def main():
saved_classifier_filename = '../classifiers/msu_mfsd' + version + extension
# load or recompute train features. If none, the train features are not loaded into memory
load_train_features = True
# retrain or load classifier
load_classifier = False
# load or recompute test features
load_test_features = True
# descriptor computer
mlbp_feature_computer = feature_computer.FrameFeatureComputer(features.MultiScaleLocalBinaryPatterns((8,1), (8,2),
(16, 2)))
#mlbp_feature_computer = feature_computer.FrameFeatureComputer(features.LocalBinaryPatterns(8,1))
(train_features, train_labels) = get_train_features_and_labels(load_train_features, mlbp_feature_computer)
if not load_classifier:
'''
param_grid = [
{'C':[0.0001, 0.001, 0.01], 'kernel':['linear'], 'class_weight':['balanced', None]},
{'C':[0.0001, 0.001, 0.01], 'kernel':['rbf'],'gamma':[0.0001, 0.001], 'class_weight':['balanced', None]}
]
'''
# C = 0.0001, kernel=linear, class_weight=balanced
param_grid = {'C':[0.0001], 'kernel':['linear'], 'gamma':[0.0001, 0.001, 1, 10],
'class_weight':['balanced'], 'decision_function_shape':['ovr']}
clf = GridSearchCV(svm.SVC(verbose=True, probability=True), param_grid, verbose=True, n_jobs=4)
#clf = svm.SVC(verbose=True, probability=True, C=0.0001, kernel='linear', class_weight='balanced')
#clf = svm.SVC(verbose=True, probability=True, C=0.001, kernel='rbf', gamma=0.001, class_weight='balanced')
# C = 5.77218597038 gamma = 9.38268773999 class_weight = balanced kernel = linear
clf = svm.SVC(verbose=True, probability=True, C=5.772185, gamma=9.3826877, class_weight='balanced',
kernel='linear')
#param_dist = {
# 'C':sp_uniform(0.00001, 10), 'kernel':['rbf', 'linear'], 'gamma':sp_uniform(0.0001, 10),
# 'class_weight':['balanced', None], 'decision_function_shape':['ovr', 'ovo']
#}
#n_iter_search = 20
#clf = RandomizedSearchCV(svm.SVC(verbose=True, probability=True), param_distributions=param_dist,
# n_iter=n_iter_search, n_jobs=4, verbose=True)
clf.fit(train_features, train_labels)
#print("Best estimator found by grid search:")
#print(clf.best_estimator_)
joblib.dump(clf, saved_classifier_filename)
else:
clf = joblib.load(saved_classifier_filename)
(test_features, test_labels) = get_test_features_and_labels(load_test_features, mlbp_feature_computer)
test_labels_bin = label_binarize(test_labels, classes=[-1,1])
pred_labels = clf.predict(test_features)
pred_confidences = clf.predict_proba(test_features)
plot_roc_curve(test_labels_bin, pred_confidences)
from sklearn.metrics import roc_curve, accuracy_score,confusion_matrix, roc_auc_score, auc
from scipy.optimize import brentq
from scipy.interpolate import interp1d
roc_auc = roc_auc_score(test_labels, pred_confidences[:,1])
print('ROC area under the curve score {}'.format(roc_auc))
# compute the equal error rate
fpr, tpr, _ = roc_curve(test_labels, pred_confidences[:,1])
eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
print('Equal error rate {}'.format(eer))
print('Accuracy score {}'.format(accuracy_score(test_labels, pred_labels)))
print('Confusion matrix {}'.format(confusion_matrix(test_labels, pred_labels)))
def main():
# pickels filenames
saved_classifier_filename = "../classifiers/msu_mfsd.pkl"
# load or recompute train features. If none, the train features are not loaded into memory
load_train_features = True
# retrain or load classifier
load_classifier = True
# load or recompute test features
load_test_features = True
# descriptor computer
mlbp_feature_computer = feature_computer.FrameFeatureComputer(
features.MultiScaleLocalBinaryPatterns((8, 1), (8, 2), (16, 2))
)
# mlbp_feature_computer = feature_computer.FrameFeatureComputer(features.LocalBinaryPatterns(8,1))
(
real_features,
spoof_features_per_dir,
