|
torch.mean(torch.abs(pred_fake[i][j][:B] - pred_fake[i][j][B:]).view(B,-1),dim=1) |
Awesome work! I was amazed about how your simple trick improved diversity so much.
Small question: Going through your code (see above) and paper I feel like that there is a minor typo in equation (6) in your paper. I think it should be the feature matching loss with respect to the generated images, right? Meaning:
Also, have you tried using the latent variable clipping trick from the BigGAN paper, meaning using standard normal for training and truncated for testing? I feel like this might be an easy way to further improve quality of generation.
DSGAN/Inpainting/train.py
Line 223 in 9747be1
Awesome work! I was amazed about how your simple trick improved diversity so much.
Small question: Going through your code (see above) and paper I feel like that there is a minor typo in equation (6) in your paper. I think it should be the feature matching loss with respect to the generated images, right? Meaning:
Also, have you tried using the latent variable clipping trick from the BigGAN paper, meaning using standard normal for training and truncated for testing? I feel like this might be an easy way to further improve quality of generation.