Fastai reinforcement learning

Deep Learning
22

Please submit your list - I’m kinda deep in reviewing the existing code out there and wtf, I need more pressure! LOL.

23

Okay sure! Will take a while to make the wiki post, it’s a google folder right now of papers (which I think I can’t share, not really sure what the proper way is).

24

I have personally used stable-baselines and it worked fairly well. I haven’t tried any of the other libraries, but I had no idea there were so many. Next time I’m working in RL again, I’ll give catalyst a try!

25

can i ask what is the current status ?

26

I’ve merged my efforts with https://github.com/josiahls/fast-reinforcement-learning

2 Likes
27

I agree with your viewpoint. Here is a project I did recently where I applied Reinforcement Learning to an optimal investment problem:

Managing Investor Decisions with Reinforcement Learning

28

Hi Everyone,

I have renamed the project to https://github.com/josiahls/fastrl from @Jumonji 's link. Please reference https://github.com/josiahls/fastrl/tree/pre_refactor as this branch is prior to a giant refactor I am doing to the package. I do not have a pypi package out yet, however I plan to have automatic nightly builds inplace for pypi and conda.

I will also be building docker images for dev prod by copying fastai’s devops methods. If you want, we can start a project in github with notes on it depending on interest.

1 Like
29

Anyone else interested. I recently finished a rough draft of the refactored RL backend using pytorch and fastai API. It is limited but tbh this is the scary part and its showing promise. You can find it here: https://josiahls.github.io/fastrl/data.block.html . this should also work in in colab. I modified the notebook headings to install the right stuff to run renderable envs

1 Like