basiclab / gngan-pytorch Goto Github PK

I am very interested in the method of the paper, but when I add GN to my own training process there is a runtime error, can you please tell me why? Thank you!

hi, great job!
But I have two concern:

1. detach torch.abs(f) maybe more faster and reasonable?
2. f_hat is a discontinuous function. How does it work?

Could you please provide me with the trained weight parameter file of CelebA_HQ128128? I can't train because I don't have a multi-card GPU, but I would like to see the generating effect of your model on the dataset CelebA_HQ128128. Thank you very much

After integrating Gradient Normalization into CycleGAN, the GPU memory usage keeps ceaselessly growing, and finally, it blows up.

with module_no_grad([netG_A2B]):
    optimizer_D_A.zero_grad()
    pred_real = normalize_gradient(net_D=netD_A, x=real_A)
    pred_fake = normalize_gradient(net_D=netD_A, x=fake_A.detach())
    loss_D_A = -pred_real.mean() + pred_fake.mean() 
    loss_D_A.backward()
    optimizer_D_A.step()

I've tried to del pred_real, pred_fake, loss_D_A in the code above, even del grad_norm, f in gradnorm.py, and it does not help.
The GN sounds great to my work.
And any tips would be appreciated.

Hi, do you have this file link? cifar100.train.npz
because it will download automatically.
Thank you!

Hello,
I found a very interesting technique.

There may be a mistake about "How to integrate Gradient Normalization into your work?" of Readme.md.

Now it's the following:

loss_real = loss_fn(pred_real, torch.ones_like(pred_real))
loss_fake = loss_fn(pred_fake, torch.ones_like(pred_fake))

Is the following correct?

loss_real = loss_fn(pred_real, torch.ones_like(pred_real))
loss_fake = loss_fn(pred_fake, torch.zeros_like(pred_fake))

The losses.py in your repository was exactly like this.

Thank you.

Hi,

Nice paper, is it any plan to release the pre-trained model?

Thanks!

execuse me, the function "get_inception_score_and_fid" could you have