[
https://issues.apache.org/jira/browse/SPARK-14174?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16025870#comment-16025870
]
zhengruifeng commented on SPARK-14174:
--------------------------------------
[~mlnick] [~sethah] I am sorry to say that I think current impl may be
incorrect according to scikit-learn's impl. So I think you don't need to go on
reviewing the PR.
MiniBatch KMeans in scikit-learn is implemented according to this
paper[https://www.eecs.tufts.edu/~dsculley/papers/fastkmeans.pdf]
and the centers are updated in an on-line style, which is quite different with
this pr now.
So current PR is lack of theoretical proof, although this impl looks like work
well in some cases.
I will still work on this issue, and I think there are three options:
1, strictly follow the paper, and we should sample the minibatch back to
driver, like
https://github.com/apache/spark/pull/1248/files#diff-f1bafa6656211b6909dbccdb443a3003R162
2, still follow the paper, the sequential update seems can be parallized by
{{foldByKey}}, and the parallelism level is {{k}}.
3, we follow the update method in the paper, but use the gradient of the
minibath to update the centers; In this way, we modify the algorithm.
If I obtain some test result, I will post it here.
> Accelerate KMeans via Mini-Batch EM
> -----------------------------------
>
> Key: SPARK-14174
> URL: https://issues.apache.org/jira/browse/SPARK-14174
> Project: Spark
> Issue Type: Improvement
> Components: ML
> Reporter: zhengruifeng
> Attachments: MiniBatchKMeans_Performance_II.pdf,
> MiniBatchKMeans_Performance.pdf
>
>
> The MiniBatchKMeans is a variant of the KMeans algorithm which uses
> mini-batches to reduce the computation time, while still attempting to
> optimise the same objective function. Mini-batches are subsets of the input
> data, randomly sampled in each training iteration. These mini-batches
> drastically reduce the amount of computation required to converge to a local
> solution. In contrast to other algorithms that reduce the convergence time of
> k-means, mini-batch k-means produces results that are generally only slightly
> worse than the standard algorithm.
> I have implemented mini-batch kmeans in Mllib, and the acceleration is realy
> significant.
> The MiniBatch KMeans is named XMeans in following lines.
> {code}
> val path = "/tmp/mnist8m.scale"
> val data = MLUtils.loadLibSVMFile(sc, path)
> val vecs = data.map(_.features).persist()
> val km = KMeans.train(data=vecs, k=10, maxIterations=10, runs=1,
> initializationMode="k-means||", seed=123l)
> km.computeCost(vecs)
> res0: Double = 3.317029898599564E8
> val xm = XMeans.train(data=vecs, k=10, maxIterations=10, runs=1,
> initializationMode="k-means||", miniBatchFraction=0.1, seed=123l)
> xm.computeCost(vecs)
> res1: Double = 3.3169865959604424E8
> val xm2 = XMeans.train(data=vecs, k=10, maxIterations=10, runs=1,
> initializationMode="k-means||", miniBatchFraction=0.01, seed=123l)
> xm2.computeCost(vecs)
> res2: Double = 3.317195831216454E8
> {code}
> The above three training all reached the max number of iterations 10.
> We can see that the WSSSEs are almost the same. While their speed perfermence
> have significant difference:
> {code}
> KMeans 2876sec
> MiniBatch KMeans (fraction=0.1) 263sec
> MiniBatch KMeans (fraction=0.01) 90sec
> {code}
> With appropriate fraction, the bigger the dataset is, the higher speedup is.
> The data used above have 8,100,000 samples, 784 features. It can be
> downloaded here
> (https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass/mnist8m.scale.bz2)
> Comparison of the K-Means and MiniBatchKMeans on sklearn :
> http://scikit-learn.org/stable/auto_examples/cluster/plot_mini_batch_kmeans.html#example-cluster-plot-mini-batch-kmeans-py
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
