Create a hyperparameter sweep

A sweep runs many trials, each with a different combination of hyperparameters. It tracks the metric you are optimizing and surfaces the best trial.

Prerequisites

• A Kubernetes cluster in ready status with at least one compute pool
• The Vantage SDK installed and configured

Choose a search strategy

Vantage supports three sweep algorithms:

Algorithm	Best for	How it works
Grid	Small, discrete spaces	Tests every combination of specified values
Random	Large spaces, quick exploration	Samples random combinations from defined ranges
Bayesian	Expensive training, adaptive search	Uses prior trial results to pick the next combination

Create a sweep with the SDK

from vantage_sdk import VantageClient

client = VantageClient()
sweep = client.sweeps.create(
    name="lr-batch-sweep",
    objective_metric="val_accuracy",
    objective_goal="maximize",
    algorithm="bayesian",
    parameters={
        "learning_rate": {"min": 1e-5, "max": 1e-2, "distribution": "log_uniform"},
        "batch_size": {"values": [16, 32, 64, 128]},
        "epochs": {"values": [10, 20, 50]},
    },
    training_job={
        "image": "pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime",
        "script": "train.py",
        "gpu": 1,
        "cpu": "4",
        "memory": "16Gi",
    },
    max_trials=50,
)
print(f"Sweep started: {sweep.id}")

Monitor a sweep in the UI

1. Click Workbench in the left sidebar, then click Sweeps under the Train section.
2. The sweep list shows each sweep with its name, algorithm, objective metric, number of trials, and best metric value so far.
3. Click a sweep to view:
   • Trials -- each trial's parameters and resulting metric, sorted by objective
   • Best trial -- the parameter combination that produced the best result
   • Presets -- saved configurations that can seed new sweeps

Use a sweep preset

Presets are pre-defined configurations for common search patterns. Select an existing preset when creating a sweep to start from a known-good configuration instead of defining every parameter from scratch.

Related

• Sweep algorithms for details on each strategy
• Submit a training job