loss function

[riyajatar37003]

i am trying to understand the loss function :

def calc_loss(self, y_true, y_pred): """ 矩阵计算batch内的cos loss """ y_true = y_true[::2] norms = (y_pred ** 2).sum(axis=1, keepdims=True) ** 0.5 y_pred = y_pred / norms y_pred = torch.sum(y_pred[::2] * y_pred[1::2], dim=1) * 20 y_pred = y_pred[:, None] - y_pred[None, :] y_true = y_true[:, None] < y_true[None, :] y_true = y_true.float() y_pred = y_pred - (1 - y_true) * 1e12 y_pred = y_pred.view(-1) y_pred = torch.cat((torch.tensor([0]).float().to(self.device), y_pred), dim=0) return torch.logsumexp(y_pred, dim=0)

1. why we are taking alternate values from true labels?
2. why we are taking dot product between alternate ypred?

if possible can you share any link or documentation of paper for this. thanks

[shibing624]

https://spaces.ac.cn/archives/8847

[riyajatar37003]

thanks , but is there any english version?

[riyajatar37003]

https://kexue.fm/archives/8847

in this article , what is mean by positive sample pairs and negative sample pairs? does it mean as follow: positve sample pair: (sent1, sent2,1) negative sample pair: {sent1, sent3,0) something like that

[shibing624]

yes