v0.17

Highlights

• LightGBM evaluation 3-4x faster!
• Spark Serving v2
• LightGBM training supports early stopping and regularization
• LIME on Spark significantly faster

New Features

Spark Serving v2:

• Both Microbatch and Continuous mode have sub-millisecond latency
• Supports fault tolerance
• Can reply from anywhere in the pipeline
• Fail fast modes for warning callers of bad JSON parsing
• Fully based on DataSource API v2

LightGBM:

• 3-4x evaluation performance improvement
• Add early stopping capabilities
• Added L1 and L2 Regularization parameters
• Made network init more robust
• Fixed bug caused by empty partitions

LIME on Spark:

• LIME Parallelization significantly faster for large datasets
• Tabular Lime now supported

Other:

• Added UnicodeNormalizer for working with complex text
• Recognize Text exposes parameters for its polling handlers

Acknowledgements

We would like to acknowledge the developers and contributors, both internal and external who helped create this version of MMLSpark.

• Ilya Matiach, Markus Cozowicz, Scott Graham, Daniel Ciborowski, Jeremy Reynolds, Miguel Fierro, Robert Alexander, Tao Wu, Sudarshan Raghunathan, Anand Raman,Casey Hong, Karthik Rajendran, Dalitso Banda, Manon Knoertzer, Lars Ahlfors, The Microsoft AI Development Acceleration Program, Cognitive Search Team, Azure Search Team

v0.16

New Features

• Added the AzureSearchWriter for integrating Spark with Azure Search
• Added the Smart Adaptive Recommender (SAR) for better recommendations in SparkML
• Added Named Entity Recognition Cognitive Service on Spark
• Several new LightGBM features (Multiclass Classification, Windows Support, Class Balancing, Custom Boosting, etc.)
• Added Ranking Train Validation Splitter for easy ranking experiments
• All Computer Vision Services can now send binary data or URLs to Cognitive Services

New Examples

• Learn how to use the Azure Search writer to create a visual search system for The Metropolitan Museum of Art with: AzureSearchIndex - Met Artworks.ipynb

Updates and Improvements

General

• MMLSpark Image Schema now unified with Spark Core
  • Now supports Query pushdown and Deep Learning Pipelines
• Bugfixes for Text Analytics services
• PageSplitter now propagates nulls
• HTTP on Spark now supports socket and read timeouts
• HyperparamBuilder python wrappers now return idiomatic python objects

LightGBM on Spark

• Added multiclass classification
• Added multiple types of boosting (Gradient Boosting Decision Tree, Random Forest, Dropout meet Multiple Additive Regression Trees, Gradient-based One-Side Sampling)
• Added windows OS support/bugfix
• LightGBM version bumped to 2.2.200
• Added native support for categorical columns, either through Spark's StringIndexer, MMLSpark's ValueIndexer or list of indexes/slot names parameter
• isUnbalance parameter for unbalanced datasets
• Added boost from average parameter

Acknowledgements

We would like to acknowledge the developers and contributors, both internal and external who helped create this version of MMLSpark.

• Ilya Matiach, Casey Hong, Daniel Ciborowski, Karthik Rajendran, Dalitso Banda, Manon Knoertzer, Sudarshan Raghunathan, Anand Raman,Markus Cozowicz, The Microsoft AI Development Acceleration Program, Cognitive Search Team, Azure Search Team

v0.15

New Features

• Add the TagImage and DescribeImage services
• Add Ranking Cross Validator and Evaluator

New Examples

• Learn how to use HTTP on Spark to work with arbitrary web services at scale in HttpOnSpark - Working with Arbitrary Web APIs.ipynb

Updates and Improvements

LightGBM

• Fix issue with raw2probabilityInPlace
• Add weight column
• Add getModel API to TrainClassifier and TrainRegressor
• Improve robustness of getting executor cores

HTTP on Spark and Spark Serving

• Improve robustness of Gateway creation and management
• Imrpove Gateway documentation

Version Bumps

• Updated to Spark 2.4.0
• LightGBM version update to 2.1.250

Misc

• Fix Flaky Tests
• Remove autogeneration of scalastyle
• Increase training dataset size in snow leopard example

Acknowledgements

We would like to acknowledge the developers and contributors, both internal and external who helped create this version of MMLSpark.

