Why batch training？

[jiaojiaoguan]

Hello everyone：

I have a error when I run e = Wh1 + Wh2.T

because the dimension of e is N by N, in my graph, n is about 400000. so it takes 400000 x 40000 x 4/1024/1024/1024 GB (about 600G). I want to know whether batch size is useful.

Thanks in advance!

[LQchen1]

Hello, I think it's useless, because GAT actually needs to compute the Attention of the entire graph

[Preston2jager]

From my understanding batch will work, for a shallow GAT example 2 layers, a single node only gathers information from nodes 2 steps away, nodes beyond that distance have no affect.