分布式LDA

Reading time ~2 minutes

1.搭建集群

建议先看下教程分布式LSA,建立一个LDA的集群完全类似,不同的是,你需要运行lda_worker 和 lda_dispatcher脚本。

2.运行LDA

运行LDA,但是打开开关 distributed=True 参数:

>>> # extract 100 LDA topics, using default parameters
>>> lda = LdaModel(corpus=mm, id2word=id2word, num_topics=100, distributed=True)
using distributed version with 4 workers
running online LDA training, 100 topics, 1 passes over the supplied corpus of 3199665 documets, updating model once every 40000 documents
..

在串行模式下(非分布式),创建online LDA的wikipedia模型需要10h56m (OS X, C2D 2.53GHz, 4GB RAM with libVec). 在4 worker的分布式模型下(Linux, Xeons of 2Ghz, 4GB RAM with ATLAS),整个过程只需要3h20m.

运行标准的batch LDA(非online),你需要进行相类似的调用:

>>> lda = LdaModel(corpus=mm, id2word=id2token, num_topics=100, update_every=0, passes=20, distributed=True)
using distributed version with 4 workers
running batch LDA training, 100 topics, 20 passes over the supplied corpus of 3199665 documets, updating model once every 3199665 documents
initializing workers
iteration 0, dispatching documents up to #10000/3199665
iteration 0, dispatching documents up to #20000/3199665
...

接着,两天之后:

iteration 19, dispatching documents up to #3190000/3199665
iteration 19, dispatching documents up to #3199665/3199665
reached the end of input; now waiting for all remaining jobs to finish

>>> lda.print_topics(20)
topic #0: 0.007*disease + 0.006*medical + 0.005*treatment + 0.005*cells + 0.005*cell + 0.005*cancer + 0.005*health + 0.005*blood + 0.004*patients + 0.004*drug
topic #1: 0.024*king + 0.013*ii + 0.013*prince + 0.013*emperor + 0.008*duke + 0.008*empire + 0.007*son + 0.007*china + 0.007*dynasty + 0.007*iii
topic #2: 0.031*film + 0.017*films + 0.005*movie + 0.005*directed + 0.004*man + 0.004*episode + 0.003*character + 0.003*cast + 0.003*father + 0.003*mother
topic #3: 0.022*user + 0.012*edit + 0.009*wikipedia + 0.007*block + 0.007*my + 0.007*here + 0.007*edits + 0.007*blocked + 0.006*revert + 0.006*me
topic #4: 0.045*air + 0.026*aircraft + 0.021*force + 0.018*airport + 0.011*squadron + 0.010*flight + 0.010*military + 0.008*wing + 0.007*aviation + 0.007*f
topic #5: 0.025*sun + 0.022*star + 0.018*moon + 0.015*light + 0.013*stars + 0.012*planet + 0.011*camera + 0.010*mm + 0.009*earth + 0.008*lens
topic #6: 0.037*radio + 0.026*station + 0.022*fm + 0.014*news + 0.014*stations + 0.014*channel + 0.013*am + 0.013*racing + 0.011*tv + 0.010*broadcasting
topic #7: 0.122*image + 0.099*jpg + 0.046*file + 0.038*uploaded + 0.024*png + 0.014*contribs + 0.013*notify + 0.013*logs + 0.013*picture + 0.013*flag
topic #8: 0.036*russian + 0.030*soviet + 0.028*polish + 0.024*poland + 0.022*russia + 0.013*union + 0.012*czech + 0.011*republic + 0.011*moscow + 0.010*finland
topic #9: 0.031*language + 0.014*word + 0.013*languages + 0.009*term + 0.009*words + 0.008*example + 0.007*names + 0.007*meaning + 0.006*latin + 0.006*form
topic #10: 0.029*w + 0.029*toronto + 0.023*l + 0.020*hockey + 0.019*nhl + 0.014*ontario + 0.012*calgary + 0.011*edmonton + 0.011*hamilton + 0.010*season
topic #11: 0.110*wikipedia + 0.110*articles + 0.030*library + 0.029*wikiproject + 0.028*project + 0.019*data + 0.016*archives + 0.012*needing + 0.009*reference + 0.009*statements
topic #12: 0.032*http + 0.030*your + 0.022*request + 0.017*sources + 0.016*archived + 0.016*modify + 0.015*changes + 0.015*creation + 0.014*www + 0.013*try
topic #13: 0.011*your + 0.010*my + 0.009*we + 0.008*don + 0.008*get + 0.008*know + 0.007*me + 0.006*think + 0.006*question + 0.005*find
topic #14: 0.073*r + 0.066*japanese + 0.062*japan + 0.018*tokyo + 0.008*prefecture + 0.005*osaka + 0.004*j + 0.004*sf + 0.003*kyoto + 0.003*manga
topic #15: 0.045*da + 0.045*fr + 0.027*kategori + 0.026*pl + 0.024*nl + 0.021*pt + 0.017*en + 0.015*categoria + 0.014*es + 0.012*kategorie
topic #16: 0.010*death + 0.005*died + 0.005*father + 0.004*said + 0.004*himself + 0.004*took + 0.004*son + 0.004*killed + 0.003*murder + 0.003*wife
topic #17: 0.027*book + 0.021*published + 0.020*books + 0.014*isbn + 0.010*author + 0.010*magazine + 0.009*press + 0.009*novel + 0.009*writers + 0.008*story
topic #18: 0.027*football + 0.024*players + 0.023*cup + 0.019*club + 0.017*fc + 0.017*footballers + 0.017*league + 0.011*season + 0.007*teams + 0.007*goals
topic #19: 0.032*band + 0.024*album + 0.014*albums + 0.013*guitar + 0.013*rock + 0.011*records + 0.011*vocals + 0.009*live + 0.008*bass + 0.008*track

如果你使用gensim的分布式实现,可以通过email告诉我。我将了解下你的应用和场景,以及遇到的相关问题,以便在将来做改进。

英文原版