Boolean data column lead to wrong predictions

[Patrick Schleiter]

• ebm2onnx version: 3.1.1
• Python version: 3.10.13
• Operating System: Windows

Description

I like to convert an ExplainableBoostingClassifier in to an onnx model. The data is containing columns with the dtype bool. Converting to onnx is working, but the resulting onnx model does not produce the same probabilities (not even close).

The workaround I found, was converting the columns of type bool to int. Then the probabilities where almost equal.

Although the workaround works fine for me, I would be thankful if this could be fixed.

What I Did

Here is a minimal example demonstrating the issue:

import pandas as pd
from interpret.glassbox import ExplainableBoostingClassifier
import ebm2onnx
import onnxruntime

# Small dataset
X = pd.DataFrame(
    {
        "feature1": [0,0,1,1] * 8,
        "feature2": [0]*16 + [1]*16,
    }
).astype("bool")    
y = pd.Series( [0]*28 + [1]*4)

# Train model
model = ExplainableBoostingClassifier(
    feature_types=["ordinal", "nominal"],
    interactions=0
)
model.fit(X=X, y=y)

# Compute probability with onnx
proba = model.predict_proba(X)

# Compute probability with onnx
onnx_model = ebm2onnx.to_onnx(
    model,
    ebm2onnx.get_dtype_from_pandas(X),
    predict_proba=True,
)
sess = onnxruntime.InferenceSession(onnx_model.SerializeToString())
onnx_proba = sess.run(
    output_names=["probabilities"],
    input_feed={
        "feature1": X.loc[:, "feature1"].to_numpy(),
        "feature2": X.loc[:, "feature2"].to_numpy(),
    }
)[0]


from numpy.testing import assert_almost_equal

# Test
assert_almost_equal(
    actual=onnx_proba,
    desired=proba,
    decimal=2, 
)

#Example

#onnx_proba[:2]
# array([[0.8545578 , 0.14544217],
#        [0.8545578 , 0.14544217]], dtype=float32)

# proba[:2]
# array([[0.90316828, 0.09683172],
#        [0.90316828, 0.09683172]])

[Romain Picard]

This is supposed to work. Thanks for the code sample, I will investigate it.

[Romain Picard]

v3.1.2 fixes this issue but I reopen it because there is still one conversion failure.
If all values of the feature were True or all values were False, then the conversion fails.

[Patrick Schleiter]

@MainRo Thanks a lot for fixing it.