File size: 38,985 Bytes
5db3c1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
#
# For licensing see accompanying LICENSE file.
# Copyright (C) 2025 Apple Inc. All Rights Reserved.
#
import io
import os
import csv
import json
import random
import torch
import numpy as np
import math
import time
import contextlib
from typing import Optional, Union
from PIL import Image
from collections import defaultdict
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from torch.utils.data import default_collate, get_worker_info
import tarfile
import tqdm
import gc
import threading
import psutil
import tempfile
import decord
from decord import VideoReader
import concurrent.futures
from concurrent.futures import ThreadPoolExecutor, TimeoutError
from misc import print, xprint
from misc.condition_utils import get_camera_condition, get_point_condition, get_wind_condition

# Initialize multiprocessing manager
manager = torch.multiprocessing.Manager()

# ==== helpers ==== #

@contextlib.contextmanager
def ram_temp_file(data, suffix=".mp4"):
    available_ram = psutil.virtual_memory().available
    video_size = len(data)
    
    # Use RAM if available, otherwise fall back to disk
    if video_size < available_ram - (500 * 1024 * 1024):
        temp_dir = "/dev/shm"  # RAM disk
    else:
        temp_dir = None  # Default system temp (disk)

    with tempfile.NamedTemporaryFile(dir=temp_dir, suffix=suffix, delete=True) as temp_file:
        temp_file.write(data)
        temp_file.flush()
        yield temp_file.name


def _nearest_multiple(x: float, base: int = 8) -> int:
    """Round x to the nearest multiple of `base`."""
    return int(round(x / base)) * base


def aspect_ratio_to_image_size(target_size, R, multiple=8):
    if R is None:
        return target_size, target_size
    if isinstance(R, str):
        rw, rh = map(int, R.split(':'))
        R = rw / rh
    area  = target_size ** 2
    out_h = _nearest_multiple(math.sqrt(area / R), multiple)
    out_w = _nearest_multiple(math.sqrt(area * R), multiple)
    return out_h, out_w


def read_tsv(filename):
    # Open the TSV file for reading
    with open(filename, 'r', newline='') as tsvfile:
        reader = csv.reader(tsvfile, delimiter='\t')
        rows = []
        while True:
            try:
                r = next(reader)
                rows.append(r)
            except csv.Error as e:
                print(f'{e}')
            except StopIteration:
                break
        return rows


def sample_clip(
    video_path: str,
    num_frames: int = 8,
    out_fps: Optional[float] = None,      # ← pass an fps here
):
    vr       = VideoReader(video_path)
    src_fps  = vr.get_avg_fps()        # native fps
    total    = len(vr)

    if out_fps is None or out_fps >= src_fps:
        step = 1                       # keep native rate or up-sample later
    else:
        target_duration = (num_frames - 1) / out_fps  # duration in seconds
        frame_span = target_duration * src_fps   # frames needed for this duration
        step = max(frame_span / (num_frames - 1), 1)

    max_start = total - step * (num_frames - 1)
    if max_start <= 1:  # video too short for requested clip
        indices = np.linspace(0, total - 1, num_frames, dtype=int)
        return vr.get_batch(indices.tolist()), indices

    max_start = int(np.floor(max_start - 1))
    start  = random.randint(0, max_start) if max_start > 0 else 0
    idxs   = [int(np.round(start + i * step)) for i in range(num_frames)]
    return vr.get_batch(idxs), idxs


class InfiniteDataLoader(torch.utils.data.DataLoader):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        # Initialize an iterator over the dataset.
        self.dataset_iterator = super().__iter__()

    def __iter__(self):
        return self

    def __next__(self):
        try:
            batch = next(self.dataset_iterator)
        except StopIteration:
            # Dataset exhausted, use a new fresh iterator.
            print('Another Loop over the dataset', flush=True)
            self.dataset_iterator = super().__iter__()
            batch = next(self.dataset_iterator)
        return batch


class DataLoaderWrapper(InfiniteDataLoader):
    def __iter__(self):
        return IterWrapper(super().__iter__())


class IterWrapper:
    def __init__(self, obj):
        self.obj = obj

    def __iter__(self):
        return self
    
    def __next__(self):
        return self.next()

    def next(self):
        return next(self.obj)


