big-ann-benchmarks
big-ann-benchmarks copied to clipboard
harsha-simhadri
add NGT to T1 and T2
I have not got results with this parameter setting yet, and our method might not surpass the baselines.
I have not got results with this parameter setting yet, and our method might not surpass the baselines.
Thanks for the PR, Masajiro. Did you test on the Azure VM? If you can please add test entries for random-xs dataset for both your t1 and t2 algorithms to https://github.com/harsha-simhadri/big-ann-benchmarks/blob/main/.github/workflows/benchmarks.yml
@harsha-simhadri Yes, I tested our method in some degree, however, I have not spotted the parameters to improve its performance yet, or our approach might be basically unpromising. Since this competition stimulated my interest in one billion target, I am going to try to improve our method regardless of our method's results. Thank you for providing this opportunity. BTW, adding test entries for random-xs dataset are necessary?
@harsha-simhadri Yes, I tested our method in some degree, however, I have not spotted the parameters to improve its performance yet, or our approach might be basically unpromising. Since this competition stimulated my interest in one billion target, I am going to try to improve our method regardless of our method's results. Thank you for providing this opportunity. BTW, adding test entries for random-xs dataset are necessary?
Masajiro, our goal is to keep this site open for longer term. You are most welcome to keep improving and we can re-evaluate later. The NeurIPS competition is just the first opportunity to start something. I will run the entries and report whatever turns up. Please do add a random-xs test so that we at least know your code runs as necessary and any later changes will not break your set up.
I have added random-xs tests.
Thanks. Could you please add URL to pre-built indices to your PR?
@harsha-simhadri We don't have indices with the same parameters as our PR, and unfortunately spent almost all of our budget to use azure VMs. We are currently focusing on improving our search algorithm, and are going to release it in a few days.
I added prebuilt indices that were built on our in house servers, and rebuilding the docker NGT image is also necessary. However, these indices might be updated because we are going to tune some parameters continuously.
@harsha-simhadri Although our results do not definitely surpass the results of the winner, did you try to run this PR?