lop_gpn_loss

probly.losses.torch.lop_gpn_loss(alpha_features: Tensor, mixture_weights: Tensor, y: Tensor, entropy_regularization: Tensor | None = None, entropy_weight: float = 0.0, reduction: str = 'sum') → Tensor[source]

LOP-GPN loss based on [DHullermeier24a].

Uses the mixture UCE objective computed by mixture_uce_loss(), with an optional caller-supplied entropy regularizer.

Parameters:

  • alpha_features – Feature-level Dirichlet concentration parameters with shape (N, C).

  • mixture_weights – Dense mixture weights with shape (B, N).

  • y – Ground-truth labels for the mixed nodes with shape (B,).

  • entropy_regularization – Optional per-sample entropy regularizer.

  • entropy_weight – Weight applied to entropy_regularization.

  • reduction – Reduction to apply, either "mean", "sum", or "none".

Returns:

Scalar or per-sample LOP-GPN loss.