class

HingeEmbeddingLoss

extendsModule

HingeEmbeddingLoss(margin: float = 1.0, reduction: str = 'mean')

source

Hinge embedding loss for binary similarity learning.

Given an input x representing a scalar distance measure and a binary label y, the per-element loss is:

\ell(x, y) = \begin{cases} x & \text{if } y = +1 \\[4pt] \max\!\bigl(0,\; \text{margin} - x\bigr) & \text{if } y = -1 \end{cases}

With $y = +1$ the loss is the distance itself (the model should produce small distances for positive pairs). With $y = -1$ the model is penalised only if the distance is smaller than margin (dissimilar pairs should be pushed apart).

Parameters

marginfloat= 1.0

Minimum required distance for dissimilar pairs. Default 1.0.

reductionstr= 'mean'

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

Attributes

marginfloat

The minimum distance for dissimilar pairs.

reductionstr

The reduction mode.

Notes

x : $(N,)$ or $(*)$ — non-negative distance values.
y : same shape as x — labels in $\{-1, +1\}$ .
Output : scalar for 'mean' / 'sum'; same shape as input for 'none'.

This loss is a one-sided generalisation of the hinge loss used in SVMs, applied here to embedding distances rather than margin-based decision functions.
The input is typically a pre-computed distance (e.g. L2 norm between representations), not raw logits.

Examples

Embedding distances with margin of 1:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.HingeEmbeddingLoss(margin=1.0)
>>> dists   = lucid.tensor([0.3, 1.5, 0.8, 2.1])
>>> labels  = lucid.tensor([1.0, 1.0, -1.0, -1.0])
>>> loss = criterion(dists, labels)
Tighter margin to encourage more separation:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.HingeEmbeddingLoss(margin=2.0)
>>> dists  = lucid.tensor([0.5, 1.0, 3.0])
>>> labels = lucid.tensor([-1.0, -1.0, 1.0])
>>> loss = criterion(dists, labels)

Methods (3)

dunder

init

→None

__init__(margin: float = 1.0, reduction: str = 'mean')

source

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

forward

→Tensor

forward(x: Tensor, y: Tensor)

source

Compute the loss between predictions and targets.

Parameters

xTensor

Input tensor.

yTensor

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.