opendacs
/
DeepBurning-MixQ
forked from liucheng/DeepBurning-MixQ
9
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
9 Commits 2 Branches 0 Tags 145 KiB
Python 100%
Go to file
fffasttime 71737900d1 add dacsdc model quant check 2022-11-14 21:03:38 +08:00
anypacking add dacsdc model quant check 2022-11-14 21:03:38 +08:00
cifar add cifar/simulate_hw.py; check&fix quantization bug 2022-11-07 00:03:09 +08:00
dacsdc add dacsdc model quant check 2022-11-14 21:03:38 +08:00
utils merge cifar dateset code (WIP) 2022-10-28 12:50:34 +08:00
.gitignore merge cifar dateset code (WIP) 2022-10-28 12:50:34 +08:00
__init__.py merge cifar dateset code (WIP) 2022-10-28 12:50:34 +08:00
readme.md init commit 2022-10-14 12:47:16 +08:00

readme.md

train

  • python train.py --multi-scale --img-size 320 --multi-scale --batch-size 32

reference

22修改

  • python train22.py 训练非量化模型时不用加其他参数
  • train22.py 80行改使用的模型，初始训练使用UltraNet_Float或RepVGG
  • QAT时把模型改为量化模型例如UltraNet_ismart，并且用--weights载入全精度权重
  • 测试精度使用python test.py --weights <权重路径> --model <模型类名称>

注意

  • 余弦学习率衰减最好跑满全部epoch
  • RepVGG全精度训练后量化训练精度更好
  • quant_dorefa.py 第18行改权重量化为[-8,7]还是[-7,7]