Search code examples
pythonkerasautoencoder

locallyconnected1D autoencoder in keras

I'm trying to create a LocallyConnected1D autoencoder in Keras by repurposing a 'simplest possible' Dense autoencoder from this tutorial.

I keep getting the error bellow, which I believe is being caused by the dimensions of my input_shape.

Traceback (most recent call last):
File "localdendritic.py", line 38, in <module>
kernel_size=6)
File "localdendritic.py", line 15, in __init__
activation='relu')(input_placeholder)
File "/Users/me/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 573, in __call__
self.assert_input_compatibility(inputs)
File "/Users/me/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 472, in assert_input_compatibility
str(K.ndim(x)))
ValueError: Input 0 is incompatible with layer encoded_layer: expected ndim=3, found ndim=2

My code is below. I've tried changing the input_shape array into [None, 1, input_size], [1, 1, input_size], [1, input_size] and [None, input_size], but it doesn't seem to change anything. I think I'm missing some insight about the shape of the input.

import numpy as np
from keras.models import Model, Sequential
from keras.layers import Input, LocallyConnected1D


class Localautoencoder:

  def __init__(self, input_size, encoded_size, kernel_size, **kwargs):

    input_shape = np.array([input_size])
    input_placeholder = Input(shape=(input_size, 1))
    encoded = LocallyConnected1D(encoded_size, kernel_size,
                                 input_shape=input_shape,
                                 name='encoded_layer',
                                 activation='relu')(input_placeholder)
    decoded = LocallyConnected1D(input_size, kernel_size,
                                activation='sigmoid',
                                name='decoded_layer')(encoded)
    self.localae = Model(input_placeholder, decoded)
    self.encoder = Model(input_placeholder, encoded)
    encoded_input = Input(shape=(1, encoded_size))
    decoded_layer = self.localae.layers[-1]
    self.decoder = Model(encoded_input, decoded_layer(encoded_input))
    self.localae.compile(optimizer='adam', loss='binary_crossentropy')


from keras.datasets import mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.astype('float32')/255.
x_test = x_test.astype('float32')/255.
x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))
x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
print(x_train.shape)
print(x_test.shape)

lae = Localautoencoder(input_size=x_train.shape[1],
                   encoded_size=100,
                   kernel_size=6)
Solution

  • A LocallyConnected1D layer takes a three-dimensional input, but input_placeholder is only two dimensions. A fix for this would be to add a Reshape layer that would convert your 2D input into a 3D one.