Error while converting onnx model to Mxnet

[ashkanaev]

Hi!

Found that it's impossible to converting any model from pytorch which has Upsample layer.

sym, arg, aux = onnx_mxnet.import_model("torch_model.onnx")
  File "/home/user/anaconda3/envs/mx/lib/python3.5/site-packages/mxnet/contrib/onnx/onnx2mx/import_model.py", line 54, in import_model
    sym, arg_params, aux_params = graph.from_onnx(model_proto.graph)
  File "/home/user/anaconda3/envs/mx/lib/python3.5/site-packages/mxnet/contrib/onnx/onnx2mx/import_onnx.py", line 115, in from_onnx
    mxnet_sym = self._convert_operator(node_name, op_name, onnx_attr, inputs)
  File "/home/user/anaconda3/envs/mx/lib/python3.5/site-packages/mxnet/contrib/onnx/onnx2mx/import_onnx.py", line 63, in _convert_operator
    raise NotImplementedError("Operator {} not implemented.".format(op_name))
NotImplementedError: Operator Upsample not implemented.

That means that majority of segmentation model are not convertible. Is there any trick to avoid this problem?

Hope for reply.
Thanks!

[lanking520]

@ashkanaev Thanks for your issue, I have added @Roshrini to help you out. In fact, if there is a missing operator not available in MXNet, it will be hard to find an alternative. I would mark this as a feature request to see if anyone would like to add this operator support.

@mxnet-label-bot [ONNX, feature]

[Roshrini]

@ashkanaev Thanks for reporting this. MXNet's Upsampling operator is not exactly similar to ONNX's upsample. So, this was not implemented until now. Apologies for the inconvenience.

But this is definitely in our plan and we will be working on adding support for Upsample operator soon. You are also welcome to contribute the operator implementation. :)

[Roshrini]

Related issue to update the operator: #12602

[vandanavk]

@ashkanaev Could you share the model or a link to the ONNX model that you are using?

[DuinoDu]

If your UpSampling is bilinear, you can convert it to deconv firstly and then convert to onnx.
I've tried this way, it worked for me.