search

ariel415el / SimplePytorch-ALAE

Implementation of Adverserial autoencoders
11 stars 1 forks source link

GAN mode collapse issue/question #3

Open derrxb opened 3 years ago

derrxb commented 3 years ago

Hi,

Thanks for this awesome implementation. I've been attempting to adopt your implementation of ALAE to work with sequential data. In my refactor, I replace the MLP layers with LSTMs since I'm working with a time dependent data (specifically music generation).

However, I've been having some issues where after training for a few minutes, the generator dies and outputs the same things over and over again. I figured this is a GAN mode collapse state.

I'm writing to ask if this is something you also experienced while working on your implementation? Or perhaps working with time dependent data just wouldn't work with this implementation? If this isn't the best place to ask this, I'll be happy to close this issue.

Thanks :)

ariel415el commented 3 years ago

Seems like the right place to ask but I haven't encountered such phenomenon In my training. I only trained the StyleALE architecture on FFHQ and MLP on MNIST trying to "reproduce" the paper results (I only have one gpu so i couldn't really reproduce.)

Mode collapse after few minutes seems odd, maybe try the GAN training (train_StyleGan.py vs train_styleALAE.py) with your LSTMS and see if you still have the problem so fast.

Please tell me if you solved this. And I'll also be glad to hear your generated audio when you get it.

Good Luck !