Why parametric init still has poorer performance than non-parametric one?

matsui528 / nanopq

Pure python implementation of product quantization for nearest neighbor search

MIT License

323 stars 43 forks source link

Why parametric init still has poorer performance than non-parametric one? #39

Closed madmtang closed 4 months ago

madmtang commented 6 months ago

Hi! I like your code a lot! But there is one question: Why is that when I change the parameters in the test_parametric_init to N, D, M, Ks = 100, 12, 4, 20, the test will fail?

Do you have any idea? I thought the rotation matrix would work with different Ms.

matsui528 commented 4 months ago

You seem right. Changing M causes an error. Here's what I can tell you now:

The specially-computed initial values are determined by setting parametric_init=True, while the random initial values are assigned if parametric_init=False. It happens that the former could be worse than the latter in some cases.
Our implementation is a direct translation of the original MATLAB code anyway.

Closing the issue now, but please let me know if you find a way to solve this issue.