Hands-On Machine Learning for Algorithmic Trading

One-layer feedforward autoencoder

We start with a vanilla feedforward autoencoder with a single hidden layer to illustrate the general design approach using the functional Keras API and establish a performance baseline.

The first step is a placeholder for the flattened image vectors with 784 elements:

input_ = Input(shape=(input_size,), name='Input')

The encoder part of the model consists of a fully-connected layer that learns the new, compressed representation of the input. We use 32 units for a compression ratio of 24.5:

encoding_size = 32 # compression factor: 784 / 32 = 24.5encoding = Dense(units=encoding_size,activation='relu',name='Encoder')(input_)

The decoding part reconstructs the compressed data to its original size in a single ...