losses

DiceLoss

Bases: nn.Module

torch.nn.Module for dice loss (given by 1-dice_score) computation.

Parameters:

Name	Type	Description	Default
smooth	float	smooth value for dice score.	1
reduction	str	reduction method for computed dice scores. Can be one of "mean", "sum" or "none".	'mean'

FocalLoss

Bases: nn.Module

torch.nn.Module for focal loss (given by :math:FL(y, t) = -\beta t(1-y)^\gamma \log{y} - (1-\beta)(1-t)(y^\gamma \log{y})) computation.

Parameters:

Name	Type	Description	Default
reduction	str	reduction method for computed dice scores. Can be one of "mean", "sum" or "none".	'mean'
beta	float	coefficient ratio for positive class. Negative class has coefficient 1-beta.	0.5
gamma	float	focusing parameter for focal loss. When gamma=0, focal loss is equivalent to binary cros-sentropy. When gamma increases, highly misclassified predictions will have higher weight.	2.0

SumLosses

Bases: nn.Module

Wrapper used to compute a weighted sum of different losses.

Parameters:

Name	Type	Description	Default
*losses_with_coef	Sequence[Tuple[nn.Module, float]]	tuples containg a loss function and a float corresponding to the coefficient to use for the weighted sum of the losses.	()

dice_loss(input, target, smooth=1, reduction='mean')

Compute dice loss (given by 1-dice_score) for given prediction and target.

Parameters:

Name	Type	Description	Default
input	torch.Tensor	predicted input tensor of shape (N, ...).	required
target	torch.Tensor	target ground truth tensor of shape (N, ...).	required
smooth	float	smooth value for dice score.	1
reduction	str	reduction method for computed dice scores. Can be one of "mean", "sum" or "none".	'mean'

Returns:

Type	Description
torch.Tensor	Computed dice loss, optionally reduced using specified reduction method.

focal_loss(input, target, reduction='mean', beta=0.5, gamma=2.0, eps=1e-07)

Compute focal loss (given by :math:FL(y, t) = -\beta t(1-y)^\gamma \log{y} - (1-\beta)(1-t)(y^\gamma \log{y})) for given prediction and target. Sigmoid is applied to input before computation.

Parameters:

Name	Type	Description	Default
input	torch.Tensor	predicted input tensor of shape (N, ...).	required
target	torch.Tensor	target ground truth tensor of shape (N, ...).	required
reduction	str	reduction method for computed dice scores. Can be one of "mean", "sum" or "none".	'mean'
beta	float	coefficient ratio for positive class. Negative class has coefficient 1-beta.	0.5
gamma	float	focusing parameter for focal loss. When gamma=0, focal loss is equivalent to binary cros-sentropy. When gamma increases, highly misclassified predictions will have higher weight.	2.0
eps	float	term added for numerical stability.	1e-07

Returns:

Type	Description
torch.Tensor	Computed focal loss, optionally reduced using specified reduction method.

get_loss(name)

Get loss object from its name. If name is not recognized, raises ValueError.

Parameters:

Name	Type	Description	Default
name	str	can either be "bce" for :class:~torch.nn.BCEWithLogitsLoss, "focal" for :class:FocalLoss, "dice" for :class:DiceLoss or "sum_loss1_coef1_loss2_coef2_..._lossN_coefN" for :class:SumLosses.	required

Returns:

Type	Description
nn.Module	Loss function as a torch.nn.Module object.

get_loss_name(loss)

Get a loss object's name. If object is not recognized, will return the class name.

Parameters:

Name	Type	Description	Default
loss	nn.Module	loss module.	required

Returns:

Type	Description
str	Either "bce" for :class:~torch.nn.BCEWithLogitsLoss, "focal" for :class:FocalLoss, "dice" for :class:DiceLoss, "sum_loss1_coef1_loss2_coef2_..._lossN_coefN" for :class:SumLosses or the class name if none of the above.