Configs/config_ft_single.yml

# ─── GLOBAL ──────────────────────────────────────────────────────────
log_dir: logs/pod_90h_30k_second_v2
device: "cuda"

batch_size: 12        # 40 GB A100, fp16
max_len: 300          # ≈ 8 s (200 × 40 ms)

epochs_1st: 25        # first-stage schedule
epochs_2nd: 20        # second-stage schedule (later)
save_freq: 1
log_interval: 50

# leave blank on first run
pretrained_model: "/workspace/styletts2/logs/pod_90h_30k_second_v2/epoch_2nd_00005.pth"
second_stage_load_pretrained: true
load_only_params: false

# ─── PRE-PROCESS ─────────────────────────────────────────────────────
preprocess_params:
  sr: 24000
  spect_params:        # required by Mel extractor
    n_fft: 2048
    win_length: 1200
    hop_length: 300

# ─── DATA ────────────────────────────────────────────────────────────
data_params:
  root_path:  /workspace
  train_data: /workspace/styletts2/data/train_list.txt
  val_data:   /workspace/styletts2/data/val_list.txt
  min_length: 50    # sample until texts with this size are obtained for OOD texts        
  OOD_data:   /workspace/styletts2/data/OOD_texts.txt

# ─── LOSS SCHEDULE ──────────────────────────────────────────────────
loss_params:
    lambda_mel: 5. # mel reconstruction loss
    lambda_gen: 1. # generator loss
    lambda_slm: 1. # slm feature matching loss
    
    lambda_mono: 1. # monotonic alignment loss (1st stage, TMA)
    lambda_s2s: 1. # sequence-to-sequence loss (1st stage, TMA)
    TMA_epoch: 14 # TMA starting epoch (1st stage)

    lambda_F0: 1. # F0 reconstruction loss (2nd stage)
    lambda_norm: 1. # norm reconstruction loss (2nd stage)
    lambda_dur: 1. # duration loss (2nd stage)
    lambda_ce: 20. # duration predictor probability output CE loss (2nd stage)
    lambda_sty: 1. # style reconstruction loss (2nd stage)
    lambda_diff: 1. # score matching loss (2nd stage)
    
    diff_epoch: 1 # style diffusion starting epoch (2nd stage)
    joint_epoch: 5 # joint training starting epoch (2nd stage)

# ─── OPTIMISER ──────────────────────────────────────────────────────
optimizer_params:
  lr: 0.0001
  bert_lr: 0.00001
  ft_lr: 0.00001
  grad_accum_steps: 2

# ─── MODEL (core network & sub-modules) ─────────────────────────────
model_params:
  multispeaker: true        # speaker-ID column present
  dim_in: 64
  hidden_dim: 512
  max_conv_dim: 512
  n_layer: 3
  n_mels: 80

  n_token: 178              # 178 phonemes 
  max_dur: 50
  style_dim: 128
  dropout: 0.2

  decoder:
    type: hifigan
    resblock_kernel_sizes: [3, 7, 11]
    upsample_rates:       [10, 5, 3, 2]
    upsample_initial_channel: 512
    resblock_dilation_sizes: [[1,3,5],[1,3,5],[1,3,5]]
    upsample_kernel_sizes: [20, 10, 6, 4]

  slm:
    model: microsoft/wavlm-base-plus
    sr: 16000
    hidden: 768
    nlayers: 13
    initial_channel: 64

  diffusion:
    embedding_mask_proba: 0.1
    transformer:
      num_layers: 3
      num_heads: 8
      head_features: 64
      multiplier: 2
    dist:
      sigma_data: 0.2        # ← placeholder; code will overwrite if
      estimate_sigma_data: true
      mean: -3.0
      std: 1.0

# ─── EXTERNAL CHECKPOINTS ───────────────────────────────────────────
F0_path: "Utils/JDC/bst.t7"
ASR_config: "Utils/ASR/config.yml"
ASR_path: "Utils/ASR/epoch_00080.pth"
PLBERT_dir: 'Utils/PLBERT/'
first_stage_path: "/workspace/styletts2/stage1_final.pth"        # filled automatically after this run

# ─── SLM ADVERSARIAL (ignored in stage-1, kept default) ─────────────
slmadv_params:
  min_len: 400
  max_len: 500
  batch_percentage: 0.5
  iter: 20
  thresh: 5
  scale: 0.01
  sig: 1.5