search

google-research / episodic-curiosity

Tensorflow/Keras code and trained models for Episodic Curiosity Through Reachability
Apache License 2.0
195 stars 34 forks source link

TPU training support #6

Closed hsezhiyan closed 4 years ago

hsezhiyan commented 4 years ago

Does this codebase currently support training on a TPU? If so, how would I train on a TPU?

nsavinov commented 4 years ago

Hi @hsezhiyan,

TPU training is currently not supported. If you implement it, please send us a pull request.

RaphaelMarinier commented 4 years ago

Hi @hsezhiyan,

To close this bug:

(1) This project is much more likely to benefit from simpler changes than switching to TPU if you want to speed up training. For instance, running more actors on the same machine than we did, or running actors on many machines in a distributed fashion is likely to provide good training speedups.

(2) Fully harnessing Cloud TPUs for RL training is multiple levels more challenging than (1). See https://github.com/google-research/seed_rl for a distributed RL framework that will soon run on Cloud TPUs.