[SOLVED] Class Weight for BCELoss

22

yes, that should solve it

23

Its not working. None of my class2 samples are classified correctly.

25

I made a loss function for this:

def BCELoss_ClassWeights(input, target, class_weights):
    # input (n, d)
    # target (n, d)
    # class_weights (1, d)
    input = torch.clamp(input,min=1e-7,max=1-1e-7)
    bce = - target * torch.log(input) - (1 - target) * torch.log(1 - input)
    weighted_bce = (bce * class_weights).sum(axis=1) / class_weights.sum(axis=1)[0]
    final_reduced_over_batch = weighted_bce.mean(axis=0)
    return final_reduced_over_batch
26

torch.clamp(input, 1e-9, 1-1e-9) won’t work with torch.float32 input. It will simply clamp to 0 or 1.
Using 1e-7 might be more appropriate.

>>> torch.clamp(torch.tensor(2.), max=1-1e-9) == 1
tensor(True)
>>> torch.clamp(torch.tensor(2.), max=1-1e-7) == 1
tensor(False)
27

Hi @miguelvr,

First of all thank you for having opened this topic.
I have a question for you: what if I initialize the criterion once for every batch with a proper weight torch tensor (that is batch dependent)?
Would it not properly work?

Thank you in advcanced