def encoder(x, comp_ratio, noise_std, activation='prelu'):
# input_dim = x.shape[-1] #x.get_shape().as_list()[-1]
num_filters = int(np.log2(comp_ratio))
h = x;
if len(x.get_shape().as_list())<3:
h = tf.expand_dims(h, axis =2)
skips = []
for i in range(num_filters):
print('enc_layer_{}'.format( i))
with tf.variable_scope('enc'):
h = mf_downconv(h, output_dim = 2**(i + 1) , stride = 2,
filter_width=filter_width, name='downconv_{}'.format(i))
print('h_dim= ', h.get_shape().as_list())
# Skip connections
if i < num_filters - 1:
skips.append(h)
print('skip_dim= ', h.get_shape().as_list())
if apply_BN:
h = segan.vbn( h, 'ae_enc_vbn_{}'.format(i))
if activation=='leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('enc'):
h , _ = prelu(h, name='prelu_{}'.format(i))
z = make_z(h.get_shape().as_list(), std = noise_std)
# h = h + z
h = tf.concat([z, h], 2)
print('h_and_z_dim', h.get_shape().as_list())
return h, skips
def decoder(x, comp_ratio, activation='leakyrelu', nonlin='segan'): #gated_conv'):
input_dim = x.shape[-1] #tf.shape(x)[-1]
num_filters = int(np.log2(comp_ratio))
h=x;
if apply_BN:
h = batch_norm (h, 'dec_batch_norm_{}'.format(0))
# with tf.variable_scope('g_dec'):
# w_full_dec = tf.get_variable('w_full_dec', [input_dim ,
# input_dim * comp_ratio / (2**num_filters) + 1],
# initializer= tf.random_normal_initializer(stddev=0.02))
# h = full_layer(x, w_full_dec, 'g_dec')
for i in range(num_filters):
print('dec_layer_number = ', i)
print('h_dim= ', h.get_shape().as_list())
with tf.variable_scope('g_dec'):
if nonlin=='segan':
#original segan
h = nn_deconv(h, filter_width=filter_width, name='nn_deconv_{}'.format(i),
dilation=2, init=tf.truncated_normal_initializer(stddev=0.02))
if apply_BN:
h = batch_norm (h, 'batch_norm_{}'.format(i))
if activation=='leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_dec'):
h,_ = prelu(h, name='prelu_{}'.format(i))
else:
# Wavenet gated convolutions
h = gated_deconv(h, filter_width=filter_width, name='gated_deconv_{}'.format(i),
dilation=2, init=tf.truncated_normal_initializer(stddev=0.02))
return h
def decoder(x, activation='leakyrelu', nonlin='segan'): #gated_conv'):
# input_dim = x.shape[-1]
num_filters = int(np.log2(comp_ratio))
h = x
for i in range(num_filters):
print('dec_layer_number = ', i)
with tf.variable_scope('g_dec'):
if nonlin == 'segan':
#original segan
h = nn_deconv(
h,
filter_width=filter_width,
name='nn_deconv_{}'.format(i),
dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02))
h = batch_norm(h, 'batch_norm_{}'.format(i))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_dec'):
h, _ = prelu(h, name='prelu_{}'.format(i))
else:
# Wavenet gated convolutions
h = gated_deconv(
h,
filter_width=filter_width,
name='gated_deconv_{}'.format(i),
dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02))
return h
def encoder(x, comp_ratio, activation='prelu'):
input_dim = x.shape[-1] #tf.shape(x)[-1]
num_filters = int(np.log2(comp_ratio))
h = x;
if apply_BN:
h = batch_norm(h, 'enc_batch_norm_{}'.format(0))
for i in range(num_filters):
print('enc_layer_number = ', i)
print('h_dim= ', h.get_shape().as_list())
with tf.variable_scope('g_enc'):
h = downconv(h, filter_width=filter_width, name='downconv_{}'.format(i))
if apply_BN:
h = batch_norm(h, 'batch_norm_{}'.format(i))
if activation=='leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_enc'):
h , _ = prelu(h, name='prelu_{}'.format(i))
# with tf.variable_scope('g_enc'):
# w_full_enc = tf.get_variable('w_full_enc', [ input_dim / (2**num_filters), input_dim / comp_ratio],
# initializer= tf.random_normal_initializer(stddev=0.01))
# h = full_layer(h, w_full_enc, 'g_enc')
return h
def conv_net(x, num_layers, activation='prelu'):
# input_dim = x.shape[-1] #x.get_shape().as_list()[-1]
h = x
if len(x.get_shape().as_list()) < 3:
h = tf.expand_dims(h, axis=2)
# skips = []
for i in range(num_layers):
