search

google-deepmind / dqn_zoo

DQN Zoo is a collection of reference implementations of reinforcement learning agents developed at DeepMind based on the Deep Q-Network (DQN) agent.
Apache License 2.0
457 stars 78 forks source link

Eating too much host memory #15

Closed yueyang130 closed 2 years ago

yueyang130 commented 2 years ago

When I run the your code, specifically Rainbow Agent on Atari 100k, it consumes too much memory. One running need about 50GB memory per hour running. The below picture shows the memory overview when running two scripts.

Do you have any idea why it need such a big memory? Thanks for any advice.

image

GeorgOstrovski commented 2 years ago

Thank you for reporting this. We have not observed such excessive memory usage with this code in our setup and default configuration - usually memory consumption lies well below 20GiB.

Can I check what modifications you have applied to the code that you run? In particular, increasing the replay buffer size, or disabling image-downscaling/greyscaling in the preprocessing pipeline could easily lead to an increase in memory requirements.

GeorgOstrovski commented 2 years ago

Closing this for now - we could not reproduce the problem. If you are still experiencing the issue, please provide us with some more context when re-opening it, in particular let us know which local modifications were applied to the code.