Skip to content

Based of paper "Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference"

Notifications You must be signed in to change notification settings

Wulingtian/yolo_quantization

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

yolo_quantization

The code is to quantization float32 network of darknet to uint8 network based of paper:

Quantization and Training of Neural Networks for Efficient

< https://arxiv.org/abs/1712.05877 >

[The Commond to Run My Project]

Train:

set GPU=1 in Makefile

make -j8

./darknet detector train cfg/voc_nok.data cfg/yolov3-tiny-mask_quant.cfg [pretrain weights file I gave to you(default in cfg folder)]

Test:

set GPU=0 in Makefile

make -j8

./darknet detector test cfg/voc_nok.data cfg/yolov3-tiny_quant.cfg [weights file] [image path]

[Pretrain Cfg file and Weights file]

https://pan.baidu.com/s/16_ULXdNPmIhoEmu7jXmkmQ 
password: qy8a 

[Performance]

quantization inference time (intel chip 64bit) recall precision f1 score
darknet 0.83s 74.43 89.45 81.25
quantization mine 0.34s 90.08 91.83 90.94

About

Based of paper "Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C 89.0%
  • Cuda 9.7%
  • Other 1.3%