def spectrogram_summary(name, audio, audio_gen, step):
"""Writes a summary of spectrograms for a batch of images."""
specgram = lambda a: spectral_ops.compute_logmag(tf_float32(a), size=768)
# Batch spectrogram operations
spectrograms = specgram(audio)
spectrograms_gen = specgram(audio_gen)
batch_size = int(audio.shape[0])
for sample_idx in range(batch_size):
# Manually specify exact size of fig for tensorboard
fig, axs = plt.subplots(2, 1, figsize=(8, 8))
ax = axs[0]
spec = np.rot90(spectrograms[sample_idx])
ax.matshow(spec, vmin=-5, vmax=1, aspect='auto', cmap=plt.cm.magma)
ax.set_title('original')
ax.set_xticks([])
ax.set_yticks([])
ax = axs[1]
spec = np.rot90(spectrograms_gen[sample_idx])
ax.matshow(spec, vmin=-5, vmax=1, aspect='auto', cmap=plt.cm.magma)
ax.set_title('synthesized')
ax.set_xticks([])
ax.set_yticks([])
# Format and save plot to image
tag = 'spectrogram/{}_{}'.format(name, sample_idx)
fig_summary(tag, fig, step)
def specplot(audio,
vmin=-5,
vmax=1,
rotate=True,
size=512 + 256,
**matshow_kwargs):
"""Plot the log magnitude spectrogram of audio."""
# If batched, take first element.
if len(audio.shape) == 2:
audio = audio[0]
logmag = spectral_ops.compute_logmag(core.tf_float32(audio), size=size)
if rotate:
logmag = np.rot90(logmag)
# Plotting.
plt.matshow(logmag,
vmin=vmin,
vmax=vmax,
cmap=plt.cm.magma,
aspect='auto',
**matshow_kwargs)
plt.xticks([])
plt.yticks([])
plt.xlabel('Time')
plt.ylabel('Frequency')
def call(self, audio, *conditioning):
if self.spectral_op == 'compute_mfcc':
z = spectral_ops.compute_mfcc(
audio,
lo_hz=20.0,
hi_hz=8000.0,
fft_size=self.fft_size,
mel_bins=128,
mfcc_bins=30,
overlap=self.overlap,
pad_end=True)
elif self.spectral_op == 'compute_logmag':
z = spectral_ops.compute_logmag(core.tf_float32(audio), size=self.fft_size)
# Normalize.
z = self.z_norm(z[:, :, tf.newaxis, :])[:, :, 0, :]
n_timesteps = z.shape[1]
conditioning = [resample(c, n_timesteps) for c in conditioning]
z = tf.concat([z] + conditioning, axis=-1)
# Run an RNN over the latents.
z = self.rnn(z)
# Bounce down to compressed z dimensions.
w = tf.math.sigmoid(self.confidence(z))
z = self.dense_out(z)
z = tf.reduce_sum(z * w, axis=1, keepdims=True) / tf.reduce_sum(w, axis=1, keepdims=True)
return z
def spectrogram(audio, sess=None, rotate=False, size=2048):
"""Compute logmag spectrogram."""
if sess is None:
sess = tf.Session()
mag = sess.run(
spectral_ops.compute_logmag(
tf.convert_to_tensor(audio, tf.float32), size=size))
if rotate:
mag = np.rot90(mag)
return mag
def get_spectrogram(audio, rotate=False, size=1024):
"""Compute logmag spectrogram."""
mag = spectral_ops.compute_logmag(tf_float32(audio), size=size)
if rotate:
mag = np.rot90(mag)
return mag
