Keras将两个模型连接到一起

神经网络玩得越久就越会尝试一些网络结构上的大改动。

先说意图

有两个模型:模型A和模型B。模型A的输出可以连接B的输入。将两个小模型连接成一个大模型,A-B,既可以同时训练又可以分离训练

流行的算法里经常有这么关系的两个模型,对GAN来说,生成器和判别器就是这样子;对VAE来说,编码器和解码器就是这样子;对目标检测网络来说,backbone和整体也是可以拆分的。所以,应用范围还是挺广的。


实现方法

首先说明,我的实现方法不一定是最佳方法。也是实在没有借鉴到比较好的方法,所以才自己手动写了一个。

第一步,我们有现成的两个模型A和B;我们想把A的输出连到B的输入,组成一个整体C。

第二步, 重构新模型C;我的方法是:读出A和B各有哪些layer,然后一层一层重新搭成C

可以看一个自编码器的代码(本人所编写):

class AE:
    def __init__(self, dim, img_dim, batch_size):
        self.dim = dim
        self.img_dim = img_dim
        self.batch_size = batch_size
        self.encoder = self.encoder_construct()
        self.decoder = self.decoder_construct()

    def encoder_construct(self):
        x_in = Input(shape=(self.img_dim, self.img_dim, 3))
        x = x_in
        x = Conv2D(self.dim // 16, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(x)
        x = BatchNormalization()(x)
        x = LeakyReLU(0.2)(x)
        x = Conv2D(self.dim // 8, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(x)
        x = BatchNormalization()(x)
        x = LeakyReLU(0.2)(x)
        x = Conv2D(self.dim // 4, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(x)
        x = BatchNormalization()(x)
        x = LeakyReLU(0.2)(x)
        x = Conv2D(self.dim // 2, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(x)
        x = BatchNormalization()(x)
        x = LeakyReLU(0.2)(x)
        x = Conv2D(self.dim, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(x)
        x = BatchNormalization()(x)
        x = LeakyReLU(0.2)(x)
        x = GlobalAveragePooling2D()(x)
        encoder = Model(x_in, x)
        return encoder

    def decoder_construct(self):
        map_size = K.int_shape(self.encoder.layers[-2].output)[1:-1]
        # print(type(map_size))
        z_in = Input(shape=K.int_shape(self.encoder.output)[1:])
        z = z_in
        z_dim = self.dim
        z = Dense(np.prod(map_size) * z_dim)(z)
        z = Reshape(map_size + (z_dim,))(z)
        z = Conv2DTranspose(z_dim // 2, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(z)
        z = BatchNormalization()(z)
        z = Activation('relu')(z)
        z = Conv2DTranspose(z_dim // 4, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(z)
        z = BatchNormalization()(z)
        z = Activation('relu')(z)
        z = Conv2DTranspose(z_dim // 8, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(z)
        z = BatchNormalization()(z)
        z = Activation('relu')(z)
        z = Conv2DTranspose(z_dim // 16, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(z)
        z = BatchNormalization()(z)
        z = Activation('relu')(z)
        z = Conv2DTranspose(3, kernel_size=(5, 5), strides=(2, 2), padding='SAME')(z)
        z = Activation('tanh')(z)
        decoder = Model(z_in, z)
        return decoder

    def build_ae(self):
        input_x = Input(shape=(self.img_dim, self.img_dim, 3))
        x = input_x
        for i in range(1, len(self.encoder.layers)):
            x = self.encoder.layers[i](x)
        for j in range(1, len(self.decoder.layers)):
            x = self.decoder.layers[j](x)
        y = x
        auto_encoder = Model(input_x, y)
        return auto_encoder

模型A就是这里的encoder,模型B就是这里的decoder。所以,连接的精髓在build_ae()函数,直接用for循环读出各层,然后一层一层重新构造新的模型,从而实现连接效果。因为keras也是基于图的框架,这个操作并不会很费时,因为没有实际地计算。

仅供参考。

你可能感兴趣的:(keras那些事儿,python,Computer,Vision)