def test_mixed_cut_fill_supervision_raises_on_two_supervisions():
cut = dummy_cut(0,
supervisions=[dummy_supervision(0),
dummy_supervision(1)])
cut = cut.pad(duration=10)
with pytest.raises(AssertionError):
fcut = cut.fill_supervision()
def test_cut_set_decompose():
c = dummy_cut(
0,
start=5.0,
duration=10.0,
supervisions=[
dummy_supervision(0, start=0.0),
dummy_supervision(1, start=6.5)
],
)
assert c.start == 5.0
assert c.end == 15.0
cuts = CutSet.from_cuts([c])
recs, sups, feats = cuts.decompose()
assert isinstance(recs, RecordingSet)
assert len(recs) == 1
assert recs[0].id == "dummy-recording-0000"
assert isinstance(sups, SupervisionSet)
assert len(sups) == 2
assert sups[0].id == "dummy-segment-0000"
assert sups[0].start == 5.0
assert sups[0].end == 6.0
assert sups[1].id == "dummy-segment-0001"
assert sups[1].start == 11.5
assert sups[1].end == 12.5
assert isinstance(feats, FeatureSet)
assert len(feats) == 1
def cutset():
return CutSet.from_cuts([
# MonoCut
dummy_cut(0, supervisions=[dummy_supervision(0)]),
# PaddingCut
PaddingCut('pad', duration=1.0, sampling_rate=16000, feat_value=-100,
num_frames=100, frame_shift=0.01, num_features=80, num_samples=16000),
# MixedCut
dummy_cut(0, supervisions=[dummy_supervision(0)]).mix(
dummy_cut(1, supervisions=[dummy_supervision(1)]),
offset_other_by=0.5,
snr=10
)
])
def test_token_collater(add_bos, add_eos):
test_sentences = [
"Testing the first sentence.",
"Let's add some more punctuation, shall we?",
"How about number 42!",
]
cuts = CutSet.from_cuts(
dummy_cut(idx,
idx,
supervisions=[dummy_supervision(idx, idx, text=sentence)])
for idx, sentence in enumerate(test_sentences))
token_collater = TokenCollater(cuts, add_bos=add_bos, add_eos=add_eos)
tokens_batch, tokens_lens = token_collater(cuts)
assert isinstance(tokens_batch, torch.LongTensor)
assert isinstance(tokens_lens, torch.IntTensor)
extend = int(add_bos) + int(add_eos)
expected_len = len(max(test_sentences, key=len)) + extend
assert tokens_batch.shape == (len(test_sentences), expected_len)
assert torch.all(tokens_lens == torch.IntTensor(
[len(sentence) + extend for sentence in test_sentences]))
reconstructed = token_collater.inverse(tokens_batch, tokens_lens)
assert reconstructed == test_sentences
def dummy_cut_with_supervisions():
return dummy_cut(
unique_id=0,
supervisions=[
dummy_supervision(unique_id=i, duration=i) for i in range(1, 7)
],
)
def test_mixed_cut_fill_supervision_expand():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.pad(duration=7.51)
fcut = cut.fill_supervision()
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 7.51
def test_mono_cut_fill_supervision_shrink():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut.duration = 0.5
fcut = cut.fill_supervision(shrink_ok=True)
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 0.5
def test_cut_set_decompose_output_dir():
c = dummy_cut(
0,
start=5.0,
duration=10.0,
supervisions=[
dummy_supervision(0, start=0.0),
dummy_supervision(1, start=6.5)
],
)
assert c.start == 5.0
assert c.end == 15.0
cuts = CutSet.from_cuts([c])
with TemporaryDirectory() as td:
td = Path(td)
recs, sups, feats = cuts.decompose(output_dir=td)
assert list(recs) == list(load_manifest(td / "recordings.jsonl.gz"))
assert list(sups) == list(load_manifest(td / "supervisions.jsonl.gz"))
assert list(feats) == list(load_manifest(td / "features.jsonl.gz"))
def test_mixed_cut_fill_supervision_shrink():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.mix(dummy_cut(1)).truncate(
duration=0.5) # cuts are 100% overlapping
fcut = cut.fill_supervision(shrink_ok=True)
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 0.5
import pytest
from lhotse import CutSet
from lhotse.cut import PaddingCut
from lhotse.testing.dummies import dummy_cut, dummy_supervision
parametrize_on_cut_types = pytest.mark.parametrize(
'cut', [
# MonoCut
dummy_cut(0, supervisions=[dummy_supervision(0)]),
# PaddingCut
PaddingCut('pad', duration=1.0, sampling_rate=16000, feat_value=-100,
num_frames=100, frame_shift=0.01, num_features=80, num_samples=16000),
# MixedCut
dummy_cut(0, supervisions=[dummy_supervision(0)]).mix(
dummy_cut(1, supervisions=[dummy_supervision(1)]),
offset_other_by=0.5,
snr=10
)
]
)
@parametrize_on_cut_types
def test_drop_features(cut):
assert cut.has_features
cut_drop = cut.drop_features()
assert cut.has_features
assert not cut_drop.has_features
def test_mono_cut_fill_supervision_identity():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
fcut = cut.fill_supervision()
assert cut == fcut
def test_mixed_cut_fill_supervision_identity():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.mix(dummy_cut(1)) # cuts are 100% overlapping
fcut = cut.fill_supervision()
assert cut == fcut
def test_mono_cut_fill_supervision_raises_on_two_supervisions():
cut = dummy_cut(0,
supervisions=[dummy_supervision(0),
dummy_supervision(1)])
with pytest.raises(AssertionError):
fcut = cut.fill_supervision()
def test_mono_cut_fill_supervision_shrink_raises_default():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut.duration = 0.5
with pytest.raises(ValueError):
fcut = cut.fill_supervision()
def test_mixed_cut_fill_supervision_shrink_raises_default():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.mix(dummy_cut(1)).truncate(
duration=0.5) # cuts are 100% overlapping
with pytest.raises(ValueError):
fcut = cut.fill_supervision()
