机器学习：线性模型学习总结（3）：基于PyTorch的线性模型

基于周志华老师的《机器学习》、上一篇学习笔记以及网络的其他资料，对线性模型的这一部分内容进行一个总结。上接：机器学习：线性模型学习总结（2）。

学习时间：2022.04.19~2022.04.20

1. 数据预处理

和用Sk-Learn一样，也用来一个专门处理表格数据的函数，不过主要还是使用之前的流程：

    # 数据预处理
    df_x = mango_processing(df_x).astype(float)
    df_y = df_y.astype(float)

    # 划分训练集和测试集
    tr_x, te_x, tr_y, te_y = train_test_split(df_x, df_y, test_size=0.2, random_state=42)

    # 全部转换成张量
    train_tensor_x, test_tensor_x, train_tensor_y, test_tensor_y = map(torch.tensor, (np.array(tr_x), np.array(te_x), np.array(tr_y), np.array(te_y)))

    # 将标签转为long或float格式（根据损失函数定）:
    # train_tensor_y = train_tensor_y.squeeze(-1).long()
    # test_tensor_y = test_tensor_y.squeeze(-1).long()
    train_tensor_y = train_tensor_y.squeeze(-1).float()
    test_tensor_y = test_tensor_y.squeeze(-1).float()

    # 返回测试集和训练集
    return train_tensor_x, test_tensor_x, train_tensor_y, test_tensor_y

2. PyTorch线性回归

因为全连接层不加激活函数，就只相当于加权求和，所以就可以当做线性回归：

class LinearModel(nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.liner = nn.Linear(14, 1)

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.liner(x)
        x = x.squeeze(-1)  # 线性回归的损失函数MSELoss要求输入维数和目标维数一致，因此做了个降维
        return x

3. PyTorch线性分类

在最后加一个Sigmoid函数实现逻辑回归分类：

class LinearModel(nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.liner = torch.nn.Linear(24, 2)

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.liner(x)
        x = torch.sigmoid(x)
        return x

4. 回归评价

这里学习了TorchMetrics这个包，直接用来调用评价回归结果，老样子，还是先写了一个函数：

    # 计算均方误差MSE
    mean_squared_error = torchmetrics.MeanSquaredError()
    mean_squared_error(y_pred, y_true)
    mse = mean_squared_error.compute()
    print('MSE:', mse, end='; ')

    # 计算平均绝对误差MAE
    mean_absolute_error = torchmetrics.MeanAbsoluteError()
    mean_absolute_error(y_pred, y_true)
    mae = mean_absolute_error.compute()
    print('MAE:', mae, end='; ')

    # 计算平均绝对百分比误差MAPE
    mean_absolute_percentage_error = torchmetrics.MeanAbsolutePercentageError()
    mean_absolute_percentage_error(y_pred, y_true)
    mape = mean_absolute_percentage_error.compute()
    print('MAPE:', mape, end='; ')

    # 计算可解释方差EV
    explained_variance = torchmetrics.ExplainedVariance()
    explained_variance(y_pred, y_true)
    ev = explained_variance.compute()
    print('EV:', ev, end='; ')

    # 计算可解释方差EV
    r2_score = torchmetrics.R2Score()
    r2_score(y_pred, y_true)
    r2 = r2_score.compute()
    print('R2-Score:', r2, end='.')

5. 分类评价

同上：

    # 计算准确率Accuracy
    accuracy = torchmetrics.Accuracy()
    accuracy(y_pred, y_true)
    acc = accuracy.compute()
    print('Accuracy:', acc, end='; ')

    # 计算精度precision
    precision = torchmetrics.Precision(average='macro', num_classes=calss_num)  # 需要根据预测的类别数量设定
    precision(y_pred, y_true)
    pre = precision.compute()
    print('Precision:', pre, end='; ')

    # 计算召回率recall
    recall = torchmetrics.Recall(average='macro', num_classes=calss_num)  # 需要根据预测的类别数量设定
    recall(y_pred, y_true)
    rec = recall.compute()
    print('Recall:', rec, end='; ')

