PointMLP/classification_ModelNet40/helper.py

import torch
import torch.nn.functional as F


def cal_loss(pred, gold, smoothing=True):
    """Calculate cross entropy loss, apply label smoothing if needed."""
    gold = gold.contiguous().view(-1)

    if smoothing:
        eps = 0.2
        n_class = pred.size(1)

        one_hot = torch.zeros_like(pred).scatter(1, gold.view(-1, 1), 1)
        one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
        log_prb = F.log_softmax(pred, dim=1)

        loss = -(one_hot * log_prb).sum(dim=1).mean()
    else:
        loss = F.cross_entropy(pred, gold, reduction="mean")

    return loss