使用 TensorFlow 和 Keras 构建 CNN 进行分类，同时结合感知机进行预测

标签：features Keras batch name 感知机 train CNN model size

引言

本教程将演示如何使用 TensorFlow 和 Keras 构建一个卷积神经网络（CNN），对图片数据进行分类。同时，我们将提取 CNN 的特征并利用感知机（Perceptron）对其进行进一步预测。数据集来源于 GitHub 仓库。

通过以下步骤，我们将实现一个完整的训练、特征提取和预测流程。

步骤 1：环境配置与必要库的导入

首先，导入所有需要的库，包括 TensorFlow、Keras、OpenCV、Scikit-learn 等。

import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dropout, Dense, Flatten, Input
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing.image import ImageDataGenerator, img_to_array
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau, TensorBoard
from tensorflow.keras.optimizers import Adam
import cv2
from sklearn.linear_model import Perceptron
import sklearn.metrics as sm
import joblib
import matplotlib.pyplot as plt

步骤 2：定义数据路径与超参数

设置训练集和测试集的路径，同时定义 CNN 模型和训练所需的超参数。

chkfilepath = "/datasets/ComplexCNN_weights.keras"
img_path = '/datasets/dataset/train/杯子/1.jpg'
train_dir = '/datasets/dataset/train'
test_dir = '/datasets/dataset/test'
model_path = "ComplexCNN_model.keras"
pkl_path = 'perceptron.pkl'

im_height, im_width = 224, 224
batch_size = 16
epochs = 50

步骤 3：数据增强与数据生成器

训练数据增强

使用 ImageDataGenerator 对训练数据进行增强。

train_image_generator = ImageDataGenerator(
    rescale=1./255,
    rotation_range=30,
    width_shift_range=0.1,
    height_shift_range=0.1,
    shear_range=0.1,
    zoom_range=0.1,
    horizontal_flip=True,
    fill_mode='nearest'
)

train_data_gen = train_image_generator.flow_from_directory(
    directory=train_dir,
    batch_size=batch_size,
    shuffle=True,
    target_size=(im_height, im_width),
    class_mode='categorical'
)

验证数据生成器

对验证数据仅进行归一化处理。

validation_image_generator = ImageDataGenerator(rescale=1./255)

val_data_gen = validation_image_generator.flow_from_directory(
    directory=test_dir,
    batch_size=batch_size,
    shuffle=False,
    target_size=(im_height, im_width),
    class_mode='categorical'
)

步骤 4：构建 CNN 模型

定义一个 3 层卷积结构的 CNN 模型，最后通过 Dense 层完成分类。

def ComplexCNN(height, width, channel, classes):
    inputs = Input(shape=(height, width, channel), name="input_layer")
    x = Conv2D(16, (3, 3), activation='relu', padding='same', name='c16')(inputs)
    x = MaxPooling2D(pool_size=2, name="maxpool1")(x)
    x = Conv2D(32, (3, 3), activation='relu', padding='same', name='c32')(x)
    x = MaxPooling2D(pool_size=2, name="maxpool2")(x)
    x = Conv2D(64, (3, 3), activation='relu', padding='same', name='c64')(x)
    x = MaxPooling2D(pool_size=2, name="maxpool3")(x)

    x = Flatten(name="flatten")(x)
    x = Dense(128, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001), name='dense128')(x)
    x = Dropout(0.75, name="dropout")(x)

    outputs = Dense(classes, activation='softmax', name="output_layer")(x)
    model = Model(inputs=inputs, outputs=outputs, name="ComplexCNN")
    return model

步骤 5：模型训练

编译和训练模型，并使用回调函数保存最优模型，动态调整学习率和早停。

model_cnn = ComplexCNN(im_height, im_width, 3, classes)
model_cnn.compile(
    optimizer=Adam(learning_rate=0.001),
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

callbacks = [
    ModelCheckpoint(chkfilepath, monitor="val_loss", verbose=1, save_best_only=True, save_weights_only=False),
    EarlyStopping(monitor="val_loss", patience=10, verbose=1, restore_best_weights=True),
    ReduceLROnPlateau(monitor="val_loss", factor=0.5, patience=5, verbose=1, min_lr=1e-6),
    TensorBoard(log_dir="./logs", histogram_freq=1)
]

history = model_cnn.fit(
    train_data_gen,
    steps_per_epoch=total_train // batch_size,
    epochs=epochs,
    validation_data=val_data_gen,
    validation_steps=total_val // batch_size,
    callbacks=callbacks
)

