Closed eitanrich closed 6 years ago
Which architecture?
From: eitanrich notifications@github.com Sent: Sunday, January 28, 2018 12:51:47 PM To: igul222/improved_wgan_training Cc: Subscribed Subject: [igul222/improved_wgan_training] Poor results in WGAN mode on CelebA (#59)
Hi,
I'm trying to train on CelebA (cropped and resized to 64x64). The results in WGAN-GP mode look great, both in quality and diversity, however, when I set the mode to 'wgan', I get very distorted faces even after 200K iterations (see attached). Any ideas?
Thanks, Eitan [celeba_samples_10199]https://user-images.githubusercontent.com/29534333/35486911-b7bbc3e4-047d-11e8-841a-870699de3fb3.png
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/igul222/improved_wgan_training/issues/59, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ABBP7qzsTMK_9vLin8IGrYDKXlHEVEEfks5tPN3jgaJpZM4Rv2We.
I'm not sure it answers your question, but I used the default GoodGenerator, GoodDiscriminator, with the following parameters:
MODE = 'wgan' # dcgan, wgan, wgan-gp, lsgan DIM = 64 # Model dimensionality CRITIC_ITERS = 5 # How many iterations to train the critic for N_GPUS = 2 # Number of GPUs BATCH_SIZE = 64 # Batch size. Must be a multiple of N_GPUS ITERS = 200000 # How many iterations to train for LAMBDA = 10 # Gradient penalty lambda hyperparameter OUTPUT_DIM = 64643 # Number of pixels in each iamge
The same configuration with MODE = 'wgan-gp' produces nice results, without the distortions. Eitan
GoodGenerator is only tested with WGAN-GP. If you want to use other gans it's recommended to use the DcganGenerator / DcganDiscriminator.
From: eitanrich notifications@github.com Sent: Sunday, January 28, 2018 2:47:40 PM To: igul222/improved_wgan_training Cc: Ishaan Gulrajani; Comment Subject: Re: [igul222/improved_wgan_training] Poor results in WGAN mode on CelebA (#59)
The same configuration with MODE = 'wgan-gp' produces nice results, without the distortions. Eitan
— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/igul222/improved_wgan_training/issues/59#issuecomment-361103532, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ABBP7srdd32pH4h93S3wdLYU0YLHOR1bks5tPPkLgaJpZM4Rv2We.
Thanks. This seems to be working.
hi,now i am also trying to train on CelebA (cropped and resized to 64x64) in WGAN-GP mode . I just modify the DATA_DIR in gan_64x64.py. But there was a mistake like this: IOError: [Errno 2] No such file or directory: '/data-4T-B/yelu/data/dcgan-completion.tensorflow/aligned/img_align_celeba_png/train_64x64/train_64x64/0927649.png' Could you show me your code? thanks so much~~~
Hi,
I'm trying to train on CelebA (cropped and resized to 64x64). The results in WGAN-GP mode look great, both in quality and diversity, however, when I set the mode to 'wgan', I get very distorted faces even after 200K iterations. Any ideas?
Thanks, Eitan