Custom operator support for conversion from MXNet to PyTorch

[farleylai]

Platform: ubuntu 16.04: Python version: 3.6 Source framework with version: mxnet-cu92mkl 1.2.0 Destination framework with version: pytorch-0.4 with GPU Running scripts:

SYM=mxnet/Deformable-ConvNets/rfcn/resnet_v1_101_rfcn_dcn.json
PARAMS=mxnet/Deformable-ConvNets/params/rfcn_dcn_coco-0000.params

mmconvert -sf mxnet -in $SYM -iw $PARAMS --dstNodeName MMdnn_Output -df pytorch -om res101_rfcn_dcn_coco.pth --inputShape 3 224 224

Error:

Traceback (most recent call last):
  File "/home/ml/farleylai/anaconda3/lib/python3.6/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/home/ml/farleylai/anaconda3/lib/python3.6/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 189, in <module>
    _main()
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 184, in _main
    ret = _convert(args)
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 68, in _convert
    parser = MXNetParser(model)
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 239, in __init__
    self.model, self.weight_data = MXNetParser._load_model(input_arg[1], input_arg[2])
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mmdnn/conversion/mxnet/mxnet_parser.py", line 172, in _load_model
    sym, arg_params, aux_params = mx.model.load_checkpoint(weights, int(epoch))
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mxnet/model.py", line 419, in load_checkpoint
    symbol = sym.load('%s-symbol.json' % prefix)
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mxnet/symbol/symbol.py", line 2535, in load
    check_call(_LIB.MXSymbolCreateFromFile(c_str(fname), ctypes.byref(handle)))
  File "/home/ml/farleylai/anaconda3/lib/python3.6/site-packages/mxnet/base.py", line 149, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [00:13:37] src/operator/custom/custom.cc:104: Check failed: CustomOperator::Get()->Find(params.op_type) != nullptr Cannot find custom operator proposal

The pre-trained model is downloaded from the Deformable-ConvNets which contains a custom operator 'proposal.'

The custom operator is a python module in the source tree. Is there any way to instruct MMdnn to import the custom operator? Or the custom operator for conversion from MXNet is not suppored yet?

[namizzz]

Hi @farleylai ,Could you provide your model file(.json and .params) for us to test?

[farleylai]

Sure, the network symbol is generated by calling tojson() as attached FYI. The network source is located at resnet_v1_101_rfcn_dcn.py.

resnet_v1_101_rfcn_dcn.json.zip

The params is quite large (~254MB) but I believe you can download from the official site (stored on OneDrive with file named rfcn_dcn_coco-0000.params).

The CustomOp proposal source is available in Python or by calling C++ contrib APIs.

Let me know if you need something else.

[kitstar]

Hi @farleylai . In current MMdnn, custom operator is not supported yet. Thanks.