def _create_iaf(self, inputs, iaf_idx):
num_stages = self.hparams.num_stages
num_layers = self.hparams.num_iaf_layers[iaf_idx]
filter_length = self.hparams.filter_length
width = self.hparams.width
out_width = self.out_width
deconv_width = self.hparams.deconv_width
deconv_config = self.hparams.deconv_config # [[l1, s1], [l2, s2]]
mel = inputs['mel']
x = inputs['x']
iaf_name = 'iaf_{:d}'.format(iaf_idx + 1)
mel_en = wavenet._deconv_stack(mel,
deconv_width,
deconv_config,
name=iaf_name)
l = masked.shift_right(x)
l = masked.conv1d(l,
num_filters=width,
filter_length=filter_length,
name='{}/start_conv'.format(iaf_name))
for i in range(num_layers):
dilation = 2**(i % num_stages)
d = masked.conv1d(l,
num_filters=2 * width,
filter_length=filter_length,
dilation=dilation,
name='{}/dilated_conv_{:d}'.format(
iaf_name, i + 1))
c = masked.conv1d(mel_en,
num_filters=2 * width,
filter_length=1,
name='{}/mel_cond_{:d}'.format(iaf_name, i + 1))
d = wavenet._condition(d, c)
assert d.get_shape().as_list()[2] % 2 == 0
m = d.get_shape().as_list()[2] // 2
d_sigmoid = tf.sigmoid(d[:, :, :m])
d_tanh = tf.tanh(d[:, :, m:])
d = d_sigmoid * d_tanh
l += masked.conv1d(d,
num_filters=width,
filter_length=1,
name='{}/res_{:d}'.format(iaf_name, i + 1))
l = tf.nn.relu(l)
l = masked.conv1d(l,
num_filters=width,
filter_length=1,
name='{}/out1'.format(iaf_name))
c = masked.conv1d(mel_en,
num_filters=width,
filter_length=1,
name='{}/mel_cond_out1'.format(iaf_name))
l = wavenet._condition(l, c)
l = tf.nn.relu(l)
out = masked.conv1d(l,
num_filters=out_width,
filter_length=1,
name='{}/out2'.format(iaf_name))
mean, scale = tf.split(out, num_or_size_splits=2, axis=2)
scale = tf.clip_by_value(scale, tf.exp(-7.0), tf.exp(7.0))
new_x = x * scale + mean
return {'x': new_x, 'mean': mean, 'scale': scale}
def _create_iaf(self, inputs, iaf_idx, init):
num_stages = self.hparams.num_stages
num_layers = self.hparams.num_iaf_layers[iaf_idx]
filter_length = self.hparams.filter_length
width = self.hparams.width
out_width = self.out_width
deconv_width = self.hparams.deconv_width
deconv_config = self.hparams.deconv_config # [[l1, s1], [l2, s2]]
use_weight_norm = self.use_weight_norm
use_resize_conv = self.use_resize_conv
upsample_act = self.upsample_act
gate_width = width
final_init, final_bias = PWNHelper.manual_finit_or_not_fn(
init, iaf_idx)
mel = inputs['mel']
x = inputs['x']
iaf_name = 'iaf_{:d}'.format(iaf_idx + 1)
mel_en = wavenet._deconv_stack(mel,
deconv_width,
deconv_config,
act=upsample_act,
use_resize_conv=use_resize_conv,
name=iaf_name,
use_weight_norm=use_weight_norm,
init=init)
l = masked.shift_right(x)
l = masked.conv1d(l,
num_filters=width,
filter_length=filter_length,
name='{}/start_conv'.format(iaf_name),
use_weight_norm=use_weight_norm,
init=init)
for i in range(num_layers):
dilation = 2**(i % num_stages)
d = masked.conv1d(l,
num_filters=gate_width,
filter_length=filter_length,
dilation=dilation,
name='{}/dilated_conv_{:d}'.format(
iaf_name, i + 1),
use_weight_norm=use_weight_norm,
init=init)
c = masked.conv1d(mel_en,
num_filters=gate_width,
filter_length=1,
name='{}/mel_cond_{:d}'.format(iaf_name, i + 1),
use_weight_norm=use_weight_norm,
init=init)
d = wavenet._condition(d, c)
assert d.get_shape().as_list()[2] % 2 == 0
m = d.get_shape().as_list()[2] // 2
d_sigmoid = tf.sigmoid(d[:, :, :m])
d_tanh = tf.tanh(d[:, :, m:])
d = d_sigmoid * d_tanh
l += masked.conv1d(d,
num_filters=width,
filter_length=1,
name='{}/res_{:d}'.format(iaf_name, i + 1),
use_weight_norm=use_weight_norm,
init=init)
l = tf.nn.relu(l)
l = masked.conv1d(l,
num_filters=width,
