input_dfm.yaml

general:
  use_gpu: True    # only if available
  # specify CUDA device, these are 0-indexed, e.g., 
  #   cuda:0, cuda:1 or others. "cuda" is the default CUDA device
  gpu_device: "cuda"
  # Parent dir to save trained models 
  models_dir: "./models"

preprocessing:
  # normalizing text to the ASCII encoding standard
  uni2ascii: True
  lowercase: True
  # removing both the leading and the trailing empty characters
  strip: True
  only_latin_letters: False
  # Accepted proportion of characters in a string that are not present in our vocabulary, i.e.,
  # String is accepted if:
  # (number of its characters found in the vocabulary)/(total number its characters) >= missing_char_threshold
  missing_char_threshold: 0.5
  # read a list of characters and add to the vocabulary
  read_list_chars: "./inputs/characters_v001.vocab"

# --- RNN/GRU/LSTM architecture/misc info
gru_lstm:
  main_architecture: "gru"    # rnn, gru, lstm
  mode:    # Tokenization mode
    # choices: "char", "ngram", "word"
    # for example: tokenize: ["char", "ngram", "word"] or ["char", "word"] 
    tokenize: ["char", "ngram"]
    # ONLY if "char" or "ngram" are slected in tokenize, the following arg will be used:
    # Strings in the inputs will be prefix + string + suffix:
    # NOTE: please use only STRINGS in prefix_suffix list, 
    #       if you want only prefix or suffix, enter the other as an empty string ""
    #       e.g., ["|", ""] would add | as prefix and no suffix
    prefix_suffix: ["<", ">"]
    # ONLY if "ngram" is selected in tokenize, the following args will be used:
    min_gram: 2
    max_gram: 3
    # ONLY if "word" is selected in tokenize, the following arg will be used:
    # Characters according to which the string will split into tokens:
    # Accepted inputs are either a string of delimiters (e.g. ", -!?:()" or
    # "default", in which case delimiters will be the white space and any
    # punctuation mark):
    token_sep: "default"
  bidirectional: True    # if True, becomes a bidirectional RNN/GRU/LSTM
  # num_layers
  # number of recurrent layers. e.g., setting num_layers=2 means stacking two 
  #   RNN/GRU/LSTMs together to form a stacked RNN/GRU/LSTM,
  #   with the second RNN/GRU/LSTM taking in outputs of the first RNN/GRU/LSTM and computing the final results.
  num_layers: 2
  # number of dimensions of the first fully connected network
  fc1_out_dim: 120    
  # pooling_mode:
  # hstates_layers_simple, hstates_layers, hstates
  # hstates_subtract, hstates_l2_distance, hstates_cosine
  # average, max
  # attention
  pooling_mode: 'hstates_layers_simple'    
  # rnn_dropout:
  # if non-zero, introduces a Dropout layer on the outputs of each RNN/LSTM/GRU layer except the last layer,
  #   with dropout probability equal to rnn_dropout.
  rnn_dropout: 0.01
  # fully-connected layers dropout depends on the number of fc layers (currently there are two)
  fc_dropout: [0.01, 0.01]  
  # attention layer dropout depends on the number of attention layers (currently there are two)
  att_dropout: [0.01, 0.01] 
  # Add bias to all learnable parameters
  bias: True

  rnn_hidden_dim: 60
  max_seq_len: 120
  embedding_dim: 60
  output_dim: 2

  learning_rate: 0.001    # 3e-4
  optimizer: adam
  epochs: 5
  batch_size: 256
  # shuffle when creating DataLoader
  dl_shuffle: True 
  random_seed: 123
  # Early stopping:
  # Number of epochs with no improvement after which training will be stopped and
  #   the model with the least validation loss will be saved
  # If 0 or negative, early stopping will be deactivated
  early_stopping_patience: -1

  # if -1 or 1, perform the validation step in every epoch; 
  # if 0, no validation will be done
  # otherwise, specify the interval (integer)
  validation: 1
  # split dataset
  train_proportion: 0.7
  val_proportion: 0.15
  test_proportion: 0.15

  # False or path to a directory to create tensor-board
  #create_tensor_board: "./tb_gru_test" 
  create_tensor_board: False

  # Layers to freeze during fine-tuning
  layers_to_freeze: ["emb", "rnn_1", "attn"]

inference:
 # Output predictions and save the results in output_preds_file
 output_preds: True
 # either a path or "default"
 #   "default" saves the prediction output inside the model directory
 output_preds_file: "default"
 # change it to true to have Mean Average Precision as an eval metric. Note that this would have an impact on computational time
 eval_map_metric: False