defaults:
  # - nn_params: nb8_embSum_NMT
  # - nn_params: remi8
  - nn_params: nb8_embSum_diff_t2m_150M_finetunning
    # - nn_params: nb8_embSum_diff_t2m_150M_pretraining
  # - nn_params: nb8_embSum_subPararell
  # - nn_params: nb8_embSum_diff_t2m_150M

    #  - nn_params: nb8_embSum_subFeedForward
    # -  nn_params: nb8_embSum_diff
  # nn_params: nb8_SA_diff
  # - nn_params: nb8_embSum_diff_main12head16dim512_ave
  # - nn_params: nb8_embSum_NMT_main12_head_16_dim512
  # - nn_params: remi8_main12_head_16_dim512
    # - nn_params: nb5_embSum_diff_main12head16dim768_sub3

dataset: SOD  # Pop1k7, Pop909, SOD, LakhClean,PretrainingDataset FinetuneDataset
captions_path: dataset/midicaps/train_set.json

# dataset:  SymphonyNet_Dataset # Pop1k7, Pop909, SOD, LakhClean
# captions_path: dataset/symphonyNet/syd-caption.json

use_ddp: True # True, False | distributed data parallel
use_fp16: True # True, False | mixed precision training
use_diff: True # True,use diffusion in subdecoder
diff_steps: 8 # number of diffusion steps
use_dispLoss: True
lambda_weight: 0.5
tau: 0.5

train_params:
  device: cuda
  batch_size: 3
  grad_clip: 1.0
  num_iter: 300000  # total number of iterations
  num_cycles_for_inference: 10 # number of cycles for inference, iterations_per_validation_cycle * num_cycles_for_inference
  num_cycles_for_model_checkpoint: 1 # number of cycles for model checkpoint, iterations_per_validation_cycle * num_cycles_for_model_checkpoint
  iterations_per_training_cycle: 10 # number of iterations for logging training loss
  iterations_per_validation_cycle: 5000 # number of iterations for validation process
  input_length: 3072 # input sequence length3072
  # you can use focal loss, it it's not used, set focal_gamma to 0
  focal_alpha: 1
  focal_gamma: 0
  # learning rate scheduler: 'cosinelr', 'cosineannealingwarmuprestarts', 'not-using', please check train_utils.py for more details
  scheduler : cosinelr
  initial_lr: 0.00005
  decay_step_rate: 0.8 # means it will reach its lowest point at decay_step_rate * total_num_iter
  num_steps_per_cycle: 20000 # number of steps per cycle for 'cosineannealingwarmuprestarts'
  warmup_steps: 2000 #number of warmup steps
  max_lr: 0.00015
  gamma: 0.6 # the decay rate for 'cosineannealingwarmuprestarts'
  # Distributed Data Parallel
  world_size: 5 # 0 means no distributed training
  gradient_accumulation_steps: 4  # 1 means no gradient accumulation
inference_params:
  num_uncond_generation: 1 # number of unconditional generation
  num_cond_generation: 3 # number of conditional generation
data_params:
  first_pred_feature: pitch # compound shifting for NB only, choose the target sub-token (remi and cp are not influenced by this argument)
  split_ratio: 0.998 # train-validation-test split ratio
  aug_type: pitch # random, null | pitch and chord augmentation type
general:
  debug: False
  make_log: True # True, False | update the log file in wandb online to your designated project and entity
  infer_and_log: True # True, False | inference and log the results