search

esteveste / dreamerV2-pytorch

Pytorch implementation of DreamerV2: Mastering Atari with Discrete World Models, based on the original implementation
17 stars 4 forks source link

Simple question: How long did it take to train it on Atari? #5

Closed jdubkim closed 1 year ago

jdubkim commented 1 year ago

Hi thank you so much for an amazing work. I think this is the best pytorch implementation of dreamer out there. Very clean and concise and reproduced similar results to the original implementation. I was simply wondering how long does it take to train this on atari environment. Thank you so much :)

esteveste commented 1 year ago

Around 2 days to do 20M steps (5M steps with frameskip) on a RTX 3060 Ti. Pretty slow :)

jdubkim commented 1 year ago

@esteveste Thanks a lot, it's quite slow, but at the same time, quite fast compared to Director paper, which is the work based on Dreamer v2 by the same author. Do you remember if the official code was faster than your pytorch implementation or not by any chance? Thank you so much :)

esteveste commented 1 year ago

Yes, I tried my best also with fp16 and cura benchmark, but I think I still was like 10/20 fps off the original implementation. I could get something like 150fps while the original TF code got 170fps. But maybe now with Pytorch 2.0 and the compile functionality maybe it could get better.

On Fri, May 5, 2023, 09:41 Jeewoo Kim @.***> wrote:

@esteveste https://github.com/esteveste Thanks a lot, it's quite slow, but at the same time, quite fast compared to Director paper, which is the work based on Dreamer v2 by the same author. Do you remember if the official code was faster than your pytorch implementation or not by any chance? Thank you so much :)

— Reply to this email directly, view it on GitHub https://github.com/esteveste/dreamerV2-pytorch/issues/5#issuecomment-1535925671, or unsubscribe https://github.com/notifications/unsubscribe-auth/AF36D65K7OHYCTE2L2RI5J3XES4LHANCNFSM6AAAAAAWWM62ZA . You are receiving this because you were mentioned.Message ID: @.***>

jdubkim commented 1 year ago

@esteveste oh that's still amazing, I was running with tf but somehow it was very slow (3~5 fps for some reason), so I thought your torch implementation could be much faster. I guess something's wrong with my tensorflow and cuda setup. Thank you so much for help! and the repo is amazing :)

esteveste commented 1 year ago

Thanks! :)

On Fri, May 5, 2023, 10:04 Jeewoo Kim @.***> wrote:

@esteveste https://github.com/esteveste oh that's still amazing, I was running with tf but somehow it was very slow (3~5 fps for some reason), so I thought your torch implementation could be much faster. I guess something's wrong with my tensorflow and cuda setup. Thank you so much for help! and the repo is amazing :)

— Reply to this email directly, view it on GitHub https://github.com/esteveste/dreamerV2-pytorch/issues/5#issuecomment-1535954234, or unsubscribe https://github.com/notifications/unsubscribe-auth/AF36D63LPNPMNI6SCKG7DZDXES7A3ANCNFSM6AAAAAAWWM62ZA . You are receiving this because you were mentioned.Message ID: @.***>