tagirshin/RxnScribe-ONNX

RxnScribe ONNX CPU bundle

ONNX conversions of the existing RxnScribe, MolScribe and English EasyOCR checkpoints. This is a community conversion, not a newly trained model or an upstream release. Use the matching RxnScribe feature branch and its RxnScribeONNX interface. PyTorch is needed for export, not inference.

Download

Download the complete repository, including manifest.json, all seven .onnx graphs, reaction_embeddings.npz and molecule_vocab.json:

hf download tagirshin/RxnScribe-ONNX --revision onnx-poc-v1 --local-dir local/onnx-models

Then use the matching source checkout with requirements-onnx.txt installed:

from rxnscribe import RxnScribeONNX
model = RxnScribeONNX("local/onnx-models", threads=2)
reactions = model.predict_image_file("scheme.png", molscribe=True, ocr=True)

The runtime reads local files and makes no Hub requests. Pin the Hub commit SHA from the source branch's README.onnx.md for reproducible deployments.

Components and provenance

Component	Source	License notice
Reaction encoder, decoder, embeddings	yujieq/RxnScribe, pix2seq_reaction_full.ckpt, revision 034bcfeaa6780624b2897f7955853271de1d1f65	MIT
Molecule encoder, decoder, bond predictor and vocabulary	yujieq/MolScribe, swin_base_char_aux_1m.pth, revision a0189776b7415b82795c7ee81eed311bf5c8724b	MIT
Text detector	EasyOCR 1.7.2, craft_mlt_25k.pth, based on CRAFT	Apache-2.0, CRAFT MIT
English text recognizer	EasyOCR 1.7.2, english_g2.pth	Apache-2.0

Original checkpoint and exported file SHA-256 hashes are recorded in manifest.json. Conversion uses ONNX opset 17. It was validated with Python 3.11, Torch 2.6.0 CPU for export, and ONNX Runtime 1.20.1 for inference on Linux ARM64 via Docker Desktop. Model artifacts occupy approximately 659 MB.

Validation and limitations

• Exact reaction tokens on five bundled schemes, the app Ugi example and a blank portrait; token log scores within 2e-4 absolute/relative tolerance.
• Exact native single-crop molecular graphs, SMILES and molfiles on 46 crops.
• Exact native floating-point OCR text on 32 condition crops.
• A Torch-free, network-disabled container completed full reaction, molecule and OCR inference; the companion app's Dagster extraction job also passed.
• CPU, one image/crop at a time. Only English greedy OCR is supported. GPU, x86 execution and production multi-user workloads were not validated.
• Batch-32 Torch inference on one 17-crop scheme gave identical SMILES but four different molfiles compared with single-crop inference.
• OCR uses floating-point weights. The quantized Torch CPU path produced corrupted text in the tested ARM environment; its text matched only 8 of 32 crops. The companion native fork also disables this quantization.
• These checks establish conversion parity, not chemical correctness. Review structures and conditions before use. Atom mapping and rule extraction are separate stages and are not supplied by this model bundle.

The original training data, methods and intended model behavior are documented by the linked upstream projects. No fine-tuning was performed for this conversion.