• Ilya Matiach, Casey Hong, Karthik Rajendran, Daniel Ciborowski, Sebastien Thomas, Eli Barzilay, Sudarshan Raghunathan, @flybywind, @wentongxin, @haal

v0.14

New Features

• The Cognitive Services on Spark: A simple and scalable integration between the Microsoft Cognitive Services and SparkML
  • Bing Image Search
  • Computer Vision: OCR, Recognize Text, Recognize Domain Specific Content,
    Analyze Image, Generate Thumbnails
  • Text Analytics: Language Detector, Entity Detector, Key Phrase Extractor,
    Sentiment Detector, Named Entity Recognition
  • Face: Detect, Find Similar, Identify, Group, Verify
• Added distributed model interpretability with LIME on Spark
• 100x lower latencies (<1ms) with Spark Serving
• Expanded Spark Serving to cover the full HTTP protocol
• Added the SuperpixelTransformer for segmenting images
• Added a Fluent API, mlTransform and mlFit, for composing pipelines more elegantly

New Examples

• Chain together cognitive services to understand the feelings of your favorite celebrities with CognitiveServices - Celebrity Quote Analysis.ipynb
• Explore how you can use Bing Image Search and Distributed Model Interpretability to get an Object Detection system without labeling any data in ModelInterpretation - Snow Leopard Detection.ipynb
• See how to deploy any spark computation as a Web service on any Spark platform with the SparkServing - Deploying a Classifier.ipynb notebook

Updates and Improvements

LightGBM

• More APIs for loading LightGBM Native Models
• LightGBM training checkpointing and continuation
• Added tweedie variance power to LightGBM
• Added early stopping to lightGBM
• Added feature importances to LightGBM
• Added a PMML exporter for LightGBM on Spark

HTTP on Spark

• Added the VectorizableParam for creating column parameterizable inputs
• Added handler parameter added to HTTP services
• HTTP on Spark now propagates nulls robustly

Version Bumps

• Updated to Spark 2.3.1
• LightGBM version update to 2.1.250

Misc

• Added Vagrantfile for easy windows developer setup
• Improved Image Reader fault tolerance
• Reorganized Examples into Topics
• Generalized Image Featurizer and other Image based code to handle Binary Files as well as Spark Images
• Added ModelDownloader R wrapper
• Added getBestModel and getBestModelInfo to TuneHyperparameters
• Expanded Binary File Reading APIs
• Added Explode and Lambda transformers
• Added SparkBindings trait for automating spark binding creation
• Added retries and timeouts to ModelDownloader
• Added ResizeImageTransformer to remove ImageFeaturizer dependence on OpenCV

Acknowledgements

We would like to acknowledge the developers and contributors, both internal and external who helped create this version of MMLSpark. (In alphabetical order)

• Abhiram Eswaran, Anand Raman, Ari Green, Arvind Krishnaa Jagannathan, Ben Brodsky, Casey Hong, Courtney Cochrane, Henrik Frystyk Nielsen, Ilya Matiach, Janhavi Suresh Mahajan, Jaya Susan Mathew, Karthik Rajendran, Mario Inchiosa, Minsoo Thigpen, Soundar Srinivasan, Sudarshan Raghunathan, @terrytangyuan

v0.13

New Functionality:

• Export trained LightGBM models for evaluation outside of Spark

• LightGBM on Spark supports multiple cores per executor

• CNTKModel works with multi-input multi-output models of any CNTK
  datatype

• Added Minibatching and Flattening transformers for adding flexible
  batching logic to pipelines, deep networks, and web clients.

• Added Benchmark test API for tracking model performance across
  versions

• Added PartitionConsolidator function for aggregating streaming data
  onto one partition per executor (for use with connection/rate-limited
  HTTP services)

Updates and Improvements:

• Updated to Spark 2.3.0

• Added Databricks notebook tests to build system

• CNTKModel uses significantly less memory

• Simplified example notebooks

• Simplified APIs for MMLSpark Serving

• Simplified APIs for CNTK on Spark

• LightGBM stability improvements

• ComputeModelStatistics stability improvements

Acknowledgements:

We would like to acknowledge the external contributors who helped create
this version of MMLSpark (in order of commit history):

• 严伟, @terrytangyuan, @ywskycn, @dvanasseldonk, Jilong Liao,
  @chappers, @ekaterina-sereda-rf

mmlspark-v0.11: v0.11

New functionality:

• TuneHyperparameters: parallel distributed randomized grid search for
  SparkML and TrainClassifier/TrainRegressor parameters. Sample
  notebook and python wrappers will be added in the near future.

• Added PowerBIWriter for writing and streaming data frames to
  PowerBI.

• Expanded image reading and writing capabilities, including using
  images with Spark Structured Streaming. Images can be read from and
  written to paths specified in a dataframe.

• New functionality for convenient plotting in Python.

• UDF transformer and additional UDFs.

