Linking the Custom plugins for the Conversion(while converting from ONNX to TRT Model)

[Darshcg]

Hi @TrojanXu,

Thanks for your wonderful work.

I have Implemented two custom Plugins for two unsupported Ops. I want to link those custom Plugins for the conversion(while converting from ONNX to TRT Model). Can you please assist me in doing the same? Would be a great help.

Thanks and Regards,
Darshan

[Darshcg]

Hi @TrojanXu ,

Actually, I am facing difficulty during converting the ONNX model to TRT with a Custom plugin. It would be a great help if you can help me in the same.

Waiting for your reply.

Thanks,
Darshan

[TrojanXu]

With latest TRT release, you can modify the onnx file to map the onnx op to a TRT plugin op, refer to https://github.com/TrojanXu/onnxparser-trt-plugin-sample/blob/master/test_plugin_result.py#L116

[Darshcg]

Hi @TrojanXu,

Thanks for your kind reply.

For now, I have created the .so files for the two custom Plugins. What are the next steps? (what are the commands?). As I am doing this for the first time, your help would be highly appreciable. Do I need to do anything with the TRT OSS Project?.

Environment:
TRT Version: 7.1.3
CUDA: 10.2
OS: Ubuntu 18.04.
GPU: RTX 2060.

Thanks and Regards,
Darshan C G

[Darshcg]

Hi @TrojanXu ,

I got to know about the same. It is about including the .so file while conversion.
Example: ./trtexec --onnx=model.onnx --explicitBatch --shapes=input1:1x1x1x1 --plugins=layer.so

But when I run this I am getting an issue as: Einsum_194: PluginV2Layer must be V2DynamicExt when there are runtime input dimensions

The image of the error has been attached below:

Can you please help me to resolve this error? Where Exactly I am going wrong? what's the solution for this?

Thanks and Regards,
Darshan C G

[TrojanXu]

You used "--shapes=input1:1x1x1x1"， it means the network is using dynamic shape. Just as the error said, you need to extend V2DynamicExt class instead of PluginV2