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_mask_time():
c, f, t = 2, 120, 80
min_size = 5
max_size = 7
sg = AudioSpectrogram(torch.rand([c, f, t]))
val = 10 # Use a value not in the original spectrogram
gradient_sg = AudioSpectrogram(
torch.linspace(0, 1, t).view(1, 1, t).repeat([c, f, 1]))
ones = torch.ones_like(sg)
# Test patching with mean
with patch(
"fastaudio.augment.functional.region_mask",
side_effect=[
torch.BoolTensor([[1] * 10 + [0] * (t - 10)]),
],
):
mask_with_mean = MaskTimeGPU(min_size=min_size,
max_size=max_size,
mask_val=None)
# Use a gradient so we can be sure the mean will never show up outside the mask
inp, out = apply_transform(mask_with_mean, gradient_sg)
channelwise_mean = inp[..., :10].mean(dim=(-2, -1)).reshape(-1, 1, 1)
_test_close(
out[..., :10],
channelwise_mean * ones[..., :10],
)
assert not (out[..., 10:]
== channelwise_mean).any(), out == channelwise_mean
# Test multiple masks (and patching with value)
with patch(
"fastaudio.augment.functional.region_mask",
side_effect=[
torch.BoolTensor([[1] * 10 + [0] * (t - 10),
[0] * (t - 10) + [1] * 10]),
],
):
mask_with_val = MaskTimeGPU(min_size=min_size,
num_masks=2,
max_size=max_size,
mask_val=val)
inp, out = apply_transform(mask_with_val, sg)
_test_eq(
out[..., :10],
val * ones[..., :10],
)
_test_eq(
out[..., t - 10:],
val * ones[..., t - 10:],
)
matches = out[..., 10:t - 10] == val
assert not matches.any(), matches
def test_upsample(audio):
"""
Make sure that the Upsampling is possible. This can
take a while depending on the target sample rate
"""
for _ in range(10):
random_sr = random.randint(16000, 72000)
random_upsample = Resample(random_sr)(audio)
num_samples = random_upsample.nsamples
_test_close(num_samples,
abs(audio.nsamples // (audio.sr / random_sr)),
eps=1.1)
def test_upsample(audio):
for _ in range(10):
random_sr = random.randint(16000, 72000)
random_upsample = Resample(random_sr)(audio)
num_samples = random_upsample.nsamples
_test_close(num_samples, abs(audio.nsamples // (audio.sr / random_sr)), eps=1.1)