dan628 / stacked-denoising-autoencoder

Stacked Denoising Autoencoder package for feature extraction of high dimensional tabular data

SDAE

Description

SDAE is a package containing a stacked denoising autoencoder built on top of Keras that can be used to quickly and conveniently perform feature extraction on high dimensional tabular data.

The SDAE is a seven layer stacked denoising autoencoder designed to pass input data through a "bottleneck" layer before outputing a reconstruction of the input data as a prediction. Noise is introduced during training using dropout, and the model is trained to minimize reconstruction loss.

Features

• Adjustable noise levels
• Custom layer sizes
• Automatically plot reconstruction loss

Motivation

Stacked denosing autoencoders can serve as very powerful method of dimensionality reduction and feature extraction; However, testing these models can be time consuming. The goal of this package is to provide a flexible and convenient means of utilizing SDAEs using Scikit-learn-like syntax while preserving the funcionality provided by Keras.

Installation

pip install SDAE

Usage

Assume
Training data: train_data
Test data: test_data

from SDAE import StackedDenoisingAutoencoder

# Instantiate
SDAE_Model = StackedDenoisingAutoencoder(50, 5, 100, verbose=True)

# Train model
SDAE_Model.fit(X_train)

#Plot reconstruction loss during training
SDAE_Model.plot_loss()

#Access Keras model and functionality such as summary()
SDAE_Model.model.summary()

#Perform feature extraction
New_X_train = SDAE_Model.transform(X_train)
New_X_test = SDAE_Model.transform(X_test)

Contributing

Pull requests are welcome.

License

MIT