add TransformerTTS and fastspeech
This commit is contained in:
lifuchen
chenfeiyu
parent
185e25fedf
commit
5b2d2a37c9
|
|
@ -0,0 +1,173 @@
|
|||
import math
|
||||
from parakeet.g2p.text.symbols import symbols
|
||||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
import paddle.fluid.layers as layers
|
||||
from parakeet.modules.layers import Conv, Pool1D, Linear
|
||||
from parakeet.modules.dynamicGRU import DynamicGRU
|
||||
import numpy as np
|
||||
|
||||
class CBHG(dg.Layer):
|
||||
def __init__(self, hidden_size, batch_size, K=16, projection_size = 256, num_gru_layers=2,
|
||||
max_pool_kernel_size=2, is_post=False):
|
||||
super(CBHG, self).__init__()
|
||||
"""
|
||||
:param hidden_size: dimension of hidden unit
|
||||
:param batch_size: batch size
|
||||
:param K: # of convolution banks
|
||||
:param projection_size: dimension of projection unit
|
||||
:param num_gru_layers: # of layers of GRUcell
|
||||
:param max_pool_kernel_size: max pooling kernel size
|
||||
:param is_post: whether post processing or not
|
||||
"""
|
||||
self.hidden_size = hidden_size
|
||||
self.projection_size = projection_size
|
||||
self.conv_list = []
|
||||
self.conv_list.append(Conv(in_channels = projection_size,
|
||||
out_channels = hidden_size,
|
||||
filter_size = 1,
|
||||
padding = int(np.floor(1/2)),
|
||||
data_format = "NCT"))
|
||||
for i in range(2,K+1):
|
||||
self.conv_list.append(Conv(in_channels = hidden_size,
|
||||
out_channels = hidden_size,
|
||||
filter_size = i,
|
||||
padding = int(np.floor(i/2)),
|
||||
data_format = "NCT"))
|
||||
|
||||
for i, layer in enumerate(self.conv_list):
|
||||
self.add_sublayer("conv_list_{}".format(i), layer)
|
||||
|
||||
self.batchnorm_list = []
|
||||
for i in range(K):
|
||||
self.batchnorm_list.append(dg.BatchNorm(hidden_size,
|
||||
data_layout='NCHW'))
|
||||
|
||||
for i, layer in enumerate(self.batchnorm_list):
|
||||
self.add_sublayer("batchnorm_list_{}".format(i), layer)
|
||||
|
||||
conv_outdim = hidden_size * K
|
||||
|
||||
self.conv_projection_1 = Conv(in_channels = conv_outdim,
|
||||
out_channels = hidden_size,
|
||||
filter_size = 3,
|
||||
padding = int(np.floor(3/2)),
|
||||
data_format = "NCT")
|
||||
|
||||
self.conv_projection_2 = Conv(in_channels = hidden_size,
|
||||
out_channels = projection_size,
|
||||
filter_size = 3,
|
||||
padding = int(np.floor(3/2)),
|
||||
data_format = "NCT")
|
||||
|
||||
self.batchnorm_proj_1 = dg.BatchNorm(hidden_size,
|
||||
data_layout='NCHW')
|
||||
self.batchnorm_proj_2 = dg.BatchNorm(projection_size,
|
||||
data_layout='NCHW')
|
||||
self.max_pool = Pool1D(pool_size = max_pool_kernel_size,
|
||||
pool_type='max',
|
||||
pool_stride=1,
|
||||
pool_padding=1,
|
||||
data_format = "NCT")
|
||||
self.highway = Highwaynet(self.projection_size)
|
||||
|
||||
h_0 = np.zeros((batch_size, hidden_size // 2), dtype="float32")
|
||||
h_0 = dg.to_variable(h_0)
|
||||
self.fc_forward1 = Linear(hidden_size, hidden_size // 2 * 3)
|
||||
self.fc_reverse1 = Linear(hidden_size, hidden_size // 2 * 3)
|
||||
self.gru_forward1 = DynamicGRU(size = self.hidden_size // 2,
|
||||
is_reverse = False,
|
||||
origin_mode = True,
|
||||
h_0 = h_0)
|
||||
self.gru_reverse1 = DynamicGRU(size = self.hidden_size // 2,
|
||||
is_reverse=True,
|
||||
origin_mode=True,
|
||||
h_0 = h_0)
|
||||
|
||||
self.fc_forward2 = Linear(hidden_size, hidden_size // 2 * 3)
|
||||
self.fc_reverse2 = Linear(hidden_size, hidden_size // 2 * 3)
|
||||
self.gru_forward2 = DynamicGRU(size = self.hidden_size // 2,
