How shifts-project/shifts works
Main Components
a) Dataset Components (Core):
• Multiple Sclerosis Lesions Segmentation (Medical Imaging)
• Vessel Power Estimation (Time Series)
• Weather Prediction (Tabular)
• Machine Translation (NLP)
• Motion Prediction (Autonomous Driving)
b) Framework Components:
• Data Loading & Processing
• Model Training
• Evaluation & Metrics
• Uncertainty Quantification
• Visualization Tools
┌─ Data Sources Layer ─┐ ┌─ Processing Layer ─┐ ┌─ Analysis Layer ─┐
• MS Lesions │ -> │ - Data Loading │ -> │ - Evaluation │
• Vessel Power │ │ - Preprocessing │ │ - Metrics │
• Weather │ │ - Transformations │ │ - Uncertainty │
• Translation │ └──────────────────┘ │ - Visualization │
• Motion Prediction │ └─────────────────┘
• ┘
Component Details
• Raw data storage
• Processed datasets
• Evaluation sets
b) Processing Components (Rectangles):
• Data loaders (PyTorch/TensorFlow)
• Preprocessing pipelines
• Feature extractors
c) Analysis Components (Hexagons):
• Metrics calculation
• Uncertainty estimation
• Visualization tools
Specific Implementation Details
Project-specific elements:
┌─ Task-Specific Modules ─┐
• mswml/ - MS lesions │
• vpower/ - Vessel power │
• weather/ - Weather │
• translation/ - MT │
• sdc/ - Motion pred. │
• ┘
Integration Points
Connections between:
• Dataset APIs (ysdc_dataset_api/)
• Evaluation metrics (metrics.py in each module)
• Uncertainty quantification (uncertainty.py)
Important Notes
• Dataset licenses and sources
• Evaluation points and metrics
• Uncertainty quantification methods
• Integration with external tools (WandB, CSV logging)
Standardized metrics across tasks
This architecture should capture both the high-level organization and the detailed implementation structure while maintaining clarity and usefulness for development and documentation purposes.