meeting_00017_20210105.md

Rust ML WG Meeting 00017

Meeting Info

Date: 20210112

Start time: 1600 ET (-5 UTC)

Zoom: https://1password.zoom.us/j/96729203181?pwd=eER6dkxoNnJxdVJoUTgvNHlrZnUyZz09

Agenda

• Rust_RL does support Double-DQN now :)

• New cifar-ten version released

  • Download and extract is now automated

Participants

• chrism
• ricky
• manuel
• John Irungu

Minutes

Rust_RL Supports Double-DQN

• https://github.com/ZuseZ4/Rust_RL/tree/ddqn
• https://medium.com/@parsa_h_m/deep-reinforcement-learning-dqn-double-dqn-dueling-dqn-noisy-dqn-and-dqn-with-prioritized-551f621a9823
• use two neural networks to predict a state

rust_rl issue

• first open issue
• ZuseZ4/Rust_RL#6

from the rust_rl issue

• do we have that rurel tracked on arewelearningyet.com?
• https://github.com/milanboers/rurel/
  • no, but last activity was August 2019

New cifar-ten version released

• https://crates.io/crates/cifar-ten
• added an option to download the compress data (from online mirror)
  • downloads and decompresses automatically

tsuga

• chrism working on hardcoding a convolutional layer with backpropigation

GPU

• any updates on OpenCL or webGPU bindings
  • https://github.com/JonathanWoollett-Light/gpu_blas
  • extreme WIP
• https://jonathanwoollett-light.github.io/blas/

Automatic differenation of GPU Kernels

• https://github.com/wsmoses/Enzyme
• some successes!
  • differenate matrix multiplies and some financial code and got the right anwsers
• you can malloc in a GPU kernel though it is very undocumented
  • will give thread-level memory
  • amount of memory per thread isn't equal to total GPU memory/number of threads
• Rust GPU code
  • lookingo for good examples :P

MPI handling

• in enzyme
• (MPI == message passing interface) for distributed computing

Actions

• Get a gpu_blas update for next meeting