def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
duration_query="*-durations.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
duration_load_fn=np.load,
mel_length_threshold=0,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
duration_query (str): Query to find duration files in root_dir.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
duration_load_fn (func): Function to load duration file.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
duration_files = sorted(find_files(root_dir, duration_query))
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert (
len(mel_files) == len(charactor_files) == len(duration_files)
), f"Number of charactor, mel and duration files are different \
({len(mel_files)} vs {len(charactor_files)} vs {len(duration_files)})."
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "")
for f in charactor_files
]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.duration_files = duration_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.duration_load_fn = duration_load_fn
self.mel_length_threshold = mel_length_threshold
def __init__(
self,
root_dir,
audio_query="*-wave.npy",
mel_query="*-raw-feats.npy",
audio_load_fn=np.load,
mel_load_fn=np.load,
audio_length_threshold=0,
mel_length_threshold=0,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
audio_query (str): Query to find audio files in root_dir.
mel_query (str): Query to find feature files in root_dir.
audio_load_fn (func): Function to load audio file.
mel_load_fn (func): Function to load feature file.
audio_length_threshold (int): Threshold to remove short audio files.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of audio and mel files.
audio_files = sorted(find_files(root_dir, audio_query))
mel_files = sorted(find_files(root_dir, mel_query))
# assert the number of files
assert len(
audio_files) != 0, f"Not found any audio files in ${root_dir}."
assert len(audio_files) == len(
mel_files
), f"Number of audio and mel files are different ({len(audio_files)} vs {len(mel_files)})."
if ".npy" in audio_query:
suffix = audio_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "") for f in audio_files
]
# set global params
self.utt_ids = utt_ids
self.audio_files = audio_files
self.mel_files = mel_files
self.audio_load_fn = audio_load_fn
self.mel_load_fn = mel_load_fn
self.audio_length_threshold = audio_length_threshold
self.mel_length_threshold = mel_length_threshold
def __init__(
self,
root_dir,
mel_query="*-raw-feats.h5",
mel_load_fn=np.load,
mel_length_threshold=0,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
mel_query (str): Query to find feature files in root_dir.
mel_load_fn (func): Function to load feature file.
mel_length_threshold (int): Threshold to remove short feature files.
"""
# find all of mel files.
mel_files = sorted(find_files(root_dir, mel_query))
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
# assert the number of files
assert len(mel_files) != 0, f"Not found any mel files in ${root_dir}."
if ".npy" in mel_query:
suffix = mel_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "") for f in mel_files
]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.mel_lengths = mel_lengths
self.mel_load_fn = mel_load_fn
self.mel_length_threshold = mel_length_threshold
def main():
args = parse_args()
args.train_dataset = args.eval_dataset = args.dataset
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
with open(config['speech_config']) as f:
mel_config = yaml.load(f, Loader=yaml.Loader)
config.update(mel_config)
config.update(vars(args))
model = GE2E(name='ge2e', **config['model'])
model.load_weights(args.restore)
window_length = (config['data']['min_frames'] + config['data']['max_frames']) // 2
suffix = '*_mel.npy'
save_as = '_gc.npy'
for dataset in args.dataset:
files = find_files(dataset, suffix)
print('files of %s:'% dataset, len(files), files[0])
for file in tqdm(files):
save_name = file.replace(suffix[1:], save_as)
# if os.path.isfile(save_name):
# continue
s = np.load(file)
len_s = len(s)
if len_s < window_length:
s = np.concatenate([s for _ in range((window_length + len_s) // len_s)])
s = batch_frames(s, window_length=window_length, overlap=0.5)
d = model.inference(s)
np.save(save_name, d)
def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
charactor_load_fn=np.load,
return_utt_id=False,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
charactor_load_fn (func): Function to load charactor file.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
# assert the number of files
assert (
len(charactor_files) != 0
), f"Not found any char or duration files in ${root_dir}."
