experiments.conf

basic {
  data_dir = /xxx/data_dir  # Edit this
  dataset_dir = ${basic.data_dir}/dataset
  log_root = ${basic.data_dir}
  use_amp = true
}

#*************** Data-specific config ***************

data_full = ${basic} {
  dataset_name = full_seed100

  only_title = false
  emb_similarity = 0.95
  similarity_layer = -2

  num_epochs = 3
  report_frequency = 100
  eval_frequency = 400
  start_eval_after_ratio = 0
}

#*************** Transformer encoder (or encoder-decoder) config ***************

segment_encoder {
  max_seq_len = 80
}

bert_large = ${segment_encoder} {
  model_type = bert
  pretrained = bert-large-uncased
  tokenizer = bert-large-uncased
}

#*************** Model config ***************

model {
  # Only for tagging-based
  do_tag = false
  opentag = false

  # Steps
  batch_size = 128
  eval_batch_size = 128
  gradient_accumulation_steps = 1
  warmup_ratio = 0.01
  # Optimizer
  bert_learning_rate = 2e-5
  task_learning_rate = 1e-4
  bert_wd = 1e-2
  task_wd = 0
  adam_eps = 1e-8
  max_grad_norm = 1  # Set 0 to disable clipping
  freeze_layers = -4  # to freeze: > 0 or < 0

  # Seed expansion
  expand_by_sim = false

  # Model hyperparameters
  dropout_rate = 0.1
  max_itr = 1
  num_views = 1
  baseline_rpl = false
  baseline_triplet_coef = false
  baseline_triplet_channels = 8
  baseline_triplet_margin = 1
  contrastive_temp = 0.1
  attr_cls_coef = false
  attr_cls_th = 5

  # Regularization
  vi_coef = false
  vi_diverse = false
  vi_only_ngram = true
  vi_asin_bow = false  # If true, use bow; false: use CLS
  vi_no_kl = false
  num_topics = 200
  vi_sampling = 4
  usp_bullet_coef = false
  usp_bullet_temp = 1

  # Gradual clustering
  cluster_sim_relax = 0.8
  dbscan_eps = 0.05
  dbscan_min_samples = 4

  # Baseline clustering
  baseline_clustering = false
}

#*************** RE experiment config ***************

exp_full_bert_large = ${data_full} ${model} ${bert_large} {
}

exp_full_bert_large_v4 = ${exp_full_bert_large} {
  num_topics = 50
  vi_coef = 0.02
  usp_bullet_coef = 0.01
}

exp_full_bert_large_bs_v6 = ${exp_full_bert_large_bs} {
  attr_cls_coef = 1
  baseline_triplet_margin = 1
  baseline_triplet_coef = 2
}

exp_full_bert_large_tag = ${exp_full_bert_large} {
  do_tag = true
  use_crf = false
  num_epochs = 1
  start_eval_after_ratio = 10
}

exp_full_bert_large_tag_v3 = ${exp_full_bert_large_tag} {
  use_crf = true
}

exp_full_bert_large_opentag = ${exp_full_bert_large} {
  do_tag = true
  opentag = true
  use_crf = true
  num_epochs = 1
  batch_size = 96
  eval_batch_size = 64
  start_eval_after_ratio = 10
}

exp_full_chunk_bert_large = ${exp_full_bert_large} {
  dataset_name = full_seed100chunk
}

exp_chunk_no_rpl_gradual = ${exp_full_chunk_bert_large} {
  dbscan_eps = 0.1
}

exp_no_rpl_baseline = ${exp_full_bert_large} {
  baseline_clustering = true
  dbscan_eps = 0.1
}

exp_no_rpl_gradual = ${exp_full_bert_large} {
  dbscan_eps = 0.1
}