skcure 0.1.0a1__tar.gz

skcure-0.1.0a1/CHANGELOG.md

@@ -0,0 +1,46 @@
+# Release notes
+
+## 0.1.0a1 — unreleased alpha
+
+- Use a consistent training memory layout in GGKM so kernel rounding from
+  NumPy versus DataFrame inputs does not alter the optimizer's fitted result.
+- Consolidate exponential helpers and handle expected tail overflow/underflow
+  consistently, preserving NaNs, infinities, and representable tiny values.
+- Support eager Polars and Narwhals DataFrame feature inputs across estimators,
+  preserving feature-name checks and avoiding conversion through pandas. Share
+  input normalization with standalone kernels and future model implementations.
+- Extend scientific validation to genuine Poisson, negative-binomial, and
+  binomial count-generating mechanisms, preserving existing simulation defaults.
+- Add survival-aware scikit-learn estimator checks and reproducible runtime
+  and process-memory benchmarks.
+- Add a complete GBSG2 real-data analysis with training-only model selection,
+  held-out evaluation, and explicit interpretation limits.
+- Prepare the first alpha metadata and document the public API and release
+  verification procedure. No package publication is performed by CI.
+
+- Recover finite generalized-gamma latency line-search failures with one
+  validated restart, addressing a Windows warm-refit failure. Keep optional
+  JAX imports compatible with type checking when the extra is absent.
+- Organize estimators in `skcure.ggkm`, `skcure.ensemble`,
+  `skcure.linear_model`, and `skcure.deep_learning`.
+- Unify gradient boosting cure families in `GradientBoostingCure` and add
+  a scikit-survival-compatible `DeepPromotionTimeCure` wrapper.
+- Stabilize generalized-gamma tails and gradients, latent count likelihoods,
+  incidence updates, and neural losses at extreme parameter values.
+- Add relative EM convergence tolerance, per-stage solver options, explicit
+  termination reasons, inner optimizer diagnostics, and warnings at fit caps.
+  All EM estimators retain `max_em_iter=1000` by default.
+- Add a total boosting tree budget (default 1000), opt-in validation stopping
+  with complete best-state restoration, and staged tree predictions.
+- Add `predict_latent_mean` consistently across estimator families and
+  document partial warm-start initialization reuse.
+- Distinguish model-implied cure ROC/AUC diagnostics from observed-outcome
+  evaluation; preserve old metric names as aliases.
+- Add IPCW survival evaluation, ground-truth cure simulations, and a
+  simulation benchmark.
+- Use SciPy gamma shape derivatives by default. JAX becomes an optional,
+  explicitly selected helper backend without import-time precision changes.
+- Require NumPy 2.0+ and SciPy 1.13+. Keep optional neural dependencies and
+  the pinned upstream DeepCure source installation separate from core metadata.
+- Add user documentation, CI and distribution checks, citation metadata,
+  and the GPL-3.0-or-later license.

skcure-0.1.0a1/CITATION.cff

@@ -0,0 +1,36 @@
+cff-version: 1.2.0
+message: "If you use scikit-cure, cite this software and the papers relevant to your models."
+type: software
+title: scikit-cure
+authors:
+  - name: scikit-cure contributors
+version: 0.1.0a1
+repository-code: https://github.com/cowvin0/scikit-cure
+url: https://github.com/cowvin0/scikit-cure
+license: GPL-3.0-or-later
+abstract: "Cure models extending the scikit-survival estimator API."
+references:
+  - type: article
+    title: "scikit-survival: A Library for Time-to-Event Analysis Built on Top of scikit-learn"
+    authors:
+      - family-names: Pölsterl
+        given-names: Sebastian
+    journal: Journal of Machine Learning Research
+    year: 2020
+    volume: 21
+    issue: 212
+    start: 1
+    end: 6
+    url: https://jmlr.org/papers/v21/20-729.html
+  - type: article
+    title: The Deep Promotion Time Cure Model
+    authors:
+      - family-names: Medina-Olivares
+        given-names: Victor
+      - family-names: Lessmann
+        given-names: Stefan
+      - family-names: Klein
+        given-names: Nadja
+    journal: arXiv
+    year: 2023
+    url: https://arxiv.org/abs/2305.11575

skcure-0.1.0a1/CONTRIBUTING.md

@@ -0,0 +1,65 @@
+# Contributing
+
+Bug reports and pull requests are welcome at
+[cowvin0/scikit-cure](https://github.com/cowvin0/scikit-cure).
+Include a small reproducible example, package versions, and the traceback
+when reporting a problem. For numerical issues, include the model settings
+and convergence diagnostics.
+
+Use a virtual environment named `.venv` and install development tools:
+
+```bash
+python -m pip install -e '.[dev]'
+python -m pytest -q
+python -m pyright
+python -m build
+python -m twine check dist/*
+```
+
+Tests using real neural dependencies skip when those dependencies are absent.
+Install `requirements-deep-learning.txt` to run them. JAX derivative comparison
+tests similarly require `.[jax]`; core imports and tests must work without
+these extras. The CI matrix checks the minimum core versions, current
+dependencies, multiple operating systems, optional backends, and installable
+wheel/source artifacts. A manual release-check workflow builds and validates
+artifacts without publishing them.
+
+`tests/test_estimator_checks.py` runs selected upstream scikit-learn checks
+with survival-aware target adapters, plus censored-target pipeline checks.
+Keep expected short-fit convergence warnings scoped to those fits; unexpected
+warnings should remain visible. The CI validation job also runs quick scientific,
+performance, and real-data workflows and stores their JSON reports.
+
+Keep constructor parameters cloneable and explicit, use the scikit-survival
+target format, and preserve public prediction semantics. Add tests for
+mathematical invariants and meaningful failure cases when changing fitting
+or numerical routines. Frozen model fixtures in `tests/data` guard numerical
+compatibility; changes to them need an explanation of the mathematical change.
+
+## Adding models and DataFrame support
+
+New cure estimators should inherit `BaseCureEstimator` and use
+`_validate_fit_data(X, y)` for fitting and `_validate_predict_data(X)` for
+prediction. These methods supply the shared DataFrame handling, feature-name
+checks, and numeric validation. Validate each prediction input once, then pass
+the validated array to internal calculations.
+
+Utilities and transformers that do not inherit the estimator base can use
+`ensure_eager_dataframe` from the private `skcure._dataframe` module before
+their usual scikit-learn validation. This helper unwraps Narwhals frames and
+rejects lazy inputs without collecting them. Preserve the native DataFrame
+until feature names have been checked; keep dtype, sparse support, and other
+numeric requirements at the validation call site. Avoid backend-specific
+branches in numerical code and imports from scikit-survival's private modules.
+
+Add new estimators to the shared fixture in `tests/test_dataframe_inputs.py`
+so they receive the same NumPy, pandas, Polars, and Narwhals compatibility
+checks. Extend those checks for any additional input contract.
+
+Update the user guide and `CHANGELOG.md` for public behavior changes. Before
+a release, set matching versions in `pyproject.toml`, `skcure/__init__.py`,
+and `CITATION.cff`, and run `python ci/check_version.py`.
+See [alpha release scope and verification](docs/release.md) before preparing
+distribution artifacts.
+
+The project is licensed under GPL-3.0-or-later; see [LICENSE](LICENSE).