train_and_evaluate 代码详解-CSDN博客

本文链接：https://blog.csdn.net/weixin_70325224/article/details/147463736

def train_and_evaluate(model, model_name, loss_fn, X_train_normalized, y_train, X_test_normalized, y_test, epochs=50,
                       batch_size=32):
    """训练模型并计算测试集的 ROC 曲线数据"""
    print(f"\n{'=' * 50}")
    print(f"训练模型: {model_name}")
    print(f"{'=' * 50}")

    # 添加时间戳到模型文件名
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_filename = f"saved_models/{model_name}_{timestamp}.keras"

    # 添加模型检查点回调
    checkpoint = ModelCheckpoint(
        model_filename,
        monitor='val_auc',
        mode='max',
        save_best_only=True,
        verbose=1
    )

    model.compile(optimizer=Adam(learning_rate=1e-4),
                  loss=loss_fn,
                  metrics=['accuracy',
                           tf.keras.metrics.AUC(name='auc'),
                           tf.keras.metrics.Precision(name='precision'),
                           tf.keras.metrics.Recall(name='recall'),
                           tf.keras.metrics.TruePositives(name='tp'),
                           tf.keras.metrics.TrueNegatives(name='tn'),
                           tf.keras.metrics.FalsePositives(name='fp'),
                           tf.keras.metrics.FalseNegatives(name='fn')])

    early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)
    history = model.fit(X_train_normalized, y_train,
                        validation_data=(X_val_normalized, y_val),
                        epochs=epochs,
                        batch_size=batch_size,
                        callbacks=[early_stop, checkpoint],
                        verbose=1)

    # 绘制训练历史
    plot_training_history(history, model_name)

    # 获取预测概率和类别
    y_score = model.predict(X_test_normalized).flatten()
    y_pred = (y_score > 0.5).astype(int)

    # 绘制ROC和PR曲线
    roc_auc, avg_precision = plot_roc_pr_curves(y_test, y_score, model_name)

    # 绘制混淆矩阵
    plot_confusion_matrix(y_test, y_pred, classes=['Class 0', 'Class 1'], model_name=model_name)

    # 打印分类报告
    print(f"\n{model_name} 分类报告:")
    print(classification_report(y_test, y_pred, target_names=['Class 0', 'Class 1']))

    # 计算更多评估指标
    tn, fp, fn, tp = confusion_matrix(y_test, y_pred).ravel()
    sensitivity = tp / (tp + fn)
    specificity = tn / (tn + fp)
    f1_score = 2 * tp / (2 * tp + fp + fn)

def train_and_evaluate(model, model_name, loss_fn, X_train_normalized, y_train, X_test_normalized, y_test, epochs=50,
                       batch_size=32):

def: Python关键字，定义函数
train_and_evaluate: 函数名称，表示"训练和评估"
model: 参数，接收要训练的Keras模型对象
model_name: 参数，字符串类型，用于标识模型名称
loss_fn: 参数，损失函数（如交叉熵或focal loss）
X_train_normalized: 参数，标准化后的训练集特征数据
y_train: 参数，训练集标签
X_test_normalized: 参数，标准化后的测试集特征数据
y_test: 参数，测试集标签
epochs=50: 默认参数，训练轮数，默认50
batch_size=32: 默认参数，批量大小，默认32

    """训练模型并计算测试集的 ROC 曲线数据"""

函数文档字符串，说明函数功能

    print(f"\n{'=' * 50}")
    print(f"训练模型: {model_name}")
    print(f"{'=' * 50}")

打印分隔线和当前训练模型名称，用于输出美化

    # 添加时间戳到模型文件名
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_filename = f"saved_models/{model_name}_{timestamp}.keras"

timestamp: 生成当前时间的字符串（格式：年月日_时分秒）
model_filename: 构建模型保存路径，包含模型名和时间戳

    # 添加模型检查点回调
    checkpoint = ModelCheckpoint(
        model_filename,
        monitor='val_auc',
        mode='max',
        save_best_only=True,
        verbose=1
    )

ModelCheckpoint: Keras回调，用于保存模型
- model_filename: 保存路径
- monitor=‘val_auc’: 监控验证集的AUC指标
- mode=‘max’: 监控指标越大越好
- save_best_only=True: 只保存最佳模型
- verbose=1: 显示保存信息

    model.compile(optimizer=Adam(learning_rate=1e-4),
                  loss=loss_fn,
                  metrics=['accuracy',
                           tf.keras.metrics.AUC(name='auc'),
                           tf.keras.metrics.Precision(name='precision'),
                           tf.keras.metrics.Recall(name='recall'),
                           tf.keras.metrics.TruePositives(name='tp'),
                           tf.keras.metrics.TrueNegatives(name='tn'),
                           tf.keras.metrics.FalsePositives(name='fp'),
                           tf.keras.metrics.FalseNegatives(name='fn')])

model.compile: 配置模型训练参数
- optimizer=Adam(learning_rate=1e-4): 使用Adam优化器，学习率0.0001
- loss=loss_fn: 使用传入的损失函数
- metrics: 监控指标列表
  - accuracy: 准确率
  - AUC: ROC曲线下面积
  - Precision: 精确率
  - Recall: 召回率
  - True/False Positives/Negatives: 混淆矩阵元素

    early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

EarlyStopping: 早停回调
- monitor=‘val_loss’: 监控验证集损失
- patience=5: 连续5轮不改善则停止
- restore_best_weights=True: 恢复最佳权重

    history = model.fit(X_train_normalized, y_train,
                        validation_data=(X_val_normalized, y_val),
                        epochs=epochs,
                        batch_size=batch_size,
                        callbacks=[early_stop, checkpoint],
                        verbose=1)

model.fit: 训练模型
- X_train_normalized/y_train: 训练数据
- validation_data: 验证数据集
- epochs: 训练轮数
- batch_size: 批量大小
- callbacks: 回调列表（早停和模型检查点）
- verbose=1: 显示进度条

    # 绘制训练历史
    plot_training_history(history, model_name)

调用自定义函数绘制训练过程中的损失和准确率曲线

    # 获取预测概率和类别
    y_score = model.predict(X_test_normalized).flatten()
    y_pred = (y_score > 0.5).astype(int)

y_score: 模型对测试集的预测概率（经过sigmoid输出）
y_pred: 将概率转为二分类结果（>0.5为1，否则为0）

    # 绘制ROC和PR曲线
    roc_auc, avg_precision = plot_roc_pr_curves(y_test, y_score, model_name)

调用自定义函数绘制ROC和PR曲线，返回AUC和AP值

    # 绘制混淆矩阵
    plot_confusion_matrix(y_test, y_pred, classes=['Class 0', 'Class 1'], model_name=model_name)

调用自定义函数绘制混淆矩阵

    # 打印分类报告
    print(f"\n{model_name} 分类报告:")
    print(classification_report(y_test, y_pred, target_names=['Class 0', 'Class 1']))

打印sklearn的分类报告（精确率、召回率、F1等）

    # 计算更多评估指标
    tn, fp, fn, tp = confusion_matrix(y_test, y_pred).ravel()
    sensitivity = tp / (tp + fn)
    specificity = tn / (tn + fp)
    f1_score = 2 * tp / (2 * tp + fp + fn)

tn/fp/fn/tp: 从混淆矩阵获取真负/假正/假负/真正
sensitivity: 敏感度（召回率）= TP/(TP+FN)
specificity: 特异度 = TN/(TN+FP)
f1_score: F1分数 = 2*(precision*recall)/(precision+recall)

该函数最终会返回包含所有评估指标的字典结果。每个步骤都包含完整的模型训练、评估和可视化流程。