base_config:
  - configs/base.yaml

task_cls: training.acoustic_task.AcousticTask

dictionaries: {}
extra_phonemes: []
merged_phoneme_groups: []
datasets: []

vocoder: NsfHifiGAN
vocoder_ckpt: checkpoints/pc_nsf_hifigan_44.1k_hop512_128bin_2025.02/model.ckpt
audio_sample_rate: 44100
audio_num_mel_bins: 128
hop_size: 512            # Hop size.
fft_size: 2048           # FFT size.
win_size: 2048           # FFT size.
fmin: 40
fmax: 16000

binarization_args:
  shuffle: true
  num_workers: 0
augmentation_args:
  random_pitch_shifting:
    enabled: false
    range: [-5., 5.]
    scale: 0.75
  fixed_pitch_shifting:
    enabled: false
    targets: [-5., 5.]
    scale: 0.5
  random_time_stretching:
    enabled: false
    range: [0.5, 2.]
    scale: 0.75

binary_data_dir: 'data/opencpop/binary'
binarizer_cls: preprocessing.acoustic_binarizer.AcousticBinarizer
spec_min: [-12]
spec_max: [0]
mel_vmin: -14.
mel_vmax: 4.
mel_base: 'e'
energy_smooth_width: 0.06
breathiness_smooth_width: 0.06
voicing_smooth_width: 0.06
tension_smooth_width: 0.06

use_lang_id: false
num_lang: 1
use_spk_id: false
num_spk: 1
use_mix_ln: false
mix_ln_layer: [0, 2]
use_energy_embed: false
use_breathiness_embed: false
use_voicing_embed: false
use_tension_embed: false
use_key_shift_embed: false
use_speed_embed: false

diffusion_type: reflow
time_scale_factor: 1000
timesteps: 1000
max_beta: 0.02
enc_ffn_kernel_size: 3
use_rope: true
rope_interleaved: false
use_stretch_embed: true
use_variance_scaling: true
rel_pos: true
sampling_algorithm: euler
sampling_steps: 20
diff_accelerator: ddim
diff_speedup: 10
hidden_size: 384
backbone_type: 'lynxnet2'
backbone_args:
  num_channels: 1024
  num_layers: 6
  kernel_size: 31
  dropout_rate: 0.0
  use_conditioner_cache: true
  glu_type: 'atanglu'
main_loss_type: l2
main_loss_log_norm: false
schedule_type: 'linear'

# shallow diffusion
use_shallow_diffusion: true
T_start: 0.4
T_start_infer: 0.4
K_step: 400
K_step_infer: 400

shallow_diffusion_args:
  train_aux_decoder: true
  train_diffusion: true
  val_gt_start: false
  aux_decoder_arch: convnext
  aux_decoder_args:
    num_channels: 512
    num_layers: 6
    kernel_size: 7
    dropout_rate: 0.1
  aux_decoder_grad: 0.1

lambda_aux_mel_loss: 0.2

# train and eval
num_sanity_val_steps: 1
optimizer_args:
  optimizer_cls: modules.optimizer.muon.Muon_AdamW
  lr: 0.0006
  muon_args:
    weight_decay: 0.1
  adamw_args:
    weight_decay: 0.0
lr_scheduler_args:
  step_size: 5000
  gamma: 0.8
max_batch_frames: 50000
max_batch_size: 64
dataset_size_key: 'lengths'
val_with_vocoder: true
val_check_interval: 4000
num_valid_plots: 10
max_updates: 100000
num_ckpt_keep: 8
permanent_ckpt_start: 60000
permanent_ckpt_interval: 10000

finetune_enabled: false
finetune_ckpt_path: null

finetune_ignored_params:
  - model.fs2.encoder.embed_tokens
  - model.fs2.txt_embed
  - model.fs2.spk_embed
finetune_strict_shapes: true

freezing_enabled: false
frozen_params: []