eulermaxwell/analogretriever_H200

AnalogRetriever — ICCAD 2026 checkpoint

English | 简体中文

Weights for AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval, by Yihan Wang, Lei Li, Yao Lai, Jing Wang and Yan Lu.

Paper · Code · Dataset · Reproducibility notes

Select the correct artifact

tic_cur_v2_best_v6_h200.pth is the final ICCAD Table 2 model. It is a custom PyTorch state dictionary for the linked code, not a standalone Transformers AutoModel/pipeline checkpoint. The file is 1,764,335,833 bytes. SHA-256: c8fe3eb1aaed32e026c0bde98692d8be059c9d70506fe355cea0cdd11fe94514. The release weight revision is 06610fa3ad069b00627d1960c7cb76957bb34719.

Architecture and training

CLIP ViT-L/14 encodes schematics and English functional descriptions. A two-layer, 512-hidden-dimensional RGCN encodes device/net graphs with 20 typed relations, GraphNorm, residual connections and learned attention pooling. All three outputs are normalized 768-dimensional embeddings. The checkpoint has 22 node type IDs; use the supplied train_data_vocab.json (also committed in the code repository). release_manifest.json records artifact identity and configuration.

Training uses 5,371 triplets from the refined 6,371-item ICCAD dataset, a seed-42 instance split with 1,000 held-out examples, and 6/2/12 epochs for graph warm-up, joint transition and curriculum. The paper uses one H200, batch size 256, CLIP LR 5e-5, graph LR 5e-4, label smoothing 0.1, and a hard-negative ratio 0.05 to 0.3. The bottom 16 of 24 vision blocks are frozen during joint training.

Download and evaluate

Follow the code README to install Python 3.10+, PyTorch, PyTorch Geometric and OpenAI CLIP, download the dataset's iccad2026/ folder as Masala-Chai/, and run python3 scripts/prepare_csv.py. Then, from that code checkout:

python3 -m pip install -e . --no-deps
hf download eulermaxwell/analogretriever_H200 tic_cur_v2_best_v6_h200.pth \
  --revision 06610fa3ad069b00627d1960c7cb76957bb34719 --local-dir checkpoints
python3 scripts/validate_release.py --checkpoint checkpoints/tic_cur_v2_best_v6_h200.pth
python3 -m analogretriever.evaluation.evaluate_TIC --ckpt checkpoints/tic_cur_v2_best_v6_h200.pth \
  --save_json training_logs/tic_cur_results.json

The evaluator loads tensor weights with weights_only=True and checks state-dict compatibility. CPU evaluation is available with --device cpu --batch_size 4, but the published results used CUDA. Do not change the node vocabulary mapping.

Reported evaluation (paper Table 2)

Each trial uses the same 200 selected held-out triplets as queries and candidates; results are averaged over 10 seeded draws from the 1,000 held-out examples. Full 1,000-candidate recall is a different metric and should not be compared directly.

Direction	R@1 (%)	R@5 (%)	R@10 (%)
Image → Code	74.7	95.5	98.1
Text → Image	78.2	96.8	97.9
Text → Code	75.6	96.3	98.6
Code → Image	72.0	94.8	97.9
Image → Text	78.5	96.7	98.3
Code → Text	72.4	96.7	98.3

Mean R@1 across all six directions: 75.2%. These values are transcribed from the paper, not a new evaluation performed during model-card maintenance. Checkpoint selection uses the held-out mean of Image→Code, Text→Image and Text→Code; this held-out set is also used for model selection.

Uses, scope and limitations

Intended for research on cross-modal analog circuit retrieval and candidate reference search. Similarity is not proof of electrical correctness, performance, schematic fidelity or safe reuse. Validate any retrieved design independently. The dataset and instance split may contain related topology/source families; out-of-domain, RF and industrial-scale performance are not established here. The linked release does not contain the AnalogCoder RAG integration or Table 3 trial artifacts. Stage 1 data-generation and baseline adaptation details are explained in the reproducibility notes; metadata checks do not reproduce 75.2%.

Captions may exceed the CLIP 77-token context and are truncated by the code. A maintenance sample found 5 such captions among 40 triplets; this is not a full-dataset prevalence estimate.

License and citation

The repository declares MIT for the weights/code release. CLIP and other third-party components retain their original licenses; this declaration does not relicense MASALA-CHAI schematics or other source data. See the code repository's docs/THIRD_PARTY.md and the dataset card for provenance.

