onnxmltools

• Fix early stopping for XGBClassifier and xgboost > 2, #597
• Fix discrepancies with XGBRegressor and xgboost > 2, #670
• Support count:poisson for XGBRegressor, #666
• Supports XGBRFClassifier and XGBRFRegressor, #665
• ONNX_DFS_PATH to be set in the spark config, #653 (by @Ironwood-Cyber)
• Sparkml converter: support type StringType and StringType(), #639
• Add check for base_score in _get_attributes function #637, #626 (by @tolleybot)
• Support for lightgbm >= 4.0, #634

• #608 fix: Spark Imputer conversion with multiple input cols
• #607 fix: getTensorTypeFromSpark fails for Spark 3.3.0+
• #606 Add onnxruntime==1.14.0 to CI
• #605 Replace real images by dummy ones
• #602 Update CI with latext onnxruntime, xgboost
• #606 convert_lightgbm: Add shape to FloatTensor probabilities

• feat: add support for SparkML CountVectorizer conversion #560
• docs: update sparkml doc; cleanups. #559
• fix: 'SparkSession' object has no attribute 'util' #557
• feat: add support for SparkML KMeansModel conversion #556
• fix: SparkML StandardScaler conversion fails when withStd or withMean is set to true #555
• fix: Converter for SparkML VectorAssembler does not support vector inputs correctly #554
• fix: ONNX conversion for Spark OneHotEncoder model #552

• Fix conversion of XGBoost model after being restored #520
• Fix test case condition for onnx=1.11.0 #527
• Update CI for ORT 1.11.0 #539
• Adjust author and email #539

• Replace #507 + fix bug with XGBoost converter when base_score is None #510
• Use assertRegex instead of assertRegexpMatches for Python 3.11 compatibility. #508
• Support for opset 15 and update version to 1.10.0 #505
• add support for quantile objective for LGBM models #503
• Support parameter shape_override and other options for convert_tensorflow #497
• Implement option split to reduce discrepancies for lightgbm regressors #496

• Add requirements.txt to MANIFEST.in #493

LightGBM

• Improves lightgbm conversion speed #491
• Fix discovering classifier objective #480
• Fix missing type in lgbm regressor #488
• Support gamma objective in LGBMRegressor #484
• Allow to add custom post transform functions that are not supported by the ONNX spec yet #463
• Enable option zipmap for LGBM converter #452

XGBoost

• Use all tree when best_ntree_limit is not specified #459
• Fix discrepencies when xgboost trees are empty #447

Keras

• Switch to tf2onnx for tensorflow>=2.0 instead of keras2onnx #492

New features

• New converters for CatBoost #392
• Integration with Hummingbird #404, #418, #427
• Support for opset 13 #437

XGBoost

• Support float type for feature_id #423
• Support unsigned integer as class type #426
• Fix the converter when the parameter best_ntree_limit is used #429
• Support multi:softmax objective #442

CoreML

• Extend CoreML: ReshapeStatic/LoadConstantND #430
• Fix PReLU conversion from CoreML #425

The major update for this release

1. Supports ONNX 1.7
2. Work with the new xgboost version
3. Remove Python 2.x support

Details:
Add the flake8 to be the default code formatter (#401)
Fixes #396, xgboost converter for xgboost >= 1.0.2 (#397)
Support onnx 1.7 in CI build (#398)
fixed the xgboost version (#395)
fix ceiling-mode defaults for pool operators (AvgPool, MaxPool) (#388)
Update documentation, add examples (#385)
Remove support of python 2.7 (#383)
upgrade to 1.7 (#384)
Fix for onnx 1.7 release (#381)
Ping h2o version==3.28.0.3 (#377)
Fix xgboost converter (#373)
xgboost not supporting 1.0 version. (#372)

Known issues:
onnxmltools tf2onnx wrapper can only work with tf2onnx <= 1.5.6.

The major updates of this release.

1. Add a new converter for H2O GBM MOJO
2. add tf2onnx wrapper with a onnxmltools converter style API.
3. Some bug fixing for the existing converters.

Support opset 11

Major updates:

1. Moving onnxconverter-common package from onnxmltools repo
2. Fix CI/nightly build
3. Fix ImageScaler bias for opset 10
4. Fix lightgbm.Booster
5. Fixed XGboost classifier converter output labels
6. Set default_batch_size to 'None'

onnxmltools version 1.5.0 is now available! This version features ONNX Opset 10 support and code coverage.

How do I use the latest onnxmltools package?

pip install onnxmltools --upgrade
python -c "import onnxmltools"

This package includes converters for LightGBM, CoreML, Spark ML, LibSVM, XGBoost, and wrappers for conversion from scikit-learn and Keras.

Highlights since the last release

• Updating onnxmltools package version and requirements to 1.5.0 (#315)
• Opset 10 Updates
  • [Opset 10] Updates for thresholded relu (#308)
  • [Opset 10] Deprecate Upsample, create Resize op (#303)
  • [Opset 10] Pooling operator updates: AveragePool, MaxPool (#296)
  • Added apply_slice function to enable multiple versions of Slice (#291)
• Include code coverage / Improve CI Builds
  • Run code coverage on linux CI (#301)
  • Add support for Py3.7, onnx 1.5, onnxruntime 0.4 (#293)
• Fixing input to CoreML multiply for LeakyReLU (#297)
• Documentation update: Spark ML readme files (#289)

with some new converters, xgboost, libsvm, and pyspark.
refactor onnxmltools structure by splitting keras and sklearn converters out.

1. Multiple opset support: Enables user to generate model based on a specified ONNX opset
2. ONNX opset 8: Adds support for ONNX opset 8 in all converters
3. ONNX model optimization: Reduces redundant operators like transpose and identity in the converted model
4. Convert channel_last to channel_first: Enables converted model to run on WinML even if the original model is channel_last, since the WinML API only supports channel_first models
5. onnxruntime as backend test: Uses onnxruntime as the backend to test the converted model; improves code quality and compatibility between converters and the inference engine
6. Separate LightGBM from scikit-learn converter: Improves code organization