search

HazyResearch / hgcn

Hyperbolic Graph Convolutional Networks in PyTorch.
602 stars 109 forks source link

How to parallelize across multiple graphs? #19

Closed nghiahhnguyen closed 4 years ago

nghiahhnguyen commented 4 years ago

For the moment, I see that the code allows one to parallelize the computation of node embedding for all nodes in a graph. However, if I want to parallelize the computation across multiple graphs (perhaps by passing multiple graphs per batch), there is an error during the computation for hyperbolic layer because torch.spmm does not support 3D Tensor.

File "~/hgcn/layers/hyp_layers.py", line 133, in forward support_t = torch.spmm(adj, x_tangent) RuntimeError: 2D tensors expected, got 3D, 3D tensors at /pytorch/aten/src/THC/generic/THCTensorMathBlas.cu:260

Is there any way for me to perform graph parallelization?

nghiahhnguyen commented 4 years ago

In case someone else needs the same functionality, this can be done by stacking multiple graphs into one giant graph, perform computation on this giant graph before decomposing them for further evaluation. You can refer to the Advanced Mini-batching session of PyTorch Geometric for further details.