SandroMartens/DBGSOM: v.1.2.3

[1.2.3]

Added

• sigma_fine parameter for independent control of neighborhood size during the fine training phase
• Configurable neighborhood function type (new parameter)
• Numba-JIT cosine similarity kernel for n_bmu=1 (performance improvement for metric="cosine")
• Optional X parameter in plot() for data-aligned PCA projections
• New example notebooks: manifold learning comparison, MiniSom comparison, metric comparison, RGB color map
• Parameter validation across all public methods with informative error messages

Changed

• Refactored sigma parameter handling with improved default values
• Growth convergence logic switched to epoch-based control for more stable training
• Fixed sigma decay during coarse training phase to prevent premature convergence
• Improved predict_proba for hierarchical SOMs
• plot(): class labels are now mapped to strings when classes_ attribute is present
• Example-specific dependencies moved to optional examples dependency group

Removed

• convergence_iter parameter — growth triggering logic has been simplified

Fixed

• Numerical precision issues in weight updates
• Parameter validation now correctly triggered on fit() call
• Vertical growth winner indexing bug
• Sigma calculation bug

Documentation

• Expanded and clarified runtime complexity analysis
• Added reference for the transform() method
• Fixed documentation URL (removed version-specific path)

Full Changelog: https://github.com/SandroMartens/DBGSOM/compare/v1.2.0...v1.2.2

Full Changelog: https://github.com/SandroMartens/DBGSOM/compare/v1.2.2...v.1.2.3