config.yaml

dataset:
  training: [
  ]

  validation: [
  ]
  noise: [
  ]
  
  speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"
  
  use_hdf5: True
  hdf5_flag: r
  validate: True

  workers: 4
  cache: True

  phones_range: [4, 512]
  duration_range: [1.0, 24.0]

  random_utterance: 1.0
  max_prompts: 3
  prompt_duration: 3.0

  sample_type: speaker

  tasks_list: ["tts"] # ["tts", "ns", "sr", "tse", "cse", "nse", "tts"]

models:
  _max_levels: 8
  _models:
  - name: "ar"
    size: "full"
    resp_levels: 1
    prom_levels: 2
    tasks: 8
    arch_type: "retnet"

  - name: "nar"
    size: "full"
    resp_levels: 3
    prom_levels: 4
    tasks: 8
    arch_type: "retnet"


hyperparameters:
  batch_size: 32
  gradient_accumulation_steps: 4
  gradient_clipping: 100
  
  optimizer: AdamW
  learning_rate: 1.0e-6
  
  scheduler_type: ""
  #scheduler_type: OneCycle
  #scheduler_params:
  #  cycle_first_step_size: 10_000
  #  cycle_first_stair_count: 10_000

  #  cycle_second_step_size: 15_000
  #  cycle_second_stair_count: 15_000

  #  decay_step_size: 5_000

  #  cycle_min_lr: 2.5e-4 # 1.0e-5
  #  cycle_max_lr: 2.5e-4 # 1.0e-4
  #  decay_lr_rate: 0.0

  #  cycle_min_mom: 0.90
  #  cycle_max_mom: 0.99
  #  decay_mom_rate: 0.0

evaluation:
  batch_size: 64
  frequency: 500
  size: 64
  
  steps: 300
  ar_temperature: 0.95
  nar_temperature: 0.25

trainer:
  iterations: 1_000_000
  
  save_tag: step
  save_on_oom: True
  save_on_quit: True
  save_frequency: 25

  keep_last_checkpoints: 2

  aggressive_optimizations: False

  load_state_dict: True
  strict_loading: False
  #load_tag: "9500"
  #load_states: False
  #restart_step_count: True
  
  gc_mode: None # "global_step"

  weight_dtype: bfloat16

  backend: deepspeed
  deepspeed:
    zero_optimization_level: 2
    use_compression_training: True

inference:
  use_vocos: True
  normalize: False

  weight_dtype: float32

bitsandbytes:
  enabled: False
  injects: True
  linear: True
  embedding: True