SmoothL1 loss

 

实现代码如下：

def smooth_l1_loss(input, target, sigma, reduce=True, normalizer=1.0):
    beta = 1. / (sigma ** 2)
    diff = torch.abs(input - target)
    cond = diff < beta
    loss = torch.where(cond, 0.5 * diff ** 2 / beta, diff - 0.5 * beta)
    if reduce:
        return torch.sum(loss) / normalizer
    return torch.sum(loss, dim=1) / normalizer