edolab 0.0.5

A command line tool for running experiments with `edo`.

edolab

A command line tool for running experiments with edo.

Installation

edolab is pip-installable:

$ python -m pip install edolab

Usage

Experiment scripts

To use edolab, you will need to write a Python script configuring the parameters of your experiment.

Required parameters

• fitness: A function that takes (at least) an edo.Individual instance to be used as the fitness function by edo
• distributions: A list of edo.distribution.Distribution subclasses that will be used to create the edo.Family instances for edo
• Variable assignments for all of the essential arguments in edo.DataOptimiser except for families

Optional parameters

• root: A directory to which data should be written (and summarised from)
• processes: A number of processes for edo to use when calculating population fitness
• Custom column distribution classes should be defined in the script
• If you wish to use a custom stop or dwindle method then define a subclass of edo.DataOptimiser and assign that class to a variable called optimiser
• Any keyword arguments to pass to fitness or the stop and dwindle methods should be assigned to the corresponding <func>_kwargs variable.

An example of such a script would be something like this:

""" /path/to/experiment/script.py """

import edo
import numpy as np
from edo.distributions import Uniform


class CustomOptimiser(edo.DataOptimiser):
    """ This is an optimiser with custom stopping and dwindling methods. """

    def stop(self, tol):
        """ Stop if the median fitness is less than `tol` away from zero. """

        self.converged = abs(np.median(self.pop_fitness)) < tol

    def dwindle(self, rate):
        """ Cut the mutation probability in half every `rate` generations. """

        if self.generation % rate == 0:
            self.mutation_prob /= 2


def fitness(individual, size, seed=0):
    """ Randomly sample `size` values from an individual and return the
    minimum. """

    np.random.seed(seed)
    values = individual.dataframe.values.flat
    sample = np.random.choice(values, size=size)
    return min(sample)


class NegativeUniform(Uniform):
    """ A copy that only takes negative values. """

    name = "NegativeUniform"
    param_limits = {"bounds": [-1, 0]}
    hard_limits = {"bounds": [-100, 0]}


size = 5
row_limits = [1, 5]
col_limits = [1, 2]
max_iter = 3
best_prop = 0.5
mutation_prob = 0.5

Uniform.param_limits["bounds"] = [0, 1]

distributions = [Uniform, NegativeUniform]
optimiser = CustomOptimiser

fitness_kwargs = {"size": 3}
stop_kwargs = {"tol": 1e-3}
dwindle_kwargs = {"rate": 10}

For more details on the parameters of edo, see its documentation: https://edo.readthedocs.io

Running the experiment

Then, to run an experiment with this script do the following:

$ edolab run /path/to/experiment/script.py

Summarising the experiment

And to summarise the data (for easy transfer):

$ edolab summarise /path/to/experiment/script.py

For further details on the commands, use the --help flag on the run and summarise commands.

A note on reproducibility

It is highly recommended that you use a virtual environment when using edo in or outside of this command line tool as edo uses pickle to store various objects created in a run that may not be retrievable with a different version of Python.

Contributing

This tool has been made to be pretty bare and could use some padding out. If you'd like to contribute then make a fork and clone the repository locally:

$ git clone https://github.com/<your-username>/edolab.git

Install the package and replicate the conda environment (or install the development dependencies manually):

$ cd edolab
$ python setup.py develop
$ conda env create -f environment.yml
$ conda activate edolab-dev

Make your changes and write tests to go with them, ensuring they pass:

$ python -m pytest --cov=edolab --cov-fail-under=100 tests

Commit, push to your fork and open a pull request!