步骤 6：提取特征

使用 CNN 的中间层提取特征，供感知机模型使用。

dense1_layer_model = Model(inputs=model_cnn.input, outputs=model_cnn.get_layer('dense128').output)

def predict_feature(base_model, datagen, directory, batch_size=16):
    generator = datagen.flow_from_directory(
        directory,
        target_size=(224, 224),
        batch_size=batch_size,
        class_mode='binary'
    )
    count = generator.n
    features = np.zeros(shape=(count, 128))
    labels = np.zeros(shape=(count))

    i = 0
    for inputs_batch, labels_batch in generator:
        features_batch = base_model.predict(inputs_batch, verbose=0)
        features[i * batch_size: (i + 1) * batch_size] = features_batch
        labels[i * batch_size: (i + 1) * batch_size] = labels_batch
        i += 1
        if i >= count // batch_size:
            break
    return features, labels, generator.class_indices

train_features, train_labels, _ = predict_feature(dense1_layer_model, train_image_generator, train_dir)
test_features, test_labels, _ = predict_feature(dense1_layer_model, validation_image_generator, test_dir)

步骤 7：训练感知机

使用提取的特征训练感知机。

train_features_np = train_features.reshape(-1, 128)
test_features_np = test_features.reshape(-1, 128)

def perceptron_report():
    model = Perceptron(random_state=43)
    per_model = model.fit(train_features_np, train_labels)
    per_predict = per_model.predict(test_features_np)
    print(sm.classification_report(per_predict, test_labels))
    joblib.dump(per_model, pkl_path)

perceptron_report()

步骤 8：模型预测

加载感知机模型并进行预测。

def perceptron_predict():
    per_model2 = joblib.load(pkl_path)
    base_preds = dense1_layer_model.predict(x)
    data = base_preds.reshape(-1, 128)
    test_label = per_model2.predict(data)
    print("Predicted Label:", test_label)

步骤 9：可视化结果

绘制训练和验证的损失、准确率曲线。

def plot(history):
    fig, ax = plt.subplots(2, 1, figsize=(8, 8))
    ax[0].plot(history.history['loss'], label='Train Loss')
    ax[0].plot(history.history['val_loss'], label='Validation Loss')
    ax[0].set_title('Loss')
    ax[1].plot(history.history['accuracy'], label='Train Accuracy')
    ax[1].plot(history.history['val_accuracy'], label='Validation Accuracy')
    ax[1].set_title('Accuracy')
    plt.legend()
    plt.show()

plot(history)

完整代码

以下是完整的代码汇总，可以直接复制运行：

import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dropout, Dense, Flatten, Input
from tensorflow.keras.models import Model
from tensorflow.keras.preprocessing.image import ImageDataGenerator, img_to_array
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau, TensorBoard
from tensorflow.keras.optimizers import Adam
import cv2
from sklearn.linear_model import Perceptron
import sklearn.metrics as sm
import joblib
import matplotlib.pyplot as plt

# 数据路径和超参数
chkfilepath = "/datasets/ComplexCNN_weights.keras"
img_path = '/datasets/dataset/train/杯子/1.jpg'
train_dir = '/datasets/dataset/train'
test_dir = '/datasets/dataset/test'
model_path = "ComplexCNN_model.keras"
pkl_path = 'perceptron.pkl'

im_height, im_width = 224, 224
batch_size = 16
epochs = 50

# 数据增强
train_image_generator = ImageDataGenerator(
    rescale=1./255,
    rotation_range=30,
    width_shift_range=0.1,
    height_shift_range=0.1,
    shear_range=0.1,
    zoom_range=0.1,
    horizontal_flip=True,
    fill_mode='nearest'
)

train_data_gen = train_image_generator.flow_from_directory(
    directory=train_dir,
    batch_size=batch_size,
    shuffle=True,
    target_size=(im_height, im_width),
    class_mode='categorical'
)

validation_image_generator = ImageDataGenerator(rescale=1./255)

val_data_gen = validation_image_generator.flow_from_directory(
    directory=test_dir,
    batch_size=batch_size,
    shuffle=False,
    target_size=(im_height, im_width),
    class_mode='categorical'
)

