Keras中组合Embedder与Recovery构建自编码器时出现'trainable_variables'属性错误的技术求助

嘿，这个问题我之前踩过坑，咱们来一步步拆解问题出在哪：

错误根源 #

你现在把embedder和recovery定义成了普通函数（或类内的普通方法），但Keras的trainable_variables是Layer或Model实例才有的属性——普通函数只是在构建计算图的逻辑，本身并没有被封装成可训练的模块，自然就没有这个属性啦。

修复方案：把模块封装成Keras子Model #

要解决这个问题，核心是把embedder和recovery这两个模块分别封装成独立的Keras Model实例，这样它们就会自带trainable_variables属性，也能更清晰地管理模块结构。

方案1：自定义Model类（推荐，适合复杂模型）

from tensorflow.keras import Model, Input, layers

class CustomAutoencoder(Model):
    def __init__(self, timesteps, feat, hidden_nodes, output_nodes, batch_size):
        super().__init__()
        # 初始化参数
        self.timesteps = timesteps
        self.feat = feat
        self.hidden_nodes = hidden_nodes
        self.output_nodes = output_nodes
        self.batch_size = batch_size
        
        # 构建两个子模块（作为Model实例）
        self.embedder = self._build_embedder()
        self.recovery = self._build_recovery()
        
    def _build_embedder(self):
        """构建编码器子模型"""
        inputs = Input(shape=[self.timesteps, self.feat], batch_size=self.batch_size)
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(inputs)
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(x)
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(x)
        outputs = layers.Dense(self.hidden_nodes, activation='sigmoid')(x)
        return Model(inputs=inputs, outputs=outputs)
    
    def _build_recovery(self):
        """构建解码器子模型（注意输入形状要和编码器输出匹配）"""
        inputs = Input(shape=[self.timesteps, self.hidden_nodes])
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(inputs)
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(x)
        x = layers.LSTM(self.hidden_nodes, return_sequences=True)(x)
        outputs = layers.Dense(self.output_nodes, activation='sigmoid', name='OUTPUT')(x)
        return Model(inputs=inputs, outputs=outputs)
    
    def call(self, inputs):
        """定义完整前向传播逻辑"""
        h = self.embedder(inputs)
        x_tilde = self.recovery(h)
        return x_tilde

使用方式：

# 替换成你的实际参数
TIMESTEPS = 10
FEAT = 5
HIDDEN_NODES = 32
OUTPUT_NODES = 5
BATCH_SIZE = 16

# 实例化并构建模型
autoencoder = CustomAutoencoder(TIMESTEPS, FEAT, HIDDEN_NODES, OUTPUT_NODES, BATCH_SIZE)
autoencoder.build(input_shape=(BATCH_SIZE, TIMESTEPS, FEAT))
autoencoder.summary()

# 现在可以正常获取可训练变量了
var_list = autoencoder.embedder.trainable_variables + autoencoder.recovery.trainable_variables

方案2：纯函数式API（适合简单场景）

如果不想写类，也可以直接用函数式API把两个模块封装成独立Model：

from tensorflow.keras import Model, Input, layers

def build_embedder(timesteps, feat, hidden_nodes, batch_size):
    inputs = Input(shape=[timesteps, feat], batch_size=batch_size)
    x = layers.LSTM(hidden_nodes, return_sequences=True)(inputs)
    x = layers.LSTM(hidden_nodes, return_sequences=True)(x)
    x = layers.LSTM(hidden_nodes, return_sequences=True)(x)
    outputs = layers.Dense(hidden_nodes, activation='sigmoid')(x)
    return Model(inputs=inputs, outputs=outputs)

def build_recovery(timesteps, hidden_nodes, output_nodes):
    inputs = Input(shape=[timesteps, hidden_nodes])
    x = layers.LSTM(hidden_nodes, return_sequences=True)(inputs)
    x = layers.LSTM(hidden_nodes, return_sequences=True)(x)
    x = layers.LSTM(hidden_nodes, return_sequences=True)(x)
    outputs = layers.Dense(output_nodes, activation='sigmoid', name='OUTPUT')(x)
    return Model(inputs=inputs, outputs=outputs)

# 构建完整自编码器
embedder = build_embedder(TIMESTEPS, FEAT, HIDDEN_NODES, BATCH_SIZE)
recovery = build_recovery(TIMESTEPS, HIDDEN_NODES, OUTPUT_NODES)

X = Input(shape=[TIMESTEPS, FEAT], batch_size=BATCH_SIZE, name='RealData')
H = embedder(X)
X_tilde = recovery(H)
autoencoder = Model(inputs=X, outputs=X_tilde)

# 正常获取可训练变量
var_list = embedder.trainable_variables + recovery.trainable_variables

总结 #

本质就是把原来的普通函数逻辑，包装成Keras官方认可的Model实例——这样不仅能解决trainable_variables的属性问题，还能让模型结构更清晰，方便后续的模块复用、单独训练等操作。

内容的提问来源于stack exchange，提问作者HenDoNR