def build(self, inp, inp_mask):
batch_size = tf.shape(inp)[0]
input_time_steps = tf.shape(inp)[1]
### Encoder [ begin ]
with tf.variable_scope("encoder"):
with tf.variable_scope("embedding"):
embed_inp = EmbeddingLayer(EMBED_CLASS, EMBED_DIM)(inp)
with tf.variable_scope("changeToVarible"):
self.single_style_token = tf.get_variable(
'style_token', (1, styles_kind, style_dim),
dtype=tf.float32)
self.style_token = tf.tile(self.single_style_token,
(batch_size, 1, 1))
with tf.variable_scope("pre-net"):
pre_ed_inp = tf.layers.dropout(tf.layers.dense(
embed_inp, 256, tf.nn.relu),
training=self.training)
pre_ed_inp = tf.layers.dropout(tf.layers.dense(
pre_ed_inp, 128, tf.nn.relu),
training=self.training)
with tf.variable_scope("CBHG"):
# batch major
encoder_output = CBHG(16,
(128, 128))(pre_ed_inp,
sequence_length=inp_mask,
is_training=self.training,
time_major=False)
with tf.variable_scope("attention"):
att_module = AttentionModule(ATT_RNN_SIZE,
encoder_output,
sequence_length=inp_mask,
time_major=False)
with tf.variable_scope("attention_style"):
att_module_style = AttentionModule(STYLE_ATT_RNN_SIZE,
self.style_token,
time_major=False)
with tf.variable_scope("decoder"):
with tf.variable_scope("attentionRnn"):
att_cell = GRUCell(ATT_RNN_SIZE)
with tf.variable_scope("acoustic_module"):
aco_cell = MultiRNNCell(
[ResidualWrapper(GRUCell(DEC_RNN_SIZE)) for _ in range(2)])
### prepare output alpha TensorArray
reduced_time_steps = tf.div(MAX_OUT_STEPS, self.r)
att_cell_state = att_cell.init_state(batch_size, tf.float32)
aco_cell_state = aco_cell.zero_state(batch_size, tf.float32)
state_tup = tuple([att_cell_state, aco_cell_state])
output_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
alpha_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
weight_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
alpha_style_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
init_indic = tf.zeros([batch_size, OUTPUT_MEL_DIM])
# init_context = tf.zeros((batch_size, 256))
time = tf.constant(0, dtype=tf.int32)
cond = lambda time, *_: tf.less(time, reduced_time_steps)
def body(time, indic, output_ta, alpha_ta, alpha_style_ta,
weight_ta, state_tup):
with tf.variable_scope("att-rnn"):
pre_ed_indic = tf.layers.dropout(tf.layers.dense(
indic, 256, tf.nn.relu),
training=self.training)
pre_ed_indic = tf.layers.dropout(tf.layers.dense(
pre_ed_indic, 128, tf.nn.relu),
training=self.training)
att_cell_out, att_cell_state = att_cell(
tf.concat([pre_ed_indic], axis=-1), state_tup[0])
with tf.variable_scope("attention"):
query = att_cell_state[0] # att_cell_out
context, alpha = att_module(query)
alpha_ta = alpha_ta.write(time, alpha)
with tf.variable_scope("attention_style"):
context_style, alpha_style = att_module_style(query)
print('context_style:', context_style)
# print('context_style22:', alpha_style)
alpha_style_ta = alpha_style_ta.write(time, alpha_style)
with tf.variable_scope("weighting"):
weighting = add_layer(query,
query.shape[-1],
1,
'weighting_w',
'weighting_b',
activation_function=tf.nn.sigmoid)
# weighting = tf.nn.softmax(weighting)
weight_ta = weight_ta.write(time, weighting)
with tf.variable_scope("acoustic_module"):
# weighting0 = tf.reshape(weighting[:, 0], (BATCH_SIZE, 1))
# weighting1 = tf.reshape(weighting[:, 1], (BATCH_SIZE, 1))
# weighting_context = weighting0 * context + weighting1 * context_style
# print('context:', context)
weighting = tf.Print(weighting, [weighting],
message='weight',
summarize=100)
context_style = tf.Print(context_style,
[context_style[0][0:5]],
message='origal_style',
summarize=100)
context_style = tf.Print(
context_style, [tf.nn.tanh(context_style)[0][0:5]],
message='tanh_style',
summarize=100)
context = tf.Print(context, [context[0][0:5]],
message='context',
summarize=100)
weighting_context = context + weighting * tf.nn.tanh(
context_style)
aco_input = tf.layers.dense(
tf.concat([att_cell_out, weighting_context], axis=-1),
DEC_RNN_SIZE)
aco_cell_out, aco_cell_state = aco_cell(
aco_input, state_tup[1])
dense_out = tf.reshape(
tf.layers.dense(aco_cell_out, OUTPUT_MEL_DIM * self.r),
shape=(batch_size, self.r, OUTPUT_MEL_DIM))
output_ta = output_ta.write(time, dense_out)
new_indic = dense_out[:, -1]
state_tup = tuple([att_cell_state, aco_cell_state])
return tf.add(
time, 1
), new_indic, output_ta, alpha_ta, alpha_style_ta, weight_ta, state_tup
### run loop
_, _, output_mel_ta, final_alpha_ta, final_alpha_style_ta, final_weight_ta, *_ = tf.while_loop(
cond, body, [
time, init_indic, output_ta, alpha_ta, alpha_style_ta,
weight_ta, state_tup
])
### time major
with tf.variable_scope("output"):
output_mel = tf.reshape(output_mel_ta.stack(),
shape=(reduced_time_steps, batch_size,
OUTPUT_MEL_DIM * self.r))
output_mel = tf.reshape(tf.transpose(output_mel, perm=(1, 0, 2)),
shape=(batch_size, MAX_OUT_STEPS,
OUTPUT_MEL_DIM))
self.out_mel = output_mel
with tf.variable_scope("post-net"):
output_post = CBHG(8, (256, OUTPUT_MEL_DIM))(
output_mel,
sequence_length=None,
is_training=self.training,
time_major=False)
output_spec = tf.layers.dense(output_post, OUTPUT_SPEC_DIM)
self.out_stftm = output_spec
final_alpha = tf.reshape(final_alpha_ta.stack(),
shape=(reduced_time_steps, batch_size,
input_time_steps))
self.final_alpha = tf.transpose(final_alpha,
perm=(1, 0, 2)) # batch major
final_alpha_style = tf.reshape(final_alpha_style_ta.stack(),
shape=(reduced_time_steps,
batch_size, styles_kind))
self.final_alpha_style = tf.transpose(final_alpha_style,
perm=(1, 0,
2)) # batch major
final_weight_ta = tf.reshape(final_weight_ta.stack(),
shape=(reduced_time_steps, batch_size,
1))
self.final_weight_ta = tf.transpose(final_weight_ta,
perm=(1, 0, 2)) # batch major
def build(self, inp, inp_mask, mel_gtruth, spec_gtruth):
batch_size = tf.shape(inp)[0]
input_time_steps = tf.shape(inp)[1]
output_time_steps = tf.shape(mel_gtruth)[1]
### Encoder [ begin
with tf.variable_scope("encoder"):
with tf.variable_scope("embedding"):
embed_inp = EmbeddingLayer(EMBED_CLASS, EMBED_DIM)(inp)
with tf.variable_scope("changeToVarible"):
self.single_style_token = tf.get_variable(
'style_token', (1, styles_kind, style_dim),
dtype=tf.float32)
self.style_token = tf.tile(self.single_style_token,
(batch_size, 1, 1))
with tf.variable_scope("pre-net"):
pre_ed_inp = tf.layers.dropout(tf.layers.dense(
embed_inp, 256, tf.nn.relu),
training=self.training)
pre_ed_inp = tf.layers.dropout(tf.layers.dense(
pre_ed_inp, 128, tf.nn.relu),
training=self.training)
with tf.variable_scope("CBHG"):
# batch major
encoder_output = CBHG(16,
(128, 128))(pre_ed_inp,
sequence_length=inp_mask,
is_training=self.training,
time_major=False)
with tf.variable_scope("attention"):
att_module = AttentionModule(ATT_RNN_SIZE,
encoder_output,
sequence_length=inp_mask,
time_major=False)
with tf.variable_scope("attention_style"):
att_module_style = AttentionModule(STYLE_ATT_RNN_SIZE,
self.style_token,
time_major=False)
with tf.variable_scope("decoder"):
with tf.variable_scope("attentionRnn"):
att_cell = GRUCell(ATT_RNN_SIZE)
with tf.variable_scope("acoustic_module"):
aco_cell = MultiRNNCell(
[ResidualWrapper(GRUCell(DEC_RNN_SIZE)) for _ in range(2)])
### prepare output alpha TensorArray
reduced_time_steps = tf.div(output_time_steps, self.r)
att_cell_state = att_cell.init_state(batch_size, tf.float32)
aco_cell_state = aco_cell.zero_state(batch_size, tf.float32)
state_tup = tuple([att_cell_state, aco_cell_state])
output_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
alpha_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
weight_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
alpha_style_ta = tf.TensorArray(size=reduced_time_steps,
dtype=tf.float32)
indic_ta = tf.TensorArray(size=self.r + output_time_steps,
dtype=tf.float32)
time_major_mel_gtruth = tf.transpose(mel_gtruth, perm=(1, 0, 2))
indic_array = tf.concat([
tf.zeros([self.r, batch_size, OUTPUT_MEL_DIM]),
time_major_mel_gtruth
],
axis=0)
indic_ta = indic_ta.unstack(indic_array)
#init_context = tf.zeros((batch_size, 256))
time = tf.constant(0, dtype=tf.int32)
cond = lambda time, *_: tf.less(time, reduced_time_steps)
def body(time, output_ta, alpha_ta, alpha_style_ta, weight_ta,
state_tup):
with tf.variable_scope("att-rnn"):
pre_ed_indic = tf.layers.dropout(tf.layers.dense(
indic_ta.read(self.r * time + self.r - 1), 256,
tf.nn.relu),
training=self.training)
pre_ed_indic = tf.layers.dropout(tf.layers.dense(
pre_ed_indic, 128, tf.nn.relu),
training=self.training)
att_cell_out, att_cell_state = att_cell(
tf.concat([pre_ed_indic], axis=-1), state_tup[0])
with tf.variable_scope("attention"):
query = att_cell_state[0] # att_cell_out
context, alpha = att_module(query)
alpha_ta = alpha_ta.write(time, alpha)
with tf.variable_scope("attention_style"):
context_style, alpha_style = att_module_style(query)
alpha_style_ta = alpha_style_ta.write(time, alpha_style)
with tf.variable_scope("weighting"):
print(query)
weighting = add_layer(query,
query.shape[-1],
1,
'weighting_w',
'weighting_b',
activation_function=tf.nn.sigmoid)
# weighting = tf.nn.softmax(weighting)
weight_ta = weight_ta.write(time, weighting)
with tf.variable_scope("acoustic_module"):
# weighting0 = tf.reshape(weighting[:, 0], (BATCH_SIZE, 1))
# weighting1 = tf.reshape(weighting[:, 1], (BATCH_SIZE, 1))
weighting_context = context + weighting * tf.nn.tanh(
context_style)
# print(weighting_context)
aco_input = tf.layers.dense(
tf.concat([att_cell_out, weighting_context], axis=-1),
DEC_RNN_SIZE)
aco_cell_out, aco_cell_state = aco_cell(
aco_input, state_tup[1])
dense_out = tf.layers.dense(aco_cell_out,
OUTPUT_MEL_DIM * self.r)
