Design and Roadmap
Current architecture (0.3)
- Native Python core (
pyhuge/core.py)
- NumPy/SciPy data model
- Native C++ kernels for MB/glasso/tiger (
pyhuge._native_core)
Data flow
- User inputs NumPy/SciPy matrices
- Python validation and preprocessing
- Method-specific native solver
- Results converted to typed dataclasses
- Plot and summary helpers consume same dataclasses
Compatibility goals
- Keep public API names aligned with earlier
pyhuge wrapper and R huge
- Keep result fields stable where possible (
lambda_path, path, opt_lambda, etc.)
- Preserve one-page-per-function manual docs under
docs/man/
Known approximation boundaries
- Native and R implementations may still differ numerically on some datasets,
though core algorithm families are aligned.
- Selection/inference implementations target practical parity, not strict
numerical identity with R code paths.
Roadmap
- Strengthen numerical parity tests against R outputs (function-by-function)
- Add optional compiled kernels for more hotspots
- Extend solver backends for tighter TIGER parity
- Keep Python and R user docs aligned on semantics and examples