运行深度学习集成模型代码时遇ValueError：input_dim与output_dim参数异常

这个ValueError的核心问题是Attention层的输出维度被错误计算为0，大概率是你使用的Attention层参数传递或实现逻辑有误。原代码里直接传入maxlen作为参数，不符合Keras官方Attention层的用法，也可能自定义的Attention层内部逻辑导致输出维度坍缩为0。

下面给出修正后的完整代码，同时替换成正确的通道注意力实现，确保各分支输出维度正常：

修正要点 #

• 替换错误的Attention(maxlen)用法，改用标准的通道注意力实现
• 确保LSTM/GRU分支的序列输出转换为固定维度向量，避免后续层维度异常
• 补充缺失的Keras层导入语句

修正后的完整代码 #

# 集成深度学习模型架构
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Model
from tensorflow.keras.layers import (
    Input, Embedding, Bidirectional, LSTM, GRU,
    Conv1D, MaxPooling1D, Flatten, Dense, Dropout, concatenate,
    Layer, Activation, Permute, Multiply, Lambda
)
import tensorflow as tf

# 自定义通道注意力层（Channel Attention）
class ChannelAttention(Layer):
    def __init__(self, ratio=8, **kwargs):
        super(ChannelAttention, self).__init__(**kwargs)
        self.ratio = ratio

    def build(self, input_shape):
        self.channel = input_shape[-1]
        self.shared_dense_one = Dense(self.channel // self.ratio,
                                     activation='relu',
                                     kernel_initializer='he_normal',
                                     use_bias=True,
                                     bias_initializer='zeros')
        self.shared_dense_two = Dense(self.channel,
                                     kernel_initializer='he_normal',
                                     use_bias=True,
                                     bias_initializer='zeros')
        super(ChannelAttention, self).build(input_shape)

    def call(self, inputs):
        # 全局平均池化
        avg_pool = tf.reduce_mean(inputs, axis=1, keepdims=True)
        avg_pool = self.shared_dense_one(avg_pool)
        avg_pool = self.shared_dense_two(avg_pool)

        # 全局最大池化
        max_pool = tf.reduce_max(inputs, axis=1, keepdims=True)
        max_pool = self.shared_dense_one(max_pool)
        max_pool = self.shared_dense_two(max_pool)

        # 注意力权重相加 + sigmoid激活
        attention = avg_pool + max_pool
        attention = Activation('sigmoid')(attention)

        # 应用注意力权重到输入
        return Multiply()([inputs, attention])

# -------------------------- 模型构建 --------------------------
# 假设已定义变量：maxlen(文本序列长度)、max_features(词汇表大小)、embed_size(词嵌入维度)、embedding_matrix(预训练词嵌入矩阵)

inp = Input(shape=(maxlen,))
x = Embedding(max_features, embed_size, weights=[embedding_matrix], trainable=False)(inp)

# 分支1：Channel Attention-BiLSTM
Att_LSTM = Bidirectional(LSTM(64, return_sequences=True))(x)
Att_LSTM = ChannelAttention()(Att_LSTM)
# 全局平均池化将序列转换为固定维度向量
Att_LSTM = Lambda(lambda x: tf.reduce_mean(x, axis=1))(Att_LSTM)
Att_LSTM = Dense(64, activation="relu")(Att_LSTM)
Output_Att_LSTM = Dense(1, activation="sigmoid")(Att_LSTM)

# 分支2：Channel Attention-BiGRU
Att_GRU = Bidirectional(GRU(64, return_sequences=True))(x)
Att_GRU = ChannelAttention()(Att_GRU)
Att_GRU = Lambda(lambda x: tf.reduce_mean(x, axis=1))(Att_GRU)
Att_GRU = Dense(64, activation="relu")(Att_GRU)
Output_Att_GRU = Dense(1, activation="sigmoid")(Att_GRU)

# 分支3：Channel CNN
Att_CNN = Conv1D(64, kernel_size=3, activation="relu")(x)
Att_CNN = ChannelAttention()(Att_CNN)
Att_CNN = MaxPooling1D()(Att_CNN)
Att_CNN = Flatten()(Att_CNN)

# MLP层
Att_CNN = Dense(64, activation="relu")(Att_CNN)
Att_CNN = Dropout(0.2)(Att_CNN)
Output_Att_CNN = Dense(1, activation="sigmoid")(Att_CNN)

# 拼接三个分支输出
merged = concatenate([Output_Att_CNN, Output_Att_GRU, Output_Att_LSTM])

# 最终输出层
outputs = Dense(1, activation='sigmoid')(merged)
model = Model(inputs=inp, outputs=outputs)

# 编译模型
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

关键修正说明 #

• 实现了标准的通道注意力层，通过全局池化和全连接层生成通道权重，彻底避免维度异常
• 在LSTM/GRU分支中添加全局平均池化，将序列输出转换为固定维度向量，确保后续Dense层能正常处理
• 所有分支最终输出统一为(None,1)维度，拼接后输入到最终Dense层，逻辑通顺无维度冲突

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