pytorch交叉熵損失函數的weight參數的使用

Posted on by

首先

必須將權重也轉為Tensor的cuda格式；

然後

將該class_weight作為交叉熵函數對應參數的輸入值。

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

補充：關於pytorch的CrossEntropyLoss的weight參數

首先這個weight參數比想象中的要考慮的多

你可以試試下面代碼

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)

這裡的手動計算是：

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

加權呢？

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)

手算發現，並不是單純的那權重相乘：

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

發現瞭麼，加權後，除以的是權重的和，不是數目的和。

我們再驗證一遍：

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)

tensor(1.5472)

手算：

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

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

loss3 = 0 + ln(e2 + e0 + e0) = 2.2395

loss4 = -0.5 + ln(e0.5 + e0 + e0) = 0.7943

求平均 = (loss1 * 1 + loss2 * 2+loss3 * 3+loss4 * 3) / 9 = 1.5472

可能有人對loss的CE計算過程有疑問，我這裡細致寫寫交叉熵的計算過程，就拿最後一個例子的loss4的計算說明

以上為個人經驗，希望能給大傢一個參考，也希望大傢多多支持WalkonNet。

推薦閱讀：