Update cluster about pqkmeans HOT 4 CLOSED

The k-means training can be rerun using new data with a single iteration, as follows.

• Set the current centers as the initial centers
• Slowly update the centers, i.e., instead of replacing the old centers with new ones, update the old centers gradually in the new direction

Then all vectors can be reassigned to the new center. It may be enough to update only the vectors whose centers have been updated significantly.

from pqkmeans.

Thanks!

from pqkmeans.

Hi, another question if you don't mind.
Can I use the pqencoder to reduce dimensionality of a feature database and then use sklearn.neighbors.NearestNeighbors on the encoded vectors?

Thanks

from pqkmeans.

You need to reconstruct the original vector from a PQ-code. See X2_reconstructed = encoder.inverse_transform(X2_pqcode) in the tutorial https://github.com/DwangoMediaVillage/pqkmeans/blob/master/tutorial/1_pqkmeans.ipynb

Then you can run sklearn search on the reconstructed vectors.

from pqkmeans.