search

AnthonyF333 / FaceLandmark_PFLD_UltraLight

Apache License 2.0
110 stars 32 forks source link

训练多少epoches PFLD_Ultralight 0.25 112 NME可以达到0.06101? #1

Open isyanan1024 opened 3 years ago

isyanan1024 commented 3 years ago

训练了200epoches，NME是0.07467567288161081

AnthonyF333 commented 3 years ago

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 commented 3 years ago

NME较高的话，可以考虑增加数据增强时的repeat次数

isyanan1024 commented 3 years ago

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

AnthonyF333 commented 3 years ago

@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 commented 3 years ago

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

lucasjinreal commented 2 years ago

@AnthonyF333 有与训练模型体提供吗

AnthonyF333 commented 2 years ago

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

lucasjinreal commented 2 years ago

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

AnthonyF333 commented 2 years ago

@jinfagang I have lost the raw pytorch models :(

lucasjinreal commented 2 years ago

nice

SankQin commented 2 years ago

@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 commented 2 years ago

@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

AnthonyF333 commented 2 years ago

@SankQin 在data/SetPreparation.py做数据增强

SankQin commented 2 years ago

@AnthonyF333 DataSet里面随机处理效果是不是更好呢？