slow training on wikidata5m knowledge graph

[saedr]

Hi,

I'm using the following code snippet to compute the embeddings of the wikidata5m dataset.

import graphvite as gv

KG = gv.graph.KnowledgeGraph()
KG.load(gv.dataset.wikidata5m.train)

solver = gv.solver.KnowledgeGraphSolver(dim=128)
solver.build(KG)
solver.train()

But this is very slow.

Is this normal because wikidata5m is very large or am I missing something?

[KiddoZhu]

The code snap seems good. How slow is it? The benchmark here shows 2.1 hrs on 4 V100 GPUs, with the same negative samples and epochs as your default hyperparameters.

If you can afford some accuracy loss, you may reduce the number of negative samples and also the training epochs. I think num_negative=32 and num_epoch=500 would be acceptably good.