PyTorch入门学习：5-Linear Regression with PyTorch

Linear Regression with Pytorch

1 PyTorch Fashion

1. Prepare dataset
2. Design model using Class: inherit from nn.Module
3. Construct loss and optimizer: using PyTorch API
4. Training cycle (forward, backward, update)

2 Prepare Dataset

3 Design Model

确认权重和张量的shape，如图中z的shape为[3,1]，x的shape为[4,1]，则w的shape为[3,4]，b的shape为[3,1]。

4 Training Cycle

import torch
# nn.Module是所有神经网络的基类，模型类应当从nn.Module继承
class LinearModel(torch.nn.Module):
    def __init__(self):
        # 调用父类的构造函数（必做）
        super(LinearModel, self).__init__()
        # nn.Linear是PyTorch的线性层类，参数为输入特征数和输出特征数
        self.linear = torch.nn.Linear(1, 1)

    def forward(self, x):
        # Torch中nn.Linear是可调用对象，可调用对象__call__会自动调用forward方法
        y_pred = self.linear(x)
        return y_pred

model = LinearModel()
# 构造损失函数：均方误差损失函数nn.MSELoss()
# 参数size_average=False表示不进行平均，返回的是一个标量
criterion = torch.nn.MSELoss(size_average=False)
# 构造优化器：随机梯度下降优化器SGD()
# 参数model.parameters()自动获取模型的所有参数（权重和偏置），lr表示学习率
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)

for epoch in range(1000):
    # 前向传播：计算模型输出y_pred
    y_pred = model(x_data)
    # 计算损失：损失函数参数为模型输出和真实标签
    loss = criterion(y_pred, y_data)
    print(epoch, loss.item())

    # 反向传播：计算梯度
    loss.backward()
    # 更新参数：优化器根据梯度更新参数
    optimizer.step()
    # 清空梯度：优化器更新参数后，需要手动将梯度置零
    optimizer.zero_grad()

print(model.linear.weight.item())
print(model.linear.bias.item())

5 Exercize

5.1 Try different optimizer

• Adagrad
• Adam
• Adamax
• RMSprop
• ASGD
• LBFGS
• Rprop
• SGD

5.2 Read more example from official tutorial