Source code for pykeen.triples.remix

# -*- coding: utf-8 -*-

"""Remixing and dataset distance utilities.

Most datasets are given in with a pre-defined split, but it's often not discussed
how this split was created. This module contains utilities for investigating the
effects of remixing pre-split datasets like :class`pykeen.datasets.Nations`.

Further, it defines a metric for the "distance" between two splits of a given dataset.
Later, this will be used to map the landscape and see if there is a smooth, continuous
relationship between datasets' splits' distances and their maximum performance.

from typing import List, Sequence

import click

from .splitting import normalize_ratios, split
from .triples_factory import CoreTriplesFactory, cat_triples

__all__ = [

[docs]def remix(*triples_factories: CoreTriplesFactory, **kwargs) -> List[CoreTriplesFactory]: """Remix the triples from the training, testing, and validation set. :param triples_factories: A sequence of triples factories :param kwargs: Keyword arguments to be passed to :func:`split` :returns: A sequence of triples factories of the same sizes but randomly re-assigned triples :raises NotImplementedError: if any of the triples factories have ``create_inverse_triples`` """ for tf in triples_factories: if tf.create_inverse_triples: raise NotImplementedError("The remix algorithm is not implemented for datasets with inverse triples") all_triples = cat_triples(*triples_factories) ratios = _get_ratios(*triples_factories) return [ triples_factories[0].clone_and_exchange_triples(triples) for triples in split(all_triples, ratios=ratios, **kwargs) ]
def _get_ratios(*triples_factories: CoreTriplesFactory) -> Sequence[float]: total = sum(tf.num_triples for tf in triples_factories) ratios = normalize_ratios([tf.num_triples / total for tf in triples_factories]) return ratios @click.command() @click.option("--trails", type=int, default=15, show_default=True) def _main(trials): import itertools as itt import numpy as np from tqdm import tqdm from pykeen.datasets import get_dataset n_comb = trials * (trials - 1) // 2 click.echo(f"Number of combinations: {trials} n Choose 2 = {n_comb}") for dataset_name in [ "nations", "umls", "kinships", "codexsmall", "wn18", ]: reference_dataset = get_dataset(dataset=dataset_name) remixed_datasets = [reference_dataset.remix(random_state=random_state) for random_state in range(trials)] similarities = [ a.similarity(b) for a, b in tqdm( itt.combinations(remixed_datasets, r=2), total=n_comb, desc=dataset_name, ) ] click.echo(f"[{dataset_name}] Similarities Mean: {np.mean(similarities):.3f}") click.echo(f"[{dataset_name}] Similarities Std.: {np.std(similarities):.3f}") click.echo(f"[{dataset_name}] Relative Std.: {np.std(similarities) / np.mean(similarities):.3%}") if __name__ == "__main__": _main()