leehyuck/abc

pytorch implementation of ABC : Auxiliary Balanced Classifier for Class-imbalanced Semi-supervised Learning

abc's Issues

A questions about abc loss

Sorry, I have use the wrong code. Please delete this issue. Thanks~

Can I replace wideresnetwithABC with another model?

Hi, thanks for this great work!

Find out that your model is using wideresnetwithABC.py
Does it contain the ABC of your paper?
Can you tell me which side is the ABC part in wideresnetwithABC.py

I want to add ABC to my model.
Is there anything in particular that needs attention?

Thank you for reading
Kelly

Regarding the Bernoulli Mask for Unlabeled Data

Hi,
Thanks for your work. I had one doubt regarding the bernoulli distribution of the unlabeled dataset.
Can you point out where do you calculate the B(N_L/N_qhat ) in the code.

It seems you only use the statistics of the labeled distribution ir2 in the formula 1 - (epoch/500) * (1-ie2)

Thanks

What is the purpose of bernoulli distribution?

Hi, thanks for this great work!
I want to ask about bernoulli distribution in code.
What is the purpose of bernoulli distribution?
What is the difference between ir2 and ir22 in the code?
Do maskbalance and maskbalanceu mean to give target weight?

maskforbalance = torch.bernoulli(torch.sum(targets_x2 * torch.tensor(ir2).cuda(0), dim=1).detach())

ir22 = 1 - (epoch / 500) * (1 - ir2)
maskforbalanceu = torch.bernoulli(torch.sum(label_u.cuda(0) * torch.tensor(ir22).cuda(0), dim=1).detach())

Thanks
Kelly

The intuition of performing the same strong transformation to the same image twice?

Hi, thanks for this great work!

In this code it performs the same strong transformation to the same image twice. And in the training step it also computes the loss of these two images. I'm aware that the two images are not identically the same (output of strong transformation will be slightly different in each time). I didn't find this mention in the paper, therefore I'm wondering are there any intuitions of why doing this? Thanks.

performs the same strong transformation to the same image twice:

class TransformTwice:
    def __init__(self, transform, transform2):
        self.transform = transform
        self.transform2 = transform2

    def __call__(self, inp):
        out1 = self.transform(inp)
        out2 = self.transform2(inp)
        out3 = self.transform2(inp)
        return out1, out2, out3