def upload(feature_manifest: Pathlike, url: str, output_manifest: Pathlike,
num_jobs: int):
"""
Read an existing FEATURE_MANIFEST, upload the feature matrices it contains to a URL location,
and save a new feature OUTPUT_MANIFEST that refers to the uploaded features.
The URL can refer to endpoints such as AWS S3, GCP, Azure, etc.
For example: "s3://my-bucket/my-features" is a valid URL.
This script does not currently support credentials,
and assumes that you have the write permissions.
"""
from concurrent.futures import ProcessPoolExecutor
from tqdm import tqdm
output_manifest = Path(output_manifest)
assert (".jsonl" in output_manifest.suffixes
), "This mode only supports writing to JSONL feature manifests."
local_features: FeatureSet = FeatureSet.from_file(feature_manifest)
with FeatureSet.open_writer(
output_manifest) as manifest_writer, ProcessPoolExecutor(
num_jobs) as ex:
futures = []
for item in tqdm(local_features,
desc="Submitting parallel uploading tasks..."):
futures.append(ex.submit(_upload_one, item, url))
for item in tqdm(futures, desc=f"Uploading features to {url}"):
manifest_writer.write(item.result())
def test_feature_set_serialization(feature_set, format, compressed):
with NamedTemporaryFile(suffix='.gz' if compressed else '') as f:
if format == 'jsonl':
feature_set.to_jsonl(f.name)
feature_set_deserialized = FeatureSet.from_jsonl(f.name)
if format == 'json':
feature_set.to_json(f.name)
feature_set_deserialized = FeatureSet.from_json(f.name)
if format == 'yaml':
feature_set.to_yaml(f.name)
feature_set_deserialized = FeatureSet.from_yaml(f.name)
assert feature_set_deserialized == feature_set
def test_feature_set_serialization(feature_set, format, compressed):
with NamedTemporaryFile(suffix=".gz" if compressed else "") as f:
if format == "jsonl":
feature_set.to_jsonl(f.name)
feature_set_deserialized = FeatureSet.from_jsonl(f.name)
if format == "json":
feature_set.to_json(f.name)
feature_set_deserialized = FeatureSet.from_json(f.name)
if format == "yaml":
feature_set.to_yaml(f.name)
feature_set_deserialized = FeatureSet.from_yaml(f.name)
assert feature_set_deserialized == feature_set
def test_feature_set_copy_feats(cuts):
feature_set = FeatureSet.from_features([cuts[0].features])
with TemporaryDirectory() as d, NumpyFilesWriter(d) as w:
cpy = feature_set.copy_feats(writer=w)
data = cpy[0].load()
assert isinstance(data, np.ndarray)
ref_data = feature_set[0].load()
np.testing.assert_almost_equal(data, ref_data)
def feature_set():
return FeatureSet(features=[
Features(recording_id='irrelevant',
channels=0,
start=0.0,
duration=20.0,
type='fbank',
num_frames=2000,
num_features=20,
frame_shift=0.01,
sampling_rate=16000,
storage_type='lilcom',
storage_path='/irrelevant/',
storage_key='path.llc')
])
def feature_set():
return FeatureSet(features=[
Features(
recording_id="irrelevant",
channels=0,
start=0.0,
duration=20.0,
type="fbank",
num_frames=2000,
num_features=20,
frame_shift=0.01,
sampling_rate=16000,
storage_type="lilcom",
storage_path="/irrelevant/",
storage_key="path.llc",
)
])
def load_kaldi_data_dir(
path: Pathlike,
sampling_rate: int,
frame_shift: Optional[Seconds] = None,
) -> Tuple[RecordingSet, Optional[SupervisionSet], Optional[FeatureSet]]:
"""
Load a Kaldi data directory and convert it to a Lhotse RecordingSet and SupervisionSet manifests.
For this to work, at least the wav.scp file must exist.
SupervisionSet is created only when a segments file exists.
All the other files (text, utt2spk, etc.) are optional, and some of them might not be handled yet.
In particular, feats.scp files are ignored.
"""
path = Path(path)
assert path.is_dir()
# must exist for RecordingSet
recordings = load_kaldi_text_mapping(path / 'wav.scp', must_exist=True)
durations = defaultdict(float)
reco2dur = path / 'reco2dur'
if not reco2dur.is_file():
raise ValueError(
f"No such file: '{reco2dur}' -- fix it by running: utils/data/get_reco2dur.sh <data-dir>"
)
with reco2dur.open() as f:
for line in f:
recording_id, dur = line.strip().split()
durations[recording_id] = float(dur)
recording_set = RecordingSet.from_recordings(
Recording(id=recording_id,
sources=[
AudioSource(type='command' if path_or_cmd.
endswith('|') else 'file',
channels=[0],
source=path_or_cmd[:-1] if path_or_cmd.
