v0.1.0 — Multi-Backend TP-GMM

First stable release of tpgmm with full support for three computational backends.

Highlights

• Multi-backend architecture — unified API across NumPy, PyTorch, and JAX with shared abstract base classes in tpgmm._core
• JIT-compiled JAX backend — entire EM iteration compiled via @jax.jit, making JAX the fastest backend on CPU (~10 ms per fit)
• Vectorized Gaussian PDF — einsum-based implementation with no Python loops in the hot path across all backends
• Model selection — BIC, AIC, silhouette score, and log-likelihood scoring built into every backend

Core

• TPGMM — Task Parameterized Gaussian Mixture Model with K-Means initialization and Expectation Maximization
• GaussianMixtureRegression — conditional regression from a fitted TP-GMM, with from_tpgmm() convenience constructor
• Abstract base classes (BaseTPGMM, BaseGMR, LearningModule) ensuring consistent interfaces and documentation

Backends

Backend	TPGMM	GMR	Notes
NumPy	yes	yes	Pure NumPy + SciPy, no GPU required
PyTorch	yes	yes	GPU-ready, eager execution
JAX	yes	yes	IT-compiled EM loop, GPU/TPU-ready

Extras

• Three example notebooks (examples/example_numpy.ipynb, example_torch.ipynb, example_jax.ipynb) with identical structure: data loading → model fitting → model selection → GMR → visualization
• Comprehensive test suite (46 tests) with pytest-benchmark runtime profiling
• Optional dependency groups: torch, jax, examples, dev, all

Installation

Currently, only supported via install from the repository.

pip install tpgmm              # NumPy backend
pip install tpgmm[torch]       # + PyTorch
pip install tpgmm[jax]         # + JAX
pip install tpgmm[all]         # Everything

Requirements

• Python >= 3.12
• NumPy, SciPy, scikit-learn (base)
• PyTorch >= 2.0 (optional)
• JAX >= 0.4 (optional)