Multiclass

Policy variant for multiclass classification. It treats each prediction head as a discrete label and configures the run accordingly: the model output is interpreted as per-class scores, the targets are integer class indices, and targets are not normalized.

The set of classes is discovered from the dataset. The policy reads the distinct target values recorded across the dataset’s files and sizes each head’s output to cover them, so the number of output channels follows the data rather than being configured by hand.

Configuration

Selected as name: multiclass. Takes no options.

policy:
  name: multiclass
  kwargs: {}