ss_loss

后天...

已于 2022-06-19 12:39:29 修改

阅读量106

点赞数

文章标签： pytorch 深度学习 python

于 2022-06-19 12:31:05 首次发布

本文链接：https://blog.csdn.net/qq_34525135/article/details/125356638

版权

该博客探讨了一种名为`Mismatch_loss`的自定义损失函数，它结合了`ss_loss`来计算真值内外的损失。`ss_loss`通过计算目标值与网络输出之间的平方差并考虑平滑项来确定损失。`Mismatch_loss`在每个位置上对网络输出和目标值进行比较，并只计算非零最大目标和位置的损失。在提供的示例中，该损失函数被应用于具有多个通道和位置的张量，并在主函数中用具体的数据进行了演示。

摘要由CSDN通过智能技术生成

import torch
import torch.nn as nn

def ss_loss(net_out, target, smooth=1e-6):
    # 计算标签真值之内的损失
    molecule_1 = ((target - net_out) ** 2) * target
    denominator_1 = target
    loss_1 = (molecule_1.sum() + smooth) / (denominator_1.sum() + smooth)
    # 计算真值之外的损失
    molecule_2 = ((target - net_out) ** 2) * (1 - target)
    denominator_2 = (1 - target)
    loss_2 = (molecule_2.sum() + smooth) / (denominator_2.sum() + smooth)

    return 1 * loss_1 + 1 * loss_2


class Mismatch_loss(nn.Module):
    def __init__(self):
        super(Mismatch_loss, self).__init__()

    def forward(self, net_out, target, max_positiones):
        losses = []
        for j in range(net_out.shape[0]):
            for i in range(0, net_out.shape[1]):
                max_target = torch.max(target[j, i, ...])
                max_position = torch.max(max_positiones[j, i, ...])
                if max_target == 0 and max_position == 0:
                    continue
                else:
                    loss = ss_loss(net_out[j, i, ...], target[j, i, ...])
                    losses.append(loss)
            losses = torch.stack(losses)
        return losses.mean()


if __name__ == '__main__':
    net_out = torch.tensor(
      [[[[0.3, 0.3, 0.3, 0.9, 0.3, 0.3, 0.3, 0.3],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.3, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9],
         [0.3, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9, 0.9]]]]
    )
    # max_netout是根据net_out进行softmax得来的
    max_netout = torch.tensor(
      [[[[0, 0, 0, 1, 0, 0, 0, 0],
         [0, 1, 1, 1, 1, 0, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1]]]]
    )
    target = torch.tensor(
        [[[[0, 0, 0, 0, 0, 0, 0, 0],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1],
         [0, 1, 1, 1, 1, 1, 1, 1]]]]
    )
    classes = Mismatch_loss()
    loss = classes(net_out, target, max_netout)
    print(loss)