|
||||
is_reverse = False,
|
||||
origin_mode = True,
|
||||
h_0 = h_0)
|
||||
self.gru_reverse2 = DynamicGRU(size = self.hidden_size // 2,
|
||||
is_reverse=True,
|
||||
origin_mode=True,
|
||||
h_0 = h_0)
|
||||
|
||||
def _conv_fit_dim(self, x, filter_size=3):
|
||||
if filter_size % 2 == 0:
|
||||
return x[:,:,:-1]
|
||||
else:
|
||||
return x
|
||||
|
||||
def forward(self, input_):
|
||||
# input_.shape = [N, C, T]
|
||||
|
||||
conv_list = []
|
||||
conv_input = input_
|
||||
|
||||
for i, (conv, batchnorm) in enumerate(zip(self.conv_list, self.batchnorm_list)):
|
||||
conv_input = self._conv_fit_dim(conv(conv_input), i+1)
|
||||
conv_input = layers.relu(batchnorm(conv_input))
|
||||
conv_list.append(conv_input)
|
||||
|
||||
conv_cat = layers.concat(conv_list, axis=1)
|
||||
conv_pool = self.max_pool(conv_cat)[:,:,:-1]
|
||||
|
||||
|
||||
conv_proj = layers.relu(self.batchnorm_proj_1(self._conv_fit_dim(self.conv_projection_1(conv_pool))))
|
||||
conv_proj = self.batchnorm_proj_2(self._conv_fit_dim(self.conv_projection_2(conv_proj))) + input_
|
||||
|
||||
# conv_proj.shape = [N, C, T]
|
||||
highway = layers.transpose(conv_proj, [0,2,1])
|
||||
highway = self.highway(highway)
|
||||
|
||||
# highway.shape = [N, T, C]
|
||||
fc_forward = self.fc_forward1(highway)
|
||||
fc_reverse = self.fc_reverse1(highway)
|
||||
out_forward = self.gru_forward1(fc_forward)
|
||||
out_reverse = self.gru_reverse1(fc_reverse)
|
||||
out = layers.concat([out_forward, out_reverse], axis=-1)
|
||||
fc_forward = self.fc_forward2(out)
|
||||
fc_reverse = self.fc_reverse2(out)
|
||||
out_forward = self.gru_forward2(fc_forward)
|
||||
out_reverse = self.gru_reverse2(fc_reverse)
|
||||
out = layers.concat([out_forward, out_reverse], axis=-1)
|
||||
out = layers.transpose(out, [0,2,1])
|
||||
return out
|
||||
|
||||
class Highwaynet(dg.Layer):
|
||||
def __init__(self, num_units, num_layers=4):
|
||||
super(Highwaynet, self).__init__()
|
||||
self.num_units = num_units
|
||||
self.num_layers = num_layers
|
||||
|
||||
self.gates = []
|
||||
self.linears = []
|
||||
|
||||
for i in range(num_layers):
|
||||
self.linears.append(Linear(num_units, num_units))
|
||||
self.gates.append(Linear(num_units, num_units))
|
||||
|
||||
for i, (linear, gate) in enumerate(zip(self.linears,self.gates)):
|
||||
self.add_sublayer("linears_{}".format(i), linear)
|
||||
self.add_sublayer("gates_{}".format(i), gate)
|
||||
|
||||
def forward(self, input_):
|
||||
out = input_
|
||||
|
||||
for linear, gate in zip(self.linears, self.gates):
|
||||
h = fluid.layers.relu(linear(out))
|
||||
t_ = fluid.layers.sigmoid(gate(out))
|
||||
|
||||
c = 1 - t_
|
||||
out = h * t_ + out * c
|
||||
|
||||
return out
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,98 @@
|
|||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
from parakeet.modules.layers import Conv1D, Linear
|
||||
from parakeet.modules.utils import *
|
||||
from parakeet.modules.multihead_attention import MultiheadAttention
|
||||
from parakeet.modules.feed_forward import PositionwiseFeedForward
|
||||
from parakeet.modules.prenet import PreNet
|
||||
from parakeet.modules.post_convnet import PostConvNet
|
||||
|
||||
class Decoder(dg.Layer):
|
||||
def __init__(self, num_hidden, config, num_head=4):
|
||||
super(Decoder, self).__init__()
|
||||
self.num_hidden = num_hidden
|
||||
param = fluid.ParamAttr()
|
||||
self.alpha = self.create_parameter(shape=(1,), attr=param, dtype='float32',
|
||||
default_initializer = fluid.initializer.ConstantInitializer(value=1.0))
|
||||
self.pos_inp = get_sinusoid_encoding_table(1024, self.num_hidden, padding_idx=0)
|
||||
self.pos_emb = dg.Embedding(size=[1024, num_hidden],
|
||||
padding_idx=0,
|
||||
param_attr=fluid.ParamAttr(
|
||||
initializer=fluid.initializer.NumpyArrayInitializer(self.pos_inp),
|
||||
trainable=False))
|
||||
self.decoder_prenet = PreNet(input_size = config.audio.num_mels,
|
||||
hidden_size = num_hidden * 2,
|
||||
output_size = num_hidden,
|
||||
dropout_rate=0.2)
|
||||
self.linear = Linear(num_hidden, num_hidden)
|
||||
|
||||
self.selfattn_layers = [MultiheadAttention(num_hidden, num_hidden//num_head, num_hidden//num_head) for _ in range(3)]
|
||||
for i, layer in enumerate(self.selfattn_layers):
|
||||
self.add_sublayer("self_attn_{}".format(i), layer)
|
||||
self.attn_layers = [MultiheadAttention(num_hidden, num_hidden//num_head, num_hidden//num_head) for _ in range(3)]
|
||||
for i, layer in enumerate(self.attn_layers):
|
||||
self.add_sublayer("attn_{}".format(i), layer)
|
||||
self.ffns = [PositionwiseFeedForward(num_hidden, num_hidden*num_head, filter_size=1) for _ in range(3)]
|
||||
for i, layer in enumerate(self.ffns):
|
||||
self.add_sublayer("ffns_{}".format(i), layer)
|
||||
self.mel_linear = Linear(num_hidden, config.audio.num_mels * config.audio.outputs_per_step)
|
||||
self.stop_linear = Linear(num_hidden, 1)
|
||||
|
||||
self.postconvnet = PostConvNet(config.audio.num_mels, config.hidden_size,
|
||||
filter_size = 5, padding = 4, num_conv=5,
|
||||
outputs_per_step=config.audio.outputs_per_step,
|
||||
use_cudnn = config.use_gpu)
|
||||
|
||||
def forward(self, key, value, query, c_mask, positional):
|
||||
|
||||
# get decoder mask with triangular matrix
|
||||
|
||||
if fluid.framework._dygraph_tracer()._train_mode:
|
||||
m_mask = get_non_pad_mask(positional)
|
||||
mask = get_attn_key_pad_mask((positional==0).astype(np.float32), query)
|
||||
triu_tensor = dg.to_variable(get_triu_tensor(query.numpy(), query.numpy())).astype(np.float32)
|
||||
mask = mask + triu_tensor
|
||||
mask = fluid.layers.cast(mask == 0, np.float32)
|
||||
|
||||
# (batch_size, decoder_len, encoder_len)
|
||||
zero_mask = get_attn_key_pad_mask(layers.squeeze(c_mask,[-1]), query)
|
||||
else:
|
||||
mask = get_triu_tensor(query.numpy(), query.numpy()).astype(np.float32)
|
||||
mask = fluid.layers.cast(dg.to_variable(mask == 0), np.float32)
|
||||
m_mask, zero_mask = None, None
|
||||
|
||||
# Decoder pre-network
|
||||
query = self.decoder_prenet(query)
|
||||
|
||||
# Centered position
|
||||
query = self.linear(query)
|
||||
|
||||
# Get position embedding
|
||||
positional = self.pos_emb(positional)
|
||||
query = positional * self.alpha + query
|
||||
|
||||
#positional dropout
|
||||
query = fluid.layers.dropout(query, 0.1)
|
||||
|
||||
# Attention decoder-decoder, encoder-decoder
|
||||
selfattn_list = list()
|
||||
attn_list = list()
|
||||
|
||||
for selfattn, attn, ffn in zip(self.selfattn_layers, self.attn_layers, self.ffns):
|
||||
query, attn_dec = selfattn(query, query, query, mask = mask, query_mask = m_mask)
|
||||
query, attn_dot = attn(key, value, query, mask = zero_mask, query_mask = m_mask)
|
||||
query = ffn(query)
|
||||
selfattn_list.append(attn_dec)
|
||||
attn_list.append(attn_dot)
|
||||
# Mel linear projection
|
||||
mel_out = self.mel_linear(query)
|
||||
# Post Mel Network
|
||||
out = self.postconvnet(mel_out)
|
||||
out = mel_out + out
|
||||
|
||||
# Stop tokens
|
||||
stop_tokens = self.stop_linear(query)
|
||||
stop_tokens = layers.squeeze(stop_tokens, [-1])
|
||||
stop_tokens = layers.sigmoid(stop_tokens)
|
||||
|
||||
return mel_out, out, attn_list, stop_tokens, selfattn_list
|
||||
|
|
@ -0,0 +1,58 @@
|
|||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
from parakeet.modules.layers import Conv1D, Linear
|
||||
from parakeet.modules.utils import *
|
||||
from parakeet.modules.multihead_attention import MultiheadAttention
|
||||
from parakeet.modules.feed_forward import PositionwiseFeedForward
|
||||
from parakeet.models.transformerTTS.encoderprenet import EncoderPrenet
|
||||
|
||||
class Encoder(dg.Layer):
|
||||
def __init__(self, embedding_size, num_hidden, config, num_head=4):
|
||||
super(Encoder, self).__init__()
|
||||
self.num_hidden = num_hidden
|
||||
param = fluid.ParamAttr(initializer=fluid.initializer.Constant(value=1.0))
|
||||
self.alpha = self.create_parameter(shape=(1, ), attr=param, dtype='float32')
|
||||
self.pos_inp = get_sinusoid_encoding_table(1024, self.num_hidden, padding_idx=0)
|
||||
self.pos_emb = dg.Embedding(size=[1024, num_hidden],
|
||||
padding_idx=0,
|
||||
param_attr=fluid.ParamAttr(
|
||||
initializer=fluid.initializer.NumpyArrayInitializer(self.pos_inp),
|
||||
trainable=False))
|
||||
self.encoder_prenet = EncoderPrenet(embedding_size = embedding_size,
|
||||
num_hidden = num_hidden,
|
||||
use_cudnn=config.use_gpu)
|
||||
self.layers = [MultiheadAttention(num_hidden, num_hidden//num_head, num_hidden//num_head) for _ in range(3)]
|
||||
for i, layer in enumerate(self.layers):
|
||||
self.add_sublayer("self_attn_{}".format(i), layer)
|
||||
self.ffns = [PositionwiseFeedForward(num_hidden, num_hidden*num_head, filter_size=1, use_cudnn = config.use_gpu) for _ in range(3)]
|
||||
for i, layer in enumerate(self.ffns):
|
||||
self.add_sublayer("ffns_{}".format(i), layer)
|
||||
|
||||
def forward(self, x, positional):
|
||||
if fluid.framework._dygraph_tracer()._train_mode:
|
||||
query_mask = get_non_pad_mask(positional)
|
||||
mask = get_attn_key_pad_mask(positional, x)
|
||||
else:
|
||||
query_mask, mask = None, None
|
||||
|
||||
# Encoder pre_network
|
||||
x = self.encoder_prenet(x) #(N,T,C)
|
||||
|
||||
|
||||
# Get positional encoding
|
||||
positional = self.pos_emb(positional)
|
||||
|
||||
x = positional * self.alpha + x #(N, T, C)
|
||||
|
||||
|
||||
# Positional dropout
|
||||
x = layers.dropout(x, 0.1)
|
||||
|
||||
# Self attention encoder
|
||||
attentions = list()
|
||||
for layer, ffn in zip(self.layers, self.ffns):
|
||||
x, attention = layer(x, x, x, mask = mask, query_mask = query_mask)
|
||||
x = ffn(x)
|
||||
attentions.append(attention)
|
||||
|
||||
return x, query_mask, attentions
|
||||
|
|
@ -0,0 +1,52 @@
|
|||
import math
|
||||
from parakeet.g2p.text.symbols import symbols
|
||||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
import paddle.fluid.layers as layers
|
||||
from parakeet.modules.layers import Conv, Linear
|
||||
import numpy as np
|
||||
|
||||
|
||||
class EncoderPrenet(dg.Layer):
|
||||
def __init__(self, embedding_size, num_hidden, use_cudnn=True):
|
||||
super(EncoderPrenet, self).__init__()
|
||||
self.embedding_size = embedding_size
|
||||
self.num_hidden = num_hidden
|
||||
self.use_cudnn = use_cudnn
|
||||
self.embedding = dg.Embedding( size = [len(symbols), embedding_size],
|
||||
padding_idx = None)
|
||||
self.conv_list = []
|
||||
self.conv_list.append(Conv(in_channels = embedding_size,
|
||||
out_channels = num_hidden,
|
||||
