DL之Attention：基于ClutteredMNIST手写数字图片数据集分别利用CNN_Init、ST_CNN算法(CNN+SpatialTransformer)实现多分类预测（二）

核心代码

   #(2)、建立ST定位网络：尝试更多的conv层，并分别在X轴和y轴上做最大池化

   # localization net. TODO: try more conv layers, and do max pooling on X- and Y-axes respectively

   locnet = Sequential()

   # locnet.add(MaxPooling2D(pool_size=(2,2), input_shape=input_shape))

   # locnet.add(Convolution2D(32, (5, 5)))

   locnet.add(Convolution2D(32, (5, 5), input_shape=input_shape))

   locnet.add(Activation('relu'))

   # locnet.add(Dropout(0.2)) # 0.2

   locnet.add(MaxPooling2D(pool_size=(2,2)))

   locnet.add(Convolution2D(64, (5, 5)))

   locnet.add(Activation('relu'))

   # locnet.add(Dropout(0.2)) # 0.3

   locnet.add(Convolution2D(64, (3, 3)))

   locnet.add(Activation('relu'))

   locnet.add(MaxPooling2D(pool_size=(2,2)))

   locnet.add(Flatten())

   locnet.add(Dense(50))

   locnet.add(Activation('relu'))

   locnet.add(Dense(6, weights=weights))

   print(locnet.summary())

   #(3)、建立CNN网络

   model = Sequential()

   model.add(SpatialTransformer(localization_net=locnet,

                                output_size=(30,30), input_shape=input_shape))

   # model.add(Convolution2D(32, (3, 3), padding='same'))

   # model.add(Activation('relu'))

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

   # model.add(Convolution2D(64, (3, 3)))

   # model.add(Activation('relu'))

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

   # model.add(Dropout(0.5)) # 0.25

   # E: removed first 3 dropout layers

   model.add(Conv2D(32, kernel_size=(3, 3), activation='relu'))

   model.add(Dropout(0.5)) # 0.5

   model.add(Conv2D(64, (3, 3), activation='relu'))

   model.add(Dropout(0.5)) # 0.5

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

   model.add(Conv2D(64, kernel_size=(3, 3),

                    activation='relu'))

   model.add(Dropout(0.5)) # 0.5

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

   # model.add(Conv2D(64, (3, 3), activation='relu'))

   # model.add(Dropout(0.5))

   model.add(Flatten())

   model.add(Dense(256)) # 256

   model.add(Dropout(0.5)) # 0.5

   model.add(Activation('relu'))

   model.add(Dense(nb_classes))

   model.add(Activation('softmax'))