trustpy-tools 2.0.9

Trustworthiness metrics and calibration tools for predictive models

TrustPy - Trustworthiness Python

Python package for validating AI/ML model reliability and uncertainty during development and before deployment. Distributed via Conda-Forge and PyPI.

The implementation is flexible and works out-the-box with any AI/ML library.

Installation

Recommended 1: Install via Conda-Forge

The easiest way to install trustpy-tools is via Conda-Forge, which handles all dependencies automatically. Run the following command:

conda install -c conda-forge trustpy-tools

Recommended 2: Install via PyPI (pip install)

If you prefer using pip (PyPI), you can install directly:

pip install trustpy-tools

Alternative: Manual Installation

If you prefer to install the package manually or are not using Conda, you can install the required dependencies and clone the repository.

Install Dependencies

• NumPy: For numerical calculations.
• Matplotlib: For plotting the trust spectrum.
• Scikit-learn: For Kernel Density Estimation (KDE) in trust density estimation.

Install them via conda:

conda install numpy matplotlib scikit-learn

or

Install them via pip:

pip install numpy matplotlib scikit-learn

Clone the Repository

git clone https://github.com/yaniker/TrustPy.git
cd TrustPy

You can verify installation by running:

python -c "from trustpy import NTS, CNTS; print('TrustPy is ready.')"

Example Usage

from trustpy import NTS, CNTS #This is how the package is imported.
import numpy as np

# Example oracle and predictions
oracle = np.array([0, 0, 1, 2, 2, 0, 1])  # True labels
predictions = np.array([
    [0.8, 0.1, 0.1],  # Correct, high confidence
    [1.0, 0.0, 0.0],  # Correct, high confidence
    [0.2, 0.7, 0.1],  # Correct, high confidence
    [0.1, 0.2, 0.7],  # Correct, high confidence
    [0.1, 0.4, 0.5],  # Correct, lower confidence
    [0.1, 0.8, 0.1],  # Incorrect, high confidence
    [0.3, 0.3, 0.4]   # Incorrect, low confidence
]
) #Replace this with your model's predictions (`predictions = model.predict()`)

# FOR NETTRUSTSCORE #
# Initialize with default parameters
nts = NTS(oracle, predictions, show_summary=True, export_summary=True, trust_spectrum=True)
nts_scores_dict = nts.compute() # Computes trustworthiness for each class and overall.

# FOR CONDITIONAL NETTRUSTSCORE #
# Initialize with default parameters
cnts = CNTS(oracle, predictions, show_summary=True, export_summary=True, trust_spectrum=True)
cnts_scores_dict = cnts.compute() # Computes trustworthiness for each class and overall.

# Sets show_summary=True to print the results table.
# Sets export_summary=True to save the results.
# Sets trust_spectrum=True to generate plots like:
# - trustpy/nts/trust_spectrum.png (for NTS)
# - trustpy/cnts/trust_spectrum.png and conditional_trust_densities.png (for CNTS)

Example Plot for Trust Spectrum (trust_spectrum = True)

Example Plot for Conditional Trust Spectrum (trust_spectrum = True)

I shared the codes for the plots Python scripts for plots for users to modify as needed.

Unit Testing

All unit tests were run using pytest with full coverage prior to release to ensure reliability and correctness.

After installation, you can run all tests to verify everything is working:

python -m pytest tests/

Make sure to install pytest first.

pip install pytest

Licence

This project is licensed under the MIT License. See LICENSE for details.

Citations

For scholarly references and the origins of the techniques used in this package, please refer to the CITATION file.