PCA is neat but surely we can do better. This time we want you to build a deep convolutional autoencoder by... stacking more layers.
Encoder
The encoder part is pretty standard, we stack convolutional and pooling layers and finish with a dense layer to get the representation of desirable size (code_size).
We recommend to use activation='elu' for all convolutional and dense layers.
We recommend to repeat (conv, pool) 4 times with kernel size (3, 3), padding='same' and the following numbers of output channels: 32, 64, 128, 256.
Remember to flatten (L.Flatten()) output before adding the last dense layer!
Decoder
For decoder we will use so-called "transpose convolution".
Traditional convolutional layer takes a patch of an image and produces a number (patch -> number). In "transpose convolution" we want to take a number and produce a patch of an image (number -> patch). We need this layer to "undo" convolutions in encoder.
You can add "transpose convolution" layer in Keras like this:
L.Conv2DTranspose(filters=?, kernel_size=(3, 3), strides=2, activation='elu', padding='same')
Our decoder starts with a dense layer to "undo" the last layer of encoder. Remember to reshape its output to "undo" L.Flatten() in encoder.
Now we're ready to undo (conv, pool) pairs. For this we need to stack 4 L.Conv2DTranspose layers with the following numbers of output channels: 128, 64, 32, 3. Each of these layers will learn to "undo" (conv, pool) pair in encoder. For the last L.Conv2DTranspose layer use activation=None because that is our final image.