search

hzwer / ICCV2019-LearningToPaint

ICCV2019 - Learning to Paint With Model-based Deep Reinforcement Learning
MIT License
2.25k stars 312 forks source link

spectral normalization GAN #18

Closed ThisIsIsaac closed 5 years ago

ThisIsIsaac commented 5 years ago

Have you tried spectral normalization GAN & adding L1 distance to WGAN loss? I wonder how these two changes could impact the performance:

1. Replacing WGAN-GP with spectral normalization

Spectral normalization has two main advantages:

  1. Slight performance improvement relative to WGAN-GP on ResNet. The inception score of spectral normalization had a slight upper hand — approximately 0.16 — with less deviation compared to WGAN-GP.

  2. Spectral normalization is ~30% more computationally efficient. Since both actors and critics use ResNet as the backbone, replacing WGAN-GP with spectral normalization can potentially yield meaningful results.

2. Combining WGAN-GP with spectral normalization

The authors of the spectral normalization paper suggest that combining WGAN-GP with spectral normalization can further improve the results compared to the baseline WGA-GP and spectral normalization GAN.

hzwer commented 5 years ago

I tried spectral normalization, but I couldn't make it work. Maybe I can try again this weekend.

hzwer commented 5 years ago

I use LearningToPaint frame to another topic, and WGAN-GP with spectral normalization can improve the stability of training. In my experiment, soft target and specnorm are not compatible in D training.

ThisIsIsaac commented 5 years ago

Thanks for sharing the insight :)