Slurm Search

Slurm search is a library for expressing, running, and managing distributed experiments on a cluster. It has a number of parts, including a storage system, a job management tool, an experiment management system, hyperparameter tuning algorithms, and integrations with a reinforcement learning library.

Installation

Install autonomous-learning-library and pytorch before installing this package. Run the following commands:

pip install .
mkdir -p ~/hyperparameters/locks # Locks assume their directory already exists.

Usage

Experiments are specified in slurm_search/experiments/. Using the run_exp command, you may:

# Launch an experiment with a series of override parameters.
run_exp hp_tuning_effects --agent=atari:a2c --env=atari:Breakout --search:threads=12

# Relaunch an interrupted experiment, including to repeat an updated result analysis.
run_exp hp_tuning_effects --agent=atari:a2c --env=atari:Breakout --search:threads=12 --resume=exp:my_exp_name

# Display the experiment AST visually
run_exp hp_tuning_effects --display-ast=true # Regular
run_exp hp_tuning_effects --display-ast=abstract # Abstract equations

Components

This package has the following parts:

• A pickle-based object database that supports safe concurrent access across a cluster.
• A session management system that integrates with the slurm job management system for parallel execution of tasks across a cluster.
• Hyperopt integrations for parallel hyperparameter tuning searches.
• An python experiment DSL and tools for configuring, running, and resuming experiments and tools for analyzing the results.
• An integration with the Autonomoun Learning Library for running reinforcement learning experiments.

Files

Core:

• experiments/: Experiment specifications, domain-of-interest integrations, and analysis tools.
• display_experiment.py: Experiment management CLI tool.
• experiment.py: Experiment DSL.
• slurm_search.py: Slurm integration, session and state management CLI tool 'ssearch'.
• search_session.py: Parallelized search session interface.
• session_state.py: Transactional string-identified state store (uses locking and pickle).
• locking.py: String-identified mutex (uses file locks).

Utilities:

• params.py: Tools for handling parameter dict trees.
• random_phrase.py: Generate random phrases / names for object identifiers.