Non-uniform vector quantization #374

marianotepper · 2024-12-18T23:27:10Z

This PR adds a new feature that uses quantized vectors for re-ranking. The quantization is computed on subvectors of each vector in the index, using a non-uniform quantizer.

…zer.

This reverts commit 4166400.

…n OptimizationResult that facilitates error checking.

…constant to each element.

…eria and averaging over 10 trials.

…exponentiateConstantMinusVector

…nTrials to 50

This reverts commit fa295a6.

jkni

Some minor nits/questions. Overall, looks quite good.

jvector-base/src/main/java/io/github/jbellis/jvector/graph/disk/NVQ.java

jvector-base/src/main/java/io/github/jbellis/jvector/graph/disk/OnDiskGraphIndexWriter.java

jvector-base/src/main/java/io/github/jbellis/jvector/optimization/LossFunction.java

jvector-base/src/main/java/io/github/jbellis/jvector/optimization/NESOptimizer.java

jvector-base/src/main/java/io/github/jbellis/jvector/vector/VectorUtilSupport.java

...sts/src/test/java/io/github/jbellis/jvector/optimization/TestNaturalEvolutionStrategies.java

jvector-tests/src/test/java/io/github/jbellis/jvector/quantization/TestCompressedVectors.java

jvector-twenty/src/main/java/io/github/jbellis/jvector/vector/SimdOps.java

jvector-base/src/main/java/io/github/jbellis/jvector/vector/VectorUtilSupport.java

…in TestNaturalEvolutionStrategies.

…ity.

…tion

marianotepper · 2024-12-30T23:09:43Z

The changes requested/suggested by @jkni have been put in place.

jkni

One minor nit to discuss and then I think we can get this merged. Can you also add a brief mention of NVQ to the README.md section about reranking ("second pass" ...) as well as UPGRADING.md?

jvector-base/src/main/java/io/github/jbellis/jvector/graph/disk/NVQ.java

…stull being taken into account

… NVQ

# Conflicts: # jvector-base/src/main/java/io/github/jbellis/jvector/graph/GraphIndexBuilder.java # jvector-base/src/main/java/io/github/jbellis/jvector/graph/similarity/BuildScoreProvider.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/ImmutableBQVectors.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/ImmutablePQVectors.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/MutableBQVectors.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/MutableCompressedVectors.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/MutablePQVectors.java # jvector-base/src/main/java/io/github/jbellis/jvector/quantization/PQVectors.java # jvector-examples/src/main/java/io/github/jbellis/jvector/example/Bench.java # jvector-examples/src/main/java/io/github/jbellis/jvector/example/SiftSmall.java # jvector-tests/src/test/java/io/github/jbellis/jvector/quantization/TestProductQuantization.java # jvector-twenty/src/main/java/io/github/jbellis/jvector/vector/SimdOps.java

jbellis · 2025-01-02T16:59:55Z

(Please use "squash and merge" when you commit.)

marianotepper · 2025-01-02T19:06:06Z

README.md and UPGRADING.md are updated now.

jkni

LGTM. As Jonathan mentioned, please squash! Minor CI fix to clean up.

@jkni

It solves the issue that @jkni caught when reviewing #374. It renders the parallel vector accesses during "encodeAll" thread safe.

marianotepper added 30 commits October 24, 2024 16:55

Adding initial files for Non-uniform Vector Quantization

20642a5

Adding methods to get the number of rows and columns.

4166400

Adding method to compute the square of a double number

28764d0

Implementation of the Exponential Natural Evolution Strategies optimi…

5e9ecbb

…zer.

Minor code cleanup

c83bd2d

Revert "Adding methods to get the number of rows and columns."

eabfad3

This reverts commit 4166400.

Minor code cleanup

1c0608b

Making the constructor of NESOptimizer public. Now optimize returns a…

aa6dfdd

…n OptimizationResult that facilitates error checking.

