def test_extract_and_store_features_from_cut_set(cut_set, executor,
mix_eagerly):
extractor = Fbank()
with TemporaryDirectory() as tmpdir, LilcomFilesWriter(tmpdir) as storage:
with executor() if executor is not None else no_executor() as ex:
cut_set_with_feats = cut_set.compute_and_store_features(
extractor=extractor,
storage=storage,
mix_eagerly=mix_eagerly,
executor=ex)
# The same number of cuts
assert len(cut_set_with_feats) == 2
for orig_cut, feat_cut in zip(cut_set, cut_set_with_feats):
# The ID is retained
assert orig_cut.id == feat_cut.id
# Features were attached
assert feat_cut.has_features
# Recording is retained unless mixing a MixedCut eagerly
should_have_recording = not (mix_eagerly
and isinstance(orig_cut, MixedCut))
assert feat_cut.has_recording == should_have_recording
cuts = list(cut_set_with_feats)
arr = cuts[0].load_features()
assert arr.shape[0] == 100
assert arr.shape[1] == extractor.feature_dim(cuts[0].sampling_rate)
arr = cuts[1].load_features()
assert arr.shape[0] == 300
assert arr.shape[1] == extractor.feature_dim(cuts[0].sampling_rate)
def test_extract_and_store_features(cut):
extractor = Fbank(FbankConfig(sampling_rate=8000))
with TemporaryDirectory() as tmpdir, LilcomFilesWriter(tmpdir) as storage:
cut_with_feats = cut.compute_and_store_features(extractor=extractor,
storage=storage)
arr = cut_with_feats.load_features()
assert arr.shape[0] == 100
assert arr.shape[1] == extractor.feature_dim(cut.sampling_rate)
def test_extract_and_store_features_from_mixed_cut(cut, mix_eagerly):
mixed_cut = cut.append(cut)
extractor = Fbank(FbankConfig(sampling_rate=8000))
with TemporaryDirectory() as tmpdir, LilcomFilesWriter(tmpdir) as storage:
cut_with_feats = mixed_cut.compute_and_store_features(
extractor=extractor, storage=storage, mix_eagerly=mix_eagerly)
arr = cut_with_feats.load_features()
assert arr.shape[0] == 200
assert arr.shape[1] == extractor.feature_dim(mixed_cut.sampling_rate)
def test_feature_set_builder_with_augmentation():
recordings: RecordingSet = RecordingSet.from_json(
'test/fixtures/audio.json')
augment_fn = WavAugmenter.create_predefined('pitch_reverb_tdrop',
sampling_rate=8000)
extractor = Fbank()
with TemporaryDirectory() as d, LilcomFilesWriter(d) as storage:
builder = FeatureSetBuilder(feature_extractor=extractor,
storage=storage,
augment_fn=augment_fn)
feature_set = builder.process_and_store_recordings(
recordings=recordings)
assert len(feature_set) == 6
feature_infos = list(feature_set)
# Assert the properties shared by all features
for features in feature_infos:
# assert that fbank is the default feature type
assert features.type == 'fbank'
# assert that duration is always a multiple of frame_shift
assert features.num_frames == round(features.duration /
features.frame_shift)
# assert that num_features is preserved
assert features.num_features == builder.feature_extractor.config.num_mel_bins
# assert that the storage type metadata matches
assert features.storage_type == storage.name
# assert that the metadata is consistent with the data shapes
arr = features.load()
assert arr.shape[0] == features.num_frames
assert arr.shape[1] == features.num_features
# Assert the properties for recordings of duration 0.5 seconds
for features in feature_infos[:2]:
assert features.num_frames == 50
assert features.duration == 0.5
# Assert the properties for recordings of duration 1.0 seconds
for features in feature_infos[2:]:
assert features.num_frames == 100
assert features.duration == 1.0
# Post-condition 2: the iterative method yields very close results to
# the "standard" method.
true_means = np.mean(np.concatenate([f.load() for f in feature_set]),
axis=0)
true_stds = np.std(np.concatenate([f.load() for f in feature_set]), axis=0)
np.testing.assert_almost_equal(stats["norm_means"], true_means, decimal=5)
np.testing.assert_almost_equal(stats["norm_stds"], true_stds, decimal=5)
# Post-condition 3: the serialization works correctly
assert (stats["norm_means"] == read_stats["norm_means"]).all()
assert (stats["norm_stds"] == read_stats["norm_stds"]).all()
@pytest.mark.parametrize(
"storage_fn",
[
lambda: LilcomFilesWriter(TemporaryDirectory().name),
lambda: LilcomHdf5Writer(NamedTemporaryFile().name),
lambda: ChunkedLilcomHdf5Writer(NamedTemporaryFile().name),
lambda: LilcomChunkyWriter(NamedTemporaryFile().name),
lambda: NumpyFilesWriter(TemporaryDirectory().name),
lambda: NumpyHdf5Writer(NamedTemporaryFile().name),
pytest.param(
lambda: KaldiWriter(TemporaryDirectory().name),
marks=pytest.mark.skipif(
not is_module_available("kaldiio"),
reason="kaldiio must be installed for scp+ark feature writing",
),
),
],
)
def test_feature_set_builder(storage_fn):