onnx_basic

ONNX Basic Example

This example demonstrates using SHAP-Go with ONNX models for binary classification.

Prerequisites

1. ONNX Runtime installed:

   # macOS
   brew install onnxruntime
   
   # Linux (Ubuntu/Debian)
   # Download from https://github.com/microsoft/onnxruntime/releases

2. ONNX Model (choose one option):

   Option A: Generate locally (recommended)

   pip install scikit-learn skl2onnx onnx
   python generate_model.py

   Option B: Download from Hugging Face

   ./download_model.sh

   Downloads from Ritual-Net/iris-classification. Note: May require adjusting input/output names in main.go.

Usage

Run the example:

go run main.go

What This Example Shows

• Loading an ONNX model with onnx.NewSession
• Using KernelSHAP for model-agnostic explanations
• Interpreting SHAP values for classification
• Verifying local accuracy

Model Details

The example uses a logistic regression model trained on the Iris dataset for binary classification (versicolor vs. non-versicolor).

Features:

• sepal_length
• sepal_width
• petal_length
• petal_width

Expected Output

See expected_output.md for detailed sample output.

ONNX Model SHAP Explanations
============================

Instance: Setosa sample
Features: [5.1 3.5 1.4 0.2]
Prediction (P(class=1)): 0.0234
Base Value: 0.3333
SHAP Values:
  sepal_length  : value= 5.10, SHAP=-0.0521
  sepal_width   : value= 3.50, SHAP=+0.0123
  petal_length  : value= 1.40, SHAP=-0.1543
  petal_width   : value= 0.20, SHAP=-0.1158
Local accuracy: sum=-0.3099, expected=-0.3099, valid=true