Adding tests.

b16b0d6

Remove unused import

751634a

Remove unused variable

3a80c1f

Increase default number of samples in NESOptimizer

8ccf14f

Add a few more tests for NESOptimizer

1141303

Add a few more tests for NESOptimizer

a421f1d

Add javadocs

98c29c5

Add one test for the case without box constraints

5f986b6

Put access modifier first

69c745e

Put access modifier first

a1f94bc

Improve javadocs

81445e4

Remove blank line

cab64ba

Initial test for the non-uniform quantizer

114d60c

Add an overload for the subInPlace vector function, that subtracts a …

836f289

…constant to each element.

Completed TestNonUniformQuantization.testGaussian with a passing crit…

b3811f5

…eria and averaging over 10 trials.

Add vectorized operations pow, constantMinusExponentiatedVector, and …

c47d2b3

…exponentiateConstantMinusVector

Add vectorized operations to TestNonUniformQuantization and increase …

93ef1d3

…nTrials to 50

Add missing blank space

e4e82c5

Use VectorUtil.squareL2Distance for computing the loss

4250711

Remove unused import

fa295a6

Revert "Remove unused import"

8be166b

This reverts commit fa295a6.

Remove unused import

a096142

marianotepper added 2 commits December 20, 2024 13:34

Replaced the logistic/logit apir with their NQT variants

649f59e

Replaced one occurrence of snake case

ab128f8

jkni requested changes Dec 23, 2024

View reviewed changes

marianotepper added 11 commits December 30, 2024 08:06

Remove multinormal placeholders from NESOptimizer. Minor refactoring …

4f1253a

…in TestNaturalEvolutionStrategies.

Fix wrong error message in LossFunction. Use Arrays/fill for readabil…

8fffca0

…ity.

Fix documentation for nvqShuffleQueryInPlace8bit

6baa399

Update error message, adding NVQ vectors

0c9aa92

Address requested changes and suggestions in PR

1bcea07

Remove NES optimization

5370204

Remove unused method in NVQuantization.NonuniformQuantizationLossFunc…

c05f515

…tion

Merge changes from main

f3002d4

Fix bug in NVQQUantization.encode

052ce69

Remove unused import

b58d281

Remove FUSED_ADC from bench

a9eed3b

marianotepper requested a review from jkni December 30, 2024 23:02

jkni requested changes Dec 31, 2024

View reviewed changes

jvector-base/src/main/java/io/github/jbellis/jvector/graph/disk/NVQ.java Outdated Show resolved Hide resolved

marianotepper added 3 commits January 2, 2025 08:25

Fix bug in NVQuantization.compressorSize, where the magic number was …

b18cb87

…stull being taken into account

Use feature-level scratch space for the vectors used for reranking in…

a268fc5

… NVQ

Add description of NVQ to README.md and UPGRADING.md

3115b42

Merge remote-tracking branch 'origin/main' into nuveq

45b8241

jkni approved these changes Jan 2, 2025

View reviewed changes

Fix error in GraphIndexBuilderTest introduced during the last auto merge

557ba8a

jkni approved these changes Jan 2, 2025

View reviewed changes

marianotepper merged commit e8d5c3c into main Jan 2, 2025
6 checks passed

marianotepper mentioned this pull request Jan 3, 2025

Make ravv usage thread-safe #381

Merged

marianotepper added a commit that referenced this pull request Jan 3, 2025

Make ravv usage thread-safe (#381)

5b0df95

It solves the issue that @jkni caught when reviewing #374. It renders the parallel vector accesses during "encodeAll" thread safe.

Non-uniform vector quantization #374

Non-uniform vector quantization #374

Uh oh!

Conversation

marianotepper commented Dec 18, 2024

Uh oh!

jkni left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

marianotepper commented Dec 30, 2024

Uh oh!

jkni left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

jbellis commented Jan 2, 2025

Uh oh!

marianotepper commented Jan 2, 2025

Uh oh!

jkni left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

jkni left a comment •

edited

Loading