dobraczka / kiez Goto Github PK

Using predefined strings etc. limits extendability. Enabling objects as arguments enhances the possibility for users to use their preferred (A)NN libraries, and own hubness reduction methods.

Use single-source for version setting in kiez/__init__.py

Using single sources is a bit awkward since Kiez.fit() expects source and target. See #7 for an example use-case.

Annoy and NGT create permanent indices in the file system. At the moment it is the user's responsibility to remove these. It would be nice to provide some function, that could be called in order to remove such files easily, when they are not needed anymore.

If I have two different embedding spaces describing the same entities, like if I train two models on the same dataset in PyKEEN, how can I use Kiez to assess how good they correspond? Or maybe there's a notion of how "good" the Kiez fit is?

A naive idea is I could I iterate through each entity and calculate the overlap coefficient of the nearest neighbors in both embedding spaces, then maybe report the average overlap coefficient. I'm sure I could come up with a few things like this, but I bet you know better! Any ideas appreciated.

I would start with code like this:

from pykeen.pipeline import pipeline
from pykeen.datasets import Nations

dataset = Nations()

# Train the same dataset with two different models
r1 = pipeline(
    model='TransE',
    dataset=dataset,
    epochs=1,  # change this to ~25 for real usage on Nations
)

r2 = pipeline(
    model='PairRE',
    dataset=dataset,
    epochs=1,  # change this to ~25 for real usage on Nations
) 

from kiez import Kiez

k_inst = Kiez()
k_inst.fit(
    r1.model.entity_representations[0]().detach().numpy(),
    r2.model.entity_representations[0]().detach().numpy(),
)

# How do I assess how well these spaces correspond? Is there a metric for how "good" the fit is?

NGT is not available for Python 3.11, Faiss installation maybe needs to be adapted.

Candidates are:

• https://github.com/facebookresearch/faiss
• https://github.com/js1010/cuhnsw

• Explain the new possibilities enabled by class-resolver
• Incorporate architecture picture with explanation
• link to readthedocs in readme
• Give new examples (including for single-source)