import tensorflow as tf
import numpy as np
def mu_law_encode(audio, quantization_channels=256):
'''Quantizes waveform amplitudes.
Adapted from https://github.com/ibab/tensorflow-wavenet
'''
with tf.name_scope('encode'):
mu = quantization_channels - 1
safe_audio_abs = tf.minimum(tf.abs(audio), 1.0)
magnitude = tf.log(1. + mu * safe_audio_abs) / tf.log(1. + mu)
signal = tf.sign(audio) * magnitude
return tf.cast((signal + 1) / 2 * mu + 0.5, tf.uint8)
# # Numpy version
# def mu_law_encode(audio, quantization_channels=256):
# '''Quantizes waveform amplitudes.
# Adapted from https://github.com/ibab/tensorflow-wavenet
# '''
# with tf.name_scope('encode'):
# mu = quantization_channels - 1
# safe_audio_abs = np.minimum(np.abs(audio), 1.0)
# magnitude = np.log1p(mu * safe_audio_abs) / np.log1p(mu)
# signal = np.sign(audio) * magnitude
# return ((signal + 1) / 2 * mu + 0.5).astype(np.uint8)
def mu_law_decode(output, quantization_channels=256):
'''Recovers waveform from quantized values.
Adapted from https://github.com/ibab/tensorflow-wavenet
'''
with tf.name_scope('decode'):
mu = quantization_channels - 1
signal = 2 * (tf.to_float(output) / mu) - 1
magnitude = (1 / mu) * ((1 + mu)**abs(signal) - 1)
return tf.sign(signal) * magnitude
def gray2jet(x):
''' NHWC (channel last) format '''
def line(x, xa, xb, ya, yb):
''' a line determined by two points '''
return ya + (x - xa) * (yb - ya) / (xb - xa)
def clip_to_boundary(line1, line2, minval, maxval):
with tf.name_scope('ClipToBoundary'):
x = tf.minimum(line1, line2)
x = tf.minimum(x, maxval)
x = tf.maximum(x, minval)
return x
with tf.name_scope('Gray2Jet'):
r = clip_to_boundary(
line(x, .3515, .66, 0., 1.),
line(x, .8867, 1., 1., .5),
minval=0.,
maxval=1.,
)
g = clip_to_boundary(
line(x, .125, .375, 0., 1.),
line(x, .64, .91, 1., 0.),
minval=0.,
maxval=1.,
)
b = clip_to_boundary(
line(x, .0, .1132, 0.5, 1.0),
line(x, .34, .648, 1., 0.),
minval=0.,
maxval=1.,
)
return tf.concat([r, g, b], axis=-1)
def spectrogram(x, frame_length, nfft=1024):
''' Spectrogram of non-overlapping window '''
with tf.name_scope('Spectrogram'):
shape = tf.shape(x)
b = shape[0]
D = frame_length
t = shape[1] // D
x = tf.reshape(x, [b, t, D])
window = tf.contrib.signal.hann_window(frame_length)
window = tf.expand_dims(window, 0)
window = tf.expand_dims(window, 0) # [1, 1, L]
x = x * window
pad = tf.zeros([b, t, nfft - D])
x = tf.concat([x, pad], -1)
x = tf.cast(x, tf.complex64)
X = tf.fft(x) # TF's API doesn't do padding automatically yet
X = tf.log(tf.abs(X) + 1e-2)
X = X[:, :, :nfft//2 + 1]
X = tf.transpose(X, [0, 2, 1])
X = tf.reverse(X, [1])
X = tf.expand_dims(X, -1)
X = (X - tf.reduce_min(X)) / (tf.reduce_max(X) - tf.reduce_min(X))
X = gray2jet(X)
tf.summary.image('spectrogram', X)
return X
def visualize_wav_prob(x, name):
with tf.name_scope('VisualizeWavProbability'):
x = tf.nn.softmax(x) # [b, T, c]
x = tf.transpose(x, [0, 2, 1])
x = tf.expand_dims(x, -1)
tf.summary.image(name, x)
def visualize_wav(x, n_symbol, name):
with tf.name_scope('VisualizeWav'):
x = tf.one_hot(x, n_symbol) # [b, T,]
x = tf.transpose(x, [0, 2, 1])
x = tf.expand_dims(x, -1)
tf.summary.image(name, x)
def visualize_latent_distr(z_posterior, name):
'''
[b, T, c]
NOTE: feeding [b, h, c] into tf.summary.image results in non-showable TB
'''
with tf.name_scope('VisualizeLatentDistr'):
z_posterior = tf.transpose(z_posterior, [0, 2, 1])
z_posterior = tf.expand_dims(z_posterior, -1)
tf.summary.image(name, z_posterior)
tf.summary.histogram(name, z_posterior)
