好像文章和源码使用的真值轨迹有区别？有大佬能帮忙解惑吗

[caokaiyan-afk]

In the training phase, the ground truth trajectory is added
to the planning vocabulary as the positive sample. Other
trajectories are regarded as negative samples. We assign
different loss weights to negative trajectories. Trajectories
close to the ground truth trajectory are less penalized.
原文中好像提到我们将gt轨迹加入词表作为正样本

def get_plan_expert_target(self,
ego_traj_preds,
ego_fut_gt,
ego_fut_masks,
ego_cls_expert_preds, # (N, 1)
plan_col_labels,
plan_bd_labels,
):

    plan_expert_labels = torch.zeros((self.plan_fut_mode), dtype=torch.long,
                                      device=ego_traj_preds.device)

    plan_expert_labels_weight = torch.zeros((self.plan_fut_mode), dtype=ego_traj_preds.dtype,
                                      device=ego_traj_preds.device)
    
    if ego_fut_masks[0] == 1.:

        neg_idx = torch.ones((self.plan_fut_mode), dtype=torch.bool,
                                      device=ego_traj_preds.device)
        #### v1
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1) \
        #     + torch.linalg.norm(ego_traj_preds[:,0,:] - ego_fut_gt.cumsum(dim=-2)[:,0,:], dim=-1) * 5.
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=2.) / 2.
        #### v2
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=1.5) / 1.5
        #### v3
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=10) / 1.5
        # #### v4
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=15) / 1.5
        # #### v5
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=30) / 1.5
        #### v6
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=30) * 2.
        #### v7
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=100.) * 2.
        #### v8
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=100.) * 4.
        #### v9
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=100.) * 10.
        # #### v10
        # traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=100.) * 20.
        #### v11
        traj_dis = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).sum(dim=-1)
        plan_expert_labels[neg_idx] = 1
        plan_expert_labels_weight[neg_idx] = torch.clip(traj_dis, min=0, max=100.) * 100.

        plan_expert_labels[plan_col_labels == 1] = 1
        plan_expert_labels[plan_bd_labels == 1] = 1
        plan_expert_labels_weight[plan_col_labels == 1] = 100.
        plan_expert_labels_weight[plan_bd_labels == 1] = 100.

        # pos_idx = torch.linalg.norm(ego_traj_preds[:,:1,:] - ego_fut_gt.cumsum(dim=-2)[:,:1,:], dim=-1).mean(dim=-1).argmin()
        pos_idx = traj_dis.argmin()
        plan_expert_labels[pos_idx] = 0

        # add weights to balance trajs
        self.traj_selected_cnt[pos_idx] += 1.
        scaling_rate = self.traj_selected_cnt.sum() / self.traj_selected_cnt[pos_idx] / self.plan_fut_mode 
        scaling_rate = torch.clamp(scaling_rate, 0.5, 2.)
        plan_expert_labels_weight[pos_idx] = 100.  # * scaling_rate   

        # global pos_idx_cnt
        # pos_idx_cnt[pos_idx] += 1
        #-------
        # pos_idx = torch.linalg.norm(ego_traj_preds[:,:,:] - ego_fut_gt.cumsum(dim=-2)[:,:,:], dim=-1).mean(dim=-1).argmin()
        # rank = (ego_cls_expert_preds[pos_idx] < ego_cls_expert_preds).sum()
        # plan_expert_labels[pos_idx] = 0
        # plan_expert_labels_weight[pos_idx] = 500. * min(rank, 10)


        # neg_idx = torch.linalg.norm(ego_traj_preds[:,:1,:] - ego_fut_gt.cumsum(dim=-2)[:,:1,:], dim=-1).mean(dim=-1) >  -10e6 # all
        # plan_expert_labels[neg_idx] = 1
        # plan_expert_labels_weight[neg_idx] = min(rank, 10) / self.plan_fut_mode


        # plan_expert_labels[plan_col_labels == 1] = 1
        # plan_expert_labels[plan_bd_labels == 1] = 1
        # plan_expert_labels_weight[plan_col_labels == 1] = 1.
        # plan_expert_labels_weight[plan_bd_labels == 1] = 1.

    return plan_expert_labels, plan_expert_labels_weight

但源码中这一段看起来是将词表中距离模型最近的作为正样本，有大佬帮忙解惑吗