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [os.path.basename(f).replace(suffix, "") for f in charactor_files]
# set global params
self.utt_ids = utt_ids
self.charactor_files = charactor_files
self.charactor_load_fn = charactor_load_fn
self.return_utt_id = return_utt_id
def __init__(self, roots, **kwargs):
super().__init__(**kwargs)
self._metadata = []
for root in roots:
self._metadata += find_files(root, '*_mel.npy')
self._metadata = sorted(self._metadata)
self._get_speaker_from_filename = lambda filename: \
filename.split('/')[-3] if 'libri' in filename.lower() else filename.split('/')[-2]
def __init__(
self,
root_dir,
audio_query="*-wave.npy",
audio_load_fn=np.load,
audio_length_threshold=None,
return_utt_id=False,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
audio_query (str): Query to find feature files in root_dir.
audio_load_fn (func): Function to load feature file.
audio_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of mel files.
audio_files = sorted(find_files(root_dir, audio_query))
audio_lengths = [audio_load_fn(f).shape[0] for f in audio_files]
# filter by threshold
if audio_length_threshold is not None:
idxs = [
idx for idx in range(len(audio_files))
if audio_lengths[idx] > audio_length_threshold
]
if len(audio_files) != len(idxs):
logging.warning(
f"Some files are filtered by mel length threshold "
f"({len(audio_files)} -> {len(idxs)}).")
audio_files = [audio_files[idx] for idx in idxs]
# assert the number of files
assert len(
audio_files) != 0, f"Not found any mel files in ${root_dir}."
if ".npy" in audio_query:
suffix = audio_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "") for f in audio_files
]
# set global params
self.utt_ids = utt_ids
self.audio_files = audio_files
self.audio_lengths = audio_lengths
self.audio_load_fn = audio_load_fn
self.return_utt_id = return_utt_id
def __init__(self,
root_dir,
mel_query="*-raw-feats.h5",
mel_load_fn=np.load,
mel_length_threshold=None,
return_utt_id=False):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
mel_query (str): Query to find feature files in root_dir.
mel_load_fn (func): Function to load feature file.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of mel files.
mel_files = sorted(find_files(root_dir, mel_query))
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
# filter by threshold
if mel_length_threshold is not None:
idxs = [
idx for idx in range(len(mel_files))
if mel_lengths[idx] > mel_length_threshold
]
if len(mel_files) != len(idxs):
logging.warning(
f"Some files are filtered by mel length threshold "
f"({len(mel_files)} -> {len(idxs)}).")
mel_files = [mel_files[idx] for idx in idxs]
# assert the number of files
assert len(mel_files) != 0, f"Not found any mel files in ${root_dir}."
if ".npy" in mel_query:
utt_ids = [
"-".join([
os.path.basename(f).split("-")[0],
os.path.basename(f).split("-")[1]
]) for f in mel_files
]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.mel_lengths = mel_lengths
self.mel_load_fn = mel_load_fn
self.return_utt_id = return_utt_id
def main():
args = parse_args()
if 'vctk' in args.dataset.lower():
depth, start = 2, 1
elif 'libri' in args.dataset.lower():
depth, start = 3, 13
else:
depth, start = 3, 13
speaker_files = find_files(args.dataset, '_gc.npy', depth)
func = lambda filename: filename.split('/')[-depth]
d = {}
for f in speaker_files:
s = func(f)
if s not in d:
d[s] = []
d[s].append(f)
speaker_files = d
if args.save:
os.makedirs(args.save, exist_ok=True)
out_v = io.open('/'.join([args.save, 'vecs.tsv']), 'w', encoding='utf-8')
out_m = io.open('/'.join([args.save, 'meta.tsv']), 'w', encoding='utf-8')
with open(args.speaker_info) as file:
for line in np.random.choice(file.readlines()[start:], size=args.n):
speaker, info = line.split(maxsplit=1)
if depth == 2:
speaker = 'p' + speaker
if speaker not in speaker_files:
continue
for i, filename in enumerate(
np.random.choice(
speaker_files[speaker],
size=args.m,
replace=args.m < len(speaker_files[speaker]))):
d_vector = np.load(filename)
out_m.write(line.replace(speaker, speaker + ' ' + str(i)))
out_v.write('\t'.join([str(x) for x in d_vector]) + '\n')
out_v.close()
out_m.close()
def main():
args = parse_args()
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
model = TFSpeechFeaturizer(config)
executor = ProcessPoolExecutor(max_workers=cpu_count())
all_filenames = find_files(args.dataset, args.suffix)
futures = []
print('num files total: %d' % len(all_filenames), all_filenames[0])
suffix = args.suffix.replace('*', '')
# for file in all_filenames:
# futures.append(executor.submit(partial(process_file, file, model, suffix)))
# results = [future.result() for future in tqdm(futures)]
for file in tqdm(all_filenames):
process_file(file, model, suffix)
def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
duration_query="*-durations.npy",
f0_query="*-raw-f0.npy",
energy_query="*-raw-energy.npy",
f0_stat="./dump/stats_f0.npy",
energy_stat="./dump/stats_energy.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
duration_load_fn=np.load,
f0_load_fn=np.load,
energy_load_fn=np.load,
mel_length_threshold=0,
speakers_map=None
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
duration_query (str): Query to find duration files in root_dir.
