Run Hyperparameter Sweeps

Once you've validated your training code in a notebook, you can launch hyperparameter sweeps to find optimal model configurations. This workflow uses Run Remote to create an execution, then converts it to a Task for parallel tuning.

Prerequisites

Before launching a sweep:

• Your notebook should run top-to-bottom without manual intervention

• Training logic should use parameters, not hardcoded values

• Model performance should be logged as metadata

Prepare Your Notebook

Define your step configuration using valohai.prepare():

import valohai

valohai.prepare(
    step='model-training',
    image='python:3.9',
    default_inputs={
        'dataset': 's3://mybucket/data/train.csv'
    },
    default_parameters={
        'learning_rate': 0.001,
        'batch_size': 32,
        'epochs': 10
    }
)

# Your training code
learning_rate = valohai.parameters('learning_rate').value
batch_size = valohai.parameters('batch_size').value
epochs = valohai.parameters('epochs').value

# Train model and log metrics
for epoch in range(epochs):
    # ... training loop ...
    with valohai.metadata.logger() as logger:
        logger.log('epoch', epoch)
        logger.log('accuracy', accuracy)
        logger.log('loss', loss)

The parameters you define here become tunable in your Task.

Run Remote

Click the Run Remote button in Jupyter. Valohai executes your notebook end-to-end and creates a standard execution.

Wait for the execution to complete successfully. If it fails, debug in your notebook and try again.

Convert to Task

Once your execution succeeds:

1. Navigate to the completed execution in the Valohai UI

2. Click the Task icon (or Create Task button)

3. Select your Task Type:

   • Grid Search — test all parameter combinations

   • Random Search — sample parameter space randomly

   • Bayesian Optimization — intelligently explore parameter space

4. Configure the parameter ranges you want to test

Valohai launches multiple executions in parallel, each testing different parameter values. Compare results in the Task view to find the best configuration.

When to Use This Workflow

Good for:

• Quick hyperparameter exploration on proven notebook code

• Prototyping tuning strategies before moving to YAML

Better alternatives:

• For production tuning workflows, define Tasks in valohai.yaml

• For complex multi-step pipelines with tuning, use pipeline-based tasks

• For recurring tuning jobs, use scheduled executions

This notebook → task workflow is convenient for one-off experiments, but production tuning should use proper YAML definitions for versioning and reproducibility.

Next Steps

• Tasks & Parallel Execution — Deep dive into hyperparameter optimization

• Grid Search — Exhaustive parameter testing

• Bayesian Optimization — Smart parameter exploration

• Convert to Production Script — Move tuning logic to YAML