8000 [MRG + 1] ENH Do not materialise CV splits when unnecessary (#7941) · scikit-learn/scikit-learn@3dcb873 · GitHub
[go: up one dir, main page]
Skip to content

Commit 3dcb873

Browse files
raghavrvjnothman
raghavrv
authored and
jnothman
committed
[MRG + 1] ENH Do not materialise CV splits when unnecessary (#7941)
* ENH Parallelize by candidates first then by splits. * ENH do not materialize a cv iterator to avoid memory blow ups.
1 parent f9c6344 commit 3dcb873

File tree

2 files changed

+46
-67
lines changed

2 files changed

+46
-67
lines changed

‎sklearn/model_selection/_search.py

Lines changed: 40 additions & 59 deletions
Original file line numberDiff line numberDiff line change
@@ -12,7 +12,7 @@
1212
# License: BSD 3 clause
1313

1414
from abc import ABCMeta, abstractmethod
15-
from collections import Mapping, namedtuple, Sized, defaultdict, Sequence
15+
from collections import Mapping, namedtuple, defaultdict, Sequence
1616
from functools import partial, reduce
1717
from itertools import product
1818
import operator
@@ -532,25 +532,41 @@ def inverse_transform(self, Xt):
532532
self._check_is_fitted('inverse_transform')
533533
return self.best_estimator_.transform(Xt)
534534

535-
def _fit(self, X, y, groups, parameter_iterable):
536-
"""Actual fitting, performing the search over parameters."""
535+
def fit(self, X, y=None, groups=None):
536+
"""Run fit with all sets of parameters.
537+
538+
Parameters
539+
----------
540+
541+
X : array-like, shape = [n_samples, n_features]
542+
Training vector, where n_samples is the number of samples and
543+
n_features is the number of features.
537544
545+
y : array-like, shape = [n_samples] or [n_samples, n_output], optional
546+
Target relative to X for classification or regression;
547+
None for unsupervised learning.
548+
549+
groups : array-like, with shape (n_samples,), optional
550+
Group labels for the samples used while splitting the dataset into
551+
train/test set.
552+
"""
538553
estimator = self.estimator
539554
cv = check_cv(self.cv, y, classifier=is_classifier(estimator))
540555
self.scorer_ = check_scoring(self.estimator, scoring=self.scoring)
541556

542557
X, y, groups = indexable(X, y, groups)
543558
n_splits = cv.get_n_splits(X, y, groups)
544-
if self.verbose > 0 and isinstance(parameter_iterable, Sized):
545-
n_candidates = len(parameter_iterable)
559+
# Regenerate parameter iterable for each fit
560+
candidate_params = list(self._get_param_iterator())
561+
n_candidates = len(candidate_params)
562+
if self.verbose > 0:
546563
print("Fitting {0} folds for each of {1} candidates, totalling"
547564
" {2} fits".format(n_splits, n_candidates,
548565
n_candidates * n_splits))
549566

550567
base_estimator = clone(self.estimator)
551568
pre_dispatch = self.pre_dispatch
552569

553-
cv_iter = list(cv.split(X, y, groups))
554570
out = Parallel(
555571
n_jobs=self.n_jobs, verbose=self.verbose,
556572
pre_dispatch=pre_dispatch
@@ -559,28 +575,25 @@ def _fit(self, X, y, groups, parameter_iterable):
559575
fit_params=self.fit_params,
560576
return_train_score=self.return_train_score,
561577
return_n_test_samples=True,
562-
return_times=True, return_parameters=True,
578+
return_times=True, return_parameters=False,
563579
error_score=self.error_score)
564-
for parameters in parameter_iterable
565-
for train, test in cv_iter)
580+
for train, test in cv.split(X, y, groups)
581+
for parameters in candidate_params)
566582

567583
# if one choose to see train score, "out" will contain train score info
568584
if self.return_train_score:
569-
(train_scores, test_scores, test_sample_counts,
570-
fit_time, score_time, parameters) = zip(*out)
585+
(train_scores, test_scores, test_sample_counts, fit_time,
586+
score_time) = zip(*out)
571587
else:
572-
(test_scores, test_sample_counts,
573-
fit_time, score_time, parameters) = zip(*out)
574-
575-
candidate_params = parameters[::n_splits]
576-
n_candidates = len(candidate_params)
588+
(test_scores, test_sample_counts, fit_time, score_time) = zip(*out)
577589

578590
results = dict()
579591

580592
def _store(key_name, array, weights=None, splits=False, rank=False):
581593
"""A small helper to store the scores/times to the cv_results_"""
582-
array = np.array(array, dtype=np.float64).reshape(n_candidates,
583-
n_splits)
594+
# When iterated first by splits, then by parameters
595+
array = np.array(array, dtype=np.float64).reshape(n_splits,
596+
n_candidates).T
584597
if splits:
585598
for split_i in range(n_splits):
586599
results["split%d_%s"
@@ -600,7 +613,7 @@ def _store(key_name, array, weights=None, splits=False, rank=False):
600613

601614
# Computed the (weighted) mean and std for test scores alone
602615
# NOTE test_sample counts (weights) remain the same for all candidates
603-
test_sample_counts = np.array(test_sample_counts[:n_splits],
616+
test_sample_counts = np.array(test_sample_counts[::n_candidates],
604617
dtype=np.int)
605618

606619
_store('test_score', test_scores, splits=True, rank=True,
@@ -924,25 +937,9 @@ def __init__(self, estimator, param_grid, scoring=None, fit_params=None,
924937
self.param_grid = param_grid
925938
_check_param_grid(param_grid)
926939

