Multimodal modeling code and analysis helpers for ADNI-style imaging + amyloid + demographics workflows. Cohort data are not included; configure local paths and agreements separately.
| Topic | Location |
|---|---|
| Environment, frozen deps, known import gaps | REPRODUCIBILITY.md |
| Exploration scripts, FreeSurfer mapping, per-seed split JSON | data_preprocessing/README.md |
Evaluation (fixed params from configs/best_hyperparameters/, 10 seeds) |
analysis/evaluation/README.md |
| Optuna tuning (baselines + Flex-MoE), prerequisites, smoke test | scripts/optuna/README.md |
Typical order: prepare data/ → experts YAML → make_splits.py → (optional) Optuna tuning or use committed best params → analysis/evaluation/ scripts for multi-seed test metrics.
python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt pip install -e .
The editable install (pip install -e .) registers the baselines package and the shared utils module at the repo root (utils.py). You do not need export PYTHONPATH="$(pwd):$(pwd)/scripts" for normal imports. If you skip pip install -e ., set PYTHONPATH to the repo root (and scripts/ if needed for legacy paths).
Then run preprocessing smoke tests (bash data_preprocessing/smoke_test.sh) and Optuna entrypoint checks (bash scripts/optuna/smoke_test.sh) as described in the docs above.