您好，关于VSUAModule.py中GNN类的一点问题

[Davidwdq]

def forward(self, obj_vecs, attr_vecs, rela_vecs, edges, rela_masks=None):
    # for easily indexing the subject and object of each relation in the tensors
    obj_vecs, attr_vecs, rela_vecs, edges, ori_shape = self.feat_3d_to_2d(obj_vecs, attr_vecs, rela_vecs, edges)

    # obj
    new_obj_vecs = obj_vecs

    # attr
    new_attr_vecs = self.gnn_attr(torch.cat([obj_vecs, attr_vecs], dim=-1)) + attr_vecs

    # rela
    # get node features for each triplet <subject, relation, object>
    s_idx = edges[:, 0].contiguous() # index of subject
    o_idx = edges[:, 1].contiguous() # index of object
    s_vecs = obj_vecs[s_idx]
    o_vecs = obj_vecs[o_idx]
    if self.opt.rela_gnn_type == 0:
        t_vecs = torch.cat([s_vecs, rela_vecs, o_vecs], dim=1)
    elif self.opt.rela_gnn_type == 1:
        t_vecs = torch.cat([s_vecs + o_vecs, rela_vecs], dim=1)
    else:
        raise NotImplementedError()
    new_rela_vecs = self.gnn_rela(t_vecs)+rela_vecs

    new_obj_vecs, new_attr_vecs, new_rela_vecs = self.feat_2d_to_3d(new_obj_vecs, new_attr_vecs, new_rela_vecs, rela_masks, ori_shape)

    return new_obj_vecs, new_attr_vecs, new_rela_vecs

您好，我想问一下这里new_obj_vecs,new_attr_vecs,new_rela_vecs是GNN优化后的图像的三种类型的特征吗

[ltguo19]

Yes, you are right.