search

google-research / batch_rl

Offline Reinforcement Learning (aka Batch Reinforcement Learning) on Atari 2600 games
https://offline-rl.github.io/
Apache License 2.0
520 stars 73 forks source link

TF Version must < 2.0? #5

Open weihongwei0586 opened 4 years ago

weihongwei0586 commented 4 years ago

The code cant work in my environment where TF version is the newest(2.2.0). because the tf.contrib moudle has been removed?

agarwl commented 4 years ago

Yeah, I think Dopamine library still uses tensorflow 1.x and the code is only compatible with that for now :/ . @psc-g for visibility about dopamine.

psc-g commented 4 years ago

we're working on both a jax version of dopamine and a TF2 compliant version, stay tuned!

weihongwei0586 commented 4 years ago

we're working on both a jax version of dopamine and a TF2 compliant version, stay tuned!

Thank you very much!

agarwl commented 3 years ago

Dopamine now supports JAX agents! The main difference in offline agents is to simply use the FixedReplayBuffer (which directly reads data from numpy) instead of the tensorflow buffer. For an example, see the implementation of QR-DQN in JAX.