Kaggle实战：Pytorch实现房价预测
近来，我一直在学习pytorch与深度学习的理论知识，但一直苦于无法深入地理解某些理论及其背后的意义，并且很难从0开始用pytorch搭建一深度学习网络来解决一个实际问题。直到偶然接触了《动手学深度学习》这本书，我感觉收获颇丰。
这本书其中一章节是讲实战Kaggle比赛：预测房价，其中涵盖非常丰富的知识，为此我将整个实现过程记录如下，不足之处还请大家批评指正。
一、问题背景
Ask a home buyer to describe their dream house, and they probably won’t begin with the height of the basement ceiling or the proximity to an east-west railroad. But this playground competition’s dataset proves that much more influences price negotiations than the number of bedrooms or a white-picket fence.
With 79 explanatory variables describing (almost) every aspect of residential homes in Ames, Iowa, this competition challenges you to predict the final price of each home.
Kaggle网址：https://www.kaggle.com/competitions/house-prices-advanced-regression-techniques
数据集下载：
链接：https://pan.baidu.com/s/126T6iPUDNRAU-pO8KJyOaA
提取码：11oc
二、数据处理
2.1 导入包
import numpy as np
import pandas as pd
import torch
from torch import nn
from d2l import torch as d2l
import seaborn as sns
from matplotlib import pyplot as plt
from scipy.stats import norm, skew
2.2 读取CSV文件
train_data = pd.read_csv("train.csv") # [1460 rows x 81 columns]
test_data = pd.read_csv("test.csv") # [1459 rows x 80 columns]
train_data.head()
Id MSSubClass MSZoning ... SaleType SaleCondition SalePrice
0 1 60 RL ... WD Normal 208500
1 2 20 RL ... WD Normal 181500
2 3 60 RL ... WD Normal 223500
3 4 70 RL ... WD Abnorml 140000
4 5 60 RL ... WD Normal 250000
train_data.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1460 entries, 0 to 1459
Data columns (total 81 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Id 1460 non-null int64
1 MSSubClass 1460 non-null int64
2 MSZoning 1460 non-null object
3 LotFrontage 1201 non-null float64
4 LotArea 1460 non-null int64
5 Street 1460 non-null object
6 Alley 91 non-null object
7 LotShape 1460 non-null object
8 LandContour 1460 non-null object
9 Utilities 1460 non-null object
10 LotConfig 1460 non-null object
11 LandSlope 1460 non-null object
12 Neighborhood 1460 non-null object
13 Condition1 1460 non-null object
14 Condition2 1460 non-null object
15 BldgType 1460 non-null object
16 HouseStyle 1460 non-null object
17 OverallQual 1460 non-null int64
18 OverallCond 1460 non-null int64
19 YearBuilt 1460 non-null int64
20 YearRemodAdd 1460 non-null int64
21 RoofStyle 1460 non-null object
22 RoofMatl 1460 non-null object
23 Exterior1st 1460 non-null object
24 Exterior2nd 1460 non-null object
25 MasVnrType 1452 non-null object
26 MasVnrArea 1452 non-null float64
27 ExterQual 1460 non-null object
28 ExterCond 1460 non-null object
29 Foundation 1460 non-null object
30 BsmtQual 1423 non-null object
31 BsmtCond 1423 non-null object
32 BsmtExposure 1422 non-null object
33 BsmtFinType1 1423 non-null object
34 BsmtFinSF1 1460 non-null int64
35 BsmtFinType2 1422 non-null object
36 BsmtFinSF2 1460 non-null int64
37 BsmtUnfSF 1460 non-null int64
38 TotalBsmtSF 1460 non-null int64
39 Heating 1460 non-null object
40 HeatingQC 1460 non-null object
41 CentralAir 1460 non-null object
42 Electrical 1459 non-null object
43 1stFlrSF 1460 non-null int64
44 2ndFlrSF 1460 non-null int64
45 LowQualFinSF 1460 non-null int64
46 GrLivArea 1460 non-null int64
47 BsmtFullBath 1460 non-null int64
48 BsmtHalfBath 1460 non-null int64
49 FullBath 1460 non-null int64
