search

timsainb / Tensorflow-MultiGPU-VAE-GAN

A single jupyter notebook multi gpu VAE-GAN example with latent space algebra and receptive field visualizations.
MIT License
440 stars 96 forks source link

The results are not good #7

Open junsukchoe opened 7 years ago

junsukchoe commented 7 years ago

Hi,

I download your network and train it but the results are not good. My results are so blurry and not human-like.

How many epochs are needed to converge? I observed that the results collapse after 70~80 epochs. All hyper-parameters are same as your version.

Is there any optimization tip? Any advice would be very helpful.

Thank you.

timsainb commented 7 years ago

Hello gildorgood, When I ran it, it would take the night to converge to something good on a titanx, not sure how many epochs. You may need to play around with hyperparameters a bit? To optimize it, I was keeping track of how the generator and discriminator loss were competing. From what I've heard from others, things have changed a bit for new Tensorflow versions. Currently with networks that I'm running, I do not use prettytensor at all and instead use Tensorflow built in functions from TF 1.0+: https://github.com/timsainb/Tensorflow-MultiGPU-VAE-GAN/issues/4#issuecomment-302512016 The same goes for optimization.

junsukchoe commented 7 years ago

I appreciate your reply. I still have some issues that I cannot resolve. If it is not offending, can you please share your pre-trained network?