skl2onnx

0.19.0 is incomplete on pypi, please use 1.19.1

• Fix QDA converter crashing on string labels and incorrect shape calculation #1169
• Remove import of split_complex_to_pairs #1182
• Removes dependency on onnxconverter-common #1179
• Implements converter for QuantileTransformer #1164,
• Refactors KNNImputer converter with local function to match scikit-learn's implementation, the code was partially automatically generated from an equivalent implementation in pytorch and exported into ONNX #1167, #1165
• Add support to sklearn TargetEncoder #1137
• Fixes missing WhiteKernel with return_std=True #1163
• Fix empty column selector #1159
• Fix conversion for XGBClassifier and XGBRegressor #1157
• Test SelectKBest + StandardScaler pipeline #1156
• Fix np.NAN into np.nan, #1148

• Add converter for TunedThresholdClassifierCV #1107
• Changes main opset in function to_onnx #1109
• Avoid concatenation if not needed #1110
• Update and Fix documentation #1113
• Support fill_value for SimpleImputer with string data #1123
• Remove unnecessary options for Regressor #1124
• OrdinalEncoder handle encoded_missing_value and unknown_value #1132
• Create output_onnx_single_probability.py #1139 #1141

• Upgrade the maximum supported opset to 21, update requirements to scikit-learn>=1.1, older versions are not tested anymore, #1098
• Support infrequent categories for OneHotEncoder #1029
• Support kernel Matern in Gaussian Process #978
• Fix for multidimensional gaussian process #1097
• Minor fixes to support scikit-learn==1.5.0 #1095
• Fix the conversion of pipeline including pipelines, issue #1069, #1072
• Fix unexpected type for intercept in PoissonRegressor and GammaRegressor #1070
• Add support for scikit-learn 1.4.0, #1058, fixes issues Many examples in the gallery are showing "broken", TFIDF vectorizer target_opset issue, Tfidfvectorizer with sublinear_tf fails, despite opset version set to greater than 11.

• Supports cosine distance (LocalOutlierFactor, ...) #1050,
• Supports multiple columns for OrdinalEncoder #1044 (by @max-509)
• Add an example on how to handle FunctionTransformer #1042, Versions of scikit-learn < 1.0 are not tested any more.
• Supports lists of strings as inputs for FeatureHasher #1036, #1036
• Fix OneHotEncoder when categories indices to drop are not None #1028
• fFx converter for AdaBoost estimators in scikit-learn==1.3.1 #1027
• Add function 'add_onnx_graph' to insert onnx graph coming from other converting,
  libraries within the converter mapped to a custom estimator #1023, #1024
• Add option 'language' to converters of CountVectorizer, TfIdfVectorizer #1020

• Add new matrix for ORT 1.15 testing #992
• Update to sklearn==1.3.0 #1002
• Improves robustness in node ordering #1004
• Fix converter for DecisionTreeClassifier in n_classed==1 #1008
• [Internal] Refactor with black #1009

what's improved:

• #988 Adapting onnx 1.14 release
• #971 Use Furo Theme (doc)
• #972 Update doc to version 1.14.0
• #979 Implement converter for _ConstantPredictor
• #983 Remove unnecessary files
• #986 Swayand/bug fix min max scaler with clip

new converters

• add OneVsOne classifier converter
• add QuadraticDiscriminantAnalysis converter
• add GammaRegressor

improvements

• improve unit test for linear classifier
• move code ovr_decision_function as a pseudo-operator
• shorten CI pipeline
• add Double/Int64 data type for GammaRegressor unittest
• support arrays for length_scale in kernel RBF
• better error messages for TfidfVectorizer
• Implements TraceableCountVecrotizer and TraceableTfidfVectorizer

bug fix

• bug fix for multiple-output not correct in Random Forest

issues closed

• #796
• #797
• #807
• #872
• #874
• #882
• #901
• #912

• Add new Gamma Regressor converter #894
• Update max supported opset, update CI to python 3.10 #892
• Add example about TfldVectorizer
• Improve some other examples and remove unnecessary examples
• Fix documents link missing bug
• Fix some GitHub issues

• LocalOutlierFactor n_neighbors bugfix #821
• MAINT compat link function and loss for sklearn 1.1 #863
• add sgd_oneclass svm converter #860

• Fix WOE converter to adapt OnnxRuntime implementation (https://github.com/onnx/sklearn-onnx/pull/849)
• Refine few documents

• Fix potential bug in IsolationForest and WOETransformer converter due to latest onnx update #835
• LocalOutlierFactor n_neighbors bugfix #836
• Fix converter for SimpleImputer when strategy='MostFrequent' and full_value is None #834
• Enable Double input type for RandomForest and ZipMap #826
• Allow empty array as feature_names_out #832
• Fixes for onnx==1.11.0 #824
• Fix unit test with unfitted models, fix PLSRegression 1.1 #827
• Fix a validation verifying dimension type #823
• Fixes issue #817 and method _propagation_status #820
• Fixes #813, support unsigned class labels in MLPClassifier #818
• Fixes #814, fix conversion of TreeEnsemble due to a change of the loss name in scikit-learn #819