# ==== Dataset Implementation, Load your own data ==== #

class ImageTarDataset(Dataset):
    def __init__(self, dataset_tsv, image_size, temporal_size=None, rank=0,  world_size=1, 
                 use_image_bucket=False, multiple=8, no_flip=False, edit=False):
        all_lines = []
                
        # get all data lines
        self.buckets = {}
        self.weights = {}
        self.image_buckets = defaultdict(lambda: 0)
        self.image_buckets['1:1'] = 0  # default bucket

        skipped = 0
        for line in tqdm.tqdm(read_tsv(dataset_tsv)[1:]):
            tsv_file = line[0]
            bucket   = line[1] if len(line) > 1 else 'mlx'
            caption  = line[2] if len(line) > 2 else 'caption'
            weights  = float(line[3] if len(line) > 3 else "1")
            all_data = read_tsv(tsv_file)
            all_maps = {all_data[0][i]: i for i in range(len(all_data[0]))}
            self.weights[all_data[1][0]] = weights
            for line in all_data[1:]:
                try:
                    if 'width' in all_maps:  # filter too small images
                        width, height = int(line[all_maps['width']]), int(line[all_maps['height']])
                        if width * height < (image_size * image_size) / 2:  # if image is smaller than half size of the target size
                            skipped += 1; continue

                    if caption != 'folder':  # input caption has higher priority
                        captions  = caption.split('|')[0].split(':') 
                        operation = caption.split('|')[1] if len(caption.split('|')) > 1 else "none"
                        caption_line = ([line[all_maps[c]] for c in captions], operation)
                    else:
                        caption_line = (line[all_maps['file']].split('/')[-2], "none")  # use folder name as caption
                    
                    items = {'tar': line[all_maps['tar']], 'file': line[all_maps['file']], 'caption': caption_line,
                             'image_bucket': line[all_maps['image_bucket']] if 'image_bucket' in all_maps else "1:1"}
                    
                    if "camera_file" in all_maps: # dl3dv data
                        items["camera_file"] = line[all_maps["camera_file"]]
                    
                    if "force_caption" in all_maps: # force dataset
                        items["force_caption"] = line[all_maps["force_caption"]]
                        if "wind_speed" in all_maps: # wind force
                            items["wind_speed"] = line[all_maps["wind_speed"]]
                            items["wind_angle"] = line[all_maps["wind_angle"]]
                        elif "force" in all_maps: # point-wise
                            items["force"] = line[all_maps["force"]]
                            items["angle"] = line[all_maps["angle"]]
                            items["coordx"] = line[all_maps["coordx"]]
                            items["coordy"] = line[all_maps["coordy"]]
                        
                    if edit:
                        if line[all_maps['visual_file']] != 'none': continue  # TODO: for now, we only support one image, no visual clue    
                        items['edit_instruction'] = line[all_maps['edit_instruction']]
                        items['edited_file'] = line[all_maps['edited_file']]
                    all_lines.append(items)
                    
                except Exception as e:
                    skipped += 1; continue
                
                image_bucket = all_lines[-1]['image_bucket']
                self.image_buckets[image_bucket] += 1
                if all_lines[-1]['tar'] not in self.buckets:
                    self.buckets[all_lines[-1]['tar']] = bucket
        
        if "force_caption" in all_lines[0]:            
            wind_forces = [l["wind_speed"] for l in all_lines] if "wind_speed" in all_lines[0] else [l["force"] for l in all_lines]
            self.min_wind_force = min(wind_forces)
            self.max_wind_force = max(wind_forces)

        self.use_image_bucket = use_image_bucket
        self.all_lines = all_lines[rank:][::world_size]   # all lines is sorted by tar file
        self.num_samples_per_rank = None
        self.image_size = image_size
        self.multiple = multiple
        self.temporal_size = tuple(map(int, temporal_size.split(':'))) if isinstance(temporal_size, str) else None
        self.edit_mode = edit
        
        def center_crop_resize(img, ratio="1:1", target_size: int = 256, multiple: int = 8):
            """
            1. Center crop `img` to the largest window with aspect ratio = ratio.
            2. Resize so  HxW ≈ target_size²  (each side a multiple of `multiple`).