f0_query (str): Query to find f0 files in root_dir.
energy_query (str): Query to find energy files in root_dir.
f0_stat (str): str path of f0_stat.
energy_stat (str): str path of energy_stat.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
duration_load_fn (func): Function to load duration file.
f0_load_fn (func): Function to load f0 file.
energy_load_fn (func): Function to load energy file.
mel_length_threshold (int): Threshold to remove short feature files.
speakers_map (dict): Speakers map generated in dataset preprocessing
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
duration_files = sorted(find_files(root_dir, duration_query))
f0_files = sorted(find_files(root_dir, f0_query))
energy_files = sorted(find_files(root_dir, energy_query))
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert (
len(mel_files)
== len(charactor_files)
== len(duration_files)
== len(f0_files)
== len(energy_files)
), f"Number of charactor, mel, duration, f0 and energy files are different"
assert speakers_map != None, f"No speakers map found. Did you set --dataset_mapping?"
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [os.path.basename(f).replace(suffix, "") for f in charactor_files]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.duration_files = duration_files
self.f0_files = f0_files
self.energy_files = energy_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.duration_load_fn = duration_load_fn
self.f0_load_fn = f0_load_fn
self.energy_load_fn = energy_load_fn
self.mel_length_threshold = mel_length_threshold
self.speakers_map = speakers_map
self.speakers = [self.speakers_map[i.split("_")[0]] for i in self.utt_ids]
print("Speaker: utt_id", list(zip(self.speakers, self.utt_ids)))
self.f0_stat = np.load(f0_stat)
self.energy_stat = np.load(energy_stat)
def __init__(
self,
dataset,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
mel_length_threshold=0,
reduction_factor=1,
mel_pad_value=0.0,
char_pad_value=0,
ga_pad_value=-1.0,
g=0.2,
use_fixed_shapes=False,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
mel_length_threshold (int): Threshold to remove short feature files.
reduction_factor (int): Reduction factor on Tacotron-2 paper.
mel_pad_value (float): Padding value for mel-spectrogram.
char_pad_value (int): Padding value for charactor.
ga_pad_value (float): Padding value for guided attention.
g (float): G value for guided attention.
use_fixed_shapes (bool): Use fixed shape for mel targets or not.
max_char_length (int): maximum charactor length if use_fixed_shapes=True.
max_mel_length (int): maximum mel length if use_fixed_shapes=True
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
char_lengths = [charactor_load_fn(f).shape[0] for f in charactor_files]
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert (
len(mel_files) == len(charactor_files) == len(mel_lengths)
), f"Number of charactor, mel and duration files are different \
({len(mel_files)} vs {len(charactor_files)} vs {len(mel_lengths)})."
