Great talk! It'd be nice to stress out the pros/cons between Flax and Haiku since both were built with the same goal in mind (to the best of my knowledge).
great talk. I enjoyed the Mark's explanation of the flax philosophy as I am a flax user and I know understand the intuition better. However, I guess for anyone who is trying to understand flax, it is very good to also check the slides that are made available somewhere in the website. I would prefer if there was a comprehensible intro to flax somewhere though. It feels I needed to gather bits of information from multiple sources to gain an understanding Also, the dopamine library seems pretty cool. I am trying to code up some RL agents and I am wondering where should I start. I guess that may help me there. However, I am experimenting with easy games. I do not know if I need the advanced tricks to be able to handle them or not.
Interested to hear how it's been, if you've been using it at work or for research projects. How easy is it compared to Pytorch, I'm currently having a very hard time in parallelizing huggingface model on Pytorch
@@geekye I'll be dead honest I eventually gave up and just went back to pytorch LOL. We're using deepspeed and training on a gpu cluster now. JAX is only worth it if you can get a TRC trial and train on the v3-8s, or just generally speaking if you have access to TPUs. fwiw if you email them and say you want to use TRC to do something with Jax they will be VERY lenient with giving you extra days and things like that. Also what kind of parallelization are you trying to do? Deepspeed takes care of multigpu for most things.
rlax is a collection of building blcoks from which you can use to construct agents, while dopamine provides a monolithic landscape with tunable parameters that can be able to modify their behaviours.
Great talk! It'd be nice to stress out the pros/cons between Flax and Haiku since both were built with the same goal in mind (to the best of my knowledge).
great talk. I enjoyed the Mark's explanation of the flax philosophy as I am a flax user and I know understand the intuition better. However, I guess for anyone who is trying to understand flax, it is very good to also check the slides that are made available somewhere in the website. I would prefer if there was a comprehensible intro to flax somewhere though. It feels I needed to gather bits of information from multiple sources to gain an understanding
Also, the dopamine library seems pretty cool. I am trying to code up some RL agents and I am wondering where should I start. I guess that may help me there. However, I am experimenting with easy games. I do not know if I need the advanced tricks to be able to handle them or not.
When I started with JAX coming from pytorch, it just seemed really confusing but the more I learn about it, the more interesting and clean it seems.
any tutorials recommended?
@@0730pleomax I've been going off docs, this series and this linen examples the second speaker mentions.
Interested to hear how it's been, if you've been using it at work or for research projects. How easy is it compared to Pytorch, I'm currently having a very hard time in parallelizing huggingface model on Pytorch
@@geekye I'll be dead honest I eventually gave up and just went back to pytorch LOL. We're using deepspeed and training on a gpu cluster now. JAX is only worth it if you can get a TRC trial and train on the v3-8s, or just generally speaking if you have access to TPUs. fwiw if you email them and say you want to use TRC to do something with Jax they will be VERY lenient with giving you extra days and things like that. Also what kind of parallelization are you trying to do? Deepspeed takes care of multigpu for most things.
The Functional and Composition API is what really pulls towards Jax / Flax ecosystem
Hi, there are currently two lib for RL using Jax, Dopamine and Rlax. Might I ask what is the difference? Thanks!
rlax is a collection of building blcoks from which you can use to construct agents, while dopamine provides a monolithic landscape with tunable parameters that can be able to modify their behaviours.