labels_real,
labels_spoof_per_dir,
) = get_features_and_labels(load_train_features, mlbp_feature_computer)
# here I should do a cross validation on the features
"""
param_grid = [
{'C': [0.0001, 0.001, 0.01], 'kernel':['linear'], 'class_weight':['balanced', None]},
{'C': [0.0001, 0.001, 0.01], 'kernel':['rbf'],'gamma':[0.0001, 0.001], 'class_weight':['balanced', None]}
]
"""
test_fold = dbfeatures.compute_msu_ussa_subjects_folds_arr()
ps = PredefinedSplit(test_fold=test_fold)
clf = svm.SVC(
verbose=True,
probability=True,
C=0.0001,
kernel="linear",
class_weight="balanced",
)
folds_eer = []
threshes = []
confusion_matrices = []
for train_index, test_index in ps:
# split the features into current train and test folds
train_features = real_features[train_index]
test_features = real_features[test_index]
train_labels = labels_real[train_index]
test_labels = labels_real[test_index]
for i in range(len(spoof_features_per_dir)):
train_features = np.concatenate(
(train_features, spoof_features_per_dir[i][train_index]), 0
)
test_features = np.concatenate(
(test_features, spoof_features_per_dir[i][test_index]), 0
)
train_labels = np.concatenate(
(train_labels, labels_spoof_per_dir[i][train_index]), 0
)
test_labels = np.concatenate(
(test_labels, labels_spoof_per_dir[i][test_index]), 0
)
# train the classifier
clf.fit(train_features, train_labels)
# use the classifier to predict the labels for test_features
pred_labels = clf.predict(test_features)
# create the roc curve
fpr, tpr, threshold = roc_curve(test_labels, pred_labels, pos_label=1)
# compute the equal error rate
eer = brentq(lambda x: 1.0 - x - interp1d(fpr, tpr)(x), 0.0, 1.0)
thresh = interp1d(fpr, threshold)(eer)
folds_eer.append(eer)
threshes.append(thresh)
conf_mat = confusion_matrix(test_labels, pred_labels)
confusion_matrices.append(conf_mat)
# print the mean and standard deviation of equal error rate across the folds
print(np.mean(folds_eer), np.std(folds_eer))
for conf_mat in confusion_matrices:
print(conf_mat)
from plot_roc_curve import plot_roc_curve
from sklearn.metrics import roc_curve, accuracy_score, confusion_matrix, roc_auc_score
from scipy.optimize import brentq
from scipy.interpolate import interp1d
import joblib
import idiap_classifier
import casia_classifier
import msu_mfsd_classifier
import msu_ussa_classifier
import feature_computer
from faceSpoofDetection import features
mlbp_feature_computer = feature_computer.FrameFeatureComputer(
features.MultiScaleLocalBinaryPatterns((8, 1), (8, 2), (16, 2)))
classifier_name = "powerful_classifier.pkl"
classifier_path = "../classifiers/" + classifier_name
(
idiap_train_features,
idiap_train_labels,
) = idiap_classifier.get_train_features_and_labels(True, mlbp_feature_computer)
print("Loaded idiap training features")
(
idiap_test_features,
idiap_test_labels,
) = idiap_classifier.get_test_features_and_labels(True, mlbp_feature_computer)
print("Loaded idiap test features")
def main():
# descriptor computer
mlbp_feature_computer = feature_computer.FrameFeatureComputer(
features.MultiScaleLocalBinaryPatterns((8, 1), (8, 2), (16, 2)))
(train_features,
train_labels) = get_train_features_and_labels(load_train_features)
if not load_classifier:
'''
param_grid = [
{'C': [0.0001, 0.001, 0.01], 'kernel':['linear'], 'class_weight':['balanced', None]},
{'C': [0.0001, 0.001, 0.01], 'kernel':['rbf'],'gamma':[0.0001, 0.001], 'class_weight':['balanced', None]}
]
'''
param_grid = {
'C': [0.0001, 0.001, 0.01],
'kernel': ['linear'],
'class_weight': ['balanced', None]
}
#clf = GridSearchCV(svm.SVC(verbose=True, probability=True), param_grid, verbose=True, n_jobs=4)
clf = svm.SVC(verbose=True,
probability=True,
C=0.001,
kernel='linear',
class_weight='balanced')
