comb-py 0.0.3

A minimal package for configuring and exploring parameter combinations of machine learning experiments.

comb is a small and simplistic library. comb is able to

• quickly and cleanly define grid searches and random searches, and handle parameter dependencies
• add configuration options to existing projects. Existing modules and packages can be turned into a registry, enabling python classes to become accessible and searchable via a string name.

For more complicated workflows, configuration packages like hydra are better suited.

Installation

Install comb-py from pypi:

$ pip install comb-py

and import it via

import comb

comb does not have any dependencies beyond the python standard library. It works for python>=3.7.

Defining an experiment (comb.sweep)

In your project, create sweeps directly in python ─ create the following file under sweep/my_first_experiment.py

from comb import sweep
from comb.types import zipped, grid

@sweep.register("example-sweep")
class MyExperiment(sweep.Sweep):

    @property
    def script(self):
        return "my/exp.py"

    def get_random_args(self):
        return dict(
            # define a method to sample arguments from
            foo = np.random.choice([42, 73])
        )

    def get_fixed_args(self):
        return dict(
            # zip N dependen arguments together
            bar = zipped("hello", "check out"),
            baz = zipped("world", "comb"),
            # define a search grid (1x2 combinations)
            # over two parameters
            blubb  = grid("star"),
            blubb2 = grid("wars", "treck"),
        )

and generate a joblist using

$ python -m comb example-sweep
exp.py --bar hello --baz world --blubb star --blubb2 wars --foo 73
exp.py --bar hello --baz world --blubb star --blubb2 treck --foo 73
exp.py --bar check out --baz comb --blubb star --blubb2 wars --foo 73
exp.py --bar check out --baz comb --blubb star --blubb2 treck --foo 73

Parametrizing an experiment (comb.registry)

comb makes it very easy to reference design choices within your experiment by names. Suppose you wanted to add a few datasets and loss functions to a machine learning experiment.

Turn your python module packages or packages into registries by a simple call to comb.registry.add_helper_functions:

# datasets.py
from comb import registry
registry.add_helper_functions(__name__)

@register("mnist")
class MNIST(): pass 

@register("svhn")
class SVHN(): pass 


# loss_functions.py
from comb import registry
registry.add_helper_functions(__name__)

@register("mse")
class MeanSquaredError(): pass

@register("infonce")
class InfoNCE(): pass

Afterwards, you can easily list and instantiate your functions:

>>> import datasets
>>> datasets.get_options("*")
mnist svhn
>>> datasets.init("mnist")
MNIST()

Scheduling experiments

comb does not attempt to provide ways to actually launch these experiments ─ there are plenty tools better suited for this. To name a few suggestions, the following workflows are possible:

Using GNU parallel

Scheduling a maximum of 2 consecutive jobs via GNU parallel (similar results can be achived via e.g. xargs):

$ python -m comb bash-example || exit 1 | parallel --jobs 2 'echo Scheduling job {#}; eval {}'

Using SLURM

Scheduling a job array via slurm:

mkdir -p submitted
python -m comb bash-example > jobs.lst
num_jobs=$(wc -l jobs.lst)
jobid=$(sbatch -a 1-${num_jobs} --wrap 'job=$(sed -n ${SLURM_ARRAY_TASK_ID}p jobs.lst); srun ${job}')
mv jobs.lst submitted/{jobid}.lst 

License

comb is released under an MIT License.