When using scipy == 1.3.3 , the error occurs.

when i use lower scipy into 1.2.1 version, it runs well

The Optimized PQ class takes a verbose flag, and this is passed to the inner PQ class.
However, it would make sense if OPQ also obeyed this flag in its fit method, which outputs lots of things about "Reconstruction error".

Hi!

Thanks for writing this package, it looks great!

I'd be interested in turning the print statements (with verbose=True) into logging statements. The verbose flag could then be used to control whether this logging is output to stdout (i.e., by setting the log level). Is this something you are interested in? if so, I could submit a PR.

How to compute codes similar to FAISS using NanoPQ without using FAISS

How can we determine the ideal value for the Ks parameter?

Not sure if this should be a feature request.

Supposed I just want to approximate distance between two PQ codes (under the same encoder of course). What is the most efficient way to perform such operation?

https://readthedocs.org/projects/nanopq/builds/

I wanted to learn how product quantization works, and this repository provided excellent code to understand how it works. As I had been learning Rust for a few months now, I decided to re-write the pq.py script in Rust to understand each step thoroughly by self-implementation. Here's the repository containing the Rust code: shubham0204/pq.rs.

The following steps are have to be taken in order to complete the project:

1. Complete README.md and add a small usage sample of the Rust API
2. Prepare a crate and upload it to crates.io

Do let me know if the repository can be included as a community resource. Just like me, many other learners would like to learn implementation of product quantization in languages other than Python, and building a section where implementations in other languages would be of great help. Moreover, I'm also working on a detailed blog which will explain product-quantization from first-concepts and with a Rust implementation.

I am looking in to do centroid of centroids using NanoPQ, is it possible?. I have a first level nanopq model M=4, K=16, D=24. The codewords that is produced is (4, 16, 6), can this output be sent as an input for the second level nanoPQ to calculate centroid of centroids? The reason for investigating centroid of centroids is due to processing large datasets and reduce processing time.

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.

thanks for your work.

`import nanopq
import numpy as np

N, Nt, D = 10000, 2000, 128
X = np.random.random((N, D)).astype(np.float32) # 10,000 128-dim vectors to be indexed
Xt = np.random.random((Nt, D)).astype(np.float32) # 2,000 128-dim vectors for training
query = np.random.random((D,)).astype(np.float32) # a 128-dim query vector

pq = nanopq.PQ(M=8, Ks=256)
pq.fit(Xt, seed=123)
X_code = pq.encode(X) # (10000, 8) with dtype=np.uint8
X_reconstructed = pq.decode(codes=X_code)

tmp = X[0]
tmp1 = X_reconstructed[0]
dis = np.sqrt(np.sum(np.square(tmp - tmp1)))`

the dis is about 2.0+ . dose it look like right?

hi,friend ,I have two question .

1. why with parametric init has poor performance than non-parametric one according to your unit test 'test_parametric_init'? it is inconsistent with the conclusion of the paper 《Optimized Product Quantization for Approximate Nearest Neighbor Search 》--Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun

`def test_parametric_init(self):
N, D, M, Ks = 100, 12, 4, 10
X = np.random.random((N, D)).astype(np.float32)
opq = nanopq.OPQ(M=M, Ks=Ks)
opq.fit(X, parametric_init=False, rotation_iter=1)
err_init = np.linalg.norm(opq.rotate(X) - opq.decode(opq.encode(X)))

opq = nanopq.OPQ(M=M, Ks=Ks)
opq.fit(X, parametric_init=True, rotation_iter=1)
err = np.linalg.norm(opq.rotate(X) - opq.decode(opq.encode(X)))

self.assertLess(err_init, err)`

2. the code compute normal not need rotate X, the decode will rotate code to original space at 255 line of opq.py

self.pq.decode(codes) @ self.R.T

I think this should be 8 bits not 256! otherwise the package is very helpful thanks!

Replace travis with github actions