Regression

Policy variant for regression. It treats each prediction head as a continuous quantity: the model output layout matches the target layout, the targets are floating point, and the model predicts target values directly.

Configuration

Selected as name: regression.

Option	Description	Type	Default
norm_targets	Transform and normalize the targets as the features are, predicting in that space and mapping predictions back to target units.	bool	true

With norm_targets on, the loss and metrics are computed on the normalized targets. A target that is NaN stays NaN. Set ignore_nan on the loss to train on such a label.

policy:
  name: regression
  kwargs:
    norm_targets: true