Tutorial

The tutorial goes from a simple example which converts a pipeline to a more complex example involving operator not actually implemented in ONNX operators or ONNX ML operators.

• The easy case
  • Train and deploy a scikit-learn pipeline
  • Benchmark ONNX conversion
  • What is the opset number?
  • One model, many possible conversions with options
  • Choose appropriate output of a classifier
  • Black list operators when converting
  • Issues when switching to float
  • Intermediate results and investigation
  • Store arrays in one onnx graph
  • Dataframe as an input
  • Modify the ONNX graph
• Using converters from other libraries
  • Convert a pipeline with a LightGBM classifier
  • Convert a pipeline with a LightGBM regressor
  • Convert a pipeline with a XGBoost model
  • Convert a pipeline with a CatBoost classifier
• A custom converter for a custom model
  • Implement a new converter
  • Two ways to implement a converter
  • Issues with FunctionTransformer
  • Implement a new converter using other converters
  • A new converter with options
  • Change the number of outputs by adding a parser
• Advanced scenarios
  • Tricky issue when converting CountVectorizer or TfidfVectorizer
  • TfIdf and sparse matrices
  • Converter for WOE
• Write converters for other libraries
  • Converter for pyod.models.iforest.IForest

The tutorial was tested with following version:

<<<

try:
    import catboost
except Exception as e:
    print("Unable to import catboost due to", e)
    catboost = None
import numpy
import scipy
import sklearn
import lightgbm
import onnx
import onnxmltools
import onnxruntime
import xgboost
import skl2onnx

mods = [
    numpy,
    scipy,
    sklearn,
    lightgbm,
    xgboost,
    catboost,
    onnx,
    onnxmltools,
    onnxruntime,
    skl2onnx,
]
mods = [(m.__name__, m.__version__) for m in mods if m is not None]
mx = max(len(_[0]) for _ in mods) + 1
for name, vers in sorted(mods):
    print("%s%s%s" % (name, " " * (mx - len(name)), vers))

>>>

    Unable to import catboost due to numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject
    lightgbm    4.5.0
    numpy       2.2.0
    onnx        1.18.0
    onnxmltools 1.13.0
    onnxruntime 1.21.0+cu126
    scipy       1.14.1
    skl2onnx    1.18.0
    sklearn     1.6.0
    xgboost     2.1.3