def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Normalize dumped raw features (See detail in parallel_wavegan/bin/normalize.py).")
parser.add_argument("--rootdir", default=None, type=str,
help="directory including feature files to be normalized. "
"you need to specify either *-scp or rootdir.")
parser.add_argument("--wav-scp", default=None, type=str,
help="kaldi-style wav.scp file. "
"you need to specify either *-scp or rootdir.")
parser.add_argument("--feats-scp", default=None, type=str,
help="kaldi-style feats.scp file. "
"you need to specify either *-scp or rootdir.")
parser.add_argument("--segments", default=None, type=str,
help="kaldi-style segments file.")
parser.add_argument("--dumpdir", type=str, required=True,
help="directory to dump normalized feature files.")
parser.add_argument("--stats", type=str, required=True,
help="statistics file.")
parser.add_argument("--skip-wav-copy", default=False, action="store_true",
help="whether to skip the copy of wav files.")
parser.add_argument("--config", type=str, required=True,
help="yaml format configuration file.")
parser.add_argument("--ftype", default='mel', type=str,
help="feature type")
parser.add_argument("--verbose", type=int, default=1,
help="logging level. higher is more logging. (default=1)")
# runtime mode
args = parser.parse_args()
# interactive mode
# args = argparse.ArgumentParser()
# args.wav_scp = None
# args.feats_scp = None
# args.segment = None
# args.dumpdir = ""
# args.skip_wav_copy = True
# args.config = 'egs/so_emo_female/voc1/conf/multi_band_melgan.v2.yaml'
# args.ftype = 'spec'
# args.verbose = 1
# args.rootdir = '/data/evs/VCTK/VCTK-wgan/spec'
# args.stats = '/data/evs/VCTK/VCTK-wgan/spec/mel_mean_std.npy'
# args.rootdir = '/data/evs/Arctic/spec'
# args.stats = '/data/evs/Arctic/spec/spec_mean_std.npy'
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
else:
logging.basicConfig(
level=logging.WARN, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
logging.warning('Skip DEBUG/INFO messages')
# load config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None and args.rootdir is not None) or \
(args.feats_scp is None and args.rootdir is None):
raise ValueError("Please specify either --rootdir or --feats-scp.")
# check directory existence
if args.dumpdir != "":
if not os.path.exists(args.dumpdir):
os.makedirs(args.dumpdir, exist_ok=True)
# get dataset
if args.rootdir is not None:
if config["format"] == "hdf5":
audio_query, mel_query = "*.h5", "*.h5"
audio_load_fn = lambda x: read_hdf5(x, "wave") # NOQA
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
audio_query, mel_query, spc_query = "*.wav.npy", "*.mel.npy", "*.spec.npy"
audio_load_fn = np.load
mel_load_fn = np.load
spc_load_fn = np.load
else:
raise ValueError("support only hdf5 or npy format.")
if not args.skip_wav_copy:
dataset = AudioMelDataset(
root_dir=args.rootdir,
audio_query=audio_query,
mel_query=mel_query,
audio_load_fn=audio_load_fn,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
else:
dataset1 = MelDatasetNew(
root_dir=args.rootdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
dataset2 = SpcDatasetNew(
root_dir=args.rootdir,
spc_query=spc_query,
spc_load_fn=spc_load_fn,
return_utt_id=True,
)
else:
if not args.skip_wav_copy:
dataset = AudioMelSCPDataset(
wav_scp=args.wav_scp,
feats_scp=args.feats_scp,
segments=args.segments,
return_utt_id=True,
)
else:
dataset = MelSCPDataset(
feats_scp=args.feats_scp,
return_utt_id=True,
)
logging.info(f"The number of files in mel dataset = {len(dataset1)}.")
logging.info(f"The number of files in spc dataset = {len(dataset2)}.")
# restore scaler
scaler = StandardScaler()
if config["format"] == "hdf5":
scaler.mean_ = read_hdf5(args.stats, "mean")
scaler.scale_ = read_hdf5(args.stats, "scale")
elif config["format"] == "npy":
scaler.mean_ = np.load(args.stats)[0]
scaler.scale_ = np.load(args.stats)[1]
else:
raise ValueError("support only hdf5 or npy format.")