            Args
            ----
            img         : PIL Image or torch tensor (CHW/HWC)
            ratio       : "3:2", (3,2), "1:1", etc.
            target_size : reference side length (area = target_size²)
            multiple    : force each output side to be a multiple of this number
            """
            # --- parse ratio ----------------------------------------------------------
            if isinstance(ratio, str):
                rw, rh = map(int, ratio.split(':'))
            else:                                 # already a tuple/list
                rw, rh = ratio
            R = rw / rh                           # width / height

            # --- crop to that aspect ratio -------------------------------------------
            w, h = img.size if hasattr(img, "size") else (img.shape[-1], img.shape[-2])
            if w / h > R:                         # image too wide → trim width
                crop_h, crop_w = h, int(round(h * R))
            else:                                 # image too tall → trim height
                crop_w, crop_h = w, int(round(w / R))
            img = transforms.functional.center_crop(img, (crop_h, crop_w))

            # --- compute output dimensions -------------------------------------------
            area  = target_size ** 2
            out_h = _nearest_multiple(math.sqrt(area / R), multiple)
            out_w = _nearest_multiple(math.sqrt(area * R), multiple)

            # --- resize & return ------------------------------------------------------
            return transforms.functional.resize(img, (out_h, out_w), antialias=True)
        
        self.transforms = {}
        self.size_bucket_maps = {}
        self.bucket_size_maps = {}
        for bucket in self.image_buckets:
            trans = [transforms.Lambda(lambda img, r=bucket: center_crop_resize(img, ratio=r, target_size=image_size, multiple=multiple))]
            if not no_flip:
                trans.append(transforms.RandomHorizontalFlip())
            trans.extend([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])
            self.transforms[bucket] = transforms.Compose(trans)

            w, h = map(int, bucket.split(':'))
            out_h, out_w = aspect_ratio_to_image_size(image_size, w / h, multiple=multiple)
            self.size_bucket_maps[(out_h, out_w)] = bucket
            self.bucket_size_maps[bucket] = (out_h, out_w)
            
        self.transform = self.transforms['1:1']  # default transform
        print(f"Rank0 -- Loading {len(self.all_lines)} lines of data | {skipped} lines are skipped due to size or error")

    def __len__(self):
        if self.num_samples_per_rank is not None:
            return self.num_samples_per_rank
        return len(self.all_lines)
    
    def __getitem__(self, idx):
        image_item = self.all_lines[idx]
        tar_file = image_item['tar']
        img_file = image_item['file']
        img_bucket = image_item['image_bucket']
        try:
            with tarfile.open(tar_file, mode='r') as tar:
                img = self._read_image(tar, img_file, img_bucket)
                H0, W0 = img.size
                scale  = self.image_size / min(H0, W0)
                state  = np.array([scale, H0, W0])
        except Exception as e:
            print(f'Reading data error {e}')
        sample = image_item.copy()
        sample.update(image=img, state=state)
        return sample
    
    def _read_image(self, tar, img_file, img_bucket):
        
        def _transform(img):
            if not self.use_image_bucket:
                return self.transform(img)
            else:
                return self.transforms[img_bucket](img)

        x_shape = aspect_ratio_to_image_size(self.image_size, img_bucket, multiple=self.multiple)
        if self.temporal_size is not None:  # read video
            num_frames, out_fps = self.temporal_size[0], self.temporal_size[1:]
            if len(out_fps) == 1:
                out_fps = out_fps[0]
            else:
                out_fps = random.choice(out_fps)  # randomly choose one fps from the list            
            assert img_file.endswith('.mp4'), "Only support mp4 video for now"   
            try:
                with tar.extractfile(img_file) as video_data:
                    with ram_temp_file(video_data.read()) as tmp_path:
                        frames, frame_inds = sample_clip(tmp_path, num_frames=num_frames, out_fps=out_fps)
                        frames = frames.asnumpy()    
            except Exception as e:
                print(f'Reading data error {e} {img_file}')
                frames = np.zeros((num_frames, x_shape[0], x_shape[1], 3), dtype=np.uint8)
            return torch.stack([_transform(Image.fromarray(frame)) for frame in frames]), out_fps, frame_inds
        
        try:           
            original_img = Image.open(tar.extractfile(img_file)).convert('RGB')
        except Exception as e:
            print(f'Reading data error {e} {img_file}')
            original_img = Image.new('RGB', (x_shape[0], x_shape[1]), (0, 0, 0))
        return _transform(original_img), 0, None

    def collate_fn(self, batch):
        batch = default_collate(batch)
        return batch

    def get_batch_modes(self, x):
        x_aspect   = self.size_bucket_maps.get(x.size()[-2:], "1:1")
        video_mode = self.temporal_size is not None
        return x_aspect, video_mode


