def main(cmd=None):
"""Run SVS model decoding."""
print(get_commandline_args(), file=sys.stderr)
parser = get_parser()
args = parser.parse_args(cmd)
kwargs = vars(args)
kwargs.pop("config", None)
inference(**kwargs)
def main():
parser = get_parser()
args = parser.parse_args()
# logging info
logfmt = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
if args.verbose > 0:
logging.basicConfig(level=logging.INFO, format=logfmt)
else:
logging.basicConfig(level=logging.WARN, format=logfmt)
logging.info(get_commandline_args())
if args.preprocess_conf is not None:
preprocessing = Transformation(args.preprocess_conf)
logging.info("Apply preprocessing: {}".format(preprocessing))
else:
preprocessing = None
# There are no necessary for matrix without preprocessing,
# so change to file_reader_helper to return shape.
# This make sense only with filetype="hdf5".
for utt, mat in file_reader_helper(args.rspecifier,
args.filetype,
return_shape=preprocessing is None):
if preprocessing is not None:
if is_scipy_wav_style(mat):
# If data is sound file, then got as Tuple[int, ndarray]
rate, mat = mat
mat = preprocessing(mat, uttid_list=utt)
shape_str = ",".join(map(str, mat.shape))
else:
if len(mat) == 2 and isinstance(mat[1], tuple):
# If data is sound file, Tuple[int, Tuple[int, ...]]
rate, mat = mat
shape_str = ",".join(map(str, mat))
args.out.write("{} {}\n".format(utt, shape_str))
def main(cmd=None):
print(get_commandline_args(), file=sys.stderr)
parser = get_parser()
args = parser.parse_args(cmd)
kwargs = vars(args)
aggregate_stats_dirs(**kwargs)
def main():
logfmt = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
logging.basicConfig(level=logging.INFO, format=logfmt)
logging.info(get_commandline_args())
parser = argparse.ArgumentParser(
description='Create waves list from "midi.scp"',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("scp")
parser.add_argument("outdir")
parser.add_argument(
"--name",
default="midi",
help="Specify the prefix word of output file name "
'such as "wav.scp"',
)
parser.add_argument("--segments", default=None)
parser.add_argument(
"--fs",
type=np.int16,
default=None,
help="If the sampling rate specified, "
"Change the sampling rate.",
)
group = parser.add_mutually_exclusive_group()
# TODO: in midi, the reference channels should be related to track, it is not implemented now
group.add_argument("--ref-channels", default=None, type=str2int_tuple)
group.add_argument("--utt2ref-channels", default=None, type=str)
args = parser.parse_args()
if args.ref_channels is not None:
def utt2ref_channels(x) -> Tuple[int, ...]:
return args.ref_channels
elif args.utt2ref_channels is not None:
utt2ref_channels_dict = read_2column_text(args.utt2ref_channels)
def utt2ref_channels(x, d=utt2ref_channels_dict) -> Tuple[int, ...]:
chs_str = d[x]
return tuple(map(int, chs_str.split()))
else:
utt2ref_channels = None
# load segments
if args.segments is not None:
segments = {}
with open(args.segments) as f:
for line in f:
if len(line) == 0:
continue
utt_id, recording_id, segment_begin, segment_end = line.strip(
).split(" ")
segments[utt_id] = (
recording_id,
float(segment_begin),
float(segment_end),
)
Path(args.outdir).mkdir(parents=True, exist_ok=True)
out_midiscp = Path(args.outdir) / f"{args.name}.scp"
if args.segments is not None:
loader = MIDIScpReader(args.scp, rate=args.fs)
