How-to guides

Each guide finishes one task. Pick the task, follow the numbered steps, and check the expected result. If you have not run Starfold yet, start with Your first run. If you want to know why a step works, follow the links to Concepts and Metrics.

Prepare

• Prepare your table for recover: Prepare your table for recover (shape, finite values, missing measurements, column choice, units).

• Choose between standard and robust scaling: Choose between standard and robust scaling.

• Narrow the search space: Narrow the search space, and see why the library defaults are always trial 0.

• Pass measurement errors to recover: Pass measurement errors so every refit redraws your values.

Run

• Set the budgets and run on a laptop: Set the search and refit budgets, and run a laptop-sized fit.

• Run Starfold on a GPU: Run on a GPU with cuML, or fall back to the CPU.

• Run fits in parallel with n_jobs: Run fits in parallel with n_jobs.

• Reproduce a run: Reproduce a run from its seed and its record.

Tune

• Narrow the search space: Narrow the ranges Optuna searches.

• Set the budgets and run on a laptop: Trade run time against how well the search explores.

Check

• Run the calibration test: Run the calibration test and read its p-value.

• Compare the CPU and cuML partitions: Compare the CPU and cuML partitions.

• Check a result against known labels: Check a result against known labels after the fact.

• Plot your result: Redraw the plots of the digits notebook for your own result.

• Report your result in a paper: Choose what you may write in a paper about a result, and what you may not.

Save

• Save a result and load it back: Save a result to disk and load it back with checksums checked.

Scale up

• Run fits in parallel with n_jobs: Use several CPU processes.

• Run Starfold on a GPU: Use one GPU.

• Run a seed sweep on a SLURM cluster: Run a seed sweep on a SLURM cluster, then calibrate the result you report.