Upload Urdu fine-tuning config with hf_pipeline block

configs/config_urdu_ft.yml

@@ -0,0 +1,197 @@
+# ===========================================================================
+# kikiri-tts — Urdu Fine-Tuning Config (Kokoro-82M via patched StyleTTS2)
+# ===========================================================================
+#
+# This config adapts configs/config_german_ft.yml for Urdu fine-tuning on
+# the humair025/koku dataset, AND wires in the FlowTTS-style HF pipeline
+# (kikiri_pipeline package) via the top-level hf_pipeline block.
+#
+# Top-level keys (batch_size, epochs_1st, save_freq, pretrained_model, etc.)
+# are read by StyleTTS2/train_first.py + train_second.py — they MUST be at
+# the top level (not nested under 'training:'). See docs/TRAINING_GUIDE.md
+# for why this matters.
+# ===========================================================================
+
+# ── Top-level training params (read by StyleTTS2) ──────────────────────────
+batch_size: 8
+epochs_1st: 10                    # Stage 1 epochs (acoustic reconstruction)
+epochs_2nd: 10                   # Stage 2 epochs (prosody + adversarial)
+save_freq: 1                    # save every N epochs
+pretrained_model: "../training/kokoro_base.pth"
+first_stage_path: "first_stage.pth"
+load_only_params: true
+second_stage_load_pretrained: false   # false = load from first_stage.pth (recommended)
+
+# ── Logging ────────────────────────────────────────────────────────────────
+log_dir: "logs/kokoro-urdu"
+
+# ── Data ───────────────────────────────────────────────────────────────────
+# These paths are relative to StyleTTS2/ (the cwd when running train_first.py)
+data_params:
+  train_data: "../training/train_list.txt"
+  val_data:   "../training/val_list.txt"
+  root_path:  "../training/audio"           # parent dir of the WAVs
+  OOD_data:   "../training/OOD_texts.txt"
+  min_length: 5                              # min phoneme string length (chars)
+  num_workers: 8
+
+# ── Preprocessing (Kokoro/StyleTTS2 fixed — do NOT change) ─────────────────
+preprocess_params:
+  sr: 24000                                   # Kokoro = 24 kHz mono
+  spect_params:
+    n_fft: 2048
+    win_length: 1200
+    hop_length: 300
+    n_mels: 80
+    fmin: 0
+    fmax: 8000
+
+# ── Model architecture (must match Kokoro-82M exactly) ─────────────────────
+model_params:
+  dim_in: 64
+  n_token: 178                                # Kokoro vocab size (IPA + punct)
+  hidden_dim: 512
+  style_dim: 128                              # voicepack = 2 * style_dim
+  max_dur: 50
+  multispeaker: false                        # Urdu single-speaker fine-tune
+  n_mels: 80
+  dropout: 0.2
+  n_layer: 3
+  text_encoder_kernel_size: 5
+
+  decoder:
+    type: istftnet                            # Kokoro uses ISTFTNet, NOT hifigan
+    upsample_rates: [10, 6]
+    upsample_kernel_sizes: [20, 12]
+    upsample_initial_channel: 512
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    gen_istft_n_fft: 20
+    gen_istft_hop_size: 5
+
+  diffusion:                                   # kept for compat — disabled via diff_epoch=999
+    embedding_mask_proba: 0.1
+    transformer:
+      num_layers: 3
+      num_heads: 8
+      head_features: 64
+      multiplier: 2
+    dist:
+      sigma_data: 0.2
+      estimate_sigma_data: true
+      mean: -3.0
+      std: 1.0
+
+  plbert:
+    hidden_size: 768
+    num_attention_heads: 12
+    intermediate_size: 2048
+    max_position_embeddings: 512
+    num_hidden_layers: 12
+    dropout: 0.1
+
+  slm:
+    model: "microsoft/wavlm-base-plus"
+    sr: 16000                                 # WavLM operates at 16 kHz
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+
+# ── Loss weights ───────────────────────────────────────────────────────────
+loss_params:
+  lambda_gen: 1.0
+  lambda_mel: 5.0
+  lambda_dur: 1.0
+  lambda_ce: 20.0
+  lambda_F0: 1.0
+  lambda_norm: 1.0
+  lambda_s2s: 1.0
+  lambda_mono: 1.0
