v2.0.0

Breaking Changes

• Transform System: Complete refactor to PyTorch named tensors for GPU acceleration. Transforms now use ('channel', 'time') or ('batch', 'channel', 'time') tensor conventions
• Dataset Architecture: Refactored datasets module with layered architecture (WindowedDataset, DataModule, DatasetCreator)
• Models Restructure: Flattened model hierarchy from 4 levels to 2 (models/raul_net/v16.py). Removed deprecated models (V1, V4, V9 variants, classification), removed online/offline distinction
• Datatypes Package: Split datatypes.py into organized package structure
• Storage Format: Enforced .zip format for datasets, removed GPU Direct Storage support

Added

• SlidingWindowTransform base class for consistent transform implementations
• MLflow experiment tracking and visualization (myoverse/tracking.py)
• RAM caching and multiprocessing support in ContinuousDataset
• Lazy dataloader initialization for faster startup
• Cache pre-loading in main process before spawning workers
• Nested dataset storage structure (split/modality/task)
• ZipStore for faster dataset I/O on Windows
• py.typed marker for PEP 561 compliance

Changed

• Lazy import of heavy dependencies in datasets module for faster imports
• Modernized codebase to Python 3.12+ type hints and modern super() calls
• Ran ruff format across entire codebase
• Removed graph visualization from _Data class
• Removed Rich auto-coloring of numbers and text for cleaner output
• Moved emg_xarray and emg_tensor to datatypes module

Removed

• Unused utils module
• tests/workflow.py
• icecream dependency
• Deprecated model variants (V1, V4, V9, classification models)
• GPU Direct Storage (GDS) support

Fixed

• MedianFilter padding
• Tests updated to match simplified _Data API

Full Changelog: v1.1.6...v2.0.0