def test_center_freq_correction(kernel_size, stride_factor):
spec = torch.randn(2, kernel_size + 2, 50)
stride = None if stride_factor is None else kernel_size // stride_factor
new_spec = transforms.centerfreq_correction(spec,
kernel_size=kernel_size,
stride=stride)
assert spec.shape == new_spec.shape
assert_allclose(transforms.mag(spec), transforms.mag(new_spec))
def test_pmsqe_pit(n_src, sample_rate):
# Define supported STFT
if sample_rate == 16000:
stft = Encoder(STFTFB(kernel_size=512, n_filters=512, stride=256))
else:
stft = Encoder(STFTFB(kernel_size=256, n_filters=256, stride=128))
# Usage by itself
ref, est = torch.randn(2, n_src, 16000), torch.randn(2, n_src, 16000)
ref_spec = transforms.mag(stft(ref))
est_spec = transforms.mag(stft(est))
loss_func = PITLossWrapper(SingleSrcPMSQE(sample_rate=sample_rate), pit_from="pw_pt")
# Assert forward ok.
loss_func(est_spec, ref_spec)
def dc_head_separate(self, x):
""" Cluster embeddings to produce binary masks, output waveforms """
kmeans = KMeans(n_clusters=self.masker.n_src)
if len(x.shape) == 2:
x = x.unsqueeze(1)
tf_rep = self.encoder(x)
mag_spec = mag(tf_rep)
proj, mask_out = self.masker(mag_spec)
active_bins = ebased_vad(mag_spec)
active_proj = proj[active_bins.view(1, -1)]
#
bin_clusters = kmeans.fit_predict(active_proj.cpu().data.numpy())
# Create binary masks
est_mask_list = []
for i in range(self.masker.n_src):
# Add ones in all inactive bins in each mask.
mask = ~active_bins
mask[active_bins] = torch.from_numpy((bin_clusters == i)).to(mask.device)
est_mask_list.append(mask.float()) # Need float, not bool
# Go back to time domain
est_masks = torch.stack(est_mask_list, dim=1)
masked = apply_mag_mask(tf_rep, est_masks)
wavs = pad_x_to_y(self.decoder(masked), x)
dic_out = dict(tfrep=tf_rep, mask=mask_out, masked_tfrep=masked, proj=proj)
return wavs, dic_out
def unpack_data(self, batch, EPS=1e-8):
mix, sources, noise = batch
# Take only the first channel
mix = mix[..., 0]
sources = sources[..., 0]
noise = noise[..., 0]
noise = noise.unsqueeze(1)
# Compute magnitude spectrograms and IRM
src_mag_spec = mag(self.model.encoder(sources))
noise_mag_spec = mag(self.model.encoder(noise))
noise_mag_spec = noise_mag_spec.unsqueeze(1)
real_mask = src_mag_spec / (noise_mag_spec +
src_mag_spec.sum(1, keepdim=True) + EPS)
# Get the src idx having the maximum energy
binary_mask = real_mask.argmax(1)
return mix, binary_mask, real_mask
def test_pmsqe(sample_rate):
# Define supported STFT
if sample_rate == 16000:
stft = Encoder(STFTFB(kernel_size=512, n_filters=512, stride=256))
else:
stft = Encoder(STFTFB(kernel_size=256, n_filters=256, stride=128))
# Usage by itself
ref, est = torch.randn(2, 1, 16000), torch.randn(2, 1, 16000)
ref_spec = transforms.mag(stft(ref))
est_spec = transforms.mag(stft(est))
loss_func = SingleSrcPMSQE(sample_rate=sample_rate)
loss_value = loss_func(est_spec, ref_spec)
# Assert output has shape (batch,)
assert loss_value.shape[0] == ref.shape[0]
