7.11用逻辑回归对Titanic数据进行训练
之前我们已经讲了逻辑回归的原理，以及如何在PyTorch里定义Dataset和DataLoader，如何在PyTorch里定义一个逻辑回归模型。这一节，我们就把它们组合起来，实现一个完整的在PyTorch里训练的代码。
7.11.1PyTorch里的优化器
之前我们手动实现多元线性回归时，我们自己对每个参数进行梯度计算，然后对每个参数根据学习率和梯度进行参数的更新。对于少量参数还可以，如果有几十亿参数需要你管理，那还是有不少工作的，另外我们之前学过的梯度更新算法还很简单，后边还有更负责的梯度更新策略。
为此PyTorch里专门提取出了优化器（Optimizer）类，它主要有以下的功能：
- 管理参数，对参数进行更新。
- 封装对参数进行梯度更新的算法。
- 提供在训练过程中管理学习率的接口。
我们之前学习的最基本的梯度下降的算法，对应PyTorch里的torch.optim.SGD优化器。我们之前讲SGD是随机梯度下降，每次只用一个样本的loss计算梯度，更新模型参数。这是严格的定义，但是PyTorch里的SGD优化器并没有要求只能利用一条数据计算loss。
使用SGD优化器的伪代码如下：
## 定义优化器，传入模型的参数，并且设置固定的学习率
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
for epoch in range(epochs):
for features,labels in DataLoader(myDataset, batch_size=256):
optimizer.zero_grad() ##优化器清理管理参数的梯度值。
loss = forward_and_compute_loss(features, labels) ##前向传播并计算loss
loss.backward() ##loss反向传播,每个参数计算梯度
optimizer.step() ##优化器进行参数更新
上边的代码中optimizer.zero_grad()你可能会疑惑，这里做一下解释。每个可训练的参数都有一个关联的梯度值。每次loss的后向传播时，新计算的梯度值都会累加在原来每个参数的梯度上。直到调用optimizer.zero_grad()，才会将这个优化器管理的参数对应的梯度值设置为0。
在每一个训练的step里：
- optimizer把自己管理参数的梯度值都置为0。
- 计算loss。
- loss反向传播，计算出每个参数的梯度值。
- optimizer根据每个参数的梯度值，和优化器算法，学习率来更新参数。
7.11.2数据集的划分
Titanic这份数据集源自Kaggle的竞赛数据。你下载的数据集里有train.csv, test.csv, gender_submission.csv，其中trian.csv是训练数据集，test.csv是测试数据集，其中只有feature，不包含Label。其中gender_submission.csv是一份示例的提交Kaggle网站的文件。其中的值并不是test.csv里样本的真实Label值。
train.csv里有891条数据，我们用前800条数据作为训练集，放入train.csv文件。后91条数据作为验证集，放入validation.csv文件，这里需要你手动构建validation数据集。
7.11.3Titanic逻辑回归代码
我们来完整实现一下对Titanic数据进行逻辑回归模型训练。
from torch.utils.data import Dataset, DataLoader
import torch
import torch.nn as nn
import pandas as pd
class LogisticRegressionModel(nn.Module):
def __init__(self, input_dim):
super().__init__()
self.linear = nn.Linear(input_dim, 1) # nn.Linear也继承自nn.Module，输入为input_dim,输出一个值
def forward(self, x):
return torch.sigmoid(self.linear(x)) # Logistic Regression 输出概率
class TitanicDataset(Dataset):
def __init__(self, file_path):
self.file_path = file_path
self.mean = {
"Pclass": 2.236695,
"Age": 29.699118,
"SibSp": 0.512605,
"Parch": 0.431373,
"Fare": 34.694514,
"Sex_female": 0.365546,
"Sex_male": 0.634454,
"Embarked_C": 0.182073,
"Embarked_Q": 0.039216,
"Embarked_S": 0.775910
}
self.std = {
"Pclass": 0.838250,
"Age": 14.526497,
"SibSp": 0.929783,
"Parch": 0.853289,
"Fare": 52.918930,
"Sex_female": 0.481921,
"Sex_male": 0.481921,
"Embarked_C": 0.386175,
"Embarked_Q": 0.194244,
"Embarked_S": 0.417274
}
self.data = self._load_data()
self.feature_size = len(self.data.columns) - 1
def _load_data(self):
df = pd.read_csv(self.file_path)
df = df.drop(columns=["PassengerId", "Name", "Ticket", "Cabin"]) ##删除不用的列
df = df.dropna(subset=["Age"])##删除Age有缺失的行
df = pd.get_dummies(df, columns=["Sex", "Embarked"], dtype=int)##进行one-hot编码
##进行数据的标准化
base_features = ["Pclass", "Age", "SibSp", "Parch", "Fare"]
for i in range(len(base_features)):
df[base_features[i]] = (df[base_features[i]] - self.mean[base_features[i]]) / self.std[base_features[i]]
return df
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
features = self.data.drop(columns=["Survived"]).iloc[idx].values
label = self.data["Survived"].iloc[idx]
return torch.tensor(features, dtype=torch.float32), torch.tensor(label, dtype=torch.float32)
train_dataset = TitanicDataset(r"E:\电子书\RethinkFun深度学习\data\titanic\train.csv")
validation_dataset = TitanicDataset(r"E:\电子书\RethinkFun深度学习\data\titanic\validation.csv")
model = LogisticRegressionModel(train_dataset.feature_size)
model.to("cuda")
model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
epochs = 100
for epoch in range(epochs):
correct = 0
step = 0
total_loss = 0
for features, labels in DataLoader(train_dataset, batch_size=256, shuffle=True):
step += 1
features = features.to("cuda")
