# Phase 1
P1 {
  use_cpu = false
  visdial_version = 1.0
  train_on_dense = false
  metrics_to_maximize = mrr

  # visdial data
  visdial_image_feats = data/visdial_img_feat.lmdb

  visdial_image_adj_matrices = data/img_adj_matrices
  visdial_question_adj_matrices = data/question_adj_matrices
  visdial_history_adj_matrices = data/history_adj_matrices

  visdial_train = data/visdial_1.0_train.json
  visdial_val = data/visdial_1.0_val.json
  visdial_test = data/visdial_1.0_test.json
  visdial_val_dense_annotations = data/visdial_1.0_val_dense_annotations.json

  visdial_train_09 = data/visdial_0.9_train.json
  visdial_val_09 = data/visdial_0.9_val.json
  visdial_test_09 = data/visdial_0.9_test.json

  visdialconv_val = data/visdial_conv.json
  visdialconv_val_dense_annotations = data/visdialconv_dense_annotations.json

  visdialvispro_val = data/vispro.json
  visdialvispro_val_dense_annotations = data/vispro_dense_annotations.json

  visdial_question_parse_vocab = data/parse_vocab.pkl

  # init
  start_path = ckpt/vdgr_visdial_v1.0_after_warmup_K2.ckpt
  model_config = config/bert_base_6layer_6conect.json

  # visdial training
  freeze_vilbert = false
  visdial_tot_rounds = 11
  num_negative_samples = 1
  sequences_per_image = 2
  batch_size = 8
  lm_loss_coeff = 1
  nsp_loss_coeff = 1
  img_loss_coeff = 1
  batch_multiply = 1
  use_trainval = false
  dense_loss = ce
  dense_loss_coeff = 0
  dataloader_text_only = false
  rlv_hst_only = false
  rlv_hst_dense_round = false

  # visdial model
  mask_prob = 0.1
  image_mask_prob = 0.1
  max_seq_len = 256
  num_options = 100
  num_options_dense = 100
  use_embedding = joint

  # visdial evaluation
  eval_visdial_on_test = true
  eval_batch_size = 1
  eval_line_batch_size = 200
  skip_mrr_eval = false
  skip_ndcg_eval = false
  skip_visdial_eval = false
  eval_visdial_every = 1
  eval_dataset = visdial # visdial_vispro  # choices = [visdial, visdial_conv, visdial_vispro ]

  continue_evaluation = false
  eval_at_start = false
  eval_before_training = false
  initializer = normal
  bert_cased = false

  # restore ckpt
  loads_best_ckpt = false
  loads_ckpt = false
  restarts = false
  resets_max_metric = false
  uses_new_optimizer = false
  sets_new_lr = false
  loads_start_path = false

  # logging
  random_seed = 42
  next_logging_pct = 1.0
  next_evaluating_pct = 50.0
  max_ckpt_to_keep = 1
  num_epochs = 20
  early_stop_epoch = 5
  skip_saving_ckpt = false
  dp_type = apex
  stack_gr_data = false
  master_port = 5122
  stop_epochs = -1
  train_each_round = false
  drop_last_answer = false
  num_samples = -1

  # predicting
  predict_split = test
  predict_each_round = false
  predict_dense_round = false
  num_test_dialogs = 8000
  num_val_dialogs = 2064
  save_score = false

  # optimizer
  reset_optim = none
  learning_rate_bert = 5e-6
  learning_rate_gnn = 2e-4
  gnn_weight_decay = 0.01
  use_diff_lr_gnn = true
  min_lr = 0
  decay_method_bert = linear
  decay_method_gnn = linear
  decay_exp = 2
  max_grad_norm = 1.0
  task_optimizer = adam
  warmup_ratio = 0.1

  # directory
  log_dir = logs/vdgr
  data_dir = data
  visdial_output_dir = visdial_output
  bert_cache_dir = transformers

  # keep track of other hparams in bert json
  v_gnn_edge_dim = 12 # 11 classes + hub_node
  q_gnn_edge_dim = 48 # 47 classes + hub_node
  num_v_gnn_layers = 2
  num_q_gnn_layers = 2
  num_h_gnn_layers = 2
  num_gnn_attention_heads = 4
  v_gnn_ids = [0, 1, 2, 3, 4, 5]
  t_gnn_ids = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]
}

# Phase 2
P2_CE = ${P1} {
  # basic
  train_on_dense = true
  use_trainval = true
  metrics_to_maximize = ndcg

  visdial_train_dense = data/visdial_1.0_train_dense.json
  visdial_train_dense_annotations = data/visdial_1.0_train_dense_annotations.json
  visdial_val_dense = data/visdial_1.0_val.json

  tr_graph_idx_mapping = data/tr_dense_mapping.json
  val_graph_idx_mapping = data/val_dense_mapping.json
  test_graph_idx_mapping = data/test_dense_mapping.json

  visdial_val = data/visdial_1.0_val.json
  visdial_val_dense_annotations = data/visdial_1.0_val_dense_annotations.json

  # data
  start_path = logs/vdgr/P1_K2_v1.0/epoch_best.ckpt
  rlv_hst_only = false

  # visdial training
  nsp_loss_coeff = 0
  dense_loss_coeff = 1
  batch_multiply = 10
  batch_size = 1

  # visdial model
  num_options_dense = 100

  # visdial evaluation
  eval_batch_size = 1
  eval_line_batch_size = 100
  skip_mrr_eval = true

  # training
  stop_epochs = 3
  dp_type = dp
  dense_loss = ce

  # optimizer
  learning_rate_bert = 1e-4
}

P2_LISTNET = ${P2_CE} {
  dense_loss = listnet
}