# from version 0.23.0, this information is needed
scaler.n_features_in_ = scaler.mean_.shape[0]
# process each file
if args.ftype == 'mel':
dataset = dataset1
elif args.ftype == 'spec':
dataset = dataset2
for items in tqdm(dataset):
if not args.skip_wav_copy:
utt_id, audio, feat = items
else:
utt_id, feat, feat_file = items
# normalize
feat = scaler.transform(feat)
# feat = (feat - scaler.mean_) / scaler.scale_ # this is identical to scaler.transform(feat)
# save
if config["format"] == "hdf5":
write_hdf5(os.path.join(args.dumpdir, f"{utt_id}.h5"),
"feats", feat.astype(np.float32))
if not args.skip_wav_copy:
write_hdf5(os.path.join(args.dumpdir, f"{utt_id}.h5"),
"wave", audio.astype(np.float32))
elif config["format"] == "npy":
if args.dumpdir == "":
feat_file = feat_file.replace('.npy', '')
np.save((feat_file + "-norm.npy"),
feat.astype(np.float32), allow_pickle=False)
if not args.skip_wav_copy:
print("Please include --skip_wav_copy in arguments")
else:
np.save(os.path.join(args.dumpdir, f"{utt_id}.npy"),
feat.astype(np.float32), allow_pickle=False)
if not args.skip_wav_copy:
np.save(os.path.join(args.dumpdir, f"{utt_id}.wav.npy"),
audio.astype(np.float32), allow_pickle=False)
else:
raise ValueError("support only hdf5 or npy format.")
def main():
"""Run decoding process."""
parser = argparse.ArgumentParser(
description=
"Decode dumped features with trained Parallel WaveGAN Generator "
"(See detail in parallel_wavegan/bin/decode.py).")
parser.add_argument("--feats-scp",
"--scp",
default=None,
type=str,
help="kaldi-style feats.scp file. "
"you need to specify either feats-scp or dumpdir.")
parser.add_argument("--dumpdir",
default=None,
type=str,
help="directory including feature files. "
"you need to specify either feats-scp or dumpdir.")
parser.add_argument("--outdir",
type=str,
required=True,
help="directory to save generated speech.")
parser.add_argument("--checkpoint",
type=str,
required=True,
help="checkpoint file to be loaded.")
parser.add_argument(
"--config",
default=None,
type=str,
help="yaml format configuration file. if not explicitly provided, "
"it will be searched in the checkpoint directory. (default=None)")
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)")
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
else:
logging.basicConfig(
level=logging.WARN,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
logging.warning("Skip DEBUG/INFO messages")
# check directory existence
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# load config
if args.config is None:
dirname = os.path.dirname(args.checkpoint)
args.config = os.path.join(dirname, "config.yml")
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None and args.dumpdir is not None) or \
(args.feats_scp is None and args.dumpdir is None):
raise ValueError("Please specify either --dumpdir or --feats-scp.")
# get dataset
if args.dumpdir is not None:
if config["format"] == "hdf5":
mel_query = "*.h5"
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
mel_query = "*-feats.npy"
mel_load_fn = np.load
else:
raise ValueError("Support only hdf5 or npy format.")
dataset = MelDataset(
args.dumpdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
else:
dataset = MelSCPDataset(
feats_scp=args.feats_scp,
return_utt_id=True,
)
logging.info(f"The number of features to be decoded = {len(dataset)}.")
# setup model
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
model = load_model(args.checkpoint, config)
logging.info(f"Loaded model parameters from {args.checkpoint}.")
model.remove_weight_norm()
model = model.eval().to(device)
# start generation
total_rtf = 0.0
with torch.no_grad(), tqdm(dataset, desc="[decode]") as pbar:
for idx, (utt_id, c) in enumerate(pbar, 1):
# generate
c = torch.tensor(c, dtype=torch.float).to(device)
start = time.time()
y = model.inference(c).view(-1)
rtf = (time.time() - start) / (len(y) / config["sampling_rate"])
pbar.set_postfix({"RTF": rtf})
total_rtf += rtf
# save as PCM 16 bit wav file
sf.write(os.path.join(config["outdir"], f"{utt_id}_gen.wav"),
y.cpu().numpy(), config["sampling_rate"], "PCM_16")
# report average RTF
logging.info(
f"Finished generation of {idx} utterances (RTF = {total_rtf / idx:.03f})."
)
def make_wav(args):
import argparse
import logging
import os
import time
import numpy as np
import soundfile as sf
import torch
import yaml
from tqdm import tqdm
import parallel_wavegan.models
from parallel_wavegan.datasets import MelDataset
from parallel_wavegan.datasets import MelSCPDataset
from parallel_wavegan.utils import read_hdf5
"""Run decoding process."""
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
else:
logging.basicConfig(
level=logging.WARN,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
logging.warning("Skip DEBUG/INFO messages")
# check directory existence
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# load config
if args.config is None:
dirname = os.path.dirname(args.checkpoint)
args.config = os.path.join(dirname, "config.yml")
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None and args.dumpdir is not None) or \
(args.feats_scp is None and args.dumpdir is None):
raise ValueError("Please specify either --dumpdir or --feats-scp.")
