def test_torch_complex_format(dim, max_tested_ndim):
# Random tensor shape
tensor_shape = [random.randint(1, 10) for _ in range(max_tested_ndim)]
# Make sure complex dimension has even shape
tensor_shape[dim] = 2 * tensor_shape[dim]
complex_tensor = torch.randn(tensor_shape)
ta_tensor = transforms.to_torch_complex(complex_tensor, dim=dim)
tensor_back = transforms.from_torch_complex(ta_tensor, dim=dim)
assert_allclose(complex_tensor, tensor_back)
def forward_encoder(self, wav):
tf_rep = self.encoder(wav)
return to_torch_complex(tf_rep)
def forward_encoder(self, wav):
tf_rep = self.encoder(wav)
# Remove Nyquist frequency bin
return to_torch_complex(tf_rep)[..., :-1, :]
SCM,
RTFMVDRBeamformer,
SDWMWFBeamformer,
GEVBeamformer,
stable_cholesky,
)
torch_has_complex_support = tuple(map(
int,
torch.__version__.split(".")[:2])) >= (1, 8)
_stft, _istft = make_enc_dec("stft",
kernel_size=512,
n_filters=512,
stride=128)
stft = lambda x: tr.to_torch_complex(_stft(x))
istft = lambda x: _istft(tr.from_torch_complex(x))
@pytest.mark.skipif(not torch_has_complex_support, "No complex support ")
def _default_beamformer_test(beamformer: Beamformer,
n_mics=4,
*args,
**kwargs):
scm = SCM()
speech = torch.randn(1, n_mics, 16000 * 6)
noise = torch.randn(1, n_mics, 16000 * 6)
mix = speech + noise
# GeV Beamforming
mix_stft = stft(mix)