def test_amplitude_to_DB_not_channelwise_clamps(self):
"""Check that clamps are applied per-item, not per channel."""
amplitude_mult = 20.
amin = 1e-10
ref = 1.0
db_mult = math.log10(max(amin, ref))
top_db = 40.
torch.manual_seed(0)
spec = torch.rand([1, 2, 100, 100]) * 200
# Make one channel blow out the other
spec[:, 0] += 50
specwise_dbs = F.amplitude_to_DB(spec,
amplitude_mult,
amin,
db_mult,
top_db=top_db)
channelwise_dbs = torch.stack([
F.amplitude_to_DB(spec[:, i],
amplitude_mult,
amin,
db_mult,
top_db=top_db) for i in range(spec.size(-3))
])
# Just check channelwise gives a different answer.
difference = (specwise_dbs - channelwise_dbs).abs()
assert (difference >= 1e-5).any()
def test_amplitude_to_DB_itemwise_clamps(self):
"""Ensure that the clamps are separate for each spectrogram in a batch.
The clamp was determined per-batch in a prior implementation, which
meant it was determined by the loudest item, thus items weren't
independent. See:
https://github.com/pytorch/audio/issues/994
"""
amplitude_mult = 20.
amin = 1e-10
ref = 1.0
db_mult = math.log10(max(amin, ref))
top_db = 20.
# Make a batch of noise
torch.manual_seed(0)
spec = torch.rand([2, 2, 100, 100]) * 200
# Make one item blow out the other
spec[0] += 50
batchwise_dbs = F.amplitude_to_DB(spec, amplitude_mult, amin,
db_mult, top_db=top_db)
itemwise_dbs = torch.stack([
F.amplitude_to_DB(item, amplitude_mult, amin,
db_mult, top_db=top_db)
for item in spec
])
self.assertEqual(batchwise_dbs, itemwise_dbs)
def test_amplitude_to_DB(self):
spec = torch.rand((6, 201))
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
# Power to DB
multiplier = 10.0
ta_out = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier,
top_db)
lr_out = librosa.core.power_to_db(spec.numpy())
lr_out = torch.from_numpy(lr_out)
self.assertEqual(ta_out, lr_out, atol=5e-5, rtol=1e-5)
# Amplitude to DB
multiplier = 20.0
ta_out = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier,
top_db)
lr_out = librosa.core.amplitude_to_db(spec.numpy())
lr_out = torch.from_numpy(lr_out)
self.assertEqual(ta_out, lr_out, atol=5e-5, rtol=1e-5)
def test_amplitude_to_DB_reversible(self, shape):
"""Round trip between amplitude and db should return the original for various shape
This implicitly also tests `DB_to_amplitude`.
"""
amplitude_mult = 20.
power_mult = 10.
amin = 1e-10
ref = 1.0
db_mult = math.log10(max(amin, ref))
torch.manual_seed(0)
spec = torch.rand(*shape, dtype=self.dtype, device=self.device) * 200
# Spectrogram amplitude -> DB -> amplitude
db = F.amplitude_to_DB(spec,
amplitude_mult,
amin,
db_mult,
top_db=None)
x2 = F.DB_to_amplitude(db, ref, 0.5)
self.assertEqual(x2, spec, atol=5e-5, rtol=1e-5)
# Spectrogram power -> DB -> power
db = F.amplitude_to_DB(spec, power_mult, amin, db_mult, top_db=None)
x2 = F.DB_to_amplitude(db, ref, 1.)
self.assertEqual(x2, spec)
def test_DB_to_amplitude(self):
# Make some noise
x = torch.rand(1000)
spectrogram = torchaudio.transforms.Spectrogram()
spec = spectrogram(x)
amin = 1e-10
ref = 1.0
db_multiplier = math.log10(max(amin, ref))
# Waveform amplitude -> DB -> amplitude
multiplier = 20.
power = 0.5
db = F.amplitude_to_DB(torch.abs(x),
multiplier,
amin,
db_multiplier,
top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertEqual(x2, torch.abs(x), atol=5e-5, rtol=1e-5)
# Spectrogram amplitude -> DB -> amplitude
db = F.amplitude_to_DB(spec,
multiplier,
amin,
db_multiplier,
top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertEqual(x2, spec, atol=5e-5, rtol=1e-5)
# Waveform power -> DB -> power
multiplier = 10.
power = 1.