+  lambda_slm: 1.0                             # SLM adversarial (Stage 2)
+  lambda_diff: 0.0                            # no diffusion
+  lambda_sty: 0.0
+  TMA_epoch: 0                                # start alignment from epoch 0
+  diff_epoch: 999                             # disable diffusion
+  joint_epoch: 3                              # start GAN + SLM at epoch 3
+
+# ── Optimizer ──────────────────────────────────────────────────────────────
+optimizer_params:
+  lr: 0.0001
+  bert_lr: 0.00001                            # lower LR for PLBERT
+  ft_lr: 0.0001
+
+# ── Utility models (StyleTTS2 expects these in Utils/) ──────────────────────
+F0_path: "Utils/JDC/bst.t7"
+ASR_config: "Utils/ASR/config.yml"
+ASR_path: "Utils/ASR/epoch_00080.pth"
+PLBERT_dir: "Utils/PLBERT/"
+
+slmadv_params:
+  min_len: 100
+  max_len: 500
+  batch_percentage: 0.5
+  iter: 10
+  thresh: 5
+  scale: 0.01
+  sig: 1.5
+
+
+# ===========================================================================
+#  FlowTTS-style HF pipeline block (read by kikiri_pipeline package)
+# ===========================================================================
+#  This is the new section that integrates the FlowTTS pipeline:
+#    - HF dataset loading (humair025/koku) with audio decode + 24 kHz resample
+#    - urdu_g2p + KokoroTokenizer.encode for phoneme -> token ID
+#    - HF Hub push of checkpoints + losses + TensorBoard logs
+#    - Resume from local path / HF Hub / 'auto'
+#
+#  All keys here are OPTIONAL — the kikiri_pipeline package provides strong
+#  defaults. Override any of them via CLI flags:
+#     --hf_repo humair025/kikiri-urdu-v1
+#     --upload_every 500
+#     --hf_dry_run
+# ===========================================================================
+hf_pipeline:
+  # ---- Dataset ----
+  dataset_id:        "humair025/koku"
+  dataset_split:     "train"
+  dataset_audio_col: "audio"               # HF datasets Audio feature column
+  dataset_text_col:  "text"                # raw transcript column
+  dataset_phonemes_col: "phonemes"          # PRE-COMPUTED IPA column (already in dataset)
+  dataset_voice_col: "voice"                # speaker label column
+  sample_rate:       24000                   # resample target (Kokoro = 24 kHz)
+  streaming:         true                    # set false only if you have >10 GB free disk
+  val_ratio:         0.05                    # 5% validation split
+  random_seed:       42
+  min_phoneme_len:   5                       # filter phoneme strings < 5 chars
+  max_phoneme_len:   510                     # PLBERT max_position_embeddings - 2
+
+  # ---- Audio preprocessing ----
+  n_fft:             2048
+  hop_length:        300
+  win_length:        1200
+  n_mels:            80
+  f_min:             0.0
+  f_max:             8000.0
+  precompute_mels:   true
+  precompute_f0:     true
+
+  # ---- HF Hub push ----
+  hf_repo:           "humair025/kikiri-urdu-v1"   # set to null to disable uploads
+  hf_token_env:      "HF_TOKEN"              # env var to read token from
+  upload_every:      1                        # push ckpt every N epochs (we save per-epoch)
+  keep_last_n_local: 2                       # prune older local ckpts (hub copies kept)
+  upload_tensorboard: true                    # also push TB logs on a separate cadence
+  tb_log_upload_freq: 200                    # push TB logs every N steps
+  losses_upload_freq: 50                     # push losses.json every N steps
+  hf_dry_run:        false                   # true = save locally but don't upload
+  ckpt_subdir:       "checkpoints"           # path inside HF repo for .pth files
+
+  # ---- Resume ----
+  resume:            null                    # local path / HF path / 'auto'
+
+  # ---- Loss log ----
+  losses_filename:   "losses.json"
+
+  # ---- Stage tagging (set per-script: stage1 in train_first, stage2 in train_second) ----
+  stage_tag:         "stage1"