MLJ Cheatsheet

Starting an interactive MLJ session

julia> using MLJ

julia> MLJ_VERSION # version of MLJ for this cheatsheet
v"0.12.1"

Model search and code loading

info("PCA") retrieves registry metadata for the model called "PCA"

info("RidgeRegressor", pkg="MultivariateStats") retrieves metadata for "RidgeRegresssor", which is provided by multiple packages

models() lists metadata of every registered model.

models(x -> x.is_supervised && x.is_pure_julia) lists all supervised models written in pure julia.

models(matching(X)) lists all unsupervised models compatible with input X.

models(matching(X, y)) lists all supervised models compatible with input/target X/y.

With additional conditions:

models() do model
    matching(model, X, y) &&
    model.prediction_type == :probabilistic &&
        model.is_pure_julia
end

tree = @load DecisionTreeClassifier to load code and instantiate "DecisionTreeClassifier" model

tree2 = DecisionTreeClassifier(max_depth=2) instantiates a model type already in scope

ridge = @load RidgeRegressor pkg=MultivariateStats loads and instantiates a model provided by multiple packages

Scitypes and coercion

scitype(x) is the scientific type of x. For example scitype(2.4) == Continuous

Figure and Table for common scalar scitypes

Use schema(X) to get the column scitypes of a table X

coerce(y, Multiclass) attempts coercion of all elements of y into scitype Multiclass

coerce(X, :x1 => Continuous, :x2 => OrderedFactor) to coerce columns :x1 and :x2 of table X.

coerce(X, Count => Continuous) to coerce all columns with Count scitype to Continuous.

Ingesting data

Splitting any table into target and input (note semicolon):

using RDatasets
channing = dataset("boot", "channing")
y, X =  unpack(channing,
               ==(:Exit),            # y is the :Exit column
               !=(:Time);            # X is the rest, except :Time
               :Exit=>Continuous,    # correcting wrong scitypes (optional)
               :Entry=>Continuous,
               :Cens=>Multiclass)

Warning. Before julia 1.2 use col -> col != :Time insead of !=(:Time).

Splitting row indices into train/validation/test:

train, valid, test = partition(eachindex(y), 0.7, 0.2, shuffle=true, rng=1234) for 70:20:10 ratio

For a stratified split:

train, test = partition(eachindex(y), 0.8, stratify=y)

Getting data from OpenML:

table = openML.load(91)

Creating synthetic classification data:

X, y = make_blobs(100, 2) (also: make_moons, make_circles)

Creating synthetic regression data:

X, y = make_regression(100, 2)

Machine construction

Supervised case:

model = KNNRegressor(K=1) and mach = machine(model, X, y)

Unsupervised case:

model = OneHotEncoder() and mach = machine(model, X)

Fitting

fit!(mach, rows=1:100, verbosity=1, force=false)

Prediction

Supervised case: predict(mach, Xnew) or predict(mach, rows=1:100)

Similarly, for probabilistic models: predict_mode, predict_mean and predict_median.

Unsupervised case: transform(mach, rows=1:100) or inverse_transform(mach, rows), etc.

Inspecting objects

@more gets detail on last object in REPL

params(model) gets nested-tuple of all hyperparameters, even nested ones

info(ConstantRegressor()), info("PCA"), info("RidgeRegressor", pkg="MultivariateStats") gets all properties (aka traits) of registered models

info(rms) gets all properties of a performance measure

schema(X) get column names, types and scitypes, and nrows, of a table X

scitype(X) gets scientific type of a table

fitted_params(mach) gets learned parameters of fitted machine

report(mach) gets other training results (e.g. feature rankings)

Saving and retrieving machines

MLJ.save("trained_for_five_days.jlso", mach) to save machine mach

predict_only_mach = machine("trained_for_five_days.jlso") to deserialize.

Performance estimation

evaluate(model, X, y, resampling=CV(), measure=rms, operation=predict, weights=..., verbosity=1)

evaluate!(mach, resampling=Holdout(), measure=[rms, mav], operation=predict, weights=..., verbosity=1)

evaluate!(mach, resampling=[(fold1, fold2), (fold2, fold1)], measure=rms)

Resampling strategies (resampling=...)

Holdout(fraction_train=0.7, rng=1234) for simple holdout

CV(nfolds=6, rng=1234) for cross-validation

StratifiedCV(nfolds=6, rng=1234) for stratified cross-validation

or a list of pairs of row indices:

[(train1, eval1), (train2, eval2), ... (traink, evalk)]

Tuning

Tuning model wrapper

tuned_model = TunedModel(model=…, tuning=RandomSearch(), resampling=Holdout(), measure=…, operation=predict, range=…)

Ranges for tuning (range=...)

If r = range(KNNRegressor(), :K, lower=1, upper = 20, scale=:log)

then Grid() search uses iterator(r, 6) == [1, 2, 3, 6, 11, 20].

lower=-Inf and upper=Inf are allowed.

Non-numeric ranges: r = range(model, :parameter, values=…)

Nested ranges: Use dot syntax, as in r = range(EnsembleModel(atom=tree), :(atom.max_depth), ...)

Can specify multiple ranges, as in range=[r1, r2, r3]. For more range options do ?Grid or ?RandomSearch

Tuning strategies

Grid(resolution=10) or Grid(goal=50) for basic grid search

RandomSearch(rng=1234) for basic random search

Learning curves

For generating plot of performance against parameter specified by range:

curve = learning_curve(mach, resolution=30, resampling=Holdout(), measure=…, operation=predict, range=…, n=1)

curve = learning_curve(model, X, y, resolution=30, resampling=Holdout(), measure=…, operation=predict, range=…, n=1)

If using Plots.jl:

plot(curve.parameter_values, curve.measurements, xlab=curve.parameter_name, xscale=curve.parameter_scale)

Performance measures (metrics)

area_under_curve, accuracy, balanced_accuracy, cross_entropy, FScore, false_discovery_rate, false_negative, false_negative_rate, false_positive, false_positive_rate, l1, l2, mae, matthews_correlation, misclassification_rate, negative_predictive_value, positive_predictive_value, rms, rmsl, rmslp1, rmsp, true_negative, true_negative_rate, true_positive, true_positive_rate, BrierScore(), confusion_matrix

Available after doing using LossFunctions:

DWDMarginLoss(), ExpLoss(), L1HingeLoss(), L2HingeLoss(), L2MarginLoss(), LogitMarginLoss(), ModifiedHuberLoss(), PerceptronLoss(), SigmoidLoss(), SmoothedL1HingeLoss(), ZeroOneLoss(), HuberLoss(), L1EpsilonInsLoss(), L2EpsilonInsLoss(), LPDistLoss(), LogitDistLoss(), PeriodicLoss(), QuantileLoss()

measures() to get full list

info(rms) to list properties (aka traits) of the rms measure

Transformers

Built-ins include: Standardizer, OneHotEncoder, UnivariateBoxCoxTransformer, FeatureSelector, FillImputer, UnivariateDiscretizer, UnivariateStandardizer, ContinuousEncoder

Externals include: PCA (in MultivariateStats), KMeans, KMedoids (in Clustering).

models(m -> !m.is_supervised) to get full list

Ensemble model wrapper

EnsembleModel(atom=…, weights=Float64[], bagging_fraction=0.8, rng=GLOBAL_RNG, n=100, parallel=true, out_of_bag_measure=[])

Pipelines

With deterministic (point) predictions:

pipe = @pipeline OneHotEncoder KNNRegressor(K=3) target=UnivariateStandardizer

pipe = @pipeline OneHotEncoder KNNRegressor(K=3) target=v->log.(V) inverse=v->exp.(v))

Unsupervised:

pipe = @pipeline Standardizer OneHotEncoder

Define a supervised learning network:

Xs = source(X) ys = source(y)

... define further nodal machines and nodes ...

yhat = predict(knn_machine, W, ys) (final node)

Exporting a learning network as stand-alone model:

Supervised, with final node yhat returning point-predictions:

@from_network machine(Deterministic(), Xs, ys; predict=yhat) begin
    mutable struct Composite
	    reducer=network_pca
		regressor=network_knn
    end

Here network_pca and network_knn are models appearing in the learning network.

Supervised, with yhat final node returning probabilistic predictions:

@from_network machine(Probabilistic(), Xs, ys; predict=yhat) begin
    mutable struct Composite
        reducer=network_pca
        classifier=network_tree
    end

Unsupervised, with final node Xout:

@from_network machine(Unsupervised(), Xs; transform=Xout) begin
    mutable struct Composite
	    reducer1=network_pca
		reducer2=clusterer
    end
end