End-to-end YOLO26s line detector and the recognition encoder split into visual (CNN+Transformer) and memory (BiLSTM + cross-attention K/V) halves, for onnxruntime-web. Built by scripts/netra-ocr-web/export.py in the netra-lab-web repo. Existing files are unchanged.
10 days ago
Card: the PyTorch checkpoints in pytorch/, images from this repo
Weights are stored as fp16 and cast to float32 at load; ONNX Runtime folds the casts. CPU text identical on 325/325 lines, layout pages identical. lines/ unchanged. Model card updated for netra-ocr 1.2.0.
