Keras Dense层报错：输入最后维度未定义，输入形状为(None, None)

问题代码 #

class modelSIAMESE(keras.Model):
def __init__(self):
    super().__init__()
    input_shape = target_shape + (3,)
    input_a = Input(shape = input_shape)
    input_b = Input(shape = input_shape)
    input_activation = 'relu'
    hidden_activation = 'relu'
    output_activation = 'sigmoid'
    self.conv1 = Conv2D(128, kernel_size = (3, 3), activation=input_activation, input_shape=input_shape)
    self.pool1 = MaxPooling2D(pool_size = (2, 2))
    self.conv2 = Conv2D(128, kernel_size = (3, 3), activation=hidden_activation)
    self.pool2 = MaxPooling2D(pool_size = (2, 2))
    self.conv3 = Conv2D(128, kernel_size = (3, 3), activation=hidden_activation)
    self.pool3 = MaxPooling2D(pool_size = (2, 2))
    self.flatten = Flatten()
    self.dense1 = Dense(128, activation="relu", input_shape=(None,))
    # Procesar las imágenes de entrada a través de las capas compartidas
    self.norm_a = layers.BatchNormalization() #(self.flatten(self.pool3(self.conv3(self.pool2(self.conv2(self.pool1(self.conv1(input_a))))))))
    self.norm_b = layers.BatchNormalization() #(self.flatten(self.pool3(self.conv3(self.pool2(self.conv2(self.pool1(self.conv1(input_b))))))))
    # Definir la capa de comparación
    self.distance = keras.layers.Subtract()
    self.prediction = Dense(units=1, activation=output_activation, input_shape=(None,None))
    self.dropout = keras.layers.Dropout(0.5)
def call(self, inputs, training=False):
    input_a = inputs["input_a"]
    input_b = inputs["input_b"]
    x = self.conv1(input_a)
    y = self.conv1(input_b)
    x = self.pool1(x)
    y = self.pool1(y)
    x = self.conv2(x)
    y = self.conv2(y)
    x = self.pool2(x)
    y = self.pool2(y)
    x = self.conv3(x)
    y = self.conv3(y)
    x = self.pool3(x)
    y = self.pool3(y)
    x = self.dense1(x)
    y = self.dense1(y)
    x = self.norm_a(x)
    y = self.norm_b(y)
    x = self.flatten(x)
    y = self.flatten(y)
    d = self.distance([x, y])
    d = K.abs(d)
    d = self.dropout(d, training=training)
    return self.prediction(d)

错误分析与修改方案 #

核心错误点 #

• 张量维度顺序错误：call方法中先调用Dense再Flatten，但pool3输出是4D张量（(batch_size, height, width, channels)），Dense层只能处理2D张量（(batch_size, features)），直接对接会导致维度不匹配。
• Dense层input_shape参数冗余且错误：子类化keras.Model时无需手动指定input_shape，框架会自动推导；self.prediction的input_shape=(None,None)完全不符合张量实际维度。
• 孪生网络参数未共享：self.norm_a和self.norm_b是两个独立的BatchNorm层，违背孪生网络共享特征提取参数的设计原则。
• 冗余Input定义：__init__中定义的input_a、input_b属于函数式API用法，在子类化Model中完全多余，会造成混淆。

修改后的完整代码 #

import keras
from keras import layers
from keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Input
import keras.backend as K

class modelSIAMESE(keras.Model):
    def __init__(self, target_shape):
        super().__init__()
        input_shape = target_shape + (3,)
        input_activation = 'relu'
        hidden_activation = 'relu'
        output_activation = 'sigmoid'
        
        # 共享特征提取网络
        self.conv1 = Conv2D(128, kernel_size=(3, 3), activation=input_activation, input_shape=input_shape)
        self.pool1 = MaxPooling2D(pool_size=(2, 2))
        self.conv2 = Conv2D(128, kernel_size=(3, 3), activation=hidden_activation)
        self.pool2 = MaxPooling2D(pool_size=(2, 2))
        self.conv3 = Conv2D(128, kernel_size=(3, 3), activation=hidden_activation)
        self.pool3 = MaxPooling2D(pool_size=(2, 2))
        self.flatten = Flatten()
        self.dense1 = Dense(128, activation="relu")
        self.norm = layers.BatchNormalization()  # 共享BatchNorm
        
        # 比较层
        self.distance = keras.layers.Subtract()
        self.dropout = keras.layers.Dropout(0.5)
        self.prediction = Dense(units=1, activation=output_activation)

    def call(self, inputs, training=False):
        input_a = inputs["input_a"]
        input_b = inputs["input_b"]
        
        # 处理输入A
        x = self.conv1(input_a)
        x = self.pool1(x)
        x = self.conv2(x)
        x = self.pool2(x)
        x = self.conv3(x)
        x = self.pool3(x)
        x = self.flatten(x)  # 先Flatten转为2D张量
        x = self.dense1(x)
        x = self.norm(x, training=training)
        
        # 处理输入B（共享同一套特征提取层）
        y = self.conv1(input_b)
        y = self.pool1(y)
        y = self.conv2(y)
        y = self.pool2(y)
        y = self.conv3(y)
        y = self.pool3(y)
        y = self.flatten(y)
        y = self.dense1(y)
        y = self.norm(y, training=training)
        
        # 计算距离并输出预测
        d = self.distance([x, y])
        d = K.abs(d)
        d = self.dropout(d, training=training)
        return self.prediction(d)

关键修改说明 #

• 调整张量处理顺序：将Flatten移至Dense之前，确保输入到Dense的是2D张量。
• 共享BatchNorm层：删除norm_a和norm_b，改用单个norm层，保证孪生网络两侧参数完全共享。
• 移除冗余参数与定义：删除__init__中多余的Input定义，以及Dense层错误的input_shape参数。
• 将target_shape作为构造参数传入：避免依赖全局变量，提升代码可复用性。

内容的提问来源于stack exchange，提问作者Alexander Pascau