1from keras.datasets import mnist 2from keras.layers import Dense, Flatten, Conv2D, MaxPooling2D, Dropout 3from keras.models import Sequential 4from keras.optimizers import RMSprop 5from keras.utils import np_utils 6 7img_rows, img_cols = 28, 28 8 9(X_train, y_train), (X_test, y_test) = mnist.load_data() 10 11X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, 1) 12X_test = X_test.reshape(X_test.shape[0], img_rows, img_cols, 1) 13input_shape = (img_rows, img_cols, 1) 14 15X_train = X_train.astype('float32') / 255. 16X_test = X_test.astype('float32') / 255. 17 18Y_train = np_utils.to_categorical(y_train, 10) 19Y_test = np_utils.to_categorical(y_test, 10) 20 21model = Sequential() 22 23model.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape)) 24model.add(MaxPooling2D()) 25model.add(Dropout(0.2)) 26 27model.add(Conv2D(64, kernel_size=(3, 3), activation='relu')) 28model.add(MaxPooling2D()) 29model.add(Dropout(0.2)) 30 31model.add(Conv2D(128, kernel_size=(3, 3), activation='relu')) 32model.add(MaxPooling2D()) 33model.add(Flatten()) 34model.add(Dropout(0.2)) 35 36model.add(Dense(625, activation='relu')) 37model.add(Dropout(0.5)) 38 39model.add(Dense(10, activation='softmax')) 40 41opt = RMSprop() 42model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy']) 43model.summary() 44 45history = model.fit(X_train, Y_train, nb_epoch=10, batch_size=128, shuffle=True, verbose=2, 46 validation_data=(X_test, Y_test)) 47 48 49_________________________________________________________________ 50Layer (type) Output Shape Param # 51================================================================= 52conv2d_1 (Conv2D) (None, 26, 26, 32) 320 53_________________________________________________________________ 54max_pooling2d_1 (MaxPooling2 (None, 13, 13, 32) 0 55_________________________________________________________________ 56dropout_1 (Dropout) (None, 13, 13, 32) 0 57_________________________________________________________________ 58conv2d_2 (Conv2D) (None, 11, 11, 64) 18496 59_________________________________________________________________ 60max_pooling2d_2 (MaxPooling2 (None, 5, 5, 64) 0 61_________________________________________________________________ 62dropout_2 (Dropout) (None, 5, 5, 64) 0 63_________________________________________________________________ 64conv2d_3 (Conv2D) (None, 3, 3, 128) 73856 65_________________________________________________________________ 66max_pooling2d_3 (MaxPooling2 (None, 1, 1, 128) 0 67_________________________________________________________________ 68flatten_1 (Flatten) (None, 128) 0 69_________________________________________________________________ 70dropout_3 (Dropout) (None, 128) 0 71_________________________________________________________________ 72dense_1 (Dense) (None, 625) 80625 73_________________________________________________________________ 74dropout_4 (Dropout) (None, 625) 0 75_________________________________________________________________ 76dense_2 (Dense) (None, 10) 6260 77================================================================= 78Total params: 179,557 79Trainable params: 179,557 80Non-trainable params: 0
卷积深度神经网络在深度神经网络的基础上，加入了卷积层和池化层。网络的第一层是二维卷积层
Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=input_shape)
这一层的卷积核的数目为32，也就是输出的维度。卷积的空间维度是3_3,也就是用一个3_3的框框去做卷积，采用relu激活函数。 接下来用了MaxPooling2D 这个二维的池化层，主要用来降维和特征提取。
在连续做了3次CNN之后，使用了一个Flatten层。这个层会把多维数据压扁成一维数据，但是不会影响数据总量也就是batch的大小，从summery里面来看，就是把(None, 1, 1, 128) 压扁成了(None, 128) 。
最后，在fit的时候我们看到了一个新的参数shuffle=True，这个参数其实是默认开启的，不写也可以，指定为True，会在每次训练的时候让训练集随机打乱