10000 Self-training metaclassifier · scikit-learn/scikit-learn@af65887 · GitHub
[go: up one dir, main page]
Skip to content

Commit af65887

Browse files
plbeckerorausch
plbecker
authored and
orausch
committed
Self-training metaclassifier
Co-authored-by: pr0duktiv <patrice@5becker.de>
1 parent 5045c9b commit af65887

File tree

2 files changed

+194
-0
lines changed

2 files changed

+194
-0
lines changed

‎examples/semi_supervised/plot_self_training_performance.py

+80
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,80 @@
1+
import numpy as np
2+
import matplotlib.pyplot as plt
3+
from sklearn.semi_supervised.self_training import SelfTraining
4+
from sklearn.utils import shuffle
5+
from sklearn.svm import SVC
6+
from sklearn.datasets import load_iris
7+
from sklearn.model_selection import StratifiedKFold
8+
from sklearn.semi_supervised.label_propagation import LabelPropagation
9+
from sklearn.metrics import f1_score
10+
from sklearn.base import clone
11+
12+
supervised_score = []
13+
self_training_score = []
14+
label_propagation_score = []
15+
x_values = []
16+
17+
clf = SVC(probability=True, C=100, gamma=0.7, kernel='rbf')
18+
self_training_clf = SelfTraining(
19+
clone(clf, safe=True), max_iter=100, threshold=0.8
20+
)
21+
ls = LabelPropagation()
22+
23+
for t in range(20, 80):
24+
x_values.append(t)
25+
X, y = load_iris(return_X_y=True)
26+
X, y = shuffle(X, y, random_state=42)
27+
y_true = y.copy()
28+
29+
lim = t
30+
y[lim:] = -1
31+
32+
supervised_score_temp = []
33+
self_training_score_temp = []
34+
label_propagation_score_temp = []
35+
36+
skfolds = StratifiedKFold(n_splits=3, random_state=42)
37+
for train_index, test_index in skfolds.split(X, y):
38+
X_train = X[train_index]
39+
y_train = y[train_index]
40+
X_test = X[test_index]
41+
y_test = y[test_index]
42+
y_test_true = y_true[test_index]
43+
44+
X_train_filtered = X_train[np.where(y_train != -1)]
45+
y_train_filtered = y_train[np.where(y_train != -1)]
46+
47+
clf.fit(X_train_filtered, y_train_filtered)
48+
y_pred = clf.predict(X_test)
49+
supervised_score_temp.append(
50+
f1_score(y_test_true, y_pred, average='macro')
51+
)
52+
53+
self_training_clf.fit(X_train, y_train)
54+
y_pred = self_training_clf.predict(X_test)
55+
self_training_score_temp.append(
56+
f1_score(y_test_true, y_pred, average='macro')
57+
)
58+
59+
ls.fit(X_train, y_train)
60+
y_pred = ls.predict(X_test)
61+
label_propagation_score_temp.append(
62+
f1_score(y_test_true, y_pred, average='macro')
63+
)
64+
65+
supervised_score.append(np.array(supervised_score_temp).mean())
66+
self_training_score.append(np.array(self_training_score_temp).mean())
67+
label_propagation_score.append(
68+
np.array(label_propagation_score_temp).mean()
69+
)
70+
71+
72+
plt.figure(1)
73+
plt.plot(x_values, supervised_score, label='Supervised')
74+
plt.plot(x_values, self_training_score, label='Self-training')
75+
plt.plot(x_values, label_propagation_score, label='Label Propagation')
76+
plt.legend()
77+
plt.ylabel("Accuracy")
78+
plt.title("Comparision of classifiers on limited labeled data")
79+
plt.xlabel("Amount of Labeled Data")
80+
plt.show()

