一.函数原理简介:
cross_entroy_loss(x, target):
x为torch.Size([m, classes_sum]),target为一维的tensor,维度为torch.Size([classes_sum])
target必须为long类型,x为float类型
softmax–>logsoftmax–>NLLLoss
二.代码直观演示:
''' 步骤: one-hot编码 softmax求解 logsoftmax求解 NLLLoss求解 '''
def cross_entroy_loss(x, target):
rows = x.size(0)
cols = x.size(1)
one_hot =torch.zeros((rows, cols))
for i in range(len(target)):
cols = target[i]
one_hot[i][cols] = 1
#下面一句
''' torch.sum(torch.exp(x), dim = 1).reshape(-1,1) torch.sum(torch.exp(x), dim = 1)本身得到的是(1, 5)维度的数据 将torch.exp(x)的每一列求和之后并将其变为(5,1)维度的数据 '''
#先进行softmax运算:
''' softmax = e(xi) / sum(e(xi))#对每行进行操作,即每行的值加起来为1 '''
#再进行logsoftmax
''' logsoftmax = log(softmax) '''
#最后进行NLLLoss
''' Nllloss = -1/N * (onehot * logsoftmax) #N代表target.shape[0] '''
softmax = torch.exp(x) / torch.sum(torch.exp(x), dim = 1).reshape(-1,1)
logsoftmax = torch.log(softmax)
nllloss = - torch.sum(one_hot * logsoftmax) / target.shape[0]
CrossEntroyLoss_value = nllloss
return CrossEntroyLoss_value
本文地址:https://blog.csdn.net/weixin_40025586/article/details/109777923
黄山市民网:https://www.huangshanshimin.com/