Losses
MultiBoxLoss
paz.optimization.losses.multi_box_loss.MultiBoxLoss(neg_pos_ratio=3, alpha=1.0, max_num_negatives=300)
Multi-box loss for a single-shot detection architecture.
Arguments
- neg_pos_ratio: Int. Number of negatives used per positive box.
- alpha: Float. Weight parameter for localization loss.
- max_num_negatives: Int. Maximum number of negatives per batch.
References
KeypointNetLoss
paz.optimization.losses.keypointnet_loss.KeypointNetLoss(num_keypoints, focal_length, rotation_noise=0.1, separation_delta=0.05, loss_weights={'consistency': 1.0, 'silhouette': 1.0, 'separation': 1.0, 'relative_pose': 0.2, 'variance': 0.5})
KeypointNet loss for discovering latent keypoints.
Arguments
- num_keypints: Int. Number of keypoints to discover.
- focal_length: Float. Focal length of camera
- rotation_noise: Float. Noise added to the estimation of the rotation.
- separation_delta: Float. Delta used for the ''separation'' loss.
- loss_weights: Dict. having as keys strings with the different losses names e.g. ''consistency'' and as value the weight used for that loss.
References
DiceLoss
paz.optimization.losses.segmentation.dice_loss.DiceLoss(beta=1.0, class_weights=1.0)
Computes the F beta loss. The F beta score is the geometric mean of the precision and recall, where the recall is B times more important than the precision.
Arguments
- beta: Float.
- class_weights: Float or list of floats of shape
(num_classes)
.
FocalLoss
paz.optimization.losses.segmentation.focal_loss.FocalLoss(gamma=2.0, alpha=0.25)
Computes the Focal loss. The Focal loss down weights properly classified examples.
Arguments
- gamma: Float.
- alpha: Float.
- class_weights: Float or list of floats of shape
(num_classes)
.
JaccardLoss
paz.optimization.losses.segmentation.jaccard_loss.JaccardLoss(class_weights=1.0)
Computes the Jaccard loss. The Jaccard score is the intersection over union of the predicted with respect to real masks.
Arguments
- class_weights: Float or list of floats of shape
(num_classes)
.
WeightedReconstruction
paz.optimization.losses.segmentation.weighted_reconstruction.WeightedReconstruction(beta=3.0)
Computes L1 reconstruction loss by multiplying positive alpha mask by beta.
Arguments
- beta: Float. Value used to multiple positive alpha mask values.
- RGBA_true: Tensor [batch, H, W, 4]. Color with alpha mask label values.
- RGB_pred: Tensor [batch, H, W, 3]. Predicted RGB values.
Returns
Tensor [batch, H, W] with weighted reconstruction loss values.
WeightedReconstructionWithError
paz.optimization.losses.segmentation.weighted_reconstruction.WeightedReconstructionWithError(beta=3.0)
Computes L1 reconstruction loss by multiplying positive alpha mask by beta.
Arguments
- RGBA_true: Tensor [batch, H, W, 4]. Color with alpha mask label values.
- RGBE_pred: Tensor [batch, H, W, 4]. Predicted RGB and error mask.
- beta: Float. Value used to multiple positive alpha mask values.
Returns
Tensor [batch, H, W] with weighted reconstruction loss values.