search

worldmodels / worldmodels.github.io

World Models
Creative Commons Attribution 4.0 International
430 stars 55 forks source link

Training time and procedure #3

Open hundredblocks opened 6 years ago

hundredblocks commented 6 years ago

Amazing work! I had a practical question about the time it took to train these models on the setup you described in the article. Would you be able to share more? In addition, would this repository be the best way to be notified of when the code would be available?

hardmaru commented 6 years ago

Hi!

I'm in the process of cleaning up the TensorFlow implementations of the experiments and will be putting them out on GitHub soon. For my implementation, VAE and MDN-RNN trains in ~ 1 hour each (on a single P100 GPU). The Doom controller takes a few hours on a 64-thread evolution strategy process (based off my previous ESTool project), inside a virtual latent-environment as discussed in the article. The car racing controller takes a bit longer, since it trains on the actual environment and that is the main bottleneck.

Once the code is released I'll update you.

hundredblocks commented 6 years ago

Thank you for the update, that is much quicker than I expected in terms of training time, very exciting.