这是indexloc提供的服务,不要输入任何密码
Skip to content

atari_dqn.py Training Speed Drops Significantly Over Time #1265

@byw0919

Description

@byw0919

Hi team,

I've encountered a performance issue while running the example/atari/atari_dqn.py script. The training speed is slow, and slows down drastically as it progresses.

Environment:
GPU: NVIDIA A6000
PyTorch Version: 2.1.1+cu113, 2.6.0+cu126 (tried both)
Script: example/atari/atari_dqn.py
Env: Alien & Pong
Seed: 1

Steps to Reproduce:
Run the example/atari/atari_dqn.py script.
Monitor the training speed (iterations per second, it/s).

Observed Behavior:
Initially, the training speed is around 40-50 it/s.
The it/s continuously decreases. Around 2,400 training steps, the speed drops to below 10 it/s, and later even below 7 it/s.
I attempted to mitigate this by setting training_num=100.
While increasing training_num might slightly delay the severe drop, the it/s still falls below 10 around 20,000 training steps.
At this reduced speed, a single epoch is estimated to take over 2 hours to complete.

Could you please look into this? Let me know if you need any more information from my end.

Thanks!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions