def _load_features_and_targets(self, base_path):
X_train = np.load(
os.path.join(base_path, (NAME_FEATURES % self.train_alias)))
X_test = np.load(
os.path.join(base_path, (NAME_FEATURES % self.test_alias)))
y_train = np.load(
os.path.join(base_path, (NAME_TARGETS % self.train_alias)))
y_test = np.load(
os.path.join(base_path, (NAME_TARGETS % self.test_alias)))
names_test = load_txt(
os.path.join(base_path, (NAME_SAMPLES % self.test_alias)))
return X_train, y_train, X_test, y_test, names_test
def _test_inter_probas_classifier(self):
features_path = os.path.join(self.meta_dataset_output, self.INTER_NAME, "test", "features")
# [CBSR, RA, NUAA]
for dataset_origin in os.listdir(features_path):
for dataset_target in os.listdir(features_path):
if dataset_target == dataset_origin:
print('Origin and target are the same, skipping.')
continue
# path_dataset = os.path.join(self.features_root_path, dataset, self.target_all)
for classifier in self.classifiers:
# [ResNet, VGG...]
for model in self.models:
base_path = join(self.meta_dataset_output, self.INTER_NAME, "test", "features", dataset_origin,
dataset_target)
probas_original = self.__load_test_probas(dataset_origin, dataset_target, "original", model,
classifier)
probas_depth = self.__load_test_probas(dataset_origin, dataset_target, "depth", model,
classifier)
probas_illumination = self.__load_test_probas(dataset_origin, dataset_target, "illumination",
model, classifier)
probas_saliency = self.__load_test_probas(dataset_origin, dataset_target, "saliency", model,
classifier)
stacked_probas = np.stack((probas_depth, probas_illumination, probas_saliency, probas_original),
axis=2)
labels = self.__load_test_labels(dataset_origin, dataset_target, "original", model, classifier)
names = load_txt(
join(base_path, "original", model.alias, classifier.get_alias(), 'names.txt'))
classifier_path = join(self.meta_dataset_output, self.INTER_NAME, "train", "probas",
dataset_origin, model.alias, classifier.get_alias(), 'model.sav')
with open(classifier_path, 'rb') as f:
model_fitted = pickle.load(f)
stacked_probas = np.reshape(stacked_probas, (stacked_probas.shape[0], -1))
y_pred, y_pred_proba = self._predict(model_fitted, stacked_probas)
results = self._evaluate_results(y_pred, labels, names)
print('HTER: %f\nAPCER: %f\nBPCER: %f' % (results[0], results[1], results[2]))
def _test_intra_probas_classifier(self):
features_path = os.path.join(self.meta_dataset_output, self.INTRA_NAME,
"test", "features")
# [CBSR, RA, NUAA]
for dataset_origin in os.listdir(features_path):
# path_dataset = os.path.join(self.features_root_path, dataset, self.target_all)
for classifier in self.classifiers:
# [ResNet, VGG...]
for model in self.models:
base_path = join(self.meta_dataset_output, self.INTRA_NAME,
"test", "features", dataset_origin)
probas_original = self.__load_test_probas_intra(
dataset_origin, "original", model, classifier)
probas_depth = self.__load_test_probas_intra(
dataset_origin, "depth", model, classifier)
probas_illumination = self.__load_test_probas_intra(
dataset_origin, "illumination", model, classifier)
probas_saliency = self.__load_test_probas_intra(
dataset_origin, "saliency", model, classifier)
stacked_probas = np.stack(
(probas_depth, probas_illumination, probas_saliency,
probas_original),
axis=2)
labels = self.__load_test_labels_intra(
dataset_origin, "original", model, classifier)
names = load_txt(
join(base_path, "original", model.alias,
classifier.get_alias(), 'names.txt'))
classifier_path = join(self.meta_dataset_output,
self.INTRA_NAME, "train", "probas",
dataset_origin, model.alias,
classifier.get_alias(), 'model.sav')
with open(classifier_path, 'rb') as f:
model_fitted = pickle.load(f)
stacked_probas = np.reshape(stacked_probas,
(stacked_probas.shape[0], -1))
y_pred, y_pred_proba = self._predict(
model_fitted, stacked_probas)
print(
'\n\nDataset %s classifier %s model %s ' %
(dataset_origin, classifier.get_alias(), model.alias))
results = self._evaluate_results(y_pred, labels, names)
print('HTER: %f\nAPCER: %f\nBPCER: %f\nACC: %f' %
(results[0], results[1], results[2], results[3]))
output_dir = join(self.meta_dataset_output,
self.INTRA_NAME, "test", "probas",
dataset_origin, model.alias,
classifier.get_alias())
self._save_artifacts(classifier, output_dir, labels,
y_pred, y_pred_proba, results)
#
save_txt(join(output_dir, 'names.txt'), names)
np.save(join(output_dir, 'labels.npy'), labels)
print('stacked done!')