endswith('|') else path_or_cmd)
],
sampling_rate=sampling_rate,
num_samples=int(durations[recording_id] * sampling_rate),
duration=durations[recording_id])
for recording_id, path_or_cmd in recordings.items())
supervision_set = None
segments = path / 'segments'
if segments.is_file():
with segments.open() as f:
supervision_segments = [l.strip().split() for l in f]
texts = load_kaldi_text_mapping(path / 'text')
speakers = load_kaldi_text_mapping(path / 'utt2spk')
genders = load_kaldi_text_mapping(path / 'spk2gender')
languages = load_kaldi_text_mapping(path / 'utt2lang')
supervision_set = SupervisionSet.from_segments(
SupervisionSegment(id=segment_id,
recording_id=recording_id,
start=float(start),
duration=float(end) - float(start),
channel=0,
text=texts[segment_id],
language=languages[segment_id],
speaker=speakers[segment_id],
gender=genders[speakers[segment_id]])
for segment_id, recording_id, start, end in supervision_segments)
feature_set = None
feats_scp = path / 'feats.scp'
if feats_scp.exists() and is_module_available('kaldiio'):
if frame_shift is not None:
import kaldiio
from lhotse.features.io import KaldiReader
feature_set = FeatureSet.from_features(
Features(type='kaldiio',
num_frames=mat.shape[0],
num_features=mat.shape[1],
frame_shift=frame_shift,
sampling_rate=sampling_rate,
start=0,
duration=mat.shape[0] * frame_shift,
storage_type=KaldiReader.name,
storage_path=str(feats_scp),
storage_key=utt_id,
recording_id=supervision_set[utt_id].recording_id
if supervision_set is not None else utt_id,
channels=0)
for utt_id, mat in kaldiio.load_scp_sequential(str(feats_scp)))
else:
warnings.warn(f"Failed to import Kaldi 'feats.scp' to Lhotse: "
f"frame_shift must be not None. "
f"Feature import omitted.")
return recording_set, supervision_set, feature_set
def load_kaldi_data_dir(
path: Pathlike,
sampling_rate: int,
frame_shift: Optional[Seconds] = None,
map_string_to_underscores: Optional[str] = None,
num_jobs: int = 1,
) -> Tuple[RecordingSet, Optional[SupervisionSet], Optional[FeatureSet]]:
"""
Load a Kaldi data directory and convert it to a Lhotse RecordingSet and
SupervisionSet manifests. For this to work, at least the wav.scp file must exist.
SupervisionSet is created only when a segments file exists.
All the other files (text, utt2spk, etc.) are optional, and some of them might
not be handled yet. In particular, feats.scp files are ignored.
:param map_string_to_underscores: optional string, when specified, we will replace
all instances of this string in SupervisonSegment IDs to underscores.
This is to help with handling underscores in Kaldi
(see :func:`.export_to_kaldi`). This is also done for speaker IDs.
"""
path = Path(path)
assert path.is_dir()
def fix_id(t: str) -> str:
if map_string_to_underscores is None:
return t
return t.replace(map_string_to_underscores, "_")
# must exist for RecordingSet
recordings = load_kaldi_text_mapping(path / "wav.scp", must_exist=True)
with ProcessPoolExecutor(num_jobs) as ex:
dur_vals = ex.map(get_duration, recordings.values())
durations = dict(zip(recordings.keys(), dur_vals))
recording_set = RecordingSet.from_recordings(
Recording(
id=recording_id,
sources=[
AudioSource(
type="command" if path_or_cmd.endswith("|") else "file",
channels=[0],
source=path_or_cmd[:-1]
if path_or_cmd.endswith("|")
else path_or_cmd,
)
],
sampling_rate=sampling_rate,
num_samples=compute_num_samples(durations[recording_id], sampling_rate),
duration=durations[recording_id],
)
for recording_id, path_or_cmd in recordings.items()
)
supervision_set = None
segments = path / "segments"
if segments.is_file():
with segments.open() as f:
supervision_segments = [sup_string.strip().split() for sup_string in f]
texts = load_kaldi_text_mapping(path / "text")
speakers = load_kaldi_text_mapping(path / "utt2spk")
genders = load_kaldi_text_mapping(path / "spk2gender")
languages = load_kaldi_text_mapping(path / "utt2lang")
supervision_set = SupervisionSet.from_segments(
SupervisionSegment(
id=fix_id(segment_id),
recording_id=recording_id,
start=float(start),
duration=add_durations(
float(end), -float(start), sampling_rate=sampling_rate
),
channel=0,
text=texts[segment_id],
language=languages[segment_id],
speaker=fix_id(speakers[segment_id]),
gender=genders[speakers[segment_id]],
)
for segment_id, recording_id, start, end in supervision_segments
)
feature_set = None
feats_scp = path / "feats.scp"
if feats_scp.exists() and is_module_available("kaldi_native_io"):
if frame_shift is not None:
import kaldi_native_io
from lhotse.features.io import KaldiReader
feature_set = FeatureSet.from_features(
Features(
type="kaldi_native_io",
num_frames=mat_shape.num_rows,
num_features=mat_shape.num_cols,
frame_shift=frame_shift,
sampling_rate=sampling_rate,
start=0,
duration=mat_shape.num_rows * frame_shift,
storage_type=KaldiReader.name,
storage_path=str(feats_scp),
storage_key=utt_id,
recording_id=supervision_set[utt_id].recording_id
if supervision_set is not None
else utt_id,
channels=0,
)
for utt_id, mat_shape in kaldi_native_io.SequentialMatrixShapeReader(
f"scp:{feats_scp}"
)
)
else:
warnings.warn(
"Failed to import Kaldi 'feats.scp' to Lhotse: "
"frame_shift must be not None. "
"Feature import omitted."
)
return recording_set, supervision_set, feature_set