Internal¶
Audience
This section is intended for power users and extenders. Symbols here are not part of the supported public surface and may change without notice. Two subtrees are excluded entirely:
pytyche.research.*— alternative implementations kept for diagnostic / comparison value (already excluded from pyright).pytyche.bcf.cpu_gfr— StochTree CPU backend used for cross- validation and bench work only; not a user runtime or extender hook.
For the curated user-facing surface, see Public API.
Top-level package¶
v2 typed contracts for the pytyche analysis pipeline. |
|
v2 metric dispatch registry. |
|
Empirical (non-Bayesian) summarization of observed experiment data. |
|
Runtime validators for v2 contracts. |
|
Visual confidence payload for generator verification. |
BCF subpackage¶
Submodules of pytyche.bcf listed explicitly so we can omit
pytyche.bcf.cpu_gfr (the parent package’s stub would otherwise pull
it in via package walk). The pytyche.bcf.hurdle and
pytyche.bcf.diagnostics subpackages are listed recursively.
Phase 2 — independent binary/probit BCF (Albert-Chib data augmentation). |
|
Configuration dataclass, result types, and small utilities for the GPU BCF. |
|
Phase 1 — independent continuous BCF (bartz-based JIT scan). |
|
BCF posterior calibration diagnostics. |
|
GPU GFR warm-start (Greedy/Grow-From-Root Forest Refinement) for the GPU BCF. |
|
Low-level GPU primitives shared across the GPU BCF modules. |
|
Pure-math sufficient statistics and log-marginal likelihoods for the GPU BCF. |
|
Covariate preprocessing for GPU BCF. |
|
Public fit entry point for hurdle BCF. |
Calibration¶
SBC calibration — sweeps, layered R(p) corrections, recalibration artifacts. |
Generators¶
Programmatic DGP generator surface ( |
Diagnostics¶
General MCMC convergence and coverage diagnostics. |