Memory leak with different libraries (numpy, pyTorch, OpenAI Gym)

[Paethon]

I am currently working with OpenAI Gym using PyCall and the following code (randomly plays games of Atari 2600 Pong) uses up memory it never frees (on the order of 100MiB a second).

using PyCall

@pyimport gym
env = gym.make("Pong-v0")
env[:reset]()

for i in 1:100000
    obs, reward, done, info = env[:step](rand([3,4]))
    env[:render]()
    done && env[:reset]()
end

env[:close]()

To try you should pip install gym[all]

The same example in Python does not have this problem. Any idea how to fix this? Or how do I find out what exactly is happening here?

This seems to occur with other libraries (numpy, PyTorch) as well. See later comments on this issue.

[stevengj]

Would be good to apply a Python tool to track which object(s) are not being garbage-collected: https://stackoverflow.com/questions/1435415/python-memory-leaks

[Paethon]

Thank's for the starting point. I will have a look and report back.

[tomchor]

I'm experiencing what I believe to be the same issue.

[c42f]

To give a reduced test case, inspired by @tomchor's problem:

using PyCall
@pyimport numpy as np
z = zeros(10000)
while true
    np.copy(z)
end

This seems ok on 0.6, but quickly exhausts the memory on 1.0

[boathit]

I could also provide a minimal code for the test

using PyCall
@pyimport torch

for i = 1:10_000
    X = torch.rand(500, 500)
    i % 1000 == 0 && (println("i: $i"); sleep(2))
end

It seems the memory held by X will never be released.

Any chance to fix this?

[Paethon]

Thank you! #558 has fixed the issue for me!

[JobJob]

Aside: @Paethon you might be interested in OpenAIGym.jl and this PR if you want performant access to the Gym from Julia.

[Paethon]

@JobJob Thanks, will have a look!