FaceLandmark_PFLD_UltraLight icon indicating copy to clipboard operation
FaceLandmark_PFLD_UltraLight copied to clipboard

Published 20 hours ago • verified publisher icon AnthonyF333

训练多少epoches PFLD_Ultralight 0.25 112 NME可以达到0.06101?

Open isyanan1024 opened this issue 4 years ago • 14 comments

训练了200epoches，NME是0.07467567288161081

isyanan1024 avatar Dec 03 '20 02:12 isyanan1024

大概训练100个epoch，学习率降到1e-5网络会达到最优，再继续训练的话，网络性能提升不大，甚至会有些过拟合

AnthonyF333 avatar Dec 03 '20 05:12 AnthonyF333

NME较高的话，可以考虑增加数据增强时的repeat次数

AnthonyF333 avatar Dec 03 '20 05:12 AnthonyF333

强，效果很好，但是还有有个疑问，用训练好的PFLD_Ultralight 0.25 112模型在Mac Pro上cpu跑需要18ms左右，用ncnn跑也需要18ms左右，readme上说是5.5ms，是单纯的前向时间嘛? 而且一个ghostmoudle下来基本上就接近2ms了，是有什么trick嘛？ 希望大神解答一下，谢谢啦 @AnthonyF333

isyanan1024 avatar Dec 04 '20 02:12 isyanan1024

@isyanan1024 我是在Ubuntu的PC上跑出来的结果，可能cpu性能、系统等都会有影响

AnthonyF333 avatar Dec 04 '20 07:12 AnthonyF333

@isyanan1024 我的CPU是i5-7500 CPU @ 3.40GHz，Ubuntu16.04

AnthonyF333 avatar Dec 04 '20 07:12 AnthonyF333

@AnthonyF333 有与训练模型体提供吗

lucasjinreal avatar Feb 24 '22 02:02 lucasjinreal

@jinfagang https://drive.google.com/drive/folders/1zWCaRRPdzNh41XsOHXywP_FFBDX3wQK5?usp=sharing These are the ncnn format model files

AnthonyF333 avatar Feb 24 '22 02:02 AnthonyF333

@AnthonyF333 Do u have raw pytorch model? so that I can transfer to any other frameworks

lucasjinreal avatar Feb 24 '22 03:02 lucasjinreal

@jinfagang I have lost the raw pytorch models :(

AnthonyF333 avatar Feb 24 '22 08:02 AnthonyF333

nice

lucasjinreal avatar Feb 24 '22 14:02 lucasjinreal

@AnthonyF333 源码里面没有看到数据增强repeat在哪里设置，暂时没有找到

SankQin avatar Mar 15 '22 06:03 SankQin

@AnthonyF333 而且源码里面数据增强全部被注释了，如何做的数据增强

SankQin avatar Mar 15 '22 06:03 SankQin

@SankQin 在data/SetPreparation.py做数据增强

AnthonyF333 avatar Mar 16 '22 05:03 AnthonyF333

@AnthonyF333 DataSet里面随机处理效果是不是更好呢？

SankQin avatar Mar 16 '22 05:03 SankQin