• Use of sigmoid in SGDClassifier converter (#811)
• Reduce import time by 1 second (#801)
• Fix output shape for nearest neighbors regressor models (#799)
• Fix LinearRegressor converter (coef dimension) + fix bug with apply_clip (#795)

• Improve error message in ColumnTransformer parser (#792)
• Add logging in OneHotEncoder (#790)
• Fix converters for MinMaxScaler, GaussianProcessRegressor (Scan) (#791)

• Support non ascii characters in variable names (#784)
• Add option class_output_labels to expose class labels for a classifier (#776)
• Fix issue while converting graph with operator If (#781)
• Implements option zipmap for MultiOutputClassifier (#775)
• Modify tfidf_transformer to enable custom vocabulary and approximate sublinear-tf scaling without sparse containers (#777) (@adam444555)
• Enable RandomForestClassifier in converter for CalibrationClassifierCV (#772)

Features

• Add parameter naming to rename all intermediate results (#756)
• Add a function to update initializers in a ONNX model (#758)
• Support for scikit-learn==1.0 (#752)

Bug fixes

• Handle double with CalibratedClassifierCV (#753)
• Fix a bug with TfIdfConverter and pruned terms (#747)

Converters

• Replace Normalizer, Scaler, LinearRegressor, LinearClassifier by standard ONNX operators (#755)
• Add converters for TweedieRegressor, PoissonRegressor, QuantileRegressor (#750)
• Add converter for RandomTreesEmbedding (#751)

Bug fixes

• Fix LinearSVC, SVC in CalibratedClassifierCV (#742)
• Add function to store constant as initializer into the ONNX graph (#743)
• Fix missing types with two subestimators in a converter (#729)
• Fix conversion when a pipeline is one estimator in parameter estimators (Voting, ...) (#723)
• Fix converter for GridSearchCV (#722)
• Fixes #703, Interpet parameter value in ConstantOfShape (#709)
• Simplify onnx_parser API (#727)
• Refactoring, remove reversed name, simplify topology (#726)

Converters

• Add converter for LocalOutlierFactor (#741)
• Add converter for KernelPCA (#737)
• Add converters for MultiOutputRegressor and MultiOutputClassifier (#714), Use Sequence instead of Tensor for MultiOutputClassifier probabilities (#730), Remove option ZipMap for MultiOutputClassifier (#721)
• Add custom WOETransformer and its conversion to ONNX (#720)
• Support option score_samples with IsolationForest converter (#735)

• #697: support certain special characters in raw name
• #701: more verbosity for long models

• #689: Support onnx==1.9.0
• #694: Support missing string values with SimpleImputer
• #692: Fix Scaler converter, fix OnnxSubEstimator not directly connected to the converter inputs
• #691: Remove trivial Runtime error in SVM converter
• #682: fix type issue in AdaBoostClassifier (apply_clip)
• #680: Support conversion of StandardScaler with Double and norm L2
• #677: Add annotation to converters and shape calculator
• #676: Remove dependency on six

• Fix topological order after the model is converted #665, #667
• Enables opset 14 and enables onnxruntime 1.8.0 on CI, remove CI for onnxruntime<1.5.0 #664
• Add support for Int8, UInt8 #648
• Add option nan to CountVectorizer to avoid discrepencies with the use of sparse matrix #651
• Implements kernel cosine #641
• Removes class OnnxSubOperator, refactoring of class OnnxOperator #631

Framework and bug fixes:

• #229: improves LabelBinarizer
• #327: add option to remove ZipMap operator
• #382: fix final opset in topology
• #445: add missing cast when using operator concat
• #447: add parameters to rename ONNX outputs
• #451: add white list, black list of operators, the converters can use this information to change the way a model is converted into ONNX
• #464: fix issues in ArrayFeatureExtractor, OneHotEncoder
• #466: BaggingClassifier supports zipmap options
• #473, #460, #458, #426, #422: sklearn-onnx works with onnx development version, works with opset 12, scikit-learn 0.23

New converters:

• #318: PLSRegression
• #341: HistGradientBoostingRegressor
• #315, #344, #345, #347, #348, #349, #356, #421: multi-label support
• #358: HistGradientBoostingClassifier
• #362, #363, #364, #366, #369, #371: support decision function
• #377: CalibratedClassifierCV
• #375: CategoricalNB
• #381, #382: StackingClassifier
• #395: KNeighborsTransformer
• #384: StackingRegressor
• #403: PowerTransformer
• #404, #469: KNNImputer
• #419: support more kernels in SVM
• #431: GaussianMixture
• #438: GaussianRandomProjection
• #450: handle OVR decision function (SVM)