filter_length=1,
name='{}/out1'.format(iaf_name),
use_weight_norm=use_weight_norm,
init=init)
c = masked.conv1d(mel_en,
num_filters=width,
filter_length=1,
name='{}/mel_cond_out1'.format(iaf_name),
use_weight_norm=use_weight_norm,
init=init)
l = wavenet._condition(l, c)
l = tf.nn.relu(l)
mean = masked.conv1d(l,
num_filters=out_width // 2,
filter_length=1,
name='{}/out2_mean'.format(iaf_name),
use_weight_norm=use_weight_norm,
init=final_init)
scale_params = masked.conv1d(
l,
num_filters=out_width // 2,
filter_length=1,
name='{}/out2_scale'.format(iaf_name),
use_weight_norm=use_weight_norm,
init=final_init,
biases_initializer=tf.constant_initializer(final_bias))
scale, log_scale = PWNHelper.scale_log_scale_fn(scale_params)
new_x = x * scale + mean
if DETAIL_LOG:
tf.summary.scalar('scale_{}'.format(iaf_idx),
tf.reduce_mean(scale))
tf.summary.scalar('log_scale_{}'.format(iaf_idx),
tf.reduce_mean(log_scale))
tf.summary.scalar('mean_{}'.format(iaf_idx), tf.reduce_mean(mean))
return {
'x': new_x,
'mean': mean,
'scale': scale,
'log_scale': log_scale
}
def _create_iaf(self, inputs, iaf_idx):
num_stages = self.hparams.num_stages
num_layers = self.hparams.num_iaf_layers[iaf_idx]
filter_length = self.hparams.filter_length
width = self.hparams.width
out_width = self.out_width
deconv_width = self.hparams.deconv_width
deconv_config = self.hparams.deconv_config # [[l1, s1], [l2, s2]]
use_log_scale = getattr(self.hparams, 'use_log_scale', True)
mel = inputs['mel']
x = inputs['x']
iaf_name = 'iaf_{:d}'.format(iaf_idx + 1)
mel_en = wavenet._deconv_stack(
mel, deconv_width, deconv_config, name=iaf_name)
l = masked.shift_right(x)
l = masked.conv1d(l, num_filters=width, filter_length=filter_length,
name='{}/start_conv'.format(iaf_name))
for i in range(num_layers):
dilation = 2 ** (i % num_stages)
d = masked.conv1d(
l,
num_filters=2 * width,
filter_length=filter_length,
dilation=dilation,
name='{}/dilated_conv_{:d}'.format(iaf_name, i + 1))
c = masked.conv1d(
mel_en,
num_filters=2 * width,
filter_length=1,
name='{}/mel_cond_{:d}'.format(iaf_name, i + 1))
d = wavenet._condition(d, c)
assert d.get_shape().as_list()[2] % 2 == 0
m = d.get_shape().as_list()[2] // 2
d_sigmoid = tf.sigmoid(d[:, :, :m])
d_tanh = tf.tanh(d[:, :, m:])
d = d_sigmoid * d_tanh
l += masked.conv1d(d, num_filters=width, filter_length=1,
name='{}/res_{:d}'.format(iaf_name, i + 1))
l = tf.nn.relu(l)
l = masked.conv1d(l, num_filters=width, filter_length=1,
name='{}/out1'.format(iaf_name))
c = masked.conv1d(mel_en, num_filters=width, filter_length=1,
name='{}/mel_cond_out1'.format(iaf_name))
l = wavenet._condition(l, c)
l = tf.nn.relu(l)
# to keep the scale in a reasonable small range if use_log_scale=True.
final_kernel_init = (tf.truncated_normal_initializer(0., 0.01) if use_log_scale
else tf.uniform_unit_scaling_initializer(1.0))
out = masked.conv1d(l, num_filters=out_width, filter_length=1,
name='{}/out2'.format(iaf_name),
kernel_initializer=final_kernel_init)
mean, scale_params = tf.split(out, num_or_size_splits=2, axis=2)
if use_log_scale:
log_scale = tf.clip_by_value(scale_params, -9.0, 7.0)
scale = tf.exp(log_scale)
else:
scale_params = tf.nn.softplus(scale_params)
scale = tf.clip_by_value(scale_params, tf.exp(-9.0), tf.exp(7.0))
log_scale = tf.log(scale)
new_x = x * scale + mean
if DETAIL_LOG:
tf.summary.scalar('scale_{}'.format(iaf_idx), tf.reduce_mean(scale))
tf.summary.scalar('log_scale_{}'.format(iaf_idx), tf.reduce_mean(log_scale))
tf.summary.scalar('mean_{}'.format(iaf_idx), tf.reduce_mean(mean))
return {'x': new_x,
'mean': mean,
'scale': scale,
'log_scale': log_scale}