def _generator_loss(self, sr_predictions,
ground_truth_predictions) -> float:
super_resolution_mean = tf.math.reduce_mean(sr_predictions)
groud_truth_mean = tf.math.reduce_mean(ground_truth_predictions)
groud_truth_average = ground_truth_predictions - super_resolution_mean
super_resolution_average = sr_predictions - groud_truth_mean
# Reshaping tensor from shape [batch_size, 1] to [1, batch_size],
# because with the original shape `apply_softmax` was buggy and was
# outputting an array of 1's like [1, 1, 1, 1, ...].
groud_truth_average = self.reshape_tensor_to_softmax(
groud_truth_average)
super_resolution_average = self.reshape_tensor_to_softmax(
super_resolution_average)
gt_relativistic_average = self._compute_binary_crossentropy(
tf.zeros_like(groud_truth_average),
apply_softmax(groud_truth_average),
)
sr_relativistic_average = self._compute_binary_crossentropy(
tf.ones_like(super_resolution_average),
apply_softmax(super_resolution_average),
)
return (gt_relativistic_average + sr_relativistic_average) / 2
def _compute_arcloss(self,
embeddings,
ground_truth,
num_classes: int,
net_type: str = 'syn') -> float:
"""Compute the ArcLoss.
### Parameters
embeddings: Batch of Embedding vectors where loss will be calculated on.
ground_truth: Batch of Ground Truth classes.
fc_weights: Weights extracted from the last Fully Connected layer of\
the network (Embedding layer).
num_classes: Total number of classes in the dataset.
scale:
margin:
### Returns
The loss value."""
original_target_embeddings = embeddings
cos_theta = original_target_embeddings / self.scale
theta = tf.acos(cos_theta)
z = theta + self.margin
marginal_target_embeddings = tf.cos(z) * self.scale
one_hot_vector = tf.one_hot(ground_truth, depth=num_classes)
difference = marginal_target_embeddings - original_target_embeddings
new_one_hot = one_hot_vector * difference
softmax_output = apply_softmax(original_target_embeddings +
new_one_hot)
return self._compute_categorical_crossentropy(softmax_output,
one_hot_vector)
def compute_arcloss(self, embeddings, ground_truth) -> float:
"""Compute the ArcLoss.
### Parameters
embeddings: Batch of Embedding vectors where loss will be calculated on.
ground_truth: Batch of Ground Truth classes.
### Returns
The loss value."""
original_target_embeddings = embeddings
cos_theta = original_target_embeddings / self.scale
theta = tf.acos(cos_theta)
z = theta + self.margin
marginal_target_embeddings = tf.cos(z) * self.scale
one_hot_vector = tf.one_hot(ground_truth, depth=self.num_classes)
difference = marginal_target_embeddings - original_target_embeddings
new_one_hot = one_hot_vector * difference
softmax_output = apply_softmax(original_target_embeddings +
new_one_hot)
return self._compute_categorical_crossentropy(softmax_output,
one_hot_vector)
def test_softmax_zeros_output(zeros_array, softmax_zeros_array_output):
output = apply_softmax(zeros_array)
assert (output == softmax_zeros_array_output).numpy().all() == True
def test_softmax_output_sum(softmax_input_minimal):
output = apply_softmax(softmax_input_minimal)
assert tf.reduce_sum(output).numpy() == 1.0
def test_softmax_output(softmax_real_input, softmax_real_output):
output = apply_softmax(softmax_real_input)
assert (output == softmax_real_output).numpy().all() == True