Build your own face model step by step, with blogs written in Chinese.
Build Your Own Face Recognition Model
Build Your Own Face Detection Model
for detection, you may find DBFace repo helpful.
NOTE that Official Pytorch ArcFace is released here
Build your own ArcFace, CenterFace/Centernet
Build your own face model step by step, with blogs written in Chinese.
Build Your Own Face Recognition Model
Build Your Own Face Detection Model
for detection, you may find DBFace repo helpful.
NOTE that Official Pytorch ArcFace is released here
老哥 在图像分类任务中 如果想得到 top1,top5之类的准确率 应该怎么修改呢
请问用fmobile模型训练网络,数据集为CASIA, 一个epoch要多长时间?
Could you make the pre-trained model available?
Thanks for the tutorial
我學習了您的代碼,
但我發現使用在視頻偵測人臉時,
Bounding box 會有蠻明顯的抖動 or 縮放。
我懷疑是 Heatmap 的問題,不知道作者有什麼看法呢?
若要改善這個問題有什麼好辦法嗎?!
博主你好,我想问问我是用的我自己的数据集训练的,模型的损失在8 左右就无法再收敛了,在你给的lfw上面做的测试,测试的acc在80%,你看这个问题是正常的吗??望回复,感谢
lm_xs = (lm_xs - bbox[0]) / original_width
lm_ys = (lm_ys - bbox[1]) / original_height
论文中应该是剪去bbox的中心坐标?这里是否有误
您好,请问您方便测试一下您的模型在验证集上的得分吗?
受益匪浅,十分感谢。
centernet loss那里好像有点问题。
热力图,中心点偏移,宽高应该不能都用统一的regloss替代把。
在https://github.com/xingyizhou/CenterNet/blob/master/src/main.py,中原作者用了
self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss()
self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else
RegLoss() if opt.reg_loss == 'sl1' else None
self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else
NormRegL1Loss() if opt.norm_wh else
RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg
首先感谢作者,写的很好!
我在看第6章的第7小节中,作者有这么一句话:“用选好的阈值进行划分,大于此阈值的是不同人的脸,小于此阈值的是同一个人的脸”。感觉是不是写反了,大于此阈值的是相同的脸。
理由:代码段“y_test = (y_score >= th)”,分值大于阈值的为1,否则为0,而对应的label:y_true的值,也是1对应的同一张脸,0是不同的脸。
人脸识别class有很多,label也是one-hot的吗?
Hi ...,
Which IDE are you using?
Best regards,
PeterPham
Really Nice Tutorial..
I have trained using your code for centerface detection and facing Accuracy Issue.
Some Modifications I have done according to https://github.com/chenjun2hao/CenterFace.pytorch
I have trained Mobilnet_10 model with pretrained weights upto 90 epochs with LR 5e-4 and facing accuracy problem as compared to original centerface model.
Also during inference if i use threshold to 0.50, 1000+ face detections are made (false detection)
Can you help?
博主你好!感谢分享文章和代码。
使用了你的代码后有一些疑问想请教一下:
基本情况:
想请教的问题:1. 要做什么调整才能让 training loss收敛呢?2. validation loss为什么不下降?
期待你的答复,感谢!
Can you give me the google drive link of cleaned_list.txt, I can't download it from baidu link. Thank you!
我训练的时候,模型的Network Setup占用大量的时间,并且执行到第一个epoch的第一个batch的时候就一直在执行,即没有执行到下一个batch也没有报错。想请教一下可能的原因。
博主你好,你24.pth那个模型在LFW灰度图上测的Acc在0.96+,请问你有在RGB图上的模型吗,效果差别大不大?
@siriusdemon 您好,感谢您的工作。我对训练过程感到困惑,
这里面只加载了webface的目录,没看到它对应的list加载。请问这个list不需要吗,还是在别的地方加载了。期待您的回复。
hi!
我尝试 test.py 代码加载您的24.pth权重
但是可能训练环境不一样(pytorch版本不同)出现报错:
RuntimeError: Error(s) in loading state_dict for DataParallel:
Missing key(s) in state_dict: "module.conv1.net.0.net.0.bias", "module.conv2.net.0.bias", ... ,... (后面一堆)
请问up有好的解决方法或者在torch1.6版本以上训练的权重文件分享吗?
我训练了MobileFaceNet但是,效果不好,完全按照您提供的方式对数据集划分,损失最终只收敛在了4左右,达不到您提供的24.pth的那个效果?请问我还有啥疏漏的地方吗
作者你好,
可以问下为什么arcface.py中__init__方法中self.weight 是随机初始化呢?
论文中提到的角度应该是最后一层全连接层的权重 也就是fmobilenet 中 self.linear层得到的权重 和 embedding_feature的角度
博主你好,感谢你提供的保姆式教程,受益匪浅!
有一个问题想请教,我看了一下fmobilenet的网络编码以及train.py中的代码,没有发现类别权重矩阵。这个实现中网络的输出直接是embedding吗?我有参考了其他的实现,发现网络的输出还会再与类别权重矩阵运算,最后才得到embedding。
谢谢!
请问下大概多少个epoch loss能够降到1呢?现在batch_size 128,训练了大概300个epoch维持在5左右降不下去
this is really a good work ,but i find the demo is not working, i use the imgs in this project
您好,我最近跑了一下您的代码,参数设置就是默认的coonfig.py中的参数,数据集用的清洗后的CASIA-WebFace,测试用的LFW中6000个样本对,但是精度达不到24.pth。
Test Model: /mnt/sunqiao/xyh/models/_0.pth
Accuracy: 0.661
Threshold: 0.704
Test Model: /mnt/sunqiao/xyh/models/_10.pth
Accuracy: 0.825
Threshold: 0.426
Test Model: /mnt/sunqiao/xyh/models/_20.pth
Accuracy: 0.835
Threshold: 0.391
Test Model: /mnt/sunqiao/xyh/models/_30.pth
Accuracy: 0.835
Threshold: 0.399
再往下训练也基本停在84%左右,达不到97%,不知道什么原因。
可以检测多个类别吗 比如单层车牌/双层车牌,同时有关键点,大佬能给点思路吗
很感谢大佬的分享!
您在Recognition文件夹里分享了identification的测试代码,所以想请教一下,如果用在人脸分类任务上,应该怎么修改呢?
首先,非常感谢你的分享,但是我非常困惑的是arcface的实现问题
我很想知道使用cosface去代替arcface,当出现越界问题的时候,是否合理?
如果使用clip操作的话效果如何
谢谢
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.