PytorchUsage.md

This md will show you how to apply onnx-tool to a pytorch model

• Pytorch export to ONNX

  import torchvision
  import torch
  model = torchvision.models.alexnet()
  x = torch.rand(1, 3, 224, 224)
  tmpfile='tmp.onnx'
  with torch.no_grad():
      torch_out = torch.onnx.export(model, x, tmpfile, opset_version=12 )  #opset 12 and opset 7 tested
      #do not use dynamic axes will simplify the process

• Pytorch MACs profile and shape inference sample

  import torchvision
  import onnx_tool
  import torch
  model = torchvision.models.alexnet()
  x = torch.rand(1, 3, 224, 224)
  tmpfile='tmp.onnx'
  with torch.no_grad():
      torch_out = torch.onnx.export(model, x, tmpfile, opset_version=12 )  #opset 12 and opset 7 tested
      #do not use dynamic axes will simplify the process
      onnx_tool.model_profile(tmpfile,saveshapesmodel='shapes.onnx')
      #you will get the print of MACs of each layer, and the hidden tensor's shapes will be export to shapes.onnx

• example code pytorch_example.py