BCELoss和BCEWithLogitsLoss要求的input都是经过sigmoid产生的分类概率,target是0或1的二分类。
假设我们有一个3×3的输入,也就是batch_size是3,target是3×3,表示有3个标签。现在用这个例子做一个演示:
import torch
import torch.nn as nn
m = nn.Sigmoid()
loss1 = nn.BCELoss()
loss2 = nn.BCEWithLogitsLoss()
input = torch.randn((3, 3), requires_grad=True)
target = torch.empty(3, 3).random_(2) # 0或1
# 在使用nn.BCELoss需要在该层前面加上Sigmoid函数
output1 = loss1(m(input), target)
output2 = loss2(input, target)
input: tensor([[ 0.2295, -0.4329, 0.8312],
[ 0.5107, -0.2862, 1.3013],
[ 0.2942, -0.7253, 0.3596]], requires_grad=True)
sigmoid(input): tensor([[0.5571, 0.3934, 0.6966],
[0.6250, 0.4289, 0.7861],
[0.5730, 0.3262, 0.5889]], grad_fn=<SigmoidBackward>)
target: tensor([[0., 1., 1.],
[1., 0., 0.],
[1., 1., 0.]])
output1: tensor(0.8052, grad_fn=<BinaryCrossEntropyBackward>)
output2: tensor(0.8052, grad_fn=<BinaryCrossEntropyWithLogitsBackward>)