OpenAI baselines in pytorch

[deontologician]

There are several nice implementations of RL algorithms in

https://github.com/openai/baselines

But they're all implemented in Tensorflow, and in addition have lots of onerous dependencies like MPI. It would be nice to have these implementations in pytorch, and using something other than MPI for communication (maybe [gRPC][(https://github.com/grpc/grpc) ? )

Someone has started a reimplementation of the baselines in pytorch here:

https://github.com/nadavbh12/baselines-pytorch

but they've only implemented A2C so far. It would be nice if we also had an implementation of ACKTR and PPO2

[shicode]

I can take a shot at this! Github username: shicode

[deontologician]

Cool, I added you to the project. What I'd suggest is cloning the repo, creating a new branch, and opening a pull request with what you're working on. We can talk in gitter if you want to discuss the process in more detail

[deontologician]

I'm thinking maybe attempt ACKTR first. The original paper is here:

https://arxiv.org/abs/1708.05144

And the OpenAI code for it is here:

https://github.com/openai/baselines/tree/master/baselines/acktr

[shicode]

Thanks! I’ll get on it.

…

On 19-Dec-2017, at 9:13 AM, Josh Kuhn ***@***.***> wrote: I'm thinking maybe attempt ACKTR first. The original paper is here: https://arxiv.org/abs/1708.05144 <https://arxiv.org/abs/1708.05144> And the OpenAI code for it is here: https://github.com/openai/baselines/tree/master/baselines/acktr <https://github.com/openai/baselines/tree/master/baselines/acktr> — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#12 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AS5i1wACdrYOEH86P4UaV2Tspl3MeVWtks5tBzD7gaJpZM4REyRD>.

[deontologician]

Let me know if you need anything. This is a kind of a large project so it could be intimidating to find where to start.

…

On Mon, Dec 18, 2017, 20:04 Shilpa Suresh ***@***.***> wrote: Hi Josh, Thanks! I’ll get on it. Sincerely, Shilpa > On 19-Dec-2017, at 9:13 AM, Josh Kuhn ***@***.***> wrote: > > I'm thinking maybe attempt ACKTR first. The original paper is here: > > https://arxiv.org/abs/1708.05144 <https://arxiv.org/abs/1708.05144> > And the OpenAI code for it is here: > > https://github.com/openai/baselines/tree/master/baselines/acktr < https://github.com/openai/baselines/tree/master/baselines/acktr> > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub < #12 (comment)>, or mute the thread < https://github.com/notifications/unsubscribe-auth/AS5i1wACdrYOEH86P4UaV2Tspl3MeVWtks5tBzD7gaJpZM4REyRD >. > — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#12 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAAFVjqU-hCJ7t5VxIP0H26oXrsld7S8ks5tBzW0gaJpZM4REyRD> .

[JACKHAHA363]

Is the project going on?

[deontologician]

No, not really