def generator(self, x):
net = self.arch['generator']
## DownSample
downS = net['downSample']
c = downS['channel']
k = downS['kernel']
s = downS['stride']
for i in range(len(c)):
x = conv2d(x,
c[i],
k[i],
s[i],
prelu,
name='downSample-L{}'.format(i))
## Residual
resid = net['residual']
c = resid['channel']
k = resid['kernel']
s = resid['stride']
for i in range(len(c) // 2):
y = x
x = conv2d(x,
c[2 * i],
k[2 * i],
s[2 * i],
prelu,
name='resid_conv1-L{}'.format(i))
x = conv2d(x,
c[2 * i + 1],
k[2 * i + 1],
s[2 * i + 1],
prelu,
name='resid_conv2-L{}'.format(i))
x = y + x
## UpSample
upS = net['UpSample']
c = upS['channel']
k = upS['kernel']
s = upS['stride']
r = upS['ratio']
for i in range(len(c)):
x = conv2d(x,
c[i],
k[i],
s[i],
prelu,
name='UpSample-L{}'.format(i))
if (i < len(r)):
x = PS(x, r[i], color=True)
return x
def discriminator(self, x):
unit = self.arch['discriminator']
c = unit['channel']
k = unit['kernel']
s = unit['stride']
for i in range(len(c)):
x = conv2d(x,
c[i],
k[i],
s[i],
prelu,
name='discriminator-L{}'.format(i))
x = tf.layers.flatten(x)
y = tf.layers.dense(x, 1)
return y
def encoder(self, x):
unit = self.arch['encoder']
c = unit['channel']
k = unit['kernel']
s = unit['stride']
for i in range(len(c)):
x = conv2d(x,
c[i],
k[i],
s[i],
prelu,
name='encoder-L{}'.format(i))
x = tf.layers.flatten(x)
z_mu = tf.layers.dense(x, self.arch['z_dim'])
z_var = tf.layers.dense(x, self.arch['z_dim'])
return z_mu, z_var
def discriminator(self, x):
x = tf.reshape(x, [-1, self.featureSize, 1, 1])
unit = self.arch['discriminator']
c = unit['channel']
k = unit['kernel']
s = unit['stride']
for i in range(len(c)):
x = conv2d(x,
c[i],
k[i],
s[i],
prelu,
name='discriminator-L{}'.format(i))
x = tf.layers.flatten(x)
y = tf.layers.dense(x,
1,
bias_initializer=tf.constant_initializer(0.1))
return y
randfile = random.choice(dat[randInd])
randind = random.randint(0, randfile.shape[0] - 4000)
data.append(randfile[randind:randind + 4000])
ind = list(dat.keys()).index(randInd)
lab = np.zeros((1, speakerN))
lab[0][ind] = 1
label.append(lab)
data = np.array(data)
label = np.array(label)
data = data.reshape([batch_size, nodeNum, 1, 1])
return data, label
x = tf.placeholder(tf.float32, shape=[None, nodeNum, 1, 1])
feature = conv2d(x, 512, [5, 1], [2, 1], prelu, 'ex_feature')
en_h1 = conv2d(feature, 256, [5, 1], [2, 1], prelu, 'en_h1')
en_h2 = conv2d(en_h1, 64, [5, 1], [2, 1], prelu, 'en_h2')
de_h1 = deconv2d(en_h2, 256, [5, 1], [2, 1], prelu, 'de_h1')
de_h2 = deconv2d(de_h1, 512, [5, 1], [2, 1], prelu, 'de_h2')
rec_x = deconv2d(de_h2, 1, [5, 1], [2, 1], prelu, 'rec_x')
mse = tf.reduce_mean(tf.pow(rec_x - x, 2))
optimizer = tf.train.AdamOptimizer(0.000001).minimize(mse)
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