class OnlineImageTarDataset(ImageTarDataset):
    max_retry_n = 20
    max_read = 4096
    tar_keys_lock = manager.Lock() if manager is not None else None
    
    def __init__(self, dataset_tsv, image_size, batch_size=None, **kwargs):
        super().__init__(dataset_tsv, image_size, **kwargs)
        
        self.tar_lists = defaultdict(lambda: [])
        self.tar_image_buckets = defaultdict(lambda: defaultdict(lambda: 0))
        for i, line in enumerate(self.all_lines):
            tar_file = line['tar']
            image_bucket = line['image_bucket']
            self.tar_lists[tar_file] += [i]
            self.tar_image_buckets[tar_file][image_bucket] += 1
        self.reset_tar_keys = []
        for key in self.tar_lists.keys():
            repeat = int(self.weights.get(key, 1))
            self.reset_tar_keys.extend([key] * repeat)
        self.tar_keys = manager.list(self.reset_tar_keys) if manager is not None else list(self.reset_tar_keys)
        
        # Use more workers for better prefetching, but limit to reasonable number
        self.worker_executors = {}
        self.worker_caches = {}  # each entry: {active:{tar,key,cnt,inner_idx}, prefetch:{future,key}} 
        self.worker_caches_lock = threading.Lock()  # Protect worker_caches access
        self.shuffle_everything()
        if self.use_image_bucket:
            assert batch_size, "batch_size should be set when use_image_bucket is True"
        self.batch_size = batch_size
        if self.temporal_size is not None:
            assert self.temporal_size[0] > 1, "temporal_size should be greater than 1 for video data" 
            self.max_read = 512
    
    def cleanup_worker_cache(self, wid):
        """Clean up worker cache entry and associated resources"""
        with self.worker_caches_lock:
            if wid in self.worker_caches:
                cache_entry = self.worker_caches[wid]
                # Cancel prefetch future if still running
                if 'prefetch' in cache_entry and hasattr(cache_entry['prefetch'], 'cancel'):
                    cache_entry['prefetch'].cancel()
                
                if cache_entry.get('tar') is not None:
                    tar = cache_entry['tar']
                    self._close_tar(tar)
                    cache_entry['tar'] = None
                # Remove the entire cache entry
                del self.worker_caches[wid]
                gc.collect()
    
    def _s3(self):
        raise NotImplementedError("Please implement your own _s3() method to return a boto3 session/client")
    
    def shuffle_everything(self):
        for key in tqdm.tqdm(self.tar_keys):
            random.shuffle(self.tar_lists[key])
        random.shuffle(self.tar_keys)
        print("shuffle everything done!")
    
    def download_tar(self, prefetch=True, wid=None):
        i = 0
        file_stream = None
        tar_file = None
        download = f'prefetch {wid}' if prefetch else 'just download'
        while True:
            if i % self.max_retry_n == 0:  # retry a different tar file
                tar_file = self._get_next_key()  # get the next tar file key
            file_stream = None
            try:
                file_stream = io.BytesIO()
                self._s3().download_fileobj(self.buckets[tar_file], tar_file, file_stream)  # hard-coded
                file_stream.seek(0)
                tar = tarfile.open(fileobj=file_stream, mode='r')
                # Store the file_stream reference so it can be closed later
                tar._file_stream = file_stream
                xprint(f'[INFO] {download} tar file: {tar_file}')
                return tar, tar_file
            except Exception as e:
                xprint(f'[ERROR] {download} tar file {tar_file} failed: {e}')
                i += 1
                if file_stream:
                    file_stream.close()
                    file_stream = None
                time.sleep(min(i * 0.1, 5))  # Exponential backoff with cap
        
    def _get_next_key(self):
        with self.tar_keys_lock:
            if not self.tar_keys or len(self.tar_keys) == 0:
                xprint(f'[WARN] all dataset exhausted... this should not happen usually')
                self.tar_keys.extend(list(self.reset_tar_keys))  # reset
                random.shuffle(self.tar_keys)
            return self.tar_keys.pop(0)  # remove and return the first key
    
    def _start_prefetch(self, wid):
        """Start prefetching the next tar file for the worker"""
        # Create executor per worker process if it doesn't exist
        if wid not in self.worker_executors:
            self.worker_executors[wid] = ThreadPoolExecutor(max_workers=1)
        future = self.worker_executors[wid].submit(self.download_tar, prefetch=True, wid=wid)  # download tar file in a separate thread
        self.worker_caches[wid]['prefetch'] = future
    
    def _close_tar(self, tar):
        # Properly close both tar and underlying file stream
        if hasattr(tar, '_file_stream') and tar._file_stream:
            tar._file_stream.close()
            tar._file_stream = None
        tar.close()
        del tar
        gc.collect()
    
    def __getitem__(self, idx):        
        try:
            wid = get_worker_info().id
        except Exception as e:
            wid = -1
        
