Simple demo for converting PaddleOCR text recognizer model to ONNX and ONNX inference
This simple demo shows how to export trained PaddleOCR recognizer model to ONNX and make inference
To export ONNX model add following lines after model.eval() in the tools\export_model.py in PaddleOCR repo
model.eval()
###
if config['Architecture']['model_type'] == "rec":
input_spec = paddle.static.InputSpec(shape=[1, 3, 48, 96], dtype='float32', name='image')
paddle.onnx.export(model, os.path.join(config['Global']['save_inference_dir'],'rec_db'), input_spec=[input_spec], opset_version=10, enable_onnx_checker=True)
###
NB Not all of the recognition models can be converted to ONNX (e.g. STARNet), because of the specific ops in Paddle which are missing from ONNX.
and run
python3 tools/export_model.py -c configs/rec/your_rec_config.yml -o Global.pretrained_model=output/rec_mv3/best_accuracy Global.save_inference_dir=output/rec_db/ Global.load_static_weights=false
To test ONNX inference
python3 demo_razmetka.py
