AttributeError: module 'torch.nn' has no attribute 'BCEWithLogitsLoss'

[wasiahmad]

I am getting the following error while trying to use the BCEWithLogitsLoss from torch.nn.

AttributeError: module 'torch.nn' has no attribute 'BCEWithLogitsLoss'

I am wondering why it is happening?

[hughperkins]

In [1]: import torch

In [2]: dir(torch.nn)
Out[2]: 
['AdaptiveAvgPool1d',
 'AdaptiveAvgPool2d',
 'AdaptiveMaxPool1d',
 'AdaptiveMaxPool2d',
 'AvgPool1d',
 'AvgPool2d',
 'AvgPool3d',
 'BCELoss',
 'BatchNorm1d',
 'BatchNorm2d',
 'BatchNorm3d',
 'Bilinear',
 'ConstantPad2d',
...

No such method, only BSELoss exists. What makes you feel BCEWithLogitsLoss exists?

[wasiahmad]

See the followings.

• http://pytorch.org/docs/master/nn.html?highlight=logitsloss#torch.nn.BCEWithLogitsLoss
• http://pytorch.org/docs/master/nn.html?highlight=binary#torch.nn.functional.binary_cross_entropy_with_logits

[hughperkins]

The first two are in master branch. Are you building master from source, or are you using a binary release version? eg what is the result of:

import torch
print(torch.version.__version__)

?

[wasiahmad]

Oh, no. I am not building master from source. I am using pytorch 0.1.12_2 binary version.

How can I use those two loss functions?

[hughperkins]

Build from source

[wasiahmad]

Can you point me to a documentation to do that?

[hughperkins]

pytorch/pytorch

[Marvin (YL)]

It’s a new feature. You may use some hacky ways to include that from github repo, or install pytorch from source.