9

I have been trying to import a model from onnx format to work with pytorch. I am finding it difficult to get an example for the same. As most of the resources in Internet talks about exporting a pytorch model to onnx.

I found that torch.onnx() can only export the model and the import method hasn't been implemented yet. A direct installation of onnx library, helps me to do onnx.load("model_name.onnx"). How do I use this model with pytorch? I am not able to move the model to GPU with model.to(device="GPU")

desertnaut
  • 57,590
  • 26
  • 140
  • 166
Ajai
  • 1,049
  • 1
  • 11
  • 23

1 Answers1

16

PyTorch doesn't currently support importing onnx models. As of writing this answer it's an open feature request.

While not guaranteed to work, a potential solution is to use a tool developed by Microsoft called MMdnn (no it's not windows only!) which supports conversion to and from various frameworks. Unfortunately onnx can only be a target of a conversion, and not a source. That said you may be able to import your model to another framework, then use MMdnn to convert from that framework to pytorch. Obviously this isn't ideal and the potential for success will depend on how other frameworks use onnx which may not be amenable to the way MMdnn works.

Update August 2022

Unfortunately it appears the feature request was rejected and Mmdnn has been abandoned. There are some more recent 3rd party tools that provide some ability to import onnx into pytorch like onnx2pytorch and onnx-pytorch. Neither of these tools appear to be actively developed, though pytorch and onnx are relatively stable at this point so hopefully these tools remain relevant in the future (official support would be better IMO). Note that both of these tools have unaddressed issues, so it may be necessary to try both of them if one doesn't work for you.

Update September 2022

Based on the comment from @DanNissenbaum there is a newer 3rd party tool onnx2torch that is being actively developed and maintained.

jodag
  • 19,885
  • 5
  • 47
  • 66