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "")
for f in charactor_files
]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.mel_lengths = mel_lengths
self.char_lengths = char_lengths
self.reduction_factor = reduction_factor
self.mel_length_threshold = mel_length_threshold
self.mel_pad_value = mel_pad_value
self.char_pad_value = char_pad_value
self.ga_pad_value = ga_pad_value
self.g = g
self.use_fixed_shapes = use_fixed_shapes
self.max_char_length = np.max(char_lengths)
if np.max(mel_lengths) % self.reduction_factor == 0:
self.max_mel_length = np.max(mel_lengths)
else:
self.max_mel_length = (np.max(mel_lengths) +
self.reduction_factor -
np.max(mel_lengths) % self.reduction_factor)
return extract_from_mel(features)
@tf.function(input_signature=[
tf.TensorSpec(shape=[None, speech_config['n_mels'], 1],
dtype=tf.float32,
name="signal")
])
def extract_from_mel(features):
with tf.device('/cpu:0'):
encoded = conformer.encoder_inference(features)
return encoded
suffix = '.wav'
mel_query = '_mel.npy'
feature_query = '_conformer_enc16.npy'
audio_files = sorted(find_files(args.dataset, '*' + suffix))
print('files:', len(audio_files), audio_files[0])
for filename in tqdm(audio_files):
mel = filename.replace(suffix, mel_query)
if os.path.exists(mel):
features = np.load(mel).reshape([-1, speech_config['n_mels'], 1])
encoded = extract_from_mel(features)
else:
signal = read_raw_audio(filename)
encoded = extract_from_audio(signal)
np.save(filename.replace(suffix, feature_query), encoded)
def __init__(self,
root_dir,
audio_query="*-wave.npy",
mel_query="*-raw-feats.npy",
audio_load_fn=np.load,
mel_load_fn=np.load,
audio_length_threshold=None,
mel_length_threshold=None,
return_utt_id=False
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
audio_query (str): Query to find audio files in root_dir.
mel_query (str): Query to find feature files in root_dir.
audio_load_fn (func): Function to load audio file.
mel_load_fn (func): Function to load feature file.
audio_length_threshold (int): Threshold to remove short audio files.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of audio and mel files.
audio_files = sorted(find_files(root_dir, audio_query))
mel_files = sorted(find_files(root_dir, mel_query))
# filter by threshold
if audio_length_threshold is not None:
audio_lengths = [audio_load_fn(f).shape[0] for f in audio_files]
idxs = [idx for idx in range(len(audio_files)) if audio_lengths[idx] > audio_length_threshold]
if len(audio_files) != len(idxs):
logging.warning(f"Some files are filtered by audio length threshold "
f"({len(audio_files)} -> {len(idxs)}).")
audio_files = [audio_files[idx] for idx in idxs]
mel_files = [mel_files[idx] for idx in idxs]
if mel_length_threshold is not None:
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
idxs = [idx for idx in range(len(mel_files)) if mel_lengths[idx] > mel_length_threshold]
if len(mel_files) != len(idxs):
logging.warning(f"Some files are filtered by mel length threshold "
f"({len(mel_files)} -> {len(idxs)}).")
audio_files = [audio_files[idx] for idx in idxs]
mel_files = [mel_files[idx] for idx in idxs]
# assert the number of files
assert len(audio_files) != 0, f"Not found any audio files in ${root_dir}."
assert len(audio_files) == len(mel_files), \
f"Number of audio and mel files are different ({len(audio_files)} vs {len(mel_files)})."
if ".npy" in audio_query:
utt_ids = [os.path.basename(f).replace("-wave.npy", "") for f in audio_files]
else:
utt_ids = [os.path.splitext(os.path.basename(f))[0] for f in audio_files]
# set global params
self.utt_ids = utt_ids
self.audio_files = audio_files
self.mel_files = mel_files
self.audio_load_fn = audio_load_fn
self.mel_load_fn = mel_load_fn
self.return_utt_id = return_utt_id
def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
duration_query="*-durations.npy",
f0_query="*-raw-f0.npy",
energy_query="*-raw-energy.npy",
f0_stat="./dump/stats_f0.npy",
energy_stat="./dump/stats_energy.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
duration_load_fn=np.load,
f0_load_fn=np.load,
energy_load_fn=np.load,
mel_length_threshold=0,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
duration_query (str): Query to find duration files in root_dir.
f0_query (str): Query to find f0 files in root_dir.
energy_query (str): Query to find energy files in root_dir.
f0_stat (str): str path of f0_stat.
energy_stat (str): str path of energy_stat.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
duration_load_fn (func): Function to load duration file.
f0_load_fn (func): Function to load f0 file.
energy_load_fn (func): Function to load energy file.
mel_length_threshold (int): Threshold to remove short feature files.
