search

LEGO999 / BigBiGAN-TensorFlow2.0

TensorFlow 2.0 implementation of Large Scale Adversarial Representation Learning(BigBiGAN)
45 stars 12 forks source link

Encoder collapse (maps all images to the same vector) #1

Open oriki77 opened 4 years ago

oriki77 commented 4 years ago

Hi,

First of all amazing work!, the generated images have great quality. I tried running it on my computer and it seems like the encoder doesn't learn properly. I tried a few things but couldn't find the bug myself.

LEGO999 commented 4 years ago

Hi,

Thanks for your issue report, could you describe the detailed behavior of the bug(Encoder collapse (maps all images to the same vector))?

oriki77 commented 4 years ago

Hi, Thanks for your response. I am training the model on CIFAR10 (unconditional), with your default parameters. It seems like the encoder is experiencing a mode collapse, mapping all input images to the same latent vector (or almost the same latent vector). During training at the end of every epoch I generate the images using "fixed_z" like in the code. than, I encode the generated images in order to get their latent representation, and reuse the generator to produce the reconstructions. It seems like the encoder doesn't really learn useful latent representations (and doesn't improve during training).

Thanks again for sharing this amazing project!

Illustration : generated images: image

reconstructions: image

oriki77 commented 4 years ago

Hi, did you see my last message by any chance? Thanks!

shreejalt commented 3 years ago

@oriki77 I also faced the same issue. Were you able to get the bug?