a2e - a library to auto tune autoencoders

This library is based on Keras, SMAC and HpBandSter. It implements AutoML methods that are designed, but not limited to auto tune autoencoder architectures. Examples of its usage can be found in experiments.

Features

• Supporting SMAC and HpBandSter optimizers:
  • Bayesian Optimization using Random Forrest
  • Bayesian Optimization using Gaussian Process
  • Hyperband
  • Bayesian Optimization with Hyperband
  • Random
• Providing various cost functions focused on autoencoders
• Providing Keras models for vanilla, convolutional and recurrent autoencoders
• Providing experiment logging including git repository state, models (including weights and activations plots), predictions, samples, metrics

Installation as library

pip install git+https://github.com/maechler/a2e@master
pip install -r https://raw.githubusercontent.com/maechler/a2e/master/requirements.txt

Dataset

All experiments are implemented using datasets from a2e-data. The datasets are automatically downloaded and cached in ~/.a2e/.

System Requirements

• Python 3.8
• macOS, Ubuntu 18.04 or Windows
• SWIG (see SMAC3)
• graphviz

macOS installation

brew install swig
brew install graphviz

Ubuntu 18.04 installation

Please note that the installation of SMAC will most probably fail on Ubuntu 20.04, use Ubuntu 18.04 instead!

sudo apt-get install swig
sudo apt-get install graphviz

If you can still not build some dependencies, please make sure you have the correct python-dev version installed:

sudo apt install python3.8-dev

Windows installation

We recommend to use WSL 2 and install Ubuntu 18.04 as you will probably not be able to run the SMAC optimizers on Windows. But if you really want to use Windows, you will have to at least install SWIG and graphviz:

• SWIG:
  • Download swigwin-4.*.* from http://www.swig.org/download.html
  • Extract the folder and add it to the PATH environment variable
• graphviz: https://graphviz.org/download/

Development

Setup virtualenv

• Mac / Ubuntu

pip install virtualenv;
python -m virtualenv env;
source env/bin/activate;

• Windows (PowerShell)

pip install virtualenv;
python -m virtualenv env;
.\env\Scripts\activate.ps1

See https://docs.python.org/3/library/venv.html for further platforms and usages.

Install dependencies

pip install -r requirements.txt;
pip install -e .;

Testing

python -m unittest discover tests/*

Documentation

pdoc --html --force --output-dir doc a2e;
open doc/a2e/index.html;

Running an experiment

python experiments/feed_forward.py;

Run on a specific CUDA GPU

python a2e/runner/cuda_runner.py experiments/feed_forward.py --gpu=1