This repository is the code for the paper: Progressive Stochastic Binarization.

@article{corr/abs-1904-02205,
  title     = {Progressive Stochastic Binarization of Deep Networks},
  author    = {David Hartmann and Michael Wand},
  journal   = {CoRR},
  volume    = {abs/1904.02205},
  year      = {2019},
  url       = {http://arxiv.org/abs/1904.02205},
}

1. Make sure you have Tensorflow v1.13.1 installed.
2. Install python requirements:

   pip install --user -r requirements.txt
   

3. Prepare the Imagenet-Dataset (as .tfrecords) as described in https://github.com/tensorflow/models/tree/master/official/resnet
4. Place the tfrecords of Imagenet in ./download/imagenet/

for the ResNet18-Tests

1. Train a ResNet18 from official Tensorflow-Models https://github.com/tensorflow/models/tree/master/official/resnet
2. Place the Checkpoints in ./ckpts_imgn/resnet18_slim

for the Classification Models

1. Run for every model from https://github.com/qubvel/classification_models/tree/master/classification_models that you want to evaluate the download script. E.g. for resnet50v2 run:

   py py/download_and_convert_keras_model.py resnet50v2
   

To run the experiments that produce the output to the tables of the paper run the scripts from the base directory. For instance: sh experiments/test_attention.sh

Please check the experiments for example scripts or check the optional arguments by

python bitnetwork.py --help