Accuracy on LFW

[yao5461]

Hi clcarwin, I run the code: python train.py, with default hyper parameters. and evaluate on lfw: python lfw_eval.py. The accuracy is: LFWACC=0.9907 std=0.0033 thd=0.3250. For your model, the accuracy is: LFWACC=0.9922 std=0.0054 thd=0.3050. Is there any advice? What else should I do? Thanks!

[clcarwin]

You can try to set phiflag to False and fine-tune more epoch on your model.

[HaoLiuHust]

@clcarwin I have tried that, after 5 epoch, loss came to nan, and acc will be 50%

[bixiang]

@yao5461 Can you give me your email? I have some questions about sphereface_pytorch. Thanks very much!

[YihangLou]

@HaoLiuHust I meet the same problem, have you solved it? Thanks, if you can share a solution for working around.

[HaoLiuHust]

@YihangLou change the learning rate to a smaller number, however, I can not get the acc reported here

[YihangLou]

thank you i solve it in the same way. the performce is still under the expectation

发自我的 iPhone

在 2018年3月2日，11:02，Liu, Hao notifications@github.com<mailto:notifications@github.com> 写道：

@YihangLouhttps://github.com/yihanglou change the learning rate to a smaller number, however, I can not get the acc reported here

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://github.com/clcarwin/sphereface_pytorch/issues/6#issuecomment-369805832, or mute the threadhttps://github.com/notifications/unsubscribe-auth/AHhbgiPtjnwM9zfW9ENg6APBrx0ilh1vks5taLZTgaJpZM4QMlGe.

[Gerkam]

I train on washed casia-web-face(~450k) and test on lfw with accuracy-61%. All params was default. What can be problem?

[vicdu]

@yao5461 Have u got a good work when set phiflag to False?

[Gerkam]

@yao5461 Did you train on original or cleaned Casisa-WebFace?

[timho102003]

@Gerkam Can u provide a valid link to download cleaned Casia-Webface washed list? I've searching it for a while and can't find a valid link. Thanks!!!!!!

[Gerkam]

@timho102003 See https://groups.google.com/forum/#!topic/cmu-openface/Xue_D4_mxDQ and https://github.com/cmusatyalab/openface/issues/119

[timho102003]

@Gerkam I've found these websites too, however the links are all dead. Do u have other valid download link? Thanks!!!!!

[Gerkam]

@timho102003 The washed list can be downloaded from http://pan.baidu.com/s/1kUdRRJT with password 3zbb . For fast download find manuals how can you do it from baidu

[timho102003]

@Gerkam when entering the password, baidu shows that the content is illegal and cannot be downloaded. I don't really understand ur last sentence. Does that mean I should find a way to download such data from baidu?? BTW Thanks for ur prompt response!

[Gerkam]

@timho102003 After enter password I see all files. So, maybe baidu block request from your country or another reason. In previous post I mean, if you want download with good speed you must see how can you do it. By default speed will be slow.

[RyanCV]

@clcarwin @yao5461 Is there an error in the lfw_eval code? https://github.com/clcarwin/sphereface_pytorch/blob/e2f11f0b0f394cf75b9f65ce8d353e4390b0948c/lfw_eval.py#L114

in the original paper it said the original and flipped image features are concatenated together, and then to calculate the cosine similarity.

As shown here: https://github.com/wy1iu/sphereface/blob/a848797f1b1da00ef756ea73b38b4926cc3563cd/test/code/evaluation.m#L112.

[zorrocai]

@RyanCV I have the same question. Do you think it should be changed to: f1,f2 = [f[0],f[1]], [f[2],f[3]]?