config.yaml

model_name: GTSinger

float32_matmul_precision: high
random_seed: 123456

extra_phonemes: [ ] # Not yet enabled.
silent_phonemes:
  - SP
  - <SP>
  - ''
  - pau
  - cl
  - CL
  - AP
  - EP
  - GS

language_prefix: false # When enabled, automatically add language prefixes for dictionary keys and isolate multiple dictionaries.
# Word /tab phoneme_1 /space phoneme_2
dictionaries:
  zh: dictionaries/opencpop-extension.txt
  ja: dictionaries/japanese_dict_full.txt
  en: dictionaries/ds_cmudict-07b.txt
#  yue: dictionaries/jyutping_dict.txt

datasets_config_paths: # datasets
  - configs/GTSinger_train.yaml

binary_folder: GTSinger_train/binary_fa # Preprocessing Results Folder
valid_set_size: 10 # When valid_set_Size is greater than 0, randomly select by quantity; otherwise, use the validation set in the dataset
max_length: 30 # Maximum duration of a single audio stream

# It is not recommended to enable the number of preprocessing threads and the size of the temporary queue if the data volume is too small (within 5 hours).
# Every 6-8g of memory and 6g of video memory can open one worker.
multiprocess_works: 0
multiprocess_max_size: 200  # size * num_works. Too large to ensure sufficient memory
multiprocess_start_size: 100  # Minimum startup data volume

# Number of training data loading processes
dataloader_workers: 6
dataloader_prefetch_factor: 4
batch_max_length: 200
binning_length: 500
drop_last: False

num_valid_plots: 20 # Maximum drawing quantity for val and evaluate items
draw_evaluate: True

augmentation_args: # Code defect, prohibited from use
  enabled: false
  random_pitch_shifting:
    range: [ -5., 5. ]
    num: 0
  blank_padding:
    range: [ 0, 5 ] # seconds
    num: 2

fa_arg:
  hidden_dims: 192
  down_sampling_factor: 2
  down_sampling_times: 3
  channels_scaleup_factor: 1.5
  dropout: 0.1

  curves_attention_dropout: 0.1

optimizer_config:
  lr: 0.0003
  gamma: 0.9999
  total_steps: 40000
  muon_args:
    weight_decay: 0.1
  adamw_args:
    weight_decay: 0.0

loss_config:
  losses:
    weights: [ 8.0, 0.1, 1.0, 6.0, 5.0 ]
    enable_RampUpScheduler: [ False,False,False,True,False ]
  function:
    num_bins: 10
    alpha: 0.999
    label_smoothing: 0.08

# trainer
accelerator: auto
devices: auto # num_devices
precision: bf16-mixed # bf16-mixed , 32-true
gradient_clip_val: 0.5
gradient_clip_algorithm: norm
val_check_interval: 198

save_top_k: 5
save_every_steps: 198

# Normally not changed
hubert_config:
  encoder: cn_hubert
  model_path: dependencies/chinese-hubert-base
  sample_rate: 16000
  hop_size: 320
  channel: 768

mel_spec_config:
  n_mels: 128
  sample_rate: 44100
  window_size: 882
  hop_size: 441
  n_fft: 1764
  f_min: 40
  f_max: 16000
  clamp: 0.00001

# Not recommended to open
merged_phoneme: false
# merge phonemes
merged_phoneme_groups:
  - - zh/f
    - en/f
  - - zh/m
    - en/m
    - ja/m
  - - zh/w
    - en/w
    - ja/w
  - - zh/s
    - en/s
    - ja/s
  - - zh/n
    - en/n
    - ja/n
  - - zh/c
    - ja/ts
  - - zh/y
    - en/y
    - ja/y
  - - zh/l
    - en/l
  - - zh/x
    - ja/sh
  - - ja/z
    - en/z