v0.20.0
MLJBase v0.20.0
- Relax and simplify scitype checks when constructing machines. The existing
fit_data_scitype
model trait encodes all allowedfit
"scitype" signatures, and scitype checks now only consider this trait. In particular, an appropriately implemented transformer can now be passed a training target without tripping the type checker. (#699, #732) @pazzo83 @ablaom - (enhancement, breaking) Redesign the serialization API to: (i) Allow use of arbitrary serialization packages for core serialization; (ii) Ensure serialization plays nicely with model composition and meta-algorithms like tuning; (iii) Ensure all traces of training data are absent in serialised models (not previously true for all composite models or if
cache=true
in machine constructor). Models with non-persistent learned parameters (fitresult
) implement a modified model API that is documented here. The new user workflow will shortly appear in the MLJ manual under "Machines". (JuliaAI/MLJSerialization.jl#15, #733, JuliaAI/MLJSerialization.jl#16) @olivierlabayle
Closed issues:
- Relax any checks that block transformers needing to see target in training. (#699)
Merged pull requests:
- Serialization (#733) (@olivierlabayle)
- Add suggestion to
err_incompatible_prediction_types
message (#748) (@ablaom) - Add a brief reminder that the fields of the
PerformanceEvaluation
struct are part of the public API (#749) (@DilumAluthge) - For a 0.20 release (#751) (@ablaom)
- Fix a problem with confmat and CategoricalValue eltype (#752) (@ablaom)
- For a 0.20 release (#754) (@ablaom)
- Bump version. (#755) (@ablaom)