utils.py

# =========================
# utils.py
# =========================

import torch


def evidential_loss(y, alpha):

    S = torch.sum(
        alpha,
        dim=1,
        keepdim=True
    )

    probs = alpha / S

    error = torch.sum(
        (y - probs) ** 2,
        dim=1,
        keepdim=True
    )

    variance = torch.sum(
        alpha * (S - alpha)
        / (S * S * (S + 1)),
        dim=1,
        keepdim=True
    )

    loss = error + variance

    return loss.mean()


def predictive_entropy(probs):

    return -torch.sum(
        probs * torch.log(probs + 1e-10),
        dim=-1
    )