v0.13.0

MLJ v0.13.0

Diff since v0.12.1

Updates requirements for MLJBase, MLJModels and MLJScientificTypes to enable new features and fix some bugs:

• (enhancement) Add fitted_params_per_fold and report_per_fold properties to the object returned by evaluate/evaluate! to give user access to the outcomes of training for each train/test pair in resampling (#400, #616)

• (enhancement) Implement logpdf for UnivariateFinite distributions (JuliaAI/MLJBase.jl#411)

• (bug fix and deprecation) Fix bug related to creating new composite models by hand in special case of non-model hyper-parameters (not an issue with @pipeline or @from_network models). Introduce new return! syntax for doing this and deprecate calling of learning network machines ( JuliaAI/MLJBase.jl#390, JuliaAI/MLJBase.jl#391, JuliaAI/MLJBase.jl#377)

• (breaking) Change the behavior of evaluate/evaluate! so that weights are only passed to measures if explicitly passed using the key-word argument weights=... (JuliaAI/MLJBase.jl#405)

• (new model) Add UnivariateTimeTypeToContinuous model for converting assorted time data into Continuous data (JuliaAI/MLJModels.jl#295)

• (breaking) LDA models from MultivariateStats that take UnivariateFinite objects as a hyper-parameters to take dictionaries instead. Also change some default hyper-parameter values and improve the report (JuliaAI/MLJModels.jl#276)

• (enhancement) Improve efficiency of FillImputer model (JuliaAI/MLJModels.jl#292)

• (bug fix) Fix issue with syntax for loading models with a user-specified name (JuliaAI/MLJModels.jl#294)

• (mildly breaking) Regard Nothing as a native scientific type and declare scitype(nothing) = Nothing (old behaviour: scitype(nothing) = Unknown (JuliaAI/ScientificTypes.jl#112)
  Also updates the manual to reflect changes, and makes some improvements to the same.

• (breaking) Remove deprecated @pipeline syntax (JuliaAI/MLJBase.jl#350)

Closed issues:

• [docs] Document metadata_pkg, metadata_model and @mlj_model (#241)
• Improved docs for the data interface between MLJ and the world (#379)
• Document workaround for @mlj_model macro issues around negative defaults (#504)
• Document: can't use multithreading for python models (#525)
• In "Working with Categorical Data" part of docs, explain about int method. (#605)
• Allow access to the outcomes of fitting models on each fold in resampling (#616)
• Errors when columns read from the CSV file have missing entries. (#622)
• On models that fit a distribution to some data. (#641)
• Meta-issue: Add the JointProbabilistic <: Probabilistic subtype (#642)

Merged pull requests:

• Documentation updates for version 0.13 (#648) (@ablaom)
• For a 0.13.0 release (#653) (@ablaom)

v0.12.1

MLJ v0.12.1

Diff since v0.12.0

Closed issues:

• Should some if any "classic" ML algorithms accepts matrix input in addition to table input? (#209)
• Dataset generation for model inspection (#214)
• Coerce fails when a column has type Vector{Missing} (#549)
• OpenML integration: Columns as Symbols (#579)
• Issue to generate new releases (#583)
• Something not right with the Binder project file? (#587)
• range(pipeEvoTreeClassifier, :(selector.features), values = cases): ArgumentError: values does not have an appropriate type. (#590)
• fitted_params(LogisticModel): linear_binary_classifier = Any[] (#597)
• Problem fetching spawned @pipeline processes (#598)
• LogisticModel: pdf(ŷ[i], 1) does not work after the last release (#599)
• Improve error message for non-functions in @pipeline ... operation=... (#600)
• unable to use predict_mode() on machine associated with pipeline since MLJ 0.12.0 release (#601)
• Unable to use functions predict(), predict_mode() (#602)
• In the latest version, how do we do range(pipeXGBoostRegressor, :(xgr.max_depth), lower=3, upper=10) ? (#603)
• Add section on creating synthetic data to the manual (#604)
• Documentation for adding models: Discourage fields with type Union{Nothing,T} where T<:Real (#606)
• ERROR: LoadError: BoundsError: pkg = DecisionTree (#607)
• Old @from_network syntax still in docs (#608)
• Can't use @load inside a package (#613)
• max_samples parameter for RamdomForestClassifier (#619)
• Ambiguous assignment in soft scope on Julia 1.5 and Julia 1.6 (#624)
• inverse_transform of a PCA (#625)
• potential bug in MLJBase.roc_curve (#630)
• MLJ 0.12.0 doesn't work with Julia 1.5.0 (Windows) (#631)
• Meta-issue: Add the JointProbabilistic supervised model type (#633)

