ai-voice-cloning/models/.template.yaml

142 lines
4.2 KiB
YAML
Raw Normal View History

name: ${voice}
2023-02-17 03:05:27 +00:00
model: extensibletrainer
scale: 1
gpu_ids: [0] # Superfluous, redundant, unnecessary, the way you launch the training script will set this
2023-02-23 06:24:54 +00:00
start_step: 0
checkpointing_enabled: true
fp16: ${half_p}
bitsandbytes: ${bitsandbytes}
gpus: ${gpus}
wandb: false
2023-02-17 03:05:27 +00:00
use_tb_logger: true
datasets:
train:
name: training
n_workers: ${workers}
batch_size: ${batch_size}
2023-02-17 03:05:27 +00:00
mode: paired_voice_audio
path: ${dataset_path}
fetcher_mode: ['lj']
2023-02-17 03:05:27 +00:00
phase: train
max_wav_length: 255995
max_text_length: 200
sample_rate: 22050
load_conditioning: True
num_conditioning_candidates: 2
conditioning_length: 44000
use_bpe_tokenizer: True
tokenizer_vocab: ./models/tortoise/bpe_lowercase_asr_256.json
2023-02-17 03:05:27 +00:00
load_aligned_codes: False
val: # I really do not care about validation right now
name: validation
n_workers: ${workers}
batch_size: ${validation_batch_size}
2023-02-17 03:05:27 +00:00
mode: paired_voice_audio
path: ${validation_path}
fetcher_mode: ['lj']
phase: val
2023-02-17 03:05:27 +00:00
max_wav_length: 255995
max_text_length: 200
sample_rate: 22050
load_conditioning: True
num_conditioning_candidates: 2
conditioning_length: 44000
use_bpe_tokenizer: True
tokenizer_vocab: ./models/tortoise/bpe_lowercase_asr_256.json
2023-02-17 03:05:27 +00:00
load_aligned_codes: False
steps:
gpt_train:
training: gpt
loss_log_buffer: 500
2023-02-17 03:05:27 +00:00
# Generally follows the recipe from the DALLE paper.
optimizer: adamw # this should be adamw_zero if you're using distributed training
optimizer_params:
lr: !!float ${learning_rate} # originally: 1e-4
2023-02-17 03:05:27 +00:00
weight_decay: !!float 1e-2
beta1: 0.9
beta2: 0.96
clip_grad_eps: 4
injectors:
2023-02-17 03:05:27 +00:00
paired_to_mel:
type: torch_mel_spectrogram
mel_norm_file: ./models/tortoise/clips_mel_norms.pth
2023-02-17 03:05:27 +00:00
in: wav
out: paired_mel
paired_cond_to_mel:
type: for_each
subtype: torch_mel_spectrogram
mel_norm_file: ./models/tortoise/clips_mel_norms.pth
2023-02-17 03:05:27 +00:00
in: conditioning
out: paired_conditioning_mel
to_codes:
type: discrete_token
in: paired_mel
out: paired_mel_codes
dvae_config: "./models/tortoise/train_diffusion_vocoder_22k_level.yml"
2023-02-17 03:05:27 +00:00
paired_fwd_text:
type: generator
generator: gpt
in: [paired_conditioning_mel, padded_text, text_lengths, paired_mel_codes, wav_lengths]
out: [loss_text_ce, loss_mel_ce, logits]
losses:
text_ce:
type: direct
weight: ${text_ce_lr_weight}
2023-02-17 03:05:27 +00:00
key: loss_text_ce
mel_ce:
type: direct
weight: 1
key: loss_mel_ce
networks:
gpt:
type: generator
which_model_G: unified_voice2 # none of the unified_voice*.py files actually match the tortoise inference code... 4 and 3 have "alignment_head" (wtf is that?), 2 lacks the types=1 parameter.
kwargs:
layers: 30 # originally: 8
model_dim: 1024 # originally: 512
heads: 16 # originally: 8
max_text_tokens: 402 # originally: 120
max_mel_tokens: 604 # originally: 250
max_conditioning_inputs: 2 # originally: 1
2023-02-17 03:05:27 +00:00
mel_length_compression: 1024
number_text_tokens: 256 # supposed to be 255 for newer unified_voice files
number_mel_codes: 8194
start_mel_token: 8192
stop_mel_token: 8193
start_text_token: 255
train_solo_embeddings: False # missing in uv3/4
use_mel_codes_as_input: True # ditto
checkpointing: True
#types: 1 # this is MISSING, but in my analysis 1 is equivalent to not having it.
#only_alignment_head: False # uv3/4
path:
strict_load: true
${source_model}
${resume_state}
2023-02-17 03:05:27 +00:00
train:
2023-02-18 15:50:51 +00:00
niter: ${iterations}
2023-02-17 03:05:27 +00:00
warmup_iter: -1
mega_batch_factor: ${gradient_accumulation_size}
val_freq: ${validation_rate}
2023-02-17 03:05:27 +00:00
ema_enabled: false # I really don't think EMA matters
${learning_rate_scheme}
2023-02-17 03:05:27 +00:00
eval:
pure: ${validation_enabled}
2023-02-17 03:05:27 +00:00
output_state: gen
logger:
2023-02-18 15:50:51 +00:00
print_freq: ${print_rate}
save_checkpoint_freq: ${save_rate}
2023-02-17 03:05:27 +00:00
visuals: [gen, mel]
2023-02-18 15:50:51 +00:00
visual_debug_rate: ${print_rate}
2023-02-17 03:05:27 +00:00
is_mel_spectrogram: true