Cannot convert models of Insight face(mxnet) to IR

[stereomatchingkiss]

Platform : win10 64 bits

Python version: Anaconda 3.6

Source framework with version: mxnet 1.3.0

Destination framework with version: Ultimate goal is convert to caffe model, but fail to convert to IR

Pre-trained model path : insight face model zoo, converting MobileFaceNet

Running scripts(try all of them):

1. python -m mmdnn.conversion._script.convertToIR -f mxnet -n model-symbol.json -w model-0000.params -d mobile_face_net --inputShape 1 3 112 112
2. python -m mmdnn.conversion._script.convertToIR -f mxnet -n model-symbol.json -w model-0000.params -d mobile_face_net --inputShape 3 112 112
3. python -m mmdnn.conversion._script.convertToIR -f mxnet -n model-symbol.json -w model-0000.params -d mobile_face_net --inputShape 112 112

Error message:

 File "C:\Users\yyyy\Anaconda3\lib\site-packages\mxnet\base.py", line 252, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: Error in operator conv0: [00:40:56] c:\jenkins\workspace\mxnet\mxnet\src\operator\nn\convolution.cc:148: Check failed: dshp.ndim() == 4U (2 vs. 4) Input data should be 4D in batch-num_filter-y-x

Any suggestions?Thanks

[rainLiuplus]

Hi @stereomatchingkiss , The --inputShape should be 3,112,112.

[stereomatchingkiss]

Hi @stereomatchingkiss , The --inputShape should be 3,112,112.

Thanks, the dimension problem solved, but I found new issues. When I convert the resnet34 models of insight face, I saw following warnings

 C:\Users\yyyy\Anaconda3\lib\site-packages\mxnet\module\base_module.py:56: UserWarning: You created Module with Module(..., label_names=['softmax_label']) but input with name 'softmax_label' is not found in symbol.list_arguments(). Did you mean one of:
        data
  warnings.warn(msg)
Warning: MXNet Parser has not supported operator null with name data.
Warning: convert the null operator with name [data] into input layer.
IR network structure is saved as [mobile_face_net.json].
IR network structure is saved as [mobile_face_net.pb].
IR weights are saved as [mobile_face_net.npy].

Although I can convert the models to IR files, but not sure this is safe or not

When I convert MobileFaceNet, saw error messages

C:\Users\yyyy\Anaconda3\lib\site-packages\mxnet\module\base_module.py:56: UserWarning: You created Module with Module(..., label_names=['softmax_label']) but input with name 'softmax_label' is not found in symbol.list_arguments(). Did you mean one of:
        data
  warnings.warn(msg)
Warning: MXNet Parser has not supported operator null with name data.
Warning: convert the null operator with name [data] into input layer.
Traceback (most recent call last):
  File "C:\Users\yyyy\Anaconda3\lib\runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "C:\Users\yyyy\Anaconda3\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\_script\convertToIR.py", line 197, in <module>
    _main()
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\_script\convertToIR.py", line 192, in _main
    ret = _convert(args)
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\_script\convertToIR.py", line 115, in _convert
    parser.run(args.dstPath)
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\common\DataStructure\parser.py", line 22, in run
    self.gen_IR()
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\mxnet\mxnet_parser.py", line 262, in gen_IR
    func(current_node)
  File "C:\Users\yyyy\Anaconda3\lib\site-packages\mmdnn\conversion\mxnet\mxnet_parser.py", line 398, in rename_FullyConnected
    weight = self.weight_data.get(source_node.name + "_weight").asnumpy().transpose((1, 0))
AttributeError: 'NoneType' object has no attribute 'asnumpy'

[namizzz]

Hi @stereomatchingkiss, this warning doesn't matter. I find there is no 'pre_fc1_weight' object in the self.weight_data. But I find the 'fc1_weight' in the self.weight_data in the shape of (128, 512). I think it may be the weight of the fully connected layer 'pre_fc1' which the num_hidden is 128. Maybe you can change this line:

weight = self.weight_data.get(source_node.name + "_weight").asnumpy().transpose((1, 0))

into

weight = self.weight_data.get(source_node.name.split('pre_')[-1] + "_weight").asnumpy().transpose((1, 0))

I tried, it works. Thanks!

[kaishijeng]

@namizzz

I am able to convert mobilefacenet with your suggestion. My goal is to convert to pytorch or tensorflow. DO you know how to convert this IR model to tensorflow or pytorhc?

Thanks,

[kaishijeng]

@namizzz

Do you know how to convert this IR model to tensorflow or pytorch?
Your help will be appreciated.

Thanks, FC

[nakohli]

I got it converted to pytorch by using the below command

mmtocode -f pytorch -n mobile_face_net.pb -w mobile_face_net.npy -d torch_converted.py -dw torch_converted.pt

However the major issue is that PRelu operation is not yet supported as a pytorch emitter.

@namizzz Can you suggest what extra changes will have to be made for this model to be successfully converted to pytorch model

Thanks

[tengerye]

Hi, any progress on this issue? I have a similar problem #499 @rainLiuplus @namizzz