={}, texture_data=None, **kwargs): - shading.update(kwargs) - sh = self.__get_shading(shading) - mesh_obj = {} - - # it is a tet - if v.shape[1] == 3 and f.shape[1] == 4: - f_tmp = np.ndarray([f.shape[0] * 4, 3], dtype=f.dtype) - for i in range(f.shape[0]): - f_tmp[i * 4 + 0] = np.array([f[i][1], f[i][0], f[i][2]]) - f_tmp[i * 4 + 1] = np.array([f[i][0], f[i][1], f[i][3]]) - f_tmp[i * 4 + 2] = np.array([f[i][1], f[i][2], f[i][3]]) - f_tmp[i * 4 + 3] = np.array([f[i][2], f[i][0], f[i][3]]) - f = f_tmp - - if v.shape[1] == 2: - v = np.append(v, np.zeros([v.shape[0], 1]), 1) - - # Type adjustment vertices - v = v.astype("float32", copy=False) - - # Color setup - colors, coloring = self.__get_colors(v, f, c, sh) - - # Type adjustment faces and colors - c = colors.astype("float32", copy=False) - - # Material and geometry setup - ba_dict = {"color": p3s.BufferAttribute(c)} - if coloring == "FaceColors": - verts = np.zeros((f.shape[0] * 3, 3), dtype="float32") - for ii in range(f.shape[0]): - # print(ii*3, f[ii]) - verts[ii * 3] = v[f[ii, 0]] - verts[ii * 3 + 1] = v[f[ii, 1]] - verts[ii * 3 + 2] = v[f[ii, 2]] - v = verts - else: - f = f.astype("uint32", copy=False).ravel() - ba_dict["index"] = p3s.BufferAttribute(f, normalized=False) - - ba_dict["position"] = p3s.BufferAttribute(v, normalized=False) - - if uv is not None: - uv = (uv - np.min(uv)) / (np.max(uv) - np.min(uv)) - if texture_data is None: - texture_data = gen_checkers(20, 20) - tex = p3s.DataTexture(data=texture_data, format="RGBFormat", type="FloatType") - material = p3s.MeshStandardMaterial(map=tex, reflectivity=sh["reflectivity"], side=sh["side"], - roughness=sh["roughness"], metalness=sh["metalness"], - flatShading=sh["flat"], - polygonOffset=True, polygonOffsetFactor=1, polygonOffsetUnits=5) - ba_dict["uv"] = p3s.BufferAttribute(uv.astype("float32", copy=False)) - else: - material = p3s.MeshStandardMaterial(vertexColors=coloring, reflectivity=sh["reflectivity"], - side=sh["side"], roughness=sh["roughness"], metalness=sh["metalness"], - flatShading=sh["flat"], - polygonOffset=True, polygonOffsetFactor=1, polygonOffsetUnits=5) - - if type(n) != type(None) and coloring == "VertexColors": # TODO: properly handle normals for FaceColors as well - ba_dict["normal"] = p3s.BufferAttribute(n.astype("float32", copy=False), normalized=True) - - geometry = p3s.BufferGeometry(attributes=ba_dict) - - if coloring == "VertexColors" and type(n) == type(None): - geometry.exec_three_obj_method('computeVertexNormals') - elif coloring == "FaceColors" and type(n) == type(None): - geometry.exec_three_obj_method('computeFaceNormals') - - # Mesh setup - mesh = p3s.Mesh(geometry=geometry, material=material) - - # Wireframe setup - mesh_obj["wireframe"] = None - if sh["wireframe"]: - wf_geometry = p3s.WireframeGeometry(mesh.geometry) # WireframeGeometry - wf_material = p3s.LineBasicMaterial(color=sh["wire_color"], linewidth=sh["wire_width"]) - wireframe = p3s.LineSegments(wf_geometry, wf_material) - mesh.add(wireframe) - mesh_obj["wireframe"] = wireframe - - # Bounding box setup - if sh["bbox"]: - v_box, f_box = self.__get_bbox(v) - _, bbox = self.add_edges(v_box, f_box, sh, mesh) - mesh_obj["bbox"] = [bbox, v_box, f_box] - - # Object setup - mesh_obj["max"] = np.max(v, axis=0) - mesh_obj["min"] = np.min(v, axis=0) - mesh_obj["geometry"] = geometry - mesh_obj["mesh"] = mesh - mesh_obj["material"] = material - mesh_obj["type"] = "Mesh" - mesh_obj["shading"] = sh - mesh_obj["coloring"] = coloring - mesh_obj["arrays"] = [v, f, c] # TODO replays with proper storage or remove if not needed - - return self.