rpn_distillation_loss

probly.losses.torch.rpn_distillation_loss(rpn_params: tuple[Tensor, Tensor, Tensor, Tensor], mus: list[Tensor], variances: list[Tensor]) → Tensor[source]

Regression Prior Network distillation loss based on [MCPG20].

Uses ensemble distribution distillation for Regression Prior Networks (RPN). This loss measures how well the RPN’s Normal-Wishart distribution matches the empirical ensemble distributions (mu_k, var_k) using normal_wishart_log_prob().

Parameters:

  • rpn_params – The RPN output parameters (m, l_precision, kappa, nu).

  • mus – Ensemble predicted means.

  • variances – Ensemble predicted variances.

Returns:

Scalar loss value.