need IndexFlatIP support float16

[yangsp5]

Summary

need IndexFlatIP support float16

when the number of vector is very very large, such as 1e10. Use IndexFlatIP of float32 is too expensive, maybe float16 is much fastter

[mdouze]

You can use IndexFlatScalarQuantizer with QT_fp16.

[yangsp5]

How to use IndexFlatScalarQuantizer with QT_fp16,

index = faiss.IndexFlatScalarQuantizer(emb_size, faiss.ScalarQuantizer.QT_fp16) got wrong

[dxsanny99]

facebook.com/profile.php?id=100054340704401

[dxsanny99]

facebook.com/profile.php?id=100054340704401

[yunjiangster]

@yangsp5 I think @mdouze meant IndexScalarQuantizer, but that still takes in float32 index vectors. I found the function knn_gpu that computes k nearest neighbors without building an index first, and it seems to support float16 input directly.

[ucas010]

IndexScalarQuantizer support float16?
and P or R @1 is what ?
I also have 1e10 datasets

[dxsanny99]

acebook.com/profile.php?id=100054340704401

[mdouze]

The conversion from fp32 to fp16 is done internally.
Note that most CPUs don't handle fp16 computations natively anyways so all computations are performed in float32.
knn_gpu supports float16 because there is hardware support for it.