def build_losses(self,
labels: Mapping[str, tf.Tensor],
model_outputs: Union[Mapping[str, tf.Tensor], tf.Tensor],
aux_losses: Optional[Any] = None):
"""Segmentation 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.
"""
loss_params = self._task_config.losses
segmentation_loss_fn = segmentation_losses.SegmentationLoss(
loss_params.label_smoothing,
loss_params.class_weights,
loss_params.ignore_label,
use_groundtruth_dimension=loss_params.use_groundtruth_dimension,
top_k_percent_pixels=loss_params.top_k_percent_pixels)
total_loss = segmentation_loss_fn(model_outputs, labels['masks'])
if aux_losses:
total_loss += tf.add_n(aux_losses)
return total_loss
def build_losses(self,
outputs: Mapping[str, Any],
labels: Mapping[str, Any],
aux_losses: Optional[Any] = None) -> Dict[str, tf.Tensor]:
"""Build Panoptic Mask R-CNN losses."""
params = self.task_config.losses
use_groundtruth_dimension = params.semantic_segmentation_use_groundtruth_dimension
segmentation_loss_fn = segmentation_losses.SegmentationLoss(
label_smoothing=params.semantic_segmentation_label_smoothing,
class_weights=params.semantic_segmentation_class_weights,
ignore_label=params.semantic_segmentation_ignore_label,
use_groundtruth_dimension=use_groundtruth_dimension,
top_k_percent_pixels=params.
semantic_segmentation_top_k_percent_pixels)
instance_segmentation_weight = params.instance_segmentation_weight
semantic_segmentation_weight = params.semantic_segmentation_weight
losses = super(PanopticMaskRCNNTask,
self).build_losses(outputs=outputs,
labels=labels,
aux_losses=None)
maskrcnn_loss = losses['model_loss']
segmentation_loss = segmentation_loss_fn(
outputs['segmentation_outputs'], labels['gt_segmentation_mask'])
model_loss = (instance_segmentation_weight * maskrcnn_loss +
semantic_segmentation_weight * segmentation_loss)
total_loss = model_loss
if aux_losses:
reg_loss = tf.reduce_sum(aux_losses)
total_loss = model_loss + reg_loss
losses.update({
'total_loss': total_loss,
'maskrcnn_loss': maskrcnn_loss,
'segmentation_loss': segmentation_loss,
'model_loss': model_loss,
})
return losses
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.
"""
loss_params = self._task_config.losses
segmentation_loss_fn = segmentation_losses.SegmentationLoss(
loss_params.label_smoothing,
loss_params.class_weights,
loss_params.ignore_label,
use_groundtruth_dimension=loss_params.use_groundtruth_dimension)
total_loss = segmentation_loss_fn(model_outputs, labels['masks'])
if aux_losses:
total_loss += tf.add_n(aux_losses)
return total_loss
