Issue with setting margin for HingeEmbeddingLoss

1

I’m having trouble with the HingeEmbeddingLoss in pytorch 0.1.11.

When I try to specify the margin it throws an error.

import torch
mrl = torch.nn.MarginRankingLoss(margin=0.5)  # this works
hel = torch.nn.HingeEmbeddingLoss()  # this works

# this throws an error - __init__() got an unexpected keyword argument margin
hel_m = torch.nn.HingeEmbeddingLoss(margin=0.5)

Looking at the source in nn/_functions/loss.py, it seems to be defined correctly.

Thanks

2

I think I got to the bottom of this. For some reason the HingeEmbeddingLoss class is not defined in: https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/loss.py#L202

As opposed to MarginRankingLoss where the class is set up correctly: https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/loss.py#L365

Not sure why though.

3

that’s weird. Both are even imported side by side: https://github.com/pytorch/pytorch/blame/master/torch/nn/modules/init.py

4

either ways this is fixed in master and will be part of next release.

5

Thanks for your response. I had a quick look at the master branch for loss.py but it has the same problem?

Instead of pass at line 220 should it not contain something like:

class HingeEmbeddingLoss(_Loss):
    # Creates a criterion ...
    def __init__(self, margin=1.0, size_average=True):
        super(HingeEmbeddingLoss, self).__init__()
        self.margin = margin
        self.size_average = size_average

    def forward(self, input1, target):
        return self._backend.HingeEmbeddingLoss(self.margin,
                                               self.size_average)(input1, target)
6

actually you’re right. this does give an error on master too. Needs to be fixed like you specified.
Can you send a pull request to patch this? thanks for finding the issue.

7

I made a pull request here:

Thanks