b6443232013e8d248fe26f59630d43bc9688df06,losses.py,,loss,#,25

Before Change

            regression_diff - 0.5 / 9.0
        )

        regression_losses.append(regression_loss[positive_indices, :].mean())


    return torch.stack(classification_losses).mean(), torch.stack(regression_losses).mean()

After Change

            0.5 * 9.0 * torch.pow(regression_diff, 2),
            regression_diff - 0.5 / 9.0
        )
        if positive_indices.sum() > 0:
            regression_losses.append(regression_loss[positive_indices, :].mean())
        else:
            regression_losses.append(torch.Tensor([0]).float().cuda())

    return torch.stack(classification_losses).mean(), torch.stack(regression_losses).mean()

Instances