nexuml_library.layers.loss.classification_metrics¶
Classification metrics layer.
Supports both multiclass and multilabel classification metrics via the
multi_label flag. In multilabel mode, sigmoid is applied to logits
before updating metrics and labels are kept as multi-hot vectors.
Bases: LayerDefinition
Accumulate classification metrics with torchmetrics.
When multi_label is False (default) the layer uses multiclass
metrics (MulticlassAccuracy, MulticlassF1Score,
MulticlassAveragePrecision). Labels are expected as integer class
indices with shape [batch].
When multi_label is True the layer uses multilabel metrics
(MultilabelAccuracy, MultilabelF1Score,
MultilabelAveragePrecision). Sigmoid is applied to logits before
passing them to the metrics and labels are expected as multi-hot float
tensors with shape [batch, num_labels].