search

clcarwin / sphereface_pytorch

A PyTorch Implementation of SphereFace.
MIT License
714 stars 172 forks source link

Acc question #8

Open bixiang opened 6 years ago

bixiang commented 6 years ago

I use your pre_trained model 'sphere20a_20171020.pth' and the result is 'LFWACC=0.6402 std=0.0146 thd=0.4775'. I'm so confused.

clcarwin commented 6 years ago

I have no idea. Maybe you can try to update your pytorch/cuda/cudnn to the newest version.

hzshuai commented 6 years ago

I got the same result as bixiang's, I use python2.7, pytorch-0.2.0_3, cudnn-6, cuda8.0.

How about your @clcarwin pytorch environment to get the result?

hzshuai commented 6 years ago

update: same to bixiang, my lfw dataset is not the original official lfw dataset

After that, I got the LFWACC=0.9915 std=0.0054 thd=0.3085, lower than 0.992 reported in your repo.

I also retrained the model from scratch, finally got the training result: Loss=2.0312 | AccT=93.5304%. I found with your default hyperparameter, there's no improvement with more epochs. BTW: the crresponding eval result is 'LFWACC=0.9913 std=0.0048 thd=0.2925'

Could you give me some advice on the training process ? @clcarwin

How's your work recently? @bixiang

bixiang commented 6 years ago

@hzshuai I have got the right result. My reason of low LFWACC is my test dataset have been done alignment. As a result, if I use the clcarwin's test code, I will do alignment for testdata twice!

bixiang commented 6 years ago

@hzshuai About this model, I am doing some other works and want to discuss with others. I'm from iscas. My wechat is 631885006.

yujun1993 commented 6 years ago

I use your pre_trained model 'sphere20a_20171020.pth' and and the result is 'LFWACC=0.5002 std=0.0005 thd=0.7655.' I use python2.7, pytorch0.3.0.post4, cudnn-6, cuda8.0. And download the lfw dataset from http://vis-www.cs.umass.edu/lfw/lfw.tgz. Anyone can help me ?

yujun1993 commented 6 years ago

@clcarwin After my test, seems like pytorch0.3.0 are not yet support , I use pytorch0.2.0 get the proper result

yulongwang12 commented 6 years ago

@hzshuai I got the exactly same test result with you (LFWACC=0.9915 std=0.0054 thd=0.3085). So do you figure out what's the reason for that lower than the reported accuracy? Looking forward to your reply, many thanks.