pytorch loss损失函数
呆萌的代Ma 人气:0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
示例代码：
以一个pytorch求解线性回归的代码为例：
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()))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__() self.mse_loss = nn.MSELoss() 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
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
改为：
model_loss = CustomLoss() # 自定义loss
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__()
self.mse_loss = nn.MSELoss()
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()))
加载全部内容