forked from mrq/ai-voice-cloning
Update 'src/utils.py'
whisper->whisperx
This commit is contained in:
yqxtqymn
parent
079cd32074
commit
9ca5192309
453
src/utils.py
453
src/utils.py
|
|
@ -1,4 +1,5 @@
|
|||
import os
|
||||
|
||||
if 'XDG_CACHE_HOME' not in os.environ:
|
||||
os.environ['XDG_CACHE_HOME'] = os.path.realpath(os.path.join(os.getcwd(), './models/'))
|
||||
|
||||
|
|
@ -36,8 +37,9 @@ from tortoise.utils.audio import load_audio, load_voice, load_voices, get_voice_
|
|||
from tortoise.utils.text import split_and_recombine_text
|
||||
from tortoise.utils.device import get_device_name, set_device_name
|
||||
|
||||
MODELS['dvae.pth'] = "https://huggingface.co/jbetker/tortoise-tts-v2/resolve/3704aea61678e7e468a06d8eea121dba368a798e/.models/dvae.pth"
|
||||
WHISPER_MODELS = ["tiny", "base", "small", "medium", "large"]
|
||||
MODELS[
|
||||
'dvae.pth'] = "https://huggingface.co/jbetker/tortoise-tts-v2/resolve/3704aea61678e7e468a06d8eea121dba368a798e/.models/dvae.pth"
|
||||
WHISPER_MODELS = ["tiny", "base", "small", "medium", "large", "large-v2"]
|
||||
WHISPER_SPECIALIZED_MODELS = ["tiny.en", "base.en", "small.en", "medium.en"]
|
||||
EPOCH_SCHEDULE = [9, 18, 25, 33]
|
||||
|
||||
|
|
@ -49,6 +51,7 @@ voicefixer = None
|
|||
whisper_model = None
|
||||
training_state = None
|
||||
|
||||
|
||||
def generate(
|
||||
text,
|
||||
delimiter,
|
||||
|
|
@ -110,13 +113,16 @@ def generate(
|
|||
if voice_samples and len(voice_samples) > 0:
|
||||
sample_voice = torch.cat(voice_samples, dim=-1).squeeze().cpu()
|
||||
|
||||
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, progress=progress, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents)
|
||||
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean,
|
||||
progress=progress, slices=voice_latents_chunks,
|
||||
force_cpu=args.force_cpu_for_conditioning_latents)
|
||||
if len(conditioning_latents) == 4:
|
||||
conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None)
|
||||
|
||||
if voice != "microphone":
|
||||
if hasattr(tts, 'autoregressive_model_hash'):
|
||||
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents_{tts.autoregressive_model_hash[:8]}.pth')
|
||||
torch.save(conditioning_latents,
|
||||
f'{get_voice_dir()}/{voice}/cond_latents_{tts.autoregressive_model_hash[:8]}.pth')
|
||||
else:
|
||||
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth')
|
||||
voice_samples = None
|
||||
|
|
@ -132,10 +138,10 @@ def generate(
|
|||
seed = None
|
||||
|
||||
if conditioning_latents is not None and len(conditioning_latents) == 2 and cvvp_weight > 0:
|
||||
print("Requesting weighing against CVVP weight, but voice latents are missing some extra data. Please regenerate your voice latents.")
|
||||
print(
|
||||
"Requesting weighing against CVVP weight, but voice latents are missing some extra data. Please regenerate your voice latents.")
|
||||
cvvp_weight = 0
|
||||
|
||||
|
||||
settings = {
|
||||
'temperature': float(temperature),
|
||||
|
||||
|
|
@ -199,7 +205,8 @@ def generate(
|
|||
beta=8.555504641634386,
|
||||
)
|
||||
|
||||
volume_adjust = torchaudio.transforms.Vol(gain=args.output_volume, gain_type="amplitude") if args.output_volume != 1 else None
|
||||
volume_adjust = torchaudio.transforms.Vol(gain=args.output_volume,
|
||||
gain_type="amplitude") if args.output_volume != 1 else None
|
||||
|
||||
idx = 0
|
||||
idx_cache = {}
|
||||
|
|
@ -380,7 +387,6 @@ def generate(
|
|||
with open(f'{outdir}/{voice}_{name}.json', 'w', encoding="utf-8") as f:
|
||||
f.write(json.dumps(info, indent='\t'))
|
||||
|
||||
|
||||
if voice and voice != "random" and conditioning_latents is not None:
|
||||
latents_path = f'{get_voice_dir()}/{voice}/cond_latents.pth'
|
||||
|
||||
|
|
@ -428,10 +434,12 @@ def generate(
|
|||
stats,
|
||||
)
|
||||
|
||||
|
||||
def cancel_generate():
|
||||
import tortoise.api
|
||||
tortoise.api.STOP_SIGNAL = True
|
||||
|
||||
|
||||
def hash_file(path, algo="md5", buffer_size=0):
|
||||
import hashlib
|
||||
|
||||
|
|
@ -458,6 +466,7 @@ def hash_file(path, algo="md5", buffer_size=0):
|
|||
|
||||
return "{0}".format(hash.hexdigest())
|
||||
|
||||
|
||||
def update_baseline_for_latents_chunks(voice):
|
||||
path = f'{get_voice_dir()}/{voice}/'
|
||||
if not os.path.isdir(path):
|
||||
|
|
@ -481,6 +490,7 @@ def update_baseline_for_latents_chunks( voice ):
|
|||
return int(total_duration / total) if total > 0 else 1
|
||||
return int(total_duration / args.autocalculate_voice_chunk_duration_size) if total_duration > 0 else 1
|
||||
|
||||
|
||||
def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm=True)):
|
||||
global tts
|
||||
global args
|
||||
|
|
@ -498,18 +508,22 @@ def compute_latents(voice, voice_latents_chunks, progress=gr.Progress(track_tqdm
|
|||
if voice_samples is None:
|
||||
return
|
||||
|
||||
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean, progress=progress, slices=voice_latents_chunks, force_cpu=args.force_cpu_for_conditioning_latents)
|
||||
conditioning_latents = tts.get_conditioning_latents(voice_samples, return_mels=not args.latents_lean_and_mean,
|
||||
progress=progress, slices=voice_latents_chunks,
|
||||
force_cpu=args.force_cpu_for_conditioning_latents)
|
||||
|
||||
if len(conditioning_latents) == 4:
|
||||
conditioning_latents = (conditioning_latents[0], conditioning_latents[1], conditioning_latents[2], None)
|
||||
|
||||
if hasattr(tts, 'autoregressive_model_hash'):
|
||||
