ablaom/mljtutorial.jl

Notebooks for introducing the machine learning toolbox MLJ (Machine Learning in Julia)

MLJTutorial.jl

Notebooks for introducing the machine learning toolbox MLJ (Machine Learning in Julia)

Based on tutorials originally part of a 3.5 hour online workshop.

Options for running the tutorials

Prerequisites

• Familiarity with basic data manipulation in Julia: vectors, tuples, dictionaries, arrays, generating random numbers, tabular data (e.g., DataDrames.jl) basic stats, Distributions.jl.

• Familiarity with Machine Learning fundamentals and best practice.

Topics covered

Basic

• Part 1 - Data Representation

• Part 2 - Selecting, Training and Evaluating Models

• Part 3 - Transformers and Pipelines

Advanced

• Part 4 - Tuning hyper-parameters

• Part 5 - Advanced model composition

The tutorials include links to external resources and exercises with solutions.

More about the tutorials

• The tutorials focus on the machine learning part of the data science workflow, and less on exploratory data analysis and other conventional "data analytics" methodology

• Here "machine learning" is meant in a broad sense, and is not restricted to so-called deep learning (neural networks)

• The tutorials are crafted to rapidly familiarize the user with what MLJ can do and how to do it, and are not a substitute for a course on machine learning fundamentals. Examples do not necessarily represent best practice or the best solution to a problem.

Additional resources

• Slides from workshop given at ResBaz 2021 (Auckland)
• HelloJulia.jl - Resources from an Introduction to Julia workshop
• DataFrames.jl Cheatsheets
• MLJ Cheatsheet
• Common MLJ Workflows
• MLJ manual
• Data Science Tutorials in Julia

Credits

The author and maintainer of this repository is @ablaom. Pluto notebooks have been adapted from the julia scripts by @roland-KA who is also a maintainer.

mljtutorial.jl's Issues

Make model in tute 02 reproducible

@roland-KA has notices that misclassification_rate is occassionially zero, which might confuse some users. Perhaps consider a reproducible model (this one uses Flux dropout which cannot be passed an RNG at time of writing).

Update Pluto notebooks for ScientificTypes 3.0 and MLJ 0.17

Breaking changes in MLJ 0.17 are detailed here

• 01
• 02
• 03
• 04
• 05
• exercises

`evaluate!` mutating?

I have a general question concerning evaluate!: As the exclamation mark in it's name shows, this is a mutating function. But it's purpose is the calculation of different measures, which basically (at least from an abstract point of view) shouldn't require to mutate the data given to evaluate!?

So, why is it nonetheless defined as being mutating? Are there calculations of specific measures which require to mutate the data (so my assumption is wrong) or is it just a way to give the implementors of this function more degrees of freedom on how to implement it?

MLJ 0.17 breaks `fit!` in tutorial 02

I've updated the Pluto notebook for tutorial 02 to MLJ 0.17 as well as all other packages to their most recent versions.

Version changes in detail

Old versions:
CSV = "~0.9.10"
DataFrames = "~1.2.2"
Distributions = "~0.25.28"
MLJ = "~0.16.11"
MLJFlux = "~0.2.5"
Plots = "~1.23.6"
PlutoUI = "~0.7.19"
Tables = "~1.6.0"
UrlDownload = "~1.0.0"

New versions:
CSV = "~0.9.11"
DataFrames = "~1.3.1"
Distributions = "~0.25.37"
MLJ = "~0.17.0"
MLJFlux = "~0.2.6"
Plots = "~1.25.4"
PlutoUI = "~0.7.27"
Tables = "~1.6.1"
UrlDownload = "~1.0.0"

This breaks the first fit! statement in the tutorial (in 'Step 3'):

fit!(mach, rows=train, verbosity=2)

It comes up with the following error message:

UndefVarError: scitype not defined

reformat(::DataFrames.DataFrame)@core.jl:167
collate(::MLJFlux.NeuralNetworkClassifier{MLJFlux.Short, typeof(NNlib.softmax), Flux.Optimise.ADAM, typeof(Flux.Losses.crossentropy)}, ::DataFrames.DataFrame, ::CategoricalArrays.CategoricalVector{String, UInt32, String, CategoricalArrays.CategoricalValue{String, UInt32}, Union{}})@core.jl:247
fit(::MLJFlux.NeuralNetworkClassifier{MLJFlux.Short, typeof(NNlib.softmax), Flux.Optimise.ADAM, typeof(Flux.Losses.crossentropy)}, ::Int64, ::DataFrames.DataFrame, ::CategoricalArrays.CategoricalVector{String, UInt32, String, CategoricalArrays.CategoricalValue{String, UInt32}, Union{}})@mlj_model_interface.jl:56
var"#fit_only!#53"(::Vector{Int64}, ::Int64, ::Bool, ::typeof(MLJBase.fit_only!), ::MLJBase.Machine{MLJFlux.NeuralNetworkClassifier{MLJFlux.Short, typeof(NNlib.softmax), Flux.Optimise.ADAM, typeof(Flux.Losses.crossentropy)}, true})@machines.jl:592
#fit!#[email protected]:659[inlined]
top-level scope@Local: 1[inlined]

I can't relate this error to any of the changes coming with the new releases.

Update Literate scripts and Juptyer notebooks for ScientificTypes 3.0 and MLJ 0.17

Breaking changes are detailed here

• 01 #28
• 02 #25
• 03 #26
• 04 #29
• 05 and 99 #31

Tutorial 04: `predict(..., rows = 1:3)`

In part 04 of the tutorial, in section 'The tuning wrapper' there is a predict statement applied to the tuned model/machine:
predict(tuned_mach, rows = 1:3)

To me it is not quite clear why the parameter 1:3 is given to rows. I would have expected that XHorse (or a subset of it) would be used.