def build_losses(self,
labels,
model_outputs,
aux_losses=None) -> tf.Tensor:
"""Interface to compute losses. Refer to base_task.Task.build_losses."""
del labels
left_logits = model_outputs['left_logits']
right_logits = model_outputs['right_logits']
batch_size = tf_utils.get_shape_list(left_logits, name='batch_size')[0]
ranking_labels = tf.range(batch_size)
loss = tf_utils.safe_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=ranking_labels, logits=left_logits))
if self.task_config.model.bidirectional:
right_rank_loss = tf_utils.safe_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=ranking_labels, logits=right_logits))
loss += right_rank_loss
return tf.reduce_mean(loss)
def build_losses(self, labels, model_outputs, aux_losses=None):
"""Sparse categorical cross entropy loss.
Args:
labels: labels.
model_outputs: Output logits of the classifier.
aux_losses: auxiliarly loss tensors, i.e. `losses` in keras.Model.
Returns:
The total loss tensor.
"""
losses_config = self.task_config.losses
if losses_config.one_hot:
total_loss = tf.keras.losses.categorical_crossentropy(
labels,
model_outputs,
from_logits=True,
label_smoothing=losses_config.label_smoothing)
else:
total_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, model_outputs, from_logits=True)
total_loss = tf_utils.safe_mean(total_loss)
if aux_losses:
total_loss += tf.add_n(aux_losses)
return total_loss
def build_losses(self, labels, model_outputs, aux_losses=None):
"""Sigmoid Cross Entropy.
Args:
labels: tensor containing truth labels.
model_outputs: output logits of the classifier.
aux_losses: tensor containing auxiliarly loss tensors, i.e. `losses` in
keras.Model.
Returns:
Tensors: The total loss, model loss tensors.
"""
losses_config = self.task_config.losses
model_loss = tf.keras.losses.binary_crossentropy(
labels,
model_outputs,
from_logits=losses_config.from_logits,
label_smoothing=losses_config.label_smoothing)
model_loss = tf_utils.safe_mean(model_loss)
total_loss = model_loss
if aux_losses:
total_loss += tf.add_n(aux_losses)
return total_loss, model_loss
def _compute_top_k_loss(self, loss):
"""Computs top k loss."""
batch_size = tf.shape(loss)[0]
loss = tf.reshape(loss, shape=[batch_size, -1])
top_k_pixels = tf.cast(self._top_k_percent_pixels *
tf.cast(tf.shape(loss)[-1], dtype=tf.float32),
dtype=tf.int32)
# shape: [batch_size, top_k_pixels]
per_sample_top_k_loss = tf.map_fn(
fn=lambda x: tf.nn.top_k(x, k=top_k_pixels, sorted=False)[0],
elems=loss,
parallel_iterations=32,
fn_output_signature=tf.float32)
# shape: [batch_size]
per_sample_normalizer = tf.reduce_sum(tf.cast(
tf.not_equal(per_sample_top_k_loss, 0.0), dtype=tf.float32),
axis=-1) + EPSILON
per_sample_normalized_loss = tf.reduce_sum(
per_sample_top_k_loss, axis=-1) / per_sample_normalizer
normalized_loss = tf_utils.safe_mean(per_sample_normalized_loss)
return normalized_loss
def build_losses(self,
labels,
model_outputs,
aux_losses=None) -> tf.Tensor:
loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, tf.cast(model_outputs, tf.float32), from_logits=True)
return tf_utils.safe_mean(loss)
def build_losses(self, labels, model_outputs, aux_losses=None) -> tf.Tensor:
if self.task_config.model.num_classes == 1:
loss = tf.keras.losses.mean_squared_error(labels, model_outputs)
else:
loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, tf.cast(model_outputs, tf.float32), from_logits=True)
if aux_losses:
loss += tf.add_n(aux_losses)
return tf_utils.safe_mean(loss)
def __call__(self, logits, labels, sample_weight=None):
_, height, width, _ = labels.get_shape().as_list()
logits = mask_ops.resize_and_rescale_offsets(
logits, target_size=[height, width])
loss = self._loss_fn(y_true=labels, y_pred=logits)
if sample_weight is not None:
loss *= sample_weight
return tf_utils.safe_mean(loss)
def __call__(self, logits, labels, sample_weight=None):
_, height, width, _ = labels.get_shape().as_list()
logits = tf.image.resize(logits,
size=[height, width],
method=tf.image.ResizeMethod.BILINEAR)
loss = self._loss_fn(y_true=labels, y_pred=logits)
if sample_weight is not None:
loss *= sample_weight
return tf_utils.safe_mean(loss)
def build_losses(self,
labels,
model_outputs,
aux_losses=None) -> tf.Tensor:
ranking_loss = tfr_losses.get(
loss=self.task_config.loss,
reduction=self.task_config.loss_reduction)
loss = ranking_loss(tf.cast(labels, tf.float32),
tf.cast(model_outputs, tf.float32))
if aux_losses:
loss += tf.add_n(aux_losses)
return tf_utils.safe_mean(loss)
def build_losses(self,
labels: Any,
model_outputs: Any,
aux_losses: Optional[Any] = None):
"""Sparse categorical cross entropy loss.
