-
Notifications
You must be signed in to change notification settings - Fork 72
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
How much time is required to train the model? #27
Comments
Hello, thank you for taking an interest! Note that we trained with early stopping on the training FID metric, so the times indicated below have been computed using the elapsed time per epoch multiplied by the number of training epochs until the early stopping was triggered: IC-GAN (BigGAN backbone) on ImageNet 64x64: ~ 14days Class-conditional IC-GAN (BigGAN backbone) on ImageNet Note that the experiments at 64x64 present a mostly flat FID curve for most of the time, but due to the early stopping it takes some time to finally get them to a halt. In practice, training for 25-30% of the specified time above for the 64x64 resolution (3-4 days) would result in a very close FID metric to the ones reported in the paper. IC-GAN (StyleGAN backbone) on COCO-Stuff Same as before, training for ~50% of the time indicated above should result in a very similar FID metric. |
Thank you for your detailed response!! But, @ArantxaCasanova, when I had run the code for ICGAN on ImageNet using BigGAN for three days using TITANRTX, which shows similar performance to V100, I obtained FID near 14~, which is expected to be around 9.2 according to the paper and your response. Can you check my log below to see whether it is correct? I did not change the code and run the default setup of icgan_res64.json.
|
I did check your log, and here is a plot to show you how the curve looks like with respect to one of my runs As an additional note, remember that the numbers reported in the paper are obtained with the tensorflow FID code and conditioning on only 1000 feature vectors selected with K-means. The FIDs logged during training are only used to monitor training and meant to compare experiments with each other, not to report the final results. For more details, refer to: https://github.com/facebookresearch/ic_gan#how-to-test-the-models. |
In supplementary materials, I found that what types of GPUs are used to obtain the results, but any information about training times could not be found.
Can you provide approximate times (or days) to train the model for ImageNet 64/128/256 for both unconditional and conditional BigGANs and COCO-stuff 128, 256 for StyleGAN2?
Thank you for sharing the code of this great work!
The text was updated successfully, but these errors were encountered: