This will be an extension to Spark, but should be builtin to MLeap.

Add in support for Spark's MultiLayerPerceptron

Like other scala based projects the release jar should have scala version in jar name or at least in the documentation.

Use Option monad for null values.
Data types should be optionally nullable.

How can I integrate custom spark transformers and estimators into mleap?

I am thinking of preprocessing steps, nan cleaning, ...

When I deserialize the bundle model (a simple Random Forest Model) from jar zip file, like
val bundle = BundleFile("jar:file:/home//userA/rf.zip").load().get
I got a bundle of type ml.combust.bundle.dsl.Bundle[Nothing],

and then when I access root, I got following exceptions:
java.lang.ClassCastException: ml.combust.mleap.runtime.transformer.Pipeline cannot be cast to scala.runtime.Nothing$

Thanks for help in advance!

Requires JDK 8
Requires Scala 2.11

So that people can evaluate things faster to see whether this works with their tech stack or not.

Add a transformer to a new module called "mleap-tensorflow" that allows passing data into a tensorflow graph for transformation.

Include an optional schema.json file in the root bundle. Only include this if there is enough information to accurately describe the input and output schemas.

For Spark-trained pipelines, we will have to include the DataFrame used to train the pipeline while we serialize the model. SparkBundleContext already has an optional DataFrame for this purpose.

Right now we only support protobuf serialization of decision trees. Let's offer JSON serialization as well for when JSON-only serialization is used.

We should add imputer to MLeap based on the Spark transformer.

1. MLeap core model
2. MLeap runtime transformer
3. MLeap transformer NodeOp
4. Spark transformer NodeOp

Currently working with MLeap from Java can be a pain. Let's make the interface nicer.

This epic is for Spark transformers that are rather tricky for one reason or another to adapt to MLeap.

This usually is because multiple data frames may be involved in the transform process. MLeap will have to come up with a solution to this as we move forward.

Coalesce transformer takes in multiple columns and chooses the first non-null value. Supports only doubles and nullable doubles.

StringMap takes in a string and outputs a double using a user-defined map.

Spark support for these two transformers will come with a later ticket.

Hi,

I just tried mleap, which is really awesome. However, I was wondering is there a way to get the schema of the exported model in a format such as PMML? So that we can have a better overview of what types of features the model is using and information like that?

Thanks

Implement these transformers for Spark

Dear sir or Madam,
When I try to save the model using:
val sbc = SparkBundleContext().withDataset(pipeline.transform(df))
there is following exception:
com.typesafe.config.ConfigException$Missing: No configuration setting found for key 'ml'

I am using iheart/ficus for configuration setting, which is built on top of the typesafe.config library.
So in the beginning of my program, i have code like:
val conf = ConfigFactory.load()
val settings = new Settings(conf)
which will read configuration from reference.conf and application.conf.

When I test the code in spark-shell, without using typesafe.config library, the code works.

How can I continue to use typesafe.config on my own without breaking MLeap?

Thanks.

This epic is to track the progress towards full Spark support. It will not include transformers that require multiple data frames (recommendation algorithm) and it will not include LDA, which is a rather large undertaking to get all the code into place for it.

Support LDA clustering algorithm

Hi,

Just want to say this project is pretty cool and thanks for your effort!

We are looking for some solution to train models offline using Spark, yet score online in real time. This is exactly what we need.

I was trying to follow the minimal doc in the wiki page. https://github.com/combust-ml/mleap/wiki/Setting-up-a-Spark-2.0-notebook-with-MLeap-an-Toree

The page seems not finished yet. Any unit test class of example that I can follow to use MLeap?

Also, a couple of corrections.

• page title should be "... MLeap and Toree"
• in build & install toree, it should be "pip install toree-0.2.0.dev1.tar.gz"
• also, it should be SPARK_HOME=... jupyter toree install

Again. Thanks!

Build out a Python module to serialize Scikit-learn + Pandas transformer pipelines to MLeap. We do not need to support deserialization to start.

1. Support Bundle.ML JSON format
2. Support several feature transformers, regression algorithms and classifiers
3. Make sure decision tree serialization is working properly
4. Publish module to PIP when we release MLeap 0.5.0
5. Add documentation for Python SK Learn integration
6. Create some notebooks showing SK learn

1. Core Model
2. Transformer
3. Spark Serialization
4. MLeap Serialization
5. reference.conf files with added ops

I have started experimenting with serializing and deserializing spark pipelines. However, I have noticed that when I override default params, they are missing after deserialization. I have narrowed down the cause to how the OpNode#load method is implemented, and specifically the use of .copy(model.extractParamMap()).

