Tests

This page documents the current repository test coverage around diffusion math, GeoTIFF dataset behavior, config wiring, inference exports, and one-batch model smoke runs.

How To Run

Run the full local suite through the repository wrapper:

bash tests/run_tests.sh

The wrapper uses unittest discovery in the configured environment. It also prepends src to PYTHONPATH so tests exercise the src-layout package without requiring a prior editable install.

Main Test Areas

Diffusion Math

tests/test_diffusion_math.py checks schedule construction, mask polarity, ambient further-corruption, target selection, masked loss behavior, and the guarded zero-loss path.

Model Dry Runs

tests/test_model_dry_runs.py instantiates small in-memory datasets and tiny models to verify one training/validation batch for pixel diffusion, latent diffusion, and the autoencoder.

Inference Exports

The global export and Cesium packaging tests verify that inference code uses public dataset metadata (rows, depth_axis_m) and produces consistent raster, GeoJSON, summary, and globe asset outputs.

Config And CLI Wiring

Training, inference, validation-summary, and override tests protect the active pixel super-config path, scenario derivation, effective config materialization, and GeoTIFF builder selection.

Notes

• Tests use unittest, not pytest.
• Some smoke tests intentionally use tiny in-memory fixtures so model wiring can be checked quickly without local source archives.