线性回归¶

Note

我们以线性回归为例，来展示使用pytorch训练模型的基本步骤:

载入数据
定义模型、损失函数和优化器
训练

载入数据¶

from sklearn.datasets import load_boston

# 载入boston房价数据集
X, y = load_boston(return_X_y=True)
X.shape, y.shape

((506, 13), (506,))

from sklearn.preprocessing import StandardScaler
import numpy as  np

# 数据标准化: z = (x - u) / s
# 为什么要标准化在《机器学习手册》中有详解，简而言之就是为了保持各个特征之间量纲一致，避免z字形梯度下降
X_scaled = StandardScaler().fit_transform(X)

import torch
from torch.utils import data

# ndarray转化成tensor
# 需使用float()将dtype转化为torch.float32
# target size should be: (n_samples, 1)
features = torch.from_numpy(X_scaled).float()
labels = torch.from_numpy(y).float().reshape(-1, 1)

#@save
def load_array(tensors, batch_size, is_train=True):
    """使用tensors创建data iter"""
    dataset = data.TensorDataset(*tensors)
    return data.DataLoader(dataset, batch_size, shuffle=is_train)

batch_size = 16
data_iter = load_array((features, labels), batch_size)

Tip

那些会复用的函数或类的开头会打上#@save标记，它们会保存在d2l.py中，这样其他地方若需要使用只需：import d2l

定义三要素¶

jupyter

from torch import nn

# 定义模型
model = nn.Sequential(nn.Linear(features.shape[1], 1))
# 定义损失函数
loss_fn = nn.MSELoss()
# 定义优化器，需指定要学习的参数和学习率
optimizer = torch.optim.SGD(model.parameters(), lr=0.03)

训练¶

num_epochs = 5
for epoch in range(num_epochs):
    # 循环获取数据
    for X, y in data_iter:
        # Compute prediction error
        y_hat = model(X)
        loss = loss_fn(y_hat, y)
        
        # Backpropagation
        # 1.梯度清0
        optimizer.zero_grad()
        # 2.自动求导
        loss.backward()
        # 3.更新参数
        optimizer.step()
    print(f'epoch {epoch + 1}, loss {loss_fn(model(features), labels):.4f}')

epoch 1, loss 33.7665
epoch 2, loss 26.7241
epoch 3, loss 23.9003
epoch 4, loss 22.6301
epoch 5, loss 23.1965

自动求导 Fashion-MNIST数据集