‎sklearn/semi_supervised/self_training.py

+114
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,114 @@
1+
import numpy as np
2+
3+
from ..base import BaseEstimator
4+
from ..utils.validation import check_X_y, check_array, check_is_fitted
5+
from ..utils import safe_mask
6+
7+
8+
def _check_estimator(estimator):
9+
"""Make sure that an estimator implements the necessary methods."""
10+
if not hasattr(estimator, "predict_proba"):
11+
raise ValueError("The base estimator should implement predict_proba!")
12+
13+
14+
class SelfTraining(BaseEstimator):
15+
16+
"""Self-Training classifier
17+
18+
Parameters
19+
----------
20+
estimator : estimator object
21+
An estimator object implementing `fit` and `predict_proba`.
22+
23+
threshold : float
24+
Threshold above which predictions are added to the labeled dataset
25+
26+
max_iter : integer
27+
Change maximum number of iterations allowed
28+
29+
"""
30+
def __init__(self, estimator, threshold=0.7, max_iter=500):
31+
self.estimator = estimator
32+
self.threshold = threshold
33+
self.max_iter = max_iter
34+
35+
def fit(self, X, y):
36+
"""
37+
Fits SelfTraining Estimator to dataset
38+
39+
Parameters
40+
----------
41+
X : array-like, shape = (n_samples, n_features)
42+
array representing the data
43+
y : array-like, shape = (n_samples, 1)
44+
array representing the labels
45+
46+
Returns
47+
-------
48+
self: returns an instance of self.
49+
"""
50+
X, y = check_X_y(X, y)
51+
_check_estimator(self.estimator)
52+
53+
# Data usable for supervised training
54+
X_labeled = X[safe_mask(X, np.where(y != -1))][0]
55+
y_labeled = y[safe_mask(y, np.where(y != -1))][0]
56+
57+
# Unlabeled data
58+
X_unlabeled = X[safe_mask(X, np.where(y == -1))][0]
59+
y_unlabeled = y[safe_mask(y, np.where(y == -1))][0]
60+
61+
iter = 0
62+
while (len(X_labeled) < len(X) and iter < self.max_iter):
63+
iter += 1
64+
self.estimator.fit(X_labeled, y_labeled)
65+
66+
# Select prediction where confidence is above the threshold
67+
pred = self.predict(X_unlabeled)
68+
max_proba = np.max(self.predict_proba(X_unlabeled), axis=1)
69+
confident = np.where(max_proba > self.threshold)[0]
70+
71+
# Add newly labeled confident predictions to the dataset
72+
X_labeled = np.append(X_labeled, X_unlabeled[confident], axis=0)
73+
y_labeled = np.append(y_labeled, pred[confident], axis=0)
74+
75+
# Remove already labeled data from unlabeled dataset
76+
X_unlabeled = np.delete(X_unlabeled, confident, axis=0)
77+
y_unlabeled = np.delete(y_unlabeled, confident, axis=0)
78+
79+
self.estimator.fit(X_labeled, y_labeled)
80+
return self.estimator
81+
82+
def predict(self, X):
83+
"""Predict on a dataset.
84+
85+
Parameters
86+
----------
87+
X : array-like, shape = (n_samples, n_features)
88+
array representing the data
89+
90+
Returns
91+
-------
92+
y : array-like, shape = (n_samples, 1)
93+
array with predicted labels
94+
"""
95+
check_is_fitted(self, 'estimator')
96+
X = check_array(X)
97+
return self.estimator.predict(X)
98+
99+
def predict_proba(self, X):
100+
"""Predict probability for each possible outcome.
101+
102+
Parameters
103+
----------
104+
X : array-like, shape = (n_samples, n_features)
105+
array representing the data
106+
107+
Returns
108+
-------
109+
y : array-like, shape = (n_samples, n_features)
110+
array with prediction probabilities
111+
"""
112+
_check_estimator(self.estimator)
113+
check_is_fitted(self, 'estimator')
114+
return self.estimator.predict_proba(X)

0 commit comments

Comments
 (0)
0