Overview

ArcFace

Usage

You need to alert your model.

• add label argument to model's forward function

• add label as classifier's input

• I recommend you to add if else into foward function, so that when you want to train without Arcface, you don't need to change code

  class Your_Model(nn.Module):
       def __init__(self):
           ..............
           self.classifier = ....
  
       def forward(self, x, label=None):
           ...................
           output = self.classifier(x) if label is None else self.classifier(x, label)
           return output
  

• Training phase

1. Replace last fully-connected layer of your model with ArcFace
   Take ResNet18 for example:

from arcface import ArcFace
from torchvision import models

model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
model_ft.fc = ArcFace(num_ftrs, num_classes, m=0.5)

2. In the training loop, add label as model's input argument.
   Generally, loss_func is nn.CrossEntropyLoss()

for img, label in torch_dataloader:
    if use_arcface:
        output = model(img, label)
    else:
        output = model(img)
    loss = loss_func(output, label)
    ......

• Evaluation/Inference phase
  Don't need to alert anything

Note

• If your training is hard to converge, you can set m to smaller(close to zero).
• If your dataset is difficult, i recommend you to set m to smaller.
• Acface's argument s is 64 in the original paper, but when s become bigger the gradient is enlarged. So I set s to 8 by default.