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents_{tts.autoregressive_model_hash[:8]}.pth')
|
||||
torch.save(conditioning_latents,
|
||||
f'{get_voice_dir()}/{voice}/cond_latents_{tts.autoregressive_model_hash[:8]}.pth')
|
||||
else:
|
||||
torch.save(conditioning_latents, f'{get_voice_dir()}/{voice}/cond_latents.pth')
|
||||
|
||||
return voice
|
||||
|
||||
|
||||
# superfluous, but it cleans up some things
|
||||
class TrainingState():
|
||||
def __init__(self, config_path, keep_x_past_datasets=0, start=True, gpus=1):
|
||||
|
|
@ -580,7 +594,8 @@ class TrainingState():
|
|||
self.cmd = ['train.bat', config_path] if os.name == "nt" else ['./train.sh', str(int(gpus)), config_path]
|
||||
|
||||
print("Spawning process: ", " ".join(self.cmd))
|
||||
self.process = subprocess.Popen(self.cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, universal_newlines=True)
|
||||
self.process = subprocess.Popen(self.cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
|
||||
universal_newlines=True)
|
||||
|
||||
def load_losses(self, update=False):
|
||||
if not os.path.isdir(f'{self.dataset_dir}/tb_logger/'):
|
||||
|
|
@ -599,7 +614,8 @@ class TrainingState():
|
|||
self.statistics = []
|
||||
|
||||
if use_tensorboard:
|
||||
logs = sorted([f'{self.dataset_dir}/tb_logger/{d}' for d in os.listdir(f'{self.dataset_dir}/tb_logger/') if d[:6] == "events" ])
|
||||
logs = sorted([f'{self.dataset_dir}/tb_logger/{d}' for d in os.listdir(f'{self.dataset_dir}/tb_logger/') if
|
||||
d[:6] == "events"])
|
||||
if update:
|
||||
logs = [logs[-1]]
|
||||
|
||||
|
|
@ -893,6 +909,7 @@ class TrainingState():
|
|||
message,
|
||||
)
|
||||
|
||||
|
||||
def run_training(config_path, verbose=False, gpus=1, keep_x_past_datasets=0, progress=gr.Progress(track_tqdm=True)):
|
||||
global training_state
|
||||
if training_state and training_state.process:
|
||||
|
|
@ -911,7 +928,8 @@ def run_training(config_path, verbose=False, gpus=1, keep_x_past_datasets=0, pro
|
|||
if training_state.killed:
|
||||
return
|
||||
|
||||
result, percent, message = training_state.parse( line=line, verbose=verbose, keep_x_past_datasets=keep_x_past_datasets, progress=progress )
|
||||
result, percent, message = training_state.parse(line=line, verbose=verbose,
|
||||
keep_x_past_datasets=keep_x_past_datasets, progress=progress)
|
||||
print(f"[Training] [{datetime.now().isoformat()}] {line[:-1]}")
|
||||
if result:
|
||||
yield result
|
||||
|
|
@ -924,6 +942,14 @@ def run_training(config_path, verbose=False, gpus=1, keep_x_past_datasets=0, pro
|
|||
return_code = training_state.process.wait()
|
||||
training_state = None
|
||||
|
||||
|
||||
def get_training_losses():
|
||||
global training_state
|
||||
if not training_state or not training_state.statistics:
|
||||
return
|
||||
return pd.DataFrame(training_state.statistics)
|
||||
|
||||
|
||||
def update_training_dataplot(config_path=None):
|
||||
global training_state
|
||||
update = None
|
||||
|
|
@ -932,22 +958,24 @@ def update_training_dataplot(config_path=None):
|
|||
if config_path:
|
||||
training_state = TrainingState(config_path=config_path, start=False)
|
||||
if training_state.statistics:
|
||||
update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics), x_lim=[0,training_state.its], x="step", y="value", title="Training Metrics", color="type", tooltip=['step', 'value', 'type'], width=600, height=350,)
|
||||
update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics))
|
||||
del training_state
|
||||
training_state = None
|
||||
elif training_state.statistics:
|
||||
training_state.load_losses()
|
||||
update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics), x_lim=[0,training_state.its], x="step", y="value", title="Training Metrics", color="type", tooltip=['step', 'value', 'type'], width=600, height=350,)
|
||||
update = gr.LinePlot.update(value=pd.DataFrame(training_state.statistics))
|
||||
|
||||
return update
|
||||
|
||||
|
||||
def reconnect_training(verbose=False, progress=gr.Progress(track_tqdm=True)):
|
||||
global training_state
|
||||
if not training_state or not training_state.process:
|
||||
return "Training not in progress"
|
||||
|
||||
for line in iter(training_state.process.stdout.readline, ""):
|
||||
result, percent, message = training_state.parse( line=line, verbose=verbose, keep_x_past_datasets=keep_x_past_datasets, progress=progress )
|
||||
result, percent, message = training_state.parse(line=line, verbose=verbose,
|
||||
keep_x_past_datasets=keep_x_past_datasets, progress=progress)
|
||||
print(f"[Training] [{datetime.now().isoformat()}] {line[:-1]}")
|
||||
if result:
|
||||
yield result
|
||||
|
|
@ -955,6 +983,7 @@ def reconnect_training(verbose=False, progress=gr.Progress(track_tqdm=True)):
|
|||
if progress is not None and message:
|
||||
progress(percent, message)
|
||||
|
||||
|
||||
def stop_training():
|
||||
global training_state
|
||||
if training_state is None:
|
||||
|
|
@ -965,7 +994,8 @@ def stop_training():
|
|||
children = []
|
||||
# wrapped in a try/catch in case for some reason this fails outside of Linux
|
||||
try:
|
||||
children = [p.info for p in psutil.process_iter(attrs=['pid', 'name', 'cmdline']) if './src/train.py' in p.info['cmdline']]
|
||||
children = [p.info for p in psutil.process_iter(attrs=['pid', 'name', 'cmdline']) if
|
||||
'./src/train.py' in p.info['cmdline']]
|
||||
except Exception as e:
|
||||
pass
|
||||
|
||||
|
|
@ -981,10 +1011,12 @@ def stop_training():
|
|||
print("Killed training process.")
