Workflows¶

All OptiConn workflows follow the same core pattern: parameter space exploration → objective + QC/QA scoring → robust selection → freeze config → downstream graph analysis. The sections below show the concrete commands for the most common variants.

Bayesian optimization → Apply (recommended)¶

1. Find optimal parameters on a pilot subset:

   python opticonn.py tune-bayes \
       -i /path/to/pilot_data \
       -o studies/bayes_opt \
       --config configs/braingraph_default_config.json \
       --modalities qa \
       --n-iterations 30 \
       --sample-subjects
   

2. Inspect selection:

   python opticonn.py select -i studies/bayes_opt --modality qa
   

3. Apply to the full dataset:

   python opticonn.py apply \
       -i /path/to/full_dataset \
       --optimal-config studies/bayes_opt/qa/bayesian_optimization_results.json \
       -o studies/final_analysis
   

Cross-validation bootstrap (with Bayes seeding)¶

1. Run Bayesian optimization (small pilot) as above.

2. Seed cross-validation with the Bayes result:

   python scripts/cross_validation_bootstrap_optimize.py \
       -i /path/to/pilot_data \
       -o studies/cv \
       --extraction-config configs/demo_config.json \
       --from-bayes studies/bayes_opt/qa/bayesian_optimization_results.json \
       --subjects 3 \
       --max-parallel 1 \
       --verbose
   

3. Uses two waves by default; metrics/atlases stay fixed from the base config.

4. Seeded parameters come from best_parameters in the Bayes results.

Apply-only with known optimal config¶

python opticonn.py apply \
    -i /path/to/full_dataset \
    --optimal-config studies/bayes_opt/qa/bayesian_optimization_results.json \
    -o studies/final_analysis