@inproceedings{wang2026analogretriever,
  title = {AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval},
  author = {Wang, Yihan and Li, Lei and Lai, Yao and Wang, Jing and Lu, Yan},
  booktitle = {IEEE/ACM International Conference on Computer-Aided Design (ICCAD)},
  year = {2026},
  doi = {10.1145/3831252.3834253}
}

---

AnalogRetriever — ICCAD 2026 模型权重

论文 AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval 的权重。作者：Yihan Wang、Lei Li、Yao Lai、Jing Wang、Yan Lu。

论文 · 代码 · 数据集 · 复现说明

选择正确的产物

tic_cur_v2_best_v6_h200.pth 是 ICCAD 论文表 2 的最终模型。 它是配合上述代码使用的自定义 PyTorch state dictionary，不是可独立通过 Transformers AutoModel 或 pipeline 加载的权重。文件大小为 1,764,335,833 字节。

SHA-256：c8fe3eb1aaed32e026c0bde98692d8be059c9d70506fe355cea0cdd11fe94514。

发布权重版本：06610fa3ad069b00627d1960c7cb76957bb34719。

架构与训练

CLIP ViT-L/14 编码原理图及英文功能描述。两层、隐藏维度 512 的 RGCN 编码器件/网络图，使用 20 种带类型关系、GraphNorm、残差连接和可学习注意力池化。三种输出均为归一化的 768 维嵌入。权重含 22 个节点类型 ID，请使用附带的 train_data_vocab.json，该文件也已提交到代码仓库。release_manifest.json 记录产物身份与配置。

训练使用精炼 ICCAD 数据集 6,371 组中的 5,371 组，以 seed 42 进行样本级划分并留出 1,000 组。图编码器预热、联合过渡、课程学习分别进行 6/2/12 epochs。论文使用一张 H200，batch size 256，CLIP 学习率 5e-5，图编码器学习率 5e-4，标签平滑 0.1，难负样本比例从 0.05 到 0.3。联合训练期间冻结 24 个视觉块中的底部 16 个。

下载与评估

按代码 README 安装 Python 3.10+、PyTorch、PyTorch Geometric 及 OpenAI CLIP；将数据集的 iccad2026/ 文件夹下载为 Masala-Chai/，再运行 python3 scripts/prepare_csv.py。然后在代码仓库目录执行：

python3 -m pip install -e . --no-deps
hf download eulermaxwell/analogretriever_H200 tic_cur_v2_best_v6_h200.pth \
  --revision 06610fa3ad069b00627d1960c7cb76957bb34719 --local-dir checkpoints
python3 scripts/validate_release.py --checkpoint checkpoints/tic_cur_v2_best_v6_h200.pth
python3 -m analogretriever.evaluation.evaluate_TIC --ckpt checkpoints/tic_cur_v2_best_v6_h200.pth \
  --save_json training_logs/tic_cur_results.json

评估器使用 weights_only=True 加载张量权重并检查 state-dict 兼容性。可通过 --device cpu --batch_size 4 使用 CPU，但论文结果使用 CUDA。请勿更改节点词表映射。

论文报告结果（表 2）

每次试验从留出的 1,000 组中抽取 200 组，同时作为查询和候选，按固定种子重复 10 次后平均。完整 1,000 候选库的召回率属于不同指标，不应直接比较。

下表 Direction 为检索方向：Image 为图片、Text 为文本、Code 为网表；R@k 为前 k 个结果中的召回率。

Direction	R@1 (%)	R@5 (%)	R@10 (%)
Image → Code	74.7	95.5	98.1
Text → Image	78.2	96.8	97.9
Text → Code	75.6	96.3	98.6
Code → Image	72.0	94.8	97.9
Image → Text	78.5	96.7	98.3
Code → Text	72.4	96.7	98.3

六方向平均 R@1 为 **75.2%**。这些数值抄录自论文，并非此次模型卡维护重新运行的评估结果。权重选择使用留出集 Image→Code、Text→Image、Text→Code 的平均指标，因此该留出集也参与模型选择。

用途、范围与限制

适用于跨模态模拟电路检索研究及候选参考电路搜索。相似度不能证明电气正确性、性能、原理图保真或可安全复用；使用检索设计前应独立验证。数据集及样本级划分可能包含相关拓扑或来源的电路族；尚未验证分布外、射频及工业规模场景性能。

所链接的代码发布不包含 AnalogCoder RAG 集成或表 3 的逐次实验产物。第一阶段数据生成及基线改编细节见复现说明；元数据检查不能复现 75.2% 的结果。

描述可能超过 CLIP 的 77-token 上下文，代码会截断。维护抽样的 40 组中有 5 条超限；这不是全数据集超限比例的估计。

许可证与引用

本仓库为权重/代码发布声明 MIT。CLIP 等第三方组件保留各自许可证；该声明不会重新授权 MASALA-CHAI 原理图或其他源数据。来源说明见代码仓库的 docs/THIRD_PARTY.md 及数据卡。

@inproceedings{wang2026analogretriever,
  title = {AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval},
  author = {Wang, Yihan and Li, Lei and Lai, Yao and Wang, Jing and Lu, Yan},
  booktitle = {IEEE/ACM International Conference on Computer-Aided Design (ICCAD)},
  year = {2026},
  doi = {10.1145/3831252.3834253}
}