# Assert support for transposed inputs.
tr_loss_value = loss_func(est_spec.transpose(1, 2), ref_spec.transpose(1, 2))
assert_allclose(loss_value, tr_loss_value)
def unpack_data(self, batch, EPS=1e-8):
mix, sources = batch
# Compute magnitude spectrograms and IRM
src_mag_spec = mag(self.model.encoder(sources))
real_mask = src_mag_spec / (src_mag_spec.sum(1, keepdim=True) + EPS)
# Get the src idx having the maximum energy
binary_mask = real_mask.argmax(1)
return mix, binary_mask, real_mask
def test_griffinlim(fb_config, feed_istft, feed_angle):
stft = Encoder(STFTFB(**fb_config))
istft = None if not feed_istft else Decoder(STFTFB(**fb_config))
wav = torch.randn(2, 1, 8000)
spec = stft(wav)
tf_mask = torch.sigmoid(torch.randn_like(spec))
masked_spec = spec * tf_mask
mag = transforms.mag(masked_spec, -2)
angles = None if not feed_angle else transforms.angle(masked_spec, -2)
griffin_lim(mag, stft, angles=angles, istft_dec=istft, n_iter=3)
def separate(self, x):
""" Separate with mask-inference head, output waveforms """
if len(x.shape) == 2:
x = x.unsqueeze(1)
tf_rep = self.encoder(x)
proj, mask_out = self.masker(mag(tf_rep))
masked = apply_mag_mask(tf_rep.unsqueeze(1), mask_out)
wavs = torch_utils.pad_x_to_y(self.decoder(masked), x)
dic_out = dict(tfrep=tf_rep, mask=mask_out, masked_tfrep=masked, proj=proj)
return wavs, dic_out
def distance(estimate, target, is_complex=True):
"""Compute the average distance in the complex plane. Makes more sense
when the network computes a complex mask.
Args:
estimate (torch.Tensor): Estimate complex spectrogram.
target (torch.Tensor): Speech target complex spectrogram.
is_complex (bool): Whether to compute the distance in the complex or
the magnitude space.
Returns:
torch.Tensor the loss value, in a tensor of size 1.
"""
if is_complex:
# Take the difference in the complex plane and compute the squared norm
# of the remaining vector.
return mag(estimate - target).pow(2).mean()
else:
# Compute the mean difference between magnitudes.
return (mag(estimate) - mag(target)).pow(2).mean()
def test_angle_mag_recompostion(dim):
""" Test complex --> (mag, angle) --> complex conversions"""
max_tested_ndim = 4
# 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)
phase = transforms.angle(complex_tensor, dim=dim)
mag = transforms.mag(complex_tensor, dim=dim)
tensor_back = transforms.from_magphase(mag, phase, dim=dim)
assert_allclose(complex_tensor, tensor_back)
def common_step(self, batch, batch_nb, train=False):
inputs, targets, masks = self.unpack_data(batch)
embeddings, est_masks = self(inputs)
spec = mag(self.model.encoder(inputs.unsqueeze(1)))
if self.mask_mixture:
est_masks = est_masks * spec.unsqueeze(1)
masks = masks * spec.unsqueeze(1)
loss, loss_dic = self.loss_func(embeddings,
targets,
est_src=est_masks,
target_src=masks,
mix_spec=spec)
return loss, loss_dic
def test_misi(fb_config, feed_istft, feed_angle):
stft = Encoder(STFTFB(**fb_config))
istft = None if not feed_istft else Decoder(STFTFB(**fb_config))
n_src = 3
# Create mixture
wav = torch.randn(2, 1, 8000)
spec = stft(wav).unsqueeze(1)
# Create n_src masks on mixture spec and apply them
shape = list(spec.shape)
shape[1] *= n_src
tf_mask = torch.sigmoid(torch.randn(*shape))
masked_specs = spec * tf_mask
# Separate mag and angle.
mag = transforms.mag(masked_specs, -2)
angles = None if not feed_angle else transforms.angle(masked_specs, -2)
est_wavs = misi(wav, mag, stft, angles=angles, istft_dec=istft, n_iter=2)
# We actually don't know the last dim because ISTFT(STFT()) cuts the end
assert est_wavs.shape[:-1] == (2, n_src)
def forward(self, x):
if len(x.shape) == 2:
x = x.unsqueeze(1)
# Compute STFT
tf_rep = self.encoder(x)
# Estimate TF mask from STFT features : cat([re, im, mag])
if self.is_complex:
to_masker = magreim(tf_rep)
else:
to_masker = mag(tf_rep)
# LSTM masker expects a feature dimension last (not like 1D conv)
est_masks = self.masker(to_masker.transpose(1, 2)).transpose(1, 2)
# Apply TF mask
if self.is_complex:
masked_tf_rep = apply_real_mask(tf_rep, est_masks)
else:
masked_tf_rep = apply_mag_mask(tf_rep, est_masks)
return masked_tf_rep
def test_mag(encoder_list):
for (enc, fb_dim) in encoder_list:
tf_rep = enc(torch.randn(2, 1, 16000)) # [batch, freq, time]
batch, freq, time = tf_rep.shape
mag = transforms.mag(tf_rep, dim=1)
assert mag.shape == (batch, freq // 2, time)
def forward(self, x):
if len(x.shape) == 2:
x = x.unsqueeze(1)
tf_rep = self.encoder(x)
final_proj, mask_out = self.masker(mag(tf_rep))
return final_proj, mask_out