50 HalfBath 1460 non-null int64
51 BedroomAbvGr 1460 non-null int64
52 KitchenAbvGr 1460 non-null int64
53 KitchenQual 1460 non-null object
54 TotRmsAbvGrd 1460 non-null int64
55 Functional 1460 non-null object
56 Fireplaces 1460 non-null int64
57 FireplaceQu 770 non-null object
58 GarageType 1379 non-null object
59 GarageYrBlt 1379 non-null float64
60 GarageFinish 1379 non-null object
61 GarageCars 1460 non-null int64
62 GarageArea 1460 non-null int64
63 GarageQual 1379 non-null object
64 GarageCond 1379 non-null object
65 PavedDrive 1460 non-null object
66 WoodDeckSF 1460 non-null int64
67 OpenPorchSF 1460 non-null int64
68 EnclosedPorch 1460 non-null int64
69 3SsnPorch 1460 non-null int64
70 ScreenPorch 1460 non-null int64
71 PoolArea 1460 non-null int64
72 PoolQC 7 non-null object
73 Fence 281 non-null object
74 MiscFeature 54 non-null object
75 MiscVal 1460 non-null int64
76 MoSold 1460 non-null int64
77 YrSold 1460 non-null int64
78 SaleType 1460 non-null object
79 SaleCondition 1460 non-null object
80 SalePrice 1460 non-null int64
dtypes: float64(3), int64(35), object(43)
memory usage: 924.0+ KB
None
2.3 数据分析
分析房价
我们的目标列是房价‘SalePrice’，因此我们先对其进行分析
train_data[‘SalePrice’].describe()
count 1460.000000
mean 180921.195890
std 79442.502883
min 34900.000000
25% 129975.000000
50% 163000.000000
75% 214000.000000
max 755000.000000
Name: SalePrice, dtype: float64
sns.distplot(train_data[‘SalePrice’])
由上图可知，SalePrice呈现正的偏度，我们需对其进行修正
print("Skewness: %f" % train_data['SalePrice'].skew())
print("Kurtosis: %f" % train_data['SalePrice'].kurt())
Skewness: 1.882876
Kurtosis: 6.536282
train_data['SalePrice'] = np.log1p(train_data['SalePrice'])
sns.distplot(train_data['SalePrice'], fit=norm)
研究部分因素对房价的影响
sns.boxplot(data=train_data, x=‘OverallQual’, y=‘SalePrice’)
plt.figure(figsize=(12,4),dpi=130)
sns.boxplot(data=train_data, x='YearBuilt', y='SalePrice')
plt.xticks(rotation=90)
plt.show()
sns.scatterplot(data=train,x=‘TotalBsmtSF’,y=‘SalePrice’)
sns.scatterplot(data=train_data, x=‘GrLivArea’, y=‘SalePrice’)
相关性分析
print(train_data.corr()[“SalePrice”].sort_values(ascending=False)[1:])
OverallQual 0.817185
GrLivArea 0.700927
GarageCars 0.680625
GarageArea 0.650888
TotalBsmtSF 0.612134
1stFlrSF 0.596981
FullBath 0.594771
YearBuilt 0.586570
YearRemodAdd 0.565608
GarageYrBlt 0.541073
TotRmsAbvGrd 0.534422
Fireplaces 0.489450
MasVnrArea 0.430809
BsmtFinSF1 0.372023
LotFrontage 0.355879
WoodDeckSF 0.334135
OpenPorchSF 0.321053
2ndFlrSF 0.319300
HalfBath 0.313982
LotArea 0.257320
BsmtFullBath 0.236224
BsmtUnfSF 0.221985
BedroomAbvGr 0.209043
ScreenPorch 0.121208
PoolArea 0.069798
MoSold 0.057330
3SsnPorch 0.054900
BsmtFinSF2 0.004832
BsmtHalfBath -0.005149
Id -0.017942
MiscVal -0.020021
OverallCond -0.036868
YrSold -0.037263
LowQualFinSF -0.037963
MSSubClass -0.073959
KitchenAbvGr -0.147548
EnclosedPorch -0.149050
Name: SalePrice, dtype: float64
corrmat = train_data.corr()
f, ax = plt.subplots(figsize=(12, 9))
sns.heatmap(corrmat, vmax=.8, square=True)
plt.show()
corr = train_data.corr()
highest_corr_features = corr.index[abs(corr["SalePrice"])>0.5]
plt.figure(figsize=(10, 10))
g = sns.heatmap(train_data[highest_corr_features].corr(), annot=True, cmap="RdYlGn")
plt.show()
从以上两张热力图，我们可得知：
- ‘OverQual’ in the top of highest correlation it’s 0.79!
