onnx model has no weights and bias about dcnv2_latest

[maxclaey]

@ArtyZe How did you manage to export to ONNX model? When I try to export to onnx, I get RuntimeError: No Op registered for Plugin with domain_version of 9
Thanks in advance for you help!

[lucasjinreal]

@ArtyZe DCN's onnx is a custom op not listed on ONNX standard, your visualization result might not accurate. For validation, you can check your onnx inference result using onnx-runtime.

[ArtyZe]

@ArtyZe How did you manage to export to ONNX model? When I try to export to onnx, I get RuntimeError: No Op registered for Plugin with domain_version of 9
Thanks in advance for you help!

Hi, You got see code below in file dcnv2_onnx.py ?

It's a torch supported way to register ops to torch/onnx.
So 1. replace your dcnv2.py file with dcnv2_onnx.py then export model to onnx or 2. add above code to your own code and then export model to onnx

[ArtyZe]

@ArtyZe DCN's onnx is a custom op not listed on ONNX standard, your visualization result might not accurate. For validation, you can check your onnx inference result using onnx-runtime.

You are right. It's a tool bug of netron for custom ops. Data exist but netron displays nothing. It's right when I export data of this op with python

[austinmw]

@maxclaey did you find a solution to this? I'm running into the same issue. @jinfagang Do you happen to know what this error comes from?

edit: Turns out I needed operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK in torch.onnx.export()

[MoaazAbdulrahman]

@ArtyZe
I have tried your solution using dcnv2_onnx.py but I am still facing the same error RuntimeError: No Op registered for Plugin with domain_version of 9. Could you please share with me the file that you used to convert to onnx.
Thanks.