Github user EntilZha commented on a diff in the pull request: https://github.com/apache/spark/pull/4047#discussion_r23502411 --- Diff: mllib/src/main/scala/org/apache/spark/mllib/clustering/LDA.scala --- @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.mllib.clustering + +import java.util.Random + +import breeze.linalg.{DenseVector => BDV, sum => brzSum, normalize, axpy => brzAxpy} + +import org.apache.spark.Logging +import org.apache.spark.annotation.DeveloperApi +import org.apache.spark.graphx._ +import org.apache.spark.mllib.linalg.Vector +import org.apache.spark.rdd.RDD +import org.apache.spark.util.Utils + + +/** + * :: DeveloperApi :: + * + * Latent Dirichlet Allocation (LDA), a topic model designed for text documents. + * + * Terminology: + * - "word" = "term": an element of the vocabulary + * - "token": instance of a term appearing in a document + * - "topic": multinomial distribution over words representing some concept + * + * Currently, the underlying implementation uses Expectation-Maximization (EM), implemented + * according to the Asuncion et al. (2009) paper referenced below. + * + * References: + * - Original LDA paper (journal version): + * Blei, Ng, and Jordan. "Latent Dirichlet Allocation." JMLR, 2003. + * - This class implements their "smoothed" LDA model. + * - Paper which clearly explains several algorithms, including EM: + * Asuncion, Welling, Smyth, and Teh. + * "On Smoothing and Inference for Topic Models." UAI, 2009. + * + * NOTE: This is currently marked DeveloperApi since it is under active development and may undergo + * API changes. + */ +@DeveloperApi +class LDA private ( + private var k: Int, + private var maxIterations: Int, + private var topicSmoothing: Double, + private var termSmoothing: Double, + private var seed: Long) extends Logging { + + import LDA._ + + def this() = this(k = 10, maxIterations = 20, topicSmoothing = -1, termSmoothing = -1, + seed = Utils.random.nextLong()) + + /** + * Number of topics to infer. I.e., the number of soft cluster centers. + * (default = 10) + */ + def getK: Int = k + + def setK(k: Int): this.type = { + require(k > 0, s"LDA k (number of clusters) must be > 0, but was set to $k") + this.k = k + this + } + + /** + * Topic smoothing parameter (commonly named "alpha"). + * + * This is the parameter to the Dirichlet prior placed on the per-document topic distributions + * ("theta"). We use a symmetric Dirichlet prior. + * + * This value should be > 0.0, where larger values mean more smoothing (more regularization). + * If set to -1, then topicSmoothing is set automatically. + * (default = -1 = automatic) + * + * Automatic setting of parameter: + * - For EM: default = (50 / k) + 1. + * - The 50/k is common in LDA libraries. + * - The +1 follows Asuncion et al. (2009), who recommend a +1 adjustment for EM. + */ + def getTopicSmoothing: Double = topicSmoothing + + def setTopicSmoothing(topicSmoothing: Double): this.type = { + require(topicSmoothing > 0.0 || topicSmoothing == -1.0, + s"LDA topicSmoothing must be > 0 (or -1 for auto), but was set to $topicSmoothing") + if (topicSmoothing > 0.0 && topicSmoothing <= 1.0) { + logWarning(s"LDA.topicSmoothing was set to $topicSmoothing, but for EM, we recommend > 1.0") + } + this.topicSmoothing = topicSmoothing + this + } + + /** + * Term smoothing parameter (commonly named "eta"). + * + * This is the parameter to the Dirichlet prior placed on the per-topic word distributions + * (which are called "beta" in the original LDA paper by Blei et al., but are called "phi" in many + * later papers such as Asuncion et al., 2009.) + * + * This value should be > 0.0. + * If set to -1, then termSmoothing is set automatically. + * (default = -1 = automatic) + * + * Automatic setting of parameter: + * - For EM: default = 0.1 + 1. + * - The 0.1 gives a small amount of smoothing. + * - The +1 follows Asuncion et al. (2009), who recommend a +1 adjustment for EM. + */ + def getTermSmoothing: Double = termSmoothing + + def setTermSmoothing(termSmoothing: Double): this.type = { + require(termSmoothing > 0.0 || termSmoothing == -1.0, + s"LDA termSmoothing must be > 0 (or -1 for auto), but was set to $termSmoothing") + if (termSmoothing > 0.0 && termSmoothing <= 1.0) { + logWarning(s"LDA.termSmoothing was set to $termSmoothing, but for EM, we recommend > 1.0") + } + this.termSmoothing = termSmoothing + this + } + + /** + * Maximum number of iterations for learning. + * (default = 20) + */ + def getMaxIterations: Int = maxIterations + + def setMaxIterations(maxIterations: Int): this.type = { + this.maxIterations = maxIterations + this + } + + /** Random seed */ + def getSeed: Long = seed + + def