# 定义 CNN 模型
def ComplexCNN(height, width, channel, classes):
    inputs = Input(shape=(height, width, channel), name="input_layer")
    x = Conv2D(16, (3, 3), activation='relu', padding='same', name='c16')(inputs)
    x = MaxPooling2D(pool_size=2, name="maxpool1")(x)
    x = Conv2D(32, (3, 3), activation='relu', padding='same', name='c32')(x)
    x = MaxPooling2D(pool_size=2, name="maxpool2")(x)
    x = Conv2D(64, (3, 3), activation='relu', padding='same', name='c64')(x)
    x = MaxPooling2D(pool_size=2, name="maxpool3")(x)

    x = Flatten(name="flatten")(x)
    x = Dense(128, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001), name='dense128')(x)
    x = Dropout(0.75, name="dropout")(x)

    outputs = Dense(classes, activation='softmax', name="output_layer")(x)
    model = Model(inputs=inputs, outputs=outputs, name="ComplexCNN")
    return model

# 提取特征
def predict_feature(base_model, datagen, directory, batch_size=16):
    generator = datagen.flow_from_directory(
        directory,
        target_size=(224, 224),
        batch_size=batch_size,
        class_mode='binary'
    )
    count = generator.n
    features = np.zeros(shape=(count, 128))
    labels = np.zeros(shape=(count))

    i = 0
    for inputs_batch, labels_batch in generator:
        features_batch = base_model.predict(inputs_batch, verbose=0)
        features[i * batch_size: (i + 1) * batch_size] = features_batch
        labels[i * batch_size: (i + 1) * batch_size] = labels_batch
        i += 1
        if i >= count // batch_size:
            break
    return features, labels, generator.class_indices

# Perceptron 模型报告
def perceptron_report():
    model = Perceptron(random_state=43)
    per_model = model.fit(train_features_np, train_labels)
    per_predict = per_model.predict(test_features_np)
    print(sm.classification_report(per_predict, test_labels))
    joblib.dump(per_model, pkl_path)

# Perceptron 预测
def perceptron_predict():
    per_model2 = joblib.load(pkl_path)
    base_preds = dense1_layer_model.predict(x)
    data = base_preds.reshape(-1, 128)
    test_label = per_model2.predict(data)
    print("Predicted Label:", test_label)

# 绘制训练结果
def plot(history):
    fig, ax = plt.subplots(2, 1, figsize=(8, 8))
    ax[0].plot(history.history['loss'], label='Train Loss')
    ax[0].plot(history.history['val_loss'], label='Validation Loss')
    ax[0].set_title('Loss')
    ax[1].plot(history.history['accuracy'], label='Train Accuracy')
    ax[1].plot(history.history['val_accuracy'], label='Validation Accuracy')
    ax[1].set_title('Accuracy')
    plt.legend()
    plt.show()

if __name__ == '__main__':
    # 创建并训练 CNN 模型
    model_cnn = ComplexCNN(im_height, im_width, 3, len(train_data_gen.class_indices))
    model_cnn.compile(
        optimizer=Adam(learning_rate=0.001),
        loss='categorical_crossentropy',
        metrics=['accuracy']
    )
    callbacks = [
        ModelCheckpoint(chkfilepath, monitor="val_loss", verbose=1, save_best_only=True, save_weights_only=False),
        EarlyStopping(monitor="val_loss", patience=10, verbose=1, restore_best_weights=True),
        ReduceLROnPlateau(monitor="val_loss", factor=0.5, patience=5, verbose=1, min_lr=1e-6),
        TensorBoard(log_dir="./logs", histogram_freq=1)
    ]

    history = model_cnn.fit(
        train_data_gen,
        steps_per_epoch=train_data_gen.n // batch_size,
        epochs=epochs,
        validation_data=val_data_gen,
        validation_steps=val_data_gen.n // batch_size,
        callbacks=callbacks
    )

    model_cnn.save(model_path)
    plot(history)

    # 提取训练和测试特征
    dense1_layer_model = Model(inputs=model_cnn.input, outputs=model_cnn.get_layer('dense128').output)
    train_features, train_labels, _ = predict_feature(dense1_layer_model, train_image_generator, train_dir)
    test_features, test_labels, _ = predict_feature(dense1_layer_model, validation_image_generator, test_dir)

    train_features_np = train_features.reshape(-1, 128)
    test_features_np = test_features.reshape(-1, 128)

    # 训练感知机并进行预测
    perceptron_report()
    perceptron_predict()

总结

本文介绍了如何使用 TensorFlow 和 Keras 构建 CNN 模型进行分类，并结合感知机完成进一步预测分析。通过分步操作，从数据增强到模型训练，再到特征提取与感知机训练，我们实现了一个完整的分类任务。

访问 GitHub 数据集以尝试运行代码！

标签：features,Keras,batch,name,感知机,train,CNN,model,size
From： https://blog.csdn.net/yaoyihe/article/details/143958148