NEWS.md

MLJ News

News for MLJ and its satellite packages: MLJBase, MLJModels, and ScientificTypes.

Latest release notes

MLJ (general users)

MLJBase | MLJModels | ScientificTypes (mainly for developers)

News

Note: New patch releases are no longer being announced below. Refer to the links above for complete release notes.

30 Oct 2019

• MLJModels 0.5.3 released.

• MLJBase 0.7.2 released.

22 Oct 2019

MLJ 0.5.1 released.

21 Oct 2019

• MLJBase 0.7.1 released.

• ScientificTypes 0.2.2 released.

• MLJModels 0.5.2 released.

17 Oct 2019

MLJBase 0.7 released.

11 Oct 2019

MLJModels 0.5.1 released.

30 Sep 2019

MLJ 0.5 released.

29 Sep 2019

MLJModels 0.5 released.

26 Sep 2019

MLJBase 0.6 released.

Older release notes

MLJ 0.4.0

• (Enhancment) Update to MLJBase 0.5.0 and MLJModels 0.4.0. In particular, this updates considerably the list of wrapped scikit-learn models available to the MLJ user:

• ScikitLearn.jl
  • SVM: SVMClassifier, SVMRegressor, SVMNuClassifier, SVMNuRegressor, SVMLClassifier, SVMLRegressor,
  • Linear Models (regressors): ARDRegressor, BayesianRidgeRegressor, ElasticNetRegressor, ElasticNetCVRegressor, HuberRegressor, LarsRegressor, LarsCVRegressor, LassoRegressor, LassoCVRegressor, LassoLarsRegressor, LassoLarsCVRegressor, LassoLarsICRegressor, LinearRegressor, OrthogonalMatchingPursuitRegressor, OrthogonalMatchingPursuitCVRegressor, PassiveAggressiveRegressor, RidgeRegressor, RidgeCVRegressor, SGDRegressor, TheilSenRegressor

• (Enhancement) The macro @pipeline allows one to construct linear (non-branching) pipeline composite models with one line of code. One may include static transformations (ordinary functions) in the pipeline, as well as target transformations for the supervised case (when one component model is supervised).

• (Breaking) Source nodes (type Source) now have a kind field, which is either :input,:target or :other, with :input the default value in the source constructor. If building a learning network, and the network is to be exported as a standalone model, then it is now necessary to tag the source nodes accordingly, as in Xs = source(X) and ys = source(y, kind=:target).

• (Breaking) By virtue of the preceding change, the syntax for exporting a learning network is simplified. Do?@from_network for details. Also, one now uses fitresults(N) instead of fit results(N, X, y) and fitresults(N, X) when exporting a learning network N "by hand"; see the updated manual for details.

• (Breaking) One must explicitly state if a supervised learning network being exported with @from_network is probabilistic by adding is_probablistic=true to the macro expression. Before, this information was unreliably inferred from the network.

• (Enhancement) Add macro-free method for loading model code into an arbitrary module. Do ?load for details.

• (Enhancement) @load now returns a mode instance with default hyperparameters (instead of nothing), as in tree_model = @load DecisionTreeRegressor

• (Breaking) info("PCA") now returns a named-tuple, instead of a dictionary, of the properties of a the model named "PCA"

• (Breaking) The list returned by models(conditional) is now a list of complete metadata entries (named-tuples, as returned by info). An entry proxy appears in the list exactly when conditional(proxy) == true. Model query is simplified; for example models() do model model.is_supervised && model.is_pure_julia end finds all pure julia supervised models.

• (Bug fix) Introduce new private methods to avoid relying on MLJBase type piracy MLJBase #30.

• (Enhancement) If composite is a a learning network exported as a model, and m = machine(composite, args...) then report(m) returns the reports for each machine in the learning network, and similarly for fitted_params(m).

• (Enhancement) MLJ.table, vcat and hcat now overloaded for AbstractNode, so that they can immediately be used in defining learning networks. For example, if X = source(rand(20,3)) and y=source(rand(20)) then MLJ.table(X) and vcat(y, y) both make sense and define new nodes.

• (Enhancement) pretty(X) prints a pretty version of any table X, complete with types and scitype annotations. Do ?pretty for options. A wrap of pretty_table from PrettyTables.jl.

• (Enhancement) std is re-exported from Statistics

• (Enhancement) The manual and MLJ cheatsheet have been updated.

• Performance measures have been migrated to MLJBase, while the model registry and model load/search facilities have migrated to MLJModels. As relevant methods are re-exported to MLJ, this is unlikely to effect many users.

MLJModels 0.4.0

• (Enhancement) Add a number of scikit-learn model wraps. See the above MLJ 0.4.0 release notes for a detailed list.

• The following have all been migrated to MLJModels from MLJ:

  • MLJ's built-in models (e.g., basic transformers such as OneHotEncoder)

  • The model registry metadata (src/registry/METADATA.toml)

  • The metadata @update facility for administrator registration of new models

  • The @load macro and load function for loading code for a registered model

  • The models and localmodels model-search functions

  • The info command for returning the metadata entry of a model

• (Breaking) MLJBase v0.5.0, which introduces some changes and additions to model traits, is a requirement, meaning the format of metadata as changed.

