CodingTips-torch.nn.functional.cross_entropy(x, y)

CodingTips-torch.nn.functional.cross_entropy(x, y) ——pytorch中提供的交叉熵损失函数

 

目录

Reference：

交叉熵损失函数与softmax，nll_loss之间的关系

补充内容：进化版交叉熵损失函数

---

Reference：

https://blog.csdn.net/m0_38133212/article/details/88087206

https://www.cnblogs.com/henry-zhao/p/13087275.html

交叉熵损失函数与softmax，nll_loss之间的关系

核心图片：

尝试代码如下：

#encoding:utf-8
import numpy as np
import torch
import torch.nn.functional as F


if __name__ == '__main__':
    x = np.array([[1, 2, 3, 4, 5],
        [1, 2, 3, 4, 5],
        [1, 2, 3, 4, 5]]).astype(np.float32) # 模拟训练出来的输出数据

    y = np.array([1,1,0]) # 三个样本，两个属于1类，一个属于0类

    x = torch.from_numpy(x)
    y = torch.from_numpy(y)

    soft_out = F.softmax(x, dim=1) # 进行softmax在每个样本上归一化，这时候可以看做概率，概率加和为1
    print("softmax_out:", soft_out) 

    log_soft_out = torch.log(soft_out) # 取对数
    print("log_softmax_out:", log_soft_out) 

    loss = F.nll_loss(log_soft_out, y) # nll_loss
    print("final loss", loss)


    print("直接使用cross entropy loss", F.cross_entropy(x, y))

补充内容：进化版交叉熵损失函数

Complement objective training：Chen, H. Y., Wang, P. H., Liu, C. H., Chang, S. C., Pan, J. Y., Chen, Y. T., ... & Juan, D. C. (2019). Complement objective training. arXiv preprint arXiv:1903.01182.

Guided Complement Entropy：Chen, H. Y., Liang, J. H., Chang, S. C., Pan, J. Y., Chen, Y. T., Wei, W., & Juan, D. C. (2019). Improving adversarial robustness via guided complement entropy. In Proceedings of the IEEE International Conference on Computer Vision (pp. 4881-4889).

曾经看到过这样两篇对于交叉熵损失函数的改进，从上边的图中可以明显看到交叉熵损失函数只利用了预测正确的类，而忽略了预测错误的类，该篇文章基于这个存在的问题，将预测错误的label也引入loss函数中进行考虑。

讲的比较好的博客：

https://blog.csdn.net/qq_36663791/article/details/103437368

https://blog.csdn.net/m0_37932351/article/details/103424303?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522160419899919724822529366%2522%252C%2522scm%2522%253A%252220140713.130102334..%2522%257D&request_id=160419899919724822529366&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~first_rank_v2~rank_v28-1-103424303.pc_first_rank_v2_rank_v28&utm_term=Complement-Cross-Entropy&spm=1018.2118.3001.4449