File: test_self_training.py

package info (click to toggle)
scikit-learn 1.4.2%2Bdfsg-8
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid, trixie
  • size: 25,036 kB
  • sloc: python: 201,105; cpp: 5,790; ansic: 854; makefile: 304; sh: 56; javascript: 20
file content (345 lines) | stat: -rw-r--r-- 12,543 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
from math import ceil

import numpy as np
import pytest
from numpy.testing import assert_array_equal

from sklearn.datasets import load_iris, make_blobs
from sklearn.ensemble import StackingClassifier
from sklearn.exceptions import NotFittedError
from sklearn.metrics import accuracy_score
from sklearn.model_selection import train_test_split
from sklearn.neighbors import KNeighborsClassifier
from sklearn.semi_supervised import SelfTrainingClassifier
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier

# Author: Oliver Rausch <rauscho@ethz.ch>
# License: BSD 3 clause

# load the iris dataset and randomly permute it
iris = load_iris()
X_train, X_test, y_train, y_test = train_test_split(
    iris.data, iris.target, random_state=0
)

n_labeled_samples = 50

y_train_missing_labels = y_train.copy()
y_train_missing_labels[n_labeled_samples:] = -1
mapping = {0: "A", 1: "B", 2: "C", -1: "-1"}
y_train_missing_strings = np.vectorize(mapping.get)(y_train_missing_labels).astype(
    object
)
y_train_missing_strings[y_train_missing_labels == -1] = -1


def test_warns_k_best():
    st = SelfTrainingClassifier(KNeighborsClassifier(), criterion="k_best", k_best=1000)
    with pytest.warns(UserWarning, match="k_best is larger than"):
        st.fit(X_train, y_train_missing_labels)

    assert st.termination_condition_ == "all_labeled"


@pytest.mark.parametrize(
    "base_estimator",
    [KNeighborsClassifier(), SVC(gamma="scale", probability=True, random_state=0)],
)
@pytest.mark.parametrize("selection_crit", ["threshold", "k_best"])
def test_classification(base_estimator, selection_crit):
    # Check classification for various parameter settings.
    # Also assert that predictions for strings and numerical labels are equal.
    # Also test for multioutput classification
    threshold = 0.75
    max_iter = 10
    st = SelfTrainingClassifier(
        base_estimator, max_iter=max_iter, threshold=threshold, criterion=selection_crit
    )
    st.fit(X_train, y_train_missing_labels)
    pred = st.predict(X_test)
    proba = st.predict_proba(X_test)

    st_string = SelfTrainingClassifier(
        base_estimator, max_iter=max_iter, criterion=selection_crit, threshold=threshold
    )
    st_string.fit(X_train, y_train_missing_strings)
    pred_string = st_string.predict(X_test)
    proba_string = st_string.predict_proba(X_test)

    assert_array_equal(np.vectorize(mapping.get)(pred), pred_string)
    assert_array_equal(proba, proba_string)

    assert st.termination_condition_ == st_string.termination_condition_
    # Check consistency between labeled_iter, n_iter and max_iter
    labeled = y_train_missing_labels != -1
    # assert that labeled samples have labeled_iter = 0
    assert_array_equal(st.labeled_iter_ == 0, labeled)
    # assert that labeled samples do not change label during training
    assert_array_equal(y_train_missing_labels[labeled], st.transduction_[labeled])

    # assert that the max of the iterations is less than the total amount of
    # iterations
    assert np.max(st.labeled_iter_) <= st.n_iter_ <= max_iter
    assert np.max(st_string.labeled_iter_) <= st_string.n_iter_ <= max_iter

    # check shapes
    assert st.labeled_iter_.shape == st.transduction_.shape
    assert st_string.labeled_iter_.shape == st_string.transduction_.shape


def test_k_best():
    st = SelfTrainingClassifier(
        KNeighborsClassifier(n_neighbors=1),
        criterion="k_best",
        k_best=10,
        max_iter=None,
    )
    y_train_only_one_label = np.copy(y_train)
    y_train_only_one_label[1:] = -1
    n_samples = y_train.shape[0]

    n_expected_iter = ceil((n_samples - 1) / 10)
    st.fit(X_train, y_train_only_one_label)
    assert st.n_iter_ == n_expected_iter