        # ─── first time this worker is used ─── #
        if wid not in self.worker_caches:
            tar, key = self.download_tar(prefetch=False)  # download tar file
            with self.worker_caches_lock:
                self.worker_caches[wid] = dict(
                    active=dict(tar=tar, key=key, cnt=0, inner_idx=0),  # active cache
                )
                self._start_prefetch(wid)  # start prefetching the next tar file
        
        cache = self.worker_caches[wid]
        active = cache['active']
        tar = active['tar']
        key = active['key']
        cnt = active['cnt']
        inner_idx = active['inner_idx']
        
        # handle image bucketting
        if self.use_image_bucket:
            if inner_idx % self.batch_size == 0:
                # sample based on local tar file statistics in case some dataset only has one image bucket
                tar_buckets = self.tar_image_buckets[key]
                target_image_bucket = random.choices(
                    list(tar_buckets.keys()), weights=list(tar_buckets.values()), k=1)[0]
                self.worker_caches[wid]['target_image_bucket'] = target_image_bucket
            
            # scan the list to find the nearest target image bucket
            target_image_bucket, t_cnt = self.worker_caches[wid]['target_image_bucket'], cnt
            while self.all_lines[self.tar_lists[key][t_cnt]]['image_bucket'] != target_image_bucket:
                t_cnt += 1
                if t_cnt >= len(self.tar_lists[key]): t_cnt = 0
            # sawp the image location
            if cnt != t_cnt:
                self.tar_lists[key][cnt], self.tar_lists[key][t_cnt] = self.tar_lists[key][t_cnt], self.tar_lists[key][cnt]
                    
        img_id = self.tar_lists[key][cnt]
        image_item = self.all_lines[img_id]    
        sample = {key: image_item[key] for key in image_item}
        image, fps, frame_inds = self._read_image(tar, image_item['file'], image_item['image_bucket'])
        sample.update(image=image, fps=fps, local_idx=img_id, inner_idx=inner_idx)
        if self.edit_mode:
            image, fps, _ = self._read_image(tar, image_item['edited_file'], image_item['image_bucket'])
            sample.update(edited_image=image, fps=fps, edit_instruction=image_item['edit_instruction'])
            
        if "camera_file" in image_item: # dl3dv data
            sample["condition"] = get_camera_condition(tar, image_item["camera_file"], width=image.shape[3], height=image.shape[2], factor=self.multiple, frame_inds=frame_inds)
        
        if "force_caption" in image_item: # force dataset
            if "wind_speed" in image_item: # wind force
                sample["condition"] = get_wind_condition(image_item["wind_speed"], image_item["wind_angle"], min_force=self.min_wind_force, max_force=self.max_wind_force, num_frames=image.shape[1], width=image.shape[3], height=image.shape[2]) 
            elif "force" in image_item: # point-wise
                sample["condition"] = get_point_condition(image_item["force"], image_item["angle"], image_item["coordx"], image_item["coordy"], min_force=self.min_wind_force, max_force=self.max_wind_force, num_frames=image.shape[1], width=image.shape[3], height=image.shape[2]) 
        
        # update cnt
        cnt, inner_idx = cnt + 1, inner_idx + 1
        if (cnt == len(self.tar_lists[key])) or (cnt == self.max_read):
            # -- active tar finished, switch to prefetched tar -- #
            self._close_tar(tar)  # close the current tar file
            
            try:
                # Wait for prefetch with timeout
                new_tar, new_key = cache['prefetch'].result()  # 5 minute timeout
            except Exception as e:
                xprint(f'[WARN] Prefetch failed, downloading new tar synchronously: {e}')
                new_tar, new_key = self.download_tar(prefetch=False)
                
            cache['active'] = dict(tar=new_tar, key=new_key, cnt=0, inner_idx=inner_idx)  # update active cache
            