    # 计算fl-score
    f1_score = torchmetrics.F1Score(num_classes=calss_num)
    f1_score(y_pred, y_true)
    f1 = f1_score.compute()
    print('F1-Score:', f1, end='; ')

    # 计算AUROC
    auroc = torchmetrics.AUROC(average='macro', num_classes=calss_num)
    auroc(y_pred, y_true)
    auc = auroc.compute()
    print('AUROC:', auc, end='.')
    auroc.reset()

6. 完整过程

6.1 线性回归预测

数据来源：New York City Taxi Fare Prediction | Kaggle。

# 读取数据集，做好预处理
df = pd.read_csv('train.csv')
df.pickup_datetime = pd.to_datetime(df.pickup_datetime).dt.tz_localize(None)
df['hour'] = df['pickup_datetime'].apply(lambda x: x.strftime('%H')).astype(int)
df['minute'] = df['pickup_datetime'].apply(lambda x: x.strftime('%M')).astype(int)
df['second'] = df['pickup_datetime'].apply(lambda x: x.strftime('%S')).astype(int)
df['date'] = df['pickup_datetime'].apply(lambda x: x.strftime('%Y%m%d')).astype(int)
print(df.info())

target = df.fare_amount
data = df.drop(['fare_amount', 'key', 'pickup_datetime'], axis=1)
# 划分训练集，转换成张量
tr_tx, te_tx, tr_ty, te_ty = data_to_tensor(data, target)


# ---------------------------------------定义网络---------------------------------------
class LinearModel(nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.liner = nn.Linear(14, 1)

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.liner(x)
        x = x.squeeze(-1)  # 线性回归的损失函数MSELoss要求输入维数和目标维数一致，因此做了个降维
        return x


# --------------------------准备训练（除超参数外，可复用）--------------------------

# 设置随机数种子，保证结果可复现
seed = 42
torch.manual_seed(seed)  # 设置CPU

# 实例化模型
model = LinearModel()

# 适应设备（CPU or GPU）
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)

# 存入DataLoader
ds = TensorDataset(tr_tx, tr_ty)
dl = DataLoader(ds, batch_size=128, shuffle=True)

# 学习率
lr = 1e-5
# 设定迭代次数
epoch = 100
# 设定每隔多少次显示一次评价指标
show_step = 10
# 选用优化器
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=0.01)
# 设置损失函数Loss
criterion = nn.MSELoss()


# ----------------------------------模型训练（可复用）----------------------------------
for epoch in range(epoch+1):
    for x, y in dl:
        pred = model(x)  # 正向传播
        loss = criterion(pred, y)  # 计算损失函数
        optimizer.zero_grad()  # 优化器的梯度清零
        loss.backward()  # 反向传播
        optimizer.step()  # 参数更新
    if epoch % show_step == 0:  # 控制输出间隔
        with torch.no_grad():
            print('epoch: ', epoch)
            tr_pred = model(tr_tx)  # 得到训练集的预测结果
            te_pred = model(te_tx)  # 得到测试集的预测结果
            all_regress_evaluation(tr_pred, tr_ty, te_pred, te_ty)


# -------------------------------输出验证集结果（基本可复用）-------------------------------
df_v = pd.read_csv('test.csv')  # 读取验证集数据
df_v.pickup_datetime = pd.to_datetime(df_v.pickup_datetime).dt.tz_localize(None)
df_v['hour'] = df_v['pickup_datetime'].apply(lambda x: x.strftime('%H')).astype(int)
df_v['minute'] = df_v['pickup_datetime'].apply(lambda x: x.strftime('%M')).astype(int)
df_v['second'] = df_v['pickup_datetime'].apply(lambda x: x.strftime('%S')).astype(int)
df_v['date'] = df_v['pickup_datetime'].apply(lambda x: x.strftime('%Y%m%d')).astype(int)
va_x = df_v.drop(['key', 'pickup_datetime'], axis=1)  # 弃列（少一个预测列）
va_x = mango_processing(va_x).astype(float)  # 数据预处理
va_tx = torch.tensor(np.array(va_x))  # 转换成张量
va_pred = model(va_tx)  # 预测
va_id = df_v['key']  # 读取索引列
va_out = pd.DataFrame({
    
