def forward(self, x: Tensor) -> Tensor:
r"""
Args:
x (Tensor): A signal to be encoded.
Returns:
x_mu (Tensor): An encoded signal.
"""
return F.mu_law_encoding(x, self.quantization_channels)
def test_torchscript_mu_law_encoding(self):
@torch.jit.script
def jit_method(tensor, qc):
# type: (Tensor, int) -> Tensor
return F.mu_law_encoding(tensor, qc)
tensor = torch.rand((10, 1))
qc = 256
jit_out = jit_method(tensor, qc)
py_out = F.mu_law_encoding(tensor, qc)
self.assertTrue(torch.allclose(jit_out, py_out))
def func(tensor):
qc = 256
return F.mu_law_encoding(tensor, qc)
def jit_method(tensor, qc):
# type: (Tensor, int) -> Tensor
return F.mu_law_encoding(tensor, qc)
#Mu-Law decoding
reconstructed = torchaudio.transforms.MuLawDecoding()(transformed)
print(f"Shape of recovered waveform: {reconstructed.size()}")
plt.figure()
plt.plot(reconstructed[0,:].numpy())
err = ((waveform-reconstructed).abs() / waveform.abs()).median()
print(f"Median relative difference between original and MuLaw reconstructed signals: {err:.2%}")
# %%
#Using torchaudio.functional
mu_law_encoding_waveform = F.mu_law_encoding(waveform, quantization_channels=256)
print(f"Shape of transformed waveform: {mu_law_encoding_waveform.size()}")
plt.figure()
plt.plot(mu_law_encoding_waveform[0,:].numpy())
# %%
#Functional spectrogram deltas
specgram = torchaudio.transforms.Spectrogram()(waveform)
computed = torchaudio.functional.compute_deltas(specgram.contiguous(), win_length=3)
print(f"Shape of computed deltas: {computed.shape}")
plt.figure()
