多尺度CNN

Question

我正在尝试创建一个类似于图像的网络，但我不知道它是如何实现的。

我希望它只接收一个输入，然后将它提供给包含卷积块的两个子网络。这段代码是我写的，但不起作用。

main_model = Sequential()
main_model.add(Convolution2D(filters=16, kernel_size=(2, 2), input_shape=(32, 32, 3)))
main_model.add(BatchNormalization())
main_model.add(Activation('relu'))
main_model.add(MaxPooling2D(pool_size=(2, 2)))

main_model.add(Convolution2D(filters=32, kernel_size=(2, 2)))
main_model.add(BatchNormalization())
main_model.add(Activation('relu'))
main_model.add(MaxPooling2D(pool_size=(2, 2)))

main_model.add(Convolution2D(filters=64, kernel_size=(2, 2)))
main_model.add(BatchNormalization())
main_model.add(Activation('relu'))
main_model.add(MaxPooling2D(pool_size=(2, 2)))

main_model.add(Flatten())

# lower features model - CNN2
lower_model = Sequential()
lower_model.add(Convolution2D(filters=16, kernel_size=(1, 1), input_shape=(32, 32, 3)))
lower_model.add(BatchNormalization())
lower_model.add(Activation('relu'))
lower_model.add(MaxPooling2D(pool_size=(2, 2)))
lower_model.add(Flatten())

lower_model.add(Convolution2D(filters=32, kernel_size=(1, 1)))
lower_model.add(BatchNormalization())
lower_model.add(Activation('relu'))
lower_model.add(MaxPooling2D(pool_size=(2, 2)))

lower_model.add(Convolution2D(filters=64, kernel_size=(1, 1)))
lower_model.add(BatchNormalization())
lower_model.add(Activation('relu'))
lower_model.add(MaxPooling2D(pool_size=(2, 2)))

lower_model.add(Flatten())

# merged model
merged_model = concatenate([main_model, lower_model])

final_model = Sequential()
final_model.add(merged_model)
final_model.add(Dense(32))
final_model.add(Activation('relu'))
final_model.add(Dropout(0.5))
final_model.add(Dense(1))
final_model.add(Activation('sigmoid'))

final_model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

我得到了一个错误：

ValueError: Input 0 of layer conv2d_4 is incompatible with the layer: expected ndim=4, found ndim=2. Full shape received: [None, 4096]

Answer 1

通过使用Keras函数API，这是可能的，您可以这样做

img_inputs = keras.Input(shape=(32, 32, 3))

branchA = Convolution2D(filters=32, kernel_size=(1, 1))(img_inputs)
branchA = BatchNormalization()(branchA)
branchA = Activation('relu')(branchA)
branchA = MaxPooling2D(pool_size=(2, 2))(branchA)
branchA = Model(inputs=img_inputs, outputs=branchA)

branchB = Convolution2D(filters=32, kernel_size=(1, 1))(img_inputs)
branchB = BatchNormalization()(branchB)
branchB = Activation('relu')(branchB)
branchB = MaxPooling2D(pool_size=(2, 2))(branchB)
branchB = Model(inputs=img_inputs, outputs=branchB)

#you may need to make sure output size of branchA and branchB are same size
combined = concatenate([branchA.output, branchB.output])

combined = Dense(2, activation="relu")(combined)
combined = Dense(1, activation="softmax")(combined)

model = Model(inputs=[branchA.input, branchB.input], outputs=combined)

这里是另一个教程，它使用多个分支，但使用两个不同的输入，但粗略的过程是相同的。