hey there, when i tried to run lfw_eval.py, i set my path to lfw dataset. then i got an error like this:

Traceback (most recent call last): File "lfw_eval.py", line 128, in best_thresh = find_best_threshold(thresholds, predicts[train]) IndexError: too many indices for array

is there any idea for this issue? may be sth wrong in channels which cause many arrays cant be matched well when inputing ? i got raw dataset of LFW(downloaded from websites) with ur landmark txt.

parser.add_argument('--dataset', default='../../dataset/face/casia/casia.zip', type=str) Place the original casia.zip in the above place，and the rest of the region has not changed. There are still some errors during run time:

File "/home/***/PycharmProjects/sphereface/dataset.py", line 39, in dataset_handle data = callback(name,filename,pindex,cacheobj,zfile) File "/home/***/PycharmProjects/sphereface/train.py", line 56, in dataset_load data = np.frombuffer(zfile.read(nameinzip),np.uint8) File "/home/***/anaconda3/lib/python3.5/zipfile.py", line 1232, in read with self.open(name, "r", pwd) as fp: File "/home/***/anaconda3/lib/python3.5/zipfile.py", line 1255, in open zinfo = self.getinfo(name) File "/home/***/anaconda3/lib/python3.5/zipfile.py", line 1199, in getinfo 'There is no item named %r in the archive' % name) KeyError: "There is no item named '0348347/018.jpg' in the archive"

Hi, thanks for your wonderful code! I have a little problem about your implementation. When you normalize w martrix, I notice you used ww = w.renorm(2,1,1e-5).mul(1e5). I think a natural implementation is just ww = F.normalize(w, dim=0). Do you have some special reasons for that? Thank you very much!!!

I used my own train.py to train the model which has same data processing and same parameters. But my test result is 98.38% using clcarwin's lfw_eval.py in Python2.7 and Pytorch0.4.0. I don't know why my result is lower than clcarwin's. Does anybody can solve my problem?Thanks.

Thanks for you codes.It really helpful. I use cleaned CASIA-WebFace, and have 455594 photos.After one epoch train,it stuck at 04:26:07 Te=0 Loss=8.7917 | AccT=2.0000% (12200/454469) 8.2977 16.70 888

when i use interrupt ,the program stop at '' File "train.py", line 142, in train(epoch,args) File "train.py", line 103, in train img,label = ds.get() File "/sphereface/dataset.py", line 113, in get index,imageshape,labelshape,pathlist = self.result.get()''

By the way,it always remain ''raise BadZipFile("Bad magic number for file header")zipfile.BadZipFile: Bad magic number for file header'' and ''zipfile.BadZipFile: File name in directory 'CASIA-WebFace/1813440/030.jpg' and header b'CASIA-WebFace/0110168/158.jpg' differ.'' is there something wrong? How to solve this issue?Thank you.

when I use

    loss = nn.CrossEntropyLoss()(output, target) 

replace the code in AngleLoss

    logpt = F.log_softmax(output)
    logpt = logpt.gather(1,target)
    logpt = logpt.view(-1)
    pt = Variable(logpt.data.exp())
    loss = -1 * (1-pt)**self.gamma * logpt
    loss = loss.mean()

I got loss=nan

I think theoretically logpt = F.log_softmax(output) here should be logpt = F.log_softmax(output,1) which is just the same as CrossEntropyLoss, but you implement is logpt = F.log_softmax(output,0) Would you help explain it?

Thanks for your nice repo. I'm trying to your codes. My question is the paper said about the annealing optimization strategy for A-Softmax loss with introducing lambda. here, your implementation is

self.lamb = max(self.LambdaMin,self.LambdaMax/(1+0.1*self.it ))
output = cos_theta * 1.0
output[index] -= cos_theta[index]*(1.0+0)/(1+self.lamb)
output[index] += phi_theta[index]*(1.0+0)/(1+self.lamb)

but, i think the cos term is to be scaled by a factor of lambda such that

output = cos_theta * self.lamb
output[index] -= cos_theta[index]*(self.lamb)/(1+self.lamb)
output[index] += phi_theta[index]*(1.0)/(1+self.lamb)

Please, give me your idea Thanks

Thanks for your code! but the code: loss = -1 * (1-pt)**self.gamma * logpt in the Angleloss,I really did not get it,because,the default gamma is 0,and the term: (1-pt)**self.gamma is useless,it is equal to: loss = -1 * logpt my question is , is 0 is your default gamma value?

Thanks for your code,this is really nice work,but I do not have matlab,can you upload your pocessed lfw image?really appreciate it if you can upload it!

I'm trying to use my own dataset but I have trouble generating facial landmarks like you created for Casia dataset. What did you use (opencv, dlib, or some other module) to get those facial landmarks?

Anyone solve same problem?

File "train.py", line 142, in train(epoch,args) File "train.py", line 101, in train bs=args.bs,shuffle=True,nthread=6,imagesize=128) File "/DL/pytorch/sphereface_pytorch/dataset.py", line 60, in init self.nparrimage = np.frombuffer(self.arrimage.get_obj(),np.float32).reshape(10,len(self.arrimage)/10) TypeError: 'float' object cannot be interpreted as an integer

closes #62 closes #63

Instead of loading the entire file into the memory, use the iterator object natively provided by the file object returned by the open function.

• This saves memory.
• Avoid __getitem__ call, improve performance.
• More pythonic.

In lfw_eval.py, the contents of pairs.txt is being loaded into memory, even though it is not used for anything other than iterating over the lines. It will be easier and more memory efficient to iterate over the file object itself.

with open('data/pairs.txt') as f:
    pairs_lines = f.readlines()[1:]

Which will also avoid the __getitem__ call in line 91, making the code more efficient.

for i in range(6000):
    p = pairs_lines[i].replace('\n','').split('\t')

Hi, I am testing the code under the environment of: pytorch == 1.9.0 torchvision == 0.10.0 python == 3.6.5

When I try torch.load(args.model) it returns an error:

File "/usr/local/cuda/lib/python3.6/site-packages/torch/serialization.py", line 777, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: invalid load key, '7'.

Could you help me with this? Thanks!

Thanks for your great work and kind sharing.

I am a beginner in the face area of CV. And I get confused in the evaluation process. According to the lfw_eval.py,https://github.com/clcarwin/sphereface_pytorch/blob/e2f11f0b0f394cf75b9f65ce8d353e4390b0948c/lfw_eval.py#L102-L103 it loads 2 face images, and the model's output is to predict is these 2 images from a single identity.

I did a lot of reading and searching, but cannot sure about is the task in lfw_eval.py exactly the face recognition?

The CASIA has 10575 different classes. I wonder:

1. why your model only has 10574 classes?
2. which label did you skipped during training.

Thank you very much.