def test_padding_after_resize(audio):
"Padding is added to the end but not the beginning"
new_duration = (audio.duration + 1) * 1000
cropsig_pad_after = ResizeSignal(new_duration, pad_mode=AudioPadType.Zeros_After)
# generate a random input signal that is 3s long
inp, out = apply_transform(cropsig_pad_after, audio)
# test end of signal is padded with zeros
_test_eq(out[:, -10:], torch.zeros_like(out)[:, -10:])
# test front of signal is not padded with zeros
_test_ne(out[:, 0:10], out[:, -10:])
def test_resize_signal_repeat(audio):
"""
Test pad_mode repeat by making sure that columns are
equal at the appropriate offsets
"""
dur = audio.duration * 1000
repeat = 3
cropsig_repeat = ResizeSignal(dur * repeat, pad_mode=AudioPadType.Repeat)
inp, out = apply_transform(cropsig_repeat, audio)
for i in range(repeat):
s = int(i * inp.nsamples)
e = int(s + inp.nsamples)
_test_eq(out[:, s:e], inp)
def test_crop_time_with_pipeline(ex_files):
"""
AudioToSpec->CropTime and ResizeSignal->AudioToSpec
will result in same size images
"""
oa = OpenAudio(ex_files)
crop_dur = random.randint(1000, 5000)
DBMelSpec = SpectrogramTransformer(mel=True, to_db=True)
pipe_cropsig = Pipeline(
[oa, DBMelSpec(hop_length=128),
CropTime(crop_dur)])
pipe_cropspec = Pipeline([
oa,
ResizeSignal(crop_dur),
DBMelSpec(hop_length=128),
])
for i in range(4):
_test_eq(pipe_cropsig(i).width, pipe_cropspec(i).width)
def test_cropping():
"Can use the ResizeSignal Transform"
audio = test_audio_tensor(seconds=10, sr=1000)
inp, out1000 = apply_transform(ResizeSignal(1000), audio.clone())
inp, out2000 = apply_transform(ResizeSignal(2000), audio.clone())
inp, out5000 = apply_transform(ResizeSignal(5000), audio.clone())
_test_eq(out1000.duration, 1)
_test_eq(out2000.duration, 2)
_test_eq(out5000.duration, 5)
_test_eq(out1000.nsamples, out1000.duration * inp.sr)
_test_eq(out2000.nsamples, out2000.duration * inp.sr)
_test_eq(out5000.nsamples, out5000.duration * inp.sr)
# Multi Channel Cropping
inp, mc1000 = apply_transform(ResizeSignal(1000), audio.clone())
inp, mc2000 = apply_transform(ResizeSignal(2000), audio.clone())
inp, mc5000 = apply_transform(ResizeSignal(5000), audio.clone())
_test_eq(mc1000.duration, 1)
_test_eq(mc2000.duration, 2)
_test_eq(mc5000.duration, 5)
def test_fail_invalid_pad_mode():
with pytest.raises(ValueError):
ResizeSignal(12000, pad_mode="tenchify")
def test_resize_same_duration(audio):
"Asking to resize to the duration should return the audio back"
resize = ResizeSignal(audio.duration * 1000)
inp, out = apply_transform(resize, audio)
_test_eq(inp, out)
def test_padding_both_side_resize(audio):
"Make sure they are padding on both sides"
new_duration = (audio.duration + 1) * 1000
cropsig_pad_after = ResizeSignal(new_duration)
inp, out = apply_transform(cropsig_pad_after, audio)
_test_eq(out[:, 0:2], out[:, -2:])
def test_fail_invalid_pad_mode():
_test_fail(ResizeSignal(12000, pad_mode="tenchify"))