« Back to API reference

neuraltrain.metrics.metrics.GroupedMetric

class neuraltrain.metrics.metrics.GroupedMetric(metric_name: str, kwargs: dict[str, Any] | None = None, reduction: Literal['mean', 'std'] | None = None)

A wrapper around a torchmetrics.Metric that allows for computing metrics per group.

With reduction set, compute reduces the per-group values to a scalar ("std" is the sample standard deviation), which LightningModule.log accepts. Without it, compute returns a dict of group id to value, which self.log() does not support; log it from the on_val_epoch_end and on_test_epoch_end methods:

metric_dict = {metric_name + “/” + k: v for k, v in grouped_metric.compute().items()} self.log_dict(metric_dict) grouped_metric.reset()

compute() → dict[str, float] | Tensor

Override this method to compute the final metric value.

This method will automatically synchronize state variables when running in distributed backend.

reset() → None

Reset metric state variables to their default value.

update(preds: Tensor, target: Tensor, groups: Tensor | None = None) → None

Update each group’s metric separately. groups: a tensor or list of group identifiers, same shape as preds/target.

« Back to API reference