rotate-wikidata5m-parallel-stratification-ARL-2@1.yaml

dataset:
  name: wikidata5m
entity_ranking:
  chunk_size: 5000
eval:
  batch_size: 256
import:
- rotate
- distributed_model
distributed_model:
  base_model:
    type: rotate
job:
  device_pool:
  - cuda:0
  - cuda:1
  distributed:
    entity_sync_level: partition
    num_partitions: 8
    num_workers: 4
    parameter_server: shared
    partition_type: stratification
    combine_mirror_blocks: False
lookup_embedder:
  dim: 128
  initialize: xavier_uniform_
  initialize_args:
    normal_:
      mean: 0.0
      std: 3.0258816720414108e-05
    uniform_:
      a: -0.5471612330226786
    xavier_normal_:
      gain: 1.0
    xavier_uniform_:
      gain: 1.0
  regularize: ''
  regularize_args:
    weighted: true
  sparse: true
model: distributed_model
modules:
- kge.model.embedder
- kge.model
- kge.job
negative_sampling:
  implementation: batch
  num_samples:
    o: 905
    s: 878
  sampling_type: pooled
  shared: true
  shared_type: naive
rotate:
  entity_embedder:
    dropout: 0.2189469924196601
    regularize_weight: 0.022287069546013984
  relation_embedder:
    dropout: -0.3986804271116853
    regularize_weight: 2.605672850011217e-05
search:
  on_error: continue
train:
  auto_correct: true
  batch_size: 512
  loss_arg: 1.0
  lr_scheduler: ReduceLROnPlateau
  lr_scheduler_args:
    factor: 0.95
    mode: max
    patience: 1
    threshold: 0.0001
  max_epochs: 300
  num_workers: 4
  optimizer:
    default:
      args:
        lr: 0.10233579996730208
      type: dist_adagrad
  type: distributed_negative_sampling
valid:
  early_stopping:
    threshold:
      epochs: 10
      metric_value: 0.1