Geoformer implementation is different from EdgeAttention described in paper

[jimmieliu]

What i see in code is separately one attention and one trianglar multiplicative update, cannot find EdgeAttention operation like the one in your paper.