def train_post():
dataset_name = request.form["path"]
epochs = request.form["epochs"]
batch_size = request.form["batch_size"]
early_stopping = request.form.get("early_stopping") is not None
metadata_path = os.path.join(paths["datasets"], dataset_name,
METADATA_FILE)
audio_folder = os.path.join(paths["datasets"], dataset_name, AUDIO_FOLDER)
checkpoint_folder = os.path.join(paths["models"], dataset_name)
pretrained_folder = os.path.join(paths["pretrained"], dataset_name)
if request.files.get("pretrained_model"):
os.makedirs(pretrained_folder, exist_ok=True)
transfer_learning_path = os.path.join(pretrained_folder,
"pretrained.pt")
request.files["pretrained_model"].save(transfer_learning_path)
else:
transfer_learning_path = None
start_progress_thread(
train,
metadata_path=metadata_path,
dataset_directory=audio_folder,
output_directory=checkpoint_folder,
transfer_learning_path=transfer_learning_path,
epochs=int(epochs),
batch_size=int(batch_size),
early_stopping=early_stopping,
)
return render_template("progress.html",
next_url=get_next_url(URLS, request.path))
def create_dataset_post():
min_confidence = request.form["confidence"]
if request.form["name"]:
output_folder = os.path.join(paths["datasets"], request.form["name"])
if os.path.exists(output_folder):
request.files = None
raise Exception("Dataset name taken")
os.makedirs(output_folder, exist_ok=True)
text_path = os.path.join(output_folder, TEXT_FILE)
audio_path = os.path.join(output_folder, AUDIO_FILE)
forced_alignment_path = os.path.join(output_folder, ALIGNMENT_FILE)
output_path = os.path.join(output_folder, AUDIO_FOLDER)
label_path = os.path.join(output_folder, METADATA_FILE)
info_path = os.path.join(output_folder, INFO_FILE)
request.files["text_file"].save(text_path)
request.files["audio_file"].save(audio_path)
start_progress_thread(
create_dataset,
text_path=text_path,
audio_path=audio_path,
forced_alignment_path=forced_alignment_path,
output_path=output_path,
label_path=label_path,
info_path=info_path,
min_confidence=float(min_confidence),
)
else:
output_folder = os.path.join(paths["datasets"], request.form["path"])
suffix = get_suffix()
text_path = os.path.join(output_folder, f"text-{suffix}.txt")
audio_path = os.path.join(output_folder, f"audio-{suffix}.mp3")
forced_alignment_path = os.path.join(output_folder,
f"align-{suffix}.json")
info_path = os.path.join(output_folder, INFO_FILE)
request.files["text_file"].save(text_path)
request.files["audio_file"].save(audio_path)
existing_output_path = os.path.join(output_folder, AUDIO_FOLDER)
existing_label_path = os.path.join(output_folder, METADATA_FILE)
start_progress_thread(
extend_existing_dataset,
text_path=text_path,
audio_path=audio_path,
forced_alignment_path=forced_alignment_path,
output_path=existing_output_path,
label_path=existing_label_path,
suffix=suffix,
info_path=info_path,
min_confidence=float(min_confidence),
)
return render_template("progress.html",
next_url=get_next_url(URLS, request.path))
def train_post():
language = request.form["language"]
symbols = get_symbols(language)
dataset_name = request.form["dataset"]
epochs = request.form["epochs"]
batch_size = request.form["batch_size"]
early_stopping = request.form.get("early_stopping") is not None
iters_per_checkpoint = request.form["checkpoint_frequency"]
iters_per_backup_checkpoint = request.form["backup_checkpoint_frequency"]
train_size = 1 - float(request.form["validation_size"])
alignment_sentence = request.form["alignment_sentence"]
multi_gpu = request.form.get("multi_gpu") is not None
checkpoint_path = (os.path.join(paths["models"], dataset_name,
request.form["checkpoint"])
if request.form.get("checkpoint") else None)
metadata_path = os.path.join(paths["datasets"], dataset_name,
METADATA_FILE)
use_metadata = os.path.isfile(metadata_path)
trainlist_path = os.path.join(paths["datasets"], dataset_name, TRAIN_FILE)
vallist_path = os.path.join(paths["datasets"], dataset_name,
VALIDATION_FILE)
audio_folder = os.path.join(paths["datasets"], dataset_name, AUDIO_FOLDER)
checkpoint_folder = os.path.join(paths["models"], dataset_name)
if request.files.get("pretrained_model"):
transfer_learning_path = os.path.join("data", "pretrained.pt")
request.files["pretrained_model"].save(transfer_learning_path)
else:
transfer_learning_path = None
start_progress_thread(
train,
metadata_path=metadata_path if use_metadata else None,
trainlist_path=trainlist_path if not use_metadata else None,
vallist_path=vallist_path if not use_metadata else None,
audio_directory=audio_folder,
output_directory=checkpoint_folder,
symbols=symbols,
checkpoint_path=checkpoint_path,
transfer_learning_path=transfer_learning_path,
epochs=int(epochs),
