options.yaml

exp:
  dir: logs/vrd/block
  resume: null
dataset:
  import: block.datasets.factory
  name: vrd
  dir: data/vrd
  debug: false
  train_split: train
  eval_split: val
  neg_ratio: 0.5
  batch_size: 64
  nb_threads: 8
  mode: predicate
model:
  name: default
  network:
    import: block.models.networks.factory
    name: vrd_net
    classeme_dim: 100
    nb_classeme: 101
    aggreg:
      type: cat
    aggreg_dropout: 0.5
    predictor:
      input_dim: 600
      dimensions:
      - 71
      activation: relu
      dropout: 0.0
    classeme:
      type: block
      input_dims:
      - 100
      - 100
      output_dim: 200
      mm_dim: 700
      chunks: 10
      rank: 5
      shared: false
    spatial:
      type: block
      input_dims:
      - 4
      - 4
      output_dim: 200
      mm_dim: 700
      chunks: 10
      rank: 5
      shared: false
    feature:
      type: block
      input_dims:
      - 2048
      - 2048
      output_dim: 200
      mm_dim: 700
      chunks: 10
      rank: 5
      shared: false
  criterion:
    import: block.models.criterions.factory
    name: vrd_bce
  metric:
    import: block.models.metrics.factory
    name: vrd_predicate
optimizer:
  import: block.optimizers.factory
  init: glorot
  name: Adamax
  lr: 0.003
  gradual_warmup_steps:
  - 0.5
  - 2.0
  - 4
  lr_decay_epochs:
  - 10
  - 20
  - 2
  lr_decay_rate: 0.25
engine:
  name: logger
  nb_epochs: 15
  debug: false
  print_freq: 10
  saving_criteria:
  - eval_epoch.predicate.R_50:max
misc:
  logs_name: null
  cuda: true
  seed: 1337
view:
- logs:train_epoch.loss+logs:eval_epoch.loss
- logs:train_epoch.predicate.R_50+logs:eval_epoch.predicate.R_50
- logs:train_epoch.predicate.R_100+logs:eval_epoch.predicate.R_100
- logs:train_epoch.accuracy_top1+logs:eval_epoch.accuracy_top1
- logs:train_epoch.accuracy_top5+logs:eval_epoch.accuracy_top5
- logs:train_epoch.lr