contributions mismatch for nominal features

[lboussengui]

• ebm2onnx version: 3.1.1
• onnxruntime : 1.16.1
• interpret : 0.4.2
• Python version: 3.10.8
• Operating System: MacOS

Description

I trained an EBM classification model. This model was initially saved in pickle format.

I used ebm2onnx as shown below to convert my model to the .onnx format.

I noticed that the contribution to the prediction for a test case is different for nominal type features when passing in onnx format; the contributions are set to zero.

Do you have an explanation for this ?

What I Did

import ebm2onnx
import pickle
import onnxruntime as rt

# load first EBM 
with open(f'{MODEL_PATH}ebm_first.pkl', 'rb') as f:
    ebm_first  = pickle.load(f)

# load dtypes saved during model training 
with open(f'{MODEL_PATH}training_dtypes_for_onnx.pkl', 'rb') as f:
    training_dtypes_for_onnx  = pickle.load(f)

# transform ebm to onnx 
onnx_model = ebm2onnx.to_onnx(
    model=ebm_first,
    predict_proba=True,  # Generate a dedicated output for probabilities
    explain=True,  # Generate a dedicated output for local explanations
    dtype=training_dtypes_for_onnx,
    name='DEFAULT',
)

Here are the result of local explanation with EBM pickle model for one example :

pred_pkl = ebm_first.explain_local(X_test, y_test)
pred_pkl.data(0)['scores']

result is

array([ 0.027,  0.416, -0.158,  0.388,  0.043,  0.   , -0.196,  0.051,
       -0.201, -0.032,  0.176,  0.151,  0.   ,  0.216,  0.2  ,  0.376,
        0.05 ,  0.022, -0.076,  0.028, -0.26 , -0.043,  0.173,  0.269,
       -0.203, -0.025,  0.037, -0.056,  0.164,  0.296,  0.089,  0.08 ,
        0.1  ,  0.098, -0.018, -0.002, -0.001, -0.001, -0.003, -0.002])

After transforming ebm_first to onnx_model i did the following to imitate inference in production:

onnx_model.ir_version = 9
ebm_onnx = rt.InferenceSession(onnx_model.SerializeToString())
pred_onnx = ebm_onnx.run(None, X_test.to_dict("list"))

# contributions of pred_onnx 
pred_onnx[2][0][:, 0]

result is

array([ 0.027,  0.416, -0.158,  0.388,  0.   ,  0.   , -0.196,  0.051,
       -0.201, -0.032,  0.176,  0.151,  0.   ,  0.216,  0.2  ,  0.376,
        0.05 ,  0.022, -0.076,  0.028, -0.26 , -0.043,  0.173,  0.269,
       -0.203, -0.025,  0.037, -0.056,  0.164,  0.296,  0.089,  0.08 ,
        0.1  ,  0.098, -0.018, -0.002, -0.001, -0.001, -0.003, -0.002],
      dtype=float32)

The two arrays are not equal in index 4 and 5; the only nominal features of the dataset.

[MainRo]

Is it possible for you to publish here a model and sample utterance that reproduces the issue?
In the meantime I will look at reproducing this in a unit test

[MainRo]

Could you confirm that the nominal features are of type boolean?
If this is the case, then can you try to explicitly convert them to 0/1 before calling ebm_first.explain_local:

X_test['feature'] = np.where(X_test['feature'] == False, 0, 1)

I suspect an issue in the interpret explain_local implementation. It looks like the boolean features are not correctly mapped, and have scores of 0.0.

[MainRo]

ok forget my last comment, the problem is that the conversion to onnx mutates the ebm model object.
if you call ebm_first.explain_local before converting to onnx you will have the same values.

obviously, this is not a normal behavior of the converter. I will fix this.