How claudiashu/sonnet works
Main Components
• CLI & Config Manager
• Entry point: scripts/run_experiment.py
• Hydra + YAML configs under configs/
• Dataset Layer
• Local: datasets/weatherbench/*.csv
• External: Darts API (ELEC, ENER, ETTh)
• Configs: configs/dataset/...
• Data Loader
• sonnet/data/weatherDataloader.py
• sonnet/data/dartsDataloader.py
• Model Zoo
• Core: sonnet/mts_model/models/Sonnet.py
• Baselines: sonnet/baselines/models/*.py
• Shared layers: sonnet/baselines/layers/*.py
• Lightning Engine
• Modules: sonnet/lightning/lightning_module.py
• Naive baselines: sonnet/lightning/lightning_naive.py
• Training & validation loops via PyTorch Lightning
• Utilities
• Config loaders: sonnet/data_config.py, sonnet/model_config.py
• Metrics: sonnet/utils/metrics.py
• Misc: sonnet/utils/utils.py, masking, RevIN, etc.
Relationships and Data Flow
1. User invokes the CLI
→ Hydra reads configs/exp/*.yaml, configs/model/*.yaml, configs/dataset/...
2. Config manager instantiates
→ Data Loader (weather or Darts)
→ Model (Sonnet or one of the baselines)
→ LightningModule
3. PyTorch Lightning Trainer
→ Feeds batches from Data Loader
→ Calls forward pass of selected Model
→ Computes loss & metrics via sonnet/utils/metrics.py
4. Checkpointing & Logging
→ TensorBoard / Hydra outputs
5. Post‐processing
→ Scripts under scripts/ can generate run‐sweep files (generate_runs.sh)
Architectural Patterns & Design Principles
• Modular “Model Zoo” for easy swap‐in of baselines vs Sonnet
• Config‐driven execution (Hydra) for reproducibility
• Separation of concerns: data loading, model definition, training loop
• Extensibility: add new datasets by placing YAML in configs/dataset, implement new DataLoader