search

microsoft / MMdnn

MMdnn is a set of tools to help users inter-operate among different deep learning frameworks. E.g. model conversion and visualization. Convert models between Caffe, Keras, MXNet, Tensorflow, CNTK, PyTorch Onnx and CoreML.
MIT License
5.8k stars 965 forks source link

Error converting Mobilenet v2 from pytorch to tensorflow #884

Open radu-diaconescu13 opened 4 years ago

radu-diaconescu13 commented 4 years ago

Platform (ubuntu 16.04):

Python version: Python 3.6.2

Source framework with version pytorch 1.6 CPU:

Destination framework with version (tensorflow 1.13):

Pre-trained model path (webpath or webdisk path):

Model - download with mmdownload

mmdownload -f pytorch -n mobilenet_v2 -o ./

Running scripts:

mmdownload -f pytorch -n mobilenet_v2 -o ./ mmtoir -f pytorch -d mobilenet_v2 --inputShape 3,224,224 -n imagenet_mobilenet_v2.pth

Error message:

/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/torch/onnx/symbolic_helper.py:438: UserWarning: ONNX export mode is set to inference mode, but operator dropout is set to training  mode. The model will be exported in inference, as specified by the export mode.
  training_mode + ", as specified by the export mode.")
Traceback (most recent call last):

  File "/home/radud/anaconda3/envs/model-converter/bin/mmtoir", line 8, in <module>
    sys.exit(_main())

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 197, in _main
    ret = _convert(args)

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 97, in _convert
    parser = PytorchParser151(model, inputshape[0])

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 523, in __init__
    self.build_graph(input_shape)

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 92, in build_graph
    self.pytorch_graph.build(self.input_shape)

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/pytorch/pytorch_graph.py", line 143, in build
    output_shape = [int(x.replace('!', '')) for x in output_shape_str.split(',')]

  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/pytorch/pytorch_graph.py", line 143, in <listcomp>
    output_shape = [int(x.replace('!', '')) for x in output_shape_str.split(',')]

ValueError: invalid literal for int() with base 10: '1:401408'

Hello. I have tried converting the mobilenetv2 model downloaded with the mmdownload command and have encountered this error. I have also had similar issues when trying to convert a mobilenetv2 which was first converted from TF to pytorch using mmdnn. By similar issues, I mean it can not parse the output shape. Can you please help me? Thank you

radu-diaconescu13 commented 4 years ago

I have also tried converting my model using pytorch 1.5.1 and get a different error now

Traceback (most recent call last):
  File "/home/radud/anaconda3/envs/model-converter/bin/mmtoir", line 8, in <module>
    sys.exit(_main())
  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 197, in _main
    ret = _convert(args)
  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/_script/convertToIR.py", line 120, in _convert
    parser.run(args.dstPath)
  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/parser.py", line 22, in run
    self.gen_IR()
  File "/home/radud/anaconda3/envs/model-converter/lib/python3.6/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 100, in gen_IR
    node_type = PytorchParser.layer_map[onnx_node_type]
KeyError: 'onnx::Clip'

seems related to issue: https://github.com/microsoft/MMdnn/issues/872