This repository represents training examples for the CVPR 2018 paper "SYQ:Learning Symmetric Quantization For Efficient Deep Neural Networks"

Python 2.7 or 3
Python bindings for OpenCV
Tensorflow == 0.12.1 (pip install tensorflow-gpu==0.12.1)

export PYTHONPATH=$PYTHONPATH:readlink -f path/to/tensorpack

python syq-imagenet-alexnet.py --data <enter path to data> --num-epochs 56 84 120 --learning-rate 1e-4 2e-5 4e-6 --load <enter checkpoint or npy path> --eta 0.0

python syq-imagenet-alexnet.py --data <enter path to data> --num-epochs 56 84 120 --learning-rate 1e-4 2e-5 4e-6 --load <enter checkpoint or npy path> --eta 0.0 --load <PATH>/checkpoint

--gpu >Set which gpu's you want to instantiate (example: --gpu 0,1)
--load >Load either npy or checkpoint file as a pretrained model by entering its path
--data >Path to training and validation data
--run >Enter image files along with a pretrained models to compute inference
--eta >Quantization Threshold value (eta=0 for binary networks and defaults to 0.05 for ternary)
--learning-rate >enter a list of learning rate factors at each step
--num-epochs >Enter the epochs when the learning rate changes, the last value is the total epochs
--in-epochs >Enter epochs which will compute the validation error. It automatically compute all epochs after the last
learning rate change. Default is computing validation error for all epochs
--eval >Evaluate the model on test or validation set
--name >Sets the name of the folder for storing training results