Skip to content

Latest commit

 

History

History
95 lines (77 loc) · 4.2 KB

File metadata and controls

95 lines (77 loc) · 4.2 KB

Models Overview

Language: English
Last updated: 2026-07-30
Switch: Chinese

This page is a navigation overview. Current solver, penalty, backend, and inference coverage is maintained in Implemented Methods and the linked model pages.

Core Framework

Page Content
Loss Functions Loss definitions and per-sample formulas
Solver Algorithms Public and internal solver implementations
Loss × Penalty × Solver Framework Dispatch logic and compatibility
Solver × Penalty Matrix Explicit solver routing and restrictions
Inference API Covariance, resampling, and inference interfaces

Regression and GLM

Survival Analysis

Choosing a Cox estimator

Need Estimator Import Contract
Full Cox fitting, baseline hazards, survival prediction, formula input, and inference CoxPH from statgpu.survival import CoxPH Breslow/Efron/Exact ties, delayed entry, (start, stop], strata, robust/cluster covariance, NumPy/CuPy/Torch
Select a non-negative L2 penalty by held-out partial likelihood CoxPHCV from statgpu.survival import CoxPHCV Uses the canonical Cox semantics during CV and performs a final CoxPH refit
Estimate with L1, L2, ElasticNet, SCAD, or MCP PenalizedCoxPHModel from statgpu.linear_model import PenalizedCoxPHModel Broad penalty and generic solver path; currently estimation-only and rejects compute_inference=True

CoxPH(penalty=...) and PenalizedCoxPHModel are not interchangeable aliases. Use the canonical CoxPH/CoxPHCV path when counting-process features, stratification, baseline prediction, or statistical inference are required. Use PenalizedCoxPHModel when the broader penalty family is the primary requirement and estimation-only output is sufficient.

The Cox model page is the authoritative user-facing source for Breslow/Efron/Exact ties, delayed-entry and (start, stop] data, strata, robust/cluster inference, subject-grouped CV, prediction boundaries, and the NumPy/CuPy/Torch support matrix. Internal module ownership and extension rules are documented in dev/design/ARCHITECTURE.md.

Specialized Statistical Modules

Unsupervised Learning

Current Coverage Principles

  • NumPy, CuPy, and Torch are distinct execution backends; explicit device requests do not silently select another backend.
  • Backend support may differ by solver, penalty, inference method, and optional dependency. Consult the detailed compatibility matrix instead of relying on a single global count.
  • Validation claims are scoped to the exact model, backend, hardware, and commit tested.
  • Historical release and benchmark records are evidence snapshots, not current support matrices.