vall-e/data/config.yaml

models:
- name: "ar+nar"
  size: "full"
  resp_levels: 8
  prom_levels: 8
  tasks: 8
  langs: 2
  tones: 1
  arch_type: llama
  training: True
  version: 4
  attention: flash_attention_2
  dropout: 0.1

  loss_factors:
    text: 0.1
    resp: 1.0

hyperparameters:
  autotune: False
  autotune_params:
    start_profile_step: 1
    end_profile_step: 50
    num_tuning_micro_batch_sizes: 8

  batch_size: 16
  gradient_accumulation_steps: 4
  gradient_clipping: 1.0
  warmup_steps: 100

  optimizer: Prodigy
  learning_rate: 1.0
  torch_optimizer: True
  
  scheduler: "" # ScheduleFree
  torch_scheduler: True

evaluation:
  batch_size: 8
  frequency: 5000
  size: 8
  
  steps: 500
  ar_temperature: 0.95
  nar_temperature: 0.25
  load_disabled_engines: True

trainer:
  #no_logger: True
  ddp: False
  #check_for_oom: False
  iterations: 1_000_000
  
  save_tag: step
  save_on_oom: True
  save_on_quit: True
  save_frequency: 250
  export_on_save: True

  keep_last_checkpoints: 4

  aggressive_optimizations: False
  load_disabled_engines: False

  #load_state_dict: True
  strict_loading: False
  #load_tag: "9500"
  #load_states: False
  #restart_step_count: True
  
  gc_mode: None # "global_step"

  weight_dtype: float32 # float16 or bfloat16
  amp: False

  backend: deepspeed
  deepspeed:
    inferencing: True
    zero_optimization_level: 0
    use_compression_training: False

    amp: False

  activation_checkpointing: True

  load_webui: False

inference:
  backend: deepspeed
  audio_backend: "dac"
  normalize: False

  weight_dtype: float32 # float16 or bfloat16
  amp: False

optimizations:
  injects: False
  replace: True

  linear: False
  embedding: False
  optimizers: True

  bitsandbytes: False
  dadaptation: False
  bitnet: False
  fp8: False

experimental: True # practically required now it seems

dataset:
  speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"
  speaker_group_getter: "lambda p: f'{p.parts[-3]}'"
  speaker_languages:
    ja: []

  use_hdf5: True
  use_metadata: True
  hdf5_flag: r
  validate: True

  workers: 2
  cache: True
  
  duration_range: [3.0, 5.0]

  random_utterance: 1.0
  max_prompts: 1
  prompt_duration: 3.0
  
  max_resps: 1
  p_resp_append: 0.25

  sample_type: path # speaker

  tasks_list: [ "tts" ] # , [ "tts", "tts-c", "ns", "sr", "tse", "cse", "nse", "tts"]

  training: []
  validation: []
  noise: []
Rewrite init 2023-08-02 21:53:35 +00:00			`models:`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`- name: "ar+nar"`
			`size: "full"`
			`resp_levels: 8`
			`prom_levels: 8`
			`tasks: 8`
			`langs: 2`
			`tones: 1`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`arch_type: llama`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`training: True`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`version: 4`
			`attention: flash_attention_2`
			`dropout: 0.1`

			`loss_factors:`
			`text: 0.1`
			`resp: 1.0`
nightmare of spaghetti that might break compat; mechanism to increase RVQ bins of an existing model without retraining, keeps sampled proms/resps at max RVQ level and trim off excess levels according to what model receives them, some other things I already forgot (I really hope no one else has weights being baked right now) 2023-08-19 20:06:33 +00:00
Rewrite init 2023-08-02 21:53:35 +00:00			`hyperparameters:`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`autotune: False`
			`autotune_params:`
			`start_profile_step: 1`
			`end_profile_step: 50`
			`num_tuning_micro_batch_sizes: 8`

			`batch_size: 16`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`gradient_accumulation_steps: 4`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`gradient_clipping: 1.0`
			`warmup_steps: 100`

