Hello everyone!
I’m happy to announce the Julia programming for Machine Learning course that I got to teach for the first time at TU Berlin this semester:
- GitHub repository: GitHub - adrhill/julia-ml-course: Julia programming for Machine Learning course at TU Berlin
- Website: Julia programming for Machine Learning
Contents
The course is taught in five weekly sessions of three hours. In each session, two lectures are taught:
| Week | Lecture | Content |
|---|---|---|
| 1 | 0 | General Information, Installation & Getting Help |
| 1 | Basics 1: Types, Control-flow & Multiple Dispatch | |
| 2 | 2 | Basics 2: Arrays, Linear Algebra |
| 3 | Plotting & DataFrames | |
| 3 | 4 | Basics 3: Data structures and custom types |
| 5 | Classical Machine Learning | |
| 4 | 6 | Automatic Differentiation |
| 7 | Deep Learning | |
| 5 | 8 | Workflows: Scripts, Experiments & Packages |
| 9 | Profiling & Debugging |
The first three weeks focus on teaching the fundamentals of the Julia programming language.
These weeks consist of longer lectures, followed up by shorter, “guided tours” of the Julia ecosystem, including plotting, data-frames and classical machine learning algorithms.
Week four is all about Deep Learning: A comprehensive lecture on automatic differentiation (AD) sheds light on differences between Julia’s various AD packages, before giving a brief overview of Flux’s Deep Learning ecosystem.
Finally, week five is all about starting your own Julia project, taking a look at the structure of Julia packages and different workflows for reproducible machine learning research.
This is followed up by a demonstration of Julia’s debugging and profiling utilities.
The lectures and the homework cover the following packages:
| Package | Lecture | Description |
|---|---|---|
| LinearAlgebra.jl | 2 | Linear algebra (standard library) |
| Plots.jl | 3 | Plotting & visualizations |
| DataFrames.jl | 3 | Working with and processing tabular data |
| MLJ.jl | 5 | Classical Machine Learning methods |
| ChainRules.jl | 6 | Forward- & reverse-rules for automatic differentiation |
| Zygote.jl | 6 | Reverse-mode automatic differentiation |
| Enzyme.jl | 6 | Forward- & reverse-mode automatic differentiation |
| ForwardDiff.jl | 6 | Forward-mode automatic differentiation |
| FiniteDiff.jl | 6 | Finite differences |
| FiniteDifferences.jl | 6 | Finite differences |
| Flux.jl | 7 | Deep Learning abstractions |
| MLDatasets.jl | 7 | Dataset loader |
| PkgTemplates.jl | 8 | Package template |
| DrWatson.jl | 8 | Workflow for scientific projects |
| Debugger.jl | 9 | Debugger |
| Infiltrator.jl | 9 | Debugger |
| ProfileView.jl | 9 | Profiler |
| Cthulhu.jl | 9 | Type inference debugger |
Implementation details
Inspired by the Introduction to Computational Thinking course, every lesson and homework is a Pluto notebook. Besides reactivity and automatic homework feedback, a big advantage of this approach is that each Pluto notebook contains a Project.toml and Manifest.toml. This gives students a reproducible Julia environment and allows us to defer the topic of environments to lesson 8.
Using PlutoSliderServer.jl, notebooks are automatically exported to HTML using GitHub Actions. The website is built using Franklin.jl and currently uses iFrames to include the Pluto notebooks. To make this look more seamless, the website re-uses a lot of Pluto’s CSS. Most of the magic then happens in this GitHub Action for deployment.
All of the code is MIT licensed.
If you are an author / contributor to one of the packages covered in this course, let me know if you would like me to make some changes. Contributions are most welcome!
Best,
Adrian