DBNet++ ONNX conversion fails

[Hungtran-pro]

I encounter the above error. Could anyone here help me?

The running command I use

paddle2onnx --model_dir ./inference/det_db
--model_filename inference.pdmodel
--params_filename inference.pdiparams
--save_file ./inference/det_onnx/model.onnx
--opset_version 11
--enable_onnx_checker True

[GreatV]

The error you encountered while converting the DBNet++ model to ONNX using paddle2onnx might be related to an unsupported operation or incompatible model format. Based on the related discussion here, users experienced similar issues when working with quantized models.

Potential Solutions

1. Ensure You're Using a Non-Quantized Model

The discussion suggests that directly converting a quantized model from PaddleOCR to ONNX may lead to invalid ONNX graph errors during inference. Instead:

• Try converting a non-quantized model first.
• If you require quantization, apply post-training quantization using the paddleslim library and check whether the issue persists.

2. Verify ONNX Opset Version

You specified --opset_version 11, but some PaddlePaddle operators may require a newer opset version. Try updating it to 13:

paddle2onnx --model_dir ./inference/det_db \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--save_file ./inference/det_onnx/model.onnx \
--opset_version 13 \
--enable_onnx_checker True

This may resolve operator incompatibility issues.

3. Debugging the Conversion

Enable additional logging to capture more details about the error:

export GLOG_v=3  # Enable detailed logging
paddle2onnx --model_dir ./inference/det_db \
--model_filename inference.pdmodel \
--params_filename inference.pdiparams \
--save_file ./inference/det_onnx/model.onnx \
--opset_version 11 \
--enable_onnx_checker True

This may provide more insight into which specific operator is causing the failure.

4. Try a Different ONNX Backend

If you're using ONNX Runtime for inference, check whether another backend like TensorRT or OpenVINO can handle the exported model better.

5. Check Your Paddle2ONNX Version

Ensure you have an updated paddle2onnx version:

pip install --upgrade paddle2onnx

Old versions may not support recent PaddlePaddle model structures.

---

Next Steps

1. Test with a non-quantized model first.
2. Try opset version 13.
3. Enable detailed logging to identify which operator fails.
4. If possible, test another inference runtime like TensorRT.

If the issue persists, please provide the exact error message and model details to further investigate. 🚀

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.