likaiucas / pytorch_model_2_mindspore

Pytorch_Model_2_Mindspore

First use comapre.py to generate 2 .txt about the dictionary of 2 state_dict.

Then, with the help of feature from https://github.com/WinMerge/winmerge, we manually create an mapping that represents the relationship of two "theorotically" same models in physical. And create transdict.py which supports the mapping.

Finally, use convert_darkent.py to translate .pth to .ckpt

some common relation:

 mindspore.ckpt <-> torch.pth
 convolution: weight <-> weight
 batchnorm:   beta   <-> weight
              gamma  <-> bias
              moving_mean <-> running_mean
              moving_variance <-> running_var
              None <-> num_batches_tracked