question about edge feature encoding

[YJiangcm]

I notice that in graphormer/modules/graphormer_layers.py you modify the spatialpos by setting 1 to 1, x > 1 to x - 1, could you please tell me the reason?

    if self.edge_type == "multi_hop":
        spatial_pos_ = spatial_pos.clone()
        spatial_pos_[spatial_pos_ == 0] = 1  # set pad to 1
        # set 1 to 1, x > 1 to x - 1
        spatial_pos_ = torch.where(spatial_pos_ > 1, spatial_pos_ - 1, spatial_pos_)

[mavisguan]

I notice that in graphormer/modules/graphormer_layers.py you modify the spatialpos by setting 1 to 1, x > 1 to x - 1, could you please tell me the reason?

    if self.edge_type == "multi_hop":
        spatial_pos_ = spatial_pos.clone()
        spatial_pos_[spatial_pos_ == 0] = 1  # set pad to 1
        # set 1 to 1, x > 1 to x - 1
        spatial_pos_ = torch.where(spatial_pos_ > 1, spatial_pos_ - 1, spatial_pos_)

Hi, sorry for the late response. (Other team members have been just too busy to answer!)

• About why we set x > 1 to x - 1: in function pad_spatial_pos_unsqueeze() in collator.py because we increment each element in spatial_pos by 1, and add zero padding. So later we set x > 1 to x - 1 to restore spatial_pos to real shortest distance between nodes.
• About why we set 1 to 1: we calculate an average value of edge embedding along the shortest path just a bit after your code snippet (see the code below). If we set 1 to 0, there'll be division by zero, so we set 1 to 1.

  edge_input = (
  edge_input.sum(-2) / (spatial_pos_.float().unsqueeze(-1))
  ).permute(0, 3, 1, 2)

[YJiangcm]

Thanks for you answering : )