prompt fine-tuning阶段中，关于pr_semi_path的指向数据集的困惑？

从clip_score.py里面的class four_margin_loss(nn.Module):类（如下面代码），结合论文公式，可以得出 $inp=Ib$ , $smi1=I{t-1}$ , $smi=I_{t}$ , $ref=I_w$ 。

    elif num==4:
        loss_inp_semi1=self.margin_loss_L(tensor0,tensor_mid[0],labels)
        loss_semi1_semi2=self.margin_loss_S(tensor_mid[0],tensor_mid[1],labels)
        loss_semi2_ref=self.margin_loss_S(tensor_mid[1],tensor3,labels)
        return loss_inp_ref+loss_inp_semi1+loss_semi1_semi2+loss_semi2_ref

但是train.py里面，有如下代码可得，semi_path[1-pr_semipath]指的是 $I{t}$ ，semi_path[pr_semipath]指的是 $I{t-1}$。即


            #prompt fine-tuning
            if cur_iteration+1==train_thre+1:
                if total_iteration+1>config.num_clip_pretrained_iters:
                    pr_semi_path=1-pr_semi_path# 此时pr_semi_path指向下一轮的存储地址，或者说前一轮的地址 I_{t-1}
                if semi_path[0]!=' ' and  semi_path[1]!=' ':
                    print(semi_path)
                    L_margin_loss = clip_score.four_margin_loss(0.9,0.1)
# semi_path[1-pr_semi_path]指的是I_{t} ，semi_path[pr_semi_path]指的是I_{t-1} 
                    prompt_train_dataset = dataloader_prompt_margin.lowlight_loader(config.lowlight_images_path,config.normallight_images_path,semi_path[1-pr_semi_path],semi_path[pr_semi_path])#,config.overlight_images_path)

3. 结合下面代码，可知：semi1_path= $I_{t}$ ,semi2_path= $I_{t-1}$。 但是这就与 ‘1’中：“  $inp=I_b$ , $smi1=I_{t-1}$ , $smi=I_{t}$ , $ref=I_w$  ”矛盾。麻烦可以解释下吗？或者是我哪里理解错误了。

class lowlight_loader(data.Dataset): def init(self, lowlight_images_path,normallight_images_path,semi1_path=None,semi2_path=None):

ZhexinLiang / CLIP-LIT

prompt fine-tuning阶段中，关于pr_semi_path的指向数据集的困惑？ #19