OmniXAI is a comprehensive Python library for explainable AI (XAI), designed to simplify the process of understanding machine learning model decisions across various data types and model architectures. It targets data scientists, ML researchers, and practitioners by offering a unified interface for a wide array of explanation methods, including feature attribution, counterfactual explanations, and feature visualization, along with a GUI dashboard for interactive exploration.

How It Works

OmniXAI provides specialized explainer classes (TabularExplainer, VisionExplainer, NLPExplainer, TimeseriesExplainer) that act as factories for multiple explanation techniques. Users define the ML model, data, and desired explanation methods. The library handles the integration and execution of techniques like SHAP, LIME, Grad-CAM, and even an experimental GPT explainer for natural language explanations, abstracting away complex implementation details. This unified approach simplifies the workflow for generating and comparing diverse explanations.

Quick Start & Requirements

• Install: pip install omnixai (or pip install omnixai[all] for full dependencies).
• Prerequisites: Python, Scikit-learn, PyTorch/TensorFlow (for vision/NLP), XGBoost (for examples). GPU recommended for vision tasks.
• Documentation: https://opensource.salesforce.com/OmniXAI/latest/index.html
• Tutorials: https://opensource.salesforce.com/OmniXAI/latest/tutorials.html

Highlighted Details

• Supports tabular, image, text, and time-series data.
• Integrates traditional ML (Scikit-learn) and deep learning (PyTorch/TensorFlow) models.
• Features a GUI dashboard for visualizing explanations and model performance.
• Includes an experimental GPT explainer for generating natural language explanations from model outputs.

Maintenance & Community

• Developed by Salesforce.
• Contact: omnixai@salesforce.com.
• Contribution guide available: https://opensource.salesforce.com/OmniXAI/latest/omnixai.html#how-to-contribute

Licensing & Compatibility

• License: BSD 3-Clause License.
• Permissive license suitable for commercial use and integration into closed-source applications.

Limitations & Caveats

The GPT explainer is experimental and may not always produce accurate explanations. Some advanced features or specific model integrations might require custom preprocessing or postprocessing logic.