Key Concepts

EasySweeps is built on a few core principles that simplify the management of large-scale experiment sweeps.

Templates & Variants

The most powerful feature of EasySweeps is the separation of Hyperparameter Templates from Configuration Variants.

The Problem

When running sweeps across different datasets (e.g., MNIST and CIFAR-10), you often want to use the same hyperparameter search space (e.g., Grid Search on Learning Rate). Traditionally, this requires: 1. Creating two separate W&B sweep files. 2. Manually duplicating the hyperparameter logic in each. 3. Keeping them in sync when you change a range.

The EasySweeps Solution

You define one template and one variants file.

1. Template: A standard W&B YAML where some values are placeholders (e.g., dataset: {dataset}).
2. Variants: A simple mapping of placeholders to list of values.

EasySweeps automatically generates and launches isolated sweeps for every combination, ensuring clean data separation while maintaining a DRY (Don't Repeat Yourself) configuration.

Systemd Scope Agents

Unlike standard background processes, EasySweeps launches agents inside systemd scope units.

• Isolation: Each agent group is tracked by the system manager.
• Reliability: If your terminal closes, the agents keep running.
• Easy Cleanup: The ez kill command uses these scopes to safely and completely terminate all processes associated with an agent, including child processes, preventing "zombie" training runs on your GPUs.

GPU Management

EasySweeps treats GPUs as first-class citizens. By providing a --gpu-list, the tool automatically handles CUDA_VISIBLE_DEVICES for you, ensuring that two agents never fight over the same hardware unless you explicitly want them to.