Skip to content

Commit 7097af0

Browse files
amuelleramueller
committed
PSL and CCA convergence check and input validation fixes
1 parent 28546c8 commit 7097af0

File tree

3 files changed

+82
-42
lines changed

3 files changed

+82
-42
lines changed

‎sklearn/cross_decomposition/pls_.py

Lines changed: 37 additions & 32 deletions
Original file line numberDiff line numberDiff line change
@@ -235,17 +235,18 @@ def fit(self, X, Y):
235235

236236
# copy since this will contains the residuals (deflated) matrices
237237
check_consistent_length(X, Y)
238-
X = check_array(X, dtype=np.float, copy=self.copy)
239-
Y = check_array(Y, dtype=np.float, copy=self.copy, ensure_2d=False)
238+
X = check_array(X, dtype=np.float64, copy=self.copy)
239+
Y = check_array(Y, dtype=np.float64, copy=self.copy, ensure_2d=False)
240240
if Y.ndim == 1:
241-
Y = Y[:, None]
241+
Y = Y.reshape(-1, 1)
242242

243243
n = X.shape[0]
244244
p = X.shape[1]
245245
q = Y.shape[1]
246246

247247
if self.n_components < 1 or self.n_components > p:
248-
raise ValueError('invalid number of components')
248+
raise ValueError('Invalid number of components: %d' %
249+
self.n_components)
249250
if self.algorithm not in ("svd", "nipals"):
250251
raise ValueError("Got algorithm %s when only 'svd' "
251252
"and 'nipals' are known" % self.algorithm)
@@ -271,6 +272,10 @@ def fit(self, X, Y):
271272

272273
# NIPALS algo: outer loop, over components
273274
for k in range(self.n_components):
275+
if np.all(np.dot(Yk.T, Yk) < np.finfo(np.double).eps):
276+
# Yk constant
277+
warnings.warn('Y residual constant at iteration %s' % k)
278+
break
274279
#1) weights estimation (inner loop)
275280
# -----------------------------------
276281
if self.algorithm == "nipals":
@@ -291,6 +296,7 @@ def fit(self, X, Y):
291296
# test for null variance
292297
if np.dot(x_scores.T, x_scores) < np.finfo(np.double).eps:
293298
warnings.warn('X scores are null at iteration %s' % k)
299+
break
294300
#2) Deflation (in place)
295301
# ----------------------
296302
# Possible memory footprint reduction may done here: in order to
@@ -335,6 +341,7 @@ def fit(self, X, Y):
335341
self.y_rotations_ = np.ones(1)
336342

337343
if True or self.deflation_mode == "regression":
344+
# FIXME what's with the if?
338345
# Estimate regression coefficient
339346
# Regress Y on T
340347
# Y = TQ' + Err,
@@ -367,23 +374,19 @@ def transform(self, X, Y=None, copy=True):
367374
x_scores if Y is not given, (x_scores, y_scores) otherwise.
368375
"""
369376
check_is_fitted(self, 'x_mean_')
377+
X = check_array(X, copy=copy)
370378
# Normalize
371-
if copy:
372-
Xc = (np.asarray(X) - self.x_mean_) / self.x_std_
373-
if Y is not None:
374-
Yc = (np.asarray(Y) - self.y_mean_) / self.y_std_
375-
else:
376-
X = np.asarray(X)
377-
Xc -= self.x_mean_
378-
Xc /= self.x_std_
379-
if Y is not None:
380-
Y = np.asarray(Y)
381-
Yc -= self.y_mean_
382-
Yc /= self.y_std_
379+
X -= self.x_mean_
380+
X /= self.x_std_
383381
# Apply rotation
384-
x_scores = np.dot(Xc, self.x_rotations_)
382+
x_scores = np.dot(X, self.x_rotations_)
385383
if Y is not None:
386-
y_scores = np.dot(Yc, self.y_rotations_)
384+
Y = check_array(Y, ensure_2d=False, copy=copy)
385+
if Y.ndim == 1:
386+
Y = Y.reshape(-1, 1)
387+
Y -= self.y_mean_
388+
Y /= self.y_std_
389+
y_scores = np.dot(Y, self.y_rotations_)
387390
return x_scores, y_scores
388391

389392
return x_scores
@@ -406,14 +409,11 @@ def predict(self, X, copy=True):
406409
be an issue in high dimensional space.
407410
"""
408411
check_is_fitted(self, 'x_mean_')
412+
X = check_array(X, copy=copy)
409413
# Normalize
410-
if copy:
411-
Xc = (np.asarray(X) - self.x_mean_)
412-
else:
413-
X = np.asarray(X)
414-
Xc -= self.x_mean_
415-
Xc /= self.x_std_
416-
Ypred = np.dot(Xc, self.coef_)
414+
X -= self.x_mean_
415+
X /= self.x_std_
416+
Ypred = np.dot(X, self.coef_)
417417
return Ypred + self.y_mean_
418418

419419
def fit_transform(self, X, y=None, **fit_params):
@@ -724,13 +724,15 @@ def __init__(self, n_components=2, scale=True, copy=True):
724724
def fit(self, X, Y):
725725
# copy since this will contains the centered data
726726
check_consistent_length(X, Y)
727-
X = check_array(X, dtype=np.float, copy=self.copy)
728-
Y = check_array(Y, dtype=np.float, copy=self.copy)
729-
730-
p = X.shape[1]
727+
X = check_array(X, dtype=np.float64, copy=self.copy)
728+
Y = check_array(Y, dtype=np.float64, copy=self.copy, ensure_2d=False)
729+
if Y.ndim == 1:
730+
Y = Y.reshape(-1, 1)
731731

732-
if self.n_components < 1 or self.n_components > p:
733-
raise ValueError('invalid number of components')
732+
if self.n_components > max(Y.shape[1], X.shape[1]):
733+
raise ValueError("Invalid number of components n_components=%d with "
734+
"X of shape %s and Y of shape %s."
735+
% (self.n_components, str(X.shape), str(Y.shape)))
734736

735737
# Scale (in place)
736738
X, Y, self.x_mean_, self.y_mean_, self.x_std_, self.y_std_ =\
@@ -742,7 +744,7 @@ def fit(self, X, Y):
742744
# components is smaller than rank(X) - 1. Hence, if we want to extract
743745
# all the components (C.shape[1]), we have to use another one. Else,
744746
# let's use arpacks to compute only the interesting components.
745-
if self.n_components == C.shape[1]:
747+
if self.n_components >= np.min(C.shape):
746748
U, s, V = linalg.svd(C, full_matrices=False)
747749
else:
748750
U, s, V = arpack.svds(C, k=self.n_components)
@@ -756,9 +758,12 @@ def fit(self, X, Y):
756758
def transform(self, X, Y=None):
757759
"""Apply the dimension reduction learned on the train data."""
758760
check_is_fitted(self, 'x_mean_')
761+
X = check_array(X, dtype=np.float64)
759762
Xr = (X - self.x_mean_) / self.x_std_
760763
x_scores = np.dot(Xr, self.x_weights_)
761764
if Y is not None:
765+
if Y.ndim == 1:
766+
Y = Y.reshape(-1, 1)
762767
Yr = (Y - self.y_mean_) / self.y_std_
763768
y_scores = np.dot(Yr, self.y_weights_)
764769
return x_scores, y_scores

‎sklearn/cross_decomposition/tests/test_pls.py

Lines changed: 36 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,6 @@
11
import numpy as np
2-
from sklearn.utils.testing import assert_array_almost_equal
2+
from sklearn.utils.testing import (assert_array_almost_equal,
3+
assert_array_equal, assert_true, assert_raise_message)
34
from sklearn.datasets import load_linnerud
45
from sklearn.cross_decomposition import pls_
56
from nose.tools import assert_equal
@@ -248,6 +249,30 @@ def test_univariate_pls_regression():
248249
assert_array_almost_equal(model1, model2)
249250

250251

252+
def test_predict_transform_copy():
253+
# check that the "copy" keyword works
254+
d = load_linnerud()
255+
X = d.data
256+
Y = d.target
257+
clf = pls_.PLSCanonical()
258+
X_copy = X.copy()
259+
Y_copy = Y.copy()
260+
clf.fit(X, Y)
261+
# check that results are identical with copy
262+
assert_array_almost_equal(clf.predict(X), clf.predict(X.copy(), copy=False))
263+
assert_array_almost_equal(clf.transform(X), clf.transform(X.copy(), copy=False))
264+
265+
# check also if passing Y
266+
assert_array_almost_equal(clf.transform(X, Y),
267+
clf.transform(X.copy(), Y.copy(), copy=False))
268+
# check that copy doesn't destroy
269+
# we do want to check exact equality here
270+
assert_array_equal(X_copy, X)
271+
assert_array_equal(Y_copy, Y)
272+
# also check that mean wasn't zero before (to make sure we didn't touch it)
273+
assert_true(np.all(X.mean(axis=0) != 0))
274+
275+
251276
def test_scale():
252277
d = load_linnerud()
253278
X = d.data
@@ -260,3 +285,13 @@ def test_scale():
260285
pls_.PLSSVD()]:
261286
clf.set_params(scale=True)
262287
clf.fit(X, Y)
288+
289+
290+
def test_pls_errors():
291+
d = load_linnerud()
292+
X = d.data
293+
Y = d.target
294+
for clf in [pls_.PLSCanonical(), pls_.PLSRegression(),
295+
pls_.PLSSVD()]:
296+
clf.n_components = 4
297+
assert_raise_message(ValueError, "Invalid number of components", clf.fit, X, Y)

‎sklearn/tests/test_common.py

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -95,25 +95,25 @@ def test_non_meta_estimators():
9595
continue
9696
if name.endswith("HMM") or name.startswith("_"):
9797
continue
98-
if name not in CROSS_DECOMPOSITION:
99-
yield check_estimators_dtypes, name, Estimator
100-
yield check_fit_score_takes_y, name, Estimator
101-
yield check_dtype_object, name, Estimator
98+
yield check_estimators_dtypes, name, Estimator
99+
yield check_fit_score_takes_y, name, Estimator
100+
yield check_dtype_object, name, Estimator
102101

103-
# Check that all estimator yield informative messages when
104-
# trained on empty datasets
105-
yield check_estimators_empty_data_messages, name, Estimator
102+
# Check that all estimator yield informative messages when
103+
# trained on empty datasets
104+
yield check_estimators_empty_data_messages, name, Estimator
106105

107106
if name not in CROSS_DECOMPOSITION + ['SpectralEmbedding']:
108107
# SpectralEmbedding is non-deterministic,
109108
# see issue #4236
109+
# cross-decomposition's "transform" returns X and Y
110110
yield check_pipeline_consistency, name, Estimator
111111

112-
if name not in CROSS_DECOMPOSITION + ['Imputer']:
112+
if name not in ['Imputer']:
113113
# Test that all estimators check their input for NaN's and infs
114114
yield check_estimators_nan_inf, name, Estimator
115115

116-
if name not in CROSS_DECOMPOSITION + ['GaussianProcess']:
116+
if name not in ['GaussianProcess']:
117117
# FIXME!
118118
# in particular GaussianProcess!
119119
yield check_estimators_overwrite_params, name, Estimator

0 commit comments

Comments
 (0)