Mmdetection: Onnx Conversion for Faster-RCNN

[ashnair1]

Has pytorch2onnx.py ever been used to convert a standard Faster-RCNN model? While I understand it's still an experimental feature, I thought I could use pytorch2onnx to convert my Faster-RCNN model but it did not work.

[ashnair1]

To provide more context, these are the things I've tried:

I tried running pytorch2onnx.py however this creates an invalid graph and fails the onnx check (onnx.checker.check_model(onnx_model)) due to two Constant nodes containing the device name ('cpu'). The same error as in #2299. As I stated in #2299, I was able to create a valid graph by removing these two nodes.

However, I was not able to load this onnx model. I tried using onnxruntime and caffe2 onnx backend.

1. Loading with onnxruntime fails with Fail: [ONNXRuntimeError] : 1 : FAIL : Fatal error: ATen is not a registered function/op
2. Loading with Caffe2 Onnx backend (opset 9) fails with Don't know how to translate op roi_align which I guess makes sense since RoiAlign support was added opset 10 onwards. However I thought it would fall back to torchvision ops to enable conversion.

I had also tried #1386 which did allow me to create a onnx model and run inference. However, the model was giving the same outputs for different inputs. Has the team thought of integrating that pull request or is pytorch2onnx enough?

[RunningLeon]

Pls. try with the latest mmcv and mmdetection. Reopen it if you still have any issue.