writer = MIDIScpWriter(
args.outdir,
out_midiscp,
format="midi",
rate=args.fs,
)
cache = (None, None, None)
for utt_id, (recording, start, end) in tqdm(segments.items()):
# TODO: specify track information here
if recording == cache[0]:
note_seq, tempo_seq = cache[1], cache[2]
else:
note_seq, tempo_seq = loader[recording]
cache = (recording, note_seq, tempo_seq)
if args.fs is not None:
start = int(start * args.fs)
end = int(end * args.fs)
if start < 0:
start = 0
if end > len(note_seq):
end = len(note_seq)
else:
start = np.searchsorted([item[0] for item in note_seq], start,
"left")
end = np.searchsorted([item[1] for item in note_seq], end,
"left")
sub_note = note_seq[start:end]
sub_tempo = tempo_seq[start:end]
writer[utt_id] = sub_note, sub_tempo
else:
# midi_scp does not need to change, when no segments is applied
# Note things will change, after finish other todos in the script
os.system("cp {} {}".format(args.scp,
Path(args.outdir / f"{args.name}.scp")))
def main():
logfmt = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
logging.basicConfig(level=logging.INFO, format=logfmt)
logging.info(get_commandline_args())
parser = argparse.ArgumentParser(
description='Create waves list from "wav.scp"',
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("scp")
parser.add_argument("outdir")
parser.add_argument(
"--name",
default="wav",
help="Specify the prefix word of output file name "
'such as "wav.scp"',
)
parser.add_argument("--segments", default=None)
parser.add_argument(
"--fs",
type=humanfriendly_or_none,
default=None,
help="If the sampling rate specified, "
"Change the sampling rate.",
)
parser.add_argument("--audio-format", default="wav")
group = parser.add_mutually_exclusive_group()
group.add_argument("--ref-channels", default=None, type=str2int_tuple)
group.add_argument("--utt2ref-channels", default=None, type=str)
args = parser.parse_args()
out_num_samples = Path(args.outdir) / f"utt2num_samples"
if args.ref_channels is not None:
def utt2ref_channels(x) -> Tuple[int, ...]:
return args.ref_channels
elif args.utt2ref_channels is not None:
utt2ref_channels_dict = read_2column_text(args.utt2ref_channels)
def utt2ref_channels(x, d=utt2ref_channels_dict) -> Tuple[int, ...]:
chs_str = d[x]
return tuple(map(int, chs_str.split()))
else:
utt2ref_channels = None
Path(args.outdir).mkdir(parents=True, exist_ok=True)
out_wavscp = Path(args.outdir) / f"{args.name}.scp"
if args.segments is not None:
# Note: kaldiio supports only wav-pcm-int16le file.
loader = kaldiio.load_scp_sequential(args.scp, segments=args.segments)
if args.audio_format.endswith("ark"):
fark = open(Path(args.outdir) / f"data_{args.name}.ark", "wb")
fscp = out_wavscp.open("w")
else:
writer = SoundScpWriter(
args.outdir,
out_wavscp,
format=args.audio_format,
)
with out_num_samples.open("w") as fnum_samples:
for uttid, (rate, wave) in tqdm(loader):
# wave: (Time,) or (Time, Nmic)
if wave.ndim == 2 and utt2ref_channels is not None:
wave = wave[:, utt2ref_channels(uttid)]
if args.fs is not None and args.fs != rate:
# FIXME(kamo): To use sox?
wave = resampy.resample(wave.astype(np.float64),
rate,
args.fs,
axis=0)
wave = wave.astype(np.int16)
rate = args.fs
if args.audio_format.endswith("ark"):
if "flac" in args.audio_format:
suf = "flac"
elif "wav" in args.audio_format:
suf = "wav"
else:
raise RuntimeError("wav.ark or flac")