    # Check labeled_iter_
    assert np.sum(st.labeled_iter_ == 0) == 1
    for i in range(1, n_expected_iter):
        assert np.sum(st.labeled_iter_ == i) == 10
    assert np.sum(st.labeled_iter_ == n_expected_iter) == (n_samples - 1) % 10
    assert st.termination_condition_ == "all_labeled"


def test_sanity_classification():
    base_estimator = SVC(gamma="scale", probability=True)
    base_estimator.fit(X_train[n_labeled_samples:], y_train[n_labeled_samples:])

    st = SelfTrainingClassifier(base_estimator)
    st.fit(X_train, y_train_missing_labels)

    pred1, pred2 = base_estimator.predict(X_test), st.predict(X_test)
    assert not np.array_equal(pred1, pred2)
    score_supervised = accuracy_score(base_estimator.predict(X_test), y_test)
    score_self_training = accuracy_score(st.predict(X_test), y_test)

    assert score_self_training > score_supervised


def test_none_iter():
    # Check that the all samples were labeled after a 'reasonable' number of
    # iterations.
    st = SelfTrainingClassifier(KNeighborsClassifier(), threshold=0.55, max_iter=None)
    st.fit(X_train, y_train_missing_labels)

    assert st.n_iter_ < 10
    assert st.termination_condition_ == "all_labeled"


@pytest.mark.parametrize(
    "base_estimator",
    [KNeighborsClassifier(), SVC(gamma="scale", probability=True, random_state=0)],
)
@pytest.mark.parametrize("y", [y_train_missing_labels, y_train_missing_strings])
def test_zero_iterations(base_estimator, y):
    # Check classification for zero iterations.
    # Fitting a SelfTrainingClassifier with zero iterations should give the
    # same results as fitting a supervised classifier.
    # This also asserts that string arrays work as expected.

    clf1 = SelfTrainingClassifier(base_estimator, max_iter=0)

    clf1.fit(X_train, y)

    clf2 = base_estimator.fit(X_train[:n_labeled_samples], y[:n_labeled_samples])

    assert_array_equal(clf1.predict(X_test), clf2.predict(X_test))
    assert clf1.termination_condition_ == "max_iter"


def test_prefitted_throws_error():
    # Test that passing a pre-fitted classifier and calling predict throws an
    # error
    knn = KNeighborsClassifier()
    knn.fit(X_train, y_train)
    st = SelfTrainingClassifier(knn)
    with pytest.raises(
        NotFittedError,
        match="This SelfTrainingClassifier instance is not fitted yet",
    ):
        st.predict(X_train)


@pytest.mark.parametrize("max_iter", range(1, 5))
def test_labeled_iter(max_iter):
    # Check that the amount of datapoints labeled in iteration 0 is equal to
    # the amount of labeled datapoints we passed.
    st = SelfTrainingClassifier(KNeighborsClassifier(), max_iter=max_iter)

    st.fit(X_train, y_train_missing_labels)
    amount_iter_0 = len(st.labeled_iter_[st.labeled_iter_ == 0])
    assert amount_iter_0 == n_labeled_samples
    # Check that the max of the iterations is less than the total amount of
    # iterations
    assert np.max(st.labeled_iter_) <= st.n_iter_ <= max_iter


def test_no_unlabeled():
    # Test that training on a fully labeled dataset produces the same results
    # as training the classifier by itself.
    knn = KNeighborsClassifier()
    knn.fit(X_train, y_train)
    st = SelfTrainingClassifier(knn)
    with pytest.warns(UserWarning, match="y contains no unlabeled samples"):
        st.fit(X_train, y_train)
    assert_array_equal(knn.predict(X_test), st.predict(X_test))
    # Assert that all samples were labeled in iteration 0 (since there were no
    # unlabeled samples).
    assert np.all(st.labeled_iter_ == 0)
    assert st.termination_condition_ == "all_labeled"


def test_early_stopping():
    svc = SVC(gamma="scale", probability=True)
    st = SelfTrainingClassifier(svc)
    X_train_easy = [[1], [0], [1], [0.5]]
    y_train_easy = [1, 0, -1, -1]
    # X = [[0.5]] cannot be predicted on with a high confidence, so training
    # stops early
    st.fit(X_train_easy, y_train_easy)
    assert st.n_iter_ == 1
    assert st.termination_condition_ == "no_change"


def test_strings_dtype():
    clf = SelfTrainingClassifier(KNeighborsClassifier())
    X, y = make_blobs(n_samples=30, random_state=0, cluster_std=0.1)
    labels_multiclass = ["one", "two", "three"]

    y_strings = np.take(labels_multiclass, y)

    with pytest.raises(ValueError, match="dtype"):
        clf.fit(X, y_strings)


