将2个CNN合并到Keras中的自动编码器中

Question

I am trying to merge 2 CNNs and then un-merge them through an autoencoder but I always get the same error message saying "ValueError: total size of new array must be unchanged", not sure what is wrong with my network.

I am replacing the LSTM and CNN by Dense layers and it works.

Thanks for helping!

from keras.layers import Input, Dense, Conv1D, MaxPooling1D, UpSampling1D, Embedding, Dropout, Flatten, Concatenate, LSTM, BatchNormalization, Reshape
from keras.models import Model
from keras import backend as K
from keras.callbacks import TensorBoard


input_address = Input(shape=(38,))
x_address = Reshape((38, 1))(input_address)
x_address = LSTM(125, activation="tanh", return_sequences=True)(x_address)
x_address = Dropout(0.4)(x_address)
x_address = LSTM(125, activation="tanh", return_sequences=True)(x_address)
x_address = Dropout(0.4)(x_address)
x_address = Dense(100, activation = 'relu')(x_address)
x_address = Conv1D(100, 2, activation='relu', padding='same')(x_address)
x_address = MaxPooling1D(2, padding='same')(x_address)
x_address = Dropout(0.5)(x_address)
x_address = Flatten()(x_address)

input_name = Input(shape=(29,))
x_name = Reshape((29, 1))(input_name)
x_name = LSTM(125, activation="tanh", return_sequences=True)(x_name)
x_name = Dropout(0.4)(x_name)
x_name = LSTM(125, activation="tanh", return_sequences=True)(x_name)
x_name = Dropout(0.4)(x_name)
x_name = Dense(100, activation = 'relu')(x_name)
x_name = Conv1D(100, 2, activation='relu', padding='same')(x_name)
x_name = MaxPooling1D(2, padding='same')(x_name)
x_name = Dropout(0.5)(x_name)
x_name = Flatten()(x_name)

merge = Concatenate(name="concat", axis=1)([x_name, x_address])
encoder = Dense(1, activation = 'relu')(merge)


xd_address = Reshape((19, 100))(encoder)
xd_address = UpSampling1D(2)(xd_address)
xd_address = Conv1D(100, 2, activation='relu', padding='same')(xd_address)
xd_address = Dropout(0.4)(xd_address)
xd_address = LSTM(125, activation="tanh", return_sequences=True)(xd_address)
xd_address = Dropout(0.4)(xd_address)
xd_address = LSTM(125, activation="tanh", return_sequences=True)(xd_address)
xd_address = Flatten()(xd_address)

xd_name = Reshape((15, 100))(encoder)
xd_name = UpSampling1D(2)(xd_name)
xd_name = Conv1D(100, 2, activation='relu', padding='same')(xd_name)
xd_name = Dropout(0.4)(xd_name)
xd_name = LSTM(125, activation="tanh", return_sequences=True)(xd_name)
xd_name = Dropout(0.4)(xd_name)
xd_name = LSTM(125, activation="tanh", return_sequences=True)(xd_name)
xd_name = Flatten()(xd_name)

autoencoder = Model(inputs=[input_name, input_address], outputs=[xd_name, xd_address])
autoencoder.compile(optimizer='adam', loss=['mse', 'mse'])

autoencoder.summary()

autoencoder.fit([name_train, address_train], [name_train, address_train]
                 ,epochs=10
                 ,batch_size=120
                 ,shuffle=True
                 ,validation_data=([name_test, address_test], [name_test, address_test])
                 ,callbacks=[TensorBoard(log_dir='/tmp/autoencoder')])

Here is the error message I get

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-132-19e453d673b6> in <module>()
     33 
     34 
---> 35 xd_address = Reshape((19, 100))(encoder)
     36 xd_address = UpSampling1D(2)(xd_address)
     37 xd_address = Conv1D(100, 2, activation='relu', padding='same')(xd_address)

~\Anaconda3\lib\site-packages\keras\engine\base_layer.py in __call__(self, inputs, **kwargs)
    472             if all([s is not None
    473                     for s in to_list(input_shape)]):
--> 474                 output_shape = self.compute_output_shape(input_shape)
    475             else:
    476                 if isinstance(input_shape, list):

~\Anaconda3\lib\site-packages\keras\layers\core.py in compute_output_shape(self, input_shape)
    392             # input shape known? then we can compute the output shape
    393             return (input_shape[0],) + self._fix_unknown_dimension(
--> 394                 input_shape[1:], self.target_shape)
    395 
    396     def call(self, inputs):

~\Anaconda3\lib\site-packages\keras\layers\core.py in _fix_unknown_dimension(self, input_shape, output_shape)
    380             output_shape[unknown] = original // known
    381         elif original != known:
--> 382             raise ValueError(msg)
    383 
    384         return tuple(output_shape)

ValueError: total size of new array must be unchanged

Answer 1

You are reshaping into a size that changes the number elements in the original encoder tensor. You need to:

1. print(encoder) and that will give you a hint on what the original shape is.
2. Make sure to Reshape while preserving the number of elements.

Here is a short example, (100,) cannot be reshaped to (2,45) because you omitted 10 elements, ie "total size of new array must be unchanged".