meeting_00009_20200909.md

Rust ML WG Meeting 00009

Meeting Info

Date: 20200909

Start time: 1600 ET

Zoom link: https://mozilla.zoom.us/j/99658259391

• Meeting ID: 996 5825 9391

Agenda

• Rust-RL: Short presentation

• Collect Information about active and advanced ML respositories not represented here on Zulip.

• Linfa

  • New tests for trees in the works
  • linfa-ensemble in the works (starting from random-forest)
  • Documentation is currently being updated

• There's currently a 'Book' repository, which is private to the Rust-ML group.

  • Documentation for linfa should be probably take priority, but it would be great if these two could influence each other
  • This seems like a good 'first issue' for people looking to get involved with the group or machine learning in general. Examples with real-life data sets being imported, applied, etc. could be used here
  • This is currently being done in mdbook format--considering the popularity of Jupyter in the machine learning domain and the evcxr kernel, should an alternate format be considered?

• The Rust London User Group ML/DS meetup

Participants

• Ricky (degausser)
• chrism (quietlychris)
• manuel (ZuseZ4)
• Daniel (xd009642)
• Lorenz (bytesnake)
• Samuel Lim

Minutes

Rust RL

• https://github.com/ZuseZ4/Rust_RL
• part of a university project
• half neural network, half reinforcement learning
• softmax
• using ndarray
• next steps, optimization adam
• MNIST 98.x% accuracy
• QLearning and DeepQ learning
• able to learn basic games
• would appreciate a review, but is still a heavy WIP
• Q: how much work would it be to work on Pac-Man?
  • depends if there's a working implimentation
  • his NN can take a 1D, 2D, or 3D
  • might be able to run it in a Android emulator and output images
  • would first have to implemtent GPU support to work
• Q: storing weights?
  • not yet
  • chrism has been looking into this as well
    • found since most things are dynamic serde doesn't like Strings
• Darknet
  • config file
    • information about the network and its parameters
  • weights file
• Onyx is a generalized weight file
  • https://github.com/onnx/onnx
  • https://onnx.ai/
  • https://crates.io/crates/tract-onnx

    • https://www.khronos.org/blog/nnef-and-onnx-similarities-and-differences#:~:text=On%20the%20technical%20side%2C%20NNEF,graph%20as%20a%20data%20structure
  • there's also Pickel and H5
    • plenty of generalized things to look into and support

Repos not Represented

• Reinforcement Learning Crates
• some kinda big projects not represented on Zulip channel
  • are they are on arewelearningyet.com?
    • maybe one or two crates
    • should we have a deeper look at crate that might exist
  • https://github.com/spearow/juice
    • main contributor is on Zulip
  • https://github.com/finalfusion
• Not reasonable for people to know about us yet
  • should make sure the projects are captured on arewelearningyet.com as soon as we know about them though

Linfa

• new tests for the trees module
  • random forest
• chrism has a PR for some new documentation under review
  • There's a Julia ML library that has some good documentation--worth taking a look at
    • Also this: http://dlib.net/ml_guide.svg
    • And this one: https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html
• lorenz did review for ensamble learning today - looking good
• Merging the PR for the 0.2 documentation
  • CI was failing, because of cargo fmt issues
• doesn't have a README in its sub crate
  • https://github.com/rust-ml/linfa/tree/master/linfa-reduction
  • https://github.com/rust-ml/linfa/tree/master/linfa-svm
  • https://github.com/rust-ml/linfa/tree/master/linfa-trees
  • https://github.com/rust-ml/linfa/tree/master/linfa-kernel

Book

• potential showcase for machine learning in rust
• would be nice to have a "cookbook" like section
• how to use basic datasets (or how we are loading them etc)
• reference equasions and different algorithms
• Definitely should have some support for plotly charts for our outputs
• might be some great getting started issue
  • should probably have a basic chapter or guides before we ask them to start it up
• luca's ndarray guide
  • https://www.lpalmieri.com/posts/2019-12-01-taking-ml-to-production-with-rust-a-25x-speedup/
  • https://www.lpalmieri.com/posts/2019-04-07-scientific-computing-a-rust-adventure-part-2-array1/
• we should just open up the book for contributions
  • done ✅

Rust London Meetup

• No firm date but end of November (after the 20th)
• 0x009642 presenting
• could do a linfa example
  • annouce the 0.2 release, or aiming for the release to coinside with it
  • compare and contrast with SkLearn (Scikit-learn)

Actions

• check on crates that might exist but are not tracked on arewelearningyet.com
• Open up book repository to public
• Make sure CI passes on linfa docs update, then push
  • chrism
• add README.md for all subcrates of linfa, describing at least the purpose of the algorithm and how to run the example