output_ta = output_ta.write(time, dense_out)
state_tup = tuple([att_cell_state, aco_cell_state])
return tf.add(
time, 1
), output_ta, alpha_ta, alpha_style_ta, weight_ta, state_tup
### run loop
_, output_mel_ta, final_alpha_ta, final_alpha_style_ta, final_weight_ta, *_ = tf.while_loop(
cond, body, [
time, output_ta, alpha_ta, alpha_style_ta, weight_ta,
state_tup
])
# print('hjhhhh', reduced_time_steps, batch_size, OUTPUT_MEL_DIM * self.r, batch_size, output_time_steps,
# OUTPUT_MEL_DIM)
# sys.stdout.flush()
### time major
with tf.variable_scope("output"):
# print('hjhhhh', reduced_time_steps, batch_size, OUTPUT_MEL_DIM * self.r, batch_size, output_time_steps, OUTPUT_MEL_DIM)
# sys.stdout.flush()
output_mel = tf.reshape(output_mel_ta.stack(),
shape=(reduced_time_steps, batch_size,
OUTPUT_MEL_DIM * self.r))
output_mel = tf.reshape(tf.transpose(output_mel, perm=(1, 0, 2)),
shape=(batch_size, output_time_steps,
OUTPUT_MEL_DIM))
self.out_mel = output_mel
with tf.variable_scope("post-net"):
output_post = CBHG(8, (256, OUTPUT_MEL_DIM))(
output_mel,
sequence_length=None,
is_training=self.training,
time_major=False)
output_spec = tf.layers.dense(output_post, OUTPUT_SPEC_DIM)
self.out_stftm = output_spec
final_alpha = tf.reshape(final_alpha_ta.stack(),
shape=(reduced_time_steps, batch_size,
input_time_steps))
final_alpha = tf.transpose(final_alpha,
perm=(1, 0, 2)) # batch major
final_alpha_style = tf.reshape(final_alpha_style_ta.stack(),
shape=(reduced_time_steps,
batch_size, styles_kind))
final_alpha_style = tf.transpose(final_alpha_style,
perm=(1, 0, 2)) # batch major
final_weight_ta = tf.reshape(final_weight_ta.stack(),
shape=(reduced_time_steps, batch_size,
1))
final_weight_ta = tf.transpose(final_weight_ta,
perm=(1, 0, 2)) # batch major
self.weighting = final_weight_ta
# self.alpha_style_hjk_img = tf.reshape(final_alpha_style, shape=(batch_size, reduced_time_steps, styles_kind))
with tf.variable_scope("loss_and_metric"):
self.loss_mel = tf.reduce_mean(tf.abs(mel_gtruth - output_mel))
self.loss_spec = tf.reduce_mean(tf.abs(spec_gtruth - output_spec))
self.loss = self.loss_mel + self.loss_spec
self.alpha_img = tf.expand_dims(final_alpha, -1)
self.alpha_style_img = tf.expand_dims(final_alpha_style, -1)
self.weight_img = tf.expand_dims(final_weight_ta, -1)
self.sums = []
self.sums.append(
tf.summary.image("train/alpha", self.alpha_img[:2]))
self.sums.append(
tf.summary.image("train/alpha_style",
self.alpha_style_img[:2]))
self.sums.append(
tf.summary.image("train/weight", self.weight_img[:2]))
self.sums.append(tf.summary.scalar("train/loss", self.loss))
self.sums.append(
tf.summary.scalar("train/style_0_0",
self.single_style_token[0][0][0]))
self.sums.append(
tf.summary.scalar("train/style_0_100",
self.single_style_token[0][0][100]))
self.sums.append(
tf.summary.scalar("train/style_5_100",
self.single_style_token[0][5][100]))
self.sums.append(
tf.summary.histogram("train/style_vec",
self.single_style_token))
self.pred_audio_holder = tf.placeholder(shape=(None, None),
dtype=tf.float32,
name='pred_audio')
self.pred_audio_summary = tf.summary.audio('pred_audio_summary',
self.pred_audio_holder,
sample_rate=sr,
max_outputs=12)
