search

junyanz / iGAN

Interactive Image Generation via Generative Adversarial Networks
MIT License
3.97k stars 587 forks source link

Training DCGAN will result in gray images #17

Open mojganaz opened 6 years ago

mojganaz commented 6 years ago

I am using train_dcgan.py to train the network over outdoor_64 images. However, after about 10 epochs of training, the generated images by the network become gray. Here is an example output file: gen_00011

Any idea on what the problem could be?

boluoweifenda commented 6 years ago

@mojganaz I have the same question, I think it may be caused by some hyperparameters and generator "learns" to produce gray images to minimize it's loss function. I haven't fix the problem yet, please let me know if you have any idea, thanks!

junyanz commented 6 years ago

It happened sometimes for DCGAN training. I would recommend that you use the latest GAN models as backend such as progressive-gans.

AbhishekSen commented 5 years ago

@mojganaz Same problem here , did you find out what was the issue and how did you fix it ?

junyanz commented 5 years ago

DCGAN is pretty old. If you just need to train a GAN model, you can probably try more recent work such as GAN-stability or style-gan .