Overall Ask:

Hi all, I am getting an attribute error and I am not quite sure what the cause is. I am following this autoencoder tutorial and trying to replicate some of the code. Can you tell me where I am going wrong?

Autoencoder Build:

# This is the dimension of the original space
input_dim = maxlen

# This is the dimension of the latent space (encoding space)
latent_dim = 2

encoder = Sequential([
    Dense(128, activation='relu', input_shape=(input_dim,)),
    Dense(64, activation='relu'),
    Dense(32, activation='relu'),
    Dense(latent_dim, activation='relu')
])

decoder = Sequential([
    Dense(64, activation='relu', input_shape=(latent_dim,)),
    Dense(128, activation='relu'),
    Dense(256, activation='relu'),
    Dense(input_dim, activation=None)
])

When the encoder and decoder are connected in the following part, this is when the error arises:

autoencoder = Model(inputs=encoder.input, outputs=decoder(encoder.output))

The error:

AttributeError: 'Sequential' object has no attribute '_nested_inputs'
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<command-55734003> in <module>
----> 1 autoencoder = Model(inputs=encoder.input, outputs=decoder(encoder.output))
      2 
      3 autoencoder.compile(loss='mse', optimizer='adam')

/databricks/python/lib/python3.8/site-packages/tensorflow/python/keras/engine/functional.py in input(self)
    241       AttributeError: If no inbound nodes are found.
    242     """
--> 243     return self._nested_inputs
    244 
    245   @property

AttributeError: 'Sequential' object has no attribute '_nested_inputs'

What I tried:

1. I googled the error, and a post suggested altering the input shape. I do not know a lot about Autoencoders, but the architecture appears to already have this parameter.

2. Another post suggested adding input_shape (which is already there), and then using build()

Any advice would be appreciated!

CodePudding user response：

The simplest way to combine your encoder and dencoder is to instantiate an Input layer and pass it through encoder and decoder.

inp = Input((input_dim,))
autoencoder = Model(inputs=inp, outputs=decoder(encoder(inp)))
autoencoder.compile('adam', 'mse')

here is the running notebook