cuTAGI is an open-source Bayesian neural networks library that is based on the Tractable Approximate Gaussian Inference (TAGI) theory. It supports various neural network architectures such as fully-connected, convolutional, and transpose convolutional layers, as well as skip connections, pooling and normalization layers. cuTAGI is capable of performing different tasks such as supervised, unsupervised, and reinforcement learning. This library has a python API called pyTAGI that allows users to easily use the C++ and CUDA libraries.
To get started with using our library, check out our:
- installation guide for Windows, MacOS, and Linux (CPU + GPU).
- quick tutorial for a 1D toy problem.
In this section, you will find a series of examples that you can use as a starting point for each of the available architecture.
Check out our API reference for a complete list of the functions and classes in our library.
pyTAGI already includes a set of modules that allow users to make their own models. Check out our modules reference for a list of classes and functions.
We welcome contributions from the community by 1) forking the project, 2) Create a feature branch, and 3) Commit your changes.
If you run into any issues or have any questions, please open an issue or contact us at [email protected] or [email protected].
@misc{cutagi2022,
Author = {Luong-Ha Nguyen and James-A. Goulet},
Title = {cu{TAGI}: a {CUDA} library for {B}ayesian neural networks with Tractable Approximate {G}aussian Inference},
Year = {2022},
journal = {GitHub repository},
howpublished = {https://github.com/lhnguyen102/cuTAGI}
}
- Tractable approximate Gaussian inference for Bayesian neural networks (James-A. Goulet, Luong-Ha Nguyen, and Said Amiri. JMLR, 2021)
- Analytically tractable hidden-states inference in Bayesian neural networks (Luong-Ha Nguyen and James-A. Goulet. JMLR, 2022)
- Analytically tractable inference in deep neural networks (Luong-Ha Nguyen and James-A. Goulet. ArXiv 2021)
- Analytically tractable Bayesian deep Q-Learning (Luong-Ha Nguyen and James-A. Goulet. ArXiv, 2021)
cuTAGI is licensed under the MIT License.
We would like to say a big thank you to Miquel Florensa who wrote and put together this document all by himself, showing hard work and a commitment to sharing clear and detailed information.