By Guangrun Wang, Jiefeng Peng, Ping Luo, Xinjiang Wang and Liang Lin.

Sun Yat-sen University (SYSU), the Chinese University of Hong Kong (CUHK), SenseTime Group Ltd.

0. Results
1. Introduction
2. Citation
3. Dependencies
4. Usage

• Under the context of micro-batches(batch size = 2), the validataion curves on CIFAR10:

top line: Batch Normalization(BN); mid line: Group Normalization; bottom line: Kalman Normalization

• Under the context of large-batches(batch size = 128), the validataion curves on CIFAR10:

top line: Batch Normalization(BN); bottom line: Kalman Normalization)

This repository contains the original models described in the paper "Batch Kalman Normalization: Towards Training Deep Neural Networks with Micro-Batches" (https://arxiv.org/abs/1802.03133). These models are those used in ILSVRC and CIFAR

If you use these models in your research, please cite:

@article{wang2018batch,
	title={Batch Kalman Normalization: Towards Training Deep Neural Networks with Micro-Batches},
	author={Wang, Guangrun and Peng, Jiefeng and Luo, Ping and Wang, Xinjiang and Lin, Liang},
	journal={arXiv preprint arXiv:1802.03133},
	year={2018}
}

• Python 2.7 or 3
• TensorFlow >= 1.3.0
• Tensorpack The code depends on Yuxin Wu's Tensorpack. For convenience, we provide a stable version 'tensorpack-installed' in this repository.

  # install tensorpack locally:
  cd tensorpack-installed
  python setup.py install --user
  

• To run Group Normalization, use:

  cd KalmanNorm
  python cifar-gn.py --gpu 0 -n 5 --log_dir gn
  

• To run Batch Normalization under the context of micro-batches, use:

  cd KalmanNorm
  python cifar-bn-microbatch.py --gpu 0 -n 5 --log_dir  bn-microbatch
  

• To run Kalman Normalization under the context of micro-batches, use:

  cd KalmanNorm
  python cifar-kn-microbatch.py --gpu 0 -n 5 --log_dir  kn-microbatch
  

• To run Batch Normalization under the context of large-batches, use:

  cd KalmanNorm
  python cifar-bn-largebatch.py   --gpu 0 -n 18 --log_dir  bn-largebatch
  

• To run Kalman Normalization under the context of large-batches, use:

  cd KalmanNorm
  python cifar-kn-largebatch.py   --gpu 0 -n 18 --log_dir  kn-largebatch