            # shuffle the image list
            random.shuffle(self.tar_lists[key])  # shuffle the list
            with self.tar_keys_lock:
                self.tar_keys.append(key)  # return the key to the list so other workers can use it
            
            self._start_prefetch(wid)  # start prefetching the next tar file
        else:
            cache['active']['cnt'] = cnt
        
        # always update inner_idx (IMPORTANT)
        cache['active']['inner_idx'] = inner_idx
        return sample


class OnlineImageCaptionDataset(OnlineImageTarDataset):
    def __getitem__(self, idx):   
        sample = super().__getitem__(idx)
        captions, caption_op = sample['caption']
        if caption_op == 'none':
            sample['caption'] = captions[0] if isinstance(captions, list) else captions
        elif ':' in caption_op:
            sample['caption'] = random.choices(captions, weights=[float(a) for a in caption_op.split(':')])[0]
        else:
            raise NotImplementedError(f"Unknown caption operation: {caption_op}") 
        return sample

    def collate_fn(self, batch):
        batch = super().collate_fn(batch)
        image = batch['image']
        caption = batch['caption']
        if self.edit_mode:
            image = torch.cat([image, batch['edited_image']], dim=0)
            caption.extend(batch['edit_instruction'])
        
        meta = {key: batch[key] for key in batch if key not in 
                ['image', 'caption', 'edited_image', 'edit_instruction']}
        return image, caption, meta


# ==== Dummy Dataset Implementation for Open Source Release ====

class DummyImageCaptionDataset(Dataset):
    """
    Dummy dataset that generates synthetic image-caption pairs for training/testing.
    Supports mixed aspect ratios and batch-wise aspect ratio consistency.
    """

    def __init__(
        self,
        num_samples: int = 10000,
        image_size: int = 256,
        temporal_size: Optional[str] = None,
        use_image_bucket: bool = False,
        batch_size: Optional[int] = None,
        multiple: int = 8,
        no_flip: bool = False,
        edit: bool = False
    ):
        """
        Args:
            num_samples: Number of samples in the dataset
            image_size: Base image size for generation
            temporal_size: Video size specification (e.g., "16:8" for frames:fps)
            use_image_bucket: Whether to use aspect ratio bucketing
            batch_size: Batch size for bucketing (required if use_image_bucket=True)
            multiple: Multiple for dimension rounding
            no_flip: Whether to disable horizontal flipping
            edit: Whether this is an editing dataset
        """
        self.num_samples = num_samples
        self.image_size = image_size
        self.temporal_size = temporal_size
        self.use_image_bucket = use_image_bucket
        self.batch_size = batch_size
        self.multiple = multiple
        self.no_flip = no_flip
        self.edit_mode = edit

        # Parse video parameters
        self.is_video = temporal_size is not None
        if self.is_video:
            frames, fps = map(int, temporal_size.split(':'))
            self.num_frames = frames
            self.fps = fps
        else:
            self.num_frames = 1
            self.fps = None

        # Aspect ratios for mixed aspect ratio training
        self.aspect_ratios = [
            "1:1", "2:3", "3:2", "16:9", "9:16",
            "4:5", "5:4", "21:9", "9:21"
        ] if use_image_bucket else ["1:1"]

        # Generate image buckets for aspect ratios
        self.image_buckets = {}
        for i, ar in enumerate(self.aspect_ratios):
            h, w = aspect_ratio_to_image_size(image_size, ar, multiple)
            self.image_buckets[ar] = (h, w, ar)

        # Sample captions for dummy data
        self.sample_captions = [
            "A beautiful landscape with mountains and trees",
            "A cute cat sitting on a wooden table",
            "A modern city skyline at sunset",
            "A vintage car parked on a street",
            "A delicious meal on a white plate",
            "A person walking in a park",
            "A colorful flower garden in bloom",
            "A cozy living room with furniture",
            "A stormy ocean with large waves",
            "A peaceful forest path in autumn",
            "A group of friends laughing together",
            "A majestic eagle flying in the sky",
            "A busy marketplace with vendors",
            "A snow-covered mountain peak",
            "A child playing with toys",
            "A romantic candlelit dinner",
            "A train traveling through countryside",
            "A lighthouse on a rocky coast",
            "A field of sunflowers under blue sky",
            "A family having a picnic outdoors"
        ]

        # Create transform pipeline
        def center_crop_resize(img, ratio="1:1", target_size: int = 256, multiple: int = 8):
            """
            1. Center crop `img` to the largest window with aspect ratio = ratio.
            2. Resize so  HxW ≈ target_size²  (each side a multiple of `multiple`).

