Incubator-mxnet: onnx converter error

[szha]

reported at https://discuss.gluon.ai/t/topic/7396/3, the loading of shufflenet from onnx model zoo into mxnet fails.

[szha]

The reported error message:

KeyError                                  Traceback (most recent call last)
<ipython-input-5-7912d84df342> in <module>()
      1 if __name__ == '__main__':
      2     #print(50000/16*15)
----> 3     loadShuffleNet()
      4 

<ipython-input-4-c7bde65da337> in loadShuffleNet()
     22     path = '/media/jiaming/Seagate Backup Plus Drive/pretrain_model/shufflenet/'
     23     onnx_path = path + 'model.onnx'
---> 24     sym, arg_params, aux_params = onnx_mxnet.import_model(onnx_path)
     25     (new_sym, new_args) = get_sn_finetune(sym, arg_params, 7)
     26 

~/.local/lib/python3.6/site-packages/mxnet/contrib/onnx/_import/import_model.py in import_model(model_file)
     51     # loads model file and returns ONNX protobuf object
     52     model_proto = onnx.load(model_file)
---> 53     sym, arg_params, aux_params = graph.from_onnx(model_proto.graph)
     54     return sym, arg_params, aux_params
     55 

~/.local/lib/python3.6/site-packages/mxnet/contrib/onnx/_import/import_onnx.py in from_onnx(self, graph)
    112             onnx_attr = self._parse_attr(node.attribute)
    113             inputs = [self._nodes[i] for i in node.input]
--> 114             mxnet_sym = self._convert_operator(node_name, op_name, onnx_attr, inputs)
    115 
    116             for k, i in zip(list(node.output), range(len(mxnet_sym.list_outputs()))):

~/.local/lib/python3.6/site-packages/mxnet/contrib/onnx/_import/import_onnx.py in _convert_operator(self, node_name, op_name, attrs, inputs)
     56         """
     57         if op_name in convert_map:
---> 58             op_name, new_attrs, inputs = convert_map[op_name](attrs, inputs, self)
     59         else:
     60             raise NotImplementedError("Operator {} not implemented.".format(op_name))

~/.local/lib/python3.6/site-packages/mxnet/contrib/onnx/_import/op_translations.py in batch_norm(attrs, inputs, cls)
    172 
    173     # in test mode "fix_gamma" should be unset.
--> 174     new_attrs['fix_gamma'] = 0 if new_attrs['fix_gamma'] == 1 else 1
    175     return 'BatchNorm', new_attrs, inputs
    176 

KeyError: 'fix_gamma'

[anirudhacharya]

The ONNX specification recently changed the definition of the BatchNorm operator, see here opset_version7 and op_setversion6.

So currently there are many models that use opset_version6 and some that use opset_version7. This error is caused due to this discrepancy in the operator definition. I will try to prioritize and get a patch fix to handle both versions of the operator and merge it into the master.

[Roshrini]

This issue for BatchNorm operator is fixed in latest code. User is using old code. Ran both opset 6 and 7 of shufflenet model. It works fine. Cannot reproduce the issue. Added comment on forum: https://discuss.mxnet.io/t/error-in-loading-pretrained-shufflenet-onnx-model/1520/3?u=roshrini

[anirudhacharya]

@szha please close this issue if resolved.

[szha]

@anirudhacharya Let's wait for @jmnie's confirmation

[jmnie]

@anirudhacharya @szha Sorry for late reply. I just load the model using the latest mxnet. It works. Thank you very much.

[keetsky]

Have solved? I met the same issue.

[keetsky]

If i let mx.sym.BatchNorm(data=fc1, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn3')
and save to onnx model,but when i load onnx mode ,find that mxnet default set fix_gamma to False in _op_translations.py # in test mode "fix_gamma" should be unset. new_attrs['fix_gamma'] = not attrs.get('is_test', 1) ,so my predict results are diffrent, how to fix it?