grv1207/deepdist

Distributed Deep Learning on Spark

Training deep belief networks requires extensive data and computation. DeepDist accelerates the training by distributing stochastic gradient descent for data stored on HDFS / Spark via a simple Python interface. Overview: deepdist.com

Quick start:

Training of a word2vec model on wikipedia in 15 lines of code:

from deepdist import DeepDist
from gensim.models.word2vec import Word2Vec
from pyspark import SparkContext

sc = SparkContext()
corpus = sc.textFile('enwiki').map(lambda s: s.split())

def gradient(model, sentences):  # executes on workers
    syn0, syn1 = model.syn0.copy(), model.syn1.copy()
    model.train(sentences)
    return {'syn0': model.syn0 - syn0, 'syn1': model.syn1 - syn1}

def descent(model, update):      # executes on master
    model.syn0 += update['syn0']
    model.syn1 += update['syn1']

with DeepDist(Word2Vec(corpus.collect())) as dd:

    dd.train(corpus, gradient, descent)
    print dd.model.most_similar(positive=['woman', 'king'], negative=['man'])

How does it work?

DeepDist implements a Downpour-like stochastic gradient descent. It start a master model server (on port 5000). On each data node, DeepDist fetches the model from the server, and then calls gradient(). After computing the gradient for each RDD partition, gradient updates are sent to the server. On the server, the master model is then updated by descent().

Python module

DeepDist provides a simple Python interface. The with statement starts the model server. Distributed gradient updates are computed on partitions of a resilient distributed dataset (RDD) data. The gradient updates are incorporated into the master model via custom descent method.

from deepdist import DeepDist
 
with DeepDist(model) as dd:    # initialized server with any model    
    
    dd.train(data, gradient, descent)
    # train with an RDD "data" by computing distributed gradients and
    # descending the model parameters space according to gradient updates
 
def gradient(model, data):
    # model is a copy of the master model
    # data is an iterator for the current partition of the data RDD
    # returns the gradient update
 
def descent(model, update):
    # model is a reference to the server model
    # update is a copy of a worker's update

Training Speed

Training speed can be greatly enhanced by adaptively adjusting the learning rate by AdaGrad. The complete Word2Vec model with 900 dimensions can be trained on the 19GB wikipedia corpus (using the words from the validation questions).

References

J Dean, GS Corrado, R Monga, K Chen, M Devin, QV Le, MZ Mao, M’A Ranzato, A Senior, P Tucker, K Yang, and AY Ng. Large Scale Distributed Deep Networks. NIPS 2012: Neural Information Processing Systems, Lake Tahoe, Nevada, 2012.

T Mikolov, I Sutskever, K Chen, G Corrado, and J Dean. Distributed Representations of Words and Phrases and their Compositionality. In Proceedings of NIPS, 2013.

T Mikolov, K Chen, G Corrado, and J Dean. Efficient Estimation of Word Representations in Vector Space. In Proceedings of Workshop at ICLR, 2013.