How to train GraphSAGE with SparseTensor

[nt1tov]

Is it possible to train graphSage with SparseTensor? I just changed type of reddit dataset to provide sparse Adj(SparseTensor class object for edges), but in first layer int forward(SageConv) after self.propagate I've got different size of data (out += self.lin_r(x_r). How to fix it?

[rusty1s]

Yes, this is definitely possible (usually by just swapping edge_index with adj_t. Can you show me an (minimal) example to reproduce?

[nt1tov]

I'm getting different sizes for input and output of matmul method -- may be it is the problem

[rusty1s]

Thank you. I will check.

[nt1tov]

Please, share some feedback if you'll find some fix. My point is that sampled adj_t produces wrong size of output when it's used in matmul

[nt1tov]

By the way this problem also appears in backward -- I got the same problems:
RuntimeError: Function torch::autograd::CppNode<SPMMMean> returned an invalid gradient at index 7 - got [232965, 256] but expected shape compatible with [105144, 256]

where 105144 is the size of input and 232965 is the size of matmul(torch_sparse lib) output

[rusty1s]

Fixed in #4320. Thanks for reporting!