No-Code Tuning Examples

This page provides a step-by-step walkthrough of no-code tuning using the demo preset.

Prerequisites

Uses intelligence/zenith_tune/nocode/fake_train.py from aibooster-examples. This script is a dummy training script whose execution time varies depending on --batch-size, --num-workers, and the OMP_NUM_THREADS environment variable.

# Verify it works
python fake_train.py --epochs 3

Starting training: epochs=3, batch_size=32, num_workers=4, OMP_NUM_THREADS=(not set)
Epoch 1/3 completed in 0.330s
Epoch 2/3 completed in 0.323s
Epoch 3/3 completed in 0.337s
Total training time: 0.99s

1. Check Available Presets

zenithtune optimize --list --dev

Presets:
  dev:demo
Strategies:
  grid
  random
Evaluators:
  duration

2. Run Tuning

Run 5 trials with the dev:demo preset.

zenithtune optimize --preset dev:demo --dev --n-trials 5 -- python fake_train.py --epochs 3

A total of 5 trials including the baseline (running the original command as-is) are executed, and the study is saved under outputs/.

3. Analyze Tuning Results

zenithtune analyze outputs/study_YYYYMMDD_HHMMSS/study.db

Analysis graphs (history.png, timeline.png, importances.png) are generated in the study directory.

4. Run Command with Optimal Parameters

zenithtune apply --db-path outputs/study_YYYYMMDD_HHMMSS/study.db --preset dev:demo --dev -- python fake_train.py --epochs 10

Applies the best parameters discovered during tuning and runs the command. You can specify different options from those used during tuning, such as --epochs 10, for production runs.

Inside the Demo Preset

The dev:demo preset optimizes the following 3 parameters.

Parameter	Applied to	Search Range
batch_size	--batch-size	8–256 (step 8)
num_workers	--num-workers	1–16
omp_num_threads	Environment variable OMP_NUM_THREADS	1–number of CPU cores

RegexEvaluator is used to extract Total training time: X.XXs from stdout and minimize it.