✨🤖 Update TransH to ER-Model

[mberr]

This PR updates TransH to new-style ERModel. Also updates the experiment configuration from JSON to YAML.

It also splits the TransH regularizer into its two components, and modernizes the tests.

Benchmark

FB15k

commit	training	evaluation
6e3eb43 (update-transh)	1.59s	14.78s (bs=512)
f5c6922 (master)	1.59s	14.71s (bs=1024)

WN18

commit	training	evaluation
6e3eb43 (update-transh)	0.52s	1.71s (bs=512)
f5c6922 (master)	0.51s	1.72s (bs=512)

FB15k

metadata:
  title: "Learn FB15K Dataset with TransH as described by Wang et al., 2014"
pipeline:
  dataset: "fb15k"
  model: "TransH"
  model_kwargs:
    predict_with_sigmoid: false
    embedding_dim: 100
    scoring_fct_norm: 2
    entity_initializer: "xavier_uniform"
    relation_initializer: "xavier_uniform"
  regularizer: "TransH"
  regularizer_kwargs:
    weight: 1.0
    epsilon: 1.0e-05
  optimizer: "SGD"
  optimizer_kwargs:
    lr: 0.005
  loss: "MarginRankingLoss"
  loss_kwargs:
    reduction: "mean"
    margin: 0.25
  training_loop: "sLCWA"
  negative_sampler: "bernoulli"
  negative_sampler_kwargs:
    num_negs_per_pos: 1
  training_kwargs:
    num_epochs: 5
    batch_size: 4800
  evaluator_kwargs:
    filtered: true

WN18

metadata:
  title: "Learn WN18 Dataset with TransH as described by Wang et al., 2014"
pipeline:
  dataset: "wn18"
  model: "TransH"
  model_kwargs:
    predict_with_sigmoid: false
    embedding_dim: 50
    scoring_fct_norm: 2
    entity_initializer: "xavier_uniform"
    relation_initializer: "xavier_uniform"
  regularizer: "TransH"
  regularizer_kwargs:
    weight: 0.25
    epsilon: 1.0e-05
  optimizer: "SGD"
  optimizer_kwargs:
    lr: 0.01
  loss: "MarginRankingLoss"
  loss_kwargs:
    reduction: "mean"
    margin: 1.0
  training_loop: "sLCWA"
  negative_sampler: "bernoulli"
  negative_sampler_kwargs:
    num_negs_per_pos: 1
  training_kwargs:
    num_epochs: 5
    batch_size: 4800
  evaluator_kwargs:
    filtered: true

Dependencies

• ✨🤖 Update RESCAL to ER-Model #952 - for default regularizer handling
• Add normalize_with_default cthoyt/class-resolver#40 - for external normalize_with_defaults

[cthoyt]

PR this into class resolver?

[mberr]

👍