def create_or_update_weights_map_main(base_dir: str,
prep_name: str,
weights_prep_name: str,
template_map: Optional[str] = None):
prep_dir = get_prepared_dir(base_dir, prep_name)
assert os.path.isdir(prep_dir)
orig_prep_dir = get_prepared_dir(base_dir, weights_prep_name)
assert os.path.isdir(orig_prep_dir)
logger = init_logger()
add_console_out_to_logger(logger)
logger.info(f"Creating/updating weights map for {weights_prep_name}...")
if template_map is not None:
_template_map = SymbolsMap.load(template_map)
else:
_template_map = None
if weights_map_exists(prep_dir, weights_prep_name):
existing_map = load_weights_map(prep_dir, weights_prep_name)
else:
existing_map = None
weights_map, symbols = create_or_update_weights_map(
orig=load_prep_symbol_converter(orig_prep_dir).get_all_symbols(),
dest=load_prep_symbol_converter(prep_dir).get_all_symbols(),
existing_map=existing_map,
template_map=_template_map,
)
save_weights_map(prep_dir, weights_prep_name, weights_map)
save_weights_symbols(prep_dir, weights_prep_name, symbols)
def create_or_update_inference_map_main(base_dir: str,
prep_name: str,
template_map: Optional[str] = None):
logger = init_logger()
add_console_out_to_logger(logger)
logger.info("Creating/updating inference map...")
prep_dir = get_prepared_dir(base_dir, prep_name)
assert os.path.isdir(prep_dir)
all_symbols = get_all_symbols(prep_dir)
if template_map is not None:
_template_map = SymbolsMap.load(template_map)
else:
_template_map = None
if infer_map_exists(prep_dir):
existing_map = load_infer_map(prep_dir)
else:
existing_map = None
infer_map, symbols = create_or_update_inference_map(
orig=load_prep_symbol_converter(prep_dir).get_all_symbols(),
dest=all_symbols,
existing_map=existing_map,
template_map=_template_map,
)
save_infer_map(prep_dir, infer_map)
save_infer_symbols(prep_dir, symbols)
def dl_pretrained(base_dir: str, train_name: str = DEFAULT_WAVEGLOW, prep_name: Optional[str] = None, version: int = 3):
train_dir = get_train_dir(base_dir, train_name, create=True)
assert os.path.isdir(train_dir)
checkpoints_dir = get_checkpoints_dir(train_dir)
dest_path = get_checkpoint_pretrained(checkpoints_dir)
print("Downloading pretrained waveglow model from Nvida...")
dl_wg(
destination=dest_path,
version=version
)
print("Pretrained model is now beeing converted to be able to use it...")
convert_glow(
origin=dest_path,
destination=dest_path,
keep_orig=False
)
if prep_name is not None:
prep_dir = get_prepared_dir(base_dir, prep_name)
wholeset = load_filelist(prep_dir)
save_testset(train_dir, wholeset)
save_valset(train_dir, wholeset)
save_prep_name(train_dir, prep_name=prep_name)
def convert_model(base_dir: str, prep_name: str, model_path: str,
custom_hparams: Optional[Dict[str, str]]):
prep_dir = get_prepared_dir(base_dir, prep_name)
convert_v1_to_v2_model(old_model_path=model_path,
custom_hparams=custom_hparams,
speakers=load_prep_speakers_json(prep_dir),
accents=load_prep_accents_ids(prep_dir),
symbols=load_prep_symbol_converter(prep_dir))
def map_to_prep_symbols(base_dir: str,
prep_name: str,
text_name: str,
ignore_arcs: bool = True):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
assert os.path.isdir(prep_dir)
assert infer_map_exists(prep_dir)
symb_map_path = get_infer_map_path(prep_dir)
map_text(base_dir, prep_name, text_name, symb_map_path, ignore_arcs)
def validate_main(base_dir: str, train_name: str, waveglow: str = DEFAULT_WAVEGLOW, entry_id: Optional[int] = None, speaker: Optional[str] = None, ds: str = "val", custom_checkpoint: Optional[int] = None, sigma: float = DEFAULT_SIGMA, denoiser_strength: float = DEFAULT_DENOISER_STRENGTH, custom_tacotron_hparams: Optional[Dict[str, str]] = None, custom_waveglow_hparams: Optional[Dict[str, str]] = None):
train_dir = get_train_dir(base_dir, train_name, create=False)
assert os.path.isdir(train_dir)
if ds == "val":
data = load_valset(train_dir)
elif ds == "test":
data = load_testset(train_dir)
else:
assert False
speaker_id: Optional[int] = None
if speaker is not None:
prep_name = load_prep_name(train_dir)
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
speakers = load_prep_speakers_json(prep_dir)
speaker_id = speakers.get_id(speaker)
entry = data.get_for_validation(entry_id, speaker_id)
checkpoint_path, iteration = get_custom_or_last_checkpoint(
get_checkpoints_dir(train_dir), custom_checkpoint)
val_dir = get_val_dir(train_dir, entry, iteration)
logger = prepare_logger(get_val_log(val_dir))
logger.info("Validating...")