|
||||
return f"Training cancelled: {return_code}"
|
||||
|
||||
|
||||
def get_halfp_model_path():
|
||||
autoregressive_model_path = get_model_path('autoregressive.pth')
|
||||
return autoregressive_model_path.replace(".pth", "_half.pth")
|
||||
|
||||
|
||||
def convert_to_halfp():
|
||||
autoregressive_model_path = get_model_path('autoregressive.pth')
|
||||
print(f'Converting model to half precision: {autoregressive_model_path}')
|
||||
|
|
@ -996,66 +1028,60 @@ def convert_to_halfp():
|
|||
torch.save(model, outfile)
|
||||
print(f'Converted model to half precision: {outfile}')
|
||||
|
||||
def whisper_transcribe( file, language=None ):
|
||||
# shouldn't happen, but it's for safety
|
||||
if not whisper_model:
|
||||
load_whisper_model(language=language)
|
||||
|
||||
if not args.whisper_cpp:
|
||||
if not language:
|
||||
language = None
|
||||
|
||||
return whisper_model.transcribe(file, language=language)
|
||||
|
||||
res = whisper_model.transcribe(file)
|
||||
segments = whisper_model.extract_text_and_timestamps( res )
|
||||
|
||||
result = {
|
||||
'segments': []
|
||||
}
|
||||
for segment in segments:
|
||||
reparsed = {
|
||||
'start': segment[0] / 100.0,
|
||||
'end': segment[1] / 100.0,
|
||||
'text': segment[2],
|
||||
}
|
||||
result['segments'].append(reparsed)
|
||||
|
||||
return result
|
||||
|
||||
|
||||
def prepare_dataset(files, outdir, language=None, progress=None):
|
||||
unload_tts()
|
||||
|
||||
global whisper_model
|
||||
if whisper_model is None:
|
||||
load_whisper_model(language=language)
|
||||
import whisperx
|
||||
|
||||
device = "cuda" # add cpu option?
|
||||
|
||||
# original whisper https://github.com/openai/whisper
|
||||
# whisperx fork https://github.com/m-bain/whisperX
|
||||
# supports en, fr, de, es, it, ja, zh, nl, uk, pt
|
||||
|
||||
# tiny, base, small, medium, large, large-v2
|
||||
|
||||
whisper_model = whisperx.load_model("medium", device)
|
||||
# some additional model features require huggingface token
|
||||
|
||||
os.makedirs(outdir, exist_ok=True)
|
||||
|
||||
idx = 0
|
||||
results = {}
|
||||
transcription = []
|
||||
|
||||
for file in enumerate_progress(files, desc="Iterating through voice files", progress=progress):
|
||||
basename = os.path.basename(file)
|
||||
result = whisper_transcribe(file, language=language)
|
||||
results[basename] = result
|
||||
print(f"Transcribing file: {file}")
|
||||
|
||||
result = whisper_model.transcribe(file)
|
||||
|
||||
print(result["segments"]) # before alignment
|
||||
|
||||
# load alignment model and metadata
|
||||
model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
|
||||
|
||||
# align whisper output
|
||||
result_aligned = whisperx.align(result["segments"], model_a, metadata, file, device)
|
||||
|
||||
print(result_aligned["segments"]) # after alignment
|
||||
print(result_aligned["word_segments"]) # after alignment
|
||||
|
||||
results[os.path.basename(file)] = result
|
||||
|
||||
print(f"Transcribed file: {file}, {len(result['segments'])} found.")
|
||||
|
||||
waveform, sampling_rate = torchaudio.load(file)
|
||||
num_channels, num_frames = waveform.shape
|
||||
|
||||
idx = 0
|
||||
for segment in result['segments']: # enumerate_progress(result['segments'], desc="Segmenting voice file", progress=progress):
|
||||
for segment in result[
|
||||
'segments']: # enumerate_progress(result['segments'], desc="Segmenting voice file", progress=progress):
|
||||
start = int(segment['start'] * sampling_rate)
|
||||
end = int(segment['end'] * sampling_rate)
|
||||
|
||||
sliced_waveform = waveform[:, start:end]
|
||||
sliced_name = basename.replace(".wav", f"_{pad(idx, 4)}.wav")
|
||||
|
||||
if not torch.any(sliced_waveform < 0):
|
||||
print(f"Error with {sliced_name}, skipping...")
|
||||
continue
|
||||
sliced_name = f"{pad(idx, 4)}.wav"
|
||||
|
||||
torchaudio.save(f"{outdir}/{sliced_name}", sliced_waveform, sampling_rate)
|
||||
|
||||
|
|
@ -1068,22 +1094,26 @@ def prepare_dataset( files, outdir, language=None, progress=None ):
|
|||
with open(f'{outdir}/whisper.json', 'w', encoding="utf-8") as f:
|
||||
f.write(json.dumps(results, indent='\t'))
|
||||
|
||||
joined = '\n'.join(transcription)
|
||||
with open(f'{outdir}/train.txt', 'w', encoding="utf-8") as f:
|
||||
f.write(joined)
|
||||
f.write("\n".join(transcription))
|
||||
|
||||
unload_whisper()
|
||||
|
||||
return f"Processed dataset to: {outdir}\n{joined}"
|
||||
return f"Processed dataset to: {outdir}"
|
||||
|
||||
|
||||
def calc_iterations(epochs, lines, batch_size):
|
||||
iterations = int(epochs * lines / float(batch_size))
|
||||
return iterations
|
||||
|
||||
|
||||
def schedule_learning_rate(iterations, schedule=EPOCH_SCHEDULE):
|
||||
return [int(iterations * d) for d in schedule]
|
||||
|
||||
def optimize_training_settings( epochs, learning_rate, text_ce_lr_weight, learning_rate_schedule, batch_size, gradient_accumulation_size, print_rate, save_rate, resume_path, half_p, bnb, workers, source_model, voice ):
|
||||
|
||||
def optimize_training_settings(epochs, learning_rate, text_ce_lr_weight, learning_rate_schedule, batch_size,
|
||||
gradient_accumulation_size, print_rate, save_rate, resume_path, half_p, bnb, workers,
|
||||
source_model, voice):
|
||||
name = f"{voice}-finetune"
|
||||
dataset_name = f"{voice}-train"
|
||||
dataset_path = f"./training/{voice}/train.txt"
|
||||
|
|
@ -1102,7 +1132,8 @@ def optimize_training_settings( epochs, learning_rate, text_ce_lr_weight, learni
|
|||
if batch_size % lines != 0:
|
||||
nearest_slice = int(lines / batch_size) + 1
|
||||
batch_size = int(lines / nearest_slice)
|
||||
messages.append(f"Batch size not neatly divisible by dataset size, adjusting batch size to: {batch_size} ({nearest_slice} steps per epoch)")
|
||||
messages.append(
|
||||
f"Batch size not neatly divisible by dataset size, adjusting batch size to: {batch_size} ({nearest_slice} steps per epoch)")
|
||||
|
||||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
|
@ -1112,13 +1143,15 @@ def optimize_training_settings( epochs, learning_rate, text_ce_lr_weight, learni
|
|||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
||||
messages.append(f"Gradient accumulation size is too large for a given batch size, clamping gradient accumulation size to: {gradient_accumulation_size}")
|
||||
messages.append(
|
||||
f"Gradient accumulation size is too large for a given batch size, clamping gradient accumulation size to: {gradient_accumulation_size}")
|
||||
elif batch_size % gradient_accumulation_size != 0:
|
||||
gradient_accumulation_size = int(batch_size / gradient_accumulation_size)
|
||||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
||||
messages.append(f"Batch size is not evenly divisible by the gradient accumulation size, adjusting gradient accumulation size to: {gradient_accumulation_size}")
|
||||
messages.append(
|
||||
f"Batch size is not evenly divisible by the gradient accumulation size, adjusting gradient accumulation size to: {gradient_accumulation_size}")
|
||||
|
||||
iterations = calc_iterations(epochs=epochs, lines=lines, batch_size=batch_size)
|
||||
|
||||
|
|
@ -1140,13 +1173,15 @@ def optimize_training_settings( epochs, learning_rate, text_ce_lr_weight, learni
|
|||
if half_p:
|
||||
if bnb:
|
||||
half_p = False
|
||||
messages.append("Half Precision requested, but BitsAndBytes is also requested. Due to redundancies, disabling half precision...")
|
||||
messages.append(
|
||||
"Half Precision requested, but BitsAndBytes is also requested. Due to redundancies, disabling half precision...")
|
||||
else:
|
||||
messages.append("Half Precision requested. Please note this is ! EXPERIMENTAL !")
|
||||
if not os.path.exists(get_halfp_model_path()):
|
||||
convert_to_halfp()
|
||||
|
||||
messages.append(f"For {epochs} epochs with {lines} lines in batches of {batch_size}, iterating for {iterations} steps ({int(iterations / epochs)} steps per epoch)")
|
||||
messages.append(
|
||||
f"For {epochs} epochs with {lines} lines in batches of {batch_size}, iterating for {iterations} steps ({int(iterations / epochs)} steps per epoch)")
|
||||
|
||||
return (
|
||||
learning_rate,
|
||||
|
|
@ -1160,7 +1195,11 @@ def optimize_training_settings( epochs, learning_rate, text_ce_lr_weight, learni
|
|||
messages
|
||||
)
|
||||
|
||||
def save_training_settings( iterations=None, learning_rate=None, text_ce_lr_weight=None, learning_rate_schedule=None, batch_size=None, gradient_accumulation_size=None, print_rate=None, save_rate=None, name=None, dataset_name=None, dataset_path=None, validation_name=None, validation_path=None, output_name=None, resume_path=None, half_p=None, bnb=None, workers=None, source_model=None ):
|
||||
|
||||
def save_training_settings(iterations=None, learning_rate=None, text_ce_lr_weight=None, learning_rate_schedule=None,
|
||||
batch_size=None, gradient_accumulation_size=None, print_rate=None, save_rate=None, name=None,
|
||||
dataset_name=None, dataset_path=None, validation_name=None, validation_path=None,
|
||||
output_name=None, resume_path=None, half_p=None, bnb=None, workers=None, source_model=None):
|
||||
if not source_model:
|
||||
source_model = f"./models/tortoise/autoregressive{'_half' if half_p else ''}.pth"
|
||||
|
||||
|
|
@ -1201,6 +1240,158 @@ def save_training_settings( iterations=None, learning_rate=None, text_ce_lr_weig
|
|||
if not output_name:
|
||||
output_name = f'{settings["name"]}.yaml'
|
||||
|
||||
with open(f'./models/.template.yaml', 'r', encoding="utf-8") as f:
|
||||
yaml = f.read()
|
||||
|
||||
# i could just load and edit the YAML directly, but this is easier, as I don't need to bother with path traversals
|
||||
for k in settings:
|
||||
if settings[k] is None:
|
||||
continue
|
||||
yaml = yaml.replace(f"${{{k}}}", str(settings[k]))
|
||||
|
||||
outfile = f'./training/{output_name}'
|
||||
with open(outfile, 'w', encoding="utf-8") as f:
|
||||
f.write(yaml)
|
||||
|
||||
return f"Training settings saved to: {outfile}"
|
||||
|
||||
def calc_iterations(epochs, lines, batch_size):
|
||||
iterations = int(epochs * lines / float(batch_size))
|
||||
return iterations
|
||||
|
||||
|
||||
def schedule_learning_rate(iterations, schedule=EPOCH_SCHEDULE):
|
||||
return [int(iterations * d) for d in schedule]
|
||||
|
||||
|
||||
def optimize_training_settings(epochs, learning_rate, text_ce_lr_weight, learning_rate_schedule, batch_size,
|
||||
gradient_accumulation_size, print_rate, save_rate, resume_path, half_p, bnb, workers,
|
||||
source_model, voice):
|
||||
name = f"{voice}-finetune"
|
||||
dataset_name = f"{voice}-train"
|
||||
dataset_path = f"./training/{voice}/train.txt"
|
||||
validation_name = f"{voice}-val"
|
||||
validation_path = f"./training/{voice}/train.txt"
|
||||
|
||||
with open(dataset_path, 'r', encoding="utf-8") as f:
|
||||
lines = len(f.readlines())
|
||||
|
||||
messages = []
|
||||
|
||||
if batch_size > lines:
|
||||
batch_size = lines
|
||||
messages.append(f"Batch size is larger than your dataset, clamping batch size to: {batch_size}")
|
||||
|
||||
if batch_size % lines != 0:
|
||||
nearest_slice = int(lines / batch_size) + 1
|
||||
batch_size = int(lines / nearest_slice)
|
||||
messages.append(
|
||||
f"Batch size not neatly divisible by dataset size, adjusting batch size to: {batch_size} ({nearest_slice} steps per epoch)")
|
||||
|
||||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
||||
if batch_size / gradient_accumulation_size < 2:
|
||||
gradient_accumulation_size = int(batch_size / 2)
|
||||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
||||
messages.append(
|
||||
f"Gradient accumulation size is too large for a given batch size, clamping gradient accumulation size to: {gradient_accumulation_size}")
|
||||
elif batch_size % gradient_accumulation_size != 0:
|
||||
gradient_accumulation_size = int(batch_size / gradient_accumulation_size)
|
||||
if gradient_accumulation_size == 0:
|
||||
gradient_accumulation_size = 1
|
||||
|
||||
messages.append(
|
||||
f"Batch size is not evenly divisible by the gradient accumulation size, adjusting gradient accumulation size to: {gradient_accumulation_size}")
|
||||
|
||||
iterations = calc_iterations(epochs=epochs, lines=lines, batch_size=batch_size)
|
||||
|
||||
if epochs < print_rate:
|
||||
print_rate = epochs
|
||||
messages.append(f"Print rate is too small for the given iteration step, clamping print rate to: {print_rate}")
|
||||
|
||||
if epochs < save_rate:
|
||||
save_rate = epochs
|
||||
messages.append(f"Save rate is too small for the given iteration step, clamping save rate to: {save_rate}")
|
||||
|
||||
if resume_path and not os.path.exists(resume_path):
|
||||
resume_path = None
|
||||
messages.append("Resume path specified, but does not exist. Disabling...")
|
||||
|
||||
if bnb:
|
||||
messages.append("BitsAndBytes requested. Please note this is ! EXPERIMENTAL !")
|
||||
|
||||
if half_p:
|
||||
if bnb:
|
||||
half_p = False
|
||||
messages.append(
|
||||
"Half Precision requested, but BitsAndBytes is also requested. Due to redundancies, disabling half precision...")
|
||||
else:
|
||||
messages.append("Half Precision requested. Please note this is ! EXPERIMENTAL !")
|
||||
if not os.path.exists(get_halfp_model_path()):
|
||||
convert_to_halfp()
|
||||
|
||||
messages.append(
|
||||
f"For {epochs} epochs with {lines} lines in batches of {batch_size}, iterating for {iterations} steps ({int(iterations / epochs)} steps per epoch)")
|
||||
|
||||
return (
|
||||
learning_rate,
|
||||
text_ce_lr_weight,
|
||||
learning_rate_schedule,
|
||||
batch_size,
|
||||
gradient_accumulation_size,
|
||||
print_rate,
|
||||
save_rate,
|
||||
resume_path,
|
||||
messages
|
||||
)
|
||||
|
||||
|
||||
def save_training_settings(iterations=None, learning_rate=None, text_ce_lr_weight=None, learning_rate_schedule=None,
|
||||
batch_size=None, gradient_accumulation_size=None, print_rate=None, save_rate=None, name=None,
|
||||
dataset_name=None, dataset_path=None, validation_name=None, validation_path=None,
|
||||
output_name=None, resume_path=None, half_p=None, bnb=None, workers=None, source_model=None):
|
||||
if not source_model:
|
||||
source_model = f"./models/tortoise/autoregressive{'_half' if half_p else ''}.pth"
|
||||
|
||||
settings = {
|
||||
"iterations": iterations if iterations else 500,
|
||||
"batch_size": batch_size if batch_size else 64,
|
||||
"learning_rate": learning_rate if learning_rate else 1e-5,
|
||||
"gen_lr_steps": learning_rate_schedule if learning_rate_schedule else EPOCH_SCHEDULE,
|
||||
"gradient_accumulation_size": gradient_accumulation_size if gradient_accumulation_size else 4,
|
||||
"print_rate": print_rate if print_rate else 1,
|
||||
"save_rate": save_rate if save_rate else 50,
|
||||
"name": name if name else "finetune",
|
||||
"dataset_name": dataset_name if dataset_name else "finetune",
|
||||
"dataset_path": dataset_path if dataset_path else "./training/finetune/train.txt",
|
||||
"validation_name": validation_name if validation_name else "finetune",
|
||||
"validation_path": validation_path if validation_path else "./training/finetune/train.txt",
|
||||
|
||||
"text_ce_lr_weight": text_ce_lr_weight if text_ce_lr_weight else 0.01,
|
||||
|
||||
'resume_state': f"resume_state: '{resume_path}'",
|
||||
'pretrain_model_gpt': f"pretrain_model_gpt: '{source_model}'",
|
||||
|
||||
'float16': 'true' if half_p else 'false',
|
||||
'bitsandbytes': 'true' if bnb else 'false',
|
||||
|
||||
'workers': workers if workers else 2,
|
||||
}
|
||||
|
||||
if resume_path:
|
||||
settings['pretrain_model_gpt'] = f"# {settings['pretrain_model_gpt']}"
|
||||
else:
|
||||
settings['resume_state'] = f"# resume_state: './training/{name if name else 'finetune'}/training_state/#.state'"
|
||||
|
||||
if half_p:
|
||||
if not os.path.exists(get_halfp_model_path()):
|
||||
convert_to_halfp()
|
||||
|
||||
if not output_name:
|
||||
output_name = f'{settings["name"]}.yaml'
|
||||
|
||||
with open(f'./models/.template.yaml', 'r', encoding="utf-8") as f:
|
||||
yaml = f.read()
|
||||
|
|
@ -1217,6 +1408,7 @@ def save_training_settings( iterations=None, learning_rate=None, text_ce_lr_weig
|
|||
|
||||
return f"Training settings saved to: {outfile}"
|
||||
|
||||
|
||||
def import_voices(files, saveAs=None, progress=None):
|
||||
global args
|
||||
|
||||
|
|
@ -1282,13 +1474,16 @@ def import_voices(files, saveAs=None, progress=None):
|
|||
|
||||
print(f"Imported voice to {path}")
|
||||
|
||||
|
||||
def get_voice_list(dir=get_voice_dir(), append_defaults=False):
|
||||
os.makedirs(dir, exist_ok=True)
|
||||
res = sorted([d for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d)) and len(os.listdir(os.path.join(dir, d))) > 0 ])
|
||||
res = sorted([d for d in os.listdir(dir) if
|
||||
os.path.isdir(os.path.join(dir, d)) and len(os.listdir(os.path.join(dir, d))) > 0])
|
||||
if append_defaults:
|
||||
res = res + ["random", "microphone"]
|
||||
return res
|
||||
|
||||
|
||||
def get_autoregressive_models(dir="./models/finetunes/", prefixed=False):
|
||||
os.makedirs(dir, exist_ok=True)
|
||||
base = [get_model_path('autoregressive.pth')]
|
||||
|
|
@ -1316,11 +1511,16 @@ def get_autoregressive_models(dir="./models/finetunes/", prefixed=False):
|
|||
|
||||
return res
|
||||
|
||||
|
||||
def get_dataset_list(dir="./training/"):
|
||||
return sorted([d for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d)) and len(os.listdir(os.path.join(dir, d))) > 0 and "train.txt" in os.listdir(os.path.join(dir, d)) ])
|
||||
return sorted([d for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d)) and len(
|
||||
os.listdir(os.path.join(dir, d))) > 0 and "train.txt" in os.listdir(os.path.join(dir, d))])
|
||||
|
||||
|
||||
def get_training_list(dir="./training/"):
|
||||
return sorted([f'./training/{d}/train.yaml' for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d)) and len(os.listdir(os.path.join(dir, d))) > 0 and "train.yaml" in os.listdir(os.path.join(dir, d)) ])
|
||||
return sorted([f'./training/{d}/train.yaml' for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d)) and len(
|
||||
os.listdir(os.path.join(dir, d))) > 0 and "train.yaml" in os.listdir(os.path.join(dir, d))])
|
||||
|
||||
|
||||
def do_gc():
|
||||
gc.collect()
|
||||
|
|
@ -1329,9 +1529,11 @@ def do_gc():
|
|||
except Exception as e:
|
||||
pass
|
||||
|
||||
|
||||
def pad(num, zeroes):
|
||||
return str(num).zfill(zeroes + 1)
|
||||
|
||||
|
||||
def curl(url):
|
||||
try:
|
||||
req = urllib.request.Request(url, headers={'User-Agent': 'Python'})
|
||||
|
|
@ -1345,6 +1547,7 @@ def curl(url):
|
|||
print(e)
|
||||
return None
|
||||
|
||||
|
||||
def check_for_updates():
|
||||
if not os.path.isfile('./.git/FETCH_HEAD'):
|
||||
print("Cannot check for updates: not from a git repo")
|
||||
|
|
@ -1379,13 +1582,16 @@ def check_for_updates():
|
|||
|
||||
return False
|
||||
|
||||
|
||||
def enumerate_progress(iterable, desc=None, progress=None, verbose=None):
|
||||
if verbose and desc is not None:
|
||||
print(desc)
|
||||
|
||||
if progress is None:
|
||||
return tqdm(iterable, disable=not verbose)
|
||||
return progress.tqdm(iterable, desc=f'{progress.msg_prefix} {desc}' if hasattr(progress, 'msg_prefix') else desc, track_tqdm=True)
|
||||
return progress.tqdm(iterable, desc=f'{progress.msg_prefix} {desc}' if hasattr(progress, 'msg_prefix') else desc,
|
||||
track_tqdm=True)
|
||||
|
||||
|
||||
def notify_progress(message, progress=None, verbose=True):
|
||||
if verbose:
|
||||
|
|
@ -1396,10 +1602,12 @@ def notify_progress(message, progress=None, verbose=True):
|
|||
|
||||
progress(0, desc=message)
|
||||
|
||||
|
||||
def get_args():
|
||||
global args
|
||||
return args
|
||||
|
||||
|
||||
def setup_args():
|
||||
global args
|
||||
|
||||
|
|
@ -1412,7 +1620,8 @@ def setup_args():
|
|||
'sample-batch-size': None,
|
||||
'embed-output-metadata': True,
|
||||
'latents-lean-and-mean': True,
|
||||
'voice-fixer': False, # getting tired of long initialization times in a Colab for downloading a large dataset for it
|
||||
'voice-fixer': False,
|
||||
# getting tired of long initialization times in a Colab for downloading a large dataset for it
|
||||
'voice-fixer-use-cuda': True,
|
||||
'force-cpu-for-conditioning-latents': False,
|
||||
'defer-tts-load': False,
|
||||
|
|
@ -1443,32 +1652,61 @@ def setup_args():
|
|||
pass
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--share", action='store_true', default=default_arguments['share'], help="Lets Gradio return a public URL to use anywhere")
|
||||
parser.add_argument("--share", action='store_true', default=default_arguments['share'],
|
||||
help="Lets Gradio return a public URL to use anywhere")
|
||||
parser.add_argument("--listen", default=default_arguments['listen'], help="Path for Gradio to listen on")
|
||||
parser.add_argument("--check-for-updates", action='store_true', default=default_arguments['check-for-updates'], help="Checks for update on startup")
|
||||
parser.add_argument("--models-from-local-only", action='store_true', default=default_arguments['models-from-local-only'], help="Only loads models from disk, does not check for updates for models")
|
||||
parser.add_argument("--low-vram", action='store_true', default=default_arguments['low-vram'], help="Disables some optimizations that increases VRAM usage")
|
||||
parser.add_argument("--no-embed-output-metadata", action='store_false', default=not default_arguments['embed-output-metadata'], help="Disables embedding output metadata into resulting WAV files for easily fetching its settings used with the web UI (data is stored in the lyrics metadata tag)")
|
||||
parser.add_argument("--latents-lean-and-mean", action='store_true', default=default_arguments['latents-lean-and-mean'], help="Exports the bare essentials for latents.")
|
||||
parser.add_argument("--voice-fixer", action='store_true', default=default_arguments['voice-fixer'], help="Uses python module 'voicefixer' to improve audio quality, if available.")
|
||||
parser.add_argument("--voice-fixer-use-cuda", action='store_true', default=default_arguments['voice-fixer-use-cuda'], help="Hints to voicefixer to use CUDA, if available.")
|
||||
parser.add_argument("--force-cpu-for-conditioning-latents", default=default_arguments['force-cpu-for-conditioning-latents'], action='store_true', help="Forces computing conditional latents to be done on the CPU (if you constantyl OOM on low chunk counts)")
|
||||
parser.add_argument("--defer-tts-load", default=default_arguments['defer-tts-load'], action='store_true', help="Defers loading TTS model")
|
||||
parser.add_argument("--prune-nonfinal-outputs", default=default_arguments['prune-nonfinal-outputs'], action='store_true', help="Deletes non-final output files on completing a generation")
|
||||
parser.add_argument("--use-bigvgan-vocoder", default=default_arguments['use-bigvgan-vocoder'], action='store_true', help="Uses BigVGAN in place of the default vocoder")
|
||||
parser.add_argument("--device-override", default=default_arguments['device-override'], help="A device string to override pass through Torch")
|
||||
parser.add_argument("--sample-batch-size", default=default_arguments['sample-batch-size'], type=int, help="Sets how many batches to use during the autoregressive samples pass")
|
||||
parser.add_argument("--concurrency-count", type=int, default=default_arguments['concurrency-count'], help="How many Gradio events to process at once")
|
||||
parser.add_argument("--autocalculate-voice-chunk-duration-size", type=float, default=default_arguments['autocalculate-voice-chunk-duration-size'], help="Number of seconds to suggest voice chunk size for (for example, 100 seconds of audio at 10 seconds per chunk will suggest 10 chunks)")
|
||||
parser.add_argument("--output-sample-rate", type=int, default=default_arguments['output-sample-rate'], help="Sample rate to resample the output to (from 24KHz)")
|
||||
parser.add_argument("--output-volume", type=float, default=default_arguments['output-volume'], help="Adjusts volume of output")
|
||||
parser.add_argument("--check-for-updates", action='store_true', default=default_arguments['check-for-updates'],
|
||||
help="Checks for update on startup")
|
||||
parser.add_argument("--models-from-local-only", action='store_true',
|
||||
default=default_arguments['models-from-local-only'],
|
||||
help="Only loads models from disk, does not check for updates for models")
|
||||
parser.add_argument("--low-vram", action='store_true', default=default_arguments['low-vram'],
|
||||
help="Disables some optimizations that increases VRAM usage")
|
||||
parser.add_argument("--no-embed-output-metadata", action='store_false',
|
||||
default=not default_arguments['embed-output-metadata'],
|
||||
help="Disables embedding output metadata into resulting WAV files for easily fetching its settings used with the web UI (data is stored in the lyrics metadata tag)")
|
||||
parser.add_argument("--latents-lean-and-mean", action='store_true',
|
||||
default=default_arguments['latents-lean-and-mean'],
|
||||
help="Exports the bare essentials for latents.")
|
||||
parser.add_argument("--voice-fixer", action='store_true', default=default_arguments['voice-fixer'],
|
||||
help="Uses python module 'voicefixer' to improve audio quality, if available.")
|
||||
parser.add_argument("--voice-fixer-use-cuda", action='store_true',
|
||||
default=default_arguments['voice-fixer-use-cuda'],
|
||||
help="Hints to voicefixer to use CUDA, if available.")
|
||||
parser.add_argument("--force-cpu-for-conditioning-latents",
|
||||
default=default_arguments['force-cpu-for-conditioning-latents'], action='store_true',
|
||||
help="Forces computing conditional latents to be done on the CPU (if you constantyl OOM on low chunk counts)")
|
||||
parser.add_argument("--defer-tts-load", default=default_arguments['defer-tts-load'], action='store_true',
|
||||
help="Defers loading TTS model")
|
||||
parser.add_argument("--prune-nonfinal-outputs", default=default_arguments['prune-nonfinal-outputs'],
|
||||
action='store_true', help="Deletes non-final output files on completing a generation")
|
||||
parser.add_argument("--use-bigvgan-vocoder", default=default_arguments['use-bigvgan-vocoder'], action='store_true',
|
||||
help="Uses BigVGAN in place of the default vocoder")
|
||||
parser.add_argument("--device-override", default=default_arguments['device-override'],
|
||||
help="A device string to override pass through Torch")
|
||||
parser.add_argument("--sample-batch-size", default=default_arguments['sample-batch-size'], type=int,
|
||||
help="Sets how many batches to use during the autoregressive samples pass")
|
||||
parser.add_argument("--concurrency-count", type=int, default=default_arguments['concurrency-count'],
|
||||
help="How many Gradio events to process at once")
|
||||
parser.add_argument("--autocalculate-voice-chunk-duration-size", type=float,
|
||||
default=default_arguments['autocalculate-voice-chunk-duration-size'],
|
||||
help="Number of seconds to suggest voice chunk size for (for example, 100 seconds of audio at 10 seconds per chunk will suggest 10 chunks)")
|
||||
parser.add_argument("--output-sample-rate", type=int, default=default_arguments['output-sample-rate'],
|
||||
help="Sample rate to resample the output to (from 24KHz)")
|
||||
parser.add_argument("--output-volume", type=float, default=default_arguments['output-volume'],
|
||||
help="Adjusts volume of output")
|
||||
|
||||
parser.add_argument("--autoregressive-model", default=default_arguments['autoregressive-model'], help="Specifies which autoregressive model to use for sampling.")
|
||||
parser.add_argument("--whisper-model", default=default_arguments['whisper-model'], help="Specifies which whisper model to use for transcription.")
|
||||
parser.add_argument("--whisper-cpp", default=default_arguments['whisper-cpp'], action='store_true', help="Leverages lightmare/whispercpp for transcription")
|
||||
parser.add_argument("--autoregressive-model", default=default_arguments['autoregressive-model'],
|
||||
help="Specifies which autoregressive model to use for sampling.")
|
||||
parser.add_argument("--whisper-model", default=default_arguments['whisper-model'],
|
||||
help="Specifies which whisper model to use for transcription.")
|
||||
parser.add_argument("--whisper-cpp", default=default_arguments['whisper-cpp'], action='store_true',
|
||||
help="Leverages lightmare/whispercpp for transcription")
|
||||
|
||||
parser.add_argument("--training-default-halfp", action='store_true', default=default_arguments['training-default-halfp'], help="Training default: halfp")
|
||||
parser.add_argument("--training-default-bnb", action='store_true', default=default_arguments['training-default-bnb'], help="Training default: bnb")
|
||||
parser.add_argument("--training-default-halfp", action='store_true',
|
||||
default=default_arguments['training-default-halfp'], help="Training default: halfp")
|
||||
parser.add_argument("--training-default-bnb", action='store_true',
|
||||
default=default_arguments['training-default-bnb'], help="Training default: bnb")
|
||||
|
||||
parser.add_argument("--os", default="unix", help="Specifies which OS, easily")
|
||||
args = parser.parse_args()
|
||||
|
|
@ -1478,7 +1716,6 @@ def setup_args():
|
|||
if not args.device_override:
|
||||
set_device_name(args.device_override)
|
||||
|
||||
|
||||
args.listen_host = None
|
||||
args.listen_port = None
|
||||
args.listen_path = None
|
||||
|
|
@ -1499,7 +1736,12 @@ def setup_args():
|
|||
|
||||
return args
|
||||
|
||||
def update_args( listen, share, check_for_updates, models_from_local_only, low_vram, embed_output_metadata, latents_lean_and_mean, voice_fixer, voice_fixer_use_cuda, force_cpu_for_conditioning_latents, defer_tts_load, prune_nonfinal_outputs, use_bigvgan_vocoder, device_override, sample_batch_size, concurrency_count, autocalculate_voice_chunk_duration_size, output_volume, autoregressive_model, whisper_model, whisper_cpp, training_default_halfp, training_default_bnb ):
|
||||
|
||||
def update_args(listen, share, check_for_updates, models_from_local_only, low_vram, embed_output_metadata,
|
||||
latents_lean_and_mean, voice_fixer, voice_fixer_use_cuda, force_cpu_for_conditioning_latents,
|
||||
defer_tts_load, prune_nonfinal_outputs, use_bigvgan_vocoder, device_override, sample_batch_size,
|
||||
concurrency_count, autocalculate_voice_chunk_duration_size, output_volume, autoregressive_model,
|
||||
whisper_model, whisper_cpp, training_default_halfp, training_default_bnb):
|
||||
global args
|
||||
|
||||
args.listen = listen
|
||||
|
|
@ -1531,6 +1773,7 @@ def update_args( listen, share, check_for_updates, models_from_local_only, low_v
|
|||
|
||||
save_args_settings()
|
||||
|
||||
|
||||
def save_args_settings():
|
||||
global args
|
||||
settings = {
|
||||
|
|
@ -1567,7 +1810,6 @@ def save_args_settings():
|
|||
f.write(json.dumps(settings, indent='\t'))
|
||||
|
||||
|
||||
|
||||
def import_generate_settings(file="./config/generate.json"):
|
||||
settings, _ = read_generate_settings(file, read_latents=False)
|
||||
|
||||
|
|
@ -1604,6 +1846,7 @@ def reset_generation_settings():
|
|||
f.write(json.dumps({}, indent='\t'))
|
||||
return import_generate_settings()
|
||||
|
||||
|
||||
def read_generate_settings(file, read_latents=True):
|
||||
j = None
|
||||
latents = None
|
||||
|
|
@ -1632,17 +1875,15 @@ def read_generate_settings(file, read_latents=True):
|
|||
latents = base64.b64decode(j['latents'])
|
||||
del j['latents']
|
||||
|
||||
|
||||
if "time" in j:
|
||||
j["time"] = "{:.3f}".format(j["time"])
|
||||
|
||||
|
||||
|
||||
return (
|
||||
j,
|
||||
latents,
|
||||
)
|
||||
|
||||
|
||||
def load_tts(restart=False, model=None):
|
||||
global args
|
||||
global tts
|
||||
|
|
@ -1650,7 +1891,6 @@ def load_tts( restart=False, model=None ):
|
|||
if restart:
|
||||
unload_tts()
|
||||
|
||||
|
||||
if model:
|
||||
args.autoregressive_model = model
|
||||
|
||||
|
|
@ -1672,8 +1912,10 @@ def load_tts( restart=False, model=None ):
|
|||
print("Loaded TorToiSe, ready for generation.")
|
||||
return tts
|
||||
|
||||
|
||||
setup_tortoise = load_tts
|
||||
|
||||
|
||||
def unload_tts():
|
||||
global tts
|
||||
|
||||
|
|
@ -1683,9 +1925,11 @@ def unload_tts():
|
|||
print("Unloaded TTS")
|
||||
do_gc()
|
||||
|
||||
|
||||
def reload_tts(model=None):
|
||||
load_tts(restart=True, model=model)
|
||||
|
||||
|
||||
def update_autoregressive_model(autoregressive_model_path):
|
||||
match = re.findall(r'^\[[a-fA-F0-9]{8}\] (.+?)$', autoregressive_model_path)
|
||||
if match:
|
||||
|
|
@ -1714,7 +1958,8 @@ def update_autoregressive_model(autoregressive_model_path):
|
|||
else:
|
||||
from tortoise.models.autoregressive import UnifiedVoice
|
||||
|
||||
tts.autoregressive_model_path = autoregressive_model_path if autoregressive_model_path and os.path.exists(autoregressive_model_path) else get_model_path('autoregressive.pth', tts.models_dir)
|
||||
tts.autoregressive_model_path = autoregressive_model_path if autoregressive_model_path and os.path.exists(
|
||||
autoregressive_model_path) else get_model_path('autoregressive.pth', tts.models_dir)
|
||||
|
||||
del tts.autoregressive
|
||||
tts.autoregressive = UnifiedVoice(max_mel_tokens=604, max_text_tokens=402, max_conditioning_inputs=2, layers=30,
|
||||
|
|
@ -1735,6 +1980,7 @@ def update_autoregressive_model(autoregressive_model_path):
|
|||
|
||||
return autoregressive_model_path
|
||||
|
||||
|
||||
def load_voicefixer(restart=False):
|
||||
global voicefixer
|
||||
|
||||
|
|
@ -1749,6 +1995,7 @@ def load_voicefixer(restart=False):
|
|||
except Exception as e:
|
||||
print(f"Error occurred while tring to initialize voicefixer: {e}")
|
||||
|
||||
|
||||
def unload_voicefixer():
|
||||
global voicefixer
|
||||
|
||||
|
|
@ -1759,6 +2006,7 @@ def unload_voicefixer():
|
|||
|
||||
do_gc()
|
||||
|
||||
|
||||
def load_whisper_model(language=None, model_name=None, progress=None):
|
||||
global whisper_model
|
||||
|
||||
|
|
@ -1787,6 +2035,7 @@ def load_whisper_model(language=None, model_name=None, progress=None):
|
|||
|
||||
print("Loaded Whisper model")
|
||||
|
||||
|
||||
def unload_whisper():
|
||||
global whisper_model
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user