Skip to content

Latest commit

 

History

History
147 lines (92 loc) · 4 KB

README.md

File metadata and controls

147 lines (92 loc) · 4 KB
Raw

Certified Cosine ()

Certified Cosine is a fast method for generating certificates for nearest neighbor methods in high dimensional spaces. Certified Cosine is not probabilistic but rather certifies that the returned result is correct.

More information can be found in our paper here

Usage (Python)

An example is contain in runtests.py.

import certified_cosine

# load 10,000 vectors of dimention 100 into memory
vectors = np.random(10000, 100)

# norm the vectors such that the inner product between two vectors is the cosine similarity
vectors /= np.linalg.norm(vectors, axis=1)[:, np.newaxis]


cc = certified_cosine.build(vectors)

# OPTIONAL: save the KNNG graph for use later
cc.save("my_file")

# OPTIONAL: load the KNNG graph /instead/ of building it new everytime
cc = certified_cosine.open(vectors, "my_file")



# RECOMMENDED: create a lookup engine to reuse internal search data structures
engine = cc.engine(vectors)



# perform a lookup for the 10-nearest neighbors
query = np.random.rand(100)     # the query vector.
query /= np.linalg.norm(query)


 # arguments:
 #   Query: the vector that we are looking for its nearest neighbor
 #   K: The number of neighbors to return
 #   limit: Budget before returning the current best guess.
 # Return values:
 #   Neighbors: A list of the integer ids for the k-top nearest neighbors
 #   Number expanded: The amount of the budget that was consumed
 #   Created Certificate: True if the certificate was created, false if the budget stops search early.

neighbors, number_expanded, _, created_certificate = engine.lookup_k_limit(query, 10, 50000)

Usage (C++)

#include <certified_cosine/lookup.hpp>
#include <certified_cosine/preprocesses.hpp>
#include <certified_cosine/storage.hpp>
#include <certified_cosine/policy.hpp>

// certified cosine depends on eigen for its matrix
#include <Eigen/Dense>


// construct a random data matrix
Eigen::MatrixXf vectors = MatrixXf::Random(10000, 100);
vectors.array().colwise() /= vectors.rowwise().norm();  // norm the vectors


// construct the KNNG during preprocessing
certified_cosine::dynamic_storage<float> storage;
certified_cosine::preprocess<float>(vectors, storage, 50 /* number neighbors */);

// OPTIONAL: to load or save the KNNG graph
storage.Save("my_file");
stroage.Load("my_file");

// OPTIONAL: compress storage to reduce memory during lookup
certified_cosine::compact_stoarge<float> compact_storage;
storage.BuildCompactStorage(compact_storage);


// construct a lookup engine to perform queries against.
// Can be reused to avoid reallocation.
// For handling multiple threads, multiple engines can be craeted (one per thread)
certified_cosine::lookup_proof<decltype(storage), Eigen::Ref<const decltype(vectors)>> engine(vectors, &storage);

// a query vector
Eigen::VectorXf query = Eigen::VectorXf::Random(100).normalized();

// policy controls how the lookup is performed, number of neighbors looking for
// and how certification interacts.  Here set top-10 nearest neighbors with a budget of 50,000
certified_cosine::LimitExpand<certified_cosine::CountingNBestPolicy<float>> policy(50000, 10);

// perform the lookup operation
engine.lookup(query, policy);

// get the top ids in a vector
std::vector<int> top_10;
while(!policy.items.is_empty()) {
  top_10.push_back(policy.items.max().id);
  policy.items.remove_max();
}

Install instructions

For python:

pip install 'git+https://github.com/matthewfl/certified-cosine.git'

Or:
git clone https://github.com/matthewfl/certified-cosine.git
cd certified-cosine
pip install .

This has been tested with g++-7 and up. This requires Eigen.

Cite

@article{certified_cosine,
 author = {Francis-Landau, Matthew and Van Durme, Benjamin},
 title = {Fast and/or Exact Nearest Neighbors},
 year = {2019},
 url = {https://arxiv.org/abs/1910.02478},
 keywords={Nearest Neighbors,Certificates,Cosine similarity}
}