__add_object(mesh_obj) - - def add_lines(self, beginning, ending, shading={}, obj=None, **kwargs): - shading.update(kwargs) - if len(beginning.shape) == 1: - if len(beginning) == 2: - beginning = np.array([[beginning[0], beginning[1], 0]]) - else: - if beginning.shape[1] == 2: - beginning = np.append( - beginning, np.zeros([beginning.shape[0], 1]), 1) - if len(ending.shape) == 1: - if len(ending) == 2: - ending = np.array([[ending[0], ending[1], 0]]) - else: - if ending.shape[1] == 2: - ending = np.append( - ending, np.zeros([ending.shape[0], 1]), 1) - - sh = self.__get_shading(shading) - lines = np.hstack([beginning, ending]) - lines = lines.reshape((-1, 3)) - return self.__add_line_geometry(lines, sh, obj) - - def add_edges(self, vertices, edges, shading={}, obj=None, **kwargs): - shading.update(kwargs) - if vertices.shape[1] == 2: - vertices = np.append( - vertices, np.zeros([vertices.shape[0], 1]), 1) - sh = self.__get_shading(shading) - lines = np.zeros((edges.size, 3)) - cnt = 0 - for e in edges: - lines[cnt, :] = vertices[e[0]] - lines[cnt + 1, :] = vertices[e[1]] - cnt += 2 - return self.__add_line_geometry(lines, sh, obj) - - def add_points(self, points, c=None, shading={}, obj=None, **kwargs): - shading.update(kwargs) - if len(points.shape) == 1: - if len(points) == 2: - points = np.array([[points[0], points[1], 0]]) - else: - if points.shape[1] == 2: - points = np.append( - points, np.zeros([points.shape[0], 1]), 1) - sh = self.__get_shading(shading) - points = points.astype("float32", copy=False) - mi = np.min(points, axis=0) - ma = np.max(points, axis=0) - - g_attributes = {"position": p3s.BufferAttribute(points, normalized=False)} - m_attributes = {"size": sh["point_size"]} - - if sh["point_shape"] == "circle": # Plot circles - tex = p3s.DataTexture(data=gen_circle(16, 16), format="RGBAFormat", type="FloatType") - m_attributes["map"] = tex - m_attributes["alphaTest"] = 0.5 - m_attributes["transparency"] = True - else: # Plot squares - pass - - colors, v_colors = self.__get_point_colors(points, c, sh) - if v_colors: # Colors per point - m_attributes["vertexColors"] = 'VertexColors' - g_attributes["color"] = p3s.BufferAttribute(colors, normalized=False) - - else: # Colors for all points - m_attributes["color"] = colors - - material = p3s.PointsMaterial(**m_attributes) - geometry = p3s.BufferGeometry(attributes=g_attributes) - points = p3s.Points(geometry=geometry, material=material) - point_obj = {"geometry": geometry, "mesh": points, "material": material, - "max": ma, "min": mi, "type": "Points", "wireframe": None} - - if obj: - return self.__add_object(point_obj, obj), point_obj - else: - return self.__add_object(point_obj) - - def remove_object(self, obj_id): - if obj_id not in self.__objects: - print("Invalid object id. Valid ids are: ", list(self.__objects.keys())) - return - self._scene.remove(self.__objects[obj_id]["mesh"]) - del self.__objects[obj_id] - self.__update_view() - - def reset(self): - for obj_id in list(self.__objects.keys()).copy(): - self._scene.remove(self.__objects[obj_id]["mesh"]) - del self.__objects[obj_id] - self.__update_view() - - def update_object(self, oid=0, vertices=None, colors=None, faces=None): - obj = self.__objects[oid] - if type(vertices) != type(None): - if obj["coloring"] == "FaceColors": - f = obj["arrays"][1] - verts = np.zeros((f.shape[0] * 3, 3), dtype="float32") - for ii in range(f.shape[0]): - # print(ii*3, f[ii]) - verts[ii * 3] = vertices[f[ii, 0]] - verts[ii * 3 + 1] = vertices[f[ii, 1]] - verts[ii * 3 + 2] = vertices[f[ii, 2]] - v = verts - - else: - v = vertices.astype("float32", copy=False) - obj["geometry"].attributes["position"].array = v - # self.wireframe.attributes["position"].array = v # Wireframe updates? - obj["geometry"].attributes["position"].needsUpdate = True - # obj["geometry"].exec_three_obj_method('computeVertexNormals') - if type(colors) != type(None): - colors, coloring = self.__get_colors(obj["arrays"][0], obj["arrays"][1], colors, obj["shading"]) - colors = colors.astype("float32", copy=False) - obj["geometry"].attributes["color"].array = colors - obj["geometry"].attributes["color"].needsUpdate = True - if type(faces) != type(None): - if obj["coloring"] == "FaceColors": - print("Face updates are currently only possible in vertex color mode.") - return - f = faces.astype("uint32", copy=False).ravel() - print(obj["geometry"].attributes) - obj["geometry"].attributes["index"].array = f - # self.wireframe.attributes["position"].array = v # Wireframe updates? - obj["geometry"].attributes["index"].needsUpdate = True - # obj["geometry"].exec_three_obj_method('computeVertexNormals') - # self.mesh.geometry.verticesNeedUpdate = True - # self.mesh.geometry.elementsNeedUpdate = True - # self.update() - if self.render_mode == "WEBSITE": - return self - - # def update(self): - # self.mesh.exec_three_obj_method('update') - # self.orbit.exec_three_obj_method('update') - # self.cam.exec_three_obj_method('updateProjectionMatrix') - # self.scene.exec_three_obj_method('update') - - def add_text(self, text, shading={}, **kwargs): - shading.update(kwargs) - sh = self.__get_shading(shading) - tt = p3s.TextTexture(string=text, color=sh["text_color"]) - sm = p3s.SpriteMaterial(map=tt) - text = p3s.Sprite(material=sm, scaleToTexture=True) - self._scene.add(text) - - # def add_widget(self, widget, callback): - # self.widgets.append(widget) - # widget.observe(callback, names='value') - - # def add_dropdown(self, options, default, desc, cb): - # widget = widgets.Dropdown(options=options, value=default, description=desc) - # self.__widgets.append(widget) - # widget.observe(cb, names="value") - # display(widget) - - # def add_button(self, text, cb): - # button = widgets.Button(description=text) - # self.__widgets.append(button) - # button.on_click(cb) - # display(button) - - def to_html(self, imports=True, html_frame=True): - # Bake positions (fixes centering bug in offline rendering) - if len(self.__objects) == 0: - return - ma = np.zeros((len(self.__objects), 3)) - mi = np.zeros((len(self.__objects), 3)) - for r, obj in enumerate(self.__objects): - ma[r] = self.__objects[obj]["max"] - mi[r] = self.__objects[obj]["min"] - ma = np.max(ma, axis=0) - mi = np.min(mi, axis=0) - diag = np.linalg.norm(ma - mi) - mean = (ma - mi) / 2 + mi - for r, obj in enumerate(self.__objects): - v = self.__objects[obj]["geometry"].attributes["position"].array - v -= mean - # v += np.array([0.0, .9, 0.0]) #! to move the obj to the center of window - - scale = self.__s["scale"] * (diag) - self._orbit.target = [0.0, 0.0, 0.0] - self._cam.lookAt([0.0, 0.0, 0.0]) - # self._cam.position = [0.0, 0.0, scale] - self._cam.position = [0.0, 0.5, scale * 1.3] #! show four complete meshes in the window - self._light.position = [0.0, 0.0, scale] - - state = embed.dependency_state(self._renderer) - - # Somehow these entries are missing when the state is exported in python. - # Exporting from the GUI works, so we are inserting the missing entries. - for k in state: - if state[k]["model_name"] == "OrbitControlsModel": - state[k]["state"]["maxAzimuthAngle"] = "inf" - state[k]["state"]["maxDistance"] = "inf" - state[k]["state"]["maxZoom"] = "inf" - state[k]["state"]["minAzimuthAngle"] = "-inf" - - tpl = embed.load_requirejs_template - if not imports: - embed.load_requirejs_template = "" - - s = embed.embed_snippet(self._renderer, state=state, embed_url=EMBED_URL) - # s = embed.embed_snippet(self.__w, state=state) - embed.load_requirejs_template = tpl - - if html_frame: - s = "\n\n" + s + "\n\n" - - # Revert changes - for r, obj in enumerate(self.__objects): - v = self.__objects[obj]["geometry"].attributes["position"].array - v += mean - self.__update_view() - - return s - - def save(self, filename=""): - if filename == "": - uid = str(uuid.uuid4()) + ".html" - else: - filename = filename.replace(".html", "") - uid = filename + '.html' - with open(uid, "w") as f: - f.write(self.to_html()) - print("Plot saved to file %s." % uid) diff --git a/hy3dshape/main.py b/hy3dshape/main.py deleted file mode 100644 index 4c1ae51f61e2578dad8ee93e99c2df150968cc73..0000000000000000000000000000000000000000 --- a/hy3dshape/main.py +++ /dev/null @@ -1,209 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from typing import Tuple, List -import warnings -warnings.filterwarnings("ignore") - -import os -import torch -import argparse -from pathlib import Path -from omegaconf import OmegaConf, DictConfig -from einops._torch_specific import allow_ops_in_compiled_graph # requires einops>=0.6.1 -allow_ops_in_compiled_graph() - -import pytorch_lightning as pl -from pytorch_lightning.callbacks import ModelCheckpoint, Callback -from pytorch_lightning.strategies import DDPStrategy, DeepSpeedStrategy -from pytorch_lightning.loggers import Logger, TensorBoardLogger -from pytorch_lightning.utilities import rank_zero_info - -from hy3dshape.utils import get_config_from_file, instantiate_from_config - - -class SetupCallback(Callback): - def __init__(self, config: DictConfig, basedir: Path, logdir: str = "log", ckptdir: str = "ckpt") -> None: - super().__init__() - self.logdir = basedir / logdir - self.ckptdir = basedir / ckptdir - self.config = config - - def on_fit_start(self, trainer: pl.trainer.Trainer, pl_module: pl.LightningModule) -> None: - if trainer.global_rank == 0: - os.makedirs(self.logdir, exist_ok=True) - os.makedirs(self.ckptdir, exist_ok=True) - - -def setup_callbacks(config: DictConfig) -> Tuple[List[Callback], Logger]: - training_cfg = config.training - basedir = Path(training_cfg.output_dir) - os.makedirs(basedir, exist_ok=True) - all_callbacks = [] - - setup_callback = SetupCallback(config, basedir) - all_callbacks.append(setup_callback) - - checkpoint_callback = ModelCheckpoint( - dirpath=setup_callback.ckptdir, - filename="ckpt-{step:08d}", - monitor=training_cfg.monitor, - mode="max", - save_top_k=-1, - verbose=False, - every_n_train_steps=training_cfg.every_n_train_steps) - all_callbacks.append(checkpoint_callback) - - if "callbacks" in config: - for key, value in config['callbacks'].items(): - custom_callback = instantiate_from_config(value) - all_callbacks.append(custom_callback) - - logger = TensorBoardLogger(save_dir=str(setup_callback.logdir), name="tensorboard") - - return all_callbacks, logger - - -def merge_cfg(cfg, arg_cfg): - for key in arg_cfg.keys(): - if key in