927-
def fit(self, X, y=None, groups=None):
928-
"""Run fit with all sets of parameters.
929-
930-
Parameters
931-
----------
932-
933-
X : array-like, shape = [n_samples, n_features]
934-
Training vector, where n_samples is the number of samples and
935-
n_features is the number of features.
936-
937-
y : array-like, shape = [n_samples] or [n_samples, n_output], optional
938-
Target relative to X for classification or regression;
939-
None for unsupervised learning.
940-
941-
groups : array-like, with shape (n_samples,), optional
942-
Group labels for the samples used while splitting the dataset into
943-
train/test set.
944-
"""
945-
return self._fit(X, y, groups, ParameterGrid(self.param_grid))
940+
def _get_param_iterator(self):
941+
"""Return ParameterGrid instance for the given param_grid"""
942+
return ParameterGrid(self.param_grid)
946943

947944

948945
class RandomizedSearchCV(BaseSearchCV):
@@ -1167,24 +1164,8 @@ def __init__(self, estimator, param_distributions, n_iter=10, scoring=None,
11671164
pre_dispatch=pre_dispatch, error_score=error_score,
11681165
return_train_score=return_train_score)
11691166

1170-
def fit(self, X, y=None, groups=None):
1171-
"""Run fit on the estimator with randomly drawn parameters.
1172-
1173-
Parameters
1174-
----------
1175-
X : array-like, shape = [n_samples, n_features]
1176-
Training vector, where n_samples in the number of samples and
1177-
n_features is the number of features.
1178-
1179-
y : array-like, shape = [n_samples] or [n_samples, n_output], optional
1180-
Target relative to X for classification or regression;
1181-
None for unsupervised learning.
1182-
1183-
groups : array-like, with shape (n_samples,), optional
1184-
Group labels for the samples used while splitting the dataset into
1185-
train/test set.
1186-
"""
1187-
sampled_params = ParameterSampler(self.param_distributions,
1188-
self.n_iter,
1189-
random_state=self.random_state)
1190-
return self._fit(X, y, groups, sampled_params)
1167+
def _get_param_iterator(self):
1168+
"""Return ParameterSampler instance for the given distributions"""
1169+
return ParameterSampler(
1170+
self.param_distributions, self.n_iter,
1171+
random_state=self.random_state)

‎sklearn/model_selection/_validation.py

Lines changed: 6 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -128,7 +128,6 @@ def cross_val_score(estimator, X, y=None, groups=None, scoring=None, cv=None,
128128
X, y, groups = indexable(X, y, groups)
129129

130130
cv = check_cv(cv, y, classifier=is_classifier(estimator))
131-
cv_iter = list(cv.split(X, y, groups))
132131
scorer = check_scoring(estimator, scoring=scoring)
133132
# We clone the estimator to make sure that all the folds are
134133
# independent, and that it is pickle-able.
@@ -137,7 +136,7 @@ def cross_val_score(estimator, X, y=None, groups=None, scoring=None, cv=None,
137136
scores = parallel(delayed(_fit_and_score)(clone(estimator), X, y, scorer,
138137
train, test, verbose, None,
139138
fit_params)
140-
for train, test in cv_iter)
139+
for train, test in cv.split(X, y, groups))
141140
return np.array(scores)[:, 0]
142141

143142

@@ -385,7 +384,6 @@ def cross_val_predict(estimator, X, y=None, groups=None, cv=None, n_jobs=1,
385384
X, y, groups = indexable(X, y, groups)
386385

387386
cv = check_cv(cv, y, classifier=is_classifier(estimator))
388-
cv_iter = list(cv.split(X, y, groups))
389387

390388
# Ensure the estimator has implemented the passed decision function
391389
if not callable(getattr(estimator, method)):
@@ -398,7 +396,7 @@ def cross_val_predict(estimator, X, y=None, groups=None, cv=None, n_jobs=1,
398396
pre_dispatch=pre_dispatch)
399397
prediction_blocks = parallel(delayed(_fit_and_predict)(
400398
clone(estimator), X, y, train, test, verbose, fit_params, method)
401-
for train, test in cv_iter)
399+
for train, test in cv.split(X, y, groups))
402400

403401
# Concatenate the predictions
404402
predictions = [pred_block_i for pred_block_i, _ in prediction_blocks]
@@ -752,8 +750,9 @@ def learning_curve(estimator, X, y, groups=None,
752750
X, y, groups = indexable(X, y, groups)
753751

754752
cv = check_cv(cv, y, classifier=is_classifier(estimator))
755-
# Make a list since we will be iterating multiple times over the folds
753+
# Store it as list as we will be iterating over the list multiple times
756754
cv_iter = list(cv.split(X, y, groups))
755+
757756
scorer = check_scoring(estimator, scoring=scoring)
758757

759758
n_max_training_samples = len(cv_iter[0][0])
@@ -961,16 +960,15 @@ def validation_curve(estimator, X, y, param_name, param_range, groups=None,
961960
X, y, groups = indexable(X, y, groups)
962961

963962
cv = check_cv(cv, y, classifier=is_classifier(estimator))
964-
cv_iter = list(cv.split(X, y, groups))
965-
966963
scorer = check_scoring(estimator, scoring=scoring)
967964

968965
parallel = Parallel(n_jobs=n_jobs, pre_dispatch=pre_dispatch,
969966
verbose=verbose)
970967
out = parallel(delayed(_fit_and_score)(
971968
estimator, X, y, scorer, train, test, verbose,
972969
parameters={param_name: v}, fit_params=None, return_train_score=True)
973-
for train, test in cv_iter for v in param_range)
970+
# NOTE do not change order of iteration to allow one time cv splitters
971+
for train, test in cv.split(X, y, groups) for v in param_range)
974972

975973
out = np.asarray(out)
976974
n_params = len(param_range)

0 commit comments

Comments
 (0)
0