class

MultiLabelMarginLoss

extendsMultilabelMarginLoss

MultiLabelMarginLoss(reduction: str = 'mean')

source

CamelCase alias for MultilabelMarginLoss.

Provided so that nn.MultiLabelMarginLoss (the capitalisation used by the reference framework) resolves to the same implementation. There is no behavioural difference — all parameters, attributes, and shapes are identical to MultilabelMarginLoss.

Parameters

reductionstr= 'mean'

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

Notes

Identical loss formula to MultilabelMarginLoss:

\text{loss}(x, y) = \sum_{ij} \frac{\max\!\bigl(0,\; 1 - (x[y[j]] - x[i])\bigr)}{|x|}

where the sum is taken over all pairs $(i, j)$ such that $i \notin \{y[0], y[1], \ldots\}$ . Targets are padded with $-1$ to mark end-of-list; entries at or beyond the first $-1$ are ignored.

Examples

Interchangeable with :class:`MultilabelMarginLoss`:
>>> import lucid
>>> import lucid.nn as nn
>>> criterion = nn.MultiLabelMarginLoss()
>>> scores = lucid.tensor([[0.5, 0.2, 0.9, 0.1]])
>>> target = lucid.tensor([[2, 0, -1, -1]])
>>> loss = criterion(scores, target)
Verifying alias equivalence:
>>> import lucid.nn as nn
>>> assert nn.MultiLabelMarginLoss is nn.MultiLabelMarginLoss
>>> c1 = nn.MultilabelMarginLoss()
>>> c2 = nn.MultiLabelMarginLoss()
>>> # Both produce identical results for the same inputs.