Add LogSoftmax, Softmax, and Dropout modules

I have used a pytorch VGG-16 pre-trained model, added 2 more layers to it to fit my needs and trained the model and saved it as a .pth file. However, when I am trying to convert it to a CrypTen model using the .load() function. It gives the following error

<ipython-input-41-8acf3ddeab54> in <module>()
      1 dummy_model = dummy_model
----> 2 plaintext_model = crypten.load('/content/drive/My Drive/vgg16-model.pth', dummy_model=dummy_model, src=ALICE)
      3 
      4 # Encrypt the model from Alice:
      5 

/usr/local/lib/python3.6/dist-packages/crypten/__init__.py in load(f, encrypted, dummy_model, src, **kwargs)
    233 
    234                 # raise error
--> 235                 raise TypeError("Unrecognized load type %s" % type(result))
    236 
    237         # Non-source party

TypeError: Unrecognized load type <class 'dict'>

I saw in the crypten/__init__.py file that this error is in the block where it is checked whether result which is the return value of torch.load(f, **kwargs), is of type tensor or torch.nn.Module and this error is raised when it is neither of them.

So to check, I ran the result = torch.load(f, **kwargs) command separately and that gave the following error -

AttributeError                            Traceback (most recent call last)

/usr/local/lib/python3.6/dist-packages/torch/serialization.py in _check_seekable(f)
    190     try:
--> 191         f.seek(f.tell())
    192         return True

5 frames

AttributeError: 'VGG' object has no attribute 'seek'


During handling of the above exception, another exception occurred:

AttributeError                            Traceback (most recent call last)

/usr/local/lib/python3.6/dist-packages/torch/serialization.py in raise_err_msg(patterns, e)
    185                                 " Please pre-load the data into a buffer like io.BytesIO and" +
    186                                 " try to load from it instead.")
--> 187                 raise type(e)(msg)
    188         raise e
    189 

AttributeError: 'VGG' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.

Is there any way to solve this issue so that I can use CrypTen with my pre-trained VGG-16 model?

[noobyogi0010]

Hi @lvdmaaten ! I would like to add the Softmax module. Thanks.

[vshobha]

The Dropout, Softmax and LogSoftmax modules have been added, and their descriptions are available in the documentation. https://crypten.readthedocs.io/

I am closing this issue, feel free to re-open if necessary.