3.3.0 Release

What's Changed

• Add instruction to build sdist in release script by @hcho3 in #462
• [CI] Use Python 3.8 by @hcho3 in #476
• Set type fields in ConcatenateModelObjects() by @hcho3 in #475
• Add API function to serialize to strings by @hcho3 in #480

Full Changelog: 3.2.0...3.3.0

3.2.0 Release

The 3.2 release implements three new features:

• predict_leaf: predict leaf node IDs
• predict_per_tree: obtain prediction per tree
• import_from_json: build Treelite model using a JSON string

What's Changed

• Add allow_unknown_field flag to XGBoost JSON parser by @hcho3 in #423
• Bump development version by @hcho3 in #437
• [CI] Fix test timeout for test_extra_field_in_xgb_json by @hcho3 in #436
• [CI] Require C++17 by @hcho3 in #440
• Implement predict_leaf and predict_per_tree by @hcho3 in #442
• Improve pytest coverage for GTIL by @hcho3 in #447
• Add warning when loading from newer minor version by @hcho3 in #450
• Fix docstring for predict_leaf by @hcho3 in #451
• Add ability to import model from JSON by @hcho3 in #448
• Initial support for HistGradientBoostingClassifier / HistGradientBoostingRegressor by @hcho3 in #444
• Fix up documentation by @hcho3 in #454
• Fix up documentation, round 2 by @hcho3 in #455
• Fix up documentation, round 3 by @hcho3 in #456
• Fix link for Doxygen by @hcho3 in #457
• Add tutorial for JSON import by @hcho3 in #459

Full Changelog: 3.1.0...3.2.0

3.1.0 Release

The 3.1 release implements a new feature to concatenate multiple models into one (#412). Also it's now compatible with scikit-learn 1.2.0 (#425).

What's Changed

• Fix OpenMP build with libomp 15+ by @hcho3 in #410
• Implement model concatenation by @hcho3 in #412
• [CI] Trigger CI daily by @hcho3 in #418
• [CI] Various CI improvements by @hcho3 in #419
• [CI] Fix the script that names wheels by @hcho3 in #421
• [CI] Add convenience script to automate PyPI release by @hcho3 in #422
• Use n_features_in_ instead of deprecated n_features_ attribute by @oliverholworthy in #425
• [CI] Replace load_boston with a randomly generated regression data by @hcho3 in #428
• Fix undefined behavior in tree's Clone() by @hcho3 in #430

New Contributors

• @oliverholworthy made their first contribution in #425

Full Changelog: 3.0.1...3.1.0

3.0.1 Patch Release

This patch release makes the following changes:

• Support XGBoost 1.7.0, by handling the new boost_from_average field in the XGBoost JSON format.
• Support building with libomp 15+ on MacOS

Full Changelog: 3.0.0...3.0.1

3.0.0 Release

The new release is 3.0 due to a major breaking change:

• [Breaking] New serializer with improved compatibility by @hcho3 in #401

Now it is possible to exchange serialized tree models between two different versions of Treelite, with certain restrictions. See the expected compatibility matrix here.

In addition, 3.0.0 incorporates the following changes:

What's Changed

• docs: add missing supported sklearn models by @tczhao in #385
• Replace all generic exceptions with treelite::Error by @hcho3 in #389
• Use development version in mainline branch by @hcho3 in #397
• [CI] Clean up pipeline definition by @hcho3 in #402
• [CI] Migrate MacOS tests to Azure pipelines by @hcho3 in #403
• Various improvements for Windows by @hcho3 in #404
• Various doc improvements by @hcho3 in #405
• A gtest to simulate forward compatibility in serializer by @hcho3 in #406

New Contributors

• @tczhao made their first contribution in #385

Full Changelog: 2.4.0...3.0.0

2.4.0 Release

The new release will incorporate the following changes:

• Test GTIL with IsolationForest (#370)
• [CI] Use supported versions of Windows (#381)
• [XGBoost 1.6] Handle pseudo-Huber loss with slope (#380)
• Support XGBoost classifiers with num_class > 2, num_parallel_tree > 1 (#383)

2.3.0 Release

The new release will incorporate the following changes:

• GTIL optimization using multiple CPU threads (#353, #355, #357, #358, #362, #367)
• Ensure compatibility with XGBoost 1.6+ in from_xgboost() method (#365)
• Remove all uses of distutils (#366)
• Windows compatibility fix (#368)

Patch release 2.2.2

This patch release incorporates two patches and is otherwise identical to 2.2.1:

• Ensure that c_api.h can be built with a C compiler (#348)
• Revert "Add MaxCategory() (#334)" (#349)

Patch release 2.2.1

This patch release incorporates a hotfix and is otherwise identical to 2.2.0:

• Fix PyBuffer serializer (#340)

2.2.0 Release

The new release will incorporate the following changes:

• Update deploy tutorial (#320)
• Add option to enable sanitizers in gtest (#317)
• Handle empty array inputs carefully, to avoid undefined behavior (#314)
• Support for isolation forests (#322, #327)
• When version mismatch happens, show versions (#325)
• Support importing directly from LightGBM model object (#332)
• Add weighted_n_node_samples field in sklearn importer (#330)
• Fix from_xgboost() for models with categorical splits (#333)
• Add MaxCategory() (#334)
• Fix handling of NaNs in categorial splits of LightGBM models (#304)
• Remove undefined behavior when predicting with invalid category value (#335)
• Add a test to prepare for integer default_left (#337)