taco_checkpoint = CheckpointTacotron.load(checkpoint_path, logger)
train_dir_wg = get_wg_train_dir(base_dir, waveglow, create=False)
wg_checkpoint_path, _ = get_last_checkpoint(get_checkpoints_dir(train_dir_wg))
wg_checkpoint = CheckpointWaveglow.load(wg_checkpoint_path, logger)
result = validate(
tacotron_checkpoint=taco_checkpoint,
waveglow_checkpoint=wg_checkpoint,
sigma=sigma,
denoiser_strength=denoiser_strength,
entry=entry,
logger=logger,
custom_taco_hparams=custom_tacotron_hparams,
custom_wg_hparams=custom_waveglow_hparams
)
orig_mel = get_mel(entry.wav_path, custom_hparams=custom_waveglow_hparams)
save_val_orig_wav(val_dir, entry.wav_path)
save_val_orig_plot(val_dir, orig_mel)
save_val_wav(val_dir, result.sampling_rate, result.wav)
save_val_plot(val_dir, result.mel_outputs)
save_val_pre_postnet_plot(val_dir, result.mel_outputs_postnet)
save_val_alignments_sentence_plot(val_dir, result.alignments)
save_val_comparison(val_dir)
logger.info(f"Saved output to: {val_dir}")
def validate(base_dir: str,
train_name: str,
entry_id: Optional[int] = None,
speaker: Optional[str] = None,
ds: str = "val",
custom_checkpoint: Optional[int] = None,
sigma: float = 0.666,
denoiser_strength: float = 0.00,
custom_hparams: Optional[Dict[str, str]] = None):
train_dir = get_train_dir(base_dir, train_name, create=False)
assert os.path.isdir(train_dir)
if ds == "val":
data = load_valset(train_dir)
elif ds == "test":
data = load_testset(train_dir)
else:
raise Exception()
speaker_id: Optional[int] = None
if speaker is not None:
prep_name = load_prep_name(train_dir)
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
speakers = load_prep_speakers_json(prep_dir)
speaker_id = speakers.get_id(speaker)
entry = data.get_for_validation(entry_id, speaker_id)
checkpoint_path, iteration = get_custom_or_last_checkpoint(
get_checkpoints_dir(train_dir), custom_checkpoint)
val_dir = get_val_dir(train_dir, entry, iteration)
logger = prepare_logger(get_val_log(val_dir))
logger.info(f"Validating {entry.wav_path}...")
checkpoint = CheckpointWaveglow.load(checkpoint_path, logger)
wav, wav_sr, wav_mel, orig_mel = infer(wav_path=entry.wav_path,
denoiser_strength=denoiser_strength,
sigma=sigma,
checkpoint=checkpoint,
custom_hparams=custom_hparams,
logger=logger)
save_val_wav(val_dir, wav_sr, wav)
save_val_plot(val_dir, wav_mel)
save_val_orig_wav(val_dir, entry.wav_path)
save_val_orig_plot(val_dir, orig_mel)
score = save_diff_plot(val_dir)
save_v(val_dir)
logger.info(f"Imagescore: {score*100}%")
logger.info(f"Saved output to: {val_dir}")
def get_infer_sentences(base_dir: str, prep_name: str,
text_name: str) -> InferSentenceList:
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print(f"The text '{text_name}' doesn't exist.")
assert False
result = InferSentenceList.from_sentences(
sentences=load_text_csv(text_dir),
accents=load_prep_accents_ids(prep_dir),
symbols=load_text_symbol_converter(text_dir))
return result
def _accent_template(base_dir: str, prep_name: str, text_name: str):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print("Please add text first.")
else:
print("Updating accent template...")
accented_symbol_list = infer_accents_template(
sentences=load_text_csv(text_dir),
text_symbols=load_text_symbol_converter(text_dir),
accent_ids=load_prep_accents_ids(prep_dir),
)
_save_accents_csv(text_dir, accented_symbol_list)
def _check_for_unknown_symbols(base_dir: str, prep_name: str, text_name: str):
infer_sents = get_infer_sentences(base_dir, prep_name, text_name)
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
logger = prepare_logger()
unknown_symbols_exist = infer_sents.replace_unknown_symbols(
model_symbols=load_prep_symbol_converter(prep_dir), logger=logger)
if unknown_symbols_exist:
logger.info(
"Some symbols are not in the prepared dataset symbolset. You need to create an inference map and then apply it to the symbols."
