PyTorch-adapt Multi-label classifications

[AlexandrByzov]

Hi Kevin,

I have a multi-label classification problem, and I wanted to use PyTorch-adapt. I wrote a code for four hooks, but the model weights do not change. Is there something I need to change in hooks for multi-label classifications to work?

Best regards,
Alex

[KevinMusgrave]

You need to change CLossHook's loss function to torch.nn.BCEWithLogitsLoss. The hook you're using should allow you to pass in a custom c_hook or something like that. I can help more with this if you tell me which hook you're using.

An alternative is to "monkey patch" the init function of CLossHook:

import torch
from pytorch_adapt.hooks import CLossHook

def init_modifier(method):
    def modify(self, *args, **kwargs):
        method(self, *args, **kwargs)
        self.loss_fn = torch.nn.BCEWithLogitsLoss()
    return modify

CLossHook.__init__ = init_modifier(CLossHook.__init__)

Then every hook that uses CLossHook will now be using torch.nn.BCEWithLogitsLoss.

You might get a type error during training with torch.nn.BCEWithLogitsLoss, in which case you can use this class instead (it just converts the input labels to floating point):

from pytorch_adapt.hooks.classification import BCEWithLogitsConvertLabels

[AlexandrByzov]

Very cool, thank you very much!!!
I do experimental research on domain adaptation techniques for ECG data, so I wanted to check most of the hooks)