aoiandroid/got-ocr2-onnx

GOT-OCR2 ONNX Export (stepfun-ai/GOT-OCR2_0)

This directory contains ONNX exports produced from stepfun-ai/GOT-OCR2_0 using BaofengZan/GOT-OCRv2-onnx (llm-export).

Contents

• got_ocr2_vision_encoder.onnx (+ .onnx.data): Vision encoder. Input: images; output: visual features. Ready for use with TranslateBlue's GOT-OCR2 ONNX integration.
• got_ocr2_decoder.onnx (optional): Single decoder ONNX; when present, full OCR runs: image → vision encoder → encoder_hidden_states → decoder (BOS then autoregressive) → logits → text. Produce it with scripts/export_got_ocr2_decoder_onnx.py.
• tokenizer.json, tokenizer_config.json: Tokenizer for decode. Generated from the Hugging Face model vocab.
• Split decoder (for reference; app does not load these as a single decoder):
  • embedding.onnx, block_0.onnx .. block_23.onnx, lm.onnx, norm.onnx, mm_projector_vary.onnx (each with .onnx.data where applicable).

App compatibility

TranslateBlue's GOTOCR2OCRService expects two ONNX files:

1. got_ocr2_vision_encoder.onnx – present; you can run the vision encoder in the app.
2. got_ocr2_decoder.onnx – a single decoder ONNX that takes decoder_input_ids and encoder_hidden_states and outputs logits. When this file is present (e.g. from scripts/export_got_ocr2_decoder_onnx.py), full OCR runs: vision encoder → encoder_hidden_states → decoder (with BOS then autoregressive tokens) → logits → decoded text.

See Docs/GOT_OCR2_ONNX_Export.md for I/O names and export options.