Keras 自定义层每次接收相同的输入形状

Question

我在扩展 keras 层的 keras 中编写 OctConv 卷积层，我编写了以下代码。

import keras.backend as K

from keras.layers import Layer, UpSampling2D, Add, Concatenate, Conv2D, Conv2DTranspose

class OCTCONV_LAYER(Layer):
    def __init__(self,
            filters=16,
            kernel_size=(3, 3),
            strides=(2, 2),
            dilation_rate=(1, 1),
            padding='same',
            alpha_in=0.6,
            alpha_out=0.6,
            **kwargs
        ):
        self.filters = filters
        self.kernel_size = kernel_size
        self.strides = strides
        self.padding = padding
        self.alpha_in = alpha_in
        self.alpha_out = alpha_out

        if dilation_rate[0] > 1:
            self.strides = (1, 1)

        self.dilation_rate = dilation_rate
        super(OCTCONV_LAYER, self).__init__(**kwargs)

    def build(self, input_shape):
        print('INPUT_SHAPE : {}'.format(input_shape))
        op_channels = self.filters
        low_op_channels = int(op_channels*self.alpha_out)
        high_op_channels = op_channels-low_op_channels

        inp_channels = input_shape[-1]
        low_inp_channels = int(inp_channels*self.alpha_in)
        high_inp_channels = inp_channels-low_inp_channels

        self.h_2_l = self.add_weight(
            name='hl',
            shape=self.kernel_size + (high_inp_channels, low_op_channels),
            initializer='he_normal'
        )
        self.h_2_h = self.add_weight(
            name='hh',
            shape=self.kernel_size + (high_inp_channels, high_op_channels),
            initializer='he_normal'
        )
        self.l_2_h = self.add_weight(
            name='lh',
            shape=self.kernel_size + (low_inp_channels, high_op_channels),
            initializer='he_normal'
        )
        self.l_2_l = self.add_weight(
            name='ll',
            shape=self.kernel_size + (low_inp_channels, low_op_channels),
            initializer='he_normal'
        )

        print('High 2 low : {}'.format(self.h_2_l.shape))
        print('High 2 high : {}'.format(self.h_2_h.shape))
        print('Low 2 high : {}'.format(self.l_2_h.shape))
        print('Low 2 low : {}'.format(self.l_2_l.shape))

        super(OCTCONV_LAYER, self).build(input_shape)

    def call(self, x):
        inp_channels = int(x.shape[-1])
        low_inp_channels = int(inp_channels*self.alpha_in)
        high_inp_channels = inp_channels-low_inp_channels

        high_inp = x[:,:,:, :high_inp_channels]
        print('High input shape : {}'.format(high_inp.shape))
        low_inp = x[:,:,:, high_inp_channels:]
        low_inp = K.pool2d(low_inp, (2, 2), strides=(2, 2), pool_mode='avg')
        print('Low input shape : {}'.format(low_inp.shape))

        out_high_high = K.conv2d(high_inp, self.h_2_h, strides=(2, 2), padding='same')
        print('OUT HIGH HIGH shape : {}'.format(out_high_high.shape))
        out_low_high = UpSampling2D((2, 2))(K.conv2d(low_inp, self.l_2_h, strides=(2, 2), padding='same'))
        print('OUT LOW HIGH shape : {}'.format(out_low_high.shape))
        out_low_low = K.conv2d(low_inp, self.l_2_l, strides=(2, 2), padding='same')
        print('OUT LOW LOW shape : {}'.format(out_low_low.shape))
        out_high_low = K.pool2d(high_inp, (2, 2), strides=(2, 2), pool_mode='avg')
        out_high_low = K.conv2d(out_high_low, self.h_2_l, strides=(2, 2), padding='same')
        print('OUT HIGH LOW shape : {}'.format(out_high_low.shape))

        out_high = Add()([out_high_high, out_low_high])

        print('OUT HIGH shape : {}'.format(out_high.shape))

        out_low = Add()([out_low_low, out_high_low])

        out_low = UpSampling2D((2, 2))(out_low)

        print('OUT LOW shape : {}'.format(out_low.shape))

        out_final = K.concatenate([out_high, out_low], axis=-1)
        print('OUT SHAPE : {}'.format(out_final.shape))

        out_final._keras_shape = self.compute_output_shape(out_final.shape)

        return out_final

    def compute_output_shape(self, inp_shape):
        return inp_shape

要创建图层，我使用以下代码来创建模型

from keras.layers import Input

inp = Input(shape=(224, 224, 3))
x = OCTCONV_LAYER(filters=16)(inp)
x = OCTCONV_LAYER()(x)
...

控制台上的输出是，

如您所见，最后一层的输入形状与输入层相同，而第一个 octconv 层的输出形状不是输入形状。 代码有什么问题？我错过了什么吗？

Answer 1

您的代码可能没问题。 K.conv2d 似乎存在问题，正如在 this answer 中观察到的那样。解决此问题的一种解决方法是从 tensorflow 导入 keras（例如，将所有 keras 导入更改为 tensorflow.keras）。

然后您需要进行以下更改：

low_inp_channels = int(int(inp_channels) * self.alpha_in)
high_inp_channels = int(inp_channels) - low_inp_channels

在这种情况下，第二个OCTCONV_LAYER 的输入维度对应于第一个OCTCONV_LAYER 的输出维度。经过前面的修改，结果是：

第一层

INPUT_SHAPE : (?, 224, 224, 3)
High 2 low : (3, 3, 2, 9)
High 2 high : (3, 3, 2, 7)
Low 2 high : (3, 3, 1, 7)
Low 2 low : (3, 3, 1, 9)
High input shape : (?, 224, 224, 2)
Low input shape : (?, 112, 112, 1)
OUT HIGH HIGH shape : (?, 112, 112, 7)
OUT LOW HIGH shape : (?, 112, 112, 7)
OUT LOW LOW shape : (?, 56, 56, 9)
OUT HIGH LOW shape : (?, 56, 56, 9)
OUT HIGH shape : (?, 112, 112, 7)
OUT LOW shape : (?, 112, 112, 9)
OUT SHAPE : (?, 112, 112, 16)

第二层

INPUT_SHAPE : (?, 112, 112, 16)
High 2 low : (3, 3, 7, 9)
High 2 high : (3, 3, 7, 7)
Low 2 high : (3, 3, 9, 7)
Low 2 low : (3, 3, 9, 9)
High input shape : (?, 112, 112, 7)
Low input shape : (?, 56, 56, 9)
OUT HIGH HIGH shape : (?, 56, 56, 7)
OUT LOW HIGH shape : (?, 56, 56, 7)
OUT LOW LOW shape : (?, 28, 28, 9)
OUT HIGH LOW shape : (?, 28, 28, 9)
OUT HIGH shape : (?, 56, 56, 7)
OUT LOW shape : (?, 56, 56, 9)
OUT SHAPE : (?, 56, 56, 16)