In the example below I get different probabilities when I run

probs_test = pipe.predict_proba(x_test)[:, 1]

and when I converted the pipeline to an ONNX object and run

res = sess.run(None, inputs)

Why is this the case? And how can I get the same probabilities from the ONNX object?

Reproducible example:

import pandas as pd
import sklearn.metrics as skmet
import sklearn.linear_model as lm
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.datasets import load_iris
from skl2onnx import to_onnx
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as rt

data = load_iris()
x, y = data.data, data.target

x_train = pd.DataFrame(x).add_prefix("column_")
x_train["target"] = y
x_train["target"] = x_train["target"].replace({2: 0})
x_test = x_train.copy()

target_column = "target"
features = ['column_0', 'column_1', 'column_2', 'column_3']

model = lm.LogisticRegression(
    C=0.08,
    fit_intercept=True,
    solver="liblinear",
    penalty="l1",
    random_state=2,
    max_iter=1000000,
    verbose=0,
)

column_transformer = ColumnTransformer(
    [
        ("num", "passthrough", features),
    ],
)

pipe = Pipeline(
    [
        ("preprocessing", column_transformer),
        ("logistic_regression", model),
    ],
)

y_train = x_train[target_column].to_numpy()
y_test = x_test[target_column].to_numpy()

pipe = pipe.fit(X=x_train, y=y_train)

probs_test = pipe.predict_proba(x_test)[:, 1]
auc_test = skmet.roc_auc_score(y_true=y_test, y_score=probs_test)
print(f"auc_test={auc_test}")

initial_types = [
    (feature, FloatTensorType([None, 1])) for feature in features
]

onx = to_onnx(pipe, initial_types=initial_types)

sess = rt.InferenceSession(
    onx.SerializeToString(),
    providers=["CPUExecutionProvider"],
)

inputs = {
    feature: x_test[[feature]].to_numpy().astype("float32")
    for feature in features
}

res = sess.run(None, inputs)

print(res[1][:2])
print(probs_test[:2])