AnalogRetriever checkpoints

Checkpoints for AnalogRetriever: Learning Cross-Modal Representations for Analog Circuit Retrieval (ICCAD 2026), which retrieves analog circuits across schematic images, functional descriptions and SPICE netlists.

Paper · Code · Dataset · Demo

Files

File	Model (paper Table 2)
analogretriever.pth	AnalogRetriever: CLIP ViT-L/14 + port-aware RGCN, three-phase curriculum
tic_rgcn.pth	TIC (RGCN) ablation
tic_gcn.pth	TIC (GCN) ablation
ti.pth	TI (bi-modal CLIP) ablation

Each checkpoint is a PyTorch state dict that includes the fine-tuned CLIP weights; its .metadata.json sidecar records the model type, training configuration, selected epoch, graph schema (24 node types, 20 relation types) and SHA-256. results/ holds the held-out evaluation of every model and of the four zero-shot baselines, and training_logs/ the training and evaluation logs.

Usage

git clone https://github.com/yihan-wa/AnalogRetriever_ICCAD && cd AnalogRetriever_ICCAD
bash scripts/install_env.sh
# download and prepare the dataset as described in the code README, then
hf download eulermaxwell/AnalogRetriever analogretriever.pth analogretriever.pth.metadata.json --local-dir checkpoints
python3 -m analogretriever.evaluation.evaluate_retrieval --checkpoint checkpoints/analogretriever.pth

analogretriever.checkpoints.load_checkpoint rebuilds any of the four models from its file; no separate CLIP download is needed.

Results

Recall@1 (%) on the 1,000-triplet held-out split, averaged over ten 200-item galleries (I: schematic image, T: description, C: netlist). The zero-shot baselines are listed for reference; per-model Recall@5/10 are in results/.

Model	I→C	T→I	T→C	C→I	I→T	C→T	Avg
CLIP (zero-shot)	1.2	2.2	2.9	2.9	2.7	3.2	2.5
CROP (zero-shot)	2.1	2.2	9.1	2.3	2.7	8.5	4.5
ChatLS (zero-shot)	1.5	2.2	11.8	1.8	2.7	8.3	4.7
NetTAG (zero-shot)	0.4	2.2	0.3	0.4	2.7	0.1	1.1
TI (ti.pth)	–	76.8	–	–	76.1	–	–
TIC (GCN) (tic_gcn.pth)	69.7	75.6	74.0	67.5	76.4	70.2	72.3
TIC (RGCN) (tic_rgcn.pth)	71.9	76.0	75.7	69.2	76.5	72.4	73.6
AnalogRetriever (analogretriever.pth)	72.9	78.5	74.6	70.7	77.3	71.7	74.3

Training

Trained with the code at commit 52435fd on one NVIDIA H200 (PyTorch 2.15.0, CUDA 13.2, bf16 autocast) using the paper settings: 6,371 refined Masala-CHAI triplets split 5,371 / 1,000 with seed 42; 20 epochs (6 / 2 / 12 for the three curriculum phases); batch size 256; AdamW with learning rate 5e-5 for CLIP and 5e-4 for the graph encoder; bottom 16 of 24 ViT blocks frozen; label smoothing 0.1; hard-negative ratio 0.05 → 0.3 over K = 30 Qwen3-Embedding clusters; seed 42. The shared CLIP parameters are registered once per encoder with AdamW (--clip_optimizer_mode legacy_cuda_duplicate). Checkpoints are selected on the held-out split by the mean R@1 of Image→Code, Text→Image and Text→Code (Text→Image for TI), and the same split is used for the reported results.

Limitations

The dataset consists mostly of textbook-style analog circuits, and the held-out split is drawn at the instance level, so related circuit families can appear in both splits. Retrieved circuits should be checked by simulation before reuse. Descriptions longer than CLIP's 77-token context are truncated.

License and citation

MIT. CLIP and the Masala-CHAI source material keep their own terms.

@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 '26)},
  year      = {2026},
  address   = {San Jose, CA, USA},
  publisher = {ACM},
  doi       = {10.1145/3831252.3834253}
}