goerch
commented
4 years ago Monkey patching StabilitySelection with
from sklearn.base import BaseEstimator, TransformerMixin, clone
from sklearn.utils import check_array, check_random_state, check_X_y, safe_mask
from sklearn.externals.joblib import Parallel, delayed
from sklearn.feature_selection import SelectFromModel
def _return_estimator_from_pipeline(pipeline):
"""Returns the final estimator in a Pipeline, or the estimator
if it is not"""
if isinstance(pipeline, Pipeline):
return pipeline._final_estimator
else:
return pipeline
def _bootstrap_generator(n_bootstrap_iterations, bootstrap_func, y,
n_subsamples, random_state=None):
for _ in range(n_bootstrap_iterations):
subsample = bootstrap_func(y, n_subsamples, random_state)
if isinstance(subsample, tuple):
for item in subsample:
yield item
else:
yield subsample
def _fit_bootstrap_sample(base_estimator, X, y, lambda_name, lambda_value,
threshold=None):
"""
Fits base_estimator on a bootstrap sample of the original data,
and returns a mas of the variables that are selected by the fitted model.
Parameters
----------
base_estimator : Estimator
Estimator to be fitted on the data
X : {array-like, sparse matrix}, shape = [n_samples, n_features]
The training input samples.
y : array-like, shape = [n_samples]
The target values.
lambda_name : str
Name of the penalization parameter of base_estimator
lambda_value : float
Value of the penalization parameter
threshold : string, float, optional default None
The threshold value to use for feature selection. Features whose
importance is greater or equal are kept while the others are
discarded. If "median" (resp. "mean"), then the ``threshold`` value is
the median (resp. the mean) of the feature importances. A scaling
factor (e.g., "1.25*mean") may also be used. If None and if the
estimator has a parameter penalty set to l1, either explicitly
or implicitly (e.g, Lasso), the threshold used is 1e-5.
Otherwise, "mean" is used by default.
Returns
-------
selected_variables : array-like, shape = [n_features]
Boolean mask of selected variables.
"""
base_estimator.set_params(**{lambda_name: lambda_value})
base_estimator.fit(X, y)
# TODO: Reconsider if we really want to use SelectFromModel here or not
selector_model = _return_estimator_from_pipeline(base_estimator)
variable_selector = SelectFromModel(estimator=selector_model,
threshold=threshold,
prefit=True)
return variable_selector.get_support()
def fit(self, X, y):
"""Fit the stability selection model on the given data.
Parameters
----------
X : {array-like, sparse matrix}, shape = [n_samples, n_features]
The training input samples.
y : array-like, shape = [n_samples]
The target values.
"""
self._validate_input()
X, y = check_X_y(X, y, accept_sparse='csr')
n_samples, n_variables = X.shape
n_subsamples = np.floor(self.sample_fraction * n_samples).astype(int)
n_lambdas = self.lambda_grid.shape[0]
base_estimator = clone(self.base_estimator)
random_state = check_random_state(self.random_state)
stability_scores = np.zeros((n_variables, n_lambdas))
for idx, lambda_value in enumerate(self.lambda_grid):
if self.verbose > 0:
print("Fitting estimator for lambda = %.5f (%d / %d) on %d bootstrap samples" %
(lambda_value, idx + 1, n_lambdas, self.n_bootstrap_iterations))
bootstrap_samples = _bootstrap_generator(self.n_bootstrap_iterations,
self.bootstrap_func, y,
n_subsamples, random_state=random_state)
selected_variables = Parallel(
n_jobs=self.n_jobs, verbose=self.verbose,
pre_dispatch=self.pre_dispatch
)(delayed(_fit_bootstrap_sample)(clone(base_estimator),
X=X[safe_mask(X, subsample), :],
y=y[subsample],
lambda_name=self.lambda_name,
lambda_value=lambda_value,
threshold=self.bootstrap_threshold)
for subsample in bootstrap_samples)
for i in range(0, self.n_bootstrap_iterations):
print("Sample ", i, " selected ", selected_variables[i].shape)
stability_scores[:, idx] = np.vstack(selected_variables).mean(axis=0)
self.stability_scores_ = stability_scores
return self
StabilitySelection.fit = fitthe debug output for the notebook is
Sample 0 selected (30969,)
Sample 1 selected (31011,)
Sample 2 selected (30959,)which indicates _fit_bootstrap_sample doesn't work with the feature selection of the notebook pipeline.
thuijskens
The following cell from my notebook
gives the output
and then the traceback
X_train is a scipy crs_matrix. Is this expected behaviour?