I am not very familiar with this copy API provided by spark, so cannot figure out if this is a Spark bug or if this is misuse of the API. So the only solution I've thought so far is just to explicitly get and set each param (as done in OpModel#load).

Here is a reproducible case, using Binarizer as an example:

import org.apache.spark.ml.feature.Binarizer
import ml.combust.mleap.spark.SparkSupport._

val bin = new Binarizer("bin")
  .setInputCol("in")
  .setOutputCol("out")
  .setThreshold(0.5)

val path = new File(...)
bin.serializeToBundle(path)

val bin2 = path.deserializeBundle()._2.asInstanceOf[Binarizer]
assert(bin.getInputCol == bin2.getInputCol)
assert(bin.getOutputCol == bin2.getOutputCol)
assert(bin.getThreshold == bin2.getThreshold) //fails

We should start supporting backwards to at least Spark 2.0 with the use of bundle registry config files specific to each version.

Need to make sure that our MLeap transformers work in complete parity with Spark transformers. Use our Spark integration to do this.

I'm playing https://github.com/combust/mleap/wiki/Serializing-a-Spark-ML-Pipeline-and-Scoring-with-MLeap#serialize-the-ml-data-pipeline-and-rf-model-to-bundleml and I struggled to realize that one should read featureModel instead of featurePipeline in

val pipeline = SparkUtil.createPipelineModel(uid = "pipeline", Array(featurePipeline, rf))

I could'nt find another way to contribute to the wiki (like PR) - feel free to tell me how you wish to receive such contribs.

mleap seems to support RandomForestClassifier. What about xgboost, especially xgboost4j?
https://github.com/komiya-atsushi/xgboost-predictor-benchmark provides a great jvm based and very quick evaluator. Maybe this would be helpful for a xgboost-mleap integration.

This should be implemented using Bundle.ML custom types.

This will require creating a converter form MLeap UDFs to Spark UDFs

Standardizing ML Pipeline Serialization

Currently there is a large array of serialization formats for machine learning models:

1. PMML is an XML-based format primarily targeting the JVM for executing ML models
2. Scikit-learn relies on Python pickling to export models
3. Spark has a serialization format based on Parquet and JSON
4. Various other libraries such as Caffe, Torch, MLDB, etc. have their own custom file formats they use to store models with

We propose a serialization format that is highly-extensible, portable across language and platforms, open-source and with a reference implementation in both Scala and Rust. We call this serialization format Bundle.ML.

Key Features

1. It should be easy for developers to add custom transformers in Scala, Java, Python, C, Rust, or any other language
2. The serialization format should be flexible and meet state-of-the-art performance requirements. This means being able to serialize arbitrarily-large random forest, linear, or neural network models.
3. Serialization should be optimized for ML Transformers and Pipelines as seen in Scikit-learn and Spark, but it should also support non-pipeline based frameworks such as H2O
4. Serialization should be accessible for all environments and platforms, including low-level languages like C, C++ and Rust
5. Provide a common, extensible serialization format for any technology to integrate with via custom transformers or core transformers
6. Serialization/Deserialization should be possible with as many technologies as possible to make the models truly portable between different platforms. ie, we should be able to train a pipeline with Scikit-learn then execute it in Spark.

Allow users to store arbitrary meta data in the bundle file.

This can be useful for:

1. Quick training summary statistics
2. Information about labels in the model that could be used later
3. Descriptions, notes, etc.

Spark has a nice feature that let's you build a Dataset from a case class. We should support this as well.

MleapReflection provides many tools that will be needed for this task.
I am thinking we should support the following conversions:

1. (case class) -> DefaultLeapFrame w/ 1 row
2. Seq(case class) -> DefaultLeapFrame w/ n rows
3. DefaultLeapFrame -> (case class), extracts first row (throw error if more than one?)
4. DefaultLeapFrame -> Seq(case class), extracts all rows into a Seq of a case class

These conversions should be implicit and included in the MleapSupport trait for easy usage.

We want to use NIO FileSystem objects to serialize Bundle.ML, this will make it much more versatile and simplify the code a great deal. Also, some small tweaks to how we serialize Bundle.ML root-level components.

1. NIO FileSystem objects for serializing
2. bundle.json should only include version, uid, serialization format
3. root-level transformer should be in a folder called root, next to bundle.json
4. get rid of custom attributes on the Bundle

1. Split off our extensions to Spark transformers into a new module
2. Support the out-of-the-box transformers by using metadata from a transformed DataFrame

Spark added support for multinomial logistic regression, we should do the same