def build_model(self):
with tf.variable_scope(
"Text2Mel"): ## keep scope names consistent with full Text2Mel
## to allow parameters to be reused more easily later
# Get S or decoder inputs. (B, T//r, n_mels). This is audio shifted 1 frame to the right.
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
## Babbler has no TextEnc
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.hp,
self.S,
training=self.training,
reuse=self.reuse)
with tf.variable_scope("Attention"):
## Babbler has no real attention. Dummy (all 0) text encoder outputs are supplied instead.
# R: concat Q with zero vector (dummy text encoder outputs)
dummy_R_prime = tf.zeros_like(
self.Q) ## R_prime shares shape of audio encoder output
self.R = tf.concat((dummy_R_prime, self.Q), -1)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.hp,
self.R,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, n_mels)
def __init__(self, num=1, mode="train"):
'''
Args:
num: Either 1 or 2. 1 for Text2Mel 2 for SSRN.
mode: Either "train" or "synthesize".
'''
# Load vocabulary
self.char2idx, self.idx2char = load_vocab()
# Set flag
training = True if mode == "train" else False
# Graph
# Data Feeding
## L: Text. (B, N), int32
## mels: Reduced melspectrogram. (B, T/r, n_mels) float32
## mags: Magnitude. (B, T, n_fft//2+1) float32
self.L = tf.placeholder(tf.int32, shape=(None, None))
self.mels = tf.placeholder(tf.float32, shape=(None, None, hp.n_mels))
self.prev_max_attentions = tf.placeholder(tf.int32, shape=(None, ))
with tf.variable_scope("Text2Mel"):
# Get S or decoder inputs. (B, T//r, n_mels)
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
# Networks
with tf.variable_scope("TextEnc"):
self.K, self.V = TextEnc(self.L,
training=training) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.S, training=training)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
self.R, self.alignments, self.max_attentions = Attention(
self.Q,
self.K,
self.V,
mononotic_attention=(not training),
prev_max_attentions=self.prev_max_attentions)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.R, training=training) # (B, T/r, n_mels)
# During inference, the predicted melspectrogram values are fed.
with tf.variable_scope("SSRN"):
self.Z_logits, self.Z = SSRN(self.Y, training=training)
with tf.variable_scope("gs"):
self.global_step = tf.Variable(0,
name='global_step',
trainable=False)
def build_model(self):
self.load_data_in_memory()
self.add_data(reuse=self.reuse)
with tf.variable_scope(
"Text2Mel"): ## keep scope names consistent with full Text2Mel
## to allow parameters to be reused more easily later
# Get S or decoder inputs. (B, T//r, n_mels). This is audio shifted 1 frame to the right.
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
# Build a latent representation of expressiveness, if we defined uee in config file (for unsupervised expressiveness embedding)
if self.hp.uee != 0:
with tf.variable_scope("Audio2Emo"):
with tf.variable_scope("AudioEnc"):
self.emos = Audio2Emo(
self.hp,
self.S,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, d=8)
self.emo_mean = tf.reduce_mean(self.emos, 1)
print(self.emo_mean.get_shape())
self.emo_mean = tf.expand_dims(self.emo_mean, axis=1)
print(self.emo_mean.get_shape())
#pdb.set_trace()
else:
print('No unsupervised expressive embedding')
self.emo_mean = None
#pdb.set_trace()
## Babbler has no TextEnc
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.hp,
self.S,
training=self.training,
reuse=self.reuse)
with tf.variable_scope("Attention"):
## Babbler has no real attention. Dummy (all 0) text encoder outputs are supplied instead.
# R: concat Q with zero vector (dummy text encoder outputs)
dummy_R_prime = tf.zeros_like(
self.Q) ## R_prime shares shape of audio encoder output
self.R = tf.concat((dummy_R_prime, self.Q), -1)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.hp,
self.R,
emos=self.emo_mean,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, n_mels)
def __init__(self, num=1):
# Load vocabulary
self.char2idx, self.idx2char = self.load_vocab()
# Set flag
training = False
# Graph
# Data Feeding
# Synthesize
self.L = tf.placeholder(tf.int32, shape=(None, None))
self.mels = tf.placeholder(tf.float32, shape=(None, None, hp.n_mels))
self.prev_max_attentions = tf.placeholder(tf.int32, shape=(None, ))
with tf.variable_scope("Text2Mel"):
# Get S or decoder inputs. (B, T//r, n_mels)
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
# Networks
with tf.variable_scope("TextEnc"):
self.K, self.V = TextEnc(self.L,
training=training) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.S, training=training)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
self.R, self.alignments, self.max_attentions = Attention(
self.Q,
self.K,
self.V,
mononotic_attention=(not training),
prev_max_attentions=self.prev_max_attentions)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.R, training=training) # (B, T/r, n_mels)
# During inference, the predicted melspectrogram values are fed.
with tf.variable_scope("SSRN"):
self.Z_logits, self.Z = SSRN(self.Y, training=training)
with tf.variable_scope("gs"):
self.global_step = tf.Variable(0,
name='global_step',
trainable=False)
def __init__(self, num=1, mode="train"):
'''
Args:
mode: Either "train" or "synthesize".
'''
# Load vocabulary
self.char2idx, self.idx2char = load_vocab()
# Set flag
training = True if mode == "train" else False
# Graph
# Data Feeding
## L: Text. (B, N), int32
## world: World Vocoder concatenate tensor.(B, 8*T/r, num_lf0+num_mgc+num_bap) float32
if mode == "train":
self.L, self.worlds, self.worlds_WSR, self.fnames, self.num_batch = get_batch(
)
self.prev_max_attentions = tf.ones(shape=(hp.B, ), dtype=tf.int32)
self.gts = tf.convert_to_tensor(guided_attention())
else: # Synthesize
self.L = tf.placeholder(tf.int32, shape=(None, None))
self.worlds = tf.placeholder(
tf.float32,
shape=(None, None,
hp.num_bap + hp.num_lf0 + hp.num_mgc + hp.num_vuv))
self.prev_max_attentions = tf.placeholder(tf.int32, shape=(None, ))
self.gts = tf.convert_to_tensor(guided_attention())
if num == 1 or (not training):
with tf.variable_scope("Text2World"):
# Get S or decoder inputs. (B, 8*T/r, num_lf0+num_mgc+num_bap)
self.S = tf.concat((tf.zeros_like(
self.worlds[:, :1, :]), self.worlds[:, :-1, :]), 1)
# Networks
with tf.variable_scope("TextEnc"):
self.K, self.V = TextEnc(self.L,
training=training) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.S, training=training)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
self.R, self.alignments, self.max_attentions = Attention(
self.Q,
self.K,
self.V,
mononotic_attention=(not training),
prev_max_attentions=self.prev_max_attentions)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.R,
training=training) # (B, T/r, num_lf0+num_mgc+num_bap)
else: # num==2 & training. Note that during training,
with tf.variable_scope("WSRN"):
self.Z_logits, self.Z = WSRN(self.worlds, training=training)
if not training:
# During inference, the predicted world values are fed.
with tf.variable_scope("WSRN"):
self.Z_logits, self.Z = WSRN(self.Y, training=training)
with tf.variable_scope("gs"):
self.global_step = tf.Variable(0,
name='global_step',
trainable=False)
if training:
if num == 1: # Text2World
# world L1 loss
self.loss_worlds = tf.losses.mean_squared_error(
self.worlds, self.Y)
#self.loss_worlds = tf.reduce_mean(tf.abs(self.Y - self.worlds))
# world binary divergence loss
#self.loss_bd1 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.Y_logits, labels=self.worlds))
# guided_attention loss
self.A = tf.pad(self.alignments, [(0, 0), (0, hp.max_N),
(0, hp.max_T)],
mode="CONSTANT",
constant_values=-1.)[:, :hp.max_N, :hp.max_T]
self.attention_masks = tf.to_float(tf.not_equal(self.A, -1))
self.loss_att = tf.reduce_sum(
tf.abs(self.A * self.gts) * self.attention_masks)
self.mask_sum = tf.reduce_sum(self.attention_masks)
self.loss_att /= self.mask_sum
# total loss
self.loss = self.loss_worlds + self.loss_att #self.loss_bd1 +
tf.summary.scalar('train/loss_worlds', self.loss_worlds)
#tf.summary.scalar('train/loss_bd1', self.loss_bd1)
tf.summary.scalar('train/loss_att', self.loss_att)
tf.summary.image(
'train/world_gt',
tf.expand_dims(tf.transpose(self.worlds[:1], [0, 2, 1]),
-1))
tf.summary.image(
'train/world_hat',
tf.expand_dims(tf.transpose(self.Y[:1], [0, 2, 1]), -1))
else: #WSRN
# world L1 loss
self.loss_WSR = tf.losses.mean_squared_error(
self.Z, self.worlds_WSR)
# world binary divergence loss
#self.loss_bd2 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.Z_logits, labels=self.worlds_WSR))
# total loss
self.loss = self.loss_WSR #+ self.loss_bd2
tf.summary.scalar('train/loss_world_SSRN', self.loss_WSR)
#tf.summary.scalar('train/loss_bd2', self.loss_bd2)
# Training Scheme
self.lr = learning_rate_decay(hp.lr, self.global_step)
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
tf.summary.scalar("lr", self.lr)
## gradient clipping
self.gvs = self.optimizer.compute_gradients(self.loss)
self.clipped = []
for grad, var in self.gvs:
grad = tf.clip_by_value(grad, -1., 1.)
self.clipped.append((grad, var))
self.train_op = self.optimizer.apply_gradients(
self.clipped, global_step=self.global_step)
# Summary
self.merged = tf.summary.merge_all()
def build_model(self):
with tf.variable_scope("Text2Mel"):
# Get S or decoder inputs. (B, T//r, n_mels). This is audio shifted 1 frame to the right.
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
# Networks
if self.hp.text_encoder_type == 'none':
assert self.hp.merlin_label_dir
self.K = self.V = self.merlin_label
elif self.hp.text_encoder_type == 'minimal_feedforward':
assert self.hp.merlin_label_dir
#sys.exit('Not implemented: hp.text_encoder_type=="minimal_feedforward"')
self.K = self.V = LinearTransformLabels(self.hp,
self.merlin_label,
training=self.training,
reuse=self.reuse)
elif self.hp.text_encoder_type == 'MerlinTextEnc':
assert self.hp.merlin_label_dir
with tf.variable_scope("MerlinTextEnc"):
self.K, self.V = MerlinTextEnc(
self.hp,
self.L,
self.merlin_label,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (N, Tx, e)
else: ## default DCTTS text encoder
with tf.variable_scope("TextEnc"):
self.K, self.V = TextEnc(self.hp,
self.L,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
if self.hp.history_type in [
'fractional_position_in_phone',
'absolute_position_in_phone'
]:
self.Q = self.position_in_phone
elif self.hp.history_type == 'minimal_history':
sys.exit(
'Not implemented: hp.history_type=="minimal_history"')
else:
assert self.hp.history_type == 'DCTTS_standard'
self.Q = AudioEnc(self.hp,
self.S,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
if self.hp.use_external_durations:
self.R, self.alignments, self.max_attentions = FixedAttention(
self.hp, self.durations, self.Q, self.V)
elif self.mode is 'synthesize':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=True,
prev_max_attentions=self.prev_max_attentions)
elif self.mode is 'train':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=False,
prev_max_attentions=self.prev_max_attentions)
elif self.mode is 'generate_attention':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=False,
prev_max_attentions=None)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.hp,
self.R,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, n_mels)
def __init__(self, num=1, mode="train"):
'''
Args:
num: Either 1 or 2. 1 for Text2Mel 2 for SSRN.
mode: Either "train" or "synthesize".
'''
# Load vocabulary
self.char2idx, self.idx2char = load_vocab()
# Set flag
training = True if mode == "train" else False
# Graph
# Data Feeding
## L: Text. (B, N), int32
## mels: Reduced melspectrogram. (B, T/r, n_mels) float32
## mags: Magnitude. (B, T, n_fft//2+1) float32
if mode == "train":
self.L, self.mels, self.mags, self.fnames, self.num_batch = get_batch(
)
self.prev_max_attentions = tf.ones(shape=(hp.B, ), dtype=tf.int32)
self.gts = tf.convert_to_tensor(guided_attention())
else: # Synthesize
self.L = tf.placeholder(tf.int32, shape=(None, None))
self.mels = tf.placeholder(tf.float32,
shape=(None, None, hp.n_mels))
self.prev_max_attentions = tf.placeholder(tf.int32, shape=(None, ))
# Training first neural net or testing
if num == 1 or (not training):
with tf.variable_scope("Text2Mel"):
# Get S or decoder inputs. (B, T//r, n_mels)
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]),
1)
# Networks
with tf.variable_scope("TextEnc"):
self.K, self.V = TextEnc(self.L,
training=training) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
self.Q = AudioEnc(self.S, training=training)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
self.R, self.alignments, self.max_attentions = Attention(
self.Q,
self.K,
self.V,
mononotic_attention=(not training),
prev_max_attentions=self.prev_max_attentions)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.R, training=training) # (B, T/r, n_mels)
# Training second neural net
else: # num==2 & training. Note that during training,
# the ground truth melspectrogram values are fed.
with tf.variable_scope("SSRN"):
self.Z_logits, self.Z = SSRN(self.mels, training=training)
if not training:
# During inference, the predicted melspectrogram values are fed.
with tf.variable_scope("SSRN"):
self.Z_logits, self.Z = SSRN(self.Y, training=training)
with tf.variable_scope("gs"):
self.global_step = tf.Variable(0,
name='global_step',
trainable=False)
if training:
if num == 1: # Text2Mel
# mel L1 loss
self.loss_mels = tf.reduce_mean(tf.abs(self.Y - self.mels))
# mel binary divergence loss
self.loss_bd1 = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=self.Y_logits, labels=self.mels))
# guided_attention loss
self.A = tf.pad(self.alignments, [(0, 0), (0, hp.max_N),
(0, hp.max_T)],
mode="CONSTANT",
constant_values=-1.)[:, :hp.max_N, :hp.max_T]
self.attention_masks = tf.to_float(tf.not_equal(self.A, -1))
self.loss_att = tf.reduce_sum(
tf.abs(self.A * self.gts) * self.attention_masks)
self.mask_sum = tf.reduce_sum(self.attention_masks)
self.loss_att /= self.mask_sum
# total loss
self.loss = self.loss_mels + self.loss_bd1 + self.loss_att
tf.summary.scalar('train/loss_mels', self.loss_mels)
tf.summary.scalar('train/loss_bd1', self.loss_bd1)
tf.summary.scalar('train/loss_att', self.loss_att)
tf.summary.image(
'train/mel_gt',
tf.expand_dims(tf.transpose(self.mels[:1], [0, 2, 1]), -1))
tf.summary.image(
'train/mel_hat',
tf.expand_dims(tf.transpose(self.Y[:1], [0, 2, 1]), -1))
else: # SSRN
# mag L1 loss
self.loss_mags = tf.reduce_mean(tf.abs(self.Z - self.mags))
# mag binary divergence loss
self.loss_bd2 = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=self.Z_logits, labels=self.mags))
# total loss
self.loss = self.loss_mags + self.loss_bd2
tf.summary.scalar('train/loss_mags', self.loss_mags)
tf.summary.scalar('train/loss_bd2', self.loss_bd2)
tf.summary.image(
'train/mag_gt',
tf.expand_dims(tf.transpose(self.mags[:1], [0, 2, 1]), -1))
tf.summary.image(
'train/mag_hat',
tf.expand_dims(tf.transpose(self.Z[:1], [0, 2, 1]), -1))
# Training Scheme
self.lr = learning_rate_decay(hp.lr, self.global_step)
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
tf.summary.scalar("lr", self.lr)
## gradient clipping
self.gvs = self.optimizer.compute_gradients(self.loss)
self.clipped = []
for grad, var in self.gvs:
grad = tf.clip_by_value(grad, -1., 1.)
self.clipped.append((grad, var))
self.train_op = self.optimizer.apply_gradients(
self.clipped, global_step=self.global_step)
# Summary
self.merged = tf.summary.merge_all()
def build_model(self):
self.load_data_in_memory()
self.add_data(reuse=self.reuse)
# Get S or decoder inputs. (B, T//r, n_mels). This is audio shifted 1 frame to the right.
self.S = tf.concat(
(tf.zeros_like(self.mels[:, :1, :]), self.mels[:, :-1, :]), 1)
# Build a latent representation of expressiveness, if we defined uee in config file (for unsupervised expressiveness embedding)
try:
if self.hp.uee != 0:
with tf.variable_scope("Audio2Emo"):
with tf.variable_scope("AudioEnc"):
self.emos = Audio2Emo(
self.hp,
self.S,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, d=8)
self.emo_mean = tf.reduce_mean(self.emos, 1)
print(self.emo_mean.get_shape())
self.emo_mean = tf.expand_dims(self.emo_mean, axis=1)
print(self.emo_mean.get_shape())
#pdb.set_trace()
except:
print('No unsupervised expressive embedding')
self.emo_mean = None
#pdb.set_trace()
with tf.variable_scope("Text2Mel"):
# Networks
if self.hp.text_encoder_type == 'none':
assert self.hp.merlin_label_dir
self.K = self.V = self.merlin_label
elif self.hp.text_encoder_type == 'minimal_feedforward':
assert self.hp.merlin_label_dir
#sys.exit('Not implemented: hp.text_encoder_type=="minimal_feedforward"')
self.K = self.V = LinearTransformLabels(self.hp,
self.merlin_label,
training=self.training,
reuse=self.reuse)
else: ## default DCTTS text encoder
with tf.variable_scope("TextEnc_emotional"):
self.K, self.V = TextEnc(self.hp,
self.L,
training=self.training,
emos=self.emo_mean,
speaker_codes=self.speakers,
reuse=self.reuse) # (N, Tx, e)
with tf.variable_scope("AudioEnc"):
if self.hp.history_type in [
'fractional_position_in_phone',
'absolute_position_in_phone'
]:
self.Q = self.position_in_phone
elif self.hp.history_type == 'minimal_history':
sys.exit(
'Not implemented: hp.history_type=="minimal_history"')
else:
assert self.hp.history_type == 'DCTTS_standard'
self.Q = AudioEnc(self.hp,
self.S,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse)
with tf.variable_scope("Attention"):
# R: (B, T/r, 2d)
# alignments: (B, N, T/r)
# max_attentions: (B,)
if self.hp.use_external_durations:
self.R, self.alignments, self.max_attentions = FixedAttention(
self.hp, self.durations, self.Q, self.V)
elif self.mode is 'synthesize':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=True,
prev_max_attentions=self.prev_max_attentions)
elif self.mode is 'train':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=False,
prev_max_attentions=self.prev_max_attentions)
elif self.mode is 'generate_attention':
self.R, self.alignments, self.max_attentions = Attention(
self.hp,
self.Q,
self.K,
self.V,
monotonic_attention=False,
prev_max_attentions=None)
with tf.variable_scope("AudioDec"):
self.Y_logits, self.Y = AudioDec(
self.hp,
self.R,
training=self.training,
speaker_codes=self.speakers,
reuse=self.reuse) # (B, T/r, n_mels)
