def test_crop_time_after_padding():
sg_orig = test_audio_tensor()
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(sg_orig)
crop_time = CropTime((sg.duration + 5) * 1000, pad_mode=AudioPadType.Zeros_After)
inp, out = apply_transform(crop_time, sg.clone())
_test_ne(sg.duration, sg_orig.duration)
def test_crop_time():
for i in [1, 2, 5]:
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
audio = test_audio_tensor(seconds=3)
crop = CropTime(i * 1000)
inp, out = apply_transform(crop, a2s(audio))
_test_eq(i, round(out.duration))
_test_close(out.width, int((i / inp.duration) * inp.width), eps=1.01)
def test_resize_int():
# Test when size is an int
size = 224
resize_int = TfmResize(size)
audio = test_audio_tensor()
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(audio)
inp, out = apply_transform(resize_int, sg)
_test_eq(out.shape[1:], torch.Size([size, size]))
def test_crop_time_repeat_padding():
"Test that repeat padding works when cropping time"
repeat = 3
audio = test_audio_tensor()
crop_12000ms_repeat = CropTime(repeat * 1000 * audio.duration,
pad_mode=AudioPadType.Repeat)
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(audio)
inp, out = apply_transform(crop_12000ms_repeat, sg)
_test_eq(inp.width, sg.width)
_test_ne(sg.width, out.width)
def test_delta_channels():
" nchannels for a spectrogram is how many channels its original audio had "
delta = Delta()
audio = test_audio_tensor(channels=1)
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(audio)
inp, out = apply_transform(delta, sg)
_test_eq(out.nchannels, inp.nchannels * 3)
_test_eq(out.shape[1:], inp.shape[1:])
_test_ne(out[0], out[1])
def test_mask_freq():
# create a random time mask and test that it is being correctly applied
size, start, val = [random.randint(1, 50) for i in range(3)]
time_mask_test = MaskTime(size=size, start=start, val=val)
audio = test_audio_tensor()
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(audio)
inp, out = apply_transform(time_mask_test, sg)
_test_eq(
out[:, :, start:start + size],
val * torch.ones_like(inp)[:, :, start:start + size],
)
def test_delta_channels():
" nchannels for a spectrogram is how many channels its original audio had "
delta = DeltaGPU()
# Explicitly check more than one channel
audio = test_audio_tensor(channels=2)
a2s = AudioToSpec.from_cfg(AudioConfig.Voice())
sg = a2s(audio)
inp, out = apply_transform(delta, sg)
_test_eq(out.nchannels, inp.nchannels * 3)
_test_eq(out.shape[-2:], inp.shape[-2:])
for i1, i2 in [(0, 2), (1, 3), (0, 4), (1, 5), (2, 4), (3, 5)]:
assert not torch.allclose(out[i1], out[i2])