cfg.training: - arg_cfg[key] = cfg.training[key] - cfg.training = DictConfig(arg_cfg) - return cfg - - -def get_args(): - parser = argparse.ArgumentParser() - parser.add_argument("--fast", action='store_true') - parser.add_argument("-c", "--config", type=str, required=True) - parser.add_argument("-s", "--seed", type=int, default=0) - parser.add_argument("-nn", "--num_nodes", type=int, default=1) - parser.add_argument("-ng", "--num_gpus", type=int, default=1) - parser.add_argument("-u", "--update_every", type=int, default=1) - parser.add_argument("-st", "--steps", type=int, default=50000000) - parser.add_argument("-lr", "--base_lr", type=float, default=4.5e-6) - parser.add_argument("-a", "--use_amp", default=False, action="store_true") - parser.add_argument("--amp_type", type=str, default="16") - parser.add_argument("--gradient_clip_val", type=float, default=None) - parser.add_argument("--gradient_clip_algorithm", type=str, default=None) - parser.add_argument("--every_n_train_steps", type=int, default=50000) - parser.add_argument("--log_every_n_steps", type=int, default=50) - parser.add_argument("--val_check_interval", type=int, default=1024) - parser.add_argument("--limit_val_batches", type=int, default=64) - parser.add_argument("--monitor", type=str, default="val/total_loss") - parser.add_argument("--output_dir", type=str, help="the output directory to save everything.") - parser.add_argument("--ckpt_path", type=str, default="", help="the restore checkpoints.") - parser.add_argument("--deepspeed", default=False, action="store_true") - parser.add_argument("--deepspeed2", default=False, action="store_true") - parser.add_argument("--scale_lr", type=bool, nargs="?", const=True, default=False, - help="scale base-lr by ngpu * batch_size * n_accumulate") - return parser.parse_args() - - -if __name__ == "__main__": - - args = get_args() - - if args.fast: - torch.backends.cudnn.allow_tf32 = True - torch.backends.cuda.matmul.allow_tf32 = True - torch.set_float32_matmul_precision('medium') - torch.utils.data._utils.MP_STATUS_CHECK_INTERVAL = 0.05 - - # Set random seed - pl.seed_everything(args.seed, workers=True) - - # Load configuration - config = get_config_from_file(args.config) - config = merge_cfg(config, vars(args)) - training_cfg = config.training - - # print config - rank_zero_info("Begin to print configuration ...") - rank_zero_info(OmegaConf.to_yaml(config)) - rank_zero_info("Finish print ...") - - # Setup callbacks - callbacks, loggers = setup_callbacks(config) - - # Build data modules - data: pl.LightningDataModule = instantiate_from_config(config.dataset) - - # Build model - model: pl.LightningModule = instantiate_from_config(config.model) - - nodes = args.num_nodes - ngpus = args.num_gpus - base_lr = training_cfg.base_lr - accumulate_grad_batches = training_cfg.update_every - batch_size = config.dataset.params.batch_size - - if 'NNODES' in os.environ: - nodes = int(os.environ['NNODES']) - training_cfg.num_nodes = nodes - args.num_nodes = nodes - - if args.scale_lr: - model.learning_rate = accumulate_grad_batches * nodes * ngpus * batch_size * base_lr - info = f"Setting learning rate to {model.learning_rate:.2e} = {accumulate_grad_batches} (accumulate)" - info += f" * {nodes} (nodes) * {ngpus} (num_gpus) * {batch_size} (batchsize) * {base_lr:.2e} (base_lr)" - rank_zero_info(info) - else: - model.learning_rate = base_lr - rank_zero_info("++++ NOT USING LR SCALING ++++") - rank_zero_info(f"Setting learning rate to {model.learning_rate:.2e}") - - # Build trainer - if args.num_nodes > 1 or args.num_gpus > 1: - if args.deepspeed: - ddp_strategy = DeepSpeedStrategy(stage=1) - elif args.deepspeed2: - ddp_strategy = 'deepspeed_stage_2' - else: - ddp_strategy = DDPStrategy(find_unused_parameters=False, bucket_cap_mb=1500) - else: - ddp_strategy = None # 'auto' - - rank_zero_info(f'*' * 100) - if training_cfg.use_amp: - amp_type = training_cfg.amp_type - assert amp_type in ['bf16', '16', '32'], f"Invalid amp_type: {amp_type}" - rank_zero_info(f'Using {amp_type} precision') - else: - amp_type = 32 - rank_zero_info(f'Using 32 bit precision') - rank_zero_info(f'*' * 100) - - trainer = pl.Trainer( - max_steps=training_cfg.steps, - precision=amp_type, - callbacks=callbacks, - accelerator="gpu", - devices=training_cfg.num_gpus, - num_nodes=training_cfg.num_nodes, - strategy=ddp_strategy, - gradient_clip_val=training_cfg.get('gradient_clip_val'), - gradient_clip_algorithm=training_cfg.get('gradient_clip_algorithm'), - accumulate_grad_batches=args.update_every, - logger=loggers, - log_every_n_steps=training_cfg.log_every_n_steps, - val_check_interval=training_cfg.val_check_interval, - limit_val_batches=training_cfg.limit_val_batches, - check_val_every_n_epoch=None - ) - - # Train - if training_cfg.ckpt_path == '': - training_cfg.ckpt_path = None - trainer.fit(model, datamodule=data, ckpt_path=training_cfg.ckpt_path) diff --git a/hy3dshape/minimal_demo.py b/hy3dshape/minimal_demo.py deleted file mode 100644 index a4bfe8fe49e866e2d66b3a7b44cfd6dbfe21fd92..0000000000000000000000000000000000000000 --- a/hy3dshape/minimal_demo.py +++ /dev/null @@ -1,30 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -from PIL import Image - -from hy3dshape.rembg import BackgroundRemover -from hy3dshape.pipelines import Hunyuan3DDiTFlowMatchingPipeline - -model_path = 'tencent/Hunyuan3D-2.1' -pipeline_shapegen = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained(model_path) - -image_path = 'demos/demo.png' -image = Image.open(image_path).convert("RGBA") -if image.mode == 'RGB': - rembg = BackgroundRemover() - image = rembg(image) - -mesh = pipeline_shapegen(image=image)[0] -mesh.export('demo.glb') diff --git a/hy3dshape/minimal_vae_demo.py b/hy3dshape/minimal_vae_demo.py deleted file mode 100644 index 2b398cb7e37fa2a81b930a35b2bb70cd293707df..0000000000000000000000000000000000000000 --- a/hy3dshape/minimal_vae_demo.py +++ /dev/null @@ -1,51 +0,0 @@ -# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT -# except for the third-party components listed below. -# Hunyuan 3D does not impose any additional limitations beyond what is outlined -# in the repsective licenses of these third-party components. -# Users must comply with all terms and conditions of original licenses of these third-party -# components and must ensure that the usage of the third party components adheres to -# all relevant laws and regulations. - -# For avoidance of doubts, Hunyuan 3D means the large language models and -# their software and algorithms, including trained model weights, parameters (including -# optimizer states), machine-learning model code, inference-enabling code, training-enabling code, -# fine-tuning enabling code and other elements of the foregoing made publicly available -# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT. - -import torch - -from hy3dshape.surface_loaders import SharpEdgeSurfaceLoader -from hy3dshape.models.autoencoders import ShapeVAE -from hy3dshape.pipelines import