An Empirical study of Binary Neural Networks' Optimisation

The source code used for experiments in the paper "An Empirical study of Binary Neural Networks' Optimisation".

The code grew organically as we tweaked more and more hyperparameters. Had I been more familiar with class-based declerations in TensorFlow (or embraded PyTorch sooner) the code would have been more elegant.

Environment

This code has been only tested with TensorFlow 1.8.0 and Python 3.5.4. The exact environment can be replicated by:

$ conda env create -f environment.yml

This would create a conda environment called studying-bnns.

Usage

$ conda activate studying-bnns

# Run an experiment by passing args
$ python run_with_args.py --model binary_connect_mlp \
    --dataset mnist --epochs 250 --batch-size 100 \
    --binarization deterministic-binary

# Run an experiment defined in a YAML file
python run_with_yaml.py some_experiment.yaml

An example experiment defintion in YAML file:

experiment-name: some_experiment

model: binary_connect_cnn
dataset: cifar10
epochs: 500
batch_size: 50

binarization: deterministic-binary

learning_rate:
  type: exponential-decay
  start: 3e-3
  finish: 2e-6

loss: 'square_hinge_loss'

optimiser:
  function: tf.train.AdamOptimizer
  kwargs: "{'beta1': 0.9, 'beta2': 0.999}"