db = F.amplitude_to_DB(x, multiplier, amin, db_multiplier, top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertEqual(x2, torch.abs(x), atol=5e-5, rtol=1e-5)
# Spectrogram power -> DB -> power
db = F.amplitude_to_DB(spec,
multiplier,
amin,
db_multiplier,
top_db=None)
x2 = F.DB_to_amplitude(db, ref, power)
self.assertEqual(x2, spec, atol=5e-5, rtol=1e-5)
def func(tensor):
multiplier = 10.0
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
return F.amplitude_to_DB(tensor, multiplier, amin, db_multiplier,
top_db)
def test_amplitude_to_DB_top_db_clamp(self, shape):
"""Ensure values are properly clamped when `top_db` is supplied."""
amplitude_mult = 20.
amin = 1e-10
ref = 1.0
db_mult = math.log10(max(amin, ref))
top_db = 40.
torch.manual_seed(0)
# A random tensor is used for increased entropy, but the max and min for
# each spectrogram still need to be predictable. The max determines the
# decibel cutoff, and the distance from the min must be large enough
# that it triggers a clamp.
spec = torch.rand(*shape, dtype=self.dtype, device=self.device)
# Ensure each spectrogram has a min of 0 and a max of 1.
spec -= spec.amin([-2, -1])[..., None, None]
spec /= spec.amax([-2, -1])[..., None, None]
# Expand the range to (0, 200) - wide enough to properly test clamping.
spec *= 200
decibels = F.amplitude_to_DB(spec,
amplitude_mult,
amin,
db_mult,
top_db=top_db)
# Ensure the clamp was applied
below_limit = decibels < 6.0205
assert not below_limit.any(), (
"{} decibel values were below the expected cutoff:\n{}".format(
below_limit.sum().item(), decibels))
# Ensure it didn't over-clamp
close_to_limit = decibels < 6.0207
assert close_to_limit.any(), (
f"No values were close to the limit. Did it over-clamp?\n{decibels}"
)
def spec_to_db(x_spec, top_db=80, amin=1e-5):
x_spec = x_spec.transpose(-3, -1).contiguous()
x_spec = torch.view_as_complex(x_spec)
x_mag = torch.abs(x_spec)
x_db = amplitude_to_DB(x_mag,
multiplier=20.,
amin=amin,
db_multiplier=math.log10(max(amin, 1.)),
top_db=top_db)
return x_db
def test_amplitude_to_DB(self):
amin = 1e-10
db_multiplier = 0.0
top_db = 80.0
multiplier = 20.0
spec = get_spectrogram(get_whitenoise(device=self.device, dtype=self.dtype), power=1)
result = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db)
expected = librosa.core.amplitude_to_db(spec[0].cpu().numpy())[None, ...]
self.assertEqual(result, torch.from_numpy(expected))
def forward(self, x: Tensor) -> Tensor:
r"""Numerically stable implementation from Librosa.
https://librosa.github.io/librosa/_modules/librosa/core/spectrum.html
Args:
x (Tensor): Input tensor before being converted to decibel scale.
Returns:
Tensor: Output tensor in decibel scale.
"""
return F.amplitude_to_DB(x, self.multiplier, self.amin, self.db_multiplier, self.top_db)
def forward(self, x: Tensor) -> Tensor:
r"""Numerically stable implementation from Librosa.
https://librosa.org/doc/latest/generated/librosa.amplitude_to_db.html
Args:
x (Tensor): Input tensor before being converted to decibel scale.
Returns:
Tensor: Output tensor in decibel scale.
"""
return F.amplitude_to_DB(x, self.multiplier, self.amin, self.db_multiplier, self.top_db)
def test_torchscript_amplitude_to_DB(self):
@torch.jit.script
def jit_method(spec, multiplier, amin, db_multiplier, top_db):
# type: (Tensor, float, float, float, Optional[float]) -> Tensor
return F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db)
spec = torch.rand((6, 201))
multiplier = 10.
amin = 1e-10
db_multiplier = 0.
top_db = 80.
jit_out = jit_method(spec, multiplier, amin, db_multiplier, top_db)
py_out = F.amplitude_to_DB(spec, multiplier, amin, db_multiplier, top_db)
self.assertTrue(torch.allclose(jit_out, py_out))
def jit_method(spec, multiplier, amin, db_multiplier, top_db):
# type: (Tensor, float, float, float, Optional[float]) -> Tensor
return F.amplitude_to_DB(spec, multiplier, amin, db_multiplier,
top_db)
def __call__(self, x_spec):
x_mag = ConvToMag()(x_spec)
x_db = amplitude_to_DB(x_mag, multiplier=20., amin=1e-5, db_multiplier=math.log10(max(1e-5, 1.)), top_db=120.)
return x_db