# NOTE(kamo): Using extended ark format style here.
# This format is incompatible with Kaldi
kaldiio.save_ark(
fark,
{uttid: (wave, rate)},
scp=fscp,
append=True,
write_function=f"soundfile_{suf}",
)
else:
writer[uttid] = rate, wave
fnum_samples.write(f"{uttid} {len(wave)}\n")
else:
if args.audio_format.endswith("ark"):
fark = open(Path(args.outdir) / f"data_{args.name}.ark", "wb")
else:
wavdir = Path(args.outdir) / f"data_{args.name}"
wavdir.mkdir(parents=True, exist_ok=True)
with Path(args.scp).open("r") as fscp, out_wavscp.open(
"w") as fout, out_num_samples.open("w") as fnum_samples:
for line in tqdm(fscp):
uttid, wavpath = line.strip().split(None, 1)
if wavpath.endswith("|"):
# Streaming input e.g. cat a.wav |
with kaldiio.open_like_kaldi(wavpath, "rb") as f:
with BytesIO(f.read()) as g:
wave, rate = soundfile.read(g, dtype=np.int16)
if wave.ndim == 2 and utt2ref_channels is not None:
wave = wave[:, utt2ref_channels(uttid)]
if args.fs is not None and args.fs != rate:
# FIXME(kamo): To use sox?
wave = resampy.resample(wave.astype(np.float64),
rate,
args.fs,
axis=0)
wave = wave.astype(np.int16)
rate = args.fs
if args.audio_format.endswith("ark"):
if "flac" in args.audio_format:
suf = "flac"
elif "wav" in args.audio_format:
suf = "wav"
else:
raise RuntimeError("wav.ark or flac")
# NOTE(kamo): Using extended ark format style here.
# This format is incompatible with Kaldi
kaldiio.save_ark(
fark,
{uttid: (wave, rate)},
scp=fout,
append=True,
write_function=f"soundfile_{suf}",
)
else:
owavpath = str(wavdir /
f"{uttid}.{args.audio_format}")
soundfile.write(owavpath, wave, rate)
fout.write(f"{uttid} {owavpath}\n")
else:
wave, rate = soundfile.read(wavpath, dtype=np.int16)
if wave.ndim == 2 and utt2ref_channels is not None:
wave = wave[:, utt2ref_channels(uttid)]
save_asis = False
elif args.audio_format.endswith("ark"):
save_asis = False
elif Path(wavpath).suffix == "." + args.audio_format and (
args.fs is None or args.fs == rate):
save_asis = True
else:
save_asis = False
if save_asis:
# Neither --segments nor --fs are specified and
# the line doesn't end with "|",
# i.e. not using unix-pipe,
# only in this case,
# just using the original file as is.
fout.write(f"{uttid} {wavpath}\n")
else:
if args.fs is not None and args.fs != rate:
# FIXME(kamo): To use sox?
wave = resampy.resample(wave.astype(np.float64),
rate,
args.fs,
axis=0)
wave = wave.astype(np.int16)
rate = args.fs
if args.audio_format.endswith("ark"):
if "flac" in args.audio_format:
suf = "flac"
elif "wav" in args.audio_format:
suf = "wav"
else:
raise RuntimeError("wav.ark or flac")
