def test_calibration_multiclass():
"""Test calibration for multiclass """
# test multi-class setting with classifier that implements
# only decision function
clf = LinearSVC()
X, y_idx = make_blobs(n_samples=100, n_features=2, random_state=42,
centers=3, cluster_std=3.0)
# Use categorical labels to check that CalibratedClassifierCV supports
# them correctly
target_names = np.array(['a', 'b', 'c'])
y = target_names[y_idx]
X_train, y_train = X[::2], y[::2]
X_test, y_test = X[1::2], y[1::2]
clf.fit(X_train, y_train)
for method in ['isotonic', 'sigmoid']:
cal_clf = CalibratedClassifierCV(clf, method=method, cv=2)
cal_clf.fit(X_train, y_train)
probas = cal_clf.predict_proba(X_test)
assert_array_almost_equal(np.sum(probas, axis=1), np.ones(len(X_test)))
# Check that log-loss of calibrated classifier is smaller than
# log-loss of naively turned OvR decision function to probabilities
# via softmax
def softmax(y_pred):
e = np.exp(-y_pred)
return e / e.sum(axis=1).reshape(-1, 1)
uncalibrated_log_loss = \
log_loss(y_test, softmax(clf.decision_function(X_test)))
calibrated_log_loss = log_loss(y_test, probas)
assert uncalibrated_log_loss >= calibrated_log_loss
# Test that calibration of a multiclass classifier decreases log-loss
# for RandomForestClassifier
X, y = make_blobs(n_samples=100, n_features=2, random_state=42,
cluster_std=3.0)
X_train, y_train = X[::2], y[::2]
X_test, y_test = X[1::2], y[1::2]
clf = RandomForestClassifier(n_estimators=10, random_state=42)
clf.fit(X_train, y_train)
clf_probs = clf.predict_proba(X_test)
loss = log_loss(y_test, clf_probs)
for method in ['isotonic', 'sigmoid']:
cal_clf = CalibratedClassifierCV(clf, method=method, cv=3)
cal_clf.fit(X_train, y_train)
cal_clf_probs = cal_clf.predict_proba(X_test)
cal_loss = log_loss(y_test, cal_clf_probs)
assert loss > cal_loss
"""
import numpy as np
import matplotlib.pyplot as plt
from mrex.datasets import make_blobs
from mrex.svm import LinearSVC
X, y = make_blobs(n_samples=40, centers=2, random_state=0)
plt.figure(figsize=(10, 5))
for i, C in enumerate([1, 100]):
# "hinge" is the standard SVM loss
clf = LinearSVC(C=C, loss="hinge", random_state=42).fit(X, y)
# obtain the support vectors through the decision function
decision_function = clf.decision_function(X)
# we can also calculate the decision function manually
# decision_function = np.dot(X, clf.coef_[0]) + clf.intercept_[0]
support_vector_indices = np.where((2 * y - 1) * decision_function <= 1)[0]
support_vectors = X[support_vector_indices]
plt.subplot(1, 2, i + 1)
plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)
ax = plt.gca()
xlim = ax.get_xlim()
ylim = ax.get_ylim()
xx, yy = np.meshgrid(np.linspace(xlim[0], xlim[1], 50),
np.linspace(ylim[0], ylim[1], 50))
Z = clf.decision_function(np.c_[xx.ravel(), yy.ravel()])
Z = Z.reshape(xx.shape)
plt.contour(xx,
