search

maum-ai / faceshifter

Unofficial PyTorch Implementation for FaceShifter (https://arxiv.org/abs/1912.13457)
BSD 3-Clause "New" or "Revised" License
610 stars 114 forks source link

Suggestion of training epochs? #4

Closed zjxgithub closed 3 years ago

zjxgithub commented 3 years ago

Thanks for the implementation! Well done! I am about to use your code to train my own model. I am curious how long it took your model to be trained? How many epochs did you use?

Thanks!

usingcolor commented 3 years ago

Hi! According to the paper, they trained 500K steps for AEI-Net. In our case, the training loss seems to be little bit random and unstable, but in the long term the loss was going down. We trained for over 500K with 32 batch size. For conclusion, there are no absolute training epoch, you should observe the loss graph and validation results. Then stop when you satisfy. I know it sounds little irresponsible, but this is a behavior of GAN.

zjxgithub commented 3 years ago

Thanks for your response! That is helpful to know! May I ask how long it took to finish 500K steps on single GPU? Several days? Cuz I am trying to estimate when I can finish the training.

On Fri, Oct 30, 2020 at 3:20 AM Changho Choi 최창호 notifications@github.com wrote:

Hi! According to the paper, they trained 500K steps for AEI-Net. In our case, the training loss seems to be little bit random and unstable, but in the long term the loss was going down. We trained for over 500K with 32 batch size. For conclusion, there are no absolute training epoch, you should observe the loss graph and validation results. Then stop when you satisfy. I know it sounds little irresponsible, but this is a behavior of GAN.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mindslab-ai/faceshifter/issues/4#issuecomment-719402228, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFMTSLPVD4CD7P44ENFNH3DSNJZMNANCNFSM4TBN6P7Q .

usingcolor commented 3 years ago

Depends on your environment. Two V100 32G with 32 batch size took more than a week.

zjxgithub commented 3 years ago

I see... Thanks!

On Fri, Oct 30, 2020 at 3:31 AM Changho Choi 최창호 notifications@github.com wrote:

Depends on your environment. Two V100 32G with 32 batch size took more than a week.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mindslab-ai/faceshifter/issues/4#issuecomment-719411146, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFMTSLLZZAGR65DKBWPD5YTSNJ2VXANCNFSM4TBN6P7Q .

y-x-c commented 3 years ago

hi @usingcolor ! I wonder how many images were used in your training/validation? thanks!

usingcolor commented 3 years ago

hi @usingcolor ! I wonder how many images were used in your training/validation? thanks! please make a new issue next time.

For the Question, I used the dataset that I mentioned in README.md (i.e. FFHQ, CeleabA-HQ and VGG face) You can add the face images as you want.