Question about angleproto.py code

[huangrunqian]

Why do you use fake labels instead of the input real labels here? ---> from angleproto.py

def forward(self, x, label=None):

    assert x.size()[1] >= 2

    out_anchor      = torch.mean(x[:,1:,:],1)
    out_positive    = x[:,0,:]
    stepsize        = out_anchor.size()[0]

    cos_sim_matrix  = F.cosine_similarity(out_positive.unsqueeze(-1),out_anchor.unsqueeze(-1).transpose(0,2))
    torch.clamp(self.w, 1e-6)
    cos_sim_matrix = cos_sim_matrix * self.w + self.b

    label   = torch.from_numpy(numpy.asarray(range(0,stepsize))).cuda()
    nloss   = self.criterion(cos_sim_matrix, label)
    prec1   = accuracy(cos_sim_matrix.detach(), label.detach(), topk=(1,))[0]

    return nloss, prec1

[huangrunqian]

Made it out finally~

[speaker-lover]

Excuse me, what is the result of using this loss function? I got eer=2.6 minDCF=0.2. Is this normal? Does it belong to self-supervised learning method or supervised learning method?

[huangrunqian]

I just use the loss function, while the model and data are different, so the result is not comparable. Maybe you could check and compare your result with that of the original paper. As you use the speaker labels, it's a supervised learning method.