PyTorch实现Kaggle房价预测

  1 %matplotlib inline
  2 import torch
  3 import torch.nn as nn
  4 import numpy as np
  5 import pandas as pd
  6 import sys
  7 sys.path.append("/home/kesci/input")
  8 import d2lzh1981 as d2l
  9 print(torch.__version__)
 10 torch.set_default_tensor_type(torch.FloatTensor)
 11 
 12 
 13 test_data = pd.read_csv("/home/kesci/input/houseprices2807/house-prices-advanced-regression-techniques/test.csv")
 14 train_data = pd.read_csv("/home/kesci/input/houseprices2807/house-prices-advanced-regression-techniques/train.csv")
 15 
 16 all_features = pd.concat((train_data.iloc[:, 1:-1], test_data.iloc[:, 1:]))
 17 
 18 # 数据预处理
 19 numeric_features = all_features.dtypes[all_features.dtypes != 'object'].index
 20 all_features[numeric_features] = all_features[numeric_features].apply(
 21     lambda x: (x - x.mean()) / (x.std()))
 22 # 标准化后，每个数值特征的均值变为0，所以可以直接用0来替换缺失值
 23 all_features[numeric_features] = all_features[numeric_features].fillna(0)
 24 
 25 
 26 # dummy_na=True将缺失值也当作合法的特征值并为其创建指示特征
 27 all_features = pd.get_dummies(all_features, dummy_na=True)
 28 all_features.shape
 29 
 30 n_train = train_data.shape[0]
 31 train_features = torch.tensor(all_features[:n_train].values, dtype=torch.float)
 32 test_features = torch.tensor(all_features[n_train:].values, dtype=torch.float)
 33 train_labels = torch.tensor(train_data.SalePrice.values, dtype=torch.float).view(-1, 1)
 34 
 35 # 训练模型
 36 loss = torch.nn.MSELoss()
 37 
 38 def get_net(feature_num):
 39     net = nn.Linear(feature_num, 1)
 40     for param in net.parameters():
 41         nn.init.normal_(param, mean=0, std=0.01)
 42     return net
 43 
 44 
 45 def log_rmse(net, features, labels):
 46     with torch.no_grad():
 47         # 将小于1的值设成1，使得取对数时数值更稳定
 48         clipped_preds = torch.max(net(features), torch.tensor(1.0))
 49         rmse = torch.sqrt(2 * loss(clipped_preds.log(), labels.log()).mean())
 50     return rmse.item()
 51 
 52 
 53 def train(net, train_features, train_labels, test_features, test_labels,
 54           num_epochs, learning_rate, weight_decay, batch_size):
 55     train_ls, test_ls = [], []
 56     dataset = torch.utils.data.TensorDataset(train_features, train_labels)
 57     train_iter = torch.utils.data.DataLoader(dataset, batch_size, shuffle=True)
 58     # 这里使用了Adam优化算法
 59     optimizer = torch.optim.Adam(params=net.parameters(), lr=learning_rate, weight_decay=weight_decay) 
 60     net = net.float()
 61     for epoch in range(num_epochs):
 62         for X, y in train_iter:
 63             l = loss(net(X.float()), y.float())
 64             optimizer.zero_grad()
 65             l.backward()
 66             optimizer.step()
 67         train_ls.append(log_rmse(net, train_features, train_labels))
 68         if test_labels is not None:
 69             test_ls.append(log_rmse(net, test_features, test_labels))
 70     return train_ls, test_ls
 71 
 72 
 73 # K折交叉验证
 74 def get_k_fold_data(k, i, X, y):
 75     # 返回第i折交叉验证时所需要的训练和验证数据
 76     assert k > 1
 77     fold_size = X.shape[0] // k
 78     X_train, y_train = None, None
 79     for j in range(k):
 80         idx = slice(j * fold_size, (j + 1) * fold_size)
 81         X_part, y_part = X[idx, :], y[idx]
 82         if j == i:
 83             X_valid, y_valid = X_part, y_part
 84         elif X_train is None:
 85             X_train, y_train = X_part, y_part
 86         else:
 87             X_train = torch.cat((X_train, X_part), dim=0)
 88             y_train = torch.cat((y_train, y_part), dim=0)
 89     return X_train, y_train, X_valid, y_valid
 90 
 91 def k_fold(k, X_train, y_train, num_epochs,
 92            learning_rate, weight_decay, batch_size):
 93     train_l_sum, valid_l_sum = 0, 0
 94     for i in range(k):
 95         data = get_k_fold_data(k, i, X_train, y_train)
 96         net = get_net(X_train.shape[1])
 97         train_ls, valid_ls = train(net, *data, num_epochs, learning_rate,
 98                                    weight_decay, batch_size)
 99         train_l_sum += train_ls[-1]
100         valid_l_sum += valid_ls[-1]
101         if i == 0:
102             d2l.semilogy(range(1, num_epochs + 1), train_ls, 'epochs', 'rmse',
103                          range(1, num_epochs + 1), valid_ls,
104                          ['train', 'valid'])
105         print('fold %d, train rmse %f, valid rmse %f' % (i, train_ls[-1], valid_ls[-1]))
106     return train_l_sum / k, valid_l_sum / k
107 
108 # 模型选择
109 k, num_epochs, lr, weight_decay, batch_size = 5, 100, 5, 0, 64
110 train_l, valid_l = k_fold(k, train_features, train_labels, num_epochs, lr, weight_decay, batch_size)
111 print('%d-fold validation: avg train rmse %f, avg valid rmse %f' % (k, train_l, valid_l))
112 
113 
114 # 预测
115 def train_and_pred(train_features, test_features, train_labels, test_data,
116                    num_epochs, lr, weight_decay, batch_size):
117     net = get_net(train_features.shape[1])
118     train_ls, _ = train(net, train_features, train_labels, None, None,
119                         num_epochs, lr, weight_decay, batch_size)
120     d2l.semilogy(range(1, num_epochs + 1), train_ls, 'epochs', 'rmse')
121     print('train rmse %f' % train_ls[-1])
122     preds = net(test_features).detach().numpy()
123     test_data['SalePrice'] = pd.Series(preds.reshape(1, -1)[0])
124     submission = pd.concat([test_data['Id'], test_data['SalePrice']], axis=1)
125     submission.to_csv('./submission.csv', index=False)
126     # sample_submission_data = pd.read_csv("../input/house-prices-advanced-regression-techniques/sample_submission.csv")
127 
128 train_and_pred(train_features, test_features, train_labels, test_data, num_epochs, lr, weight_decay, batch_size)