v0.19.7

MLJBase v0.19.7

Diff since v0.19.6

• (enhancement) Add Cohen's kappa, Kappa (#738) @Rahulub3r

Closed issues:

• Add Cohen's kappa to performance measures. (#689)
• Add a doc-string for Node type (#722)

Merged pull requests:

• Refactor model scitype check2 (#732) (@ablaom)
• Fix the test that addresses #377 (#737) (@ablaom)
• Added Kappa Measure (#738) (@Rahulub3r)
• Fix iteration_parameter for pipelines in case supervised component has value nothing (#740) (@ablaom)
• Small re-org for clarity (#741) (@ablaom)
• Fix docs for nodes (#742) (@ablaom)
• For a 0.19.7 release (#743) (@ablaom)

v0.19.6

MLJBase v0.19.6

Diff since v0.19.5

Closed issues:

• MLJBase fails to precompile (#734)

Merged pull requests:

• Drop filename arg from save(filename, model, ...) (#724) (@ablaom)
• Fix CategoricalDistributions.jl reexports (#735) (@davnn)
• For a 0.19.6 release (#736) (@ablaom)

v0.19.5

MLJBase v0.19.5

Diff since v0.19.4

• (enhancement) Add cross-validation scores for Stack base learners to the training report based on user-specified measures (#712) @olivierlabayle

Closed issues:

• Documentation is not deploying properly. (#726)

Merged pull requests:

• Stacking internal evaluation (#712) (@olivierlabayle)
• Get rid of ErrorNodes and fix corner case bug in replace (#723) (@ablaom)
• Fix the docs deployment (#727) (@DilumAluthge)
• Build the dev docs on pushes to the dev branch (#729) (@DilumAluthge)
• For a 0.19.5 release (#730) (@ablaom)

v0.19.4

MLJBase v0.19.4

Diff since v0.19.3

• Drop support for julia < 1.6
• Add doc strings to built-in measure instances #719
• Programatically re-export all measure types, aliases and instances #720

Closed issues:

• Fail to precompile (#715)
• documentation of log_loss/LogLoss (#716)
• Bind measure doc-string to all aliases (#717)
• Programmatically re-export all measures (#718)

Merged pull requests:

• Add docstrings for instances (aliases) of docstrings (#719) (@ablaom)
• Programmatically export measure types, type aliases, and built-in instances (#720) (@ablaom)
• For a 0.19.4 release (#721) (@ablaom)

v0.19.3

MLJBase v0.19.3

Diff since v0.19.2

• overload the target_scitype trait to give sensible value for the Stack model type (#711) @olivierlabayle

Closed issues:

• Yank MLJBase 0.19.0 and 0.19.1 from registry (#708)
• target_scitype for SupervisedPipelines (#710)

Merged pull requests:

• add target_scitype trait for supervised pipelines (#711) (@olivierlabayle)
• For a 0.19.3 release (#714) (@ablaom)

v0.19.2

MLJBase v0.19.2

Diff since v0.19.1

Merged pull requests:

• extend MLJModelInterface.scitype method (#706) (@OkonSamuel)
• For a 0.19.2 release (#707) (@ablaom)

v0.19.1

MLJBase v0.19.1

Note: Yanked from the General.jl registry. Superseded by 0.19.2.

Diff since v0.19.0

Closed issues:

• Bug: evaluate/evaluate! is still returning a named tuple (#702)

Merged pull requests:

• Fix PerformanceEvaluation bug (#703) (@ablaom)
• For a 0.19.1 release (#704) (@ablaom)

v0.19.0

MLJBase v0.19.0

Note: Yanked from the General.jl registry. Superseded by 0.19.2, for which notes below still apply.

Diff since v0.18.26

• (enhancement) Add a new Pipeline type for constructing pipelines without macros. Pipelines are to be constructed using the syntax model1 |> model2 |> ... or with the constructor Pipeline which exposes more options. The @pipeline macro is deprecated (#664)
• (enhancement) Add the metamodel TransformedTargetModel for wrapping supervised models in transformations of the target variable, which can be learned transformations (eg, standardisation). Previously this functionality was available as part of @pipeline (#678)
• (breaking) The |> syntax is no longer available for creating learning networks as it is now reserved for pipeline creation and concatenation. Also, model |> f has a new meaning when model is a Model subtype or instance. Previously, it meant f(model), but now it is syntactic sugar for Pipeline(model, f) (#664)
• (breaking) The UnivariateFinite distribution and constructor, and the UnivariateFiniteArray type, are no longer available, as they are now provided by a separate package CategoricalDistributions.jl (which includes generalization to signed and complex measures, and a better display) (#688)
• (breaking) schema(X) no longer includes the nrows property. Use nrows(X) instead. This follows from changes in the dependency ScientificTypes.jl (#698)
• Progress meters are no longer displayed for processes concluding in 0.1 s or less (#698)
• (mildly breaking) To improve robustness, the range(model, :hyperparameter, ...) constructor now tries to infer type information for the range from the corresponding fieldtype for field :hyperparameter, rather than from the type of the current value. Fix a bug with the display. (#666)
• (mildly breaking) unpack(table, p1, p2, ...) now includes an extra component in its return value, namely a table with all columns not selected by any of the predicates p1, p2, ... Generally existing code will just safely ignore the extra component (#691)
• (breaking) Because more plotting algorithms support it, the default scale for unbounded NumericRanges is changed from :log to :log10 (#677).
• (breaking) Remove deprecated code for exporting learning networks by hand (#643), which is now achieved using return! method (docs).
• (enhancement) The partition function can now be called with a tuple of data arguments, for "synchronised" partitioning, but this requires specifying multi=true (because some tables are tuples) as in (Xtrain, ytrain), (Xtest, ytest) = partition((X, y), 0.6, rng=123, multi=true) #696
• (mildly breaking) Make the objects returned by evaluate/evaluate! into a struct (formerly a named tuple) and add informative doc-string (#697)
• (enhancement) Create a way to include the state, after training, of arbitrary nodes of a learning network, in the report of a model created by exporting the learning network (#644)

Closed issues:

• metrics when the target contains missing (#93)
• Implement inverse_transform for @pipeline models where this makes sense (#384)
• Migrate UnivariateFinite (for categorical distributions) out to new package (#504)
• Tracking issue for a bunch of smaller stuff (#573)
• Change unpack to avoid colname -> true mystery (#621)
• Making measures array friendly (#631)
• Add a supervised model wrapper to implement target transformations (#642)
• Remove legacy code for Composite models whose fitresult is a node instead of a signature (#643)
• For a 0.19 release (#665)
• Pipelines should pass through training losses when appropriate (#672)
• For unbounded NumericRange change scale defaults from :log to :log10 (#677)
• Further generalisation of partition (#682)

Merged pull requests:

• Add interface points for accessing the internal state of an exported learning network composite model (#644) (@ablaom)
• For a 0.19 release (target = dev) (#669) (@ablaom)
• Add wrapper to transform targets of (semi)supervised models (#678) (@ablaom)
• Add pipeline dep warnings (#687) (@ablaom)
• Remove UnivariateFinite stuff, replacing with CategoricalDistributions.jl dependency (#688) (@ablaom)
• Address fact that UnivariateFinite is no longer subtype of Distributions.Distribution (#690) (@ablaom)
• Modify unpack to return an extra value with all remaining columns (#691) (@ablaom)
• Change :log default for unbounded range scales to :log10 (#692) (@ablaom)
• Remove some deprecated code (#693) (@ablaom)
• Allow target to be a model type in TransformedTargetModel (#694) (@ablaom)
• Move _wrapped_function.jl contents into deprecated.jl (#695) (@ablaom)
• Add synchronised partitioning of multiple objects (#696) (@ablaom)
• Make PerformanceEvaluation object a struct (#697) (@ablaom)
• Avoid depending on TRAIT_FUNCTION_GIVEN_NAME functionality. (#698) (@OkonSamuel)
• For a 0.19 release (#701) (@ablaom)

v0.18.26

MLJBase v0.18.26

Diff since v0.18.25

• Fix the orientation trait of RSquared (#684)
• Fix a bug in from_network_preprocess responsible for @pipeline not working outside of Main (#685) @Pangoraw

Merged pull requests:

• Add R-square (coefficient of determination) (#679) (@rikhuijzer)
• Update ci.yml (#680) (@rikhuijzer)
• Prohibit use of target as kwarg in Pipeline(...) (#681) (@ablaom)
• Rsquared fix plus more docs (#684) (@ablaom)
• fix @pipeline when MLJ is not in Main (#685) (@Pangoraw)
• For a 0.18.26 release (#686) (@ablaom)

v0.18.25

MLJBase v0.18.25

Diff since v0.18.24

Closed issues:

• Improvements for unsupervised models that make probabilistic predictions (#656)
• Bug in display for one-dimensional range over functions (#667)
• In measures: When aggregating an empty collection of observation scores, issue warning instead of error (#673)

Merged pull requests:

• Macro free pipelines (PR onto for-0-point-19-release) (#664) (@ablaom)
• Allow nominal ranges consisting of different functions. (#666) (@davnn)
• Fix sequence_string bug (#668) (@ablaom)
• Implement conditional testing (#670) (@davnn)
• Remove empty collection catch in aggregation of measurements (#674) (@ablaom)
• Allow automatic access to training_losses in iterative models wrapped in a pipeline (#675) (@ablaom)
• For a 0.18.25 (#676) (@ablaom)