search

chenchens190009 / CSProm-KG

32 stars 0 forks source link

GPU memory problem #4

Open meaningful96 opened 6 months ago

meaningful96 commented 6 months ago

Hello.

I'm facing an out-of-memory (OOM) problem while training the wikidata5m-Trans model, even with an A6000 GPU. The batch size I'm using is 450, which is the same as the commands you provided on the main page.

I'd like to know if the Wikidata5M results in the paper were also obtained with a batch size of 450.

chenchens190009 commented 6 months ago

Hi @meaningful96 ,

Yes, we use the batch size of 450 for the wikidata5m dataset. We observe that using larger batch size enhances performance with this dataset. However, our device can only handle up to 450 as the maximum batch size.