Bayesnets

Some experiments and visualisations into Neural Nets inspired by Bayesian thinking.

See our Contributing Guide for an overview on the structure of this project + guidelines on doing things.

See our Blog for more info on our experiments, written in a way for others to consume.

Nomenclature

• Bee : A neural net training sequence
• Swarm : A group of networks trained the same way, with the only difference defined by starting conditions
• Hive : A set of swarms with some some training/initialisation parameter varied.

Animation Usage

Use the ./animate_training.py script for your a quick viz. You can specify depth (hidden layers) and width, even functions at the commandline. ./animate_training.py --help should give you some guidance on what can be done.

Example usages

./animate_training.py -h 3 -w 10  --func exp
./animate_training.py -h 3 -w 10  --func sin -n 800 --xdomain 0:6.2
./animate_training.py -h 3 -w 10  --func exp --numtrains 3
./animate_training.py -h 3 -w 15  --func sin -n 800 --xdomain -6.1:6.2 --lr 0.004

Most of the time is spent in generating the animation ~ 30s for the default settings

Getting Started

Tooling

1. pre-commit - Follow, https://pre-commit.com/.
   • Preferred method is to install using brew or to system python
   • pre-commit install so it will run and pre-commit run --all-files will initialise everything
   • This runs some formatting and simple static checks on the code.
   • This simply manages git hooks and is optional

Environment

If you use a venv, the --user flag won't be necessary

pip3 install --user -r requirements.txt
pip3 install --user -r requirements-dev.txt

For the Animation Backend (FFMPEG)

# OS X
brew install ffmpeg

# Ubuntu
sudo apt install ffmpeg