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
duration_files = sorted(find_files(root_dir, duration_query))
f0_files = sorted(find_files(root_dir, f0_query))
energy_files = sorted(find_files(root_dir, energy_query))
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert (
len(mel_files)
== len(charactor_files)
== len(duration_files)
== len(f0_files)
== len(energy_files)
), f"Number of charactor, mel, duration, f0 and energy files are different"
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [os.path.basename(f).replace(suffix, "") for f in charactor_files]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.duration_files = duration_files
self.f0_files = f0_files
self.energy_files = energy_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.duration_load_fn = duration_load_fn
self.f0_load_fn = f0_load_fn
self.energy_load_fn = energy_load_fn
self.mel_length_threshold = mel_length_threshold
self.speakers_map = {} # TODO
sp_id = 0
for i in self.utt_ids:
sp_name = i.split("_")[0]
if sp_name not in self.speakers_map:
self.speakers_map[sp_name] = sp_id
sp_id += 1
self.speakers = [
self.speakers_map[i.split("_")[0]] for i in self.utt_ids
] # TODO change but at the moment mfa folder name = speaker name
self.f0_stat = np.load(f0_stat)
self.energy_stat = np.load(energy_stat)
def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
duration_query="*-durations.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
duration_load_fn=np.load,
mel_length_threshold=None,
return_utt_id=False,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
duration_query (str): Query to find duration files in root_dir.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
duration_load_fn (func): Function to load duration file.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
duration_files = sorted(find_files(root_dir, duration_query))
# filter by threshold
if mel_length_threshold is not None:
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
idxs = [
idx
for idx in range(len(mel_files))
if mel_lengths[idx] > mel_length_threshold
]
if len(mel_files) != len(idxs):
logging.warning(
f"Some files are filtered by mel length threshold "
f"({len(mel_files)} -> {len(idxs)})."
)
mel_files = [mel_files[idx] for idx in idxs]
charactor_files = [charactor_files[idx] for idx in idxs]
duration_files = [duration_files[idx] for idx in idxs]
mel_lengths = [mel_lengths[idx] for idx in idxs]
# bucket sequence length trick, sort based-on mel-length.
idx_sort = np.argsort(mel_lengths)
# sort
mel_files = np.array(mel_files)[idx_sort]
charactor_files = np.array(charactor_files)[idx_sort]
duration_files = np.array(duration_files)[idx_sort]
mel_lengths = np.array(mel_lengths)[idx_sort]
# group
idx_lengths = [
[idx, length]
for idx, length in zip(np.arange(len(mel_lengths)), mel_lengths)
]
groups = [
list(g) for _, g in itertools.groupby(idx_lengths, lambda a: a[1])
]
# group shuffle
random.shuffle(groups)
# get idxs affter group shuffle
idxs = []
for group in groups:
for idx, _ in group:
idxs.append(idx)
# re-arange dataset
mel_files = np.array(mel_files)[idxs]
charactor_files = np.array(charactor_files)[idxs]
duration_files = np.array(duration_files)[idxs]
mel_lengths = np.array(mel_lengths)[idxs]
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert (
len(mel_files) == len(charactor_files) == len(duration_files)
), f"Number of charactor, mel and duration files are different \
({len(mel_files)} vs {len(charactor_files)} vs {len(duration_files)})."
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [os.path.basename(f).replace(suffix, "") for f in charactor_files]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.duration_files = duration_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.duration_load_fn = duration_load_fn
self.return_utt_id = return_utt_id
def __init__(
self,
root_dir,
charactor_query="*-ids.npy",
mel_query="*-norm-feats.npy",
charactor_load_fn=np.load,
mel_load_fn=np.load,
mel_length_threshold=None,
return_utt_id=False,
return_guided_attention=True,
reduction_factor=1,
mel_pad_value=0.0,
char_pad_value=0,
ga_pad_value=-1.0,
g=0.2,
use_fixed_shapes=False,
):
"""Initialize dataset.
