Cannot import alexnet model from pytorch with ONNX

[ieee8023]

I cannot import the alexnet model from pytorch. I receive this error:

  File "test.py", line 5, in <module>
    sym, arg, aux = mx.contrib.onnx.import_model('alexnet.onnx')
  File "/Users/ieee8023/Library/Python/2.7/lib/python/site-packages/mxnet/contrib/onnx/onnx2mx/import_model.py", line 53, in import_model
    sym, arg_params, aux_params = graph.from_onnx(model_proto.graph)
  File "/Users/ieee8023/Library/Python/2.7/lib/python/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 "/Users/ieee8023/Library/Python/2.7/lib/python/site-packages/mxnet/contrib/onnx/onnx2mx/import_onnx.py", line 61, in _convert_operator
    op_name, new_attrs, inputs = convert_map[op_name](attrs, inputs, self)
  File "/Users/ieee8023/Library/Python/2.7/lib/python/site-packages/mxnet/contrib/onnx/onnx2mx/_op_translations.py", line 418, in reshape
    reshape_shape = list(proto_obj._params[inputs[1].name].asnumpy())
KeyError: 'concat0'

Here is code to reproduce the issue

import torch, torchvision
import mxnet, mxnet.contrib.onnx

model = torchvision.models.alexnet(pretrained=True)
model.train(False)

dummy_input = torch.autograd.Variable(torch.randn(1, 3, 224, 224))
torch.onnx.export(model, dummy_input, 'alexnet.onnx')

sym, arg, aux = mxnet.contrib.onnx.import_model('alexnet.onnx')

>>> import mxnet
>>> mxnet.__version__
'1.3.0'
>>> import onnx
>>> onnx.__version__
u'1.3.0'

[vdantu]

@mxnet-label-bot add [question, onnx]

[vandanavk]

@ieee8023 Thanks for reporting this issue. I'm able to replicate this issue on my setup, will debug this and update you soon.

[vandanavk]

Related issues: #10789, onnx/onnx-mxnet#50.

MXNet expects shape to be an attribute but the more recent versions of ONNX have defined the shape to be a tensor.

[piyushghai]

@vandanavk Is this issue still persisting ?

[vandanavk]

Yes @piyushghai. This can't be fixed until Dynamic shape is implemented in MXNet. Ref: #10789 (comment)