export_to_trimesh - - -vae = ShapeVAE.from_pretrained( - 'tencent/Hunyuan3D-2.1', - use_safetensors=False, - variant='fp16', -) - - -loader = SharpEdgeSurfaceLoader( - num_sharp_points=0, - num_uniform_points=81920, -) -mesh_demo = 'demos/demo.glb' -surface = loader(mesh_demo).to('cuda', dtype=torch.float16) -print(surface.shape) - -latents = vae.encode(surface) -latents = vae.decode(latents) -mesh = vae.latents2mesh( - latents, - output_type='trimesh', - bounds=1.01, - mc_level=0.0, - num_chunks=20000, - octree_resolution=256, - mc_algo='mc', - enable_pbar=True -) - -mesh = export_to_trimesh(mesh)[0] -mesh.export('output.obj') diff --git a/hy3dshape/requirements.txt b/hy3dshape/requirements.txt deleted file mode 100644 index abdab84e043a82dc588767e52e925971dd183d4f..0000000000000000000000000000000000000000 --- a/hy3dshape/requirements.txt +++ /dev/null @@ -1,40 +0,0 @@ -ninja -pybind11 - -diffusers -einops -opencv-python -numpy -torch -transformers -torchvision -#taming-transformers-rom1504 -#ConfigArgParse -#ipdb -omegaconf - -#sentencepiece -tqdm - -# Mesh Processing -trimesh -pymeshlab -pygltflib -xatlas -#kornia -#facexlib - -# Training -accelerate -#pytorch_lightning -#scikit-learn -#scikit-image - -# Demo only -gradio -fastapi -uvicorn -rembg -onnxruntime -#gevent -#geventhttpclient \ No newline at end of file diff --git a/hy3dshape/scripts/train_deepspeed.sh b/hy3dshape/scripts/train_deepspeed.sh deleted file mode 100644 index ed6b7c7c3b8dded37891016fe6bcbab3fb9ba282..0000000000000000000000000000000000000000 --- a/hy3dshape/scripts/train_deepspeed.sh +++ /dev/null @@ -1,70 +0,0 @@ -# If: ImportError: /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.20' not found -# Do: ln /usr/local/gcc-8.3/lib64/libstdc++.so.6 -sf /usr/lib64/libstdc++.so.6 - -export NCCL_IB_TIMEOUT=24 -export NCCL_NVLS_ENABLE=0 -NET_TYPE="high" -if [[ "${NET_TYPE}" = "low" ]]; then - export NCCL_SOCKET_IFNAME=eth1 - export NCCL_IB_GID_INDEX=3 - export NCCL_IB_HCA=mlx5_2:1,mlx5_2:1 - export NCCL_IB_SL=3 - export NCCL_CHECK_DISABLE=1 - export NCCL_P2P_DISABLE=0 - export NCCL_LL_THRESHOLD=16384 - export NCCL_IB_CUDA_SUPPORT=1 -else - export NCCL_IB_GID_INDEX=3 - export NCCL_IB_SL=3 - export NCCL_CHECK_DISABLE=1 - export NCCL_P2P_DISABLE=0 - export NCCL_IB_DISABLE=0 - export NCCL_LL_THRESHOLD=16384 - export NCCL_IB_CUDA_SUPPORT=1 - export NCCL_SOCKET_IFNAME=bond1 - export UCX_NET_DEVICES=bond1 - export NCCL_IB_HCA=mlx5_bond_1,mlx5_bond_5,mlx5_bond_3,mlx5_bond_7,mlx5_bond_4,mlx5_bond_8,mlx5_bond_2,mlx5_bond_6 - export NCCL_COLLNET_ENABLE=0 - export SHARP_COLL_ENABLE_SAT=0 - export NCCL_NET_GDR_LEVEL=2 - export NCCL_IB_QPS_PER_CONNECTION=4 - export NCCL_IB_TC=160 - export NCCL_PXN_DISABLE=0 -fi -export NCCL_DEBUG=WARN - -node_num=$1 -node_rank=$2 -master_ip=$3 -config=$4 -output_dir=$5 - -# config='configs/dit-from-scratch-overfitting-flowmatching-dinog518-bf16-lr1e4-1024.yaml' -# output_dir='output_folder/dit/overfitting_10' - -echo node_num $node_num -echo node_rank $node_rank -echo master_ip $master_ip -echo config $config -echo output_dir $output_dir - -if test -d "$output_dir"; then - cp $config $output_dir -else - mkdir -p "$output_dir" - cp $config $output_dir -fi - -NODE_RANK=$node_rank \ -HF_HUB_OFFLINE=0 \ -MASTER_PORT=12348 \ -MASTER_ADDR=$master_ip \ -NCCL_SOCKET_IFNAME=bond1 \ -NCCL_IB_GID_INDEX=3 \ -NCCL_NVLS_ENABLE=0 \ -python3 main.py \ - --num_nodes $node_num \ - --num_gpus 8 \ - --config $config \ - --output_dir $output_dir \ - --deepspeed