@pytest.mark.parametrize("verbose", [True, False])
def test_verbose(capsys, verbose):
    clf = SelfTrainingClassifier(KNeighborsClassifier(), verbose=verbose)
    clf.fit(X_train, y_train_missing_labels)

    captured = capsys.readouterr()

    if verbose:
        assert "iteration" in captured.out
    else:
        assert "iteration" not in captured.out


def test_verbose_k_best(capsys):
    st = SelfTrainingClassifier(
        KNeighborsClassifier(n_neighbors=1),
        criterion="k_best",
        k_best=10,
        verbose=True,
        max_iter=None,
    )

    y_train_only_one_label = np.copy(y_train)
    y_train_only_one_label[1:] = -1
    n_samples = y_train.shape[0]

    n_expected_iter = ceil((n_samples - 1) / 10)
    st.fit(X_train, y_train_only_one_label)

    captured = capsys.readouterr()

    msg = "End of iteration {}, added {} new labels."
    for i in range(1, n_expected_iter):
        assert msg.format(i, 10) in captured.out

    assert msg.format(n_expected_iter, (n_samples - 1) % 10) in captured.out


def test_k_best_selects_best():
    # Tests that the labels added by st really are the 10 best labels.
    svc = SVC(gamma="scale", probability=True, random_state=0)
    st = SelfTrainingClassifier(svc, criterion="k_best", max_iter=1, k_best=10)
    has_label = y_train_missing_labels != -1
    st.fit(X_train, y_train_missing_labels)

    got_label = ~has_label & (st.transduction_ != -1)

    svc.fit(X_train[has_label], y_train_missing_labels[has_label])
    pred = svc.predict_proba(X_train[~has_label])
    max_proba = np.max(pred, axis=1)

    most_confident_svc = X_train[~has_label][np.argsort(max_proba)[-10:]]
    added_by_st = X_train[np.where(got_label)].tolist()

    for row in most_confident_svc.tolist():
        assert row in added_by_st


def test_base_estimator_meta_estimator():
    # Check that a meta-estimator relying on an estimator implementing
    # `predict_proba` will work even if it does not expose this method before being
    # fitted.
    # Non-regression test for:
    # https://github.com/scikit-learn/scikit-learn/issues/19119

    base_estimator = StackingClassifier(
        estimators=[
            ("svc_1", SVC(probability=True)),
            ("svc_2", SVC(probability=True)),
        ],
        final_estimator=SVC(probability=True),
        cv=2,
    )

    assert hasattr(base_estimator, "predict_proba")
    clf = SelfTrainingClassifier(base_estimator=base_estimator)
    clf.fit(X_train, y_train_missing_labels)
    clf.predict_proba(X_test)

    base_estimator = StackingClassifier(
        estimators=[
            ("svc_1", SVC(probability=False)),
            ("svc_2", SVC(probability=False)),
        ],
        final_estimator=SVC(probability=False),
        cv=2,
    )

    assert not hasattr(base_estimator, "predict_proba")
    clf = SelfTrainingClassifier(base_estimator=base_estimator)
    with pytest.raises(AttributeError):
        clf.fit(X_train, y_train_missing_labels)


def test_self_training_estimator_attribute_error():
    """Check that we raise the proper AttributeErrors when the `base_estimator`
    does not implement the `predict_proba` method, which is called from within
    `fit`, or `decision_function`, which is decorated with `available_if`.

    Non-regression test for:
    https://github.com/scikit-learn/scikit-learn/issues/28108
    """
    # `SVC` with `probability=False` does not implement 'predict_proba' that
    # is required internally in `fit` of `SelfTrainingClassifier`. We expect
    # an AttributeError to be raised.
    base_estimator = SVC(probability=False, gamma="scale")
    self_training = SelfTrainingClassifier(base_estimator)

    with pytest.raises(AttributeError, match="has no attribute 'predict_proba'"):
        self_training.fit(X_train, y_train_missing_labels)

    # `DecisionTreeClassifier` does not implement 'decision_function' and
    # should raise an AttributeError
    self_training = SelfTrainingClassifier(base_estimator=DecisionTreeClassifier())

    outer_msg = "This 'SelfTrainingClassifier' has no attribute 'decision_function'"
    inner_msg = "'DecisionTreeClassifier' object has no attribute 'decision_function'"
    with pytest.raises(AttributeError, match=outer_msg) as exec_info:
        self_training.fit(X_train, y_train_missing_labels).decision_function(X_train)
    assert isinstance(exec_info.value.__cause__, AttributeError)
    assert inner_msg in str(exec_info.value.__cause__)