Treelite Release Notes 
======================

## 0.32 (04/20/2018)
Usability improvements
* Add `__version__` attribute (#17)
* Shape results appropriately for multiclass classification (#16)
* Add default LOG handler, to make C API easier to use (#19)
* Allow batches smaller than `nthread`

Bug fixes
* Python 2.7 compatibility fix (#14)
* handle categorical splits with empty `left_categories`
* Fix `create_shared()` to be compatible with shells `zsh` and `fish` (#18)

New feature
* (Experimental) Export model as XGBoost model file, as per request in #14.

## 0.31 (03/30/2018)
* Implement a custom thread pool to improve prediction performance
  - Amortize thread launch cost
  - Pin down threads to physical cores and prevent migration
* Accommodate infinite thresholds in LightGBM models

## 0.30 (03/23/2018)
**Major re-factoring of the backend code**
* New code generator, based on Abstract Syntax Trees (AST)
  * Build AST first
  * Iteratively transform AST
  * Finally convert AST into C/JAVA program
* Improved abstraction and modularity (no more spaghetti code!)
  * It's now much easier to "edit" tree prediction logic: you'd just need to
    operate on a tree graph. Re-factoring was a worthwhile investment for
    future research.
    For instance, we could create a new node type in AST representing an AVX
    operation. A group of conditional nodes can be converted into the AVX node
    with equivalent semantics.
* Support for Java: see example at
  https://github.com/hcho3/treelite-java-example

**BREAKING CHANGES**
* Prediction library generated by previous versions are NOT compatible with
  the current version. You will have to run `export_lib()` or `compile()`
  again.

## 0.21 (03/08/2018)
New features:
* Now categorical features with more than 64 categories are supported.
* It is now possible to specify variants of GCC and Clang for `export_lib`;
  `gcc-5`, `gcc-7`, `arm-linux-gnueabi-gcc`, and so forth.

Bug fixes:
* Fix segmentation fault when `parallel_comp` is set to a high value

## 0.2 (02/12/2018)
A few minor fixes:
* Disable Link-Time Optimization (LTO) by default, to decrease compilation time
* Use /bin/sh when the environment variable SHELL is not set
* Enable relative path in predictor constructor
* Increase precision for floating-point values

## 0.1rc1 (11/15/2017)

* Initial release