With the LunarLander and HalfCheetah envs requiring single iteration batch sizes of > 10k for convergence, these runs take quite a long time. Gym allows multiprocessing for multiple envs at the same time in the form of the VectorEnv base class. Here's my proposal for an implementation of a vectorized environment to speed up batch sizes.
I would make a PR but the code in this repo contains many unfilled blanks (TODOs), some of which would need to be implemented in order to properly create a PR. Thus, I'm making an issue with the specified commits here.
An example of a test you can do is here. We run on the cartpole environment for 10 iterations with a batch size of 10000. We compare the single env and the vectorized env cases where we use 10 envs in the vectorized case.
Obviously the staff implementation and my implementation are different, but here are my outputs for the above test
# Without vectorization
real 1m17.553s
user 1m16.253s
sys 0m1.270s
# With vectorization, # of envs = 10
real 0m13.258s
user 0m13.899s
sys 0m2.867s
We achieve a speedup of ~6x. Let's say my LunarLander experiments take 2 hours to run. This would in theory reduce the time it takes to run the same experiment to 20 min, much more tractable, especially for students who typically only have access to CPU or colab GPUs.
bri25yu
commented
2 years ago
With the LunarLander and HalfCheetah envs requiring single iteration batch sizes of > 10k for convergence, these runs take quite a long time. Gym allows multiprocessing for multiple envs at the same time in the form of the VectorEnv base class. Here's my proposal for an implementation of a vectorized environment to speed up batch sizes.
I would make a PR but the code in this repo contains many unfilled blanks (TODOs), some of which would need to be implemented in order to properly create a PR. Thus, I'm making an issue with the specified commits here.
An example of a test you can do is here. We run on the cartpole environment for 10 iterations with a batch size of 10000. We compare the single env and the vectorized env cases where we use 10 envs in the vectorized case.
Obviously the staff implementation and my implementation are different, but here are my outputs for the above test
We achieve a speedup of ~6x. Let's say my LunarLander experiments take 2 hours to run. This would in theory reduce the time it takes to run the same experiment to 20 min, much more tractable, especially for students who typically only have access to CPU or colab GPUs.