在 Keras 中组合模型（输出）

Posted 2023-02-16

【中文标题】在 Keras 中组合模型（输出）

【英文标题】：Combine Models (outputs) in Keras

【发布时间】：2021-12-22 06:58:40

【问题描述】：

我正在尝试构建以下论文中介绍的网络：link

基本上，自动编码器是其他两个模型的组合，嵌入器和恢复器如下所述：

X = Input(shape=[TIMESTEPS, FEAT], batch_size=BATCH_SIZE, name='RealData')

def recovery(self, H):

    L1 = LSTM(HIDDEN_NODES, return_sequences=True)(H)
    L2 = LSTM(HIDDEN_NODES, return_sequences=True)(L1)
    L3 = LSTM(HIDDEN_NODES, return_sequences=True)(L2)  
    O = Dense(OUTPUT_NODES, activation='sigmoid', name='OUTPUT')(L3)

    return O

def embedder(self, X):
    L1 = LSTM(HIDDEN_NODES, return_sequences=True)(X)
    L2 = LSTM(HIDDEN_NODES, return_sequences=True)(L1)
    L3 = LSTM(HIDDEN_NODES, return_sequences=True)(L2)      
    O = Dense(HIDDEN_NODES, activation='sigmoid')(L3)
    return O 

最后，将它们与以下几行结合起来：

    H = self.embedder(X) 

    X_tilde = self.recovery(H)

    self.autoencoder = Model(inputs=X, outputs=X_tilde)

显示自动编码器的.summary 我有以下内容：

然后出现以下错误：

var_list = self.embedder.trainable_variables + self.recovery.trainable_variables
AttributeError: 'function' object has no attribute 'trainable_variables'

我做错了什么？

我复制的基线代码可以在here找到

【参考方案1】：

问题是embedder 和recovery 不是带有trainable_variables 的模型。这两个函数只是返回最后一层的输出。也许尝试这样的事情：

import tensorflow as tf

X = tf.keras.layers.Input(shape=[10, 10], batch_size=2, name='RealData')

def recovery():
    model = tf.keras.Sequential([
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.Dense(10, activation='sigmoid', name='OUTPUT')
    ])
    return model

def embedder():
    model = tf.keras.Sequential([
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.LSTM(10, return_sequences=True),
      tf.keras.layers.Dense(10, activation='sigmoid')
    ])
    return model 


embedder_model = embedder() 
H = embedder_model(X)

recovery_model = recovery() 
X_tilde = recovery_model(H)

autoencoder = tf.keras.Model(inputs=X, outputs=X_tilde)

var_list = embedder_model.trainable_variables + embedder_model.trainable_variables

tf.print(var_list[:2])

[[[0.343916416 0.310338378 0.34440577 ... 0.0633761585 0.0405358076 0.276733816]
 [0.245998859 0.197870493 0.0333348215 ... -0.136249736 0.271893084 -0.0605607331]
 [-0.290359527 0.240957797 0.117871583 ... 0.172593892 0.113803834 0.0506341457]
 ...
 [0.15672195 -0.161336392 -0.13484776 ... 0.306486845 -0.0707859397 0.245753765]
 [0.00567743182 0.181330919 0.206510961 ... 0.0141542256 0.205756843 -0.074064374]
 [0.299010575 -0.236641362 0.272176802 ... 0.0658480823 0.04648754 -0.342863292]], [[0.224076748 -0.112819761 -0.114276126 ... -0.190908 -0.282466382 -0.0711786151]
 [-0.0689174235 0.203702673 -0.248280779 ... -0.0145524191 0.202952 0.0797807127]
 [0.0919017 0.108805738 -0.124872617 ... 0.26839748 0.21041657 0.251440644]
 ...
 [-0.117122218 -0.0974424109 -0.17138055 ... 0.150875479 0.0454813093 0.0753096]
 [-0.115990438 -0.360190183 -0.0988362879 ... -0.0655761734 0.11425022 0.0291871373]
 [-0.00164104556 -0.0442082509 0.135109842 ... -0.182655513 -0.0121813752 0.0497299805]]]

【讨论】：

成功了。谢谢！