search

lrjconan / GRAN

Efficient Graph Generation with Graph Recurrent Attention Networks, Deep Generative Model of Graphs, Graph Neural Networks, NeurIPS 2019
MIT License
469 stars 96 forks source link

RuntimeError: invalid argument 2: Index dimension is out of bounds #1

Closed hmt2014 closed 5 years ago

hmt2014 commented 5 years ago

Error when testing at code att_edge_feat = att_edge_feat.scatter(1, att_idx[[edges[:, 0]]], 1)

lrjconan commented 5 years ago

Hi

Did you test it on cpu? I noticed this while I was testing using CPU. In my case, testing with GPU is fine. If I remember correctly about my case, it happens in the beginning of generation where att_idx[[edges[:, 0]]] is empty. I believe it is caused by the inconsistency between the CPU and GPU implementations of scatter in PyTorch.

hmt2014 commented 5 years ago

Thanks :-D

DBraun commented 2 years ago

Follow up question. If I do self.att_edge_dim = 16 instead of 64, then on the following line relative to the one mentioned https://github.com/lrjconan/GRAN/blob/43cb4433e6f69401c3a4a6e946ea75da6ec35d72/model/gran_mixture_bernoulli.py#L246-L247 I get this error C:\w\b\windows\pytorch\aten\src\ATen\native\cuda\ScatterGatherKernel.cu:276: block: [13,0,0], thread: [32,0,0] Assertion `idx_dim >= 0 && idx_dim < index_size && "index out of bounds"` failed.

But I should warn you that this is with some other changes to the project. Still I'm curious if there's a reason att_edge_dim is hard coded to 64? 32 and 64 work somehow. And also, does self.has_rand_feat work? Because I'm having trouble with that too.