class

MultiMarginLoss

extendsModule

MultiMarginLoss(p: int = 1, margin: float = 1.0, weight: Tensor | None = None, reduction: Reduction = 'mean')

source edit

Multi-class hinge (SVM-style) margin loss.

For each sample with predicted scores $x \in \mathbb{R}^C$ and correct class index $y$ , the per-sample loss sums hinge contributions from all incorrect classes:

\mathcal{L}(x, y) = \frac{1}{C} \sum_{\substack{j=1 \\ j \neq y}}^{C} \max\!\bigl(0,\; \text{margin} - x[y] + x[j]\bigr)^p

When weight is provided, each class contribution is scaled by weight[y].

Parameters

pint= 1

Exponent of the hinge term; 1 (default) gives the standard hinge, 2 gives the squared hinge.

marginfloat= 1.0

Minimum required score gap between the true class and all others. Default 1.0.

weightTensor of shape (C,)= None

Manual rescaling weight per class.

reductionstr= 'mean'

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

Attributes

pint

Hinge exponent.

marginfloat

The score margin.

weightTensor or None

Per-class weight tensor.

reductionstr

The reduction mode.

Notes

Input x : $(N, C)$ — class scores.
Target y : $(N,)$ — integer class indices in $[0, C)$ .
Output : scalar for 'mean' / 'sum'; $(N,)$ for 'none'.

This is the multi-class generalisation of the SVM hinge loss.
With p=2 the loss is the squared hinge, which penalises margin violations more aggressively.

Examples

Standard hinge loss over 3 classes:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.MultiMarginLoss(margin=1.0)
>>> scores  = lucid.tensor([[0.3, 2.1, 0.5], [1.8, 0.2, 1.0]])
>>> targets = lucid.tensor([1, 0])
>>> loss = criterion(scores, targets)
Squared hinge with class weights:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.MultiMarginLoss(
...     p=2,
...     weight=lucid.tensor([1.0, 2.0, 1.0]),
... )
>>> scores  = lucid.tensor([[0.5, 1.5, 0.1]])
>>> targets = lucid.tensor([1])
>>> loss = criterion(scores, targets)

Used by 1

lucid.nn.modules

Constructors

dunder

init

→None

__init__(p: int = 1, margin: float = 1.0, weight: Tensor | None = None, reduction: Reduction = 'mean')

source edit

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

Instance methods

extra_repr

→str

extra_repr()

source edit

Return a string representation of the layer's configuration.

forward

→Tensor

forward(x: Tensor, target: Tensor)

source edit

Compute the loss between predictions and targets.

Parameters

xTensor

Input tensor.

targetTensor

Input tensor.

Returns

Tensor

Scalar loss (or unreduced tensor depending on reduction).

Standard hinge loss over 3 classes: >>> import lucid >>> import lucid.nn as nn >>> criterion = nn.MultiMarginLoss(margin=1.0) >>> scores = lucid.tensor([[0.3, 2.1, 0.5], [1.8, 0.2, 1.0]]) >>> targets = lucid.tensor([1, 0]) >>> loss = criterion(scores, targets) Squared hinge with class weights: >>> import lucid >>> import lucid.nn as nn >>> criterion = nn.MultiMarginLoss( ... p=2, ... weight=lucid.tensor([1.0, 2.0, 1.0]), ... ) >>> scores = lucid.tensor([[0.5, 1.5, 0.1]]) >>> targets = lucid.tensor([1]) >>> loss = criterion(scores, targets)