Python API 0.13.0

0.13.0 - 2025-07-15

API Changes

• For Random Forest models, .out_of_bag_evaluations() now returns a
  TrainingLogs object. The content is identical to the object previously
  returned, but the number_of_trees property has been renamed to
  iteration for consistency with Gradient Boosted Trees Training Logs.
• mode="tf" is now the default on model.to_tensorflow_saved_model(). The
  previous default is still available by setting mode="keras".
• model.label() returns None for models trained without a label.
• Remove deprecated evaluation_task argument for model.evaluate(). Use
  task instead.

Feature

• Add standalone C++ export with model.to_standalone_cc(). Standalone models
  are super flexible, fast and memory-efficient. They only depend on the C++
  standard library.
• Add model.training_logs() method to return the training logs of the model.
• Expose Mean Average Precision for Ranking tasks.
• Add hyperparameters
  numerical_vector_sequence_enable_closer_than_conditions and
  numerical_vector_sequence_enable_projected_more_than_conditions.
• Clear error messages when attempting to evaluate models without label.
• Faster training with sparse oblique splits for datasets with many numerical
  features
• Many documentation improvements.
• Increase default number of threads to 256 or number of CPU cores.
• Enable cross-validation for hyperparameter tuning.
• Add thresholds to classification plots.
• Explicitly disable custom losses for hyperparameter tuning.
• Disable parallel evaluation for cross-validation custom losses.

Fix

• Distributed Training: recvmsg: Connection reset to isTransientError.
• Enable SHAP values when training with BEST_FIRST_GLOBAL.
• Predictions with cross-entropy LambdaMART no longer need the slow engine.
• Disable the generic engine for oblique splits without global imputation.
  This may fix a very rare bug in the way predictions are computed.

Release music

Sinfonie Nr. 4 in A-Dur, op. 90. Felix Mendelssohn

Python API 0.12.0

0.12.0 - 2025-05-20

Feature

• Enable support for Python 3.13.
• Add custom fields to model metadata.
• Add SHAP value variable importances with model.analyze().
• Add SHAP values for a dataset with model.predict_shap().
• Speed-up (up to 20x) training of models with CATEGORICAL_SET features.
• Add hyper-parameter to limit the mask size for CATEGORICAL_SET features.
• Add hyper-parameter total_max_num_nodes to limit the total number of nodes in a model.
• Add support for na_replacements in python tree editor API.
• Add support for include_all_columns in FeatureSelector.
• Add the ydf.utils.LogBook to manage and track experiments.
• Speed-up training of NDCG ranking model when a single example per group
  is non-zero.
• Speed-up training on datasets with few columns on a computer with a
  large amount of cores.
• Speed-up loss computation multi-threading code.
• Improve distributed training error messages.
• Remove need for label columns for deep learning models.

Fix

• Log message if early stopping is not used.
• Fix force_numerical_discretization errors and documentation.
• Fix handling of empty list columns in the dataset.

Release music

Te Deum in D major, H.146. Marc-Antoine Charpentier

v1.11.0

1.11.0 - 2025-03-12

Features

• Speed-up training of GBT models by ~10%.
• Support for categorical and boolean features in Isolation Forests.
• Rename LAMBDA_MART_NDCG5 to LAMBDA_MART_NDCG. The old name is deprecated but
  can still be used.
• Allow configuring the truncation of NDCG losses.
• Add support for distributed training for ranking gradient boosted tree
  models.
• Add support for NUMERICAL_VECTOR_SEQUENCE features.
• Add support for AVRO data file using the "avro:" prefix.
• Additional hyperparameters restricting weights of sparse oblique splits
  to integers or powers of 2.
• Facilitate training on VertexAI.
• Deprecated SparseObliqueSplit.binary_weights hyperparameter in favor of
  SparseObliqueSplit.weights.
• Add Gzip-compressed BLOB_SEQUENCE serialization
• Enable Poisson loss for model analysis and fast inference.
• Add config for compatibility with protobuf lite.

Fix

• Fix structural variable importances for oblique splits.
• Deflake tests.
• Remove CHECK/FATAL from training code.
• Fix crash in YDF distributed training.

Misc

• Loss options are now defined
  model/gradient_boosted_trees/gradient_boosted_trees.proto (previously
  learner/gradient_boosted_trees/gradient_boosted_trees.proto)
• Remove C++14 support.
• Various documentation improvements.

Python API 0.11.0

0.11.0 - 2025-03-12

Feature

• Expose losses for distributed training.
• Add class_weights parameter to the learners.
• Support for Google Cloud paths for datasets and model IO.
• Add utility to facilitate distributed training on VertexAI.
• Improved support for non-unicode data in categorical features.
• Add support for saving and analyzing deep models.

Fix

• Fix incorrectly transposed confusion table in HTML.
• Various documentation fixes.
• Better requirements management.

Documentation

• Add tutorial for Categorical Set features.
• Add tutorial for training on VertexAI.

Release music

3. Sinfonie in d-Moll. Gustav Mahler

Python API 0.10.0

0.10.0 - 2025-02-11

Feature

• Expose model.save(..., pure_serving=True) for saving a model without debug
  information.
• Allow users to provide a training proto configuration to the learner.
• Add vector sequence feature support.
• Add Variable importances for Isolation Forest Models.
• Add ydf.help.loading_data() to print information about the type of
  supported dataset formats.
• Add experimental Tabular Transformer implementation.
• Add gzipped blob sequence as new model format (still optional).
• Enabled Poisson Loss for model analysis and fast inference.

