What does this line in NPCFace mean? I didn't find it in the paper.

[marigoold]

FaceVerLoss/classifier.py

Line 326 in 82a7b9a

final_gt= torch.where(gt > 0, cos_theta_m, gt) #gt>0即角度<90°

I didn't find it in the paper, and I don't know what it means.
In my opinion, gt < 0 means that the angle between positive case and its corresponding class weight center is large enough so there is no need to add margin, am I right?

[bobo0810]

FaceVerLoss/classifier.py

Line 326 in 82a7b9a

final_gt= torch.where(gt > 0, cos_theta_m, gt) #gt>0即角度<90°

I didn't find it in the paper, and I don't know what it means.
In my opinion, gt < 0 means that the angle between positive case and its corresponding class weight center is large enough so there is no need to add margin, am I right?

I think so too. In the implementation of NPCFace, ArcFace is referred to, and the same measures are taken.

[marigoold]

Thanks!
By the way, do you have plan to implement NPCFace in DDP like what insightface does? I tried to implement one but it didn't converge.
😭

[bobo0810]

Thanks!
By the way, do you have plan to implement NPCFace in DDP like what insightface does? I tried to implement one but it didn't converge.
😭

In my experiment, NPCFace works well under DP training. Maybe you first make sure to converge under DP？

Although I haven't tried DDP training, there should be no problem in theory. The following two suggestions:

1. Since DDP runs a separate project code for each GPU, it only needs to control the tensor on the same device.
   For example, create a new tensor mi and move it to the same device as the input tensor.
   mi = torch.full([batch_size], self.m0).to(input.device)
2. A recently implemented another version of NPCFace