💾💽 Change serialization format

[mberr]

This PR changes the serialization format of the triples factory to store the key components, triples (and label to ID mappings), in a compressed, but human-readable format. This allows easy inspection outside of PyKEEN.

For FB15k237, I saw that the total file size was also significantly reduced (from 6.9MiB on ef769a8 = master to 1.1MiB for bf47eb5)

Based upon #655 (comment)

[cthoyt]

@mberr is this covered by unit tests? Have you checked it's working on a variety of datasets?

[mberr]

@mberr is this covered by unit tests? Have you checked it's working on a variety of datasets?

pykeen/tests/test_triples_factory.py

Line 501 in 4084a1a

class TestUtils(unittest.TestCase):

comprises the serialization tests.

So far I only tested this for FB15k237, but I can try a few other tomorrow.

[mberr]

@mberr is this covered by unit tests? Have you checked it's working on a variety of datasets?

pykeen/tests/test_triples_factory.py

Line 501 in 4084a1a

class TestUtils(unittest.TestCase):

comprises the serialization tests.

So far I only tested this for FB15k237, but I can try a few other tomorrow.

The ten smallest datasets work all.

from docdata import get_docdata
from pykeen.datasets import dataset_resolver, get_dataset
from pykeen.triples.triples_factory import TriplesFactory


def _triples(d: str) -> int:
    return get_docdata(dataset_resolver.lookup_dict[d])["statistics"]["triples"]


dataset_list = sorted(dataset_resolver.lookup_dict, key=_triples)
for name in dataset_list[:10]:
    print(name)
    dataset = get_dataset(dataset=name)
    path = f"/tmp/{name}-temp"
    dataset.training.to_path_binary(path=path)
    TriplesFactory.from_path_binary(path=path)