torchgeo.losses

TorchGeo losses.

class torchgeo.losses.QRLoss(eps=1e-08)

Bases: Module

The QR (forward) loss between class probabilities and predictions.

This loss is defined in ‘Resolving label uncertainty with implicit generative models’.

New in version 0.2.

__init__(eps=1e-08)

Initialize a new QRLoss instance.

Parameters:

eps (float) – small constant for numerical stability to prevent log(0) when computing the loss. Must be greater than or equal to 0.

Raises:

ValueError – If eps is less than 0.

New in version 0.8: The eps parameter.

forward(probs, target)

Computes the QR (forwards) loss on prior.

Parameters:

• probs (Tensor) – probabilities of predictions, expected shape B x C x H x W.

• target (Tensor) – prior probabilities, expected shape B x C x H x W.

Returns:

qr loss

Return type:

Tensor

New in version 0.8: The eps parameter.

class torchgeo.losses.RQLoss(eps=1e-08)

Bases: Module

The RQ (backwards) loss between class probabilities and predictions.

This loss is defined in ‘Resolving label uncertainty with implicit generative models’.

New in version 0.2.

__init__(eps=1e-08)

Initialize a new RQLoss instance.

Parameters:

eps (float) – small constant for numerical stability to prevent division by zero and log(0) when computing the loss. Must be greater than or equal to 0.

Raises:

ValueError – If eps is less than 0.

forward(probs, target)

Computes the RQ (backwards) loss on prior.

Parameters:

• probs (Tensor) – probabilities of predictions, expected shape B x C x H x W

• target (Tensor) – prior probabilities, expected shape B x C x H x W

Returns:

qr loss

Return type:

Tensor