Examples
The notebook pages retain code, tables and figures. Download a notebook or clone the repository to run it locally. The 2.0.0 examples below use a few ensembles and at most two workers; their timings are illustrative laptop measurements.
Executed with 2.0.0
| Notebook | Demonstrates | Inputs |
|---|---|---|
| 10. Execution and memory | Database helpers, native/memmap training and prediction, mean/std equivalence, separate file-backed ensemble output, and persistent timing/memory reports. | Bundled docs/mini_data/mini_data.pkl; no external dataset. |
| 11. Local calibration and profiling | Within-stixel blocked calibration, fixed/tuned recipe choices, held-out diagnostics, and native/memmap prediction equivalence. | Small seeded synthetic observations with known marginal probabilities. |
These notebooks retain fresh executed outputs, the stemflow/Python versions, and execution provenance. The first checks execution equivalence; the second illustrates calibration APIs and sparse-bin diagnostics. Neither establishes universally better calibration or accurate copula inference.
To reproduce from the repository root after installing the package and notebook tools:
python -m pip install -e ".[monitoring,tuning]" nbformat nbclient ipykernel matplotlib-inline
python scripts/01_documentation/execute_notebook.py docs/Examples/10.Execution_and_memory.ipynb --output docs/Examples/10.Execution_and_memory.ipynb --in-process
python scripts/01_documentation/execute_notebook.py docs/Examples/11.Local_calibration_and_profiling.ipynb --output docs/Examples/11.Local_calibration_and_profiling.ipynb --in-process
Omit --in-process to use a registered Jupyter kernel. Temporary databases/maps/models are cleaned up; tables and figures are embedded in the saved notebooks. Arial is used when installed; a message explicitly flags its absence on other systems.
Retained historical examples
Notebooks 01–09 keep their original code and saved outputs. They have not been rerun for 2.0.0. Several contain unexecuted cells; their displayed outputs are not evidence of current-version results. Most require Sample_data_Mallard.csv; the main demo also uses Predset_2020.csv, and some comparisons require additional estimators or previously generated metric files. Those external inputs are not bundled here.
| Notebook | Topic |
|---|---|
| 01 | AdaSTEM demo |
| 02 | Learning-curve analysis |
| 03 | Maxent base model |
| 04 | Spherical indexing |
| 05 | Hurdle-model composition |
| 06 | Base-model comparisons |
| 07 | Stixel-size optimization |
| 08 | Lazy loading |
| 09 | Original manual DuckDB/Parquet preparation |
When rerunning, consult the 2.0.0 execution options. Corrected bootstrap, logit aggregation, spatial jitter and rotation defaults can change results. Existing saved geometry retains its documented compatibility; new fits should be evaluated again rather than compared to old figures as if only runtime changed.