config.yml

model:
  configuration: "ggnn"
  base:
    hidden_dim: 256
    dropout_rate: 0.1
    num_edge_types: 24
  evaluation:
    top: 1
    node_of_line: 7
    window_size: 9
  ggnn:
    time_steps: [3, 1, 3, 1]
    residuals:  # Note: keys must be strings for TF checkpointing
      "1": [0]
      "3": [0, 1]
    add_type_bias: true
  transformer:
    ff_dim: 2048
    num_layers: 6
    attention_dim: 512
    num_heads: 8
model_2:
  configuration: "transformer"
  base:
    hidden_dim: 256
    dropout_rate: 0.1
    num_edge_types: 24
  ggnn:
    time_steps: [3, 1, 3, 1]
    residuals:  # Note: keys must be strings for TF checkpointing
      "1": [0]
      "3": [0, 1]
    add_type_bias: true
  transformer:
    ff_dim: 2048
    num_layers: 6
    attention_dim: 512
    num_heads: 8
data:
  max_batch_size: 1250
  max_buffer_size: 100  # In terms of no. of (maximum efficiency) batches.
  max_node_size: 1024
  valid_interval: 5000 # 25000 for LOC pre-training, 50000 for CLS fine-tuning*, 2500 for LOC fine-tuning, 5000 for CLS fine-tuning
  max_valid_samples: 25000
  max_token_length: 10  # In terms of (BPE) sub-tokens.
  w2v_dimension: 256
training:
  max_steps: 50 # 10 for pre-training, 50 for fine-tuning
  print_freq: 500
  learning_rate: 0.00001 # 0.0001 is for pre-training, 0.00001 is for fine-tuning