Args:
labels: labels.
model_outputs: Output logits of the classifier.
aux_losses: auxiliarly loss tensors, i.e. `losses` in keras.Model.
Returns:
The total loss tensor.
"""
all_losses = {}
losses_config = self.task_config.losses
total_loss = None
if self._is_multilabel():
entropy = -tf.reduce_mean(
tf.reduce_sum(
model_outputs * tf.math.log(model_outputs + 1e-8), -1))
total_loss = tf.keras.losses.binary_crossentropy(labels,
model_outputs,
from_logits=False)
all_losses.update({
'class_loss': total_loss,
'entropy': entropy,
})
else:
if losses_config.one_hot:
total_loss = tf.keras.losses.categorical_crossentropy(
labels,
model_outputs,
from_logits=False,
label_smoothing=losses_config.label_smoothing)
else:
total_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, model_outputs, from_logits=False)
total_loss = tf_utils.safe_mean(total_loss)
all_losses.update({
'class_loss': total_loss,
})
if aux_losses:
all_losses.update({
'reg_loss': aux_losses,
})
total_loss += tf.add_n(aux_losses)
all_losses[self.loss] = total_loss
return all_losses
def build_losses(self,
labels: tf.Tensor,
model_outputs: tf.Tensor,
aux_losses: Optional[Any] = None) -> tf.Tensor:
"""Builds sparse categorical cross entropy loss.
Args:
labels: Input groundtruth labels.
model_outputs: Output logits of the classifier.
aux_losses: The auxiliarly loss tensors, i.e. `losses` in tf.keras.Model.
Returns:
The total loss tensor.
"""
losses_config = self.task_config.losses
is_multilabel = self.task_config.train_data.is_multilabel
if not is_multilabel:
if losses_config.one_hot:
total_loss = tf.keras.losses.categorical_crossentropy(
labels,
model_outputs,
from_logits=True,
label_smoothing=losses_config.label_smoothing)
elif losses_config.soft_labels:
total_loss = tf.nn.softmax_cross_entropy_with_logits(
labels, model_outputs)
else:
total_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, model_outputs, from_logits=True)
else:
# Multi-label weighted binary cross entropy loss.
total_loss = tf.nn.sigmoid_cross_entropy_with_logits(
labels=labels, logits=model_outputs)
total_loss = tf.reduce_sum(total_loss, axis=-1)
total_loss = tf_utils.safe_mean(total_loss)
if aux_losses:
total_loss += tf.add_n(aux_losses)
total_loss = losses_config.loss_weight * total_loss
return total_loss
def build_losses(self,
labels: tf.Tensor,
model_outputs: tf.Tensor,
aux_losses: Optional[Any] = None) -> tf.Tensor:
"""Builds losses for training and validation.
Args:
labels: Input groundtruth labels.
model_outputs: Output of the model.
aux_losses: The auxiliarly loss tensors, i.e. `losses` in tf.keras.Model.
Returns:
The total loss tensor.
"""
total_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, model_outputs, from_logits=True)
total_loss = tf_utils.safe_mean(total_loss)
if aux_losses:
total_loss += tf.add_n(aux_losses)
return total_loss
def __call__(self, predicted_scores, logits, labels):
actual_scores = get_actual_mask_scores(logits, labels,
self._ignore_label)
loss = tf_utils.safe_mean(
self._mse_loss(actual_scores, predicted_scores))
return loss