Args:
root_dir (str): Root directory including dumped files.
charactor_query (str): Query to find charactor files in root_dir.
mel_query (str): Query to find feature files in root_dir.
charactor_load_fn (func): Function to load charactor file.
mel_load_fn (func): Function to load feature file.
mel_length_threshold (int): Threshold to remove short feature files.
return_utt_id (bool): Whether to return the utterance id with arrays.
reduction_factor (int): Reduction factor on Tacotron-2 paper.
mel_pad_value (float): Padding value for mel-spectrogram.
char_pad_value (int): Padding value for charactor.
ga_pad_value (float): Padding value for guided attention.
g (float): G value for guided attention.
use_fixed_shapes (bool): Use fixed shape for mel targets or not.
max_char_length (int): maximum charactor length if use_fixed_shapes=True.
max_mel_length (int): maximum mel length if use_fixed_shapes=True
"""
# find all of charactor and mel files.
charactor_files = sorted(find_files(root_dir, charactor_query))
mel_files = sorted(find_files(root_dir, mel_query))
mel_lengths = [mel_load_fn(f).shape[0] for f in mel_files]
char_lengths = [charactor_load_fn(f).shape[0] for f in charactor_files]
# filter by threshold
if mel_length_threshold is not None:
idxs = [
idx for idx in range(len(mel_files))
if mel_lengths[idx] > mel_length_threshold
]
if len(mel_files) != len(idxs):
logging.warning(
f"Some files are filtered by mel length threshold "
f"({len(mel_files)} -> {len(idxs)}).")
mel_files = [mel_files[idx] for idx in idxs]
charactor_files = [charactor_files[idx] for idx in idxs]
mel_lengths = [mel_lengths[idx] for idx in idxs]
char_lengths = [char_lengths[idx] for idx in idxs]
# bucket sequence length trick, sort based-on mel-length.
idx_sort = np.argsort(mel_lengths)
# sort
mel_files = np.array(mel_files)[idx_sort]
charactor_files = np.array(charactor_files)[idx_sort]
mel_lengths = np.array(mel_lengths)[idx_sort]
char_lengths = np.array(char_lengths)[idx_sort]
# group
idx_lengths = [[
idx, length
] for idx, length in zip(np.arange(len(mel_lengths)), mel_lengths)]
groups = [
list(g)
for _, g in itertools.groupby(idx_lengths, lambda a: a[1])
]
# group shuffle
random.shuffle(groups)
# get idxs affter group shuffle
idxs = []
for group in groups:
for idx, _ in group:
idxs.append(idx)
# re-arange dataset
mel_files = np.array(mel_files)[idxs]
charactor_files = np.array(charactor_files)[idxs]
mel_lengths = np.array(mel_lengths)[idxs]
char_lengths = np.array(char_lengths)[idxs]
# assert the number of files
assert len(mel_files) != 0, f"Not found any mels files in ${root_dir}."
assert len(mel_files) == len(charactor_files) == len(mel_lengths), \
f"Number of charactor, mel and duration files are different \
({len(mel_files)} vs {len(charactor_files)} vs {len(mel_lengths)})."
if ".npy" in charactor_query:
suffix = charactor_query[1:]
utt_ids = [
os.path.basename(f).replace(suffix, "")
for f in charactor_files
]
# set global params
self.utt_ids = utt_ids
self.mel_files = mel_files
self.charactor_files = charactor_files
self.mel_load_fn = mel_load_fn
self.charactor_load_fn = charactor_load_fn
self.return_utt_id = return_utt_id
self.return_guided_attention = return_guided_attention
self.mel_lengths = mel_lengths
self.char_lengths = char_lengths
self.reduction_factor = reduction_factor
self.mel_pad_value = mel_pad_value
self.char_pad_value = char_pad_value
self.ga_pad_value = ga_pad_value
self.g = g
self.use_fixed_shapes = use_fixed_shapes
self.max_char_length = np.max(char_lengths) + 1 # +1 for eos
self.max_mel_length = np.max(
mel_lengths
) + self.reduction_factor - np.max(mel_lengths) % self.reduction_factor
