深度学习中常见的损失函数——关于使用与prediction与label的shape说明

最新推荐文章于 2024-07-30 10:51:24 发布

不变强不改名

最新推荐文章于 2024-07-30 10:51:24 发布

阅读量5k

点赞数

分类专栏： Pytorch 文章标签： pytorch 计算机视觉人工智能

本文链接：https://blog.csdn.net/m0_58772523/article/details/124417341

版权

Pytorch 专栏收录该内容

14 篇文章 2 订阅

订阅专栏

1. MSELoss

2. NLLLloss

3. CrossEntropyLoss

4. BCELoss与BCEWithLogitsLoss

1. MSELoss

用于一系列点的回归预测

prediction和label的shape应相同，否则可能会出错，例：

import torch
import torch.nn as nn

prediction = torch.ones(2, 5, 5) // 仅在最后一个维度计算，只需要保证prediction与labels维度一直即可
labels = torch.zeros(2, 5, 5) 
loss_func = nn.MSELoss(reduction="mean") // 默认去均值，可设置为sum
loss = loss_func(prediction, labels)
print(loss)

2. NLLLloss

用于多分类

NLLLoss是指对预测结果取负数，然后以label值为索引取出预测的值，prediction比label多一个维度，label中的值表示prediction各预测结果的索引，举例：

prediction = tensor([[-2.4076, -1.4076, -0.4076],

[-2.4076, -1.4076, -0.4076],

[-1.0986, -1.0986, -1.0986]])

label = tensor([0, 0, 1]) // 分别对应-2.4076，-2.4076， -1.0986

输出为：

((-(-2.4076)) + (-(-2.4076)) + (-(-1.0986))) / 3 = tensor(5.9138)

先对prediction做softmax再取log然后使用NLLLoss进行多分类预测

import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
__prediction = np.array([1, 2, 3, 3, 4, 5, 6, 6, 6]).reshape(3, 3)
_label = np.array([0, 0, 1])
_prediction = torch.tensor(__prediction, dtype=torch.float32)
prediction = torch.log(F.softmax(input=_prediction, dim=1))
label = torch.tensor(_label, dtype=torch.int64)
loss_func = nn.NLLLoss()
loss = loss_func(prediction, label)
print(loss)

3. CrossEntropyLoss

用于多分类

CrossEntropyLoss即为softmax+log+NLLLoss，prediction比label多一个维度，label中的值表示prediction各预测结果的索引

import torch
import torch.nn as nn
import numpy as np
_prediction = np.array([1, 2, 3, 3, 4, 5, 6, 6, 6]).reshape(3, 3)
_label = np.array([0, 0, 1])
prediction = torch.tensor(_prediction, dtype=torch.float32)
label = torch.tensor(_label, dtype=torch.int64)
loss_func = nn.CrossEntropyLoss()
loss = loss_func(prediction, label)
print(loss)

4. BCELoss与BCEWithLogitsLoss

用于二分类，如逻辑回归

prediction与label的shape一致，prediction的shape应为[batch_size, 1]，预测结果经sigmoid函数，趋向于为0或者1，label值为0或1

sigmoid + nn.BCELoss = nn.BCEWithLogitsLoss

import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
_prediction = (np.array([1, 2, 3])).reshape(3, 1)
_label = np.array([0, 1, 1]).reshape(3, 1)
prediction = torch.tensor(_prediction, dtype=torch.float32)
# prediction = F.sigmoid(torch.tensor(_prediction, dtype=torch.float32))
print(prediction)
label = torch.tensor(_label, dtype=torch.float32)
loss_func = nn.BCEWithLogitsLoss(reduction="sum")
loss = loss_func(prediction, label)
print(loss)