Fast sLCWA

[mberr]

This PR aims at adding fast sLCWA training to PyKEEN.

For this, we exploit that negative samples are created by corruption, and thus share some parts of the triple, e.g., (h, r), with the positive triple. This allows saving computation. Such optimizations already exist is other packages, e.g.,

• DGL (and respective OGB), called, e.g., mode="tail-batch"
• LibKGE, called, e.g., combine="_sp_"

Tasks

• add score_{h,r,t}s to score multiple heads/relations/tails at once
• group samples with same e.g. (h, r) -> problem: variable number -> no easy batching

Dependencies:

• 📦 💾 Batch data loader #817

[cthoyt]

@mberr related to #943?

[mberr]

@mberr related to #943?

Only marginally. For Fast sLCWA we group within a batch which includes positive and negative triples, while in #943, we group only over positive ones but across all training triples.