			`optimizer: Prodigy`
			`learning_rate: 1.0`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`torch_optimizer: True`
Rewrite init 2023-08-02 21:53:35 +00:00
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`scheduler: "" # ScheduleFree`
			`torch_scheduler: True`
Rewrite init 2023-08-02 21:53:35 +00:00
			`evaluation:`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`batch_size: 8`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`frequency: 5000`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`size: 8`
Rewrite init 2023-08-02 21:53:35 +00:00
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`steps: 500`
nightmare of spaghetti that might break compat; mechanism to increase RVQ bins of an existing model without retraining, keeps sampled proms/resps at max RVQ level and trim off excess levels according to what model receives them, some other things I already forgot (I really hope no one else has weights being baked right now) 2023-08-19 20:06:33 +00:00			`ar_temperature: 0.95`
			`nar_temperature: 0.25`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`load_disabled_engines: True`
Rewrite init 2023-08-02 21:53:35 +00:00
			`trainer:`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`#no_logger: True`
			`ddp: False`
			`#check_for_oom: False`
repaired auraloss loss calc during eval/val 2023-08-19 02:19:47 +00:00			`iterations: 1_000_000`
Rewrite init 2023-08-02 21:53:35 +00:00
			`save_tag: step`
			`save_on_oom: True`
			`save_on_quit: True`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`save_frequency: 250`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`export_on_save: True`
repaired auraloss loss calc during eval/val 2023-08-19 02:19:47 +00:00
			`keep_last_checkpoints: 4`
Rewrite init 2023-08-02 21:53:35 +00:00
			`aggressive_optimizations: False`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`load_disabled_engines: False`
Rewrite init 2023-08-02 21:53:35 +00:00
			`#load_state_dict: True`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`strict_loading: False`
repaired auraloss loss calc during eval/val 2023-08-19 02:19:47 +00:00			`#load_tag: "9500"`
			`#load_states: False`
Rewrite init 2023-08-02 21:53:35 +00:00			`#restart_step_count: True`

			`gc_mode: None # "global_step"`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`weight_dtype: float32 # float16 or bfloat16`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`amp: False`
Rewrite init 2023-08-02 21:53:35 +00:00
big cleanup 2023-08-04 01:26:36 +00:00			`backend: deepspeed`
			`deepspeed:`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`inferencing: True`
repaired auraloss loss calc during eval/val 2023-08-19 02:19:47 +00:00			`zero_optimization_level: 0`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`use_compression_training: False`
adjustments 2023-08-02 22:01:49 +00:00
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`amp: False`

fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`activation_checkpointing: True`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`load_webui: False`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
big cleanup 2023-08-04 01:26:36 +00:00			`inference:`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`backend: deepspeed`
			`audio_backend: "dac"`
fixed training stats not loading from exported weights, a bit of a readme cleanup, updated example training yaml 2023-09-24 00:59:00 +00:00			`normalize: False`
big cleanup 2023-08-04 01:26:36 +00:00
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`weight_dtype: float32 # float16 or bfloat16`
updated setup script 2023-10-07 01:08:28 +00:00			`amp: False`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`optimizations:`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`injects: False`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`replace: True`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
			`linear: False`
			`embedding: False`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`optimizers: True`

			`bitsandbytes: False`
			`dadaptation: False`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`bitnet: False`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`fp8: False`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`experimental: True # practically required now it seems`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
			`dataset:`
			`speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'"`
			`speaker_group_getter: "lambda p: f'{p.parts[-3]}'"`
			`speaker_languages:`
			`ja: []`

			`use_hdf5: True`
			`use_metadata: True`
			`hdf5_flag: r`
			`validate: True`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`workers: 2`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00			`cache: True`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`duration_range: [3.0, 5.0]`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
			`random_utterance: 1.0`
added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`max_prompts: 1`
			`prompt_duration: 3.0`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
			`max_resps: 1`
			`p_resp_append: 0.25`

added option to split between text loss and audio loss (to-do: document this better), because it may or may not be a problem with LLaMA-backed models because my loss hovers around 3.9 / 56% accuracy despite sounding decent at the moment 2024-05-19 16:23:56 +00:00			`sample_type: path # speaker`
dataset preparation script updates, caved and am using HF tokenizer now 2024-04-21 19:49:18 +00:00
			`tasks_list: [ "tts" ] # , [ "tts", "tts-c", "ns", "sr", "tse", "cse", "nse", "tts"]`

			`training: []`
			`validation: []`
			`noise: []`