# get dataset
if args.dumpdir is not None:
if config["format"] == "hdf5":
mel_query = "*.h5"
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
mel_query = "*-feats.npy"
mel_load_fn = np.load
else:
raise ValueError("support only hdf5 or npy format.")
dataset = MelDataset(
args.dumpdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
else:
dataset = MelSCPDataset(
feats_scp=args.feats_scp,
return_utt_id=True,
)
logging.info(f"The number of features to be decoded = {len(dataset)}.")
# setup
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
model_class = getattr(
parallel_wavegan.models,
config.get("generator_type", "ParallelWaveGANGenerator"))
model = model_class(**config["generator_params"])
model.load_state_dict(
torch.load(args.checkpoint, map_location="cpu")["model"]["generator"])
logging.info(f"Loaded model parameters from {args.checkpoint}.")
model.remove_weight_norm()
model = model.eval().to(device)
use_noise_input = not isinstance(model,
parallel_wavegan.models.MelGANGenerator)
pad_fn = torch.nn.ReplicationPad1d(config["generator_params"].get(
"aux_context_window", 0))
# start generation
total_rtf = 0.0
with torch.no_grad(), tqdm(dataset, desc="[decode]") as pbar:
for idx, (utt_id, c) in enumerate(pbar, 1):
x = ()
#c = c.T
if use_noise_input:
z = torch.randn(1, 1,
np.shape(c)[2] * config["hop_size"]).to(device)
x += (z, )
print(c.shape)
c = torch.from_numpy(c)
c = c.type(torch.cuda.FloatTensor).to(device)
c = pad_fn(c)
x += (c, )
# setup input
#---------------------------------------------------------------------
'''
x = ()
print(c.shape)
if use_noise_input:
print('len(c).shape: ', len(c))
z = torch.randn(1, 1, np.shape(c)[2] * config["hop_size"]).to(device)
x += (z,)
c = c.type(torch.cuda.FloatTensor).to(device)
c = pad_fn(c)
x += (c,)
#c = pad_fn(torch.from_numpy(c).unsqueeze(0).transpose(2, 1)).to(device)
'''
#---------------------------------------------------------
'''
import pickle
x_ = ()
with open('test2.pickle', 'rb') as f:
c_ = pickle.load(f)
print(c_.shape)
if use_noise_input:
#print('c_.shape : ', np.shape(c_)[2])
z = torch.randn(1, 1, np.shape(c_)[2] * config["hop_size"]).to(device)
x_ += (z,)
c_ = c_.type(torch.cuda.FloatTensor).to(device)
c_ = pad_fn(c_)
x_ += (c_,)
'''
#---------------------------------------------------------
# generate
start = time.time()
y = model(*x).view(-1).cpu().numpy()
rtf = (time.time() - start) / (len(y) / config["sampling_rate"])
pbar.set_postfix({"RTF": rtf})
total_rtf += rtf
# save as PCM 16 bit wav file
sf.write(os.path.join(config["outdir"], f"{utt_id}_gen.wav"), y,
config["sampling_rate"], "PCM_16")
# report average RTF
logging.info(
f"Finished generation of {idx} utterances (RTF = {total_rtf / idx:.03f})."
)
def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Normalize dumped raw features (See detail in parallel_wavegan/bin/normalize.py)."
)
parser.add_argument(
"--rootdir",
default=None,
type=str,
help="directory including feature files to be normalized. "
"you need to specify either *-scp or rootdir.",
)
parser.add_argument(
"--wav-scp",
default=None,
type=str,
help="kaldi-style wav.scp file. "
"you need to specify either *-scp or rootdir.",
)
parser.add_argument(
"--feats-scp",
default=None,
type=str,
help="kaldi-style feats.scp file. "
"you need to specify either *-scp or rootdir.",
)
parser.add_argument(
"--segments",
default=None,
type=str,
help="kaldi-style segments file.",
)
parser.add_argument(
"--dumpdir",
type=str,
required=True,
help="directory to dump normalized feature files.",
)
parser.add_argument(
"--stats",
type=str,
required=True,
help="statistics file.",
)
parser.add_argument(
"--skip-wav-copy",
default=False,
action="store_true",
help="whether to skip the copy of wav files.",
)
parser.add_argument(
"--config", type=str, required=True, help="yaml format configuration file."
