searchgrid documentation

Helps building parameter grids for scikit-learn grid search.

Specifying a parameter grid for GridSearchCV in Scikit-Learn can be annoying, particularly when:

• you change your code to wrap some estimator in, say, a Pipeline and then need to prefix all the parameters in the grid using lots of __s
• you are searching over multiple grids (i.e. your param_grid is a list) and you want to make a change to all of those grids

searchgrid allows you to define (and change) the grid together with the esimator, reducing effort and sometimes code. It stores the parameters you want to search on each particular estimator object. This makes it much more straightforward to specify complex parameter grids, and means you don’t need to update your grid when you change the structure of your composite estimator.

It provides two main functions:

• searchgrid.set_grid() is used to specify the parameter values to be searched for an estimator or GP kernel.
• searchgrid.make_grid_search() is used to construct the GridSearchCV object using the parameter space the estimator is annotated with.

Quick Start

If scikit-learn is installed, then, in a terminal:

pip install searchgrid

and use in Python:

from search_grid import set_grid, make_grid_search
estimator = set_grid(MyEstimator(), param=[value1, value2, value3])
search = make_grid_search(estimator, cv=..., scoring=...)
search.fit(X, y)

Or search for the best among multiple distinct estimators/pipelines:

search = make_grid_search([estimator1, estimator2], cv=..., scoring=...)
search.fit(X, y)

Motivating examples

Let’s look over some of the messy change cases. We’ll get some imports out of the way.:

>>> from sklearn.pipeline import Pipeline
>>> from sklearn.linear_model import LogisticRegression
>>> from sklearn.feature_selection import SelectKBest
>>> from sklearn.decomposition import PCA
>>> from searchgrid import set_grid, make_grid_search
>>> from sklearn.model_selection import GridSearchCV

Wrapping an estimator in a pipeline.

You had code which searched over parameters for a classifier. Now you want to search for that classifier in a Pipeline. With plain old scikit-learn, you have to insert __s and change:

>>> gs = GridSearchCV(LogisticRegression(), {'C': [.1, 1, 10]})

to:

>>> gs = GridSearchCV(Pipeline([('reduce', SelectKBest()),
...                             ('clf', LogisticRegression())]),
...                   {'clf__C': [.1, 1, 10]})

With searchgrid we only have to wrap our classifier in a Pipeline, and do not have to change the parameter grid, adding the clf__ prefix. From:

>>> lr = set_grid(LogisticRegression(), C=[.1, 1, 10])
>>> gs = make_grid_search(lr)

to:

>>> lr = set_grid(LogisticRegression(), C=[.1, 1, 10])
>>> gs = make_grid_search(Pipeline([('reduce', SelectKBest()),
...                                 ('clf', lr)]))

You want to change the estimator being searched in a pipeline.

With scikit-learn, to use PCA instead of SelectKBest, you change:

>>> pipe = Pipeline([('reduce', SelectKBest()),
...                  ('clf', LogisticRegression())])
>>> gs = GridSearchCV(pipe,
...                   {'reduce__k': [5, 10, 20],
...                    'clf__C': [.1, 1, 10]})

to:

>>> pipe = Pipeline([('reduce', PCA()),
...                  ('clf', LogisticRegression())])
>>> gs = GridSearchCV(pipe,
...                   {'reduce__n_components': [5, 10, 20],
...                    'clf__C': [.1, 1, 10]})

Note that reduce__k became reduce__n_components.

With searchgrid it’s easier because you change the estimator and the parameters in the same place:

>>> reduce = set_grid(SelectKBest(), k=[5, 10, 20])
>>> lr = set_grid(LogisticRegression(), C=[.1, 1, 10])
>>> pipe = Pipeline([('reduce', reduce),
...                  ('clf', lr)])
>>> gs = make_grid_search(pipe)

becomes:

>>> reduce = set_grid(PCA(), n_components=[5, 10, 20])
>>> lr = set_grid(LogisticRegression(), C=[.1, 1, 10])
>>> pipe = Pipeline([('reduce', reduce),
...                  ('clf', lr)])
>>> gs = make_grid_search(pipe)

Searching over multiple grids.

You want to take the code from the previous example, but instead search over feature selection and PCA reduction in the same search.

Without searchgrid:

>>> pipe = Pipeline([('reduce', None),
...                  ('clf', LogisticRegression())])
>>> gs = GridSearchCV(pipe, [{'reduce': [SelectKBest()],
...                           'reduce__k': [5, 10, 20],
...                           'clf__C': [.1, 1, 10]},
...                          {'reduce': [PCA()],
...                           'reduce__n_components': [5, 10, 20],
...                           'clf__C': [.1, 1, 10]}])

With searchgrid:

>>> kbest = set_grid(SelectKBest(), k=[5, 10, 20])
>>> pca = set_grid(PCA(), n_components=[5, 10, 20])
>>> lr = set_grid(LogisticRegression(), C=[.1, 1, 10])
>>> pipe = set_grid(Pipeline([('reduce', None),
...                           ('clf', lr)]),
...                 reduce=[kbest, pca])
>>> gs = make_grid_search(pipe)

API Reference

searchgrid.build_param_grid(estimator)

Determine the parameter grid annotated on the estimator

Parameters:

estimator : scikit-learn compatible estimator

Should have been annotated using set_grid()

Notes

Most often, it is unnecessary for this to be used directly, and make_grid_search() should be used instead.

searchgrid.make_grid_search(estimator, **kwargs)

Construct a GridSearchCV with the given estimator and its set grid

Parameters:

estimator : (list of) estimator

When a list, the estimators are searched over.

kwargs

Other parameters to the sklearn.model_selection.GridSearchCV constructor.

searchgrid.set_grid(estimator, **grid)

Set the grid to search for the specified estimator

Overwrites any previously set grid.

Parameters:

grid : dict (str -> list of values)

Keyword arguments define the values to be searched for each specified parameter.

Returns:

estimator

Useful for chaining

Changelog

v0.2

• Fixed a bug where the grid of the default estimator in a Pipeline step was attributed to alternatives for that step. #10.