ONNX conversion failed for text detector - DB++

[VishyAnand28]

Dear Paddle Community,

I was able to use Paddle2ONNX from here:https://github.com/PaddlePaddle/PaddleOCR/blob/main/docs/ppocr/infer_deploy/paddle2onnx.en.md to convert text detector method DB (Differentiable Binarization). But the conversion fails for DB++ (Differentiable Binarization with ASF).

Problem: Currently Opset versions of 7-18 are supported. How to overcome this issue? Any other ways to convert it?

Following is the error message:
[Paddle2ONNX] Start to parse PaddlePaddle model...
[Paddle2ONNX] Model file path: C:/Users/db++_trained\inference.pdmodel
[Paddle2ONNX] Parameters file path: C:/Users/04_db++_trained\inference.pdiparams[Paddle2ONNX]
Start to parsing Paddle model...
[Paddle2ONNX] [reduce_mean: mean_0.tmp_0] Requires the minimal opset version of 11.
[Paddle2ONNX] Due to the operator: deformable_conv, requires opset_version >= 19.
[Paddle2ONNX] The Opset Version must be between 7 and 18
[Paddle2ONNX] Opset version will change to 19 from 11
[ERROR] Due to opset version, the model exporting is aborted.

Thanks for reading this!

I humbly request guidance from @GreatV @WenmuZhou @LDOUBLEV @MissPenguin @tink2123 @UserWangZz and others ........ for guidance.

[GreatV]

The error you encountered during the ONNX conversion for the DB++ text detector model arises due to certain operators in the model requiring ONNX opset versions greater than 18, specifically for the deformable_conv operator, which needs opset version >= 19. However, Paddle2ONNX currently only supports opset versions between 7 and 18. Below are possible solutions and workarounds to address this issue:

---

1. Upgrade Dependency for Higher Opset Support

• The main issue is that the Paddle2ONNX tool does not currently support opset versions higher than 18, but your model requires opset 19.
• Check if there’s a newer version of Paddle2ONNX or an alternative tool that supports higher opset versions. Monitor updates in the PaddleOCR GitHub repository or Paddle2ONNX repository for announcements about extended opset support.

If no updates are yet available:

• Consider manually modifying the Paddle2ONNX codebase to allow exporting with opset 19. This involves updating the tool to recognize and support opset 19 for operators like deformable_conv. However, this approach requires advanced knowledge of ONNX and Paddle2ONNX internals.
• Alternatively, export the model to other formats (e.g., TensorRT or ONNXRuntime) that bypass strict opset constraints.

---

2. Replace or Simplify Unsupported Operators

If the deformable_conv operator is the bottleneck, you can:

• Replace the operator with a functionally equivalent operator that is supported within opset <= 18. This requires modifying the model architecture or retraining the model.
• Use PaddleOCR’s custom operator mapping mechanism to define the behavior of unsupported operators when converting to ONNX.

---

3. Quantization Workaround (If Applicable)

In some cases, using quantized models with dynamic post-quantization can simplify operator graphs and reduce opset version requirements. Follow these steps:

• Apply dynamic quantization to your DB++ model using PaddleSlim:

  import paddle
  import paddleslim
  
  paddle.enable_static()
  model_dir = "./db++_model"
  model_filename = 'inference.pdmodel'
  params_filename = 'inference.pdiparams'
  quant_model_dir = "./db++_quantized"
  
  paddleslim.quant.quant_post_dynamic(
      model_dir=model_dir,
      model_filename=model_filename,
      params_filename=params_filename,
      save_model_dir=quant_model_dir,
      save_model_filename=model_filename,
      save_params_filename=params_filename,
      weight_bits=8,
  )

• Convert the quantized model using Paddle2ONNX with a lower opset version:

  paddle2onnx --model_dir ./db++_quantized \
              --model_filename inference.pdmodel \
              --params_filename inference.pdiparams \
              --save_file ./db++_quantized.onnx \
              --opset_version 11 \
              --enable_onnx_checker True

• Test the resulting ONNX model in ONNXRuntime or another execution environment.

Caveat: Quantization may degrade model performance slightly. Verify the inference results to ensure they meet your requirements.

---

4. Export with Paddle Inference or TensorRT

If ONNX conversion is not strictly necessary for your deployment pipeline, consider using Paddle Inference or TensorRT directly:

• Paddle Inference supports deploying PaddleOCR models without conversion. Refer to PaddleOCR deployment documentation.
• Use TensorRT for optimized inference. Export the model to a format compatible with TensorRT using Paddle2ONNX or directly from Paddle.

---

5. Alternative Tools

If Paddle2ONNX cannot meet your requirements, explore alternative tools that might support opset 19 or higher:

• ONNX Simplifier: Simplifies ONNX models and can sometimes handle unsupported opsets more gracefully.
• MMdnn: A model conversion tool that supports multiple frameworks, though its Paddle support may be limited.

---

References

• PaddleOCR Paddle2ONNX Guide
• PaddleSlim Quantization Documentation
• Related discussion: PaddleOCR量化模型转onnx报错

---

Conclusion

Your best options are:

1. Monitor updates for Paddle2ONNX with higher opset support.
2. Quantize the model or replace unsupported operators.
3. Use Paddle Inference or TensorRT for deployment if ONNX is not mandatory.

If you need further clarification or assistance, feel free to ask!

---

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

[Hungtran-pro]

@VishyAnand28 Did you solve the problem?