pytorch交叉熵损失函数的weight参数的使用

发布时间：2022-12-25 20:23:34 来源：网络

.

这是初夏，陽光灿烂，海水湛蓝，海鸥轻巧地向船桅间滑过，生命是多彩的，充满了青春的欢乐。

首先

必须将权重也转为Tensor的cuda格式；

然后

将该class_weight作为交叉熵函数对应参数的输入值。

class_weight = torch.FloatTensor([0.13859937, 0.5821059, 0.63871904, 2.30220396, 7.1588294, 0]).cuda()

补充：关于pytorch的CrossEntropyLoss的weight参数

首先这个weight参数比想象中的要考虑的多

你可以试试下面代码

import torch
import torch.nn as nn
inputs = torch.FloatTensor([0,1,0,0,0,1])
outputs = torch.LongTensor([0,1])
inputs = inputs.view((1,3,2))
outputs = outputs.view((1,2))
weight_CE = torch.FloatTensor([1,1,1])
ce = nn.CrossEntropyLoss(ignore_index=255,weight=weight_CE)
loss = ce(inputs,outputs)
print(loss)

tensor(1.4803)

这里的手动计算是：

loss1 = 0 + ln(e0 + e0 + e0) = 1.098

loss2 = 0 + ln(e1 + e0 + e1) = 1.86

求平均 = (loss1 *1 + loss2 *1)/ 2 = 1.4803

加权呢？

import torch
import torch.nn as nn
inputs = torch.FloatTensor([0,1,0,0,0,1])
outputs = torch.LongTensor([0,1])
inputs = inputs.view((1,3,2))
outputs = outputs.view((1,2))
weight_CE = torch.FloatTensor([1,2,3])
ce = nn.CrossEntropyLoss(ignore_index=255,weight=weight_CE)
loss = ce(inputs,outputs)
print(loss)

tensor(1.6075)

手算发现，并不是单纯的那权重相乘：

loss1 = 0 + ln(e0 + e0 + e0) = 1.098

loss2 = 0 + ln(e1 + e0 + e1) = 1.86

求平均 = (loss1 * 1 + loss2 * 2)/ 2 = 2.4113

而是

loss1 = 0 + ln(e0 + e0 + e0) = 1.098

loss2 = 0 + ln(e1 + e0 + e1) = 1.86

求平均 = (loss1 *1 + loss2 *2) / 3 = 1.6075

发现了么，加权后，除以的是权重的和，不是数目的和。

我们再验证一遍：

import torch
import torch.nn as nn
inputs = torch.FloatTensor([0,1,2,0,0,0,0,0,0,1,0,0.5])
outputs = torch.LongTensor([0,1,2,2])
inputs = inputs.view((1,3,4))
outputs = outputs.view((1,4))
weight_CE = torch.FloatTensor([1,2,3])
ce = nn.CrossEntropyLoss(weight=weight_CE)
# ce = nn.CrossEntropyLoss(ignore_index=255)
loss = ce(inputs,outputs)
print(loss)