lix19937/yolo_quantization

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

yolo_quantization

https://blog.csdn.net/qq_29462849/article/details/118159941

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 >

[Requirments]

MKL:

If you want to use mkl to accelerate, you need to install mkl by yourself, else do not need to install mkl

1. Download and Install MKL: https://pan.baidu.com/s/1vl8W7gp1MS_E_owgc6zrkA password: v37i
2. Fit MKL
   1. For Linux, change MKLROOT in Makefile to your own mkl install path
   2. For Windows, fit mkl path in setting of vs, follow this blog if you have no experience: https://www.cnblogs.com/Mayfly-nymph/p/11617651.html

[The Commond to Run My Project]

[Linux] 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)]

[Linux] Test:

set GPU=0, QUANTIZATION=1 in Makefile and OPENBLAS=1 if you use mkl make -j8

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

[Windows]
Test:

1. close macro OPENBLAS in vs, else open OPENBLAS to use mkl

2. yolo_quantization.exe detector test [abs path to data file] [abs path to cfg file] [abs path to weights file] [abs path to image file]

[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