Fix

• Fix recognition of multidimensional features for Numpy arrays of type
  object.
• Fix subsample count for small number of training examples for Isolation
  Forests.
• Fix NUM_NODES variable importance for oblique splits.

Other

• Updated OSS dependencies of protobuf, grpc and abseil.

Release music

3. Sinfonie in Es-Dur "Sinfonia Eroica", op. 55. Ludwig van Beethoven

Python API 0.9.0

0.9.0 - 2024-12-02

Breaking

• Classification Label classes are now consistently ordered lexicographically
  (for string labels) or increasingly (for integer labels).
• Change typo partial_depepence_plot to partial_dependence_plot on
  model.analyze().

Feature

• Add support for Avro file for path / distributed training with the "avro:"
  prefix.
• Add support for discretized numerical features for in-memory datasets.
• Expose MRR for ranking models.
• Add model.predict_class to generate the most likely predicted class of
  classification models.
• Add support for automatic feature selection with the feature_selector
  learner constructor argument. See the feature selection tutorial for
  more details.
• Add standalone prediction evaluation ydf.evaluate_predictions().
• Add new hyperparameter sparse_oblique_max_num_projections.
• Add options "POWER_OF_TWO" and "INTEGER" for sparse oblique weights.
• Emit proper errors when using lists for multi-dimensional features.

Fix

• Regression and Ranking CEPs scaling corrected.

Release music

The John B. Sails. Traditional

Python API 0.8.0

0.8.0 - 2024-09-23

Breaking

• Disallow positional parameters for the learners, except for label and task.
• Remove the unsupported / invalid hyperparameters from the Isolation Forest
  learner.
• Remove parameters for distributed training and resuming training from
  learners that do not support these capabilities.
• By default, model.analyze for a maximum of 20 seconds (i.e.
  maximum_duration=20 by default).
• Convert boolean values in categorical sets to lowercase, matching the
  treatment of categorical features.

Feature

• Warn if training on a VerticalDataset and fail if attempting to modify the
  columns in a VerticalDataset during training.
• User can override the model's task, label or group during evaluation.
• Add num_examples_per_tree() method to Isolation Forest models.
• Expose the slow engine for debugging predictions and evaluations with
  use_slow_engine=True.
• Speed-up training of GBT models by ~10%.
• Support for categorical and boolean features in Isolation Forests.
• Add ydf.util.read_tf_record and ydf.util.write_tf_record to facilitate
  TF Record datasets usage.
• Rename LAMBDA_MART_NDCG5 to LAMBDA_MART_NDCG. The old name is deprecated but
  can still be used.
• Allow configuring the truncation of NDCG losses.
• Enable multi-threading when using model.predict and model.evaluate.
• Default number of threads of model.analyze is equal to the number of
  cores.
• Add multi-threaded results in model.benchmark.
• Add argument to control the maximum duration of model.analyze.
• Add support for Unicode strings, normalize categorical set values in the
  same way as categorical values, and validate their types.
• Add support for distributed training for ranking gradient boosted tree
  models.

Fix

• Fix labels of regression evaluation plots
• Improved errors if Isolation Forest training fails.

Release music

Perpetuum Mobile "Ein musikalischer Scherz", Op. 257. Johann Strauss (Sohn)

v1.10.0

1.10.0 - 2024-08-21

Features

• Add support for Isolation Forests model.
• The default value of num_candidate_attributes in the CART learner is
  changed from 0 (Random Forest style sampling) to -1 (no sampling). This is
  the generally accepted logic of CART.
• Added support for GCS for file I/O.

Python API 0.7.0

Python API 0.7.0 - 2024-08-21

Feature

• Expose validate_hyperparameters() on the learner.
• Clarify which parameters in the learner are optional.
• Add support in JAX FeatureEncoder for non-string categorical feature values.
• Improve performance of Isolation Forests.
• Models can be serialized/deserialized to/from bytes with model.serialize()
  and ydf.deserialize_model.
• Models can be pickled safely.
• Native support for Xarray as a dataset format for all operations (e.g.,
  training, evaluation, predictions).
• The output of model.to_jax_function can be converted to a TensorFlow Lite
  model.
• Change the default number of examples to scan when training on files to
  determine the semantic and dictionaries of columns from 10k to 100k.
• Various improvements of error messages.
• Evaluation for Anomaly Detection models.
• Oblique splits for Anomaly Detection models.

Fix

• Fix parsing of multidimensional ragged inputs.
• Fix isolation forest hyperparameter defaults.
• Fix bug causing distributed training to fail on a sharded dataset containing
  an empty shard.
• Handle unordered categorical sets in training.
• Fix dataspec ignoring definitions of unrolled columns, such as
  multidimensional categorical integers.
• Fix error when defining categorical sets for non-ragged multidimensional
  inputs.
• MacOS: Fix compatibility with other protobuf-using libraries such as
  Tensorflow.

Release music

Rondo Alla ingharese quasi un capriccio "Die Wut über den verlorenen Groschen",
Op. 129. Ludwig van Beethoven

Python API 0.6.0

Feature

• model.to_jax_function now always outputs a FeatureEncoder to help feeding
  data to the JAX model.
• The default value of num_candidate_attributes in the CART learner is
  changed from 0 (Random Forest style sampling) to -1 (no sampling). This is
  the generally accepted logic of CART.
• model.to_tensorflow_saved_model support preprocessing functions which have
  a different signature than the YDF model.
• Improve error messages when feeding wrong size Numpy arrays.
• Add option for weighted evaluation in model.evaluate.

Fix

• Fix display of confusion matrix with floating point weights.

Known issues

• MacOS build is broken.