def compute_object_vanishing_gradient_and_loss(self, x, detections=None):
detections_ = np.asarray([])
encoded_labels = preprocess_true_boxes(detections_,
input_shape=self.model_img_size,
anchors=self.anchors,
num_classes=self.num_classes)
return K.get_session().run(
[self.object_vanishing_gradient, self.object_vanishing_loss],
feed_dict={
self.encoded_labels[0]: encoded_labels[0],
self.encoded_labels[1]: encoded_labels[1],
self.encoded_labels[2]: encoded_labels[2],
self.model.input: x
})
def compute_object_mislabeling_gradient_and_loss(self, x, detections):
detections_ = np.asarray([
detections[:, [-4, -3, -2, -1, 0]]
if len(detections) > 0 else detections
])
encoded_labels = preprocess_true_boxes(detections_,
input_shape=self.model_img_size,
anchors=self.anchors,
num_classes=self.num_classes)
return K.get_session().run(
[self.object_mislabeling_gradient, self.object_mislabeling_loss],
feed_dict={
self.encoded_labels[0]: encoded_labels[0],
self.encoded_labels[1]: encoded_labels[1],
self.encoded_labels[2]: encoded_labels[2],
self.model.input: x
})
def compute_object_fabrication_gradient(self, x, detections=None):
detections_ = np.asarray([])
encoded_labels = preprocess_true_boxes(detections_,
input_shape=self.model_img_size,
anchors=self.anchors,
num_classes=self.num_classes)
encoded_labels[0][..., 4] = 1
encoded_labels[1][..., 4] = 1
encoded_labels[2][..., 4] = 1
return K.get_session().run(self.object_fabrication_gradient,
feed_dict={
self.encoded_labels[0]:
encoded_labels[0],
self.encoded_labels[1]:
encoded_labels[1],
self.encoded_labels[2]:
encoded_labels[2],
self.model.input: x
})
