ikostrikov/pytorch-a2c-ppo-acktr-gail

get different results when I set the same seed

Closed this issue · 3 comments

Dear Author,
When I run PPO, I notice that sometimes I would get different results when I set the same random seed. It happens when I use multiple workers in Atari, however, the result is always fixed when I use a single worker in Mujoco. So I guess multiprocess might be the problem? I wonder how could I fix the final result with multiple workers in Atari, thank you!

I tried to use a single worker in Atari just now and it also shows different results...

Hi @naivety77 !

Did you set the flag cuda_deterministic to True?

Asynchronous execution on GPU is usually the main reason under getting different results for different runs.

Hi @naivety77 !

Did you set the flag cuda_deterministic to True?

Asynchronous execution on GPU is usually the main reason under getting different results for different runs.

It works, thanks for your reply! I forgot to change the default of cuda_deterministic in arguments before...