在PyTorch中进行二分类,有三种主要的全连接层,激活函数和loss function组合的方法,分别是:torch.nn.Linear+torch.sigmoid+torch.nn.BCELoss,torch.nn.Linear+BCEWithLogitsLoss,和torch.nn.Linear(输出维度为2)+torch.nn.CrossEntropyLoss,BCEWithLogitsLoss集成了Sigmoid,但是CrossEntropyLoss集成了Softmax。
下面重点写写几点区别:
- CrossEntropyLoss的输入logits=(3,2),target=(3)就够了,但是BCELoss、BCEWithLogitsLoss的输入得是logits=(3,2),target=(3,2)。也就是说BCE系列在设计的时候是期待把输出压缩成一维再过;但是CrossEntropyLoss是可以多维且每一维对应某一个类别的logit。
- CrossEntropyLoss的target是LongTensor,表示是哪一类;但是BCE系列是0到1之间的FloatTensor
- CrossEntropyLoss和BCE系列从数值上看除了0.5的情况下其他情况完全不一样。BCE系列的数值计算思路是target*log(logits)+(1-target)*log(1-logits);但是CrossEntropyLoss实际上是Softmax+NLLLoss, 最后数值计算思路变成-logits[sample_index][选中类别]+sum(exp(logits[sample_index][i]) for i in all)
来点代码:
import torch
from torch import nn
import math
loss_f = nn.CrossEntropyLoss(reduction='none')
output = torch.randn(2, 3) # 表示2个样本,3个类别
# target = torch.from_numpy(np.array([1, 0])).type(torch.LongTensor)
target = torch.LongTensor([0, 2]) # 表示label0和label2
loss = loss_f(output, target)
print('CrossEntropy loss: ', loss)
print(f'reduction=none,所以可以看到每一个样本loss,输出为[{loss}]')
nll = nn.NLLLoss(reduction='none')
logsoftmax = nn.LogSoftmax(dim=-1)
print('logsoftmax(output) result: {}'.format(logsoftmax(output)))
#可以清晰地看到nll这个loss在pytorch多分类里作用就是取个负号,同时去target对应下标拿一下已经算好的logsoftmax的值
print('nll(logsoftmax(output), target) :{}'.format(nll(logsoftmax(output), target)))
def manual_cal(sample_index, target, output):
# 输入是样本下标
sample_output = output[sample_index]
sample_target = target[sample_index]
x_class = sample_output[sample_target]
sample_output_len = len(sample_output)
log_sigma_exp_x = math.log(sum(math.exp(sample_output[i]) for i in range(sample_output_len)))
sample_loss = -x_class + log_sigma_exp_x
print(f'交叉熵手动计算loss{sample_index}:{sample_loss}')
return sample_loss
for i in range(2):
manual_cal(i, target, output)
# 如果nn.CrossEntropyLoss(reduction='mean')模式,刚好是手动计算的每个样本的loss取平均,最后输出的是一个值
# 如果nn.CrossEntropyLoss(reduction='none')模式,手动计算的loss0和loss1都会被列出来
'''
贴一个输出
CrossEntropy loss: tensor([2.7362, 0.9749])
reduction=none,所以可以看到每一个样本loss,输出为[tensor([2.7362, 0.9749])]
logsoftmax(output) result: tensor([[-2.7362, -1.4015, -0.3726],
[-0.8505, -1.6319, -0.9749]])
nll(logsoftmax(output), target) :tensor([2.7362, 0.9749])
交叉熵手动计算loss0:2.736179828643799
交叉熵手动计算loss1:0.9749272465705872
'''
如果用Pytorch来实现,可以看以下脚本,顺带连rce(logit和pred对换)和sce(ce和rce加强)也实现了:
import torch.nn.functional as F
import torch
import torch.nn as nn
# nn.CrossEntropyLoss() 和 KLDivLoss 关系
class SCELoss(nn.Module):
def __init__(self, num_classes=10, a=1, b=1, eps=1e-18):
super(SCELoss, self).__init__()
self.num_classes = num_classes
self.a = a #两个超参数
self.b = b
self.cross_entropy = nn.CrossEntropyLoss()
self.cross_entropy_none = nn.CrossEntropyLoss(reduction="none")
self.eps = eps
def forward(self, raw_pred, labels):
# CE 部分,正常的交叉熵损失
ce = self.cross_entropy(raw_pred, labels)
# RCE
pred = F.softmax(raw_pred, dim=1)
pred = torch.clamp(pred, min=self.eps, max=1.0)
label_one_hot = F.one_hot(labels, self.num_classes).float().to(pred.device)
label_one_hot = torch.clamp(label_one_hot, min=self.eps, max=1.0) #最小设为 1e-4,即 A 取 -4
my_ce = (-1 * torch.sum(label_one_hot * torch.log(pred), dim=1))
print('pred={} label_one_hot={} my_ce={}'.format(pred, label_one_hot, my_ce))
print('raw_pred={} labels={} official_ce={}'.format(raw_pred, labels, self.cross_entropy_none(raw_pred, labels)))
rce = (-1 * torch.sum(pred * torch.log(label_one_hot), dim=1))
print('pred={} label_one_hot={} rce={}'.format(pred, label_one_hot, rce))
loss = self.a * ce + self.b * rce.mean()
return loss
y_pred = torch.tensor([[10.0, 5.0, -6.0], [8.0, 8.0, 8.0]])
y_true = torch.tensor([0, 2])
ce1 = SCELoss(num_classes=3)(y_pred, y_true)
----------------------------------------