compute_metric

Evaluate metric(output, target) over dataset under torch.no_grad.

Parameters:

Name	Type	Description	Default
model	Module	Model to evaluate. Moved to device in-place.	required
dataset	Dataset	Iterable yielding (data, target) pairs.	required
metric	Callable[[Tensor, Tensor], float]	Callable (output, target) -> value invoked once per batch.	required
batch_size	int	DataLoader batch size.	128
device	str	Device to run evaluation on.	'cuda'
**kwargs	Any	Extra keyword arguments forwarded to model(data, ...).	{}

Returns:

Type	Description
list	List of per-batch metric values in dataset order.