print('conv_layer_{}'.format(i))
with tf.variable_scope('conv'):
h = mf_downconv(h,
output_dim=1,
stride=1,
filter_width=31,
name='downconv_{}'.format(i))
print('h_dim= ', h.get_shape().as_list())
# Skip connections
# if i < num_filters - 1:
# skips.append(h)
# print('skip_dim= ', h.get_shape().as_list())
if apply_BN:
h = segan.vbn(h, 'ae_enc_vbn_{}'.format(i))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('enc'):
h, _ = prelu(h, name='prelu_{}'.format(i))
h = tf.squeeze(h)
return h #, skips
def decoder(x, activation='leakyrelu'):
num_filters = int(np.log2(comp_ratio))
h = x
for i in range(num_filters):
print('dec_layer_number = ', i)
with tf.variable_scope('g_dec'):
h = nn_deconv(h,
filter_width=filter_width,
name='nn_deconv_{}'.format(i),
dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_dec'):
h, _ = prelu(h, name='prelu_{}'.format(i))
return h
def encoder(x, activation='prelu'):
num_filters = int(np.log2(comp_ratio))
h = x
hs = []
for i in range(num_filters):
print('enc_layer_number = ', i)
hs.append(h)
with tf.variable_scope('g_enc'):
h = downconv(h,
filter_width=filter_width,
name='downconv_{}'.format(i))
h = batch_norm(h, 'batch_norm_{}'.format(i))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_enc'):
h, _ = prelu(h, name='prelu_{}'.format(i))
return h, hs
def decoder(x,
comp_ratio,
activation='leakyrelu',
nonlin='segan'): #gated_conv'):
input_dim = x.shape[-1] #tf.shape(x)[-1]
num_filters = int(np.log2(comp_ratio))
h = x
if apply_BN:
h = batch_norm(h, 'dec_batch_norm_{}'.format(0))
for i in range(num_filters):
print('dec_layer_number = ', i)
print('h_dim= ', h.get_shape().as_list())
with tf.variable_scope('g_dec'):
if nonlin == 'segan':
#original segan
h = mf_nn_deconv(
h,
output_dim=2**(num_filters - i - 1),
filter_width=filter_width,
name='nn_deconv_{}'.format(i),
dilation=4,
init=tf.truncated_normal_initializer(stddev=0.02))
if apply_BN:
h = batch_norm(h, 'batch_norm_{}'.format(i))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_dec'):
h, _ = prelu(h, name='prelu_{}'.format(i))
else:
# Wavenet gated convolutions
h = gated_deconv(
h,
filter_width=filter_width,
name='gated_deconv_{}'.format(i),
dilation=2,
init=tf.truncated_normal_initializer(stddev=0.02))
# making h of size [num_batch, input_dim]
h = tf.squeeze(h, axis=-1)
return h
def discriminator(x, comp_ratio):
# input_dim = x.get_shape().as_list()[-1] #x.shape[-1]
num_filters = int(np.log2(comp_ratio))
h = x;
with tf.variable_scope('disc'):
# if apply_BN:
# h = segan.vbn(h, 'disc_batch_norm_{}'.format(0))
for i in range(num_filters):
print('disc_layer_{} = '.format( i))
h = mf_downconv(h, output_dim = 2 ** (i + 1) , stride = 2,
filter_width=filter_width, name='disc_downconv_{}'.format(i))
print('h_dim= ', h.get_shape().as_list())
if apply_BN:
h = segan.vbn( h, 'disc_vbn_{}'.format(i))
h = leakyrelu(h)
# Last layer
h = flatten(h)
h_logit_out = conv1d(h, filter_width=1, output_dim=1,
w_init = tf.truncated_normal_initializer(stddev=0.02),
name='disc_logits_conv') # 1 x 1 convolution
d_logit_out = tf.squeeze(h_logit_out) # Remove dimension of 1 comming from conv1d
disc_output = fully_connected(d_logit_out, 1, activation_fn = None) #tf.tanh)
return disc_output
def encoder(x, comp_ratio, activation='prelu'):
input_dim = x.shape[-1] #x.get_shape().as_list()[-1]
num_filters = int(np.log2(comp_ratio))
h = x
if apply_BN:
h = batch_norm(h, 'enc_batch_norm_{}'.format(0))
for i in range(num_filters):
print('enc_layer_number = ', i)
print('h_dim= ', h.get_shape().as_list())
with tf.variable_scope('g_enc'):
h = mf_downconv(h,
output_dim=2**i,
stride=4,
filter_width=filter_width,
name='downconv_{}'.format(i))
if apply_BN:
h = batch_norm(h, 'batch_norm_{}'.format(i))
if activation == 'leakyrelu':
h = leakyrelu(h)
else:
with tf.variable_scope('g_enc'):
h, _ = prelu(h, name='prelu_{}'.format(i))
return h