diff --git a/models/model_3d_generator.py b/models/model_3d_generator.py index c92c507b42442a38297592ccb0655434e8b577dd..c0ea6668ede30c056155a72c5ce7b92c6daeb90d 100644 --- a/models/model_3d_generator.py +++ b/models/model_3d_generator.py @@ -93,23 +93,35 @@ class Hunyuan3DGenerator: # Try to set up the model pipeline logger.info("🔧 Setting up Hunyuan3D pipeline...") - # Import necessary modules + # Import necessary modules using importlib for robustness import sys - sys.path.append(self.model_path) + import os + import importlib.util - # Try to import the main modules try: - from hy3dshape.infer import predict_shape - from hy3dpaint.infer import predict_texture - - self.predict_shape = predict_shape - self.predict_texture = predict_texture - self.model = "direct_model" + # Add the model path to sys.path to allow internal imports within Hunyuan3D code + if self.model_path not in sys.path: + sys.path.insert(0, self.model_path) + + # Dynamically load the modules from the downloaded paths + shape_infer_path = os.path.join(self.model_path, 'hy3dshape', 'infer.py') + paint_infer_path = os.path.join(self.model_path, 'hy3dpaint', 'infer.py') + + spec_shape = importlib.util.spec_from_file_location("hy3dshape.infer", shape_infer_path) + shape_module = importlib.util.module_from_spec(spec_shape) + spec_shape.loader.exec_module(shape_module) + self.predict_shape = shape_module.predict_shape - logger.info("✅ Hunyuan3D modules loaded successfully") + spec_paint = importlib.util.spec_from_file_location("hy3dpaint.infer", paint_infer_path) + paint_module = importlib.util.module_from_spec(spec_paint) + spec_paint.loader.exec_module(paint_module) + self.predict_texture = paint_module.predict_texture - except ImportError as e: - logger.warning(f"⚠️ Could not import Hunyuan3D modules directly: {e}") + self.model = "direct_model" + logger.info("✅ Hunyuan3D modules loaded successfully via importlib") + + except (ImportError, FileNotFoundError) as e: + logger.warning(f"⚠️ Could not import Hunyuan3D modules dynamically: {e}") logger.info("🔄 Using simplified implementation...") self.model = "simplified" diff --git a/test_3d_generation.py b/test_3d_generation.py new file mode 100644 index 0000000000000000000000000000000000000000..2d433a97e9c4cbd97ee50b5eee0a8b11081cd465 --- /dev/null +++ b/test_3d_generation.py @@ -0,0 +1,37 @@ +import sys +from pathlib import Path + +# Add the project root to the Python path +sys.path.insert(0, str(Path(__file__).resolve().parent)) + +import asyncio +from app import generate_monster +from core.ai_pipeline import MonsterGenerationPipeline + +class MockOAuthProfile: + def __init__(self, username): + self.username = username + +async def main(): + # Create a mock OAuth profile + mock_profile = MockOAuthProfile(username="test_user") + + # Call the generate_monster function with a text prompt + # This will trigger the 3D generation pipeline + image_path, model_path, stats, dialogue, message = await asyncio.to_thread( + generate_monster, + oauth_profile=mock_profile, + text_input="a cute, fluffy, blue monster" + ) + + # Print the results + print(f"Message: {message}") + print(f"Image Path: {image_path}") + print(f"3D Model Path: {model_path}") + print(f"Stats: {stats}") + print(f"Dialogue: {dialogue}") + +if __name__ == "__main__": + # The generate_monster function is a regular function, but since it can be run in a separate thread by Gradio, + # it's better to run it in an asyncio event loop to be safe. + asyncio.run(main())