BWN_Shift

Codes for “Towards State-Aware Computation in ReRAM Neural Networks,”, Yintao He, Ying Wang, Xiandong Zhao, Huawei Li, Xiaowei Li, in IEEE/ACM Proceedings of Design, Automation Conference (DAC), 2020.

Install

git clone <.git>

cd BWN_Shift

export PYTHONPATH=$PYTHONPATH:`pwd`

pip install -r requirements.txt

Codes

The folder contains three important parts:

• utils/ contains scripts help us get started quickly
• examples/ contains *.py for training and validation of network
• models/ contains files of BWN-Shift, BWN without Shift, and full-precision model structure

Run

Run scripts with following (available dataset: "mnist", "cifar10", "svhn"):

cd scripts
./[scripts_name].sh [GPU_id] [log_name] [learning_rate]

1. Obtain full-precision model for follow-up binary training

./train_mnist_lenet.sh 0 baseline 1e-3

2. Train BWN-Shift model

./train_mnist_lenet_bwnshift.sh 0 test 1e-8

3. Train BWN without Shift to compare

./train_mnist_lenet_bwnall.sh 0 test 1e-5

• All the models will be stored in /scripts/logger
• All the layers has been binarized in this work