• (Breaking) The model method for retrieving model metadata has been renamed back to info, but continues to return a named-tuple. (The MLJBase.info method, returning the dictionary form of the metadata, is now called MLJBase.info_dic).

MLJBase 0.5.0

• Bump ScientificTypes requirement to v0.2.0

• (Enhancement) The performance measures API (built-in measures + adaptor for external measures) from MLJ has been migrated to MLJBase. MLJ.

• (Breaking) info, which returns a dictionary (needed for TOML serialization) is renamed to info_dic. In this way "info" is reserved for a method in MLJModels/MLJ that returns a more-convenient named-tuple

• (Breaking) The is_probabilistic model trait is replaced with prediction_type, which can have the values :deterministic, :probabilistic or :interval, to allow for models predicting real intervals, and for consistency with measures API.

• (Bug fix, mildly breaking) The package_license model trait is now included in info_dict in the case of unsupervisd models.

• (Enhancement, mildly breaking) Add new model traits hyperparameters, hyperparameter_types, docstring, and implemented_operations (fit, predict, inverse_transform, etc) (#36, #37, #38)

• (Enhancement) The MLJBase.table and MLJBase.matrix operations are now direct wraps of the corresponding Tables.jl operations for improved performance. In particular MLJBase.matrix(MLJBase.table(A)) is essentially a non-operation, and one can pass MLJBase.matrix the keyword argument transpose=... .

• (Breaking) The built-in dataset methods load_iris, load_boston, load_ames, load_reduced_ames, load_crabs return a raw DataFrame, instead of an MLJTask object, and continue to require import CSV to become available. However, macro versions @load_iris, etc, are always available, automatically triggering import CSV; these macros return a tuple (X, y) of input DataFrame and target vector y, with scitypes appropriately coerced. (MLJ #224)

• (Enhancement) selectrows now works for matrices. Needed to allow matrices as "node type" in MLJ learning networks; see MLJ #209.

• (Bug) Fix problem with == for MLJType objects (#35)

• (Breaking) Update requirement on ScientficTypes.jl to v0.2.0 to mitigate bug with coercion of column scitypes for tables that are also AbstractVectors, and to make coerce more convenient.

• (Enhancement) Add new method unpack for splitting tables, as in y, X = unpack(df,==(:target),!=(:dummy)). See doc-string for details.

• (Bug fix) Remove type piracy in get/setproperty! (#30)

ScientificTypes 0.2.0

• (Breaking) The argument order is switched in coerce methods. So now use coerce(v, T) for a vector v and scientific type T and coerce(X, d) for a table X and dictionary d.

• (Feature) You can now call coerce on tables without needing to wrap specs in a dictionary, as in scitype(X, :age => Continuous, :ncalls => Count).

ScientficTypes 0.1.3

Release notes

MLJ 0.4.0

• Introduction of traits for measures (loss functions, etc); see top of /src/measures.jl for definitions. This

  • allows user to use loss functions from LossFunctions.jl,
  • enables improved measure checks and error message reporting with measures
  • allows evaluate! to report per-observation measures when available (for later use by Bayesian optimisers, for example)
  • allows support for sample-weighted measures playing nicely with rest of API

• Improvements to resampling:

  • evaluate! method now reports per-observation measures when available
  • sample weights can be passed to evaluate! for use by measures that support weights
  • user can pass a list of train/evaluation pairs of row indices directly to evaluate!, in place of a ResamplingStrategy object
  • implementing a new ResamplingStrategy is now straightforward (see docs)
  • one can call evaluate (no exclamation mark) directly on model + data without first constructing a machine, if desired

• Doc strings and the manual have been revised and updated. The manual includes a new section "Tuning models", and extra material under "Learning networks" explaining how to export learning networks as stand-alone models using the @from_network macro.

• Improved checks and error-reporting for binding models to data in machines.

• (Breaking) CSV is now an optional dependency, which means you now need to import CSV before you can load tasks with load_boston(), load_iris(), load_crabs(), load_ames(), load_reduced_ames()

• Added schema method for tables (re-exported from ScientificTypes.jl). Returns a named tuple with keys :names, :types, :scitypes and :nrows.

• (Breaking) Eliminate scitypes method. The scientific types of a table are returned as part of ScientificTypes schema method (see above)

MLJModels 0.3.0

Release notes

MLJBase v0.4.0

Release notes

ScientificTypes 0.1.2

• New package to which the scientific types API has been moved (from MLJBase).

MLJBase v0.3.0

• Make CSV an optional dependency (breaking). To use load_iris(), load_ames(), etc, need first to import CSV.

MLJBase v0.2.4

• Add ColorImage and GreyImage scitypes

• Overload in method for subtypes of Model (apparently causing Julia crashes in an untagged commit, because of a method signature ambiguity, now resolved).