#clf = svm.SVC(verbose=True, probability=True, C = 0.001, kernel='rbf', gamma=0.1, class_weight='balanced')
clf.fit(train_features, train_labels)
#print("Best estimator found by grid search:")
#print(clf.best_estimator_)
joblib.dump(clf, saved_classifier_filename)
else:
clf = joblib.load(saved_classifier_filename)
(test_features,
test_labels) = get_test_features_and_labels(load_test_features)
test_labels_bin = label_binarize(test_labels, classes=[-1, 1])
pred_labels = clf.predict(test_features)
pred_confidences = clf.predict_proba(test_features)
plot_roc_curve(test_labels_bin, pred_confidences)
from sklearn.metrics import roc_curve, accuracy_score, confusion_matrix, roc_auc_score, auc
from scipy.optimize import brentq
from scipy.interpolate import interp1d
roc_auc = roc_auc_score(test_labels, pred_confidences[:, 1])
print('ROC area under the curve score {}'.format(roc_auc))
# compute the equal error rate
fpr, tpr, _ = roc_curve(test_labels, pred_confidences[:, 1])
eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
print('Equal error rate {}'.format(eer))
print('Accuracy score {}'.format(accuracy_score(test_labels, pred_labels)))
print('Confusion matrix {}'.format(
confusion_matrix(test_labels, pred_labels)))
def main():
saved_classifier_filename = "../classifiers/idiap.pkl"
# load or recompute train features. If none, the train features are not loaded into memory
load_train_features = True
# retrain or load classifier
load_classifier = False
# load or recompute test features
load_test_features = True
# descriptor computer
mlbp_feature_computer = feature_computer.FrameFeatureComputer(
features.MultiScaleLocalBinaryPatterns((8, 1), (8, 2), (16, 2))
)
(train_features, train_labels) = get_train_features_and_labels(
load_train_features, mlbp_feature_computer
)
size = len(train_features)
test_features = train_features[-size / 10 :]
test_labels = train_labels[-size / 10 :]
train_features = train_features[: -size / 10]
train_labels = train_labels[: -size / 10]
if not load_classifier:
"""
param_grid = [
{'C':[0.0001, 0.001, 0.01], 'kernel':['linear'], 'class_weight':['balanced', None]},
{'C':[0.0001, 0.001, 0.01], 'kernel':['rbf'],'gamma':[0.0001, 0.001], 'class_weight':['balanced', None]}
]
# C = 0.0001, kernel=linear, class_weight=balanced
clf = GridSearchCV(svm.SVC(verbose=True, probability=True), param_grid, verbose=True)
"""
# clf = svm.SVC(verbose=True, probability=True, C=0.0001, kernel='linear', class_weight='balanced')
clf = svm.SVC(
verbose=True,
probability=True,
C=0.0001,
kernel="rbf",
gamma=0.001,
class_weight="balanced",
)
clf.fit(train_features, train_labels)
# print("Best estimator found by grid search:")
# print(clf.best_estimator_)
# joblib.dump(clf, saved_classifier_filename)
else:
clf = joblib.load(saved_classifier_filename)
# (test_features, test_labels) = get_test_features_and_labels(load_test_features)
test_labels_bin = label_binarize(test_labels, classes=[-1, 1])
pred_labels = clf.predict(test_features)
pred_confidences = clf.predict_proba(test_features)
plot_roc_curve(test_labels_bin, pred_confidences)
from sklearn.metrics import (
roc_curve,
accuracy_score,
confusion_matrix,
roc_auc_score,
auc,
)
from scipy.optimize import brentq
from scipy.interpolate import interp1d
roc_auc = roc_auc_score(test_labels, pred_confidences[:, 1])
print("ROC area under the curve score {}".format(roc_auc))
# compute the equal error rate
fpr, tpr, _ = roc_curve(test_labels, pred_confidences[:, 1])
eer = brentq(lambda x: 1.0 - x - interp1d(fpr, tpr)(x), 0.0, 1.0)
print("Equal error rate {}".format(eer))
print("Accuracy score {}".format(accuracy_score(test_labels, pred_labels)))
print("Confusion matrix {}".format(confusion_matrix(test_labels, pred_labels)))