20way1shot.yaml

name: omniglot_protonet 
data:
    data_path: ./data
    validation_portion: 0.25
    tasks_per_epoch_train: 100
    tasks_per_epoch_val: 1000
    train_workers: 8
    val_workers: 4

hyperparameters:
  learning_rate: 1.0e-3
  weight_decay: 0
  reduce_every: 200
  lr_gamma: 0.5
  global_batch_size: 2 # how many tasks to train before performing a meta-update
  val_batch_size: 2 # how many tasks to evaluate on
  # Meta-training
  num_classes_train: 60
  num_support_train: 1
  num_query_train: 5
  # Meta-test
  num_classes_val: 20 #n-way
  num_support_val: 1 #k-shot
  # Model
  img_resize_dim: 28 # input will be 1 x img_resize_dim x img_resize_dim
  hidden_dim: 64 # intermediate number of channels
  embedding_dim: 64 # embedding number of channels

resources:
  slots_per_trial: 2

searcher:
  name: single
  metric: loss
  smaller_is_better: true
  # Original paper trained for 10,000 epochs with a plateau stopping condition
  max_length:
    batches: 30000

entrypoint: model_def:OmniglotProtoNetTrial
min_validation_period:
  batches: 5000
checkpoint_policy: none