TF-KMeans
Description
A Simple JavaScript Library to make it easy for people to use KMeans algorithms with Tensorflow JS.
The library was born out of another project in which except KMeans, our code completely depended on TF.JS
As such, moving to TF.JS helped standardise our code base substantially and reduce dependency on other libraries
Sample Code
When you are using a browser at frontend!
Example takes cosineDistance.
const KMeans = require('./tf-kmeans').default
const tf = require('@tensorflow/tfjs')
function testCosineCluster() {
tf.tidy(() => {
const kmeans = new KMeans({
k: 3,
maxIter: 50,
distanceFunction: KMeans.cosineDistance,
})
const data1 = [
[1, 23, 3],
[1, 23, 3],
[4, 5, 2.1],
[2, 3, 1],
[4, 5, 2],
[4, 5, 2],
[4, 5, 2],
[4, 5, 2],
[4, 5, 2],
[4, 5, 2.1],
[4, 5, 2.1],
[4, 5, 2.1],
[4, 5, 2.1],
]
const data2 = [
[-0.026, 0.0533, 0.1],
[0.1, 0.033, 0.032],
[0.12, -0.2, 0.123],
[0.333333, 0.21, 0.21],
[-0.76, -0.321, 0.228],
[-0.26, -0.321, 0.22],
[0.1, 0.3, 0.28],
[0.1, 0.06, 0.22],
[-0.00000001, 0.01, 0.0211],
[0.02, -0.009, -0.0211],
[0.12, 0.01, 0.0211],
[0.02, 0.01, -0.111],
[-0.02333, -0.043, -0.12001],
]
const dataset = tf.tensor(data2)
const train = kmeans.train(dataset)
console.log('Train Classify', train.arraySync())
console.log('Centers', kmeans.centroids.arraySync())
console.log('Memory Used', tf.memory())
console.log('Predict:')
const ys = kmeans.predict(
tf.tensor([
[0.1, 0.22, 0.21],
[-0.02, -0.01, 0.02001],
]),
)
console.log('--------category index--------')
console.log(ys.index.arraySync())
console.log('--------category center-------')
console.log(ys.center.arraySync())
console.log('--------category ditance-------')
console.log(ys.distance.arraySync())
// dispose
kmeans.dispose()
train.dispose()
dataset.dispose()
})
}
testCosineCluster()When you are using nodejs at backend!
Example takes euclideanDistance.
const KMeans = require('./tf-kmeans-node').default
const tf = require('@tensorflow/tfjs-node')
const PATH = './kmeans.json'
function test() {
tf.tidy(() => {
const kmeans = new KMeans({
k: 3,
maxIter: 50,
})
console.log(kmeans)
const dataset = tf.tensor([
[2, 2, 2],
[5, 5, 5],
[3, 3, 3],
[4, 4, 4],
[7, 8, 7],
])
const train = kmeans.train(dataset)
console.log('Train Classify', train.arraySync())
console.log('Centers', kmeans.centroids.arraySync())
console.log('Memory Used', tf.memory())
console.log('Predict:')
const pre = kmeans.predict(tf.tensor([2, 3, 2]))
console.log('Category index:', pre.index.arraySync())
console.log('Category distance:', pre.distance.arraySync())
console.log('Category center:', pre.center.arraySync())
kmeans.save(PATH)
// dispose
kmeans.dispose()
train.dispose()
dataset.dispose()
})
}
function testLoad() {
console.log('====================Test load model=======================')
const model = require(PATH)
const kmeans = new KMeans(model)
console.log('Predict:')
const pre = kmeans.predict(tf.tensor([2, 3, 2]))
console.log('Category index:', pre.index.arraySync())
console.log('Category distance:', pre.distance.arraySync())
console.log('Category center:', pre.center.arraySync())
}
// train
test()
// load
testLoad()Functions
-
ConstructorTakes 4 Optional parameters
- k:- Number of Clusters
- maxIter:- Max Iterations
- distanceFunction:- The Distance function Used Currently:
euclideanDistanceandcosineDistance - centroids?:- Always when loading from a save json model, you don't need to train again.
-
trainTakes Dataset as Parameter
Performs Training on This Dataset
Sync callback function is optional
-
trainAsyncTakes Dataset as Parameter
Performs Training on This Dataset
Also takes async callback function called at the end of every iteration
-
predictPerforms Predictions on the data Provided as Input Output:
{ index: category index, distance: distance to category center, center: category center } -
saveSave trained k-means to a json file. Pls give a '/path/to/xxx.json' into it.
PEER DEPENDENCIES
Typings
As the code is originally written in TypeScript, Type Support is provided out of the box
Contact Me
You could contact me devilyouwei
Thanks to pratikpc
You could file issues or add features via Pull Requests on GitHub
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent