AnomalyMatch v1.3.0

[v1.3.0] – 2026-02-13

Added

• Multispectral image support for arbitrary channel count images with configurable channel_combination matrices
• UI separation into standalone anomaly_match_ui package for cleaner architecture
• Flux conversion configuration (apply_flux_conversion) ensuring training/prediction consistency
• timm model backend replacing efficientnet-specific packages for broader model support
• test-cnn model for fast unit/integration testing without heavy model downloads

Changed

• Replaced black and flake8 with ruff for linting and formatting
• Restructured test suite into unit/integration/e2e/ui directories with pytest markers and CI caching
• Deduplicated prediction code into shared prediction_utils.py module
• Auto-inference of n_output_channels from channel_combination matrix or FITS extension count
• PIL resize for CONVERSION_ONLY normalisation achieving up to 73x faster image loading
• Faster catalogue validation by skipping per-chunk FITS existence checks and using parquet metadata

Fixed

• Double normalisation in cutana streaming pipeline
• Normalisation consistency between cutana and training paths with channel_weights passthrough
• Session logging with eager directory creation and per-session log files
• Gallery not updating after prediction chunks complete
• Cutana source ID handling for non-string int64 source_ids
• Albumentations 2.0 compatibility renaming deprecated mode to border_mode
• Prediction progress bar with phase tracking for better user feedback
• Identity channel_combination auto-creation for multi-extension FITS configs
• ASinh parameters missing in cutana format config
• Filter name resolution from catalogue for cutana streaming
• Primary HDU validation raising ValueError when no image data found

Documentation

• Normalisation README with improved channel_combination and flux conversion documentation
• Auto-inference documentation updating multispectral config examples

---