---
language:
  - km
license: apache-2.0
tags:
  - ocr
  - transformer
  - vision
pipeline_tag: image-to-text
---

# Khmer OCR CNN + Transformer (ONNX)

This repository contains a ResNet + Transformer decoder checkpoint for Khmer OCR,
the exported ONNX graph, the serialized `config.json` (vocab + hyperparameters),
and the standalone `inference_onnx.py` helper.

## Files

- `khmer_ocr.onnx` – ONNX model
- `config.json` – hyperparameters plus serialized vocabulary
- `inference_onnx.py` and `model.py` – inference helper and architecture

## Usage

```python
from huggingface_hub import hf_hub_download
import importlib.util

repo_id = "metythorn/ocr-stn-cnn-transformer-base"
onnx_path = hf_hub_download(repo_id=repo_id, filename="khmer_ocr.onnx")
config_path = hf_hub_download(repo_id=repo_id, filename="config.json")
inference_path = hf_hub_download(repo_id=repo_id, filename="inference_onnx.py")

spec = importlib.util.spec_from_file_location("khmer_ocr_infer", inference_path)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
ONNXPredictor = module.ONNXPredictor

predictor = ONNXPredictor(model_path=onnx_path, config_path=config_path)
print(predictor.predict("path/to/image.jpg"))
```