How to build caffe2 with ONNX opset version greater than 9?

deployment
1

Hello,
I’ve currently worked with freshly merged feature pytorch/vision#1401 and won’t able to find a way to make Caffe2 work with ONNX operation set 10?

Is there a way to build a Caffe2 from source with this opset?

2

I don’t know but I would venture that arguably the go-to library for ONNX model serving is ONNXRuntime these days.

See also Caffe2 vs libtorch, realistically .

Best regards

Thomas

3

Indeed. We’ve removed tutorials around this as the Caffe2 path really isn’t being maintained. Assuming you are focused on using ONNX (and not deploying via Torchscript), to leverage the latest opset, it makes more sense to export an ONNX model from PyTorch and then deploy in ONN Runtime as @tom suggests. Here is a tutorial that is maintained by the MSFT team: https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html

4

Thanks a lot, I’ve got it. Caffe2 right now is totally out maintenance

5

Thank you too. Let me clarify why I was eager to find a way to make Caffe2 work with recent opset. I’ve recently worked with your repo with AICamera that manages PyTorch v1.0.0 to work with Android devices using Caffe2 library. So I know that right now it is TorchScript available in mobile and onnx-runtime doesn’t work with mobile. So I wondered how to deploy on mobile ONNX model with recent opset using Caffe2 backend.
P.S. The last statement is not a question just a theme to discuss

1 Like