README
¶
gann
gann (go-approximate-nearest-neighbor) is a library for approximate nearest neighbor search purely written in golang.
The implemented algorithm is truly inspired by Annoy (https://github.com/spotify/annoy).
feature
- purely written in Go: no dependencies out of Go world.
- easy to tune with a bit of parameters
installation
go get github.com/mathetake/gann
parameters
setup phase parameters
| name | type | description | run-time complexity | space complexity | accuracy |
|---|---|---|---|---|---|
| dim | int | dimension of target vectors | the larger, the more expensive | the larger, the more expensive | N/A |
| nTree | int | # of trees | the larger, the more expensive | the larger, the more expensive | the larger, the more accurate |
| k | int | maximum # of items in a single leaf | the larger, the less expensive | N/A | the larger, the less accurate |
runtime (search phase) parameters
| name | type | description | time complexity | accuracy |
|---|---|---|---|---|
| searchNum | int | # of requested neighbors | the larger, the more expensive | N/A |
| bucketScale | float64 | affects the size of bucket |
the larger, the more expensive | the larger, the more accurate |
bucketScale affects the size of bucket which consists of items for exact distance calculation.
The actual size of the bucket is calculated by int(searchNum * bucketScale).
In the search phase, we traverse index trees and continuously put items on reached leaves to the bucket until the bucket becomes full. Then we calculate the exact distances between a item in the bucket and the query vector to get approximate nearest neighbors.
Therefore, the larger bucketScale, the more computational complexity while the more accurate result to be produced.
example
package main
import (
"fmt"
"math/rand"
"time"
"github.com/mathetake/gann"
"github.com/mathetake/gann/metric"
)
var (
dim = 3
nTrees = 2
k = 10
nItem = 1000
)
func main() {
rawItems := make([][]float64, 0, nItem)
rand.Seed(time.Now().UnixNano())
for i := 0; i < nItem; i++ {
item := make([]float64, 0, dim)
for j := 0; j < dim; j++ {
item = append(item, rand.Float64())
}
rawItems = append(rawItems, item)
}
m, err := metric.NewCosineMetric(dim)
if err != nil {
// err handling
return
}
// create index
idx, err := gann.CreateNewIndex(rawItems, dim, nTrees, k, m)
if err != nil {
// error handling
return
}
// search
var searchNum = 5
var bucketScale float64 = 10
q := []float64{0.1, 0.02, 0.001}
res, err := idx.GetANNbyVector(q, searchNum, bucketScale)
if err != nil {
// error handling
return
}
fmt.Printf("res: %v\n", res)
}
references
- https://github.com/spotify/annoy
- https://en.wikipedia.org/wiki/Nearest_neighbor_search#Approximate_nearest_neighbor
License
MIT
Documentation
¶
Overview ¶
Package gann can be used for approximate nearest neighbor search.
By calling gann.CreateNewIndex function, we can obtain a search index. Its interface is defined in gann.Index:
type Index interface {
GetANNbyItemID(id int64, searchNum int, bucketScale float64) (ann []int64, err error)
GetANNbyVector(v []float64, searchNum int, bucketScale float64) (ann []int64, err error)
}
GetANNbyItemID allows us to pass id of specific item for search execution and instead GetANNbyVector allows us to pass a vector.
See README.md for more details.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Index ¶
type Index interface {
// GetANNbyItemID ... search approximate nearest neighbors by a given itemID
GetANNbyItemID(id int64, searchNum int, bucketScale float64) (ann []int64, err error)
// GetANNbyVector ... search approximate nearest neighbors by a given query vector
GetANNbyVector(v []float64, searchNum int, bucketScale float64) (ann []int64, err error)
}
Index is the interface of gann's search index. GetANNbyItemID and GetANNbyVector are different in the form of query. GetANNbyItemID can be executed by passing a certain item's id contained in the list of items used in the index building phase. GetANNbyVector allows us to pass any vector of proper dimension.
searchNum is the number of requested approximated nearest neighbors, and bucketScale can be tuned to make balance between the search result's accuracy and computational complexity in the search phase.
see README.md for more details.
func CreateNewIndex ¶
CreateNewIndex build a new search index for given vectors. rawItems should consist of search target vectors and its slice index corresponds to the first argument id of GetANNbyItemID. For example, if we want to search approximate nearest neighbors of rawItems[3], it can simply achieved by calling index.GetANNbyItemID(3, ...).
dim is the dimension of target spaces. nTree and k are tunable parameters which affects performances of the index (see README.md for details.)
The last argument m is type of metric.Metric and represents the metric of the target search space. See https://godoc.org/github.com/mathetake/gann/metric for details.
Source Files
Directories