)
else:
logger.info(
"All symbols are in the prepared dataset symbolset. You can now synthesize this text."
)
def accent_apply(base_dir: str, prep_name: str, text_name: str):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print("Please add text first.")
else:
print("Applying accents...")
updated_sentences = infer_accents_apply(
sentences=load_text_csv(text_dir),
accented_symbols=_load_accents_csv(text_dir),
accent_ids=load_prep_accents_ids(prep_dir),
)
print("\n" + updated_sentences.get_formatted(
symbol_id_dict=load_text_symbol_converter(text_dir),
accent_id_dict=load_prep_accents_ids(prep_dir)))
_save_text_csv(text_dir, updated_sentences)
_check_for_unknown_symbols(base_dir, prep_name, text_name)
def normalize_text(base_dir: str, prep_name: str, text_name: str):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print("Please add text first.")
else:
print("Normalizing text...")
symbol_ids, updated_sentences = infer_norm(
sentences=load_text_csv(text_dir),
text_symbols=load_text_symbol_converter(text_dir))
print("\n" + updated_sentences.get_formatted(
symbol_id_dict=symbol_ids,
accent_id_dict=load_prep_accents_ids(prep_dir)))
_save_text_csv(text_dir, updated_sentences)
save_text_symbol_converter(text_dir, symbol_ids)
_accent_template(base_dir, prep_name, text_name)
_check_for_unknown_symbols(base_dir, prep_name, text_name)
def add_text(base_dir: str, prep_name: str, text_name: str, filepath: str,
lang: Language):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
if not os.path.isdir(prep_dir):
print("Please prepare data first.")
else:
print("Adding text...")
symbol_ids, data = infer_add(
text=read_text(filepath),
lang=lang,
)
print(
"\n" +
data.get_formatted(symbol_id_dict=symbol_ids,
accent_id_dict=load_prep_accents_ids(prep_dir)))
text_dir = get_text_dir(prep_dir, text_name, create=True)
_save_text_csv(text_dir, data)
save_text_symbol_converter(text_dir, symbol_ids)
_accent_template(base_dir, prep_name, text_name)
_check_for_unknown_symbols(base_dir, prep_name, text_name)
def start_new_training(base_dir: str,
train_name: str,
prep_name: str,
test_size: float = 0.01,
validation_size: float = 0.01,
custom_hparams: Optional[Dict[str, str]] = None,
split_seed: int = 1234,
warm_start_train_name: Optional[str] = None,
warm_start_checkpoint: Optional[int] = None):
prep_dir = get_prepared_dir(base_dir, prep_name)
wholeset = load_filelist(prep_dir)
trainset, testset, valset = split_prepared_data_train_test_val(
wholeset,
test_size=test_size,
validation_size=validation_size,
seed=split_seed,
shuffle=True)
train_dir = get_train_dir(base_dir, train_name, create=True)
save_trainset(train_dir, trainset)
save_testset(train_dir, testset)
save_valset(train_dir, valset)
logs_dir = get_train_logs_dir(train_dir)
logger = prepare_logger(get_train_log_file(logs_dir), reset=True)
warm_model = try_load_checkpoint(base_dir=base_dir,
train_name=warm_start_train_name,
checkpoint=warm_start_checkpoint,
logger=logger)
save_prep_name(train_dir, prep_name)
train(
custom_hparams=custom_hparams,
logdir=logs_dir,
trainset=trainset,
valset=valset,
save_checkpoint_dir=get_checkpoints_dir(train_dir),
debug_logger=logger,
warm_model=warm_model,
)
def map_text(base_dir: str,
prep_name: str,
text_name: str,
symbols_map_path: str,
ignore_arcs: bool = True):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print("Please add text first.")
