I am trying to export this segmentation model to ONNX. On export I get the error:
/home/diego/.pyenv/versions/gsoc/lib/python3.6/site-packages/torch/onnx/utils.py:662: UserWarning: ONNX export failed on ATen operator upsample_bilinear2d because torch.onnx.symbolic_opset9.upsample_bilinear2d does not exist
.format(op_name, opset_version, op_name))
*** KeyError: 'upsample_bilinear2d'
I found this issue and this fix saying the problem was fixed on master. However I am working on PyTorch Nightly and I am still getting this error. In this model it comes its various nn.function.interpolate
operations, same as in the example referenced in the issue.
I am also facing this problem when exporting the ‘fcn_resnet101’ from torchvision, since it has these operations as well.