MLJ v0.2.5

• Add MLJ cheatsheet

• Allow models to query specific traits, in addition to tasks. Query ?models for details

• add @from_networks macro for exporting learning networks as models (experimental).

MLJModels v0.2.4

• Add compatibility requirement MLJBase="0.2.3"

MLJBase v0.2.3

• Small changes on definitions of == and isequal for MLJType objects. In particular, fields that are random number generators may change state without effecting an object's == equivalence class.

• Add @set_defaults macro for generating keywork constructors for Model subtypes.

• Add abstract type UnsupervisedNetwork <: Unsupervised.

MLJ v0.2.3

• Fixed bug in models(::MLJTask) method which excluded some relevant models. (#153)

• Fixed some broken links to the tour.ipynb.

MLJ v0.2.2

• Resolved these isssues:

  • Specifying new rows in calls to fit! on a Node not triggering retraining. (#147)

  • fit! of Node sometimes calls update on model when it should call fit on model (#146)

  • Error running the tour.ipynb notebook (#140)

  • For reproducibility, include a Manifest.toml file with all examples. (#137)

• Activated overalls code coverage (#131)

• Removed local version of MultivariateStats (now in MLJModels, see below).

• Minor changes to OneHotEncoder, in line with scitype philosophy.

MLJBase v0.2.2

• Fix some minor bugs.

• Added compatibility requirement CSV v0.5 or higher to allow removal of allowmissing keyword in CSV.read, which is to be depreciated.

Announcement: MLJ tutorial and development sprint

• Details here Applications close May 29th 5pm (GMTT + 1 = London)

MLJModels v0.2.3

• The following support vector machine models from LIBSVM.jl have been added: EpsilonSVR, LinearSVC, NuSVR, NuSVC, SVC, OneClassSVM.

MLJModels v0.2.2

• MulitivariateStats models RidgeRegressor and PCA migrated here from MLJ. Addresses: MLJ #125.

MLJModels v0.2.1

• ScikitLearn wraps ElasticNet and ElasticNetCV now available (and registered at MLJRegistry). Resolves: MLJ #112

MLJ v0.2.1

• Fix a bug and related problem in "Getting Started" docs: [#126](#126 .

MLJBase 0.2.0, MLJModels 0.2.0, MLJ 0.2.0

• Model API refactored to resolve #93 and #119 and hence simplify the model interface. This breaks all implementations of supervised models, and some scitype methods. However, for the regular user the effects are restricted to: (i) no more target_type hyperparameter for some models; (ii) Deterministic{Node} is now DeterministicNetwork and Probabillistic{Node} is now ProbabilisticNetwork when exporting learning networks as models.

• New feature: Task constructors now allow the user to explicitly specify scitypes of features/target. There is a coerce method for vectors and tables for the user who wants to do this manually. Resolves: #119

Official registered versions of MLJBase 0.1.1, MLJModels 0.1.1, MLJ 0.1.1 released

• Minor revisions to the repos, doc updates, and a small breaking change around scitype method names and associated traits. Resolves: #119

unversioned commits 12 April 2019 (around 00:10, GMT)

• Added out-of-bag estimates for performance in homogeneous ensembles. Resolves: #77

unversioned commits 11 April 2019 (before noon, GMT)

• Removed dependency on unregistered package TOML.jl (using, Pkg.TOML instead). Resolves #113

unversioned commits 8 April 2019 (some time after 20:00 GMT)

• Addition of XGBoost models XGBoostRegressor, XGBoostClassifier and XGBoostCount. Resolves #65.

• Documentation reorganized as GitHub pages. Includes some additions but still a work in progress.

unversioned commits 1 March 2019 (some time after 03:50 GMT)

• Addition of "scientific type" hierarchy, including Continuous, Discrete, Multiclass, and Other subtypes of Found (to complement Missing). See Getting Started for more one this. Resolves: #86

• Revamp of model traits to take advantage of scientific types, with output_kind replaced with target_scitype_union, input_kind replaced with input_scitype. Also, output_quantity dropped, input_quantity replaced with Bool-valued input_is_multivariate, and is_pure_julia made Bool-valued. Trait definitions in all model implementations and effected meta-algorithms have been updated. Related: #81

• Substantial update of the core guide Adding New Models to reflect above changes and in response to new model implementer queries. Some design "decisions" regarding multivariate targets now explict there.

• the order the y and yhat arguments of measures (aka loss functions) have been reversed. Progress on: #91

• Update of Standardizer and OneHotEncoder to mesh with new scitypes.

• New improved task constructors infer task metadata from data scitypes. This brings us close to a simple implementation of basic task-model matching. Query the doc-strings for SupervisedTask and UnsupervisedTask for details. Machines can now dispatch on tasks instead of X and y. A task, task, is now callable: task() returns (X, y) for supervised models, and X for unsupervised models. Progress on: #86

• the data in the load_ames() test task has been replaced by the full data set, and load_reduced_ames() now loads a reduced set.