class

SoftMarginLoss

extendsModule

SoftMarginLoss(reduction: str = 'mean')

source

Soft-margin binary hinge loss.

Computes the element-wise two-class logistic loss between the input x and the binary label y \in \{-1, +1\}:

\ell(x, y) = \log\!\bigl(1 + e^{-y \cdot x}\bigr)

This is the logistic (log-sigmoid) surrogate for the binary hinge. It is smooth and differentiable everywhere, unlike the hard hinge.

Parameters

reductionstr= 'mean'

'none' | 'mean' (default) | 'sum'.

Attributes

reductionstr

The reduction mode.

Notes

Input x : $(*)$ — real-valued scores.
Target y : $(*)$ — binary labels in $\{-1, +1\}$ .
Output : scalar for 'mean' / 'sum'; $(*)$ for 'none'.

As $x \to +\infty$ with $y = +1$ the loss approaches zero; as $x \to -\infty$ it grows linearly, which provides outlier robustness compared to cross-entropy.
This loss is essentially BCEWithLogitsLoss re-labelled for $\{-1, +1\}$ targets.

Examples

Positive and negative labels:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.SoftMarginLoss()
>>> x = lucid.tensor([ 1.5, -1.0,  0.5, -2.0])
>>> y = lucid.tensor([ 1.0,  1.0, -1.0, -1.0])
>>> loss = criterion(x, y)
Element-wise output:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.SoftMarginLoss(reduction="none")
>>> x = lucid.tensor([0.0, 2.0])
>>> y = lucid.tensor([1.0, -1.0])
>>> loss = criterion(x, y)  # shape (2,)

Methods (3)

dunder

init

→None

__init__(reduction: str = 'mean')

source

Initialise the SoftMarginLoss module. See the class docstring for parameter semantics.

forward

→Tensor

forward(x: Tensor, target: Tensor)

source

Compute the loss between predictions and targets.

Parameters

xTensor

Input tensor.

targetTensor

Input tensor.

Returns

Tensor

Scalar loss (or unreduced tensor depending on reduction).

extra_repr

→str

extra_repr()

source

Return a string representation of the layer's configuration.