File PolysemousTraining.h

namespace faiss

Implementation of k-means clustering with many variants.

Copyright (c) Facebook, Inc. and its affiliates.

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

IDSelector is intended to define a subset of vectors to handle (for removal or as subset to search)

PQ4 SIMD packing and accumulation functions

The basic kernel accumulates nq query vectors with bbs = nb * 2 * 16 vectors and produces an output matrix for that. It is interesting for nq * nb <= 4, otherwise register spilling becomes too large.

The implementation of these functions is spread over 3 cpp files to reduce parallel compile times. Templates are instantiated explicitly.

This file contains callbacks for kernels that compute distances.

Throughout the library, vectors are provided as float * pointers. Most algorithms can be optimized when several vectors are processed (added/searched) together in a batch. In this case, they are passed in as a matrix. When n vectors of size d are provided as float * x, component j of vector i is

x[ i * d + j ]

where 0 <= i < n and 0 <= j < d. In other words, matrices are always compact. When specifying the size of the matrix, we call it an n*d matrix, which implies a row-major storage.

I/O functions can read/write to a filename, a file handle or to an object that abstracts the medium.

The read functions return objects that should be deallocated with delete. All references within these objectes are owned by the object.

Definition of inverted lists + a few common classes that implement the interface.

Since IVF (inverted file) indexes are of so much use for large-scale use cases, we group a few functions related to them in this small library. Most functions work both on IndexIVFs and IndexIVFs embedded within an IndexPreTransform.

In this file are the implementations of extra metrics beyond L2 and inner product

Implements a few neural net layers, mainly to support QINCo

Defines a few objects that apply transformations to a set of vectors Often these are pre-processing steps.

struct SimulatedAnnealingParameters

#include <PolysemousTraining.h>

parameters used for the simulated annealing method

Subclassed by faiss::PolysemousTraining, faiss::SimulatedAnnealingOptimizer

Public Functions

inline SimulatedAnnealingParameters()

Public Members

double init_temperature = 0.7

double temperature_decay = 0.9997893011688015

int n_iter = 500000

int n_redo = 2

int seed = 123

int verbose = 0

bool only_bit_flips = false

bool init_random = false

struct PermutationObjective

#include <PolysemousTraining.h>

abstract class for the loss function

Subclassed by faiss::ReproduceDistancesObjective

Public Functions

virtual double compute_cost(const int *perm) const = 0

virtual double cost_update(const int *perm, int iw, int jw) const

inline virtual ~PermutationObjective()

Public Members

int n

struct ReproduceDistancesObjective : public faiss::PermutationObjective 

Public Functions

double dis_weight(double x) const

double get_source_dis(int i, int j) const

virtual double compute_cost(const int *perm) const override

virtual double cost_update(const int *perm, int iw, int jw) const override

ReproduceDistancesObjective(int n, const double *source_dis_in, const double *target_dis_in, double dis_weight_factor)

void set_affine_target_dis(const double *source_dis_in)

inline ~ReproduceDistancesObjective() override

Public Members

double dis_weight_factor

std::vector<double> source_dis: “real” corrected distances (size n^2)

const double *target_dis: wanted distances (size n^2)

std::vector<double> weights: weights for each distance (size n^2)

int n

Public Static Functions

static inline double sqr(double x)

static void compute_mean_stdev(const double *tab, size_t n2, double *mean_out, double *stddev_out)

struct SimulatedAnnealingOptimizer : public faiss::SimulatedAnnealingParameters 

#include <PolysemousTraining.h>

Simulated annealing optimization algorithm for permutations.

Public Functions

SimulatedAnnealingOptimizer(PermutationObjective *obj, const SimulatedAnnealingParameters &p): logs values of the cost function

double optimize(int *perm)

double run_optimization(int *best_perm)

virtual ~SimulatedAnnealingOptimizer()

Public Members

PermutationObjective *obj

int n: size of the permutation

FILE *logfile

RandomGenerator *rnd

double init_cost: remember initial cost of optimization

double init_temperature = 0.7

double temperature_decay = 0.9997893011688015

int n_iter = 500000

int n_redo = 2

int seed = 123

int verbose = 0

bool only_bit_flips = false

bool init_random = false

struct PolysemousTraining : public faiss::SimulatedAnnealingParameters 

#include <PolysemousTraining.h>

optimizes the order of indices in a ProductQuantizer

Public Types

enum Optimization_type_t

Values:

enumerator OT_None

enumerator OT_ReproduceDistances_affine: default

enumerator OT_Ranking_weighted_diff: same as _2, but use rank of y+ - rank of y-

Public Functions

PolysemousTraining()

void optimize_pq_for_hamming(ProductQuantizer &pq, size_t n, const float *x) const: reorder the centroids so that the Hamming distance becomes a good approximation of the SDC distance (called by train)

void optimize_ranking(ProductQuantizer &pq, size_t n, const float *x) const: called by optimize_pq_for_hamming

void optimize_reproduce_distances(ProductQuantizer &pq) const: called by optimize_pq_for_hamming

size_t memory_usage_per_thread(const ProductQuantizer &pq) const: make sure we don’t blow up the memory

Public Members

Optimization_type_t optimization_type

int ntrain_permutation: use 1/4 of the training points for the optimization, with max. ntrain_permutation. If ntrain_permutation == 0: train on centroids

double dis_weight_factor: decay of exp that weights distance loss

size_t max_memory: refuse to train if it would require more than that amount of RAM

std::string log_pattern

double init_temperature = 0.7

double temperature_decay = 0.9997893011688015

int n_iter = 500000

int n_redo = 2

int seed = 123

int verbose = 0

bool only_bit_flips = false

bool init_random = false