search

ikostrikov / pytorch-a3c

PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from "Asynchronous Methods for Deep Reinforcement Learning".
MIT License
1.23k stars 279 forks source link

Running with pytorch 0.2.0 #40

Closed ShaniGam closed 7 years ago

ShaniGam commented 7 years ago

Have you tried running it with pytorch 0.2.0? seem like the multiprocessing doesn't work.

ikostrikov commented 7 years ago

Hi. Yes, I works for me for pytorch 0.2.0. What do you mean by mp doesn't work?

ShaniGam commented 7 years ago

all the training processes exit after: (policy_loss + args.value_loss_coef * value_loss).backward() in some cases I get a core file that says: Program terminated with signal 11, Segmentation fault. but it doesn't say where it happens exactly I did manage to run other python files using pytorch 0.2

ikostrikov commented 7 years ago

Interesting. How many RAM do you have?

ShaniGam commented 7 years ago

252 GB it worked perfectly with pytorch 1.2 (but I have to upgrade)

ikostrikov commented 7 years ago

Does it happen immediately or after several hours of training?

What arguments do you use?

ShaniGam commented 7 years ago

Immediately, default arguments.

ikostrikov commented 7 years ago

Do you use python2 or python3?

At the moment, I cannot reproduce the issue. I need more information.

ShaniGam commented 7 years ago

python 2.7 I deleted all the torch files and installing it from scratch now, It might be a conflict between the versions (even though it does work on other files)

ikostrikov commented 7 years ago

This code works only for python 3. I'm not planning to add support for python 2 any time soon (it's already scheduled to be deprecated anyway).

ShaniGam commented 7 years ago

Oh, so that's why! btw, thank you for the code, it's been very useful in my research :)