labels = labels.to("cuda")
optimizer.zero_grad()
outputs = model(features).squeeze()
correct += torch.sum(((outputs >= 0.5) == labels))
loss = torch.nn.functional.binary_cross_entropy(outputs, labels)
total_loss += loss.item()
loss.backward()
optimizer.step()
print(f'Epoch {epoch + 1}, Loss: {total_loss/step:.4f}')
print(f'Training Accuracy: {correct / len(train_dataset)}')
model.eval()
with torch.no_grad():
correct = 0
for features, labels in DataLoader(validation_dataset, batch_size=256):
features = features.to("cuda")
labels = labels.to("cuda")
outputs = model(features).squeeze()
correct += torch.sum(((outputs >= 0.5) == labels))
print(f'Validation Accuracy: {correct / len(validation_dataset)}')经过训练，我们得到的结果在训练集上模型的正确率为0.80，在验证集上模型的正确率为0.83。可以看到我们模型目前没有过拟合，我们还可以尝试增加模型的复杂度，比如这里我们增加一些特征的二次项。
我们修改TitanicDataset代码如下：
class TitanicDataset(Dataset):
def __init__(self, file_path):
self.file_path = file_path
self.mean = {
"Pclass": 2.236695,
"Age": 29.699118,
"SibSp": 0.512605,
"Parch": 0.431373,
"Fare": 34.694514,
"Sex_female": 0.365546,
"Sex_male": 0.634454,
"Embarked_C": 0.182073,
"Embarked_Q": 0.039216,
"Embarked_S": 0.775910,
"Pclass_Pclass": 5.704482,
"Pclass_Age": 61.938151,
"Pclass_SibSp": 1.198880,
"Pclass_Parch": 0.983193,
"Pclass_Fare": 53.052327,
"Pclass_Sex_female": 0.754902,
"Age_Age": 1092.761169,
"Age_SibSp": 11.066415,
"Age_Parch": 10.470476,
"Age_Fare": 1104.142053,
"Age_Sex_female": 10.204482,
"SibSp_SibSp": 1.126050,
"SibSp_Parch": 0.525210,
"SibSp_Fare": 24.581262,
"SibSp_Sex_female": 0.233894,
"Parch_Parch": 0.913165,
"Parch_Fare": 24.215465,
"Parch_Sex_female": 0.259104,
"Fare_Fare": 4000.200255,
"Fare_Sex_female": 17.393698,
"Sex_female_Sex_female": 0.365546
}
self.std = {
"Pclass": 0.838250,
"Age": 14.526497,
"SibSp": 0.929783,
"Parch": 0.853289,
"Fare": 52.918930,
"Sex_female": 0.481921,
"Sex_male": 0.481921,
"Embarked_C": 0.386175,
"Embarked_Q": 0.194244,
"Embarked_S": 0.417274,
"Pclass_Pclass": 3.447593,
"Pclass_Age": 34.379609,
"Pclass_SibSp": 2.603741,
"Pclass_Parch": 2.236945,
"Pclass_Fare": 52.407209,
"Pclass_Sex_female": 1.118572,
"Age_Age": 991.079188,
"Age_SibSp": 19.093099,
"Age_Parch": 29.164503,
"Age_Fare": 1949.356185,
"Age_Sex_female": 15.924481,
"SibSp_SibSp": 3.428831,
"SibSp_Parch": 1.561298,
"SibSp_Fare": 70.185369,
"SibSp_Sex_female": 0.639885,
"Parch_Parch": 3.008314,
"Parch_Fare": 77.207321,
"Parch_Sex_female": 0.729143,
"Fare_Fare": 19105.110593,
"Fare_Sex_female": 43.568303,
"Sex_female_Sex_female": 0.481921
}
self.data = self._load_data()
self.feature_size = len(self.data.columns) - 1
def _load_data(self):
df = pd.read_csv(self.file_path)
df = df.drop(columns=["PassengerId", "Name", "Ticket", "Cabin"])
df = df.dropna(subset=["Age"])
df = pd.get_dummies(df, columns=["Sex", "Embarked"], dtype=int)
base_features = ["Pclass", "Age", "SibSp", "Parch", "Fare", "Sex_female"]
for i in range(len(base_features)):
for j in range(i, len(base_features)):
df[base_features[i] + "_" + base_features[j]] = ((df[base_features[i]] * df[base_features[j]]
- self.mean[
base_features[i] + "_" + base_features[j]])
/ self.std[base_features[i] + "_" + base_features[j]])
for i in range(len(base_features)):
df[base_features[i]] = (df[base_features[i]] - self.mean[base_features[i]]) / self.std[base_features[i]]
return df
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
features = self.data.drop(columns=["Survived"]).iloc[idx].values
label = self.data["Survived"].iloc[idx]
return torch.tensor(features, dtype=torch.float32), torch.tensor(label, dtype=torch.float32)我们在上边的代码里增加了一些特征的二次项，并对这些新的特征也进行了标准化，然后我们再次训练模型，查看模型在训练集和验证集上的表现。训练结果为模型在训练集上正确率为0.82，在验证集上正确率为0.87。可以看到模型的性能有提升，而且没有发生过拟合。这很好，你可以继续尝试进一步增加模型的复杂度，比如增加模型的三次项。