search

krisrs1128 / clouds_dist

Simulation of low-clouds, from weather measures.
4 stars 1 forks source link

Discriminator not learning #26

Closed vict0rsch closed 5 years ago

vict0rsch commented 5 years ago

I ran an experiment trying to overfit 100 samples, over 100 epochs.

Here's the link to the comet exp.

Pb: discriminator loss is constantly 0.5.

There must be a bug in the code, something's not right. Trying to investigate issues that may be related to backward or detach

krisrs1128 commented 5 years ago

I guess this is what you're referring to, but just in case, wanted to point out that the generated image is not detached when calling the discriminator.

vict0rsch commented 5 years ago

@mustafaghali @krisrs1128 I've worked on a fix; check out the branch debug-discriminator basically I replaced Tianle's discriminator by a more trusted one: spectral-normalized resnet from https://github.com/christiancosgrove/pytorch-spectral-normalization-gan/blob/master/model_resnet.py (see file res_discriminator.py)

I ran an "overfitting" experiment with 500 datapoints: generator converges to 0 but at least loss is not stuck at 0: https://www.comet.ml/vict0rsch/clouds/fe23187c420c47819677ba610c31e4fd?experiment-tab=chart

vict0rsch commented 5 years ago

Thoughts on the fix? Merge to master?