About the use of large matrix causes Memory Error

[zhangyikaii]

不好意思打搅了, 我在使用较大的图, 调用deeprobust.graph.utils.preprocess时, 出现了内存超限的错误, 在调用时将sparse参数置为True会影响后面的训练吗, 谢谢!

Excuse me, I used a larger graph and called the function deeprobust.graph.utils.preprocess, which caused https://github.com/DSE-MSU/DeepRobust/blob/master/deeprobust/graph/utils.py#L32 to throw the Memory Error.

If I set sparse to True: adj, features, labels = preprocess(adj, features, labels, sparse=True), does it affect the training behind? Thank you!

[ChandlerBang]

Hi ZhangYikaii,

Sorry for the late reply.

It depends on what model you are using now. Some models require the input graph to be dense graph. If you are using Metattack/PGD/MinMax/ProGNN, changing it to sparse=True would throw an error.

[zhangyikaii]

Hi ChandlerBang, Thank you very much for your reply.

ProGNN seems to use Metattack when generating Attack samples generate_attack.py. Actually I have tried to change the Metattack source code in many places when I use sparse tensor. For example, when adj is sparse, the type conversion of this code is repeated: https://github.com/DSE-MSU/DeepRobust/blob/master/deeprobust/graph/global_attack/mettack.py#L256

Is there any other way to deal with large matrices?

Anyway, thank you very much for the method proposed in your paper!

[ChandlerBang]

Actually ProGNN directly loads the pre-attacked graph in link without calling generate_attack.py.

The operation that requires dense input graph is torch.svd() and the reconstructed graph from SVD is a also dense matrix. Although you may use scipy.sparse.linalg.svds to do a sparse SVD, to reproduce the reported results in the paper, I would suggest you to still use dense graph and try to find a machine with larger RAM and large GPU memory (12GB).

[zhangyikaii]

Thank you very much!