Releases: esa/AnomalyMatch
Releases · esa/AnomalyMatch
AnomalyMatch v1.2.0
[v1.2.0] – 2025-01-22
Added
- Cutana streaming integration for catalogue-based predictions with parquet and CSV support
- FitsBolt integration for consistent FITS normalization across training and prediction
- Iteration score storage for tracking unlabeled and test data scores per iteration
- Automatic batch size estimation using exponential and binary search for optimal GPU memory usage
- Full resolution image preview button in the UI for detailed inspection
- Dead code detection CI workflow using Vulture for codebase maintenance
Changed
- Refactored Widget architecture by extracting PreviewWidget for better code organization
- FitsBolt config persistence in model checkpoints for reproducible normalization
- Parquet format for Cutana buffer instead of CSV for improved performance
- Black line-length updated to 100 characters for better readability
Fixed
- Gallery filename display for long filenames with improved shortening (#237)
- Duplicate result accumulation in prediction process (#238)
- Error handling for iteration score CSV saves (#236)
- FITS extension handling in Cutana streaming
- Tensor handling improvements throughout the codebase
Removed
- Dead code cleanup removing unused functions and imports identified by Vulture
- IDE/editor files from repository with updated .gitignore
AnomalyMatch v1.1.0
[v1.1.0] – 2025-07-07
Added
- Zarr file format support for scalable array storage and processing
- Session tracking and management with comprehensive iteration history
- Metadata handling for associating metadata with images in labeled_data.csv
- Configuration validation to ensure proper setup before training
- Label caching for improved performance in active learning loops
- Unlabeling functionality allowing users to remove labels from images
- Interpolation order configuration for image resizing operations
- ASinh normalization with grayscale/multichannel and RGB functionality
Changed
- Improved UI responsiveness with optimized image loading and display
- Enhanced session saving to capture training results and configurations
- Streamlined prediction process with better file type detection
- Reduced logging verbosity to minimize output spam
- Unified image resizing to use BILINEAR interpolation consistently
- Improved error handling for insufficient labeled data scenarios
- Better memory management in prediction processes
Fixed
- RGB display reset when using brightness/contrast sliders
- Train iterations slider usability issues
- Test ratio image reading bugs
- Cached image normalization not updating after training
- Channel ordering in TurboJPEG decoded files
- CPU fallback when CUDA is not available
- NaN/inf value handling in image processing
- Top image preservation across prediction batches
- Label count display in UI
Removed
- ZIP file support (kept for benchmarking, removed from prediction process)
- Redundant configuration options and deprecated functions
Performance
- Faster label lookups through intelligent caching mechanisms
- Optimized batch processing for HDF5 and Zarr formats
- Reduced memory usage in prediction workflows
- Improved UI responsiveness in ESA Datalabs environment