how to add a converter for custom operateion written by c++

[deemoo-wang]

the model i used has some opration written by c++, using torch.utils.cpp_extention with a .cpp file. I want to convert the model into trt model but failed with the error NVALID_ARGUMENT: Cannot find binding of given name: input_0. how can i write a converter for the operation written by c++?

[grimoire]

• First, you need to implement the custom ops you need as a TensorRT plugin, here are some examples.
• After you have your plugin, add a creator in torch2trt_dynamic/plugins, note that if you did not implement your plugin inside amirstan_plugin, you need to load library ctypes.CDLL('xxxxx.so')
• Finally, add your convertor in torch2trt_dynamic/converters. You can read cummax.py as an example.

Note that since there are some limits on TensorRT, I can not promise that your custom ops can be implemented as a TensorRT plugin.