)
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)",
)
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
logging.basicConfig(
level=logging.WARN,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
logging.warning("Skip DEBUG/INFO messages")
# load config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None and args.rootdir is not None) or (
args.feats_scp is None and args.rootdir is None
):
raise ValueError("Please specify either --rootdir or --feats-scp.")
# check directory existence
if not os.path.exists(args.dumpdir):
os.makedirs(args.dumpdir)
# get dataset
if args.rootdir is not None:
if config["format"] == "hdf5":
audio_query, mel_query = "*.h5", "*.h5"
audio_load_fn = lambda x: read_hdf5(x, "wave") # NOQA
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
audio_query, mel_query = "*-wave.npy", "*-feats.npy"
audio_load_fn = np.load
mel_load_fn = np.load
else:
raise ValueError("support only hdf5 or npy format.")
if not args.skip_wav_copy:
dataset = AudioMelDataset(
root_dir=args.rootdir,
audio_query=audio_query,
mel_query=mel_query,
audio_load_fn=audio_load_fn,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
else:
dataset = MelDataset(
root_dir=args.rootdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn,
return_utt_id=True,
)
else:
if not args.skip_wav_copy:
dataset = AudioMelSCPDataset(
wav_scp=args.wav_scp,
feats_scp=args.feats_scp,
segments=args.segments,
return_utt_id=True,
)
else:
dataset = MelSCPDataset(
feats_scp=args.feats_scp,
return_utt_id=True,
)
logging.info(f"The number of files = {len(dataset)}.")
# restore scaler
scaler = StandardScaler()
if config["format"] == "hdf5":
scaler.mean_ = read_hdf5(args.stats, "mean")
scaler.scale_ = read_hdf5(args.stats, "scale")
elif config["format"] == "npy":
scaler.mean_ = np.load(args.stats)[0]
scaler.scale_ = np.load(args.stats)[1]
else:
raise ValueError("support only hdf5 or npy format.")
# from version 0.23.0, this information is needed
scaler.n_features_in_ = scaler.mean_.shape[0]
# process each file
for items in tqdm(dataset):
if not args.skip_wav_copy:
utt_id, audio, mel = items
else:
utt_id, mel = items
# normalize
mel = scaler.transform(mel)
# save
if config["format"] == "hdf5":
write_hdf5(
os.path.join(args.dumpdir, f"{utt_id}.h5"),
"feats",
mel.astype(np.float32),
)
if not args.skip_wav_copy:
write_hdf5(
os.path.join(args.dumpdir, f"{utt_id}.h5"),
"wave",
audio.astype(np.float32),
)
elif config["format"] == "npy":
np.save(
os.path.join(args.dumpdir, f"{utt_id}-feats.npy"),
mel.astype(np.float32),
allow_pickle=False,
)
if not args.skip_wav_copy:
np.save(
os.path.join(args.dumpdir, f"{utt_id}-wave.npy"),
audio.astype(np.float32),
allow_pickle=False,
)
else:
raise ValueError("support only hdf5 or npy format.")
def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Compute mean and variance of dumped raw features "
"(See detail in parallel_wavegan/bin/compute_statistics.py).")
parser.add_argument(
"--feats-scp",
"--scp",
default=None,
type=str,
help="kaldi-style feats.scp file. "
"you need to specify either feats-scp or rootdir.",
)
parser.add_argument(
"--rootdir",
type=str,
help="directory including feature files. "
"you need to specify either feats-scp or rootdir.",
)
parser.add_argument(
"--config",
type=str,
required=True,
help="yaml format configuration file.",
)
parser.add_argument(
"--dumpdir",
default=None,
type=str,
required=True,
help="directory to save statistics. if not provided, "
"stats will be saved in the above root directory. (default=None)",
)
parser.add_argument(
"--verbose",
type=int,
default=1,
help="logging level. higher is more logging. (default=1)",
)
args = parser.parse_args()
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
logging.basicConfig(
level=logging.WARN,
format=
"%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
logging.warning("Skip DEBUG/INFO messages")
# load config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None
and args.rootdir is not None) or (args.feats_scp is None
and args.rootdir is None):
raise ValueError("Please specify either --rootdir or --feats-scp.")
# check directory existence
if not os.path.exists(args.dumpdir):
os.makedirs(args.dumpdir)
# get dataset
if args.feats_scp is None:
if config["format"] == "hdf5":
mel_query = "*.h5"
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
mel_query = "*-feats.npy"
mel_load_fn = np.load
else:
raise ValueError("support only hdf5 or npy format.")
dataset = MelDataset(args.rootdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn)
else:
dataset = MelSCPDataset(args.feats_scp)
logging.info(f"The number of files = {len(dataset)}.")
