def box_classification_loss(inputs,
outputs,
is_intermediate=False,
is_balanced=False,
mine_hard_negatives=False,
hard_negative_score_threshold=0.5):
"""Calculates the voxel level classification loss.
Args:
inputs: A dictionary of tf.Tensors with our input label data.
outputs: A dictionary of tf.Tensors with the network output.
is_intermediate: If True, loss will be computed on intermediate tensors.
is_balanced: If True, the per-voxel losses are re-weighted to have equal
total weight for foreground vs. background voxels.
mine_hard_negatives: If True, mines hard negatives and applies loss on them
too.
hard_negative_score_threshold: A prediction is a hard negative if its label
is 0 and the score for the 0 class is less than this threshold.
Returns:
loss: A tf.float32 scalar corresponding to softmax classification loss.
Raises:
ValueError: If the size of the third dimension of the predicted logits is
unknown at graph construction.
"""
return loss_utils.apply_unbatched_loss_on_voxel_tensors(
inputs=inputs,
outputs=outputs,
unbatched_loss_fn=functools.partial(
_box_classification_loss_unbatched,
is_intermediate=is_intermediate,
is_balanced=is_balanced,
mine_hard_negatives=mine_hard_negatives,
hard_negative_score_threshold=hard_negative_score_threshold))
def box_classification_using_center_distance_loss(
inputs,
outputs,
is_intermediate=False,
is_balanced=False,
max_positive_normalized_distance=0.3):
"""Calculates the loss based on predicted center distance from gt center.
Computes the loss using the object properties of the voxel tensors.
Args:
inputs: A dictionary of tf.Tensors with our input label data.
outputs: A dictionary of tf.Tensors with the network output.
is_intermediate: If True, loss will be computed on intermediate tensors.
is_balanced: If True, the per-voxel losses are re-weighted to have equal
total weight for foreground vs. background voxels.
max_positive_normalized_distance: Maximum distance of a predicted box from
the ground-truth box that we use to classify the predicted box as
positive.
Returns:
loss: A tf.float32 scalar corresponding to distance confidence loss.
"""
return loss_utils.apply_unbatched_loss_on_voxel_tensors(
inputs=inputs,
outputs=outputs,
unbatched_loss_fn=functools.partial(
_box_classification_using_center_distance_loss_unbatched,
is_intermediate=is_intermediate,
is_balanced=is_balanced,
max_positive_normalized_distance=max_positive_normalized_distance))
def box_corner_distance_loss_on_voxel_tensors(
inputs,
outputs,
loss_type,
delta=1.0,
is_balanced=False,
is_intermediate=False):
"""Computes regression loss on object corner locations using object tensors.
Args:
inputs: A dictionary of tf.Tensors with our input data.
outputs: A dictionary of tf.Tensors with the network output.
loss_type: Loss type.
delta: float, the voxel where the huber loss function changes from a
quadratic to linear.
is_balanced: If True, the per-voxel losses are re-weighted to have equal
total weight for each object instance.
is_intermediate: If True, intermediate tensors are used for computing
the loss.
Returns:
localization_loss: A tf.float32 scalar corresponding to localization loss.
"""
standard_fields.check_input_voxel_fields(inputs=inputs)
standard_fields.check_output_voxel_fields(outputs=outputs)
def fn(inputs_1, outputs_1):
return _box_corner_distance_loss_on_voxel_tensors_unbatched(
inputs_1=inputs_1,
outputs_1=outputs_1,
loss_type=loss_type,
delta=delta,
is_balanced=is_balanced,
is_intermediate=is_intermediate)
return loss_utils.apply_unbatched_loss_on_voxel_tensors(
inputs=inputs, outputs=outputs, unbatched_loss_fn=fn)
def hard_negative_classification_loss(inputs,
outputs,
is_intermediate=False,
gamma=1.0):
"""Calculates the loss based on predicted center distance from gt center.
Computes the loss using the object properties of the voxel tensors.
Args:
inputs: A dictionary of tf.Tensors with our input label data.
outputs: A dictionary of tf.Tensors with the network output.
is_intermediate: If True, loss will be computed on intermediate tensors.
gamma: Gamma similar to how it is used in focal loss.
Returns:
loss: A tf.float32 scalar corresponding to distance confidence loss.
"""
return loss_utils.apply_unbatched_loss_on_voxel_tensors(
inputs=inputs,
outputs=outputs,
unbatched_loss_fn=functools.partial(
_voxel_hard_negative_classification_loss_unbatched,
is_intermediate=is_intermediate,
gamma=gamma))