- ‘GarageCars’ & ‘GarageArea’ like each other (correlation between them is 0.88)
- ‘TotalBsmtSF’ & ‘1stFlrSF’ also like each other (correlation betwwen them is 0.82), so we can keep either one of them or add the1stFlrSF to the Toltal.
- ‘TotRmsAbvGrd’ & ‘GrLivArea’ also has a strong correlation (0.83), I decided to keep ‘GrLivArea’ because it’s correlation with ‘SalePrice’ is higher.
接下来我们分析与房价相关性高的特征
cols = ['SalePrice', 'OverallQual', 'GrLivArea', 'GarageCars', 'TotalBsmtSF', 'FullBath', 'YearBuilt']
sns.pairplot(train_data[cols])
plt.show()
2.4 处理缺失数据
y_train = train_data['SalePrice']
test_id = test_data['Id']
all_data = pd.concat([train_data, test_data], axis=0, sort=False)
all_data = all_data.drop(['Id', 'SalePrice'], axis=1)
Total = all_data.isnull().sum().sort_values(ascending=False)
percent = (all_data.isnull().sum() / all_data.isnull().count() * 100).sort_values(ascending=False)
missing_data = pd.concat([Total, percent], axis=1, keys=['Total', 'Percent'])
print(missing_data.head(25))
Total Percent
PoolQC 2909 99.657417
MiscFeature 2814 96.402878
Alley 2721 93.216855
Fence 2348 80.438506
FireplaceQu 1420 48.646797
LotFrontage 486 16.649538
GarageYrBlt 159 5.447071
GarageFinish 159 5.447071
GarageQual 159 5.447071
GarageCond 159 5.447071
GarageType 157 5.378554
BsmtExposure 82 2.809181
BsmtCond 82 2.809181
BsmtQual 81 2.774923
BsmtFinType2 80 2.740665
BsmtFinType1 79 2.706406
MasVnrType 24 0.822199
MasVnrArea 23 0.787941
MSZoning 4 0.137033
Functional 2 0.068517
BsmtHalfBath 2 0.068517
BsmtFullBath 2 0.068517
Utilities 2 0.068517
SaleType 1 0.034258
BsmtFinSF1 1 0.034258
由上述结果我们发现，数据缺失率高的特征同时也是不重要的特征（correlation < 0.5），因此删除这些特征所在的列不会影响模型的预测。
all_data.drop((missing_data[missing_data['Total'] > 5]).index, axis=1, inplace=True)
print(all_data.isnull().sum().max())
4
查看剩余的包含缺失数据的特征
total = all_data.isnull().sum().sort_values(ascending=False)
total.head(19)
MSZoning 4
Functional 2
BsmtFullBath 2
BsmtHalfBath 2
Utilities 2
BsmtFinSF2 1
Exterior2nd 1
GarageCars 1
GarageArea 1
BsmtFinSF1 1
BsmtUnfSF 1
Exterior1st 1
TotalBsmtSF 1
Electrical 1
SaleType 1
KitchenQual 1
HalfBath 0
FullBath 0
BedroomAbvGr 0
dtype: int64
填充缺失数据
填充numeric型数据
numeric_missed = ['BsmtFinSF1',
'BsmtFinSF2',
'BsmtUnfSF',
'TotalBsmtSF',
'BsmtFullBath',
'BsmtHalfBath',
'GarageArea',
'GarageCars']
for feature in numeric_missed:
all_data[feature] = all_data[feature].fillna(0)
填充categorical型数据
categorical_missed = ['Exterior1st',
'Exterior2nd',
'SaleType',
'MSZoning',
'Electrical',
'KitchenQual']
for feature in categorical_missed:
all_data[feature] = all_data[feature].fillna(all_data[feature].mode()[0])
