https://github.com/dnv-opensource/gp-diagnostics
Tools for diagnosing Gaussian Process models
https://github.com/dnv-opensource/gp-diagnostics
Last synced: 5 months ago
JSON representation
Tools for diagnosing Gaussian Process models
- Host: GitHub
- URL: https://github.com/dnv-opensource/gp-diagnostics
- Owner: dnv-opensource
- License: mit
- Created: 2025-01-23T13:50:44.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2025-02-11T11:28:21.000Z (over 1 year ago)
- Last Synced: 2025-08-20T06:52:37.386Z (11 months ago)
- Language: Python
- Homepage:
- Size: 2.44 MB
- Stars: 0
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# gp-diagnostics
**gp-diagnostics** is a Python library for diagnosing Gaussian Process (GP) models. It provides tools for advanced
cross-validation, residual analysis, and visualization to support the evaluation and interpretation of GP regression
models.
---
## Key Features
- **Cross-Validation**
- **Leave-One-Out (LOO)** or **Multifold** CV for GP regressions with optional fixed observational noise.
- Fast computation of CV residuals (including covariance) based on the approach by
[Ginsbourger and Schaerer (2021)](https://arxiv.org/abs/2101.03108).
- Cholesky-based implementations with fallback checks for matrix definiteness.
- **GP Evaluation Metrics**
- Functions to compute **log marginal likelihood**, **pseudo-likelihood** (for CV), and **mean squared error**.
- Easy integration with existing GP code to measure model performance and residual normality.
- **Diagnostic Plots**
- **Histogram of residuals**, **QQ plots**, **predictions vs. true** values, and more.
- Interactive **Plotly** figures for easy exploration and diagnostics.
---
## Installation
```bash
pip install gp-diagnostics
```
---
## Usage
See [examples/example.ipynb](examples/example.ipynb) and [examples/exact_gp_example.ipynb](examples/exact_gp_example.ipynb) for detailed usage examples.
---
## Core Modules
**1. `cv.py`**
- `loo()`: Leave-One-Out CV residuals and covariance.
- `multifold()`: Multifold CV for grouped data (e.g., multiple time series).
- Both handle GP prior covariance plus optional noise variance.
**2. `metrics.py`**
- `evaluate_GP()`: Calculates log marginal likelihood, pseudo-likelihood, and MSE.
- `log_prob_normal()` and `log_prob_standard_normal()`: Useful log probability helpers.
**3. `plots.py`**
- Plotly-based functions to visualize residuals (histogram, QQ), predictive intervals, etc.
- `qq_residuals()`, `hist_residuals()`, `pred_vs_error()`, and more for quick diagnostics.
**4. `utils`**
- **`checks.py`**: Validations (e.g., numeric array, lower-triangular).
- **`linalg.py`**: Cholesky utilities, triangular solves, matrix inversion.
- **`stats.py`**: QQ data generation, partitioning data folds, etc.
---
## Contributing
Please feel free to open [issues](https://github.com/dnv-opensource/gp-diagnostics/issues) or submit pull requests if
you have ideas for improvements or bug fixes.
For local development (requires [`uv`](https://docs.astral.sh/uv/))
1. Clone this repository.
2. Install dependencies with `uv sync`
3. Implement your changes and add tests.
4. Submit a pull request to `main` when ready.
---
## Testing
All tests live under the `tests/` directory. To run them:
```bash
# With pytest:
pytest
# Or with tox (runs multiple Python versions, if available):
tox
```
---
## Authors
- Christian Agrell ([christian.agrell@dnv.com](mailto\:christian.agrell@dnv.com))
- Magnus Kristiansen ([magnus.kristiansen@dnv.com](mailto\:magnus.kristiansen@dnv.com))
---
## License
[MIT License](LICENSE)
© 2024 [DNV](https://www.dnv.com). See [LICENSE](LICENSE) for details.
---