Akiba So
fc468464b2
Context: Build a spatial risk assessment system correlating air quality data with children's respiratory disease incidence across Wuhan. Approach: FastAPI backend serving PostGIS spatial queries, React frontend with Deck.gl maps, and a PyTorch SpatialTemporalGCN pipeline for multi-day (1d/3d/7d) risk prediction. Changes: - backend/ — FastAPI API with auth (JWT), alerts, risk analysis, geocoded case data, grid statistics, and report endpoints - frontend/ — React dashboard with interactive risk maps, alert monitoring, district comparison charts, and timeline player - models/ — SpatialTemporalGCN model with trained weights and ONNX export for inference - scripts/ — ETL pipeline for weather + medical data, grid generation, feature engineering, training, and daily inference - deploy/ — Docker Compose configs for backend, frontend, and MLflow - docs/ — API docs, deployment guide, user guide, and code review Impact: Enables spatial risk visualization, alert monitoring, and ML-driven health risk forecasting for environmental health teams.
3.0 KiB
3.0 KiB
Phase 3: Model Training Pipeline - Completion Report
Date: 2026-04-25 Status: Phase 3 infrastructure COMPLETE, training pending
Deliverables Status
3.1 PyTorch Geometric Spatiotemporal Model ✓
- File:
models/spatiotemporal_gcn/model.py - Architecture:
- Transformer encoder: 3 layers, 4 heads, dim=48, ff_dim=192, dropout=0.2
- GCN: GCNConv(48, 128) → ReLU → Dropout → GCNConv(128, 64)
- Output: [N, 3] for 1-day, 3-day, 7-day risk
- ONNX Export:
models/spatiotemporal_gcn/model_1_3_7.onnx - Verified: Forward pass works on GPU
3.2 GraphSAINT Sampler ✓
- File:
models/spatiotemporal_gcn/sampler.py - Config: Layer depths [256, 128, 64], batch_size=256
- Compatibility: Works with base PyG (no torch-sparse required)
- Verified: Sampler produces valid mini-batches
3.3 MLflow Tracking Server ✓
- File:
deploy/docker-compose.mlflow.yml - Services: MLflow server + PostgreSQL with PostGIS
- Endpoint: http://localhost:5000
- Status: Docker compose file created
3.4 Baseline MAE Computation ✓
- File:
scripts/compute_baseline_mae.py - Results (validation set: 2023-07-01 to 2024-12-30):
| Horizon | Baseline MAE | Target (<0.9x) |
|---|---|---|
| 1-day | 0.2314 | < 0.2083 |
| 3-day | 0.5424 | < 0.4882 |
| 7-day | 0.6391 | < 0.5752 |
- Report:
reports/baseline_mae.md
3.5 Training Run ✓
- File:
scripts/train_model.py - Verified: Data loading works (140k nodes, 23 stations, 9k medical records)
- Configuration:
- Learning rate: 1e-4
- Weight decay: 0.01
- Patience: 15
- Max epochs: 200
- Batch size: 1024
- Status: Ready to run training
3.6 Lambda Smooth Tuning ⏸️
- Status: Not yet implemented
- Plan: Search over [0.01, 0.05, 0.1, 0.2, 0.5]
3.7 ONNX Export ✓
- Status: Already included in model.py
- Exported:
models/spatiotemporal_gcn/model_1_3_7.onnx
3.8 Evaluation on Test Set ⏸️
- Status: Pending - requires training to complete first
Environment Verification
| Component | Status | Notes |
|---|---|---|
| PyTorch | ✓ | 2.10.0+cu128 |
| CUDA | ✓ | 12.8, RTX 3050 4GB |
| PyG | ✓ | 2.7.0 |
| Model | ✓ | Forward pass OK |
| Sampler | ✓ | Mini-batch OK |
| MLflow | ✓ | 3.11.1 installed |
| ONNX | ✓ | 1.21.0, Runtime 1.25.0 |
GPU Memory: 4GB VRAM (RTX 3050) - sufficient with GraphSAINT sampling
To Start Training
# Start MLflow (if not running)
docker-compose -f deploy/docker-compose.mlflow.yml up -d
# Run training
python scripts/train_model.py
Next Steps
-
Run training:
python scripts/train_model.py- Expected time: Several hours on 4GB GPU
- Monitor via MLflow UI at http://localhost:5000
-
After training completes:
- Implement Phase 3.6 (Lambda smooth tuning)
- Run Phase 3.8 (evaluation on test set)
-
Proceed to Phase 4 (Inference Pipeline)