first commit

2026-08-15 14:43:56 +08:00
commit 63b8c28e1d
340 changed files with 59515 additions and 0 deletions
@@ -0,0 +1,76 @@
+train_data:
+  train_audio_sample_rate: 48000
+  audio_samples_per_llm_token: 7680
+  sources:
+    - name: ljspeech_basic
+      weight: 1.0
+      pipeline: basic
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_train.jsonl
+          shuffle: true
+    - name: ljspeech_interleave
+      weight: 1.0
+      pipeline: interleave
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_train.jsonl
+          shuffle: true
+    # append other sources here if need
+  num_tokens_per_epoch: 2000000
+  num_workers: 20
+  pin_memory: true
+  max_audio_seconds_in_batch: 30.0
+  max_text_tokens_in_batch: 2048
+  max_samples_per_batch: null
+  bucketing_pool_size: 100
+val_data:
+  train_audio_sample_rate: 48000
+  audio_samples_per_llm_token: 7680
+  sources:
+    - name: ljspeech_valid_basic
+      weight: 1.0
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_valid.jsonl
+          shuffle: false
+      pipeline: basic
+    - name: ljspeech_valid_interleave
+      weight: 1.0
+      pipeline: interleave
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_valid.jsonl
+          shuffle: false
+      pipeline: interleave
+    # append other sources here if need
+  num_workers: 4
+  pin_memory: true
+  max_audio_seconds_in_batch: 30.0
+  max_text_tokens_in_batch: 2048
+  max_samples_per_batch: null
+  bucketing_pool_size: 64
+train:
+  pretrained_model_path: pretrained_models/pretrain_cpt_decay/latest/model/
+  output_dir: debug_train/run_003
+  seed: 42
+  learning_rate: 1.0e-05
+  weight_decay: 0.01
+  warmup_steps: 50
+  max_train_steps: 500
+  gradient_accumulation_steps: 2
+  grad_clip_norm: 1
+  save_interval: 500
+  max_checkpoints_to_keep: 40
+  log_interval: 10
+  eval_interval: 100
+  max_eval_batches: null
+  run_eval_on_start: false
+loss:
+  ce_weight: 1.0
+  fm_weight: 1.0
+  eos_weight: 1.0
@@ -0,0 +1,61 @@
+train_data:
+  train_audio_sample_rate: 48000
+  audio_samples_per_llm_token: 7680
+  sources:
+    - name: ljspeech_meanflow_basic
+      weight: 1.0
+      pipeline: basic
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_train.jsonl
+          shuffle: true
+    # append other sources here if need
+  num_tokens_per_epoch: 1000000
+  num_workers: 4
+  pin_memory: true
+  max_audio_seconds_in_batch: 10.0
+  max_text_tokens_in_batch: 1024
+  max_samples_per_batch: 1
+  bucketing_pool_size: 64
+val_data:
+  train_audio_sample_rate: 48000
+  audio_samples_per_llm_token: 7680
+  sources:
+    - name: ljspeech_meanflow_valid_basic
+      weight: 1.0
+      pipeline: basic
+      adapter:
+        class_name: JsonlManifestSourceAdapter
+        params:
+          manifest_path: downloaded_data/ljspeech_48khz_manifest_valid.jsonl
+          shuffle: false
+    # append other sources here if need
+  num_workers: 2
+  pin_memory: true
+  max_audio_seconds_in_batch: 10.0
+  max_text_tokens_in_batch: 1024
+  max_samples_per_batch: 1
+  bucketing_pool_size: 32
+train:
+  pretrained_model_path: pretrained_models/dots.tts-soar
+  output_dir: debug_train/meanflow_run_001
+  seed: 42
+  learning_rate: 1.0e-05
+  cfg_droprate: 0.0
+  xvec_drop_rate: 0.5
+  weight_decay: 0.01
+  warmup_steps: 50
+  max_train_steps: 500
+  gradient_accumulation_steps: 1
+  grad_clip_norm: 1
+  save_interval: 500
+  max_checkpoints_to_keep: 10
+  log_interval: 10
+  eval_interval: 100
+  max_eval_batches: 10
+  run_eval_on_start: false
+loss:
+  ce_weight: 1.0
+  fm_weight: 1.0
+  eos_weight: 1.0