SEML is the missing link between the open-source workload scheduling system Slurm, the experiment management tool sacred, and a MongoDB experiment database. It is lightweight, hackable, written in pure Python, and scales to thousands of experiments.
Keeping track of computational experiments can be annoying and failure to do so can lead to lost results, duplicate running of the same experiments, and lots of headaches.
While workload scheduling systems such as Slurm make it easy to run many experiments in parallel on a cluster, it can be hard to keep track of which parameter configurations are running, failed, or completed.
sacred is a great tool to collect and manage experiments and their results, especially when used with a MongoDB. However, it is lacking integration with workload schedulers.
SEML enables you to
- very easily define hyperparameter search spaces using YAML files,
- run these hyperparameter configurations on a compute cluster using
Slurm, - and to track the experimental results using
sacredandMongoDB.
In addition, SEML offers many more features to make your life easier, such as
- automatically saving and loading your source code for reproducibility,
- easy debugging on Slurm or locally,
- automatically checking your experiment configurations,
- extending Slurm with local workers,
- and keeping track of resource usage (experiment runtime, RAM, etc.).
The fastest way to get started with SEML is via uv:
- Install
uv:curl -LsSf https://astral.sh/uv/install.sh | sh - Setup a new project
# uvx will execute `SEML` in a temporary virtual environment # and run it to setup your new project. uvx seml project init my_new_project
- Setup a virtual environment
cd my_new_project uv sync - Activate your virtual environment
source .venv/bin/activate - Configure
SEML:seml configure
When executing SEML make sure to always use the seml command from your project's virtual environment and only use uvx seml for high-level commands that do not affect experiments (like setting up new projects).
If you want to include SEML into existing projects, you can install it via:
pip install semlThen configure your MongoDB via:
seml configureIf your MongoDB is only accessible via an SSH port forward, SEML allows you to directly configure this as well if you install the ssh_forward dependencies via:
pip install seml[ssh_forward]It remains to configure the SSH settings:
seml configure --ssh_forwardFor development, we recommend uv which you can install via
curl -LsSf https://astral.sh/uv/install.sh | shSetup the right environment use and activate it:
uv sync --locked
source .venv/bin/activateAlternatively, you can install the repository in any Python environment via:
pip install -e .[dev]Make sure to install the pre-commit hooks via
pre-commit installDocumentation is available in our docs.md or via the CLI:
seml --helpSee our simple example to get familiar with how SEML works.
SEML supports command line completion. To install this feature run:
seml --install-completion {shell}If you are using the zsh shell, you might have to append compinit -D to the ~/.zshrc file (see this issue).
SEML should work with Slurm 18.08 and above out of the box. Version 17.11 and earlier do not have a SIGNALING job state, which you have to remove from the SLURM_STATES defined in SEML's settings (seml/settings.py). Earlier versions have not been tested and might have other issues.
Contact us at zuegnerd@in.tum.de, johannes.gasteiger@tum.de, or n.gao@tum.de for any questions.
When you use SEML in your own work, please cite the software along the lines of the following bibtex:
@software{seml_2023,
author = {Z{\"u}gner, Daniel and Gasteiger, Johannes and Gao, Nicholas and Dominik Fuchsgruber},
title = {{SEML: Slurm Experiment Management Library}},
url = {https://github.com/TUM-DAML/seml},
version = {0.4.0},
year = {2023}
}
Copyright (C) 2023 Daniel Zügner, Johannes Gasteiger, Nicholas Gao, Dominik Fuchsgruber Technical University of Munich