### 使用TensorFlow实现LeNet-5手写数字识别的完整过程 以下是使用TensorFlow实现LeNet-5模型进行MNIST手写数字识别的具体流程，包括样本下载、模型定义、训练、测试验证以及参数计算等内容。 --- #### 1. 样本下载 MNIST数据集可以通过`tensorflow.datasets`模块或者手动加载的方式获取。通常情况下，可以直接利用`input_data.read_data_sets()`方法完成数据集的下载和预处理[^4]。 ```python import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data # 下载并加载MNIST数据集 mnist = input_data.read_data_sets("data/mnist", one_hot=True) # 获取训练集、测试集和验证集 train_images, train_labels = mnist.train.images, mnist.train.labels test_images, test_labels = mnist.test.images, mnist.test.labels validation_images, validation_labels = mnist.validation.images, mnist.validation.labels print(f"Training set size: {train_images.shape}") print(f"Test set size: {test_images.shape}") print(f"Validation set size: {validation_images.shape}") ``` --- #### 2. LeNet-5 模型定义 LeNet-5 是一种经典的卷积神经网络架构，适用于图像分类任务。其核心结构包括两组卷积层+池化层，随后接两个全连接层[^3]。 ```python class LeNet5(tf.keras.Model): def __init__(self): super(LeNet5, self).__init__() # 卷积层 C1 (输入: 32x32x1 -> 输出: 28x28x6) self.conv1 = tf.keras.layers.Conv2D(filters=6, kernel_size=(5, 5), strides=(1, 1), activation='relu', padding="valid") # 池化层 S2 (输入: 28x28x6 -> 输出: 14x14x6) self.pool1 = tf.keras.layers.AvgPool2D(pool_size=(2, 2), strides=(2, 2)) # 卷积层 C3 (输入: 14x14x6 -> 输出: 10x10x16) self.conv2 = tf.keras.layers.Conv2D(filters=16, kernel_size=(5, 5), strides=(1, 1), activation='relu', padding="valid") # 池化层 S4 (输入: 10x10x16 -> 输出: 5x5x16) self.pool2 = tf.keras.layers.AvgPool2D(pool_size=(2, 2), strides=(2, 2)) # 展平层 Fc5 输入: 5x5x16 -> 输出: 400 self.flatten = tf.keras.layers.Flatten() # 全连接层 Fc6 (输入: 400 -> 输出: 120) self.fc1 = tf.keras.layers.Dense(units=120, activation='relu') # 全连接层 Fc7 (输入: 120 -> 输出: 84) self.fc2 = tf.keras.layers.Dense(units=84, activation='relu') # 输出层 Fc8 (输入: 84 -> 输出: 10) self.out = tf.keras.layers.Dense(units=10, activation=None) def call(self, inputs): x = self.conv1(inputs) # Convolutional Layer C1 x = self.pool1(x) # Pooling Layer S2 x = self.conv2(x) # Convolutional Layer C3 x = self.pool2(x) # Pooling Layer S4 x = self.flatten(x) # Flatten Layer x = self.fc1(x) # Fully Connected Layer Fc6 x = self.fc2(x) # Fully Connected Layer Fc7 output = self.out(x) # Output Layer Fc8 return output ``` --- #### 3. 数据预处理 由于原始MNIST图片尺寸为 `28x28`，而LeNet-5期望输入尺寸为 `32x32`，因此需要对图片进行填充操作[^3]。 ```python def preprocess_mnist(data): data = tf.reshape(data, [-1, 28, 28, 1]) padded_data = tf.pad(data, [[0, 0], [2, 2], [2, 2], [0, 0]], mode="CONSTANT") # 填充至32x32 return padded_data / 255.0 # 归一化到[0, 1] # 预处理训练集和测试集 train_images = preprocess_mnist(train_images) test_images = preprocess_mnist(test_images) ``` --- #### 4. 模型编译与训练 配置优化器、损失函数和评估指标，并启动训练循环[^2]。 ```python model = LeNet5() loss_fn = tf.keras.losses.CategoricalCrossentropy(from_logits=True) optimizer = tf.keras.optimizers.Adam() @tf.function def train_step(images, labels): with tf.GradientTape() as tape: predictions = model(images) loss = loss_fn(labels, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) epochs = 10 batch_size = 64 steps_per_epoch = len(train_images) // batch_size for epoch in range(epochs): for step in range(steps_per_epoch): start_idx = step * batch_size end_idx = (step + 1) * batch_size images_batch = train_images[start_idx:end_idx] labels_batch = train_labels[start_idx:end_idx] train_step(images_batch, labels_batch) print(f"Epoch {epoch + 1}/{epochs} completed.") ``` --- #### 5. 测试验证 在测试集上评估模型性能[^1]。 ```python accuracy_metric = tf.keras.metrics.CategoricalAccuracy() predictions = model(test_images) accuracy_metric.update_state(test_labels, predictions) print(f"Test Accuracy: {accuracy_metric.result().numpy():.4f}") ``` --- #### 6. 参数计算 统计整个模型中的可训练参数数量[^3]。 ```python total_params = sum([tf.size(var).numpy() for var in model.trainable_variables]) print(f"Total Trainable Parameters: {total_params}") ``` --- #### 7. 代码运行结果示例 假设经过10轮训练后，最终得到如下结果： ``` Epoch 1/10 completed. ... Epoch 10/10 completed. Test Accuracy: 0.9876 Total Trainable Parameters: 61706 ``` --- ###