Skip to content

v0.12

v0.12
@elibarzilay elibarzilay tagged this 17 Apr 02:46 
New functionality:

* MMLSpark Serving: a RESTful computation engine built on Spark
  streaming.  See `docs/mmlspark-serving.md` for details.

* New LightGBM Binary Classification and Regression learners and
  infrastructure with a Python notebook for examples.

* MMLSpark Clients: a general-purpose, distributed, and fault tolerant
  HTTP Library usable from Spark, Pyspark, and SparklyR.  See
  `docs/http.md`.

* Add `MinibatchTransformer` and `FlattenBatch` to enable efficient,
  buffered, minibatch processing in Spark.

* Added Python wrappers and a notebook example for the
  `TuneHyperparameters` module, demonstrating parallel distributed
  hyperparameter tuning through randomized grid search.

* Add a `MultiNGram` transformer for efficiently computing variable
  length n-grams.

* Added DataType parameter for building models that are parameterized by
  Spark data types.

Updates:

* Update per-instance statistics module so it works for any Spark ML
  estimators.

* Update CNTK to version 2.4.

* Updated Spark to version v2.2.1 (the following release is likely to be
  based on Spark 2.3).

* Also updated SBT and JVM.

* Refactored readers directory into `io` directory

Improvements:

* Fix the Conda installation in our Docker image, resolving issues with
  importing `numpy`.

* Fix a regression in R wrappers with the latest SparklyR version.

* Additional bugfixes, stability, and notebook improvements.
Assets 2
Loading