Convert a pipeline with ColumnTransformer

scikit-learn recently shipped ColumnTransformer which lets the user define complex pipeline where each column may be preprocessed with a different transformer. sklearn-onnx still works in this case as shown in Section Convert complex pipelines.

Create and train a complex pipeline

We reuse the pipeline implemented in example Column Transformer with Mixed Types. There is one change because ONNX-ML Imputer does not handle string type. This cannot be part of the final ONNX pipeline and must be removed. Look for comment starting with --- below.

import pprint
import os
import onnx
import sklearn
import matplotlib.pyplot as plt
from onnx.tools.net_drawer import GetPydotGraph, GetOpNodeProducer
import numpy
import onnxruntime as rt
from skl2onnx import convert_sklearn, __version__
from skl2onnx.common.data_types import FloatTensorType, StringTensorType
from skl2onnx.common.data_types import Int64TensorType
import pandas as pd
import numpy as np
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

titanic_url = ('https://raw.githubusercontent.com/amueller/'
               'scipy-2017-sklearn/091d371/notebooks/datasets/titanic3.csv')
data = pd.read_csv(titanic_url)
X = data.drop('survived', axis=1)
y = data['survived']

# SimpleImputer on string is not available for
# string in ONNX-ML specifications.
# So we do it beforehand.
for cat in ['embarked', 'sex', 'pclass']:
    X[cat].fillna('missing', inplace=True)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

numeric_features = ['age', 'fare']
numeric_transformer = Pipeline(steps=[
    ('imputer', SimpleImputer(strategy='median')),
    ('scaler', StandardScaler())])

categorical_features = ['embarked', 'sex', 'pclass']
categorical_transformer = Pipeline(steps=[
    # --- SimpleImputer is not available for strings in ONNX-ML specifications.
    # ('imputer', SimpleImputer(strategy='constant', fill_value='missing')),
    ('onehot', OneHotEncoder(handle_unknown='ignore'))])

preprocessor = ColumnTransformer(
    transformers=[
        ('num', numeric_transformer, numeric_features),
        ('cat', categorical_transformer, categorical_features),
    ])

clf = Pipeline(steps=[('preprocessor', preprocessor),
                      ('classifier', LogisticRegression(solver='lbfgs'))])


clf.fit(X_train, y_train)

Out:

Pipeline(steps=[('preprocessor',
                 ColumnTransformer(transformers=[('num',
                                                  Pipeline(steps=[('imputer',
                                                                   SimpleImputer(strategy='median')),
                                                                  ('scaler',
                                                                   StandardScaler())]),
                                                  ['age', 'fare']),
                                                 ('cat',
                                                  Pipeline(steps=[('onehot',
                                                                   OneHotEncoder(handle_unknown='ignore'))]),
                                                  ['embarked', 'sex',
                                                   'pclass'])])),
                ('classifier', LogisticRegression())])

Define the inputs of the ONNX graph

sklearn-onnx does not know the features used to train the model but it needs to know which feature has which name. We simply reuse the dataframe column definition.

print(X_train.dtypes)

Out:

pclass         int64
name          object
sex           object
age          float64
sibsp          int64
parch          int64
ticket        object
fare         float64
cabin         object
embarked      object
boat          object
body         float64
home.dest     object
dtype: object

After conversion.

def convert_dataframe_schema(df, drop=None):
    inputs = []
    for k, v in zip(df.columns, df.dtypes):
        if drop is not None and k in drop:
            continue
        if v == 'int64':
            t = Int64TensorType([None, 1])
        elif v == 'float64':
            t = FloatTensorType([None, 1])
        else:
            t = StringTensorType([None, 1])
        inputs.append((k, t))
    return inputs


inputs = convert_dataframe_schema(X_train)

pprint.pprint(inputs)

Out:

[('pclass', Int64TensorType(shape=[None, 1])),
 ('name', StringTensorType(shape=[None, 1])),
 ('sex', StringTensorType(shape=[None, 1])),
 ('age', FloatTensorType(shape=[None, 1])),
 ('sibsp', Int64TensorType(shape=[None, 1])),
 ('parch', Int64TensorType(shape=[None, 1])),
 ('ticket', StringTensorType(shape=[None, 1])),
 ('fare', FloatTensorType(shape=[None, 1])),
 ('cabin', StringTensorType(shape=[None, 1])),
 ('embarked', StringTensorType(shape=[None, 1])),
 ('boat', StringTensorType(shape=[None, 1])),
 ('body', FloatTensorType(shape=[None, 1])),
 ('home.dest', StringTensorType(shape=[None, 1]))]

Merging single column into vectors is not the most efficient way to compute the prediction. It could be done before converting the pipeline into a graph.