Merged pull requests:

• links to mybinder.org added (#585) (@darenasc)
• Add binder links in readme and documentation (#586) (@ablaom)
• binder option (#589) (@darenasc)
• Sanitize filename of workflows_learning_curves.png to fix git checkouts on Windows (#595) (@Sh4pe)
• Binder with instantiate (#596) (@ablaom)
• Minor readme figure changes (not for new release) (#609) (@ablaom)
• Binder change. Not for new release (#610) (@ablaom)
• Binder again (#611) (@ablaom)
• Dump JIT demo from binder for now (only works in regular notebook) (#612) (@ablaom)
• paper.md edit . No release (#614) (@ablaom)
• switch to new @from_network synthax in docs (#615) (@OkonSamuel)
• Add arXiv version 1 of paper. No release (#617) (@ablaom)
• Update links to arXiv version 1 of paper. No release (#618) (@ablaom)
• Add links to JuliaCon 2020 workshop. No release. (#623) (@ablaom)
• Add Generating Synthetic Data section to manual (#626) (@ablaom)
• Manual updates. No new release. (#627) (@ablaom)
• fix readme links for manual. no release (#629) (@ablaom)
• corrected wrong method signatures in docs (#635) (@OkonSamuel)
• Doc fixes. No release. (#638) (@ablaom)
• Discourage Union types for fields (#639) (@ablaom)
• Doc fixes. No release (#640) (@ablaom)
• Create the "PullRequestBaseBranch" GitHub Action (#643) (@DilumAluthge)
• Install CompatHelper on this repository (#644) (@DilumAluthge)
• New GitHub action. No new release. (#645) (@ablaom)
• Export JointProbabilistic and predict_joint (#646) (@DilumAluthge)
• For a 0.12.1 release (#647) (@ablaom)

v0.11.6

MLJ v0.11.6 (NOT LATEST RELEASE see 0.12 below)

Diff since v0.11.5

Patch removing redundant files causing problems for Windows #591

Closed issues:

• Integrate flux models (#33)
• Conflict w Scikitlearn.jl (#502)
• Accessing nested machines is (still) awkward (#553)
• Standardization of multi-targets (#568)
• Add note at MLJ landing page that MLJ wraps a majority of sk-learn models (#573)
• OpenML integration: Columns as Symbols (#579)
• OpenML integration: Kmeans is not fitting (#580)

Merged pull requests:

• export brier_score (#572) (@OkonSamuel)
• Add binder notebooks to the repo (#577) (@darenasc)
• Add MLJFlux to Model List, alphabetize model list (#578) (@azev77)
• For a 0.12.0 release (#581) (@ablaom)
• For a 0.12.0 release (#582) (@ablaom)
• links to mybinder.org added (#585) (@darenasc)
• Sanitize filename of workflows_learning_curves.png to fix git checkouts on Windows (#595) (@Sh4pe)

v0.12.0

MLJ v0.12.0

Diff since v0.11.5

This release provides updates to breaking changes in MLJBase 0.14.0, and MLJModels 0.11.0 and MLJTuning 0.4.0. For a complete list of changes, closed issues and pull requests, refer to the linked release notes.

It also updates the MLJ documentation to reflect the new changes.

Summary

Main breaking changes:

• Adds restrictions to acceleration options - nesting
  distributed processes within multithread process is now disallowed.

• Adds more user-friendly interface for inspecting training reports and
  fitted parameters of composite models. For example, if composite = @pipeline OneHotEncoder KNNRegressor and mach = machine(composite, X, y), then access the fitted parameters of the machine associated
  with KNNRegressor using fitted_params(mach).knn_regressor.