# Fill in the remaining missing values with the values that are most common for this feature.
all_data['Functional'] = all_data['Functional'].fillna('Typ')
查看是否还有缺失值
print(all_data.isnull().sum().max())
0
2.5 修正特征偏度
查看特征偏度
numeric_feats = all_data.dtypes[all_data.dtypes != 'object'].index
skewed_feats = all_data[numeric_feats].apply(lambda x: skew(x)).sort_values(ascending=False)
high_skew = skewed_feats[abs(skewed_feats) > 0.5]
print(high_skew)
MiscVal 21.947195
PoolArea 16.898328
LotArea 12.822431
LowQualFinSF 12.088761
3SsnPorch 11.376065
KitchenAbvGr 4.302254
BsmtFinSF2 4.146143
EnclosedPorch 4.003891
ScreenPorch 3.946694
BsmtHalfBath 3.931594
OpenPorchSF 2.535114
WoodDeckSF 1.842433
1stFlrSF 1.469604
BsmtFinSF1 1.425230
MSSubClass 1.375457
GrLivArea 1.269358
TotalBsmtSF 1.156894
BsmtUnfSF 0.919339
2ndFlrSF 0.861675
TotRmsAbvGrd 0.758367
Fireplaces 0.733495
HalfBath 0.694566
BsmtFullBath 0.624832
OverallCond 0.570312
YearBuilt -0.599806
dtype: float64
修正偏度
for feature in high_skew.index:
all_data[feature] = np.log1p(all_data[feature])
添加一个新特征
all_data['TotalSF'] = all_data['TotalBsmtSF'] + all_data['1stFlrSF'] + all_data['2ndFlrSF']
2.6 将类别型特征独热编码
all_data = pd.get_dummies(all_data)
all_data.head()
MSSubClass LotArea ... SaleCondition_Normal SaleCondition_Partial
0 4.110874 9.042040 ... 1 0
1 3.044522 9.169623 ... 1 0
2 4.110874 9.328212 ... 1 0
3 4.262680 9.164401 ... 0 0
4 4.110874 9.565284 ... 1 0
最后得到的训练集和测试集：
x_train = all_data[:len(y_train)] # (1460, 219)
x_test = all_data[len(y_train):] # (1459, 219)
转为tensor：
train_features = torch.tensor(x_train.values, dtype=torch.float32) # torch.Size([1460, 219])
test_features = torch.tensor(x_test.values, dtype=torch.float32) # torch.Size([1459, 219])
train_labels = torch.tensor(
train_data.SalePrice.values.reshape(-1, 1), dtype=torch.float32) # torch.Size([1460, 1])
三、模型建立
3.1 定义网络模型
首先，我们训练一个带有损失平方的线性模型。 显然线性模型很难让我们在竞赛中获胜，但线性模型提供了一种健全性检查， 以查看数据中是否存在有意义的信息。 如果我们在这里不能做得比随机猜测更好，那么我们很可能存在数据处理错误。 如果一切顺利，线性模型将作为基线（baseline）模型， 让我们直观地知道最好的模型有超出简单的模型多少。
loss = nn.MSELoss()
in_features = train_features.shape[1]
def get_net():
net = nn.Sequential(nn.Linear(in_features, 1))
# 模型参数初始化
for param in net.parameters():
nn.init.normal_(param, mean=0, std=0.01)
return net
3.2 定义训练函数
我们的训练函数将借助Adam优化器 ,Adam优化器的主要吸引力在于它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(loss(net(train_features), train_labels).item())
