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
528 stars 74 forks source link

No module named 'dopamine.google' #25

Closed jzitovsky closed 3 years ago

jzitovsky commented 3 years ago

The train.py file located in batch_rl/fixed_replay has on line 41 from dopamine.google import xm_utils. When I run the train.py file (e.g. by running the fixed_replay_runner_test test as shown on the README), I get the following error: ModuleNotFoundError: No module named 'dopamine.google'. I downloaded the dopamine library as instructed in the README file. I found that replacing this line with what was there from a previous version, from dopamine.discrete_domains import train as base_train allowed the fixed_replay_runner_test test to run successfully. Training offline agents also appears to work correctly, in that no errors are thrown and performance appears to improve over iterations according to the saved logs.

Is the dopamine.google module from a previous or experimental version of dopamine-rl, or have I done something incorrectly? Would changing line 41 to what was there before introduce any issues, such as computational slowdowns or inaccuracies?

agarwl commented 3 years ago

This is an internal import and should not be externally imported. Any line xm_utils or dopamine.google should be removed.

agarwl commented 3 years ago

https://github.com/google-research/batch_rl/commit/80bbc006719faca3b002568d37d759ae9f339cd5 resolves this issue. Thanks for bring it to my attention.