PyTorch
- I used pure manual linear regression before. The problem of linear regression is actually to find the value of w when the loss function is minimized.
- Many functions in PyTorch are encapsulated, and we can use them directly.
loss function
Handwritten loss function
def loss(y, y_pred):
"""损失函数"""
# (真实值 - 预测值)^2 的平均值
return ((y_pred - y)**2).mean()
Loss function encapsulated by PyTorch
损失函数就是计算 (预测值-真实值)^2的平均值
(均方误差)
torch.nn.MSELoss()方法参数:
- Predicted values for all data
- all true values
- (The same goes for swapping positions, because the mean square error is calculated)
import torch
X = torch.tensor([1,2,3,4],dtype=torch.float32)
Y = torch.tensor([2,4,6,8],dtype=torch.float32)
w = torch.tensor(0.0,dtype=torch.float32,requires_grad=True)
def forward(x):
return w * x
# 均方差计算预测值和真实值之间的距离
loss = torch.nn.MSELoss()
# 计算此时的损失
y_pre = forward(X)
l = loss(y_pre,Y)
print(f"此时的损失值:{
l}")
optimizer
The optimizer replaces the handwritten gradient descent algorithm. PyTorch will automatically calculate the gradient and update the parameters.
Define optimizer
optim模块内包含多个优化器,都是基于基本的梯度下降算法改进的算法,可以更快的求出最优的参数解,例如: SGD, Adam,Momentum,RMSProp,这里使用的是SGD梯度下降算法
- Parameters that need to be updated through backpropagation. There may be multiple parameters that need to be updated, so parameter 1 is a list.
lr
The named parameter is learning rate
optimizer = torch.optim.SGD([W], lr=learning_rate)
Complete linear regression
l.backward()损失函数反向传播计算梯度
optimizer.step() 通过优化器 更新w参数 向梯度的方向走一步 (求解loss损失值关于w的偏导值)
optimizer.zero_grad() 清空梯度计算,防止梯度累加导致结果错误
# 创建x、y数据和自定义w参数
X = torch.tensor([1,2,3,4],dtype=torch.float32)
Y = torch.tensor([2,4,6,8],dtype=torch.float32)
w = torch.tensor(0.0,dtype=torch.float32,requires_grad=True)
# 定义学习率 和 训练模型迭代次数
learning_rate = 0.001
n_iters = 1000
# 创建损失函数
loss = torch.nn.MSELoss()
# 创建优化器 把w参数扔进去,要计算损失值关于w的关系(偏导数) 把学习率扔进去
optimizer = torch.optim.SGD([w],lr=learning_rate)
# 正向传播函数(模型)
def forward(x):
"""正向传播函数"""
return w * x
# 训练模型
for epoch in range(n_iters):
# 通过正向传播获取到预测值
y_pred = forward(X)
# 通过损失函数获取到损失值
l = loss(y_pred,Y)
# 反向传播计算梯度
l.backward()
# 通过优化器 更新w参数 向梯度的方向走一步
optimizer.step()
# 清空梯度计算,防止梯度累加导致结果错误
optimizer.zero_grad()
if epoch % 100 == 0:
# 打印纪元 w参数的变化和损失值的变化
print(f'epoch: {
epoch},w: {
w},loss: {
l:.8f}')
Model building
Building the model eliminates the need for handwritten forward propagation functions.
torch.nn.Linear(input_size,output_size)
Function representing a linear model
- input_size: dimension of input data
- output_size: dimension of output data
model.parameters()模型中的参数
# 创建x、y数据和自定义w参数
X = torch.tensor([[1],[2],[3],[4]],dtype=torch.float32)
Y = torch.tensor([[2],[4],[6],[8]],dtype=torch.float32)
# 测试集
X_test = torch.tensor([5],dtype=torch.float32)
# 定义模型参数
n_samples,n_features = X.shape
print(n_features)
# 因为当前定义的输入和输出维度一致 所以参数输入和输出都为n_features
model = torch.nn.Linear(n_features,n_features)
# 定义学习率 和 训练模型迭代次数
learning_rate = 0.01
n_iters = 1000
# 创建损失函数
loss = torch.nn.MSELoss()
# 创建优化器 把模型需要更新参数扔进去,要计算损失值关于w的关系(偏导数) 把学习率扔进去
optimizer = torch.optim.SGD(model.parameters(),lr=learning_rate)
# 训练模型
for epoch in range(n_iters):
# 通过正向传播获取到预测值
y_pred = model(X)
# 通过损失函数获取到损失值
l = loss(y_pred,Y)
# 反向传播计算梯度
l.backward()
# 通过优化器 更新w参数 向梯度的方向走一步
optimizer.step()
# 清空梯度计算,防止梯度累加导致结果错误
optimizer.zero_grad()
if epoch % 100 == 0:
# 获取到模型中w参数的值和b的值
w,b = model.parameters()
# 打印纪元 w参数的变化和损失值的变化
print(f'epoch: {
epoch},w: {
w[0,0].item()},loss: {
l}')
After training is completed, use the test set to test the w parameter:
test_model = model(X_test)
y = 2x + b
Data 5 is infinitely close to 10
Summarize:
PyTorch training model process:
- Get the training set, get the input dimensions and output dimensions
- Create an appropriate model based on input and output dimensions
- Create loss function
- Create optimizer
- Training model forward propagation ➡ back propagation ➡ gradient descent ➡ loop
- Get the test set to evaluate the model
Machine learning is not about finding the solution at once, but a process of slowly approaching the solution.