batch_size=int(batch_size),
early_stopping=early_stopping,
multi_gpu=multi_gpu,
iters_per_checkpoint=int(iters_per_checkpoint),
iters_per_backup_checkpoint=int(iters_per_backup_checkpoint),
train_size=train_size,
alignment_sentence=alignment_sentence,
)
return render_template("progress.html",
next_url=get_next_url(URLS, request.path),
voice=Path(checkpoint_folder).stem)
def upload_dataset():
dataset = request.files["dataset"]
dataset.save(TEMP_DATASET_UPLOAD)
dataset_name = request.values["name"]
dataset_directory = os.path.join(paths["datasets"], dataset_name)
audio_folder = os.path.join(dataset_directory, AUDIO_FOLDER)
assert not os.path.isdir(dataset_directory), "Output folder already exists"
start_progress_thread(import_dataset,
dataset=TEMP_DATASET_UPLOAD,
dataset_directory=dataset_directory,
audio_folder=audio_folder)
return render_template("progress.html", next_url="/import-export")
def train_hifigan_post():
dataset_name = request.form["dataset"]
epochs = request.form["epochs"]
batch_size = request.form["batch_size"]
iters_per_checkpoint = request.form["checkpoint_frequency"]
iters_per_backup_checkpoint = request.form["backup_checkpoint_frequency"]
train_size = 1 - float(request.form["validation_size"])
audio_folder = os.path.join(paths["datasets"], dataset_name, AUDIO_FOLDER)
output_directory = os.path.join(paths["hifigan_training"], dataset_name)
if request.form.get("checkpoint_iteration"):
checkpoint_g = os.path.join(
paths["hifigan_training"], dataset_name,
f"g_{request.form['checkpoint_iteration']}")
checkpoint_do = os.path.join(
paths["hifigan_training"], dataset_name,
f"do_{request.form['checkpoint_iteration']}")
elif request.files.get("pretrained_model_g"):
checkpoint_g = os.path.join("data", "pretrained_model_g.pt")
checkpoint_do = os.path.join("data", "pretrained_model_do.pt")
request.files["pretrained_model_g"].save(checkpoint_g)
request.files["pretrained_model_do"].save(checkpoint_do)
else:
checkpoint_g = None
checkpoint_do = None
start_progress_thread(
train_hifigan,
audio_folder=audio_folder,
output_directory=output_directory,
checkpoint_g=checkpoint_g,
checkpoint_do=checkpoint_do,
epochs=int(epochs),
batch_size=int(batch_size),
iters_per_checkpoint=int(iters_per_checkpoint),
iters_per_backup_checkpoint=int(iters_per_backup_checkpoint),
train_size=train_size,
)
return render_template("progress.html", next_url="/synthesis")
def create_dataset_post():
min_confidence = float(request.form["confidence"])
language = request.form["language"]
combine_clips = request.form.get("combine_clips") is not None
min_length = float(request.form["min_length"])
max_length = float(request.form["max_length"])
transcription_model = (
Silero(language) if language in SILERO_LANGUAGES else DeepSpeech(
os.path.join(paths["languages"], language, TRANSCRIPTION_MODEL)))
symbols = get_symbols(language)
text_file = SUBTITLE_FILE if request.files["text_file"].filename.endswith(
".srt") else TEXT_FILE
if request.form["name"]:
output_folder = os.path.join(paths["datasets"], request.form["name"])
if os.path.exists(output_folder):
request.files = None
raise Exception("Dataset name taken")
os.makedirs(output_folder, exist_ok=True)
text_path = os.path.join(output_folder, text_file)
audio_path = os.path.join(output_folder,
request.files["audio_file"].filename)
with open(text_path, "w", encoding=CHARACTER_ENCODING) as f:
f.write(request.files["text_file"].read().decode(
CHARACTER_ENCODING, "ignore").replace("\r\n", "\n"))
request.files["audio_file"].save(audio_path)
start_progress_thread(
create_dataset,
text_path=text_path,
audio_path=audio_path,
transcription_model=transcription_model,
output_folder=output_folder,
min_length=min_length,
max_length=max_length,
min_confidence=min_confidence,
combine_clips=combine_clips,
symbols=symbols,
)
else:
output_folder = os.path.join(paths["datasets"],
request.form["dataset"])
suffix = get_suffix()
text_path = os.path.join(output_folder, add_suffix(text_file, suffix))
audio_path = os.path.join(
output_folder,
add_suffix(request.files["audio_file"].filename, suffix))
with open(text_path, "w", encoding=CHARACTER_ENCODING) as f:
f.write(request.files["text_file"].read().decode(
CHARACTER_ENCODING, "ignore").replace("\r\n", "\n"))
request.files["audio_file"].save(audio_path)
start_progress_thread(
extend_existing_dataset,
text_path=text_path,
audio_path=audio_path,
transcription_model=transcription_model,
output_folder=output_folder,
suffix=suffix,
min_length=min_length,
max_length=max_length,
min_confidence=min_confidence,
combine_clips=combine_clips,
symbols=symbols,
)
return render_template("progress.html",
next_url=get_next_url(URLS, request.path))
