Keras - 密集层与卷积2D层​​的融合

Question

我想制作一个自定义图层，用于将密集图层的输出与Convolution2D图层融合。

这个想法来自this paper，这是网络：

融合层试图将Convolution2D张量（256x28x28）与密集张量（256）融合。这是它的等式：

y_global => Dense layer output with shape 256 y_mid => Convolution2D layer output with shape 256x28x28

以下是有关Fusion流程的论文说明：

我最终制作了一个新的自定义图层，如下所示：

class FusionLayer(Layer):

    def __init__(self, output_dim, **kwargs):
        self.output_dim = output_dim
        super(FusionLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        input_dim = input_shape[1][1]
        initial_weight_value = np.random.random((input_dim, self.output_dim))
        self.W = K.variable(initial_weight_value)
        self.b = K.zeros((input_dim,))
        self.trainable_weights = [self.W, self.b]

    def call(self, inputs, mask=None):
        y_global = inputs[0]
        y_mid = inputs[1]
        # the code below should be modified
        output = K.dot(K.concatenate([y_global, y_mid]), self.W)
        output += self.b
        return self.activation(output)

    def get_output_shape_for(self, input_shape):
        assert input_shape and len(input_shape) == 2
        return (input_shape[0], self.output_dim)

我认为我的__init__和build方法正确，但我不知道如何将y_global（256个像素）与y-mid（256x28x28维度）连接起来）在call层中，以便输出与上面提到的等式相同。

如何在call方法中实现此等式？

非常感谢...

更新：成功整合这两层数据的任何其他方式对我来说也是可以接受的......它并非完全必须是文中提到的方式，但它至少需要返回一个可接受的方式输出...

Answer 1

我当时正在研究“图像着色”项目，最终遇到了融合层问题，然后我找到了一个包含融合层的模型。 希望可以在某种程度上解决您的问题。

    embed_input = Input(shape=(1000,))
    encoder_input = Input(shape=(256, 256, 1,))

    #Encoder
    encoder_output = Conv2D(64, (3,3), activation='relu', padding='same', strides=2,
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_input)
    encoder_output = Conv2D(128, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(128, (3,3), activation='relu', padding='same', strides=2,
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(256, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(256, (3,3), activation='relu', padding='same', strides=2,
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(512, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(512, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)
    encoder_output = Conv2D(256, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(encoder_output)

    #Fusion
    fusion_output = RepeatVector(32 * 32)(embed_input)
    fusion_output = Reshape(([32, 32, 1000]))(fusion_output)
    fusion_output = concatenate([encoder_output, fusion_output], axis=3)
    fusion_output = Conv2D(256, (1, 1), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(fusion_output)

    #Decoder
    decoder_output = Conv2D(128, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(fusion_output)
    decoder_output = UpSampling2D((2, 2))(decoder_output)
    decoder_output = Conv2D(64, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(decoder_output)
    decoder_output = UpSampling2D((2, 2))(decoder_output)
    decoder_output = Conv2D(32, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(decoder_output)
    decoder_output = Conv2D(16, (3,3), activation='relu', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(decoder_output)
    decoder_output = Conv2D(2, (3, 3), activation='tanh', padding='same',
                            bias_initializer=TruncatedNormal(mean=0.0, stddev=0.05))(decoder_output)
    decoder_output = UpSampling2D((2, 2))(decoder_output)

    model = Model(inputs=[encoder_input, embed_input], outputs=decoder_output)

以下是源链接： https://github.com/hvvashistha/Auto-Colorize

Answer 2

在我看来，实现一种新的层是一种复杂化这项任务的方法。我强烈建议您使用以下图层：

• Flatten，
• Merge，
• Dense，

以获得预期的行为。