lovasz

Lovász-Softmax loss.

Classes:

LovaszLoss –

Lovász-Softmax loss: a smooth surrogate for the mean-IoU objective.

LovaszLoss [source] ¶

LovaszLoss(
    ignore_index: int = -1,
    classes: Literal["present", "all"] = "present",
)

Bases: Module

Lovász-Softmax loss: a smooth surrogate for the mean-IoU objective.

Optimizes segmentation overlap directly, and is typically summed with cross-entropy. See The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks (Berman et al., 2018). The function form on probabilities is lovasz_softmax.

Parameters:

ignore_index (int, default: -1 ) –

Label value excluded from the loss.
classes (Literal['present', 'all'], default: 'present' ) –

"present" averages only classes present in the targets; "all" averages every class.

Methods:

forward –

Compute the loss from per-point logits \((N, C)\) and labels \((N,)\).

forward [source] ¶

forward(logits: Tensor, labels: Tensor) -> Tensor

Compute the loss from per-point logits \((N, C)\) and labels \((N,)\).

lovasz_softmax [source] ¶

lovasz_softmax(
    probas: Tensor,
    labels: Tensor,
    *,
    classes: Literal["present", "all"] = "present",
    ignore_index: int = -1,
) -> Tensor

Multi-class Lovász-Softmax loss over per-point class probabilities.

The smooth surrogate of the mean-IoU objective of The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks (Berman et al., 2018); the function form of LovaszLoss, which takes logits.

Parameters:

probas (Tensor) –

Per-point class probabilities (post-softmax), shape \((N, C)\).
labels (Tensor) –

Per-point ground-truth labels, shape \((N,)\).
classes (Literal['present', 'all'], default: 'present' ) –

"present" averages only classes present in labels; "all" averages every class.
ignore_index (int, default: -1 ) –

Label value excluded from the loss.

Returns:

Tensor –

The scalar loss (zero when no point or no class contributes).

Shape

probas: \((N, C)\)
labels: \((N,)\)
output: scalar

Example

>>> labels = torch.tensor([0, 1, 2])
>>> probas = torch.nn.functional.one_hot(labels, 3).float()
>>> lovasz_softmax(probas, labels).item()
0.0