gplite 2.1.2__tar.gz

gplite-2.1.2/.gitignore

@@ -0,0 +1,18 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+build/
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+.python-version
+uv.lock
+
+# OS files
+.DS_Store
+
+# misc
+TODO

gplite-2.1.2/CHANGELOG.md

@@ -0,0 +1,59 @@
+# Changelog
+
+Notable changes to this project will be documented in this file.
+
+## [2.1.2] - 2026-03-07
+
+### Added
+- **Benchmarking** [Benchmark suite](benchmarks/) for testing compute-heavy sections of this package
+
+## [2.1.1] - 2026-03-07
+
+### Added
+
+- **Performance** Minor changes to computation utilities and optimization algorithms for increased performance
+
+### Fixed
+
+- v2.1.0 changelog now correctly reflects the date it was released
+
+## [2.1.0] - 2026-03-05
+
+### Added
+
+- **Matérn kernel** (`MaternKernel`) with support for ν=1.5 (once differentiable) and ν=2.5 (twice differentiable), including isotropic and anisotropic variants
+- **Expected Improvement** selection strategies for active learning: `ei_max` (maximize) and `ei_min` (minimize) for Bayesian optimization use cases
+- **Model save/load**: `gp.save(filepath)` and `GaussianProcess.load(filepath)` using pickle serialization
+- Efficient `_compute_diag` method on all kernels for O(n) predictive variance computation instead of O(n²)
+- [Example files for usage reference](examples/)
+- **Logging capability** to give ActiveLearner updates in a file rather than (or along with) stdout
+
+### Fixed
+
+- Predictive variance now correctly computed for non-unit-diagonal kernels (e.g., `ConstantKernel * RBFKernel`); previously hardcoded `k(x,x) = 1`
+- Active learning optimizer now refits the model after setting final hyperparameters, matching GP optimizer behavior
+- Active learning optimizer fallback path now refits when all screening runs fail
+- Missing `raise` in `validate_variable_names` — wrong number of variable names was silently accepted
+- Active learning `max_points` budget now correctly accounts for initial training points and clamps `batch_size` to prevent overshooting
+- Installation instructions now have the correct link for installing the package with uv and standalone pip
+
+### Changed
+
+- Refactored Cholesky decomposition into two clear phases (exponential noise retry, then eigenvalue fallback), fixing incorrect noise return value after eigenvalue correction
+- Renamed `SMALL_EPSILON` to `EPSILON` in `_utils/_constants.py`
+- `_validate_input_data` in base `Kernel` class changed from abstract to concrete method, reducing boilerplate in concrete kernel subclasses
+- Default `max_points` in `ActiveLearner` now uses `len(y_full)` instead of `np.floor(len(y_full))`, returning an int instead of float
+- Removed unused `self.x_test` and `self.y_test` attributes from `GaussianProcess`
+- Simplified `fit()` method by removing redundant `_fit_without_optimization()` call in the optimize branch
+
+## [2.0.0] - 2026-02-15
+
+
+### Features
+
+- Gaussian Process regression with automatic hyperparameter optimization via log marginal likelihood
+- Kernels: RBF, Periodic, Constant, with additive and product composition (`+` and `*`)
+- Anisotropic (ARD) kernel variants with per-dimension hyperparameters when applicable 
+- Active learning with uncertainty, max absolute error, and random selection strategies
+- Input and target normalization
+- String export of fitted GP expressions for integration with external tools (e.g., OpenMM)