Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Gradient penalty "interpolates" term #176

Open
rl-max opened this issue Nov 3, 2022 · 1 comment
Open

Gradient penalty "interpolates" term #176

rl-max opened this issue Nov 3, 2022 · 1 comment

Comments

@rl-max
Copy link

rl-max commented Nov 3, 2022

Hi,

This is 'cal_grad_penalty' function in /src/utils/losses.py

def cal_grad_penalty(real_images, real_labels, fake_images, discriminator, device):
    batch_size, c, h, w = real_images.shape
    alpha = torch.rand(batch_size, 1)
    alpha = alpha.expand(batch_size, real_images.nelement() // batch_size).contiguous().view(batch_size, c, h, w)
    alpha = alpha.to(device)

    real_images = real_images.to(device)
    interpolates = alpha * real_images + ((1 - alpha) * fake_images)
    interpolates = interpolates.to(device)
    interpolates = autograd.Variable(interpolates, requires_grad=True)
    fake_dict = discriminator(interpolates, real_labels, eval=False)
    grads = cal_deriv(inputs=interpolates, outputs=fake_dict["adv_output"], device=device)
    grads = grads.view(grads.size(0), -1)

    grad_penalty = ((grads.norm(2, dim=1) - 1)**2).mean() + interpolates[:,0,0,0].mean()*0
    return grad_penalty

In the last line, grad_penalty = ((grads.norm(2, dim=1) - 1)**2).mean() + interpolates[:,0,0,0].mean()*0, I wanted to know what additive term + interpolates[:,0,0,0].mean()*0 means. Since it's zero-multiplicated, I think it has actually no effect for code.

I'll be waiting for your answer

Thank you!

@mingukkang
Copy link
Collaborator

The implementation has no effect on the code and was introduced to address a PyTorch bug that arises during DDP training with R1 regularization.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants