def vector_loss(self, targets, predictions, logits):
with tf.variable_scope('loss_{}'.format(self.name)):
if self.loss['type'] == MEAN_SQUARED_ERROR:
train_loss = tf.reduce_sum(get_squared_error(
targets, predictions, self.name),
axis=1)
elif self.loss['type'] == MEAN_ABSOLUTE_ERROR:
train_loss = tf.reduce_sum(get_absolute_error(
targets, predictions, self.name),
axis=1)
elif self.loss['type'] == SOFTMAX_CROSS_ENTROPY:
train_loss = weighted_softmax_cross_entropy(
logits, targets, self.loss)
else:
train_mean_loss = None
train_loss = None
raise ValueError('Unsupported loss type {}'.format(
self.loss['type']))
train_mean_loss = tf.reduce_mean(train_loss,
name='train_mean_loss_{}'.format(
self.name))
return train_mean_loss, train_loss
def _get_loss(self, targets, hidden, logits, probabilities, class_weights,
class_biases):
with tf.variable_scope('loss_{}'.format(self.name)):
if ('class_similarities' in self.loss
and self.loss['class_similarities'] is not None):
class_similarities = self.loss['class_similarities']
if (class_similarities.shape[0] != self.num_classes
or class_similarities.shape[1] != self.num_classes):
logging.info(
'Class similarities is {} while num classes is {}'.
format(class_similarities.shape, self.num_classes))
if (class_similarities.shape[0] > self.num_classes and
class_similarities.shape[1] > self.num_classes):
# keep only the first num_classes rows and columns
class_similarities = class_similarities[:self.
num_classes, :
self.
num_classes]
elif (class_similarities.shape[0] < self.num_classes
and class_similarities.shape[1] < self.num_classes):
# fill the missing parts of the matrix with 0s and 1
# on the diagonal
diag = np.diag((self.num_classes, self.num_classes))
diag[:class_similarities.shape[0], :class_similarities.
shape[1]] = class_similarities
class_similarities = diag
class_similarities = tf.constant(
class_similarities,
dtype=tf.float32,
name='class_similarities_{}'.format(self.name))
vector_labels = tf.gather(class_similarities,
targets,
name='vector_labels_{}'.format(
self.name))
else:
vector_labels = tf.one_hot(targets,
self.num_classes,
name='vector_labels_{}'.format(
self.name))
if self.loss['type'] == 'sampled_softmax_cross_entropy':
train_loss, eval_loss = sampled_softmax_cross_entropy(
targets, hidden, logits, vector_labels, class_weights,
class_biases, self.loss, self.num_classes)
elif self.loss['type'] == 'softmax_cross_entropy':
train_loss = weighted_softmax_cross_entropy(
logits, vector_labels, self.loss)
eval_loss = train_loss
else:
train_mean_loss = None
eval_loss = None
raise ValueError('Unsupported loss type {}'.format(
self.loss['type']))
if self.loss['robust_lambda'] > 0:
train_loss = ((1 - self.loss['robust_lambda']) * train_loss +
self.loss['robust_lambda'] / self.num_classes)
train_mean_loss = tf.reduce_mean(train_loss,
name='train_mean_loss_{}'.format(
self.name))
if self.loss['confidence_penalty'] > 0:
mean_penalty = mean_confidence_penalty(probabilities,
self.num_classes)
train_mean_loss += (self.loss['confidence_penalty'] *
mean_penalty)
return train_mean_loss, eval_loss