Not able to cross 70% validation accuracy for Graph Neural Network

[adeepak7]

Hello folks!

I tried training the rgat (25M) model on tiny, small, and medium datasets on >2 epochs, and the highest validation accuracy (computed for every 5% of validation data) that I could achieve was 70.28%.

Is the required target of 72% defined for the full dataset only, or can we achive the same with the above 3 datasets as well?

If yes, can anyone share what parameters they used to achieve the target 72% for each dataset?

Thanks

[ShriyaRishab]

Hi,

The expectation is to use the full dataset, specifically IGBH-full to reach the target accuracy. MLPerf training requires submitters to run using the exact dataset used by the reference to submit to the closed category.