wikikg2_ax_search.yaml

# This is the hyperparameter search config file of wikikg2.
# We choose the blm structure from the previous study(https://github.com/AutoML-Research/AutoSF). 

job.type: search
dataset.name: wikikg2
model: autoblm
search.type: ax
valid.metric: mean_reciprocal_rank

ax_search:
  num_sobol_trials: 20
  num_trials: 100  
  parameters: 
    - name: 1vsAll.relation_prediction_weight
      type: range
      bounds: [0.0, 0.3]
    - name: lookup_embedder.regularize_weight
      type: range
      bounds: [1.0e-7, 3.0e-6]
      log_scale: true
    - name: train.optimizer.default.args.lr
      type: range
      bounds: [0.15, 0.6]
      log_scale: true

1vsAll:
  relation_prediction: true

autoblm:
  A: [0, 1, 0, 0, 2, 0, 0, 0, 0, 0, -3, 4, 1, 0, 4, 0]
lookup_embedder:
  dim: 128
  dropout: 0.1

train:
  batch_size: 512
  optimizer.default:
    type: Adagrad
  type: 1vsAll
  max_epochs: 5   # We use little max epoch to make full use of time budget, which may influence results a little.