def _classify_inter_dataset(self, dataset_origin: str, dataset_target: str,
classifier: BaseClassifier, model: CnnModel,
prop: PropertyExtractor):
origin_path = join(self.features_root_path,
dataset_origin, self.target_all,
prop.get_property_alias(), model.alias)
target_path = join(self.features_root_path,
dataset_target, self.target_all,
prop.get_property_alias(), model.alias)
X_train = features_utils.concatenate_features(origin_path)
X_test = features_utils.concatenate_features(target_path)
y_train = features_utils.concatenate_labels(origin_path)
y_test = features_utils.concatenate_labels(target_path)
names_train = features_utils.concatenate_names(origin_path)
names_test = features_utils.concatenate_names(target_path)
y_pred, y_proba = self._fit_and_predict(classifier, X_train, y_train,
X_test)
results = self._evaluate_results(y_pred, y_test, names_test)
print('HTER: %f\nAPCER: %f\nBPCER: %f' %
(results[0], results[1], results[2]))
output_dir = join(self.inter_dataset_output, dataset_origin,
dataset_target, self.target_all,
prop.get_property_alias(), model.alias,
classifier.get_alias())
self._save_artifacts(classifier, output_dir, y_pred, y_proba, results)
def _test_inter_features_classifier(self) -> None:
"""
STEP 2.1
Used to predict on the test set with the already trained first classifier
"""
for dataset_origin in os.listdir(self.features_root_path):
for dataset_target in os.listdir(self.features_root_path):
if dataset_origin == dataset_target:
print('Origin and target are the same. Skipping.')
continue
# [ResNet, VGG]
for model in self.models:
base_path_target = join(self.features_root_path, dataset_target, self.target_all)
for prop in self.properties:
for classifier in self.classifiers:
print('origin %s target %s model %s prop %s classifier %s' % (dataset_origin,
dataset_target,
model.alias,
prop.get_property_alias(),
classifier.get_alias()))
classifier_path = join(self.meta_dataset_output, self.INTER_NAME, "train", "features",
dataset_origin, prop.get_property_alias(),
model.alias, classifier.get_alias(), 'model.sav')
with open(classifier_path, 'rb') as f:
model_fitted = pickle.load(f)
path_features = join(base_path_target, prop.get_property_alias(), model.alias)
features_concatenated = features_utils.concatenate_features(path_features)
names_concatenated = features_utils.concatenate_names(path_features)
labels_concatenated = features_utils.concatenate_labels(path_features)
y_pred, y_pred_proba = self._predict(model_fitted, features_concatenated)
results = self._evaluate_results(y_pred, labels_concatenated, names_concatenated)
print('HTER: %f\nAPCER: %f\nBPCER: %f' % (results[0], results[1], results[2]))
output_dir = join(self.meta_dataset_output,
self.INTER_NAME,
"test",
"features",
dataset_origin,
dataset_target,
prop.get_property_alias(),
model.alias,
classifier.get_alias())
self._save_artifacts(classifier, output_dir, y_pred, y_pred_proba, results)
#
save_txt(join(output_dir, 'names.txt'), names_concatenated)
np.save(join(output_dir, 'labels.npy'), labels_concatenated)
def _train_inter_feature_classifier(self) -> None:
"""
STEP 1.1
Used to train the first classifier (feature classifier) and generate the probabilities for each property map.
"""
# [CBSR, RA, NUAA]
for dataset in os.listdir(self.features_root_path):
# [ResNet, VGG, MobileNet, etc]
for model in self.models:
base_path = join(self.features_root_path, dataset,
self.target_all)
# [Depth, Illum, Saliency]
for prop in self.properties:
property_path = join(base_path, prop.get_property_alias(),
model.alias)
features_concatenated = features_utils.concatenate_features(
property_path)
names_concatenated = features_utils.concatenate_names(
property_path)
labels_concatenated = features_utils.concatenate_labels(
property_path)
for classifier in self.classifiers:
output_dir = join(self.meta_dataset_output,
self.INTER_NAME, "train", "features",
dataset, prop.get_property_alias(),
model.alias, classifier.get_alias())
if exists(output_dir):
print('Already generated, skipping.')
continue
y_pred, y_proba = self._fit_and_predict(
classifier, features_concatenated,
labels_concatenated, features_concatenated)
results = self._evaluate_results(
y_pred, labels_concatenated, names_concatenated)
print('HTER: %f\nAPCER: %f\nBPCER: %f' %
(results[0], results[1], results[2]))
self._save_artifacts(classifier, output_dir,
labels_concatenated, y_pred,
y_proba, results)
save_txt(join(output_dir, 'names.txt'),
names_concatenated)
np.save(join(output_dir, 'labels.npy'),
labels_concatenated)