data:
  block_size: 512
  duration: 1.5
  encoder: dpwavlmbase
  encoder_ckpt: models/pretrained/dphubert/DPWavLM-sp0.75.pth
  encoder_hop_size: 320
  encoder_out_channels: 768
  encoder_sample_rate: 16000
  extensions:
  - wav
  - flac
  - mp3
  - m4a
  f0_extractor: rmvpe
  f0_max: 1200
  f0_min: 65
  sampling_rate: 44100
  spk_embed_channels: 256
  spk_embed_encoder: pyannote.audio
  spk_embed_encoder_ckpt: ./models/pretrained/pyannote.audio/wespeaker-voxceleb-resnet34-LM/pytorch_model.bin
  spk_embed_encoder_sample_rate: 16000
  volume_window_size: 8
device: cuda
env:
  gpu_id: 0
loss:
  beta: 1.0
  fft_max: 2048
  fft_min: 128
  gamma: 0.0
  n_ffts:
  - 32
  - 64
  - 128
  - 256
  - 512
  - 1024
  - 2048
  overlap: 0.5
  use_multi_scale_log_freq: true
model:
  f0_input_variance: 0.0
  f0_offset_size_downsamples: 16
  harmonic_env_size_downsamples: 16
  no_use_embed_conv: false
  noise_env_size_downsamples: 16
  noise_seed: 289
  noise_to_harmonic_phase: false
  type: CombSubMinimumNoisedPhase
  units_hidden_channels: 256
  units_layers:
  - - 10
    - 11
  use_add_noise_env: false
  use_discriminator: true
  use_f0_offset: false
  use_harmonic_env: true
  use_noise_env: false
  use_speaker_embed: true
  win_length: 2048
train:
  accelerator:
    log_with: tensorboard
  accelerator_project_config:
    total_limit: 10
  allow_tf32: true
  amp_dtype: fp32
  batch_size: 32
  cache_all_data: true
  cache_device: cpu
  cache_fp16: true
  epochs: 100
  frame_hop_random_max: 64
  frame_hop_random_min: 32
  interval_log: 10
  interval_val: 2000
  loss_variation: 0.1
  low_similar_loss_variation: 0.7
  lr: 0.0003
  num_workers: 2
  only_u2c_stack: false
  save_states: true
  sched_cooldown: 2
  sched_factor: 0.5
  sched_gamma: 0.99999
  sched_min_lr: 1.0e-05
  sched_patience: 50
  sched_threshold: 1.0e-05
  sched_threshold_mode: rel
  weight_decay: 1.0e-05