Does Faiss distance function support cosine distance?

[Qiaogx]

Currently, I see faiss support L2 distance and inner product distance. My question is whether faiss distance function support cosine distance.

Thanks.
Gary

Summary

Platform

OS:

Faiss version:

Faiss compilation options:

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

[Enet4]

This seems like a duplicate of #95. The cosine distance is obtained with the inner product after normalizing all vectors to unit norm.

[Qiaogx]

Thanks Eduardo!

Does that mean if I want to use cosine similarity, I need to convert both the query vectors and the centroids vectors to unit vector beforehand?

[Enet4]

Does that mean if I want to use cosine similarity, I need to convert both the query vectors and the centroids vectors to unit vector beforehand?

That is correct.

[Qiaogx]

Thanks for your quick reply. And I will try to add cosine distance option to faiss, since sometimes the input vectors are not unit vectors and converting them beforehand on CPU cost more time.

[mdouze]

@Qiaogx please don't spend effort on that. Anything you can do will be more expensive than normalizing the vectors.

[Qiaogx]

Hi Matthijs,

When using cosine similarity, except normalizing the query and centroid vectors, are there any other factors I need to pay attention to? I normalized the vectors and set l2Distance_ to False in faiss/gpu/impl/FlatIndex.cu, but It did not work.

Thanks.

[liqima]

Hi Matthijs,

When using cosine similarity, except normalizing the query and centroid vectors, are there any other factors I need to pay attention to? I normalized the vectors and set l2Distance_ to False in faiss/gpu/impl/FlatIndex.cu, but It did not work.

Thanks.

this works.

quantizer = faiss.IndexFlatIP(d)
index = faiss.IndexIVFFlat(quantizer, d, nlist, faiss.METRIC_INNER_PRODUCT)

[Qiaogx]

Thank you@liqima. However, it seems IndexIVFPQ does not fully support inner product.

[mdouze]

Inner product con IndexIVFPQ is fully supported on CPU, otherwise it is a bug.
Another note: for normalized vectors, L2 and inner product search are equivalent so you can use L2 search.

[mdouze]

No activity. Closing.

[ZQ-XPY]

Hi,@Qiaogx,can you tell me how do you achieve the cosine distance? normalized vector or other ways?

[mdouze]

Note that there is now a FAQ entry for this question:
https://github.com/facebookresearch/faiss/wiki/MetricType-and-distances#how-can-i-index-vectors-for-cosine-similarity

[gabrer]

Note that there is now a FAQ entry for this question:
https://github.com/facebookresearch/faiss/wiki/FAQ#how-can-i-index-vectors-for-cosine-distance

This is the updated link:
https://github.com/facebookresearch/faiss/wiki/MetricType-and-distances#how-can-i-index-vectors-for-cosine-similarity

[mdouze]

Thanks! updated the link above to avoid confusion.

[thomasahle]

Note that normalizing the vectors before K-means, and the cluster centers after K-means, is not quite the same as cosine-sim K-means.
For the best performance, you should normalize the norm of the cluster centers after each step of Lloyd's algorithm.