This readme describes how to reproduce the PokeBNN paper on a google cloud TPU virtual machine.

Please refer to the section Reproduce AQT Experiments on GCloud TPUs in this documentation for setting up the gcloud command line tool, the cloud storage bucket, and downloading the ImageNet dataset.

The provided bash script cloudtpu_train_pokebnn.sh automates this step. Please copy the bash script to your local machine with gcloud command line tool installed and run the following command. Note that this script is "stand alone". It only requires gcloud CLI tool.

source cloudtpu_train_pokebnn.sh

1. Navigate to https://colab.research.google.com/
2. Open a new colab file. Load and run the provided tensorboard.ipynb. Please remember filling in the correct cloud storage bucket name and directory name that stores the data.

If you find PokeBNN and this repository useful, please cite:

@article{zhanglew2021pokebnn,
  title={PokeBNN: A Binary Pursuit of Lightweight Accuracy},
  author={Zhang, Yichi and Zhang, Zhiru and Lew, Lukasz},
  journal={arXiv preprint arXiv:2112.00133},
  year={2021}
}