alglab 0.7.0

A framework for running experiments on algorithms.

AlgLab

Framework for performing experiments with algorithms in Python.

This package provides a framework for building extendable and reproducible experiments with algorithms in python. This high-level workflow is to

1. Specify a dataset on which to perform an experiment.
2. Implement the algorithms which are to be compared.
3. Implement any evaluation methods by which the algorithm outputs will be scored.
4. Specify the experimental setup, including dataset and algorithm parameters.
5. Run the experiments, and
6. Analyse the results.

This preserves a separation between the dataset, algorithm, evaluation methods, and experimental setup. This makes it easy to extend and modify experimental setups by changing the datasets, adding new algorithms, or changing parameters.

Example

Here's a simple example of the workflow using datasets and evaluation functions already provided by alglab.

import alglab

# First, implement the algorithms that you would like to compare.
# Note that the signature of the implemented algorithms should take a dataset as the first argument,
# followed by the algorithm parameters as keyword arguments, with default values.
def kmeans(data: alglab.dataset.PointCloudDataset, k=10):
    sklearn_km = KMeans(n_clusters=k)
    sklearn_km.fit(data.data)
    return sklearn_km.labels_

def spectral_clustering(data: alglab.dataset.PointCloudDataset, k=10):
    sklearn_sc = SpectralClustering(n_clusters=k)
    sklearn_sc.fit(data.data)
    return sklearn_sc.labels_

# Configure the experiments. As well as the algorithms, we specify which dataset class to use,
# and the parameters for the algorithms and dataset.
#
# We also specify any functions which should be used to evaluate the algorithms, and give a
# filename in which to store the results.
experiments = alglab.experiment.ExperimentalSuite(
    [kmeans, spectral_clustering],
    alglab.dataset.TwoMoonsDataset,
    "results/twomoonsresults.csv",
    parameters={
        "k": 2,
        "dataset.n": np.linspace(1000, 5000, 6).astype(int),
    },
    evaluators=[alglab.evaluation.adjusted_rand_index],
)

def main():
    # Run the experiments
    experiments.run_all()

    # Now, we can visualise the results
    results = alglab.results.Results("results/twomoonsresults.csv")
    results.line_plot("n", "running_time_s")
    results.line_plot("n", "adjusted_rand_index")