    'key': va_id, 'fare_amount': va_pred.detach().numpy()})  # 构建输出数据的DataFrame
va_out['fare_amount'] = va_out['fare_amount'].apply(lambda x: round(x, 2))  # 数字保留两位小数
va_out.to_csv('Valid Prediction.csv', index=False)  # 输出到CSV，并取消索引列

6.2 逻辑回归分类

数据来源：Spaceship Titanic | Kaggle。

# 读取数据集，做好预处理
df = pd.read_csv('train.csv')
target = df.Transported
data = df.drop(['PassengerId', 'Transported', 'Name', 'Cabin'], axis=1)
# 划分训练集，转换成张量
tr_tx, te_tx, tr_ty, te_ty = data_to_tensor(data, target)


# ------------------------------------------定义网络------------------------------------------
class LinearModel(nn.Module):
    def __init__(self):
        super(LinearModel, self).__init__()
        self.liner = torch.nn.Linear(24, 2)

    def forward(self, x):
        x = x.to(torch.float32)
        x = self.liner(x)
        x = torch.sigmoid(x)
        return x


# --------------------------准备训练（除超参数外，可完全复用）--------------------------

# 设置随机数种子，保证结果可复现
seed = 42
torch.manual_seed(seed)  # 设置CPU

# 实例化模型
model = LinearModel()

# 适应设备（CPU or GPU）
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)

# 存入DataLoader
ds = TensorDataset(tr_tx, tr_ty)
dl = DataLoader(ds, batch_size=256, shuffle=True)

# 学习率
lr = 1e-3
# 设定迭代次数
epoch = 70
# 设定每隔多少次显示一次评价指标
show_step = 10
# 选用优化器
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=0.01)
# 设置损失函数Loss
criterion = nn.CrossEntropyLoss()


# ----------------------------------模型训练（可完全复用）----------------------------------
for epoch in range(epoch+1):
    for x, y in dl:
        pred = model(x)  # 正向传播
        loss = criterion(pred, y)  # 计算损失函数
        optimizer.zero_grad()  # 优化器的梯度清零
        loss.backward()  # 反向传播
        optimizer.step()  # 参数更新
    if epoch % show_step == 0:  # 控制输出间隔
        with torch.no_grad():
            print('epoch: ', epoch)
            tr_pred = model(tr_tx)  # 得到训练集的预测结果
            te_pred = model(te_tx)  # 得到测试集的预测结果
            all_classify_evaluation(tr_pred, tr_ty, te_pred, te_ty, 2)


# 输出曲线
tr_pred = model(tr_tx)
torch_plot_curve(tr_pred, tr_ty)


# -------------------------------输出验证集结果（基本可复用）-------------------------------
df_v = pd.read_csv('test.csv')  # 读取验证集数据
va_x = df_v.drop(['PassengerId', 'Name', 'Cabin'], axis=1)  # 弃列（少一个预测列）
va_x = mango_processing(va_x).astype(float)  # 数据预处理
va_tx = torch.tensor(np.array(va_x))  # 转换成张量
va_pred = model(va_tx)  # 预测
_, va_y = torch.max(va_pred.data, 1)  # 分类数据，需要选取概率最大项的索引填充到第1列
va_id = df_v['PassengerId']  # 读取索引列
va_out = pd.DataFrame({
    
    'PassengerId': va_id, 'Transported': va_y})  # 构建输出数据的DataFrame
va_out['Transported'] = va_out['Transported'].astype(bool)  # 分类数据，标签转换
va_out.to_csv('Valid Prediction.csv', index=False)  # 输出到CSV，并取消索引列