【pytorch】——crossentropy如何求出每个样本的loss

pytorch, crossentropy

思路：

• 自己写crossentropy的计算过程
• 采用pytorch自身的crossentropy，需要将input变为：bc-1, target:b*-1

1. 自己实现

def Self_cross_entropy(pred, target):
    probability=F.softmax(pred, dim=1)#shape [num_samples,num_classes]
    log_P=torch.log(probability)
    '''对输入的target标签进行 one-hot编码，使用_scatter方法'''
    a=torch.unsqueeze(target,dim=0)
    # print(a.shape,probability.shape)
    one_hot = torch.zeros(probability.shape, device=target.device).scatter_(1, torch.unsqueeze(target,dim=1), 1)
    loss3 = - one_hot * log_P
    loss3 = loss3.sum(dim=1)
    loss3 = loss3.mean(1).mean(1)
    return loss3

2. 采用pytorch的crossentropy

• 将crossentropy的reduction='none'
• 改变input，target的shape：需要将input变为：bc-1, target:b*-1

example具体代码：

import torch
import torch.nn as nn
import torch.nn.functional as F

def Self_cross_entropy(pred, target):
    probability=F.softmax(pred, dim=1)#shape [num_samples,num_classes]
    log_P=torch.log(probability)
    '''对输入的target标签进行 one-hot编码，使用_scatter方法'''
    a=torch.unsqueeze(target,dim=0)
    # print(a.shape,probability.shape)
    one_hot = torch.zeros(probability.shape, device=target.device).scatter_(1, torch.unsqueeze(target,dim=1), 1)
    loss3 = - one_hot * log_P
    loss3 = loss3.sum(dim=1)
    loss3 = loss3.mean(1).mean(1)
    return loss3

loss = nn.CrossEntropyLoss()
input = torch.randn(4, 5, 10, 10, requires_grad=True)
target = torch.empty((4, 10, 10), dtype=torch.long).random_(5)
output = loss(input, target)
print("pytorch CEloss:", output)

selfout = Self_cross_entropy(input, target)
print("self CE", selfout.mean())
print("every CE", selfout)

loss2= nn.CrossEntropyLoss(reduction='none')
b, c = input.shape[:2]
output3 = loss2(input.view(b, c, -1), target.view(b, -1))
print("pytorch CE", output3.mean())
print("pytorch every CE", output3.mean(-1))

输出

pytorch CEloss: tensor(1.9572, grad_fn=)
self CE tensor(1.9572, grad_fn=)
every CE tensor([1.8493, 2.0463, 1.9736, 1.9595], grad_fn=)
pytorch CE tensor(1.9572, grad_fn=)
pytorch every CE tensor([1.8493, 2.0463, 1.9736, 1.9595], grad_fn=)