-
Notifications
You must be signed in to change notification settings - Fork 2k
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
Holes showing up while training #156
Comments
Hi! I've been plagued by the same problems for a long time. Have u solved this problem? |
Unfortunately not. I am still getting the same problem. |
Sorry to hear that. After my comment on this issue, I find two issues highly related to our problems and the author of CycleGAN has answered them. He thinks these artifacts are caused by mode collapse and more training data, larger loss weights of identity/cycle consistency loss, or smaller learning rate can solve this problem. I'm trying them now, maybe you can try these solutions, too. Good luck for us! BTW, here are the issues I mentioned: |
Thank you! That helps a lot! |
Hi, so I've been trying to replicate your paper by creating a PyTorch model from scratch and training it on the original vangogh2photo dataset provided by Berkeley. Admittedly, it's for fun and not for any research, but I still hate it when it doesn't work out. So, this is the architecture of the model I've made:
Now around the third epoch of training, I am getting these "holes" in the generated pictures. Could anyone tell me why these are showing up and how I can prevent it? These are my hyperparameters:
'name': 'CycleGan_VanGogh_Checkpoint', 'n_epochs': 20, 'batch_size': 4, 'lr': 0.0002, 'decay_start_epoch': 19, 'b1': 0.5, 'b2': 0.999, 'img_size': 256, 'channels': 3, 'num_residual_blocks': 9, 'lambda_cyc': 10.0, 'lambda_id': 5.0}
The text was updated successfully, but these errors were encountered: