Testing Philosophy (Read This First)

Each module in IPSL-AID comes with its own dedicated tests.

Before running large-scale training, users are strongly encouraged to:

Explore the test suite
Understand data handling and normalization
Validate diffusion and architecture choices
Tune hyperparameters using small debug runs

This approach minimizes wasted compute time and ensures correct scientific usage.

Why Testing is Essential

Climate downscaling models involve:

Complex neural architectures
Sophisticated diffusion formulations
Large-scale data processing
Multi-GPU parallelization
Numerical precision requirements

Without proper testing, errors can:

Waste computational resources (days of GPU time)
Produce scientifically invalid results
Mask underlying implementation bugs
Compromise reproducibility

Recommended Workflow

Unit Tests: Run tests for individual components
```
python tests/test_all.py
```
Integration Tests: Test data loading and model initialization
```
python tests/test_integration.py
```

Small-scale Debug Runs: Train on a small subset

# Use debug mode in setup script
debug=true
num_epochs=1
batch_size=4

Validation: Check against known baselines

Test Coverage

IPSL-AID includes tests for:

Data loading and preprocessing
Diffusion model implementations (VE, VP, EDM, iDDPM)
UNet architectures and variants
Loss functions and training loops
Inference and sampling procedures
Evaluation metrics and diagnostics

Debugging Tips

Use torch.autograd.set_detect_anomaly(True) during development
Monitor GPU memory usage with nvidia-smi
Check numerical stability with gradient norms
Validate data normalization and scaling

Remember: Test first, scale later. A few hours of testing can save days of wasted computation.