search

google-deepmind / meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Apache License 2.0
624 stars 122 forks source link

Did you use grayscale images as inputs? #40

Closed GoingMyWay closed 2 years ago

GoingMyWay commented 2 years ago

Dear authors,

Melting Pot provides RGB obs as inputs. In your experiments, did you use grayscale obs as inputs?

duenez commented 2 years ago

No, we use RGB as provided by the substrate directly into a CNN on the agent side.

GoingMyWay commented 2 years ago

No, we use RGB as provided by the substrate directly into a CNN on the agent side.

I see. I will take much computation compared with Atari games as there are many agents.

GoingMyWay commented 2 years ago

Hi, @duenez, do grayscale images affect the performance? RGB images take much memory and computation cost for researchers who have a limited computation budget. I am considering using grayscale images and scaling RGB images for MARL training. I appreciate it if you can provide some best practices for melting pot training. Thank you!

jzleibo commented 2 years ago

I think grayscale should work just fine for the majority of the substrates. It might be an issue for allelopathic harvest. But I think all the others should be fine.

I agree we need to do more to guide people toward best practices. I think that's what we are working out how to do together right now actually ;) -- thank you! Without your help it would be hard for us to know what the best practices are for using the open source RL libraries since they differ so much from what we use internally.

Try grayscale and let us know what you find!

We can run an experiment internally on this sometime also perhaps.

On Sun, Jun 19, 2022, 4:50 AM GoingMyWay @.***> wrote:

Hi, @duenez https://github.com/duenez, do grayscale images affect the performance? RGB images take much memory and computation cost for researchers who have a limited computation budget. I am considering using grayscale images and scaling RGB images for MARL training. I appreciate it if you can provide some best practices for melting pot training. Thank you!

— Reply to this email directly, view it on GitHub https://github.com/deepmind/meltingpot/issues/40#issuecomment-1159610708, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAHFJWQ3GN7U4HNG2FNBDLLVP2KJ7ANCNFSM5YP3H2SQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

GoingMyWay commented 2 years ago

I think grayscale should work just fine for the majority of the substrates. It might be an issue for allelopathic harvest. But I think all the others should be fine. I agree we need to do more to guide people toward best practices. I think that's what we are working out how to do together right now actually ;) -- thank you! Without your help it would be hard for us to know what the best practices are for using the open source RL libraries since they differ so much from what we use internally. Try grayscale and let us know what you find! We can run an experiment internally on this sometime also perhaps. … On Sun, Jun 19, 2022, 4:50 AM GoingMyWay @.> wrote: Hi, @duenez https://github.com/duenez, do grayscale images affect the performance? RGB images take much memory and computation cost for researchers who have a limited computation budget. I am considering using grayscale images and scaling RGB images for MARL training. I appreciate it if you can provide some best practices for melting pot training. Thank you! — Reply to this email directly, view it on GitHub <#40 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAHFJWQ3GN7U4HNG2FNBDLLVP2KJ7ANCNFSM5YP3H2SQ . You are receiving this because you are subscribed to this thread.Message ID: @.>

Hey, dear Dr Leibo @jzleibo, thanks for the quick, kind and helpful reply. Melting Pot provides powerful and impactful evaluation environments. I will let you know once I get the results. Currently, it will take me 1.5 days to train Melting Pot for 200M steps.