Skip to content

Latest commit

 

History

History
107 lines (96 loc) · 4.46 KB

readme.md

File metadata and controls

107 lines (96 loc) · 4.46 KB
Raw

Introduction

This is the source code of the variance reduced k-means clustering methods. It is implemented based on a packet: sofia-ml (http://code.google.com/p/sofia-ml). If this project is helpful for you, please cite our paper as follows.

@article{Zhao:2018NEU,
 author = {Zhao, Yawei and Ming, Yuewei and Zhao, Kaikai and Liu, Xinwang and Zhu, En and  Yin, Jianping},
 title = {Large-scale k-means clustering via variance reduction},
 journal = {Elsevier Neurocomputing},
 year = {2018},
}

Source code

The main implementation of our algorithm is in the file "cluster-src/sf-kmeans-methods.cc". We add a function in the file and implement our algorithm in it.

void SVRGKmeans(int num_iterations,
                int num_m,
                const SfDataSet& data_set,
                SfClusterCenters* cluster_centers,
                float eta)

num_iterations
  Number of optimization iterations to take.
num_m
  Number of optimization inner iterations to take.
data_set
  Dataset including example number and dimensionality.
cluster_centers
  Cluster centers including cluster center number.
eta
  Learning rate.

Use

  • Compile
cd /path/to/cluster-src/
make

Note that the configurations in cluster-src/Makefile should be modified for your own runtime environment.

  • Execution

The usage of this program is exactly the same as sofia-ml packet. The file, run.sh, gives an example for using the program.

./sofia-kmeans --training_file demo/demo.train --test_file demo/demo.train --cluster_assignments_out result.txt --random_seed 1 --k 2 --init_type optimized_kmeans_pp --opt_type svrg_kmeans --sample_size 1000 --mini_batch_size 300 --iterations 10 --m 100 --eta 0.02 --dimensionality 47697 --objective_after_init --objective_after_training

--training_file
  File to be used for training.
--test_file
  File to be used for testing.
--cluster_assignments_out
  Assign each example in the --test_file to its closest cluster center, and write these results to this file. Format of the file is <nearest center id>TAB<true label (if any)>. Default: no output file.
--random_seed
  When set to non-zero value, use this seed instead of seed from system clock. This can be useful for parameter tuning in cross-validation, as setting a fixed seed by hand forces examples to be sampled in the same order. However for actual training/test, this should never be used. Default: 0
--k
  The number of cluster centers to find. Must be set.
--init_type
  Initialization procedure for seeding the kmeans optimization. Options are:
  random          random selection of cluster centers
  kmeans_pp       kmeans++ initialization method (naive)
  optimized_kmeans_pp   optimized kmeans++
  Default: random
--opt_type
  Optimization procedure for kmeans objective. Options are: batch_kmeans, sgd_kmeans, mini_batch_kmeans, svrg_kmeans
Default: mini_batch_kmeans
--sample_size
  When using sampling_kmeans_pp, the number of examples to sample on each round. Default: 1000
--mini_batch_size
  "When using mini_batch_kmeans, the number of examples to sample on each round. Default: 100
--iterations
  Number of optimization iterations to take. Default: 1000
--m
  Number of optimization inner iterations to take. Default: 1000
--eta
  Learning rate. Default: 0.02
--dimensionality
  Index value of largest feature index in training data set.
--objective_after_init
  Compute value of the kmeans objective function on training data, after initializing the cluster centers. Default is not to do this.
--objective_after_training
  Compute value of the kmeans objective function on training data, after completing training the cluster centers. Default is not to do this.

Numerical results

The acceleration is significant. For example, we conduct k-means clustering on the dataset: Pittsbour (http://www.ok.ctrl.titech.ac.jp/~torii/project/repttile/).

Then we obtain the more than 5x speedup. The evaluations are presented in the following figure.

image

Contact

Please contact us ([email protected], [email protected]) when you have some problems about the source code.

More details

If this project is helpful for you, please cite our paper as follows.

@article{Zhao:2018NEU,
 author = {Zhao, Yawei and Ming, Yuewei and Zhao, Kaikai and Liu, Xinwang and Zhu, En and  Yin, Jianping},
 title = {Large-scale k-means clustering via variance reduction},
 journal = {Elsevier Neurocomputing},
 year = {2018},
}