            Args
            ----
            img         : PIL Image or torch tensor (CHW/HWC)
            ratio       : "3:2", (3,2), "1:1", etc.
            target_size : reference side length (area = target_size²)
            multiple    : force each output side to be a multiple of this number
            """
            # --- parse ratio ----------------------------------------------------------
            if isinstance(ratio, str):
                rw, rh = map(int, ratio.split(':'))
            else:                                 # already a tuple/list
                rw, rh = ratio
            R = rw / rh                           # width / height

            # --- crop to that aspect ratio -------------------------------------------
            w, h = img.size if hasattr(img, "size") else (img.shape[-1], img.shape[-2])
            if w / h > R:                         # image too wide → trim width
                crop_h, crop_w = h, int(round(h * R))
            else:                                 # image too tall → trim height
                crop_w, crop_h = w, int(round(w / R))
            img = transforms.functional.center_crop(img, (crop_h, crop_w))

            # --- compute output dimensions -------------------------------------------
            area  = target_size ** 2
            out_h = _nearest_multiple(math.sqrt(area / R), multiple)
            out_w = _nearest_multiple(math.sqrt(area * R), multiple)

            # --- resize & return ------------------------------------------------------
            return transforms.functional.resize(img, (out_h, out_w), antialias=True)
        
        self.transforms = {}
        self.size_bucket_maps = {}
        self.bucket_size_maps = {}
        for bucket in self.image_buckets:
            trans = [transforms.Lambda(lambda img, r=bucket: center_crop_resize(img, ratio=r, target_size=image_size, multiple=multiple))]
            if not no_flip:
                trans.append(transforms.RandomHorizontalFlip())
            trans.extend([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])
            self.transforms[bucket] = transforms.Compose(trans)

            w, h = map(int, bucket.split(':'))
            out_h, out_w = aspect_ratio_to_image_size(image_size, w / h, multiple=multiple)
            self.size_bucket_maps[(out_h, out_w)] = bucket
            self.bucket_size_maps[bucket] = (out_h, out_w)
            
        self.transform = self.transforms['1:1']  # default transform

    def __len__(self) -> int:
        return self.num_samples

    def __getitem__(self, idx: int) -> dict:
        """Get a single sample from the dataset."""
        # Choose aspect ratio
        if self.use_image_bucket:
            bucket_name = random.choice(list(self.image_buckets.keys()))
            h, w, aspect_ratio = self.image_buckets[bucket_name]
        else:
            h, w, aspect_ratio = self.image_size, self.image_size, "1:1"
            bucket_name = aspect_ratio

        # Generate dummy image
        if self.is_video:
            # Generate video tensor (T, C, H, W)
            image = torch.randn(self.num_frames, 3, h, w)
            # Normalize to [-1, 1] range
            image = torch.tanh(image)
        else:
            # Generate RGB image
            image = Image.new('RGB', (w, h), color=(
                random.randint(50, 200),
                random.randint(50, 200),
                random.randint(50, 200)
            ))

            # Add some random patterns for variety
            if random.random() > 0.5:
                # Add gradient
                pixels = []
                for y in range(h):
                    for x in range(w):
                        r = int(255 * x / w)
                        g = int(255 * y / h)
                        b = int(255 * (x + y) / (w + h))
                        pixels.append((r, g, b))
                image.putdata(pixels)

            image = self.transform(image)

        # Generate caption
        caption = random.choice(self.sample_captions)

        # Add some variation to captions
        if random.random() > 0.7:
            adjectives = ["beautiful", "stunning", "amazing", "incredible", "magnificent"]
            caption = f"{random.choice(adjectives)} {caption.lower()}"

        sample = {
            'image': image,
            'caption': caption,
            'image_bucket': bucket_name,
            'aspect_ratio': aspect_ratio,
            'idx': idx
        }

        # Add video-specific metadata
        if self.is_video:
            sample.update({
                'num_frames': self.num_frames,
                'fps': self.fps,
                'temporal_size': self.temporal_size
            })

        # Add editing data if needed
        if self.edit_mode:
            # Generate slightly modified image for editing tasks
            edited_image = image + torch.randn_like(image) * 0.1
            edited_image = torch.clamp(edited_image, -1, 1)
            sample.update({
                'edited_image': edited_image,
                'edit_instruction': f"Edit this image to make it more {random.choice(['colorful', 'bright', 'artistic', 'realistic'])}"
            })

        return sample

    def collate_fn(self, batch: list) -> tuple:
        """Collate function for batching samples."""
        # Group by aspect ratio if using image buckets
        if self.use_image_bucket:
            # Sort batch by image bucket for consistency
            batch = sorted(batch, key=lambda x: x['image_bucket'])

        # Standard collation
        collated = {}
        images = torch.stack([item['image'] for item in batch], dim=0)
        captions = [item['caption'] for item in batch]

        # Collect metadata
        for key in ['image_bucket', 'aspect_ratio', 'idx']:
            if key in batch[0]:
                collated[key] = [item[key] for item in batch]

        # Handle video metadata
        if self.is_video:
            for key in ['num_frames', 'fps', 'temporal_size']:
                if key in batch[0]:
                    collated[key] = [item[key] for item in batch]

        # Handle editing data
        if self.edit_mode and 'edited_image' in batch[0]:
            edited_images = torch.stack([item['edited_image'] for item in batch], dim=0)
            collated['edited_image'] = edited_images
            collated['edit_instruction'] = [item['edit_instruction'] for item in batch]

        return images, captions, collated

    def get_batch_modes(self, x):
        x_aspect   = self.size_bucket_maps.get(x.size()[-2:], "1:1")
        video_mode = self.temporal_size is not None
        return x_aspect, video_mode


class DummyDataLoaderWrapper:
    """
    Wrapper that mimics the DataLoaderWrapper functionality.
    Provides infinite iteration over the dataset.
    """

    def __init__(self, dataset, batch_size=1, num_workers=0, **kwargs):
        self.dataset = dataset
        self.batch_size = batch_size
        self.dataloader = DataLoader(
            dataset,
            batch_size=batch_size,
            num_workers=num_workers,
            collate_fn=dataset.collate_fn,
            shuffle=True,
            drop_last=True,
            **kwargs
        )
        self.iterator = None
        self.secondary_loader = None

    def __iter__(self):
        """Infinite iteration over the dataset."""
        while True:
            if self.iterator is None:
                self.iterator = iter(self.dataloader)
            try:
                yield next(self.iterator)
            except StopIteration:
                self.iterator = iter(self.dataloader)
                yield next(self.iterator)

    def __len__(self):
        return len(self.dataloader)


def create_dummy_dataloader(
    dataset_name: str,
    img_size: int,
    vid_size: Optional[str] = None,
    batch_size: int = 16,
    use_mixed_aspect: bool = False,
    multiple: int = 8,
    num_samples: int = 10000,
    infinite: bool = False
) -> Union[DataLoader, DummyDataLoaderWrapper]:
    """
    Create a dummy dataloader that mimics the original functionality.

    Args:
        dataset_name: Name of the dataset (used for deterministic seeding)
        img_size: Base image size
        vid_size: Video specification (e.g., "16:8")
        batch_size: Batch size
        use_mixed_aspect: Whether to use mixed aspect ratio training
        multiple: Multiple for dimension rounding
        num_samples: Number of samples in the dataset
        infinite: Whether to create infinite dataloader

    Returns:
        DataLoader or DummyDataLoaderWrapper
    """
    # Set seed based on dataset name for reproducibility
    seed = hash(dataset_name) % (2**32 - 1)
    random.seed(seed)
    np.random.seed(seed)

    # Create dataset
    dataset = DummyImageCaptionDataset(
        num_samples=num_samples,
        image_size=img_size,
        temporal_size=vid_size,
        use_image_bucket=use_mixed_aspect,
        batch_size=batch_size,
        multiple=multiple,
        edit='edit' in dataset_name.lower()
    )

    # Set dataset attributes expected by training code
    dataset.total_num_samples = num_samples
    dataset.num_samples_per_rank = num_samples

    # Create dataloader
    if infinite:
        return DummyDataLoaderWrapper(
            dataset,
            batch_size=batch_size,
            num_workers=2,
            pin_memory=True,
            drop_last=True,
            persistent_workers=True
        )
    else:
        return DataLoader(
            dataset,
            batch_size=batch_size,
            num_workers=2,
            pin_memory=True,
            drop_last=True,
            shuffle=True,
            collate_fn=dataset.collate_fn,
            persistent_workers=True
        )