TensorFlow Convolutional AutoEncoder

This project provides utilities to build a deep Convolutional AutoEncoder (CAE) in just a few lines of code.

This project is based only on TensorFlow.

Experiments

convolutional_autoencoder.py shows an example of a CAE for the MNIST dataset.

The structure of this conv autoencoder is shown below:

The encoding part has 2 convolution layers (each followed by a max-pooling layer) and a fully connected layer. This part would encode an input image into a 20-dimension vector (representation). And then the decoding part, which has 1 fully connected layer and 2 convolution layers, would decode the representation back to a 28x28 image (reconstruction).

Training was done using GTX1070 GPU, batch size 100, 100000 passes.

Trained weights (saved in the saver directory) of the 1st convolutional layer are shown below:

And here's some of the reconstruction results:

Implementation

Un-pooling

Since the max-pooling operation is not injective, and TensorFlow does not have a built-in unpooling method, we have to implement our own approximation. But it is actually easy to do so using TensorFlow's tf.nn.conv2d_transpose() method.

The idea was to replace each entry in the pooled map with an NxM kernel with the original entry in the upper left, where N and M are the shape of the pooling kernel.

This is equivalent to doing transpose of conv2d on the input map with a kernel that has 1 on the upper left and 0 elsewhere. Therefore we could do this trick with tf.nn.conv2d_transpose() method.