Skip to content

[SPARK-20797][MLLIB]fix LocalLDAModel.save() bug. #18034

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
d0evi1 wants to merge 2 commits into from
Closed

[SPARK-20797][MLLIB]fix LocalLDAModel.save() bug. #18034

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
69e514e
LDAModel.save() bug fix.
d0evi1 May 19, 2017
edd999d
LDAModel.save() bug fix, add some code like word2vec's save method.
d0evi1 May 19, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 11 additions & 1 deletion mllib/src/main/scala/org/apache/spark/mllib/clustering/LDAModel.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@ import org.apache.spark.mllib.util.{Loader, Saveable}
import org.apache.spark.rdd.RDD
import org.apache.spark.sql.{Row, SparkSession}
import org.apache.spark.util.BoundedPriorityQueue
import org.apache.spark.util.Utils

/**
* Latent Dirichlet Allocation (LDA) model.
Expand Down Expand Up @@ -468,7 +469,16 @@ object LocalLDAModel extends Loader[LocalLDAModel] {
val topics = Range(0, k).map { topicInd =>
Data(Vectors.dense((topicsDenseMatrix(::, topicInd).toArray)), topicInd)
}
spark.createDataFrame(topics).repartition(1).write.parquet(Loader.dataPath(path))

val bufferSize = Utils.byteStringAsBytes(
spark.conf.get("spark.kryoserializer.buffer.max", "64m"))
// We calculate the approximate size of the model
// We only calculate the array size, considering an
// average string size of 15 bytes, the formula is:
// (floatSize * vectorSize + 15) * numWords
val approxSize = (4L * k + 15) * topicsMatrix.numRows
val nPartitions = ((approxSize / bufferSize) + 1).toInt
spark.createDataFrame(topics).repartition(nPartitions).write.parquet(Loader.dataPath(path))
Copy link
Member

@srowen srowen May 19, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The problem is that this writes multiple files. I don't think you can do that.

Copy link
Author

@d0evi1 d0evi1 May 20, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why not? i think it does works. the multiple parquet files will may be in random order, but it will save topic indices. when u call load process, parquet will restore dataframe, u can check the LocalLDAModel's load method, it will scan all dataframe's row with the topic indices to rebuild the (topic x vocab) breeze matrix.

Copy link
Member

@srowen srowen May 20, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe the point of repartition(1) is to export this to a single file for external consumption. Of course writing it succeeds, but I don't think this is what it is intended to do.

Copy link
Author

@d0evi1 d0evi1 May 20, 2017 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I've read mllib's online lda's code implements and the paper.

i found a simple way to test it , u can try this code snippet, simulation a matrix to save:

val vocabSize = 500000
val k = 300
val random = new Random()
val topicsDenseMatrix = DenseMatrix.fill[Double](vocabSize, k)(random.nextDouble())

Copy link
Member

@srowen srowen May 22, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@d0evi1 I'm not sure how that's related to my comment.
@hhbyyh what do you think -- am I wrong about the issue?

Copy link
Contributor

@hhbyyh hhbyyh May 22, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the PR is trying to do something similar to https://github.com/apache/spark/pull/9989/files.

Copy link
Contributor

@Krimit Krimit Jun 29, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@srowen what's your concern with saving multiple files instead of one?

We've encountered crippling errors saving large LDA models in the ml api as well. The problem there is even worse, since the entire matrix is treated as one single datum

Copy link
Contributor

@WeichenXu123 WeichenXu123 Sep 26, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you add a testcase to save model into multiple files and load back and check the correctness ?

}

def load(
Expand Down