# Models — SpatialTemporalGCN

## Architecture

Spatiotemporal GCN for Wuhan respiratory disease risk prediction:

- **Temporal**: Transformer encoder (3 layers, 4 heads) over 14-day weather windows
- **Spatial**: 2-layer GCN (48→128→64) with elevation/population scaling
- **Output**: `[N, 3]` risk probabilities (1-day, 3-day, 7-day horizons)

## Files

```
models/spatiotemporal_gcn/
  model.py          # SpatialTemporalGCN class + ONNX export
  sampler.py        # Graph sampling utilities
  best_model.pt     # Trained weights (gitignored)
```

## Input Shape

- Node features: `[N, T=14, 48]` — N nodes, 14 timesteps, 48 weather features
- Edge index: `[2, E]` — sparse adjacency from 100m grid graph
- Spatial scalars: elevation + population density per node

## Training

```bash
python scripts/train_model.py   # Full pipeline with MLflow tracking
```

Baseline MAE targets: 1-day=0.2314, 3-day=0.5424, 7-day=0.6391

## Anti-Patterns

- Don't change model architecture without updating `scripts/train_model.py` and `scripts/inference_*.py`
- Don't load `best_model.pt` without matching the exact `SpatialTemporalGCN` constructor args
- Don't skip ONNX export validation after architecture changes
- Don't train without MLflow logging