Why can't I load a PySpark RandomForestClassifier model?

Question

I can't load a RandomForestClassificationModel saved by Spark.

Environment: Apache Spark 2.0.1, standalone mode running on a small (4 machine) cluster. No HDFS - everything is saved to local disks.

Build and save model:

classifier = RandomForestClassifier(labelCol="label", featuresCol="features", numTrees=50)
model = classifier.fit(train)
result = model.transform(test)
model.write().save("/tmp/models/20161030-RF-topics-cats.model")

Later, in a separate program:

model = RandomForestClassificationModel.load("/tmp/models/20161029-RF-topics-cats.model")

gives:

Py4JJavaError: An error occurred while calling o81.load.
: org.apache.spark.sql.AnalysisException: Unable to infer schema for ParquetFormat at /tmp/models/20161029-RF-topics-cats.model/treesMetadata. It must be specified manually;
    at org.apache.spark.sql.execution.datasources.DataSource$$anonfun$16.apply(DataSource.scala:411)
    at org.apache.spark.sql.execution.datasources.DataSource$$anonfun$16.apply(DataSource.scala:411)
    at scala.Option.getOrElse(Option.scala:121)
    at org.apache.spark.sql.execution.datasources.DataSource.resolveRelation(DataSource.scala:410)
    at org.apache.spark.sql.DataFrameReader.load(DataFrameReader.scala:149)
    at org.apache.spark.sql.DataFrameReader.parquet(DataFrameReader.scala:439)
    at org.apache.spark.sql.DataFrameReader.parquet(DataFrameReader.scala:423)
    at org.apache.spark.ml.tree.EnsembleModelReadWrite$.loadImpl(treeModels.scala:441)
    at org.apache.spark.ml.classification.RandomForestClassificationModel$RandomForestClassificationModelReader.load(RandomForestClassifier.scala:301

I'd note that the same code works when I use a Naive Bayes classifier.

Do you remember what the problem was, or what the solution was? — Mike Williamson
@MikeWilliamson no I don't remember - it was nearly 4 years ago, sorry. From memory each node had access to a node-local save of the model. Speculating some, I wonder is Spark is checking across nodes to verify the model is the same on each, and errors when it finds they are different. In the JVM there is a serialVersionUID that is used for this kind of thing. But as I said this is speculative, I don't know how to fix it, and it is some years since I looked at it. — Nick Lothian
The error itself is about the ParquetFormat schema. Maybe there is a way to extract the schema once when writing the code, and then apply it when loading? — Nick Lothian

Yaron Yaron · Accepted Answer · 2016-10-31T09:49:49

Saving the model to HDFS, and later reading the model from HDFS might solve your problem.

You have 4 nodes, each node has its own local-disk. You are using model.write().save("/temp/xxx")

Later, in a separate program: You are using load("/temp/xxx")

Since there are 4 nodes, with 4 different local disks, it isn't clear to me what exactly is being saved (and to which local-disk) during the write.save() operation, and what exactly is being load() and from which local-disk.

Why can't I load a PySpark RandomForestClassifier model?

1 Answers