Comparing Different models

[mohammadshahabuddin]

Hi,
Could you please tell me how you compared different models? Did you use the same learning rate, number of epochs, Number of decay epochs, image size, optimizer among all models? Also, did you collect test results using the final saved generator or did you use the best results testing all saved generators at different epochs?

[Skylion007]

I'll have to look through my notes later, but from what I remember:

• Image Size: Kept constant among all models if possible. Some like DiscoGAN required lower resolution input and so they were downsampled.
• Optimizer: Whatever optimizer was used for the model, same for LR (fidelling with those can make GANs mode collapse really easily)
• Number of epochs: until the epoch decay was exhausted for that model. (We did try adjusting num epochs / decay epochs on them, but they didn't improve quality.
• Number of decay epochs were kept constant among all ablations. For baselines, we used the recommended number of decay epochs from their respective papers.
• We used the final saved generator.

[mohammadshahabuddin]

Thank you for the information! These will be very helpful. Thanks, Shahab

…

On Fri, Jan 29, 2021, 2:59 PM Aaron Gokaslan ***@***.***> wrote: I'll have to look through my notes later, but from what I remember: - Image Size: Kept constant among all models if possible. Some like DiscoGAN required lower resolution input and so they were downsampled. - Optimizer: Whatever optimizer was used for the model, same for LR (fidelling with those can make GANs mode collapse really easily) - Number of epochs: until the epoch decay was exhausted for that model. (We did try adjusting num epochs / decay epochs on them, but they didn't improve quality. - Number of decay epochs were kept constant among all ablations. For baselines, we used the recommended number of decay epochs from their respective papers. - We used the final saved generator. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANB7Q22W5P72WMBPJYIS6CDS4MHRRANCNFSM4WYHHQDA> .