else:
symbol_ids, updated_sentences = sents_map(
sentences=load_text_csv(text_dir),
text_symbols=load_text_symbol_converter(text_dir),
symbols_map=SymbolsMap.load(symbols_map_path),
ignore_arcs=ignore_arcs)
print("\n" + updated_sentences.get_formatted(
symbol_id_dict=symbol_ids,
accent_id_dict=load_prep_accents_ids(prep_dir)))
_save_text_csv(text_dir, updated_sentences)
save_text_symbol_converter(text_dir, symbol_ids)
_accent_template(base_dir, prep_name, text_name)
_check_for_unknown_symbols(base_dir, prep_name, text_name)
def ipa_convert_text(base_dir: str,
prep_name: str,
text_name: str,
ignore_tones: bool = False,
ignore_arcs: bool = True):
prep_dir = get_prepared_dir(base_dir, prep_name, create=False)
text_dir = get_text_dir(prep_dir, text_name, create=False)
if not os.path.isdir(text_dir):
print("Please add text first.")
else:
print("Converting text to IPA...")
symbol_ids, updated_sentences = infer_convert_ipa(
sentences=load_text_csv(text_dir),
text_symbols=load_text_symbol_converter(text_dir),
ignore_tones=ignore_tones,
ignore_arcs=ignore_arcs)
print("\n" + updated_sentences.get_formatted(
symbol_id_dict=symbol_ids,
accent_id_dict=load_prep_accents_ids(prep_dir)))
_save_text_csv(text_dir, updated_sentences)
save_text_symbol_converter(text_dir, symbol_ids)
_accent_template(base_dir, prep_name, text_name)
_check_for_unknown_symbols(base_dir, prep_name, text_name)
def eval_checkpoints_main(base_dir: str, train_name: str, select: int,
min_it: int, max_it: int):
train_dir = get_train_dir(base_dir, train_name, create=False)
assert os.path.isdir(train_dir)
prep_name = load_prep_name(train_dir)
prep_dir = get_prepared_dir(base_dir, prep_name)
symbols_conv = load_prep_symbol_converter(prep_dir)
speakers = load_prep_speakers_json(prep_dir)
accents = load_prep_accents_ids(prep_dir)
logger = prepare_logger()
eval_checkpoints(custom_hparams=None,
checkpoint_dir=get_checkpoints_dir(train_dir),
select=select,
min_it=min_it,
max_it=max_it,
n_symbols=len(symbols_conv),
n_speakers=len(speakers),
n_accents=len(accents),
valset=load_valset(train_dir),
logger=logger)
def train_main(base_dir: str,
train_name: str,
prep_name: str,
warm_start_train_name: Optional[str] = None,
warm_start_checkpoint: Optional[int] = None,
test_size: float = 0.01,
validation_size: float = 0.05,
custom_hparams: Optional[Dict[str, str]] = None,
split_seed: int = 1234,
weights_train_name: Optional[str] = None,
weights_checkpoint: Optional[int] = None,
use_weights_map: Optional[bool] = None,
map_from_speaker: Optional[str] = None):
prep_dir = get_prepared_dir(base_dir, prep_name)
train_dir = get_train_dir(base_dir, train_name, create=True)
logs_dir = get_train_logs_dir(train_dir)
taco_logger = Tacotron2Logger(logs_dir)
logger = prepare_logger(get_train_log_file(logs_dir), reset=True)
checkpoint_logger = prepare_logger(
log_file_path=get_train_checkpoints_log_file(logs_dir),
logger=logging.getLogger("checkpoint-logger"),
reset=True)
save_prep_name(train_dir, prep_name)
trainset, valset = split_dataset(prep_dir=prep_dir,
train_dir=train_dir,
test_size=test_size,
validation_size=validation_size,
split_seed=split_seed)
weights_model = try_load_checkpoint(base_dir=base_dir,
train_name=weights_train_name,
checkpoint=weights_checkpoint,
logger=logger)
weights_map = None
if use_weights_map is not None and use_weights_map:
weights_train_dir = get_train_dir(base_dir, weights_train_name, False)
weights_prep_name = load_prep_name(weights_train_dir)
weights_map = load_weights_map(prep_dir, weights_prep_name)
warm_model = try_load_checkpoint(base_dir=base_dir,
train_name=warm_start_train_name,
checkpoint=warm_start_checkpoint,
logger=logger)
save_callback = partial(
save_checkpoint,
save_checkpoint_dir=get_checkpoints_dir(train_dir),
logger=logger,
)
train(
custom_hparams=custom_hparams,
taco_logger=taco_logger,
symbols=load_prep_symbol_converter(prep_dir),
speakers=load_prep_speakers_json(prep_dir),
accents=load_prep_accents_ids(prep_dir),
trainset=trainset,
valset=valset,
save_callback=save_callback,
weights_map=weights_map,
weights_checkpoint=weights_model,
warm_model=warm_model,
map_from_speaker_name=map_from_speaker,
logger=logger,
checkpoint_logger=checkpoint_logger,
)
