doddle-model

doddle-model is an in-memory machine learning library that can be summed up with three main characteristics:

it is built on top of Breeze
it provides immutable estimators that are a doddle to use in parallel code
it exposes its functionality through a scikit-learn-like API [2] in idiomatic Scala using typeclasses

How does it compare to existing solutions?

doddle-model takes the position of scikit-learn in Scala and as a consequence, it’s much more lightweight than e.g. Spark ML. Fitted models can be deployed anywhere, from simple applications to concurrent, distributed systems built with Akka, Apache Beam or a framework of your choice. Training of estimators happens in-memory, which is advantageous unless you are dealing with enormous datasets that absolutely cannot fit into RAM.

You can chat with us on gitter.

Installation

The project is published for Scala versions 2.11, 2.12 and 2.13. Add the dependency to your SBT project definition:

libraryDependencies  ++= Seq(
  "io.github.picnicml" %% "doddle-model" % "<latest_version>",
  // add optionally to utilize native libraries for a significant performance boost
  "org.scalanlp" %% "breeze-natives" % "1.0"
)

Note that the latest version is displayed in the maven central badge above and that the v prefix should be removed from the SBT definition.

Getting Started

This is a complete list of code examples, for an example of how to serve a trained doddle-model in a pipeline implemented with Apache Beam see doddle-beam-example.

1. Feature Preprocessing

2. Metrics

3. Baseline models

4. Linear models

5. Model Selection

6. Miscellaneous

7. Use Cases

Kaggle House Prices

Contributing

Want to help us? We have a document that will make deciding how to do that much easier. Be sure to also check the roadmap.

Performance

doddle-model is developed with performance in mind.

1. Native Linear Algebra Libraries

Breeze utilizes netlib-java for accessing hardware optimised linear algebra libraries (note that the breeze-natives dependency needs to be added to the SBT project definition). TL;DR seeing something like

INFO: successfully loaded /var/folders/9h/w52f2svd3jb750h890q1x4j80000gn/T/jniloader3358656786070405996netlib-native_system-osx-x86_64.jnilib

means that BLAS/LAPACK/ARPACK implementations are used. For more information see the Breeze documentation.

2. Memory

If you encounter java.lang.OutOfMemoryError: Java heap space increase the maximum heap size with -Xms and -Xmx JVM properties. E.g. use -Xms8192m -Xmx8192m for initial and maximum heap space of 8Gb. Note that the maximum heap limit for the 32-bit JVM is 4Gb (at least in theory) so make sure to use 64-bit JVM if more memory is needed. If the error still occurs and you are using hyperparameter search or cross validation, see the next section.

3. Parallelism

To limit the number of threads running at one time (and thus memory consumption) when doing cross validation and hyperparameter search, a FixedThreadPool executor is used. By default maximum number of threads is set to the number of system’s cores. Set the -DmaxNumThreads JVM property to change that, e.g. to allow for 16 threads use -DmaxNumThreads=16.

Benchmarks

All experiments ran multiple times (iterations) for all implementations and with fixed hyperparameters, selected in a way such that models yielded similar test set performance.

1. Linear Regression

dataset with 150000 training examples and 27147 test examples (10 features)
each experiment ran for 100 iterations
scikit-learn code, doddle-model code

Implementation	RMSE	Training Time	Prediction Time
scikit-learn	3.0936	0.042s (+/- 0.014s)	0.002s (+/- 0.002s)
doddle-model	3.0936	0.053s (+/- 0.061s)	0.002s (+/- 0.004s)

2. Logistic Regression

dataset with 80000 training examples and 20000 test examples (250 features)
each experiment ran for 100 iterations
scikit-learn code, doddle-model code

Implementation	Accuracy	Training Time	Prediction Time
scikit-learn	0.8389	2.789s (+/- 0.090s)	0.005s (+/- 0.006s)
doddle-model	0.8377	3.080s (+/- 0.665s)	0.025s (+/- 0.025s)

3. Softmax Classifier

MNIST dataset with 60000 training examples and 10000 test examples (784 features)
each experiment ran for 50 iterations
scikit-learn code, doddle-model code

Implementation	Accuracy	Training Time	Prediction Time
scikit-learn	0.9234	21.243s (+/- 0.303s)	0.074s (+/- 0.018s)
doddle-model	0.9223	25.749s (+/- 1.813s)	0.042s (+/- 0.032s)

Core Maintainers

This is a collaborative project which wouldn’t be possible without all the awesome contributors. The core team currently consists of the following developers:

Picnic Machine Learning