• #277 GaussianNB
• #281 BaggingClassifier
• #282 BaggingRegressor
• #283 MultiTaskElasticNet, OrthogonalMatchingPursuit, PassiveAggressiveRegressor
• #284 Clip opset 11
• #290 PassiveAggressiveClassifier
• #294 RANSACRegressor
• #296 RidgeClassifier, RidgeClassifierCV, MultiTaskElasticNetCV and OrthogonalMatchingPursuitCV
• #293 Batch predictions for nearest neighbors
• #315 Multi-label support in MLPClassifier converter

Bug fixes and features included in this release:

• New converters for linear models (#269)
• Enables batch prediction for PolynomialFeatures (#272)
• Add converter for VotingRegressor (#268)
• Support lightgbm and xgboost model in scikit-learn pipelines (#262)
• Fixed column transformer converter to handle remainder=passthrough (#263)
• Fixes #254 + discrepancies of AdaBoostRegressor (#258)
• Fixes VotingClassifier discrepencies + enable batch predictions (#261)
• Modifies GaussianProcessRegressor to use cdist custom operator (#256)
• Fixed gradient boosting binary classification mismatches (#255)
• Support transformer_weights in ColumnTransformer converter (#250)
• Added GridSearchCV converter (#246)

1. Support Gaussian Process
2. Fixing for working with ORT 0.5.
3. A couple of converter bug fixings, including 5493c0a Fix KNeighboursClassifier, KNeighboursRegressor and NearestNeighbours converters (#240) 121b1b4 Fixed NaN probability scores in BernoulliNB converter when alpha=0 (#238) 11ac5ef Fix a couple of mismatches for every tree in scikit-learn. (#237) 517b0bb Added Perceptron classifier converter (#236) d313123 Fixed LabelEncoder converter (#224) f95ce23 Added support for int features in kmeans and mini-batch kmeans converters (#231) 160200e Fixed gradient boosting classifier converter mismatch on binary dataset (#230) d026f89 Fixed label binariser output for binary dataset to align with scikit (#228) a889074 Update nightly build (#227) 33abc90 Fixes #184, #201, fixes GradientBoostingClassifier for multi class (#226) aae4f7a GaussianProcessRegressor with float and double in ONNX models (#220) 2a9399f Fixed Adaboost converter's incorrect results with non-default learning rate (#223) 5a065b7 Handle transformer_weights hyper-parameter in FeatureUnion (#222) 6c24406 Fixed gradient boosting converters for different values of init hyper-parameter (#219) 111e49d Add GaussianMixture (#169) b5333f5 Use n_estimators_ when set in GradientBoostingClassifier (#213) 01b37e3 Fix SVC output (#209) edf8eb4 Add more linear models (#216) 8009430 Install onnxconverter-common from source (#217) 064eb31 Extend coverage for OneHotEncoder (#188) a64cc5b Removed unused methods in _topology (#210) 65951d5 Support feature input types other than float in function transformer (#208) ad1987d Improve OnnxOperator to support more types (#205) a760bf0 Support int features in BernoulliNB, MultinomialNB and SGDClassifier <E2><80><A6> (#185) 6fc5a0e Fix missing node names (#187) 200cb00 Fixed GradientBoostingRegression model conversion failure with init=zero (#164) 1204066 Extend unit test coverage (#180) 4eb0217 Update TfIdf converter to reflect changes in Tokenizer specifications (#178) 4200f11 Test topology pruning (#179) bf5311a Add an example to create a custom converter for a NMF transformer (#167) f427759 Review error message raised in exceptions (#173)

skl2onnx version 1.5.0 is now available! This version features ONNX Opset 10 support, increased test and operator coverage, improved documentation, and a new way to write converters.

How do I install the latest skl2onnx package?

pip install skl2onnx --upgrade

Highlights since the last release

• Update package version and onnxconverter-common dependency to 1.5.0 (#177)
• New way to write converters based on OnnxOperators (#128)
• Clean OnnxOperator before converting into ONNX (#149)
• Increase operator coverage for SciKitLearn
  • Fixed multi-class LR with intercept set to None (#135)
  • Fixes #113, convert OneVsRestClassifier multilabel (#142)
  • Supports passthrough and remainder='passthrough' for ColumnTransformer converter (#152)
  • Support int features in linear and non-linear regressor converter(s) (#154, #157)
  • Fixes #147, fix probabilities when multi_class == 'multinomial' (#150)
• ONNX Opset 10 updates
  • Fixed slice op to align with opset 10 changes (#116)
  • Update CountVectorizer and TfidfVectorizer (#111)
  • Updated TopK node to align with changes in opset 10 (#91)
• Extend documentation, clarify error messages
  • Follow Pep8 conventions (#141, #143, #144, #145)
  • Better exceptions, better error message, extend documentation (#146)
• Increase test coverage
  • Added SGDClassifier converter unit tests (#138)
  • Handle scikit-learn dev version when using StrictVersion in tests (#139)
  • Added unit tests to increase coverage (#140, #170)
  • Add unit tests for investigate.py (#161)
  • Fix issues raised by scikit-learn 0.22. (#175)

the new release for some dependencies.