pcl_loss | Prototypical Verbalizer for Prompt-based Few-shot Tuning | Prototypical Contrastive Learning

[Charon-HN]

在读了贵作Prototypical Verbalizer for Prompt-based Few-shot Tuning 是个很好的文章，但是我比较好奇代码是怎么实现的，然后我需要定位到程序的运行入口在哪里，我找了半天，感觉是在项目的experiments/cli.py这个[文件]中。而这个文件是需要读取配置文件的，那么原型学习的配置文件是在experiments/classification_proto_verbalizer.yaml中。

最终我在openprompt/prompts/prototypical_verbalizer.py文件中找到了这个pcl_loss的计算，结合论文来看，是由两个loss组成的，我根据自己的理解看了看代码，出现了一些困惑，下面是pcl_loss的计算

 def pcl_loss(self, v_ins):
        # instance-prototype loss

        sim_mat = torch.exp(self.sim(v_ins, self.proto))
        num = sim_mat.shape[1]
        loss = 0.
        for i in range(num):
            pos_score = torch.diag(sim_mat[:,i,:])
            neg_score = (sim_mat[:,i,:].sum(1) - pos_score)
            loss += - torch.log(pos_score / (pos_score + neg_score)).sum()
        loss = loss / (num * self.num_classes * self.num_classes)

        # instance-instance loss

        loss_ins = 0.
        for i in range(v_ins.shape[0]):
            sim_instance = torch.exp(self.sim(v_ins, v_ins[i]))
            pos_ins = sim_instance[i]
            neg_ins = (sim_instance.sum(0) - pos_ins).sum(0)
            loss_ins += - torch.log(pos_ins / (pos_ins + neg_ins)).sum()
        loss_ins = loss_ins / (num * self.num_classes * num * self.num_classes)
        loss = loss + loss_ins

        return loss

对于 instance-prototype loss的计算我有一些困惑，我在代码中加了一些说明：

 # instance-prototype loss
sim_mat = torch.exp(self.sim(v_ins, self.proto)) # 维度应该为（num_classes, batch_size, num_classes）
num = sim_mat.shape[1] # 获取batch_size
loss = 0.
for i in range(num):
    pos_score = torch.diag(sim_mat[:,i,:]) #
    neg_score = (sim_mat[:,i,:].sum(1) - pos_score)
    loss += - torch.log(pos_score / (pos_score + neg_score)).sum()
loss = loss / (num * self.num_classes * self.num_classes)

困惑

1. 对于这个loss的实现的感觉这不就变成了loss += - torch.log(pos_score / (sim_mat[:,i,:].sum(1))).sum()了吗。
2. pos_score = torch.diag(sim_mat[:,i,:])表示每个样本的在每个类别上的损失。但是从公式来看$\mathbf{v}_i^n$表示表征$\mathbf{v}_i$所属的类别为$n$，应该只需要考虑样本i所属的类别就可以了呀？
3. 论文中的exp体现在了什么地方呢？