• Expanded pipeline support for arbitrary user code and libraries such
  as NLTK through UDFTransformer.

• Refactored fuzzing system and added test coverage.

• GPU training supports multiple VMs.

Updates:

• Updated to Conda 4.3.31, which comes with Python 3.6.3.

• Also updated SBT and JVM.

Improvements:

• Additional bugfixes, stability, and notebook improvements.

mmlspark-v0.10: v0.10

New functionality:

• We now provide initial support for training on a GPU VM, and an ARM
  template to deploy an HDI Cluster with an associated GPU machine. See
  docs/gpu-setup.md for instructions on setting this up.

• New auto-generated R wrappers for estimators and transformers. To
  import them into R, you can use devtools to import from the uploaded
  zip file. Tests and sample notebooks to come.

• A new RenameColumn transformer for renaming columns within a
  pipeline.

New notebooks:

• Notebook 104: An experiment to demonstrate regression models to
  predict automobile prices. This notebook demonstrates the use of
  Pipeline stages, CleanMissingData, and
  ComputePerInstanceStatistics.

• Notebook 105: Demonstrates DataConversion to make some columns Categorical.

• There us a 401 notebook in notebooks/gpu which demonstrates CNTK
  training when using a GPU VM. (It is not shown with the rest of the
  notebooks yet.)

Updates:

• Updated to use CNTK 2.2. Note that this version of CNTK depends on
  libpng12 and libjasper1 -- which are included in our docker images.
  (This should get resolved in the upcoming CNTK 2.3 release.)

Improvements:

• Local builds will always use a "0.0" version instead of a version
  based on the git repository. This should simplify the build process
  for developers and avoid hard-to-resolve update issues.

• The TextPreprocessor transformer can be used to find and replace all
  key value pairs in an input map.

• Fixed a regression in the image reader where zip files with images no
  longer displayed the full path to the image inside a zip file.

• Additional minor bug and stability fixes.

v0.9

New functionality:

• Refactor ImageReader and BinaryFileReader to support streaming
  images, including a Python API. Also improved performance of the
  readers. Check the 302 notebook for usage example.

• Add ClassBalancer estimator for improving classification performance
  on highly imbalanced datasets.

• Create an infrastructure for automated fuzzing, serialization, and
  python wrapper tests.

• Added a DropColumns pipeline stage.

New notebooks:

• 305: A Flowers sample notebook demonstrating deep transfer learning
  with ImageFeaturizer.

Updates:

• Our main build is now based on Spark 2.2.

Improvements:

• Enable streaming through the EnsembleByKey transformer.

• ImageReader, HDFS issue, etc.

v0.8

New functionality:

• We are now uploading MMLSpark as a Azure/mmlspark spark package.
  Use --packages Azure:mmlspark:0.8 with the Spark command-line tools.

• Add a bi-directional LSTM medical entity extractor to the
  ModelDownloader, and new jupyter notebook for medical entity
  extraction using NLTK, PubMed Word embeddings, and the Bi-LSTM.

• Add ImageSetAugmenter for easy dataset augmentation within image
  processing pipelines.

Improvements:

• Optimize the performance of CNTKModel. It now broadcasts a loaded
  model to workers and shares model weights between partitions on the
  same worker. Minibatch padding (an internal workaround of a CNTK bug)
  is now no longer used, eliminating excess computations when there is a
  mismatch between the partition size and minibatch size.

• Bugfix: CNTKModel can work with models with unnamed outputs.

Docker image improvements:

• Environment variables are now part of the docker image (in addition to
  being set in bash).

• New docker images:

  • microsoft/mmlspark:latest: plain image, as always,
  • microsoft/mmlspark:gpu: GPU variant based on an nvidia/cuda image.
  • microsoft/mmlspark:plus and microsoft/mmlspark:plus-gpu: these
    images contain additional packages for internal use; they will
    probably be based on an older Conda version too in future releases.

Updates:

• The Conda environment now includes NLTK.

• Updated Java and SBT versions.

v0.7

New functionality:

• New transforms: EnsembleByKey, Cacher Timer; see the documentation.

Updates:

• Miniconda version 4.3.21, including Python 3.6.

• CNTK version 2.1, using Maven Central.

• Use OpenCV from the OpenPnP project from Maven Central.

Improvements:

• Spark's binaryFiles function had a regression in version 2.1 from
  version 2.0 which would lead to performance issues; work around that
  for now. Data frame operations after a use of BinaryFileReader (eg,
  reading images) are significantly faster with this.

• The Spark installation is now patched with hadoop-azure and
  azure-storage.

• Includes additional bug fixes and improvements.