I'm trying to use edge attributes in a GATConv network. I'm facing issue when adding self_loop.

[arake001]

I am facing issues with the GATConv when using edge_attributes.

I tried using single-dimension edge_attribute and multiple-dimension edge attributes.

Example graph structure

Data(edge_attr=[13848], edge_index=[2, 13848], x=[2295, 100], y=[1])

Data(edge_attr=[13848,1], edge_index=[2, 13848], x=[2295, 100], y=[1])

Data(edge_attr=[13848, 2], edge_index=[2, 13848], x=[2295, 100], y=[1])

The following is the error.

File "/home/ankith/graph_files/cvpr/casme2/pytorch_geometric/torch_geometric/nn/conv/edge_gat.py", line 289, in forward
edge_index, edge_attr = add_self_loops(
File "/home/ankith/anaconda3/envs/gconv-cuda/lib/python3.8/site-packages/torch_geometric/utils/loop.py", line 89, in add_self_loops
loop_weight = edge_weight.new_full((N, ), fill_value)
TypeError: new_full(): argument 'fill_value' (position 2) must be Number, not str

Any clue how to solve this issue?

Thank you in advance!

[rusty1s]

Can you show me how you call add_self_loops in edge_gat.py?

[arake001]

I haven't done any changes to the add_self_loop section of the GATConv layer.

The other thing I noticed is that the code works on the CPU version of PyTorch 1.10 and when I try the same code on the GPU version of PyTorch 1.6, it fails.

Is the issue with the version of the Pytorch or when using GPU trigger an error?

[rusty1s]

Mh, PyTorch 1.6 is now over two years old, so we don't give any guarantees that latest PyG works with it. I would recommend updating your PyTorch version.