pytorch自定义loss损失函数

自定义loss的方法有很多，但是在博主查资料的时候发现有挺多写法会有问题，靠谱一点的方法是把loss作为一个pytorch的模块，比如：

class CustomLoss(nn.Module): # 注意继承 nn.Module
    def __init__(self):
        super(CustomLoss, self).__init__()

    def forward(self, x, y):
        # .....这里写x与y的处理逻辑，即loss的计算方法
        return loss # 注意最后只能返回Tensor值，且带梯度，即 loss.requires_grad == True
1
2
3
4
5
6
7

示例代码

以一个pytorch求解线性回归的代码为例（参考：https://blog.csdn.net/weixin_35757704/article/details/117395205）：

import torch
import torch.nn as nn
import numpy as np
import os

os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"


def get_x_y():
    np.random.seed(0)
    x = np.random.randint(0, 50, 300)
    y_values = 2 * x + 21
    x = np.array(x, dtype=np.float32)
    y = np.array(y_values, dtype=np.float32)
    x = x.reshape(-1, 1)
    y = y.reshape(-1, 1)
    return x, y


class LinearRegressionModel(nn.Module):
    def __init__(self, input_dim, output_dim):
        super(LinearRegressionModel, self).__init__()
        self.linear = nn.Linear(input_dim, output_dim)  # 输入的个数，输出的个数

    def forward(self, x):
        out = self.linear(x)
        return out


if __name__ == '__main__':
    input_dim = 1
    output_dim = 1
    x_train, y_train = get_x_y()

    model = LinearRegressionModel(input_dim, output_dim)
    epochs = 1000  # 迭代次数
    optimizer = torch.optim.SGD(model.parameters(), lr=0.001)
    model_loss = nn.MSELoss() # 使用MSE作为loss
    # 开始训练模型
    for epoch in range(epochs):
        epoch += 1
        # 注意转行成tensor
        inputs = torch.from_numpy(x_train)
        labels = torch.from_numpy(y_train)
        # 梯度要清零每一次迭代
        optimizer.zero_grad()
        # 前向传播
        outputs: torch.Tensor = model(inputs)
        # 计算损失
        loss = model_loss(outputs, labels)
        # 返向传播
        loss.backward()
        # 更新权重参数
        optimizer.step()
        if epoch % 50 == 0:
            print('epoch {}, loss {}'.format(epoch, loss.item()))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

步骤1：添加自定义的类

我们就用自定义的写法来写与MSE相同的效果，MSE计算公式如下：

添加一个类

class CustomLoss(nn.Module):
    def __init__(self):
        super(CustomLoss, self).__init__()

    def forward(self, x, y):
        mse_loss = torch.mean(torch.pow((x - y), 2)) # x与y相减后平方，求均值即为MSE
        return mse_loss
1
2
3
4
5
6
7

步骤2：修改使用的loss函数

只需要把原始代码中的：

model_loss = nn.MSELoss() # 使用MSE作为loss
1

改为：

model_loss = CustomLoss()  # 自定义loss
1

即可

完整代码

import torch
import torch.nn as nn
import numpy as np
import os

os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"


def get_x_y():
    np.random.seed(0)
    x = np.random.randint(0, 50, 300)
    y_values = 2 * x + 21
    x = np.array(x, dtype=np.float32)
    y = np.array(y_values, dtype=np.float32)
    x = x.reshape(-1, 1)
    y = y.reshape(-1, 1)
    return x, y


class LinearRegressionModel(nn.Module):
    def __init__(self, input_dim, output_dim):
        super(LinearRegressionModel, self).__init__()
        self.linear = nn.Linear(input_dim, output_dim)  # 输入的个数，输出的个数

    def forward(self, x):
        out = self.linear(x)
        return out


class CustomLoss(nn.Module):
    def __init__(self):
        super(CustomLoss, self).__init__()

    def forward(self, x, y):
        mse_loss = torch.mean(torch.pow((x - y), 2))
        return mse_loss


if __name__ == '__main__':
    input_dim = 1
    output_dim = 1
    x_train, y_train = get_x_y()

    model = LinearRegressionModel(input_dim, output_dim)
    epochs = 1000  # 迭代次数
    optimizer = torch.optim.SGD(model.parameters(), lr=0.001)
    # model_loss = nn.MSELoss() # 使用MSE作为loss
    model_loss = CustomLoss()  # 自定义loss
    # 开始训练模型
    for epoch in range(epochs):
        epoch += 1
        # 注意转行成tensor
        inputs = torch.from_numpy(x_train)
        labels = torch.from_numpy(y_train)
        # 梯度要清零每一次迭代
        optimizer.zero_grad()
        # 前向传播
        outputs: torch.Tensor = model(inputs)
        # 计算损失
        loss = model_loss(outputs, labels)
        # 返向传播
        loss.backward()
        # 更新权重参数
        optimizer.step()
        if epoch % 50 == 0:
            print('epoch {}, loss {}'.format(epoch, loss.item()))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66