Skip to content

sklearn.md

Latest commit

 

History

History
60 lines (37 loc) · 4.89 KB

sklearn.md

File metadata and controls

60 lines (37 loc) · 4.89 KB
Raw

scikit-learn (sklearn) — real native Cython build

Version: 1.9.0 | Type: Native iOS arm64 (real upstream Cython cross-compile) | Extensions: 69 .cpython-314-iphoneos.so | Submodules: 39 (full API) | Location: sklearn/

This is the real, full scikit-learn 1.9.0 — the upstream Cython/C++ package cross-compiled for iOS arm64 — not a reimplementation. It replaces the previous pure-NumPy reimpl (1.8.0-offlinai, ~40 modules / "85% of workflows") with the complete upstream library: every estimator, every submodule, the genuine Cython kernels, and the libsvm/liblinear C++. As far as we can tell, the first public iOS build of scikit-learn.

It sits on the bundled native numpy (Accelerate/AMX) and scipy, plus pure-Python joblib, threadpoolctl, and narwhals.

Quick Start

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score

X, y = make_classification(n_samples=200, n_features=4, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

clf = RandomForestClassifier(n_estimators=10, max_depth=5)
clf.fit(X_train, y_train)
print(f"Accuracy: {accuracy_score(y_test, clf.predict(X_test)):.3f}")

Because it's the real library, the full upstream API is available and behaves exactly as documented at scikit-learn.org — fit / predict / transform / score / predict_proba / decision_function / get_params / set_params, pipelines, grid search, the lot.

What's included — all 39 submodules

Every public submodule imports and runs (verified on-device, see below):

calibration · cluster · covariance · cross_decomposition · datasets · decomposition · discriminant_analysis · dummy · ensemble · exceptions · experimental · feature_extraction · feature_selection · gaussian_process · impute · inspection · isotonic · kernel_approximation · kernel_ridge · linear_model · manifold · metrics · mixture · model_selection · multiclass · multioutput · naive_bayes · neighbors · neural_network · pipeline · preprocessing · random_projection · semi_supervised · svm · tree · utils (+ base, pipeline, compose, top-level clone/get_config/set_config/config_context/show_versions).

The heavy compiled kernels are all present and numerically correct: tree/ensemble (decision trees, gradient boosting, HistGradientBoosting), svm (libsvm + liblinear C++), neighbors (KD-/Ball-tree), cluster (_kmeans), linear_model (_cd_fast, _sgd_fast, _sag), metrics (_pairwise, _dist_metrics), manifold (Barnes-Hut TSNE), decomposition, _loss, and the shared _cyutility module.

iOS-specific notes & caveats

  • Serial (no OpenMP). iOS has no libomp, so the OpenMP-parallel kernels are built serial (SKLEARN_OPENMP_PARALLELISM_ENABLED=0, gated on _OPENMP). Results are identical; only intra-estimator multi-core is unavailable. (A cross-compiled libomp could enable it later.)
  • n_jobs > 1 uses joblib's default loky backend, which needs fork() — unavailable on iOS. Use n_jobs=1 (default) or joblib.parallel_backend("threading").
  • Built against numpy 2.4.x / scipy 1.17.x headers; the device runs numpy 2.3.5 / scipy 1.15.0. The numpy/scipy C-API is capsule-based and ABI-stable across 2.x (NPY_NO_DEPRECATED_API), so this is compatible — and confirmed on-device.
  • Extensions resolve Python / numpy / scipy at runtime (-undefined dynamic_lookup + capsules); they link only libSystem + libc++.

How it's built

Cross-compiled with a BeeWare cross-venv (host venv patched to report iOS sysconfig) + a meson cross-file pointing at the arm64-apple-ios-clang wrappers, via pip wheel --no-build-isolation --config-settings=setup-args=--cross-file=…. Harness + recipe: sklearn_ios/ (ios-cross.ini, build notes). Two gotchas worth knowing: meson must use Cython 3.1.x (the host's 3.0.12 rejects sklearn's cdef const), and the OpenMP dependency must be forced not-found (else meson links the macOS Homebrew libomp). The previous pure-NumPy reimpl is kept at sklearn_ios/sklearn_reimpl_backup/.

Verification

Validated on-device (CodeBench, "Designed for iPad" / real device) with sklearn_ios/test_sklearn_on_device.py: it imports every submodule and runs real-value correctness checks — recovered linear coefficients exactly, MSE=1/3, ROC-AUC=0.75, PCA explained-variance 0.9777 + lossless reconstruction, KMeans/GMM ARI=1.000, the libsvm/liblinear C++, KD-trees, and Barnes-Hut TSNE all correct. 66/66 checks pass.

See also