预测房价学习

1. 实现函数来方便下载数据

import hashlib
import os
import tarfile
import zipfile
import requests

DATA_HUB = dict()
DATA_URL = 'http://d2l-data.s3-accelerate.amazonaws.com/'

def download(name, cache_dir=os.path.join('..', 'data')):
    """下载一个DATA_HUB中的文件，返回本地文件名"""
    assert name in DATA_HUB, f"{name} 不存在于 {DATA_HUB}"
    url, sha1_hash = DATA_HUB[name]
    os.makedirs(cache_dir, exist_ok=True)
    fname = os.path.join(cache_dir, url.split('/')[-1])
    if os.path.exists(fname):
        sha1 = hashlib.sha1()
        with open(fname, 'rb') as f:
            while True:
                data = f.read(1048576)
                if not data:
                    break
                sha1.update(data)
        if sha1.hexdigest() == sha1_hash:
            return fname  # 命中缓存
    print(f'正在从{url}下载{fname}...')
    r = requests.get(url, stream=True, verify=True)
    with open(fname, 'wb') as f:
        f.write(r.content)
return fname

2.使用pandas读入并处理数据

%matplotlib inline
import numpy as np
import pandas as pd
import torch
from torch import nn
from d2l import torch as d2l

DATA_HUB['kaggle_house_train'] = (
    DATA_URL + 'kaggle_house_pred_train.csv',
    '585e9cc93e70b39160e7921475f9bcd7d31219ce')

DATA_HUB['kaggle_house_test'] = (
    DATA_URL + 'kaggle_house_pred_test.csv',
    'fa19780a7b011d9b009e8bff8e99922a8ee2eb90')

train_data = pd.read_csv(download('kaggle_house_train'))
test_data = pd.read_csv(download('kaggle_house_test'))

print(train_data.shape)
print(test_data.shape)

3.查看前四个和最后两个特征以及相应标签（房价）

print(train_data.iloc[0:4, [0, 1, 2, 3, -3, -2, -1]])

4. 在每个样本中，第一个特征是ID，有助于模型识别每个训练样本。 虽然这很方便，但它不携带任何用于预测的信息。因此，在将数据提供给模型之前，将其从数据集中删除。

all_features = pd.concat((train_data.iloc[:, 1:-1], test_data.iloc[:, 1:]))

5.将所有缺失的值替换为相应特征的平均值，通过将特征重新缩放到零均值和单位方差来标准化数据

numeric_features = all_features.dtypes[all_features.dtypes != 'object'].index
all_features[numeric_features] = all_features[numeric_features].apply(
    lambda x: (x - x.mean()) / (x.std()))
all_features[numeric_features] = all_features[numeric_features].fillna(0)

6.处理离散值，诸如“MSZoning”之类特征，用独热编码替换

all_features = pd.get_dummies(all_features, dummy_na=True)
all_features.shape

7.用价格预测的对数来衡量差异

def log_rmse(net, features, labels):
    # 为了在取对数时进一步稳定该值，将小于1的值设置为1
    clipped_preds = torch.clamp(net(features), 1, float('inf'))
    rmse = torch.sqrt(loss(torch.log(clipped_preds),
                           torch.log(labels)))
    return rmse.item()

8.训练函数借助Adam优化器（对初始学习率不那么敏感）

def train(net, train_features, train_labels, test_features, test_labels,
          num_epochs, learning_rate, weight_decay, batch_size):
    train_ls, test_ls = [], []
    train_iter = d2l.load_array((train_features, train_labels), batch_size)
    # 这里使用的是Adam优化算法
    optimizer = torch.optim.Adam(net.parameters(),
                                 lr = learning_rate,
                                 weight_decay = weight_decay)
    for epoch in range(num_epochs):
        for X, y in train_iter:
            optimizer.zero_grad()
            l = loss(net(X), y)
            l.backward()
            optimizer.step()
        train_ls.append(log_rmse(net, train_features, train_labels))
        if test_labels is not None:
            test_ls.append(log_rmse(net, test_features, test_labels))
    return train_ls, test_ls

9.K折交叉验证

def get_k_fold_data(k, i, X, y):
    assert k > 1
    fold_size = X.shape[0] // k
    X_train, y_train = None, None
    for j in range(k):
        idx = slice(j * fold_size, (j + 1) * fold_size)
        X_part, y_part = X[idx, :], y[idx]
        if j == i:
            X_valid, y_valid = X_part, y_part
        elif X_train is None:
            X_train, y_train = X_part, y_part
        else:
            X_train = torch.cat([X_train, X_part], 0)
            y_train = torch.cat([y_train, y_part], 0)
    return X_train, y_train, X_valid, y_valid

10. 在K折交叉验证中训练K次后，返回训练和验证误差的平均值

def k_fold(k, X_train, y_train, num_epochs, learning_rate, weight_decay,
           batch_size):
    train_l_sum, valid_l_sum = 0, 0
    for i in range(k):
        data = get_k_fold_data(k, i, X_train, y_train)
        net = get_net()
        train_ls, valid_ls = train(net, *data, num_epochs, learning_rate,
                                   weight_decay, batch_size)
        train_l_sum += train_ls[-1]
        valid_l_sum += valid_ls[-1]
        if i == 0:
            d2l.plot(list(range(1, num_epochs + 1)), [train_ls, valid_ls],
                     xlabel='epoch', ylabel='rmse', xlim=[1, num_epochs],
                     legend=['train', 'valid'], yscale='log')
        print(f'折{i + 1}，训练log rmse{float(train_ls[-1]):f}, '
              f'验证log rmse{float(valid_ls[-1]):f}')
    return train_l_sum / k, valid_l_sum / k