mnick / scikit-kge Goto Github PK

It would be nice to see a more detailed example.
The README says:

N, M, xs, ys = load_data('path to data')

But how exactly do these parameters look like? Something like:

N = 12 
M = 20
xs = [(s1, o1, p1), (s2, o2, p2), ...]
ys = [1, -1, ...]

...with concrete examples would be nice

Dear @mnick ,

i really appriciate this package. However, it is somewhat difficult to fully wrap my head arround it.
It would help if you could:

• Create an example that solves a classic machine learning problem with sckit-kge, as it is commonly done.
• create a detailed documentation of all available algorithms and other functions
• explain more about the negative sampling you are talking about in the readme file.
• explain how to access the embeddings
• maybe support RDF Graph input
• make scikit-kge available at pypi

Best regards
Chris

The sample code in README.md has typo. I figured it out after seeing IndexError.

from skge import HolE, StochasticTrainer

# Load knowledge graph 
# N = number of entities
# M = number of relations
# xs = list of (subject, object, predicte) triples
# ys = list of truth values for triples (1 = true, -1 = false)
N, M, xs, ys = load_data('path to data')

# instantiate HolE with an embedding space of size 100
model = HolE((N, N, M), 100)

The order of (N, N, M) must be (N, M, N).

model = HolE((N, M, N), 100)

When I run setup.py, it says "setuptools module not found."

@mnick Hi, thanks for publishing your code!

The example in the README says that the input for trainer should be composed of "xs = list of (subject, object, predicte) triples".

But when digging into the implementation, it seems that this is not correlated with what the HolE model is expecting to see in the data:

ss, ps, os, ys = unzip_triples(xys, with_ys=True)

and then:

return np.sum(self.R[ps] * ccorr(self.E[ss], self.E[os]), axis=1)

see here

This seems to indicate that the HolE model (and I think that also TransE) expects the triples to be ordered by (subject, predicate, object), i.e, switched order of predicate and object compared with the readme, right?

Thanks!

Here: https://github.com/mnick/scikit-kge/blob/master/skge/base.py#L130
self.batch_size = np.ceil(len(xys) / self.nbatches)
self.batch_size = int(np.ceil(len(xys) / self.nbatches))

In the latest stable version of numpy (1.13.0) np.ceil() returns a float value not int.

Because of this error, my code was crashing with:

TypeError: slice indices must be integers or None or have an index method

Please make the change.

i can not find this package in the anaconda?if you have free time,please help me solve the issue！

How do you access the embedding vectors after the training of the model? The documentation is unclear on this matter.