# calculate statistics
scaler = StandardScaler()
for mel in tqdm(dataset):
scaler.partial_fit(mel)
if config["format"] == "hdf5":
write_hdf5(
os.path.join(args.dumpdir, "stats.h5"),
"mean",
scaler.mean_.astype(np.float32),
)
write_hdf5(
os.path.join(args.dumpdir, "stats.h5"),
"scale",
scaler.scale_.astype(np.float32),
)
else:
stats = np.stack([scaler.mean_, scaler.scale_], axis=0)
np.save(
os.path.join(args.dumpdir, "stats.npy"),
stats.astype(np.float32),
allow_pickle=False,
)
def main():
"""Run preprocessing process."""
parser = argparse.ArgumentParser(
description="Compute mean and variance of dumped raw features "
"(See detail in parallel_wavegan/bin/compute_statistics.py).")
parser.add_argument("--feats-scp", "--scp", default=None, type=str,
help="kaldi-style feats.scp file. "
"you need to specify either feats-scp or rootdir.")
parser.add_argument("--rootdir", type=str, required=True,
help="directory including feature files. "
"you need to specify either feats-scp or rootdir.")
parser.add_argument("--config", type=str, required=True,
help="yaml format configuration file.")
parser.add_argument("--dumpdir", default=None, type=str,
help="directory to save statistics. if not provided, "
"stats will be saved in the above root directory. (default=None)")
parser.add_argument("--ftype", default='mel', type=str,
help="feature type")
parser.add_argument("--verbose", type=int, default=1,
help="logging level. higher is more logging. (default=1)")
# runtime mode
args = parser.parse_args()
# interactive mode
# args = argparse.ArgumentParser()
# args.feats_scp = None
# args.config = 'egs/so_emo_female/voc1/conf/multi_band_melgan.v2.yaml'
# args.verbose = 1
# args.ftype = 'spec'
# args.rootdir = '/data/evs/VCTK/VCTK-wgan/spec'
# args.rootdir = '/data/evs/Arctic/spec'
# args.dumpdir = os.path.join(args.rootdir, "")
# set logger
if args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
elif args.verbose > 0:
logging.basicConfig(
level=logging.INFO, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
else:
logging.basicConfig(
level=logging.WARN, format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
logging.warning('Skip DEBUG/INFO messages')
# load config
with open(args.config) as f:
config = yaml.load(f, Loader=yaml.Loader)
config.update(vars(args))
# check arguments
if (args.feats_scp is not None and args.rootdir is not None) or \
(args.feats_scp is None and args.rootdir is None):
raise ValueError("Please specify either --rootdir or --feats-scp.")
# check directory existence
if args.dumpdir is None:
args.dumpdir = os.path.dirname(args.rootdir)
if not os.path.exists(args.dumpdir):
os.makedirs(args.dumpdir)
# get dataset
if args.feats_scp is None:
if config["format"] == "hdf5":
mel_query = "*.h5"
mel_load_fn = lambda x: read_hdf5(x, "feats") # NOQA
elif config["format"] == "npy":
mel_query = "*.mel.npy"
mel_load_fn = np.load
spc_query = "*.spec.npy"
spc_load_fn = np.load
else:
raise ValueError("support only hdf5 or npy format.")
dataset1 = MelDataset(
args.rootdir,
mel_query=mel_query,
mel_load_fn=mel_load_fn)
dataset2 = SpcDataset(
args.rootdir,
spc_query=spc_query,
spc_load_fn=spc_load_fn)
else:
dataset = MelSCPDataset(args.feats_scp)
logging.info(f"The number of files in mel dataset = {len(dataset1)}.")
logging.info(f"The number of files in spc dataset = {len(dataset2)}.")
# calculate statistics
scaler = StandardScaler()
if args.ftype == 'mel':
for mel in tqdm(dataset1):
scaler.partial_fit(mel)
elif args.ftype == 'spec':
for spc in tqdm(dataset2):
scaler.partial_fit(spc)
if config["format"] == "hdf5":
write_hdf5(os.path.join(args.dumpdir, "{}_mean_std.h5".format(args.ftype)),
"mean", scaler.mean_.astype(np.float32))
write_hdf5(os.path.join(args.dumpdir, "{}_mean_std.h5".format(args.ftype)),
"scale", scaler.scale_.astype(np.float32))
else:
stats = np.stack([scaler.mean_, scaler.scale_], axis=0)
np.save(os.path.join(args.dumpdir, "{}_mean_std.npy".format(args.ftype)),
stats.astype(np.float32), allow_pickle=False)