ablaom/koalalow.jl

A module for exposing the low-level interface of the Koala machine learning environment

KoalaLow

For exposing the low-level interface of the Koala machine learning environment.

Koala's high-level interface operates primarily on machines. Presently, the Koala objects which can be wrapped in machines are supervised models and transformers. Koala's low-level interface amounts to definitions of the following methods, for each supervised model type SMT <: Koala.SupervisedModel, and transformer type TT <: Koala.Transformer:

• default_transformer_X(model::SMT)

• default_transformer_y(model::SMT)

• clean!(model::SMT) (optional)

• setup(model::SMT, Xt, yt, scheme_X, parallel, verbosity)

• fit(model::SMT, cache, add, parallel, verbosity; args...)

• predict(model::SMT, predictor, Xt, parallel, verbosity)

• fit(transformer::TT, X, parallel, verbosity; args...)

• transform(transformer::TT, scheme, X)

• inverse_transform(transformer::TT, scheme, Xt) (optional)

The first two methods on the list have fall-backs defined in Koala, but each of the others require explicit implementations, unless marked "optional".

It is our intention that the low-level implementation of most transformers appear in KoalaTransforms (a few are in Koala itself but are re-exported by KoalaTransforms) while each supervised model (or family of similar such models) gets their own dedicated module (e.g., KoalaTrees).

The code,

using Koala

brings default_transformer_X, default_transformer_y, clean! and predict into scope (the last because it is overloaded in the high-level interface). To bring any other low-level method into scope, one can import it explicitly from Koala; to bring all of them into scope, use

using KoalaLow

For more on the low-level interface, see the template for implementing new supervised learning algorithms at SupervisedModelTemplate.jl, and the template for new transformers at TransformerTemplate.jl.