Best loss for selecting node or edge?

[antoni4040]

I want to use GNNs to eventually select a single node or edge as the "best" for a given scenario. I'm currently doing it with BCELoss and I was wondering if anyone has tried a better method, as it can get somewhat unbalanced I think (all 0 except a single 1).

[rusty1s]

You can also apply a softmax in the node dimension via torch_geometric.utils.softmax and only train against the ground-truth node? WDYT?

[antoni4040]

It errors out this:

RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
  File "/data1/antoni4040/conda/lib/python3.8/site-packages/torch_geometric/utils/softmax.py", line 41, in softmax
    elif index is not None:
        N = maybe_num_nodes(index, num_nodes)
        src_max = scatter(src, index, dim, dim_size=N, reduce='max')
                  ~~~~~~~ <--- HERE
        src_max = src_max.index_select(dim, index)
        out = (src - src_max).exp()
  File "/data1/antoni4040/conda/lib/python3.8/site-packages/torch_scatter/scatter.py", line 160, in scatter
        return scatter_min(src, index, dim, out, dim_size)[0]
    elif reduce == 'max':
        return scatter_max(src, index, dim, out, dim_size)[0]
               ~~~~~~~~~~~ <--- HERE
    else:
        raise ValueError
  File "/data1/antoni4040/conda/lib/python3.8/site-packages/torch_scatter/scatter.py", line 72, in scatter_max
        out: Optional[torch.Tensor] = None,
        dim_size: Optional[int] = None) -> Tuple[torch.Tensor, torch.Tensor]:
    return torch.ops.torch_scatter.scatter_max(src, index, dim, out, dim_size)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
RuntimeError: The expanded size of the tensor (12) must match the existing size (340) at non-singleton dimension 1.  Target sizes: [204, 12].  Tensor sizes: [2, 340]

I call it with x3 = softmax(x3, edge_index)

[rusty1s]

You need to call it via x3 = softmax(x3, batch) where batch is the vector to distinguish between different examples.