re-factoring: estimator registry (incl off-shelf estimators)

[fkiraly]

currently, estimators are loaded via load_default_estimators that has an implicit in-code registry of available estimators and pipelines with default parameter settings and composition. It would be nicer if there were a queriable registry, i.e., an API for queries that are not hard-coded as currently via the load_default_estimators function and its internal case distinctions.

[ViktorKaz]

The current functionality allows queries. The user can search by any of the attributes of the estimators. For example:

from mlaut.estimators import default_estimators

#returns all available estimators
default_estimators.instantiate_default_estimators(['all'])

#searches by task
default_estimators.instantiate_default_estimators(['Classification'])

#by family
default_estimators.instantiate_default_estimators(['Ensemble methods'])

#by name
default_estimators.instantiate_default_estimators(['RandomForest','Bagging'])

The properties as specified as a dictionary as follows:

  properties = {'estimator_family':['Ensemble methods'], 
            'tasks':['Classification'], 
            'name':'RandomForestClassifier'}

The query looks for matches under any of these three dictionary keys.

Did you have something else in mind?