Learning rate decay for cat2dog dataset

[korawat-tanwisuth]

Learning rate decay for cat2dog dataset

[HelenMao]

Hi, in fact, in our CVPR setting we do not include any decay in the cat2dog dataset and only run 2000 epochs.
You can try decay=1000 if you like, but make sure you should make the training of the baseline model and MSGAN consistent if you want to compare these two models.
BTW, my experience when I am working on the Journal version is that when using decay learning rate the mode collapse of DRIT can be slightly mitigated. However, MSGAN still further improves the baseline model.

[korawat-tanwisuth]

Hi @HelenMao,

Thank you so much for your response! Do you still update the content discriminator for every 3 iteration for the cat2dog dataset?

[HelenMao]

Yes, we use the default setting of the DRIT configuration.

[korawat-tanwisuth]

Thanks!