• The @from_network syntax has been changed to make it more
  expressive. In particular, through the new concept of learning
  network machines it is possible to export a learning network to a
  composite type supporting multiple operations (e.g., predict and
  transform, as in clustering). See the
  manual
  for details. The old syntax is no longer supported.

Other enhancements of note:

• Adds MLJFlux
  models to the registry for incorporating neural network models.

• A more economic @pipeline syntax has been introduced. For example,
  pipe = @pipeline OneHotEncoder PCA(maxoutdim=3) defines model
  pipe with automatically generated field names and model type name. Target inverse
  transformations now ooccur immediately after the supervised model in
  a @pipeline, instead of at the end, unless invert_last=true. The old syntax is
  available but deprecated.

• It is now possible to simulataneously load model code for models
  having the same name but from different packages, when using @load
  or load.

• Removes the requirement to specify the kind of source node, as in
  source(y, kind=:target). The role of source nodes in learning
  networks is now inferred from the order in which they appear in
  learning network machine constructors (see above).

Deprecations:

• The old @pipeline syntax.

• The specification of kind when constructing a source node.

• The use of fitresults() when exporting learning networks "by
  hand". See the manual for the new way to do this.

Closed issues:

• Integrate flux models (#33)
• Conflict w Scikitlearn.jl (#502)
• Accessing nested machines is (still) awkward (#553)
• Add note at MLJ landing page that MLJ wraps a majority of sk-learn models (#573)
• OpenML integration: Kmeans is not fitting (#580)

Merged pull requests:

• export brier_score (#572) (@OkonSamuel)
• Add binder notebooks to the repo (#577) (@darenasc)
• Add MLJFlux to Model List, alphabetize model list (#578) (@azev77)
• For a 0.12.0 release (#581) (@ablaom)
• For a 0.12.0 release (#582) (@ablaom)

v0.11.5

MLJ v0.11.5

Diff since v0.11.4

Closed issues:

• computing UnivariateFinite matrix seems to be substantially slow (#511)
• AMES tutorial doesn't work (UndefVarError) if ScikitLearn.jl or StatsBase.jl are loaded (#534)
• DimensionMismatch in evaluate() (#540)
• Hyperparameter tuning of KNN classifier (#543)
• Decision trees from ScikitLearn.jl not available (#545)
• Export supports_weights() and prediction_type() (#547)
• Testing for type of values in a range too restrictive (#548)
• SVC won't tune cost (#551)
• Implementation of Tversky Loss (#554)
• Fix broken MLJ logo in the manual (MLJ github pages) (#555)
• Add configuration options for RandomForestClassifier.n_subfeatures that depend on the data size (#557)
• Change DecisionTree.jl n_subfeatures default to -1 for random forest classifier and regressor (#558)
• Tutorial link in Getting Started doesn't link to right spot (#560)
• Old documentation deployed on github pages (#561)
• Document how to load models without the @load macro (#562)
• Request for monte-carlo cross validation (#564)
• Loading SKLearn packages causes Julia to crash (#565)

Merged pull requests:

• typos (#541) (@ablaom)
• typo (#544) (@OkonSamuel)
• update tutorial link (#563) (@OkonSamuel)
• Update the documentation at /docs to reflect recent changes at MLJBase (#566) (@ablaom)
• Documentation update (not to trigger a new release) (#567) (@ablaom)
• For a 0.11.5 release - Bump [compat] MLJModels="^0.10" (#569) (@ablaom)
• For a 0.11.5 release (#570) (@ablaom)

v0.11.4

MLJ v0.11.4

Diff since v0.11.3

Closed issues:

• Working with models with the same name from different packages (#446)
• Update readme: MLJModels does not need to be in user's env after MLJModels 0.9.10 (#520)
• More informative error for supplying model type instead of instance in range method (#521)
• Readme inconsistency (#524)
• Not loading to all workers: @Everywhere @load RandomForestClassifier pkg = DecisionTree (#527)
• Re-export mape and MAPE from MLJBase (#532)

Merged pull requests:

• docs only (#531) (@ablaom)
• Re-export mape (#533) (@azev77)
• Update paper (#535) (@ablaom)
• Typo in paper (#536) (@ablaom)
• Export make_moons, etc; re-organize documentation (#537) (@ablaom)
• For a 0.11.4 release (#538) (@ablaom)
• teak readme again (#539) (@ablaom)

v0.11.3

MLJ v0.11.3

Diff since v0.11.2

• Update CategoricalArrays compatibility requirement to "^0.8" (#528) (@ablaom)

v0.11.2

MLJ v0.11.2

Diff since v0.11.1

Fix bug in defining MLJ_VERSION (#508 PR #509)

Merged pull requests:

• A more correct and robust way of getting the MLJ version number (#509) (@DilumAluthge)
• For a 0.11.2 release (#510) (@ablaom)

v0.11.1

MLJ v0.11.1

Diff since v0.11.0

Minor issues only:

• #497
• Revise cheatsheet (#474)

Closed issues:

• Add sample-weight interface point? (#177)
• Add default_measure to learning_curve! (#283)
• Flush out unsupervised models in "Adding models for general use" section of manual (#285)
• is_probabilistic=true in @pipeline syntax is clunky (#305)
• [suggestions] Unroll the network in @from_network (#311)
• Towards stabilisation of the core API (#318)
• failure on nightly (1.4) (#384)
• Documentation of extracting best fitted params (#386)
• incorporate input_scitype and target_scitype declarations for @pipeline models (#412)
• "Supervised" models with no predict method (#460)
• Use OpenML.load to iris data set in the Getting Started page of docs? (#461)
• Review cheatsheet (#474)
• Re-export UnsupervisedNetwork from MLJBase (#497)
• Broken link for MLJ tour in documentation (#501)

Merged pull requests:

• For a 0.11.1 release (#506) (@ablaom)

v0.11.0

MLJ v0.11.0

Diff since v0.10.3

Make compatibility updates to MLJBase and MLJModels to effect the following changes to MLJ (see the linked release notes for links to the issues/PRs)):

• (new model) Add LightGBM models LightGBMClassifier and
  LightGBMRegressor

• (new model) Add new built-in model, ContinuousEncoder, for
  transforming all features of a table to Continuous scitype,
  dropping any features that cannot be so transformed

• (new model) Add ParallelKMeans model, KMeans, loaded with
  @load KMeans pkg=ParallelKMeans

• (mildly breaking enhancement) Arrange for the CV
  resampling strategyto spread fold "remainders" evenly among folds in
  train_test_pairs(::CV, ...) (a small change only noticeable in
  small datasets)

• (breaking) Restyle report and fitted_params for exported
  learning networks (e.g., pipelines) to include a dictionary of reports or
  fitted_params, keyed on the machines in the underlying learning
  network. New doc-strings detail the new behaviour.

• (enhancement) Allow calling of transform on machines with Static models without
  first calling fit!

• Allow machine constructor to work on supervised models that take nothing for
  the input features X (for models that simply fit a
  sampler/distribution to the target data y) (#51)

Also:

• (documentation) In the "Adding New Models for General Use"
  section of the manual, add detail on how to wrap unsupervised
  models, as well as models that fit a sampler/distribution to data

• (documentation) Expand the "Transformers" sections of the
  manual, including more material on static transformers and
  transformers that implement predict (#393)

Closed issues:

• Add tuning by stochastic search (#37)
• Improve documentation around static transformers (#393)
• Error in docs for model search (#478)
• Update [compat] StatsBase="^0.32,^0.33" (#481)
• For a 0.10.3 release (#483)
• Help with coercing strings for binary data into Continuous variables (#489)
• EvoTree Error (#490)
• Add info with workaround to avoid MKL error (#491)
• LogisticClassifier pkg = MLJLinearModels computes a number of coefficients but not the same number of mean_and_std_given_feature (#492)
• MethodError: no method matching... (#493)
• For a 0.10.4 release (#495)
• Error: fitted_params(LogisticModel) (#498)

Merged pull requests:

• Fixes the docs for model search (#486) (@ven-k)
• Readme update (#494) (@ablaom)
• paper added to MLJ repo (#496) (@darenasc)
• For a 0.11.0 release (#499) (@ablaom)
• For a 0.11.0 release (#500) (@ablaom)