Face_Pytorch icon indicating copy to clipboard operation
Face_Pytorch copied to clipboard

使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错

Open frankqianghe opened this issue 5 years ago • 13 comments

您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况

frankqianghe avatar Nov 05 '19 12:11 frankqianghe

so, you solved this problem?

foocker avatar Dec 06 '19 06:12 foocker

emm, for data : face occlusion, side face, ... for align : face bbox, face align ...

Ontheway361 avatar Mar 17 '20 08:03 Ontheway361

lfw的结果高并说明不了什么,能在megaface上刷高分数才行,还有就是要看你使用的什么训练集

wujiyang avatar Mar 17 '20 08:03 wujiyang

lfw的结果高并说明不了什么,能在megaface上刷高分数才行,还有就是要看你使用的什么训练集

请问adgdb-30 cfp 测试集在哪下载呀? 我用insightface代码对训练集做了对齐,然后从你提供的云盘下载了lfw cfp adgdb30测试集, 这样行吗? 我担心的是训练集对齐方式和测试集对齐方式不一致

smartparrot avatar Apr 29 '20 18:04 smartparrot

do you did face alignment?

smartparrot avatar May 01 '20 17:05 smartparrot

请问您使用预训练模型了吗?

ypw-lbj avatar Jun 12 '20 08:06 ypw-lbj

您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况

hello,when i training,the lfw acc was as follows,how i can solve it? image

wolfworld6 avatar Dec 21 '20 08:12 wolfworld6

Maybe the lr you used is too big. What machine are you using?

---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43)

您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况

hello,when i training,the lfw acc was as follows,how i can solve it?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

QuasarLight avatar Dec 21 '20 08:12 QuasarLight

Maybe the lr you used is too big. What machine are you using? ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) 您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况 hello,when i training,the lfw acc was as follows,how i can solve it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

tesla v100, however i changed the lr=0.01,0.001,it is still so

wolfworld6 avatar Dec 22 '20 03:12 wolfworld6

did you turn the option shuffle in dataloader into true? 

---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 中午11:36 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43)

Maybe the lr you used is too big. What machine are you using? … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) 您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况 hello,when i training,the lfw acc was as follows,how i can solve it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

tesla v100, however i changed the lr=0.01,0.001,it is still so

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

QuasarLight avatar Dec 22 '20 04:12 QuasarLight

did you turn the option shuffle in dataloader into true?  ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 中午11:36 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) Maybe the lr you used is too big. What machine are you using? … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) 您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况 hello,when i training,the lfw acc was as follows,how i can solve it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe. tesla v100, however i changed the lr=0.01,0.001,it is still so — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

no,the shuffle value is false, trainloader = data.DataLoader(dataset, batch_size=64, shuffle=False, num_workers=2, drop_last=False),did you use lfw.py or lfw_2.py?

wolfworld6 avatar Dec 22 '20 06:12 wolfworld6

try to turn the shuffle into true ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 下午2:18 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43)

did you turn the option shuffle in dataloader into true?  … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 中午11:36 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) Maybe the lr you used is too big. What machine are you using? … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) 您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况 hello,when i training,the lfw acc was as follows,how i can solve it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe. tesla v100, however i changed the lr=0.01,0.001,it is still so — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

no,the shuffle value is false, trainloader = data.DataLoader(dataset, batch_size=64, shuffle=False, num_workers=2, drop_last=False),did you use lfw.py or lfw_2.py?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

QuasarLight avatar Dec 22 '20 06:12 QuasarLight

try to turn the shuffle into true ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 下午2:18 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) did you turn the option shuffle in dataloader into true?  … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月22日(周二) 中午11:36 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "MaoCai"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) Maybe the lr you used is too big. What machine are you using? … ---原始邮件--- 发件人: "wolfworld6"<[email protected]> 发送时间: 2020年12月21日(周一) 下午4:03 收件人: "wujiyang/Face_Pytorch"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: Re: [wujiyang/Face_Pytorch] 使用您的代码,我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错 (#43) 您好: 使用您的代码我在lfw的得分可以达到99.6,但是实际应用场景会出现大批量的认错。为了不认错把阀值调高,又会出现认不出的情况。我初步分析是两个原因,一个是实际使用中有复杂的光照,还一个是训练数据集中大部分是欧美人脸,亚洲人脸很少。用欧美人脸训练出来的模型对亚洲人脸识别不好。请问您感觉还会有什么原因造成这种情况 hello,when i training,the lfw acc was as follows,how i can solve it? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe. tesla v100, however i changed the lr=0.01,0.001,it is still so — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. no,the shuffle value is false, trainloader = data.DataLoader(dataset, batch_size=64, shuffle=False, num_workers=2, drop_last=False),did you use lfw.py or lfw_2.py? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

sorry,my bad,i found my pairs.txt is wrong in the gdb mode.

wolfworld6 avatar Dec 22 '20 07:12 wolfworld6