/feder

Visualize hnsw, faiss and other anns index

Primary LanguageJupyter Notebook

Feder

What is feder

Feder is a JavaScript tool designed to aid in the comprehension of embedding vectors. It visualizes index files from Faiss, HNSWlib, and other ANN libraries to provide insight into how these libraries function and the concept of high-dimensional vector embeddings. Currently, Feder is primarily focused on the IVF_FLAT index file type from Faiss and the HNSW index file type from HNSWlib, though additional index types will be added in the future.

Feder is written in javascript, and we also provide a python library federpy, which is based on federjs.

NOTE:

  • In IPython environment, it supports users to generate the corresponding visualization directly.
  • In other environments, it supports outputting visualizations as html files, which can be opened by the user through the browser with web service enabled.

Online demos

How feder works

Wiki

HNSW visualization screenshots

image

IVF_Flat visualization screenshots

image image image

Quick Start

Installation

Use npm or yarn.

yarn install @zilliz/feder

Material Preparation

Make sure that you have built an index and dumped the index file by Faiss or HNSWlib.

Init Feder

Specifying the dom container that you want to show the visualizations.

import { Feder } from '@zilliz/feder';

const feder = new Feder({
  filePath: 'faiss_file', // file path
  source: 'faiss', // faiss | hnswlib
  domSelector: '#container', // attach dom to render
  viewParams: {}, // optional
});

Visualize the index structure.

  • HNSW - Feder will show the top-3 levels of the hnsw-tree.
  • IVF_Flat - Feder will show all the clusters.
feder.overview();

Explore the search process.

Set search parameters (optional) and Specify the query vector.

feder
  .setSearchParams({
    k: 8, // hnsw, ivf_flat
    ef: 100, // hnsw (ef_search)
    nprobe: 8, // ivf_flat
  })
  .search(target_vector);

Examples

We prepare a simple case, which is the visualizations of the hnsw and ivf_flat with 17,000+ vectors that embedded from VOC 2012).

git clone git@github.com:zilliztech/feder.git
cd feder
yarn install
yarn dev

Then open http://localhost:12355/

It will show 4 visualizations:

  • hnsw overview
  • hnsw search view
  • ivf_flat overview
  • ivf_flat search view

Feder for Large Index

Feder consists of three components:

  • FederIndex - parse the index file. It requires a lot of memory.
  • FederLayout - layout calculations. It consumes a lot of computational resources.
  • FederView - render and interaction.

In case of excessive amount of data, we support separating the computation part and running it on a node server. We have two solutions for you:

  • oneServer
    • federServer (with FederIndex and FederLayout).
  • twoServer
    • indexServer (with FederIndex)
    • layoutServer (with FederLayout)

Referring to case/oneServer and case/twoServer.

Example with One Server

  1. launch the server
yarn test_one_server_backend
  1. launch the front web service
yarn test_one_server_front
  1. open http://localhost:8000

Example with Two Servers

  1. launch the FederIndex server
yarn test_two_server_feder_index
  1. launch the FederLayout server
yarn test_two_server_feder_layout
  1. launch the front web service
yarn test_two_server_front
  1. open http://localhost:8000

Pipeline - explore a new dataset with feder

Step 1. Dataset preparation

Put all images to test/data/images/. (example dataset VOC 2012)

You can also generate random vectors without embedding for index building and skip to step 3.

Step 2. Generate embedding vectors

Recommend to use towhee, one line of code to generating embedding vectors!

We have the encoded vectors ready for you.

Step 3. Build an index and dump it.

You can use faiss or hnswlib to build the index.

(*Detailed procedures please refer to their tutorials.)

Referring to test/data/gen_hnswlib_index_*.py or test/data/gen_faiss_index_*.py

Or we have the index file ready for you.

Step 4. Init Feder.

import { Feder } from '@zilliz/feder';
import * as d3 from 'd3';

const domSelector = '#container';
const filePath = [index_file_path];
const source = "hnswlib"; // "hnswlib" or "faiss"

const mediaCallback = (rowId) => mediaUrl;

const feder = new Feder({
  filePath,
  source,
  domSelector,
  viewParams: {
    mediaType: 'img',
    mediaCallback,
  },
});

If use the random_data, no need to specify the mediaType.

import { Feder } from '@zilliz/feder';
import * as d3 from 'd3';

const domSelector = '#container';
const filePath = [index_file_path];

const feder = new Feder({
  filePath,
  source: 'hnswlib',
  domSelector,
});

Step 5. Explore the index!

Visualize the overview

feder.overview();

or visualize the search process.

feder.search(target_vector[, targetMediaUrl]);

or randomly select an vector as the target to visualize the search process.

feder.searchRandTestVec();

More cases refer to the test/test.js

Blogs

Roadmap

We're still in the early stages, we will support more types of anns index, and more unstructured data viewer, stay tuned.

Acknowledgments