Consine similarities about the context-aware classifier?

[HuadongTang]

In the paper, you mentioned, "We find that adopting the context-aware classifier to calculate the cosine similarities yields better results than the commonly used dot product...". However, did you only use the dot product in the code as the get_pred function below? I don't know where is the cosine similarity.

def get_pred(self, x, proto):
b, c, h, w = x.size()[:]
if len(proto.shape[:]) == 3:
# x: [b, c, h, w]
# proto: [b, cls, c]
cls_num = proto.size(1)
x = x / (torch.norm(x, 2, 1, True) + 1e-12)
proto = proto / (torch.norm(proto, 2, -1, True) + 1e-12) # b, n, c
x = x.contiguous().view(b, c, hw) # b, c, hw
pred = proto @ x # b, cls, hw
elif len(proto.shape[:]) == 2:
# x: [b, c, h, w]
# proto: [cls, c]
cls_num = proto.size(0)
x = x / (torch.norm(x, 2, 1, True)+ 1e-12)
proto = proto / (torch.norm(proto, 2, 1, True)+ 1e-12)
x = x.contiguous().view(b, c, hw) # b, c, hw
proto = proto.unsqueeze(0) # 1, cls, c
pred = proto @ x # b, cls, hw
pred = pred.contiguous().view(b, cls_num, h, w)
return pred * 15