if test_labels is not None:
test_ls.append(loss(net(test_features), test_labels).item())
return train_ls, test_ls
四、K折交叉验证
K折交叉验证有助于模型选择和超参数调整。 我们首先需要定义一个函数，在K折交叉验证过程中返回第i折的数据。 具体地说，它选择第i个切片作为验证数据，其余部分作为训练数据。 注意，这并不是处理数据的最有效方法，如果我们的数据集大得多，会有其他解决办法。
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
当我们在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
五、模型选择
找到一组调优的超参数可能需要时间，这取决于我们优化了多少变量。 有了足够大的数据集和合理设置的超参数，K折交叉验证往往对多次测试具有相当的稳定性。 然而，如果我们尝试了不合理的超参数，我们可能会发现验证效果不再代表真正的误差。
k, num_epochs, lr, weight_decay, batch_size = 5, 200, 0.0002, 0, 64
train_l, valid_l = k_fold(k, train_features, train_labels, num_epochs, lr,
weight_decay, batch_size)
print(f'{k}-折验证: 平均训练log rmse: {float(train_l):f}, '
f'平均验证log rmse: {float(valid_l):f}')
d2l.plt.show()
折1，训练log rmse0.010107, 验证log rmse0.015819
折2，训练log rmse0.008903, 验证log rmse0.022045
折3，训练log rmse0.008702, 验证log rmse0.019592
折4，训练log rmse0.009519, 验证log rmse0.014886
折5，训练log rmse0.009213, 验证log rmse0.026204
5-折验证: 平均训练log rmse: 0.009289, 平均验证log rmse: 0.019709
请注意，有时一组超参数的训练误差可能非常低，但折交叉验证的误差要高得多， 这表明模型过拟合了。 在整个训练过程中，你将希望监控训练误差和验证误差这两个数字。 较少的过拟合可能表明现有数据可以支撑一个更强大的模型， 较大的过拟合可能意味着我们可以通过正则化技术来获益。
六、房价预测
当我们知道应该选择什么样的超参数， 我们不妨使用所有数据对其进行训练 （而不是仅使用交叉验证中使用的的数据）。 然后，我们通过这种方式获得的模型可以应用于测试集。 将预测保存在CSV文件中可以简化将结果上传到Kaggle的过程。
def train_and_pred(train_features, test_features, train_labels, test_data,
num_epochs, lr, weight_decay, batch_size):
net = get_net()
train_ls, _ = train(net, train_features, train_labels, None, None,
num_epochs, lr, weight_decay, batch_size)
d2l.plot(np.arange(1, num_epochs + 1), [train_ls], xlabel='epoch',
ylabel='log rmse', xlim=[1, num_epochs], yscale='log')
print(f'训练log rmse：{float(train_ls[-1]):f}')
# 将网络应用于测试集。
preds = net(test_features).detach().numpy()
preds = numpy.exp(preds) - 1
# 将其重新格式化以导出到Kaggle
test_data['SalePrice'] = pd.Series(preds.reshape(1, -1)[0])
submission = pd.concat([test_data['Id'], test_data['SalePrice']], axis=1)
submission.to_csv('submission.csv', index=False)
如果测试集上的预测与K倍交叉验证过程中的预测相似， 那就是时候把它们上传到Kaggle了。 下面的代码将生成一个名为submission.csv的文件。
train_and_pred(train_features, test_features, train_labels, test_data,
num_epochs, lr, weight_decay, batch_size)
d2l.plt.show()
训练log rmse：0.009800
将submission.csv文件提交到Kaggle网址，得到最终的分数与排名
博主目前还是一位AI领域的小白，预测的效果不是很好，大佬勿喷
代码已上传至github：https://github.com/Ming7959/Kaggle
2376
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
