使用 Keras 的用于 MNIST 的 LeNet CNN

让我们重新审视具有相同数据集的相同 LeNet 架构，以在 Keras 中构建和训练 CNN 模型：

1. 导入所需的 Keras 模块：

import keras
from keras.models import Sequential
from keras.layers import Conv2D,MaxPooling2D, Dense, Flatten, Reshape
from keras.optimizers import SGD

1. 定义每个层的过滤器数量：

n_filters=[32,64]

1. 定义其他超参数：

learning_rate = 0.01
n_epochs = 10
batch_size = 100

1. 定义顺序模型并添加层以将输入数据重新整形为形状(n_width,n_height,n_depth)：

model = Sequential()
model.add(Reshape(target_shape=(n_width,n_height,n_depth), 
                  input_shape=(n_inputs,))
         )

1. 使用 4 x 4 内核过滤器，SAME填充和relu激活添加第一个卷积层：

model.add(Conv2D(filters=n_filters[0],kernel_size=4, 
                 padding='SAME',activation='relu') 
         )

1. 添加区域大小为 2 x 2 且步长为 2 x 2 的池化层：

model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2)))

1. 以与添加第一层相同的方式添加第二个卷积和池化层：

model.add(Conv2D(filters=n_filters[1],kernel_size=4, 
                 padding='SAME',activation='relu') 
         )
model.add(MaxPooling2D(pool_size=(2,2),strides=(2,2)))

1. 添加层以展平第二个层的输出和 1024 个神经元的完全连接层，以处理展平的输出：

model.add(Flatten())
model.add(Dense(units=1024, activation='relu'))

1. 使用softmax激活添加最终输出层：

model.add(Dense(units=n_outputs, activation='softmax'))

1. 使用以下代码查看模型摘要：

model.summary()

该模型描述如下：

Layer (type)                 Output Shape              Param #   
=================================================================
reshape_1 (Reshape)          (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 32)        544       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 14, 14, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 14, 14, 64)        32832     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 7, 7, 64)          0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 3136)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 1024)              3212288   
_________________________________________________________________
dense_2 (Dense)              (None, 10)                10250     
=================================================================
Total params: 3,255,914
Trainable params: 3,255,914
Non-trainable params: 0
_________________________________________________________________

1. 编译，训练和评估模型：

model.compile(loss='categorical_crossentropy',
              optimizer=SGD(lr=learning_rate),
              metrics=['accuracy'])
model.fit(X_train, Y_train,batch_size=batch_size,
          epochs=n_epochs)
score = model.evaluate(X_test, Y_test)
print('\nTest loss:', score[0])
print('Test accuracy:', score[1])

我们得到以下输出：

Epoch 1/10
55000/55000 [===================] - 267s - loss: 0.8854 - acc: 0.7631   
Epoch 2/10
55000/55000 [===================] - 272s - loss: 0.2406 - acc: 0.9272   
Epoch 3/10
55000/55000 [===================] - 267s - loss: 0.1712 - acc: 0.9488   
Epoch 4/10
55000/55000 [===================] - 295s - loss: 0.1339 - acc: 0.9604   
Epoch 5/10
55000/55000 [===================] - 278s - loss: 0.1112 - acc: 0.9667   
Epoch 6/10
55000/55000 [===================] - 279s - loss: 0.0957 - acc: 0.9714   
Epoch 7/10
55000/55000 [===================] - 316s - loss: 0.0842 - acc: 0.9744   
Epoch 8/10
55000/55000 [===================] - 317s - loss: 0.0758 - acc: 0.9773   
Epoch 9/10
55000/55000 [===================] - 285s - loss: 0.0693 - acc: 0.9790   
Epoch 10/10
55000/55000 [===================] - 217s - loss: 0.0630 - acc: 0.9804
Test loss: 0.0628845927377
Test accuracy: 0.9785

准确性的差异可归因于我们在这里使用 SGD 优化器这一事实，它没有实现我们用于 TensorFlow 模型的AdamOptimizer提供的一些高级功能。