sigmaDL

sigmaDL is an open-source library for deep learning and probabilistic modeling. It provides a tensor API, automatic differentiation, and modular neural network components designed for both standard supervised learning and statistical methods. The project emphasizes readability, interoperability, and extensibility for research and production.

The library targets a Python-centric interface and aims to integrate with NumPy and SciPy. It provides a lightweight core with interchangeable backends for tensor operations, enabling experiments on CPU or accelerators. The design emphasizes clear APIs and stable model serialization.

Key features include automatic differentiation, probabilistic programming tools such as variational inference and Bayesian layers, data loading and preprocessing, training utilities, and experiment tracking. It also supports custom loss functions, metrics, and modular components that can be assembled into complex architectures.

Implementation details: sigmaDL is primarily Python-based, with performance-critical parts in C++ and CUDA where needed. It provides backend abstraction to support alternative tensor engines and a lightweight graph executor to optimize runs. The project follows an open governance model and regular releases.

History and status: sigmaDL was initiated by SigmaTech researchers in 2023 and released publicly in 2024. It has an active contributor community and is maintained on a public repository under a permissive license. Documentation includes tutorials, API references, and examples.

Reception and usage: The project is used in academia for teaching probabilistic deep learning and in early research projects. It aims to complement existing frameworks by offering integrated probabilistic modeling and a cohesive developer experience.