复值CNN报错：ComplexConv2D层输入维度不兼容问题求助

问题详情 #

我正在使用复值卷积神经网络处理复值IQ信号数据，运行如下代码时出现错误：

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from cvnn import *
import cvnn.layers as complex_layers
import tensorflow as tf


def createSB_modrelu():
    inputs = complex_layers.complex_input((2, 288))
    c0 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(inputs)
    c1 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(c0)
    c2 = complex_layers.ComplexMaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(c1)
    t01 = complex_layers.ComplexConv2DTranspose(5, kernel_size=2, strides=(2, 2), activation='cart_relu')(c2)
    concat01 = tf.keras.layers.concatenate([t01, c1], axis=-1)
    c3 = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(concat01)
    out = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(c3)
    return tf.keras.Model(inputs, out)

报错信息：

ValueError: Input 0 of layer "complex_conv2d" is incompatible with the layer: expected min_ndim=4, found ndim=3. Full shape received: (None, 2, 288)

错误原因 #

ComplexConv2D是2D卷积层，要求输入必须是4维张量，格式为(batch_size, height, width, channels)。但当前输入(None, 2, 288)是3维的，缺少必要的维度，导致层输入不兼容。

解决方法 #

根据IQ信号的特性，有两种可行的调整方式：

方式一：补充通道维度，适配2D卷积 #

将输入层的形状修改为4维，新增通道维度，让输入满足2D卷积的维度要求：

def createSB_modrelu():
    # 新增通道维度，输入形状变为(None, 2, 288, 1)
    inputs = complex_layers.complex_input((2, 288, 1))
    c0 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(inputs)
    c1 = complex_layers.ComplexConv2D(32, activation='cart_relu', kernel_size=3)(c0)
    c2 = complex_layers.ComplexMaxPooling2D(pool_size=(2, 2), strides=(2, 2), padding='valid')(c1)
    t01 = complex_layers.ComplexConv2DTranspose(5, kernel_size=2, strides=(2, 2), activation='cart_relu')(c2)
    concat01 = tf.keras.layers.concatenate([t01, c1], axis=-1)
    c3 = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(concat01)
    out = complex_layers.ComplexConv2D(4, activation='cart_relu', kernel_size=3)(c3)
    return tf.keras.Model(inputs, out)

方式二：改用1D卷积层处理序列IQ信号 #

如果IQ信号中288是时间步长、2是IQ分量，更适合用1D卷积处理序列数据，无需调整输入维度：

def createSB_modrelu():
    inputs = complex_layers.complex_input((2, 288))
    # 替换所有2D层为1D版本
    c0 = complex_layers.ComplexConv1D(32, activation='cart_relu', kernel_size=3)(inputs)
    c1 = complex_layers.ComplexConv1D(32, activation='cart_relu', kernel_size=3)(c0)
    c2 = complex_layers.ComplexMaxPooling1D(pool_size=2, strides=2, padding='valid')(c1)
    t01 = complex_layers.ComplexConv1DTranspose(5, kernel_size=2, strides=2, activation='cart_relu')(c2)
    concat01 = tf.keras.layers.concatenate([t01, c1], axis=-1)
    c3 = complex_layers.ComplexConv1D(4, activation='cart_relu', kernel_size=3)(concat01)
    out = complex_layers.ComplexConv1D(4, activation='cart_relu', kernel_size=3)(c3)
    return tf.keras.Model(inputs, out)

额外说明 #

• 若坚持使用2D卷积，需确保输入的每个维度对应空间维度（高度、宽度），通道维度不可省略。
• cvnn库的complex_input已自动处理复值数据的实部和虚部存储，无需额外拆分维度。

内容的提问来源于stack exchange，提问作者Fahmida Afrin