def forward(self, cosine: torch.Tensor, label, x):
# a_i
a_i = torch.norm(x, dim=1, keepdim=True).clamp(self.l_a, self.u_a)
# m(a_i)
m = self.margin(a_i).mean().item()
index = torch.where(label != -1)[0]
# make varaible margin
m_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot.scatter_(1, label[index, None], m)
# make fixed margin
m_hot2 = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot2.scatter_(1, label[index, None], self.m)
# add to variable margin in angular space
cosine.acos_()
cosine[index] += m_hot
cosine.cos_()
# add to fixed margin in cosine space
cosine[index] -= m_hot2
cosine = cosine * self.s
g = self.calc_loss_G(a_i) * self.lambda_g
return cosine, g
def forward(self, cosine: torch.Tensor, label, x):
index = torch.where(label != -1)[0] # label에서 GT 위치
# SphereFace
m1_hot = torch.ones(index.size()[0],
cosine.size()[1],
device=cosine.device)
m1_hot.scatter_(1, label[index, None], self.m1)
# ArcFace
m2_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m2_hot.scatter_(1, label[index, None], self.m2)
# CosFace
m3_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m3_hot.scatter_(1, label[index, None], self.m3)
cosine.acos_()
cosine[index] *= m1_hot
cosine[index] += m2_hot
cosine.cos_()
cosine[index] -= m3_hot
ret = cosine * self.s
return ret, 0
def forward(self, cosine: torch.Tensor, label):
index = torch.where(label != -1)[0]
m_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot.scatter_(1, label[index, None], self.m)
cosine.acos_()
cosine[index] += m_hot
cosine.cos_().mul_(self.s)
return cosine
def forward(self, cosine: torch.Tensor, label, x):
#print(cosine)
index = torch.where(label != -1)[0] # label에서 GT 위치
m_hot = torch.ones(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot.scatter_(1, label[index, None], self.m)
#print(index.size()[0], cosine.size(), cosine.size()[1], label.size(), m_hot.size())
cosine.acos_()
cosine[index] *= m_hot
cosine.cos_().mul_(self.s)
#print(cosine)
return cosine, 0
def forward(self, cosine: torch.Tensor, label, x):
a_i = torch.norm(x, dim=1, keepdim=True).clamp(self.l_a, self.u_a)
m = self.margin(a_i).mean().item()
index = torch.where(label != -1)[0]
m_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot.scatter_(1, label[index, None], m)
cosine.acos_()
cosine[index] += m_hot
cosine.cos_().mul_(self.s)
g = self.calc_loss_G(a_i) * self.lambda_g
return cosine, g.item()
def forward(self, input: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
"""
:param input: Arccos of cosine similarity between weight matrix and features output.
:param target: correct labels of each angle data
:return: CrossEntropyLoss of added margin cosine similarity (rescaled)
"""
# input.cpu()
# target.cpu()
index = torch.where(target != -1)[0]
m_hot = torch.zeros(index.size()[0], input.size()[1], device=input.device)
m_hot.scatter_(1, target[index, None], self.margin)
input.acos_()
input[index] += m_hot
input.cos_().mul_(self.scale_factor)
return super(ArcFace, self).forward(input, target)
def forward(self, cosine: torch.Tensor, label):
index = torch.where(label != -1)[0]
m_hot = torch.zeros(index.size()[0],
cosine.size()[1],
device=cosine.device)
m_hot.scatter_(1, label[index, None], self.m)
cosine.clamp_(min=-1.0 + 1e-7, max=1.0 - 1e-7)
cosine.acos_()
# cosine_np = cosine.data.cpu().numpy()
# cosine_np = np.arccos(cosine_np)
# cosine.data = torch.from_numpy(cosine_np).to(device=cosine.device)
cosine[index] += m_hot
cosine.cos_().mul_(self.s)
if torch.isnan(cosine).sum() > 0:
index = torch.where(torch.isnan(cosine))
print(3, index)
print(3, cosine[index])
assert False, torch.isnan(cosine).sum()
return cosine