setSeed(seed: Long): this.type = { + this.seed = seed + this + } + + /** + * Learn an LDA model using the given dataset. + * + * @param documents RDD of documents, where each document is represented as a vector of term + * counts plus an ID. Document IDs must be >= 0. + * @return Inferred LDA model + */ + def run(documents: RDD[Document]): DistributedLDAModel = { + val topicSmoothing = if (this.topicSmoothing > 0) { + this.topicSmoothing + } else { + (50.0 / k) + 1.0 + } + val termSmoothing = if (this.termSmoothing > 0) { + this.termSmoothing + } else { + 1.1 + } + var state = LDA.initialState(documents, k, topicSmoothing, termSmoothing, seed) + var iter = 0 + while (iter < maxIterations) { + state = state.next() + iter += 1 + } + new DistributedLDAModel(state) + } +} + + +object LDA { + + /* + DEVELOPERS NOTE: + + This implementation uses GraphX, where the graph is bipartite with 2 types of vertices: + - Document vertices + - indexed {0, 1, ..., numDocuments-1} + - Store vectors of length k (# topics). + - Term vertices + - indexed {-1, -2, ..., -vocabSize} + - Store vectors of length k (# topics). + - Edges correspond to terms appearing in documents. + - Edges are directed Document -> Term. + - Edges are partitioned by documents. + + Info on EM implementation. + - We follow Section 2.2 from Asuncion et al., 2009. We use some of their notation. + - In this implementation, there is one edge for every unique term appearing in a document, + i.e., for every unique (document, term) pair. + - Notation: + - N_{wkj} = count of tokens of term w currently assigned to topic k in document j + - N_{*} where * is missing a subscript w/k/j is the count summed over missing subscript(s) + - gamma_{wjk} = P(z_i = k | x_i = w, d_i = j), + the probability of term x_i in document d_i having topic z_i. + - Data graph + - Document vertices store N_{kj} + - Term vertices store N_{wk} + - Edges store N_{wj}. + - Global data N_k + - Algorithm + - Initial state: + - Document and term vertices store random counts N_{wk}, N_{kj}. + - E-step: For each (document,term) pair i, compute P(z_i | x_i, d_i). + - Aggregate N_k from term vertices. + - Compute gamma_{wjk} for each possible topic k, from each triplet. + using inputs N_{wk}, N_{kj}, N_k. + - M-step: Compute sufficient statistics for hidden parameters phi and theta + (counts N_{wk}, N_{kj}, N_k). + - Document update: + - N_{kj} <- sum_w N_{wj} gamma_{wjk} + - N_j <- sum_k N_{kj} (only needed to output predictions) + - Term update: + - N_{wk} <- sum_j N_{wj} gamma_{wjk} + - N_k <- sum_w N_{wk} + */ + + /** + * :: DeveloperApi :: + * + * Document with an ID. + * + * @param counts Vector of term (word) counts in the document. + * This is the "bag of words" representation. + * @param id Unique ID associated with this document. + * Documents should be indexed {0, 1, ..., numDocuments-1}. + * + * TODO: Can we remove the id and still be able to zip predicted topics with the Documents? + * + * NOTE: This is currently marked DeveloperApi since it is under active development and may + * undergo API changes. + */ + @DeveloperApi + case class Document(counts: Vector, id: Long) + + /** + * Vector over topics (length k) of token counts. + * The meaning of these counts can vary, and it may or may not be normalized to be a distribution. + */ + private[clustering] type TopicCounts = BDV[Double] --- End diff -- I used Ints primarily because in the Gibbs implementation I always used integer counts and didn't use a normalized distribution anywhere except in the final output. To me, it seems like using an integer type (either Int or Long) is simpler, but I recognize maybe I am missing something. Could you reference some of those places (I am curious if its general to LDA or specific to EM that makes it easier). Secondary reason, since I always use integers, Int/Long guarantees no loss of precision, granted, loss of precision would be very rare and you would need to get very high in counts for it to happen (unlikely). On Int vs Long, agreed if you want to manage overflow, then Long is good. The tradeoff is that for cases where a single given topic doesn;t exceed MAX_INT (which is ~2 billion), you are effectively doubling memory usage since most of the memory in LDA is consumed by Arrays (or Vectors) of topic counts. Again though, since the data type is tied to the implementation of EM, this would very neatly fit within EMLearningState and have the methods of the LearningState trait return regular types instead of aliased types (Arrays, Ints, Longs, Doubles, Vectors...), so it may be a moot point with regards to affecting Gibbs.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org