pomegranate

Treats all probabilistic models as distributions, enabling easy creation of complex structures like mixture models with heterogeneous components or Bayes classifiers with HMMs, as highlighted in the modular design philosophy.

Leverages PyTorch for GPU acceleration, mixed precision, and efficient serialization, providing modern computational features and better integration with deep learning workflows.

Uses PyTorch's MaskedTensor to support incomplete data without imputation bias, available for key distributions and models, enhancing real-world data applicability.

Supports prior probabilities for labeled and partially labeled data in mixture models and HMMs, allowing nuanced semi-supervised training scenarios.