Faiss: For Kmeans clustering in Python, how can I initialize centroids with computed array instead of random initialization ?

[weixuansun]

Summary

When doing kmeans in python, I am trying to set pre-computed centroids before training instead of random initialization, I wonder how can I do that?

Platform

OS: ubuntu18.04

Running on: GPU

Interface: Python

[QwertyJack]

+1

[wickedfoo]

I believe setting centroids with your desired starting centroids on the Clustering object before performing k-means clustering will start with that centroid set.

https://github.com/facebookresearch/faiss/blob/master/Clustering.h#L73

[QwertyJack]

How can I do so for IndexIVF.train()?

[QwertyJack]

How can I do so for IndexIVF.train()?

Might be difficult for now; add a Clustering object to IndexIVF::Level1Quantizer would help.

[wickedfoo]

All IndexIVF classes have a cp member which is the ClusteringParameters used for the IVF clustering.

https://github.com/facebookresearch/faiss/blob/master/IndexIVF.h#L45

[mdouze]

Right. This should be documented better.

[mdouze]

See here:
https://github.com/facebookresearch/faiss/wiki/Python-C---code-snippets#how-can-i-force-the-k-means-initialization

[sthalles]

What version of faiss is the 'init_centroids' parameter available? I am using faiss-gpu 1.6.3 and when I run the notebook, I get.

got an unexpected keyword argument 'init_centroids'

[mdouze]

Ah right, it may be available only in the latest version of Faiss.
As a quick workaround, you can just patch your faiss.py with the code here:
https://github.com/facebookresearch/faiss/blob/750d43f/faiss/python/__init__.py#L693

[sthalles]

Thanks.