Fast Pytorch Kmeans

this is a pytorch implementation of K-means clustering algorithm

Installation

pip install fast-pytorch-kmeans

Quick Start

from fast_pytorch_kmeans import KMeans
import torch

kmeans = KMeans(n_clusters=8, mode='euclidean', verbose=1)
x = torch.randn(100000, 64, device='cuda')
labels = kmeans.fit_predict(x)

Speed Comparison

Tested on google colab with Intel(R) Xeon(R) CPU @ 2.00GHz and Nvidia Tesla T4 GPU

sklearn: sklearn.cluster.KMeans

• n_init = 1
• max_iter = 100
• tol = -1 (to force 100 iterations)

faiss: faiss.Clustering

• nredo = 1
• niter = 100
• max_point_per_centroid = 10**9 (to prevent subsample from dataset)

note: time cost for transfering data from cpu to gpu is also included

fast-pytorch: fast_pytorch_kmeans.KMeans

• max_iter = 100
• tol = -1 (to force 100 iterations)
• minibatch = None

1. n_samples=100,000, n_features=256, time spent for 100 iterations

2. n_samples=100,000, n_clusters=256, time spent for 100 iterations

3. n_features=256, n_clusters=256, time spent for 100 iterations

4. n_features=32, n_clusters=1024, time spent for 100 iterations

5. n_features=1024, n_clusters=32, time spent for 100 iterations