Convert the pipeline into ONNX

try:
    convert_sklearn(clf, 'pipeline_titanic', inputs, target_opset=12)
except Exception as e:
    print(e)

Out:

Isolated variables exist: {'parch', 'boat', 'sibsp', 'cabin', 'ticket', 'home_dest', 'body', 'name'}

Predictions are more efficient if the graph is small. That’s why the converter checks that there is no unused input. They need to be removed from the graph inputs.

to_drop = {'parch', 'sibsp', 'cabin', 'ticket',
           'name', 'body', 'home.dest', 'boat'}
inputs = convert_dataframe_schema(X_train, to_drop)
try:
    model_onnx = convert_sklearn(clf, 'pipeline_titanic', inputs,
                                 target_opset=12)
except Exception as e:
    print(e)

scikit-learn does implicit conversions when it can. sklearn-onnx does not. The ONNX version of OneHotEncoder must be applied on columns of the same type.

X_train['pclass'] = X_train['pclass'].astype(str)
X_test['pclass'] = X_test['pclass'].astype(str)
inputs = convert_dataframe_schema(X_train, to_drop)

model_onnx = convert_sklearn(clf, 'pipeline_titanic', inputs,
                             target_opset=12)


# And save.
with open("pipeline_titanic.onnx", "wb") as f:
    f.write(model_onnx.SerializeToString())

Out:

/home/dupre/github_xadupre/sklearn-onnx/docs/examples/plot_complex_pipeline.py:157: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  X_train['pclass'] = X_train['pclass'].astype(str)
/home/dupre/github_xadupre/sklearn-onnx/docs/examples/plot_complex_pipeline.py:158: SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  X_test['pclass'] = X_test['pclass'].astype(str)

Compare the predictions

Final step, we need to ensure the converted model produces the same predictions, labels and probabilities. Let’s start with scikit-learn.

print("predict", clf.predict(X_test[:5]))
print("predict_proba", clf.predict_proba(X_test[:1]))

Out:

predict [0 0 1 0 0]
predict_proba [[0.65654207 0.34345793]]

Predictions with onnxruntime. We need to remove the dropped columns and to change the double vectors into float vectors as onnxruntime does not support double floats. onnxruntime does not accept dataframe. inputs must be given as a list of dictionary. Last detail, every column was described not really as a vector but as a matrix of one column which explains the last line with the reshape.

X_test2 = X_test.drop(to_drop, axis=1)
inputs = {c: X_test2[c].values for c in X_test2.columns}
for c in numeric_features:
    inputs[c] = inputs[c].astype(np.float32)
for k in inputs:
    inputs[k] = inputs[k].reshape((inputs[k].shape[0], 1))

We are ready to run onnxruntime.

sess = rt.InferenceSession("pipeline_titanic.onnx")
pred_onx = sess.run(None, inputs)
print("predict", pred_onx[0][:5])
print("predict_proba", pred_onx[1][:1])

Out:

predict [1 0 1 0 0]
predict_proba [{0: 0.4371914267539978, 1: 0.5628085732460022}]

Display the ONNX graph

Finally, let’s see the graph converted with sklearn-onnx.

pydot_graph = GetPydotGraph(model_onnx.graph, name=model_onnx.graph.name,
                            rankdir="TB",
                            node_producer=GetOpNodeProducer("docstring",
                                                            color="yellow",
                                                            fillcolor="yellow",
                                                            style="filled"))
pydot_graph.write_dot("pipeline_titanic.dot")

os.system('dot -O -Gdpi=300 -Tpng pipeline_titanic.dot')

image = plt.imread("pipeline_titanic.dot.png")
fig, ax = plt.subplots(figsize=(40, 20))
ax.imshow(image)
ax.axis('off')

Out:

(-0.5, 5708.5, 7574.5, -0.5)

Versions used for this example

print("numpy:", numpy.__version__)
print("scikit-learn:", sklearn.__version__)
print("onnx: ", onnx.__version__)
print("onnxruntime: ", rt.__version__)
print("skl2onnx: ", __version__)

Out:

numpy: 1.19.1
scikit-learn: 0.23.2
onnx:  1.7.0
onnxruntime:  1.4.0
skl2onnx:  1.7.1