omniopt2 6605

Automatic highly parallelized hyperparameter optimizer based on Ax/Botorch

OmniOpt2 - Hyperparameter Optimizer for SLURM-based Systems

OmniOpt2 is a tool designed to assist researchers, engineers, and data scientists with hyperparameter optimization on SLURM-based clusters, even though it works without it as well. It simplifies large-scale optimization tasks with built-in fault tolerance and flexibility. A graphical user interface (GUI) is available for command creation, accessible at OmniOpt2 GUI. For tutorials on configuration, exit codes, and debugging, visit OmniOpt2 Tutorials.

Main program

omniopt --partition=alpha --experiment_name=example --mem_gb=1 --time=60 \
    --worker_timeout=60 --max_eval=500 --num_parallel_jobs=500 --gpus=1 \
    --follow --run_program=$(echo 'echo "RESULT: %(param)"' | base64 -w0) \
    --parameter param range 0 1000 float

This command initiates OmniOpt2 and installs dependencies if not already installed. The parameter --run_program uses a Base64-encoded string to specify commands. It is recommended to use the GUI, though.

Plot Results

Generates visualizations, such as scatter and hex scatter plots. --min and --max adjust the plotted result value range.

Or, with --min and --max:

omniopt_plot --run_dir runs/example/0
omniopt_plot --run_dir runs/example/0 --min 0 --max 100

Run Tests (Developer Use Only)

The test suite simulates various scenarios, including handling faulty jobs and ensuring program resilience.

./tests/main

See the automated tests tutorial page for more details.

Install from pypi

This may not use the latest version.

pip3 install omniopt2

Install from repo

pip3 install -e git+https://github.com/NormanTUD/OmniOpt2.git#egg=OmniOpt2

Alternatively, it can be executed directly, as OmniOpt2 will install its dependencies automatically if required.

Error Codes

For common issues and exit codes, see the exit codes tutorial-page.

Autocompletions

Autocomplete files for zsh and bash are in .shells. Run

bash .shells/install

to install them.

Contributions

I'd be glad to see your contributions!

Old OmniOpt

The old OmniOpt version, based on HyperOpt, is not supported anymore. It is still available, though, at https://github.com/NormanTUD/LegacyOmniOpt.