Python API 0.4.3

Python API - Changelog

Feature

• Add model.to_jax_function() function to convert a YDF model into a JAX
  function that can be combined with other JAX operations.
• Print warnings when categorical features look like numbers.
• Add support for Python 3.12.

Fix

• Fix cross-validation for non-classification learners.
• Fix missing ydf/model/tree/plotter.js
• Solve dependency collision of YDF Proto between PYDF and TF-DF.

Python API 0.4.1

Python API - Changelog

Fix

• Solve dependency collision to YDF between PYDF and TF-DF. If TF-DF is
  installed after PYDF, importing YDF will fails with a has no attribute 'DType' error.
• Allow for training on cached TensorFlow dataset.

Python API 0.4.0

Python API - 0.4.0 - 2024-04-10

Feature

• Multi-dimensional features can be selected / configured with the features=
  training argument.
• Programmatic access to partial dependence plots and variable importances.
• Add model.to_tensorflow_function() function to convert a YDF model into a
  TensorFlow function that can be combined with other TensorFlow operations.
  This function is compatible with Keras 2 and Keras 3.
• Add arguments servo_api=False and feed_example_proto=False for
  model.to_tensorflow_function(mode="tf") to export TensorFlow SavedModel
  following respectively the Servo API and consuming serialized TensorFlow
  Example protos.
• Add pre_processing and post_processing arguments to the
  model.to_tensorflow_function function to pack pre/post processing
  operations in a TensorFlow SavedModel.

Tutorials

• Add tutorial
  Vertex AI with TF Serving
• Add tutorial
  Deep-learning with YDF and TensorFlow

Python API 0.3.0

Python API 0.3.0 - 2024-03-15

Breaking

• Custom losses now require to provide the gradient, instead of the negative
  of the gradient.
• Clarified that YDF may modify numpy arrays returned by a custom loss
  function.

Features

• Allow using Jax for custom loss definitions.
• Allow setting may_trigger_gc on custom losses.
• Add support for MHLD oblique decision trees.
• Expose hyperparameter sparse_oblique_max_num_projections.
• HTML plots for trees with model.plot_tree().
• Fix protobuf version to 4.24.3 to fix some incompatibilities when using
  conda.
• Allow to list compatible engines with model.list_compatible_engines().
• Allow to choose a fast engine with model.force_engine(...).

Fix

• Fix slow engine creation for some combination of oblique splits.
• Improve error message when feeding multi-dimensional labels.

Documentation

• Clarified documentation of hyperparameters for oblique splits.
• Fix plots, typos.

Release music

Doctor Gradus ad Parnassum from "Children's Corner" (L. 113). Claude Debussy

v1.9.0

1.9.0 - 2024-03-12

Feature

• Add "parallel_trials" parameter in the hyper-parameter tuner to control the number of trials to run in parallel.
• Add support for custom losses.

v1.9.0rc0

1.9.0rc0 - 2024-02-26

Feature

• Add "parallel_trials" parameter in the hyper-parameter tuner to control the number of trials to run in parallel.
• Add support for custom losses.

PYDF 0.1.0

0.1.0 - 2024-01-25

Features

• Added model validation evaluation (for GBTs) and OOB evaluation (for RFs).
• Expose winner-takes-all for Random Forests.
• Added model self evaluation.
• Added ydf.from_tensorflow_decision_forests() for importing TF-DF models.
• Allow feeding datasets as sequence of strings.

Fixes

• Fixes a plotting issue for GBTs without validation loss

Release music

Flötenuhren von 1772 und 1793 - Vivace (Hob XIX:13). Joseph Haydn

v1.8.0

1.8.0 - 2023-11-17

Feature

• Support for GBT distances.
• Remove old snapshots automatically for GBT training.

Fix

• Regression with Mean Squared Error loss and Mean Average error loss
  incorrectly clamped the gradients, leading to incorrect predictions.
• Change dependency from boost to boost_math for faster builds.

Note

The commit associated with this release has a typo in its description.

1.7.0 - 2023-10-20

Feature

• Add support for Mean average error (MAE) loss for GBT.
• Add pairwise distance between examples.
• By default, only keep the last three snapshots when training with a working
  cache to be resilient to training interruptions.

New interface

• Check out the new Python interface in port/python! It's still experimental
  but you can already install it from PyPi with pip install ydf.

v1.6.0

Breaking changes

• The dependency to the distributed gradient boosted trees learner is renamed
  from
  //third_party/yggdrasil_decision_forests/learner/distributed_gradient_boosted_trees
  to
  //third_party/yggdrasil_decision_forests/learner/distributed_gradient_boosted_trees:dgbt.
  Note most case, importing the learners with
  //third_party/yggdrasil_decision_forests/learner:all_learners is
  recommended.
• The training configuration must contain a label. A missing label is no
  longer interpreted as the label being the input feature "".

Feature

• Add support for monotonic constraints for gradient boosted trees.
• Improve speed of dataset reading and writing.

Fix

• Proper error message when using distributed training on more than 2^31
  (i.e., ~2B) examples while compiling YDF with 32-bits example index.
• Fix Window compilation with Visual Studio 2019
• Improved error messages for invalid training configuration
• Replaced outdated dependencies

1.5.0

Feature

• Rename experimental_analyze_model_and_dataset to analyze_model_and_dataset
• Add new GBT loss function POISSON for Poisson log likelihood.
• Go API: Categorical string values available for inspection.
• Improved training speed for unit-weight datasets.
• Support for MHLD oblique decision trees.
• Multi-threaded RMSE computation.
• Added Uint8 inference engine.
• Added Multi-task learning where the output of models trained as "secondary"
  are used as input for the models trained as "primary"

Fix

• Go API: fixed typo on OutOfVocabulary constant.
• Error messages for Uplift models.
• Remove owner leakage in the model compiler.
• Fix buggy restriction for SelGB sampling
• Improve documentation.