Traffic Autoencoder

What is this?

This repository contains the code used for encoding traffic data in my Master's Thesis (which can be found here).

What is it useful for?

The code has commands for:

• Tune: perform a hyperparamater grid search to look for optimal hyperparameters for the autoencoder.
• Train: use the data to train an autoencoder and export both the autoencoder and the encoder part as pickled Keras models ready to use for encoding.
• Encode: load a pickled encoder and use it on some data to obtained the encoded representation.

How to run?

The repository is designed as a command line application.

First, install the package. Within a shell, cd into this directory and:

pip install .

There are three commands you can use: autoencoder-evaluate, autoencoder-train and autoencoder-train.

autoencoder-evaluate

autoencoder-evaluate --config-file /path/to/config.json \
    --experiment-definition-path /path/to/definitions/directory/

autoencoder-train

autoencoder-train --config-file /path/to/config.json \
    --model_path /path/to/model/

autoencoder-encode

autoencoder-encode --config-file /path/to/config.json \
    --definitions-path /path/to/definitions/directory/ \
    --encode-file /path/to/encoder.pickle

All commands require:

• A config file which must be in JSON format. A template can be found in the example-config.json file.
• The path to a folder that contains one or more "definition" files, in JSON format. In autoencoder-evaluate, one different definition file should be created for every hyperparameter combination that needs to be tested. In autoencoder-train, a single definition file defining the hyperparameters to use while training the model is required. In autoencoder-encode, a single definition file is required, and only for describing the data preprocessing bits, since the model already exists by then. A template can be found in the example-definition.json.