Train, convert and predict with ONNX Runtime

This example demonstrates an end to end scenario starting with the training of a scikit-learn pipeline which takes as inputs not a regular vector but a dictionary { int: float } as its first step is a DictVectorizer.

Train a pipeline

The first step consists in creating a dummy datasets.

import pandas
from sklearn.datasets import make_regression
from sklearn.model_selection import train_test_split

X, y = make_regression(1000, n_targets=1)

X_train, X_test, y_train, y_test = train_test_split(X, y)
X_train_dict = pandas.DataFrame(X_train[:, 1:]).T.to_dict().values()
X_test_dict = pandas.DataFrame(X_test[:, 1:]).T.to_dict().values()

We create a pipeline.

from sklearn.ensemble import GradientBoostingRegressor
from sklearn.feature_extraction import DictVectorizer
from sklearn.pipeline import make_pipeline

pipe = make_pipeline(DictVectorizer(sparse=False), GradientBoostingRegressor())

pipe.fit(X_train_dict, y_train)

Pipeline(steps=[('dictvectorizer', DictVectorizer(sparse=False)),
                ('gradientboostingregressor', GradientBoostingRegressor())])

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

We compute the prediction on the test set and we show the confusion matrix.

from sklearn.metrics import r2_score

pred = pipe.predict(X_test_dict)
print(r2_score(y_test, pred))

0.9213965130497556

Conversion to ONNX format

We use module sklearn-onnx to convert the model into ONNX format.

from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import DictionaryType, FloatTensorType, Int64TensorType, SequenceType

# initial_type = [('float_input', DictionaryType(Int64TensorType([1]), FloatTensorType([])))]
initial_type = [("float_input", DictionaryType(Int64TensorType([1]), FloatTensorType([])))]
onx = convert_sklearn(pipe, initial_types=initial_type, target_opset=17)
with open("pipeline_vectorize.onnx", "wb") as f:
    f.write(onx.SerializeToString())

We load the model with ONNX Runtime and look at its input and output.

import onnxruntime as rt
from onnxruntime.capi.onnxruntime_pybind11_state import InvalidArgument

sess = rt.InferenceSession("pipeline_vectorize.onnx", providers=rt.get_available_providers())

import numpy

inp, out = sess.get_inputs()[0], sess.get_outputs()[0]
print("input name='{}' and shape={} and type={}".format(inp.name, inp.shape, inp.type))
print("output name='{}' and shape={} and type={}".format(out.name, out.shape, out.type))

input name='float_input' and shape=[] and type=map(int64,tensor(float))
output name='variable' and shape=[None, 1] and type=tensor(float)

We compute the predictions. We could do that in one call:

try:
    pred_onx = sess.run([out.name], {inp.name: X_test_dict})[0]
except (RuntimeError, InvalidArgument) as e:
    print(e)

[ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Unexpected input data type. Actual: ((seq(map(int64,tensor(float))))) , expected: ((map(int64,tensor(float))))

But it fails because, in case of a DictVectorizer, ONNX Runtime expects one observation at a time.

pred_onx = [sess.run([out.name], {inp.name: row})[0][0, 0] for row in X_test_dict]

We compare them to the model’s ones.

print(r2_score(pred, pred_onx))

0.9999999999999661

Very similar. ONNX Runtime uses floats instead of doubles, that explains the small discrepencies.