moli232777144 / mobilefacenet-mxnet Goto Github PK
View Code? Open in Web Editor NEW基于insightface训练mobilefacenet的相关步骤及ncnn转换流程
基于insightface训练mobilefacenet的相关步骤及ncnn转换流程
如果有的话能否发我一下? @moli232777144
首先必须感谢楼主分享的训练经验.我想确认这几步训练之间有什么联系?
请问您在手机上运行的,使用的输入大小是多少的,所有网络都是一致的输入吗,可以看一下benchncnn.cpp吗
在枭龙625 mobliefacenet 推理时间可以达到60-70ms,这么快吗,cmake编译参数里面加了什么啊
Thanks for your work.
Could you share your mxnet model for arcloss. I would like to convert it to caffemodel.
Thanks.
有ncnn测试代码吗?我这里测试自己的图片时输出的值的正确率很低,但是在insightFace那里验证时正确率99.5%左右
您好,参考您的demo,训练好的模型经过slim后为4.1M,在电脑cpu测试实时视频,识别时间大约为37ms/帧,比论文中手机端测试速度还慢,我想请问,这个结果正常么?
提前谢谢了
请问:softmax的fc7配置wd_mult=10.0 这样步骤该怎么设置呢?按照您的说明步骤进行,输入运行命令后会提示“ rain_softmax.py: error: unrecognized arguments: --fc7-wd-mult 10 ”
训练时候,INFO: 我的acc始终等于0.00000,但是其他都有更新。。
有的时候突然就死机了,训练的日志都没了,请问能屏幕变输出,边保存所有日志吗。试了tee重定向,不能保存INFO: 等输出
您好,请问这个light_cnn_small的怎么样的模型?没有参考可以看?
Hi,
I'm trying to train mobilefacenet-res model from this repo: https://github.com/zuoqing1988/ZQCNN. To be specific, the mobilefacenet-res4-8-16-4 model.
The first stage softmax training went pretty well, lfw 0.99617, agedb 0.95517. But the second stage didn't went well. I used your hyper parameters for training mobilefacenet. During arcface loss training, after 2 epochs, acc is around 0.15, loss value is around 12, lfw 0.9933, agedb 0.947, which is not as good as your model.
Could you please share your strategy for training this model, like hyper parameters?
Many thanks!
按着楼主思路,训练了vggface2的数据集。[lfw][266000]Accuracy-Flip: 0.99550+-0.00334
想用google的tf-android架构做到手机端,所以没有转换ncnn,而是利用MMDNN把mxnet模型转成了PB模型,但使用转换后的PB模型在https://github.com/davidsandberg/facenet中测试acc只有0.68+.有同样苦恼的小伙伴吗?
我采用按照您所说的命令执行,得到的训练测试结果也不错。但是发现经过softmax训练得到的model竟然有40M左右,这正常吗?毕竟原作论文中训练得到的模型是4M呢?
I have try to convert this model into Linux_x64(Ubuntu 17.10 LTS x86_64)
also, after transfer mxnet model into ncnn, I add the mobilefacenet into ./benchmark.cpp
run this conmmand: ./benchmark 8 4 0
I got this:
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
find_blob_index_by_name prob failed
mobilefacenet min = 0.03 max = 0.03 avg = 0.03
I know the author is developing in the windows, maybe someone face this trouble in the Linux can give me some advice.
请问大佬,我从官网下载了arcface loss的模型,然后从自己的数据用arcface loss微调,发现一开始的train acc 为0.00,但是随着训练起来,train acc逐渐上涨,请问这是正常的么?或者说这个大概是什么原因,
The mobilefacenet benchmark test result in the device: raspberry pi 3B.
loop_count = 8
num_threads = 4
powersave = 0
mobilefacenet min = 145.20 max = 148.46 avg = 145.86
以上结果没有对网络结构做任何优化,模型大小为4.1MB,测试结果相当不错
你好,我在自己的数据集上训练时出现了很多错误,我认为是我的数据转化rec文件时候出了问题
因为每次训练导入rec文件时候都会花费很多时间。
我想问的是在读取rec文件时候,label[0],label[1]和id2range分别是什么作用,因为我发现我的数据label[0]和label[1]相差非常大,导致id2range数值很大。
我认为是这里出了问题,谢谢你的解答。
I used mxnet-cu80 to train mobilefacenet model. However there are many [lfw][4000] Accuracy-Flip: 0.50000+-0.0000 during the training, and the log is different yours. Did you use mxnet-cu90 to train you model?
请问作者,caffe训练的话,同样是这个步骤吗,先只用softmax 损失进行lr=0.1的训练到12万步骤,然后添加arcface损失,在之前训练的模型上继续训练吗。。。
i have not a p40 gpu card,but i have two 1080 gpu cards; you can set batch-size 512;but i only can set 128 instead.
in your readme article ,you train 40 thousand times in your first step; so how many times i can train in my first step? in the first step ,how much can you get accuracy in lfw or agedb-30 ?
大佬,模型转换的第二步的时候,mobilefacenet.bin这个文件是哪里的啊。望大佬解答
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.