# NOTE(kamo): Using extended ark format style here.
# This format is not supported in Kaldi.
kaldiio.save_ark(
fark,
{uttid: (wave, rate)},
scp=fout,
append=True,
write_function=f"soundfile_{suf}",
)
else:
owavpath = str(wavdir /
f"{uttid}.{args.audio_format}")
soundfile.write(owavpath, wave, rate)
fout.write(f"{uttid} {owavpath}\n")
fnum_samples.write(f"{uttid} {len(wave)}\n")
def main(cmd=None):
print(get_commandline_args(), file=sys.stderr)
parser = get_parser()
args = parser.parse_args(cmd)
kwargs = vars(args)
split_scps(**kwargs)
def main(cmd=None):
logfmt = "%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s"
logging.basicConfig(level=logging.INFO, format=logfmt)
logging.info(get_commandline_args())
parser = get_parser()
args = parser.parse_args(cmd)
args.cmd = shlex.split(args.cmd)
if args.host is None and shutil.which(args.cmd[0]) is None:
raise RuntimeError(
f"The first args of --cmd should be a script path. e.g. utils/run.pl: "
f"{args.cmd[0]}"
)
# Specify init_method:
# See: https://pytorch.org/docs/stable/distributed.html#initialization
if args.host is None and args.num_nodes <= 1:
# Automatically set init_method if num_node=1
init_method = None
else:
if args.master_port is None:
# Try "shared-file system initialization" if master_port is not specified
# Give random name to avoid reusing previous file
init_file = args.init_file_prefix + str(uuid.uuid4())
init_file = Path(init_file).absolute()
Path(init_file).parent.mkdir(exist_ok=True, parents=True)
init_method = ["--dist_init_method", f"file://{init_file}"]
else:
init_method = ["--dist_master_port", str(args.master_port)]
# This can be omitted if slurm mode
if args.master_addr is not None:
init_method += ["--dist_master_addr", args.master_addr]
elif args.host is not None:
init_method += [
"--dist_master_addr",
args.host.split(",")[0].split(":")[0],
]
# Log-rotation
for i in range(args.max_num_log_files - 1, -1, -1):
if i == 0:
p = Path(args.log)
pn = p.parent / (p.stem + ".1" + p.suffix)
else:
_p = Path(args.log)
p = _p.parent / (_p.stem + f".{i}" + _p.suffix)
pn = _p.parent / (_p.stem + f".{i + 1}" + _p.suffix)
if p.exists():
if i == args.max_num_log_files - 1:
p.unlink()
else:
shutil.move(p, pn)
processes = []
# Submit command via SSH
if args.host is not None:
hosts = []
ids_list = []
# e.g. args.host = "host1:0:2,host2:0:1"
for host in args.host.split(","):
# e.g host = "host1:0:2"
sps = host.split(":")
host = sps[0]
if len(sps) > 1:
ids = [int(x) for x in sps[1:]]
else:
ids = list(range(args.ngpu))
hosts.append(host)
ids_list.append(ids)
world_size = sum(max(len(x), 1) for x in ids_list)
logging.info(f"{len(hosts)}nodes with world_size={world_size} via SSH")
if args.envfile is not None:
env = f"source {args.envfile}"
else:
env = ""
if args.log != "-":
Path(args.log).parent.mkdir(parents=True, exist_ok=True)
f = Path(args.log).open("w", encoding="utf-8")
else:
# Output to stdout/stderr
f = None
rank = 0
for host, ids in zip(hosts, ids_list):
ngpu = 1 if len(ids) > 0 else 0
ids = ids if len(ids) > 0 else ["none"]
for local_rank in ids:
cmd = (
args.args
+ [
"--ngpu",
str(ngpu),
"--multiprocessing_distributed",
"false",
"--local_rank",
str(local_rank),
"--dist_rank",
str(rank),
"--dist_world_size",
str(world_size),
]
+ init_method
)
if ngpu == 0:
# Gloo supports both GPU and CPU mode.
# See: https://pytorch.org/docs/stable/distributed.html
cmd += ["--dist_backend", "gloo"]
heredoc = f"""<< EOF
set -euo pipefail
cd {os.getcwd()}
{env}
{" ".join([c if len(c) != 0 else "''" for c in cmd])}
EOF
"""
# FIXME(kamo): The process will be alive
# even if this program is stopped because we don't set -t here,
# i.e. not assigning pty,
# and the program is not killed when SSH connection is closed.
process = subprocess.Popen(
["ssh", host, "bash", heredoc],
stdout=f,
stderr=f,
)
processes.append(process)
rank += 1
# If Single node
elif args.num_nodes <= 1:
if args.ngpu > 1:
if args.multiprocessing_distributed:
# NOTE:
# If multiprocessing_distributed=true,
# -> Distributed mode, which is multi-process and Multi-GPUs.
# and TCP initializetion is used if single-node case:
# e.g. init_method="tcp://localhost:20000"
logging.info(f"single-node with {args.ngpu}gpu on distributed mode")
else:
# NOTE:
# If multiprocessing_distributed=false
# -> "DataParallel" mode, which is single-process
# and Multi-GPUs with threading.
# See:
# https://discuss.pytorch.org/t/why-torch-nn-parallel-distributeddataparallel-runs-faster-than-torch-nn-dataparallel-on-single-machine-with-multi-gpu/32977/2
logging.info(f"single-node with {args.ngpu}gpu using DataParallel")
# Using cmd as it is simply
cmd = (
args.cmd
# arguments for ${cmd}
+ ["--gpu", str(args.ngpu), args.log]
# arguments for *_train.py
+ args.args
+ [
"--ngpu",
str(args.ngpu),
"--multiprocessing_distributed",
str(args.multiprocessing_distributed),
]
)
process = subprocess.Popen(cmd)
processes.append(process)
elif Path(args.cmd[0]).name == "run.pl":
raise RuntimeError("run.pl doesn't support submitting to the other nodes.")
elif Path(args.cmd[0]).name == "ssh.pl":
raise RuntimeError("Use --host option instead of ssh.pl")
# If Slurm
elif Path(args.cmd[0]).name == "slurm.pl":
logging.info(f"{args.num_nodes}nodes and {args.ngpu}gpu-per-node using srun")
cmd = (
args.cmd
# arguments for ${cmd}
+ [
"--gpu",
str(args.ngpu),
"--num_threads",
str(max(args.ngpu, 1)),
"--num_nodes",
str(args.num_nodes),
args.log,
"srun",
# Inherit all environment variable from parent process
"--export=ALL",
]
# arguments for *_train.py
+ args.args
+ [
"--ngpu",
str(args.ngpu),
"--multiprocessing_distributed",
"true",
"--dist_launcher",
"slurm",
]
+ init_method
)
if args.ngpu == 0:
# Gloo supports both GPU and CPU mode.
# See: https://pytorch.org/docs/stable/distributed.html
cmd += ["--dist_backend", "gloo"]
process = subprocess.Popen(cmd)
processes.append(process)
else:
# This pattern can also works with Slurm.
logging.info(f"{args.num_nodes}nodes and {args.ngpu}gpu-per-node using mpirun")
cmd = (
args.cmd
# arguments for ${cmd}
+ [
"--gpu",
str(args.ngpu),
"--num_threads",
str(max(args.ngpu, 1)),
# Make sure scheduler setting, i.e. conf/queue.conf
# so that --num_nodes requires 1process-per-node
"--num_nodes",
str(args.num_nodes),
args.log,
"mpirun",
# -np option can be omitted with Torque/PBS
"-np",
str(args.num_nodes),
]
# arguments for *_train.py
+ args.args
+ [
"--ngpu",
str(args.ngpu),
"--multiprocessing_distributed",
"true",
"--dist_launcher",
"mpi",
]
+ init_method
)
if args.ngpu == 0:
# Gloo supports both GPU and CPU mode.
# See: https://pytorch.org/docs/stable/distributed.html
cmd += ["--dist_backend", "gloo"]
process = subprocess.Popen(cmd)
processes.append(process)
logging.info(f"log file: {args.log}")
logging.info(f"submitting cmd {cmd}")
failed = False
while any(p.returncode is None for p in processes):
for process in processes:
# If any process is failed, try to kill the other processes too
if failed and process.returncode is not None:
process.kill()
else:
try:
process.wait(0.5)
except subprocess.TimeoutExpired:
pass
if process.returncode is not None and process.returncode != 0:
failed = True
for process in processes:
if process.returncode != 0:
print(
subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd),
file=sys.stderr,
)
p = Path(args.log)
if p.exists():
with p.open() as f:
lines = list(f)
raise RuntimeError(
f"\n################### The last 1000 lines of {args.log} "
f"###################\n" + "".join(lines[-1000:])
)
else:
raise RuntimeError