filter_size = 5,
|
||||
padding = int(np.floor(5/2)),
|
||||
use_cudnn = use_cudnn,
|
||||
data_format = "NCT"))
|
||||
for _ in range(2):
|
||||
self.conv_list.append(Conv(in_channels = num_hidden,
|
||||
out_channels = num_hidden,
|
||||
filter_size = 5,
|
||||
padding = int(np.floor(5/2)),
|
||||
use_cudnn = use_cudnn,
|
||||
data_format = "NCT"))
|
||||
|
||||
for i, layer in enumerate(self.conv_list):
|
||||
self.add_sublayer("conv_list_{}".format(i), layer)
|
||||
|
||||
self.batch_norm_list = [dg.BatchNorm(num_hidden,
|
||||
data_layout='NCHW') for _ in range(3)]
|
||||
|
||||
for i, layer in enumerate(self.batch_norm_list):
|
||||
self.add_sublayer("batch_norm_list_{}".format(i), layer)
|
||||
|
||||
self.projection = Linear(num_hidden, num_hidden)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.embedding(x) #(batch_size, seq_len, embending_size)
|
||||
x = layers.transpose(x,[0,2,1])
|
||||
for batch_norm, conv in zip(self.batch_norm_list, self.conv_list):
|
||||
x = layers.dropout(layers.relu(batch_norm(conv(x))), 0.2)
|
||||
x = layers.transpose(x,[0,2,1]) #(N,T,C)
|
||||
x = self.projection(x)
|
||||
|
||||
return x
|
||||
|
|
@ -0,0 +1,33 @@
|
|||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
from parakeet.models.transformerTTS.encoder import Encoder
|
||||
from parakeet.models.transformerTTS.decoder import Decoder
|
||||
|
||||
class TransformerTTS(dg.Layer):
|
||||
def __init__(self, config):
|
||||
super(TransformerTTS, self).__init__()
|
||||
self.encoder = Encoder(config.embedding_size, config.hidden_size, config)
|
||||
self.decoder = Decoder(config.hidden_size, config)
|
||||
self.config = config
|
||||
|
||||
def forward(self, characters, mel_input, pos_text, pos_mel):
|
||||
# key (batch_size, seq_len, channel)
|
||||
# c_mask (batch_size, seq_len)
|
||||
# attns_enc (channel / 2, seq_len, seq_len)
|
||||
|
||||
key, c_mask, attns_enc = self.encoder(characters, pos_text)
|
||||
|
||||
# mel_output/postnet_output (batch_size, mel_len, n_mel)
|
||||
# attn_probs (128, mel_len, seq_len)
|
||||
# stop_preds (batch_size, mel_len, 1)
|
||||
# attns_dec (128, mel_len, mel_len)
|
||||
mel_output, postnet_output, attn_probs, stop_preds, attns_dec = self.decoder(key, key, mel_input, c_mask, pos_mel)
|
||||
|
||||
return mel_output, postnet_output, attn_probs, stop_preds, attns_enc, attns_dec
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
|
@ -0,0 +1,29 @@
|
|||
import paddle.fluid.dygraph as dg
|
||||
import paddle.fluid as fluid
|
||||
from parakeet.modules.layers import Conv1D, Linear
|
||||
from parakeet.modules.utils import *
|
||||
from parakeet.models.transformerTTS.CBHG import CBHG
|
||||
|
||||
class Vocoder(dg.Layer):
|
||||
"""
|
||||
CBHG Network (mel -> linear)
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(Vocoder, self).__init__()
|
||||
self.pre_proj = Conv1D(in_channels = config.audio.num_mels,
|
||||
out_channels = config.hidden_size,
|
||||
filter_size=1,
|
||||
data_format = "NCT")
|
||||
self.cbhg = CBHG(config.hidden_size, config.batch_size)
|
||||
self.post_proj = Conv1D(in_channels = config.hidden_size,
|
||||
out_channels = (config.audio.n_fft // 2) + 1,
|
||||
filter_size=1,
|
||||
data_format = "NCT")
|
||||
|
||||
def forward(self, mel):
|
||||
mel = layers.transpose(mel, [0,2,1])
|
||||
mel = self.pre_proj(mel)
|
||||
mel = self.cbhg(mel)
|
||||
mag_pred = self.post_proj(mel)
|
||||
mag_pred = layers.transpose(mag_pred, [0,2,1])
|
||||
return mag_pred
|
||||
Loading…
Reference in New Issue