CS224W - Adds TransD KGE, and Bernoulli corruption strategy for all KGE

[mattjhayes3]

Implements TransD and Bernoulli corruption strategy (used in TransD and TransH papers).

Details

• Did not see any PRs for these ones yet :)
• Used the KGEModel base and tried to keep consistent with other KGE
• Tried to implement everything as efficiently as possible while following the paper
• Incorporated into examples/kge_fb15k_237.py, also adding duration calculation there
• The Bernoulli parameters are currently computed for each batch rather than the whole training set
  • This is a little slower but seemed to work much better in terms of metrics
  • Also seems a little more intuitive for the user to simply flip a bool as opposed to requiring they pass their whole training set for initialization, or requiring they use loader().
  • We might also try cumulating the statistics during training, but these initial experiments suggest this would be the worst of both worlds (slower as we won't know when first epoch has ended, ultimately similar to precomputing in terms of metrics)
• Happy to split Bernoulli into a separate PR, or make any other edits you might prefer.

Benchmarks

I was only able to compare it to TransE on examples/kge_fb15k_237.py and do 3 runs each, but even without any hyperparameter tuning, findings seem fairly consistent with paper, i.e. significant improvements in all metrics [colab]:

Method/Evalset	Mean Rank	MRR	Hits@10
TransD Bernoulli Val	176.19 ± 0.74	0.223 ± 0.000	0.408 ± 0.001
TransD Bernoulli Test	180.65 ± 0.71	0.220 ± 0.001	0.401 ± 0.002
TransD Val	180.08 ± 0.37	0.211 ± 0.003	0.400 ± 0.002
TransD Test	184.06 ± 1.78	0.210 ± 0.003	0.396 ± 0.002
TransE Val	258.74 ± 5.76	0.221 ± 0.000	0.366 ± 0.004
TransE Test	268.30 ± 6.39	0.217 ± 0.001	0.362 ± 0.004