L1

Loss variant computing the mean absolute error between the model output and the targets. It penalizes errors linearly, making it less sensitive to outliers than the squared error. Compatible with the regression policy.

Configuration

Selected as name: l1.

Option	Description	Type	Default
reduction	How per-sample losses are reduced.	mean | sum	mean
ignore_nan	Leave out NaN targets, the events a label has no value for, so they add nothing to the loss.	bool	false

With ignore_nan, each head’s mean is taken over the targets kept, and a head with no target in a batch adds nothing rather than NaN. Without it, a NaN target makes the loss NaN.

components:
  loss:
    name: l1
    kwargs: {}