def get_logprobs(self, f, G, G_normalized, Y_S, FCE = None):
if FCE is None:
FCE = self.FCE
F = FCE(f, G)
F_norm = torch.norm(F,p=2, dim =1).unsqueeze(1).expand_as(F)
F_normalized = F.div(F_norm+ 0.00001)
#scores = F.mm(G_normalized.transpose(0,1)) #The implementation of Ross et al., but not consistent with origin paper and would cause large norm feature dominate
scores = self.relu( F_normalized.mm(G_normalized.transpose(0,1)) ) *100 # The original paper use cosine simlarity, but here we scale it by 100 to strengthen highest probability after softmax
softmax = self.softmax(scores)
logprobs =(softmax.mm(Y_S)+1e-6).log()
return logprobs
def get_logprobs(self, f, G, G_normalized, Y_S, FCE=None):
if FCE is None:
FCE = self.FCE
F = FCE(f, G)
F_norm = torch.norm(F, p=2, dim=1).unsqueeze(1).expand_as(F)
F_normalized = F.div(F_norm + 0.00001)
scores = self.relu(
F_normalized.mm(G_normalized.transpose(0, 1))
) * 100 # The original paper use cosine simlarity, but here we scale it by 100 to strengthen highest probability after softmax
softmax = self.softmax(scores)
logprobs = (softmax.mm(Y_S) + 1e-6).log()
return logprobs