Monitoring#

class kornia.metrics.AverageMeter[source]#

Keep the running average of a metric over batches.

stats = AverageMeter() starts empty; stats.update(val, n) adds a value that stands for n samples, and stats.avg returns their n-weighted mean as a Python float.

Convention:

update(val, n) adds val * n to sum and n to count, so avg is the mean over samples when every val is a batch mean and n its batch size. val keeps the last value passed, unweighted.
Values are not rescaled: accuracy() percentages average to a percentage. A tensor val keeps sum a tensor that holds its autograd graph; pass a detached tensor or a Python number to avoid keeping it.
Float16 and bfloat16 values are weighted and accumulated in float32. Later updates can promote the accumulator to a wider dtype; val always retains the original value and dtype.

See also

accuracy(): a batch metric to feed it, as a percentage.

Example

>>> stats = AverageMeter()
>>> acc1 = torch.tensor(99.0)  # top-1 accuracy of a batch in percent, as in K.metrics.accuracy(...)[0]
>>> stats.update(acc1, n=1)  # where n is batch size usually
>>> round(stats.avg, 2)
99.0