wdlctc / QR-DNN

This repository represents the quantization and reconstruction algorithm (QR-DNN) for the FPL 2018 paper "RNA: An Accurate Residual Network Accelerator for Quantized and Reconstructed Deep Neural Networks"

Quantization and Reconstruction algorithm for Quantized and Reconstructed Deep Neural Networks

Code example for the FPL 2018 short paper. We modify the original caffe to implement our QR-DNN algorithm. And we also thank the example code provied by Ristretto.

Prerequisites

• NVIDIA GPU + CUDA + CuDNN
• python2.7
• Imagenet dataset

QR-DNN usage

Preparsion

make all

make pycaffe

Command to generate a fixed-point AlexNet model with 8-bit actiavtions and weight,

python python/merge_all.py --model=test/train_val.prototxt --mode=dynamic_fixed_point --solver_first=test/train_alex.solver

Command to generate a logrithmic AlexNet model with 8-bit actiavtions and 4-bit weight,

python python/merge_all.py --model=test/train_val.prototxt --mode=Power-of-two --solver_first=test/train_alex.solver --solver_second=test/train_alex_second.solver

Command to generate a fixed-point ResNet-50 model with 8-bit actiavtions and weight,

python python/merge_all.py --model=test_resnet/ResNet_50_train_val.prototxt --weight=test_resnet/ResNet-50-model.caffemodel --mode=dynamic_fixed_point --skip_first_phase=yes