TTS-HierSpeech_TTS / ttv_v1 /t2w2v_transformer.py
Sang-Hoon Lee
Upload 70 files
0164e4a
import torch
from torch import nn
from torch.nn import functional as F
from ttv_v1 import modules
import attentions
from torch.nn import Conv1d, ConvTranspose1d
from torch.nn.utils import weight_norm, remove_weight_norm
from commons import init_weights
import typing as tp
import transformers
import math
from ttv_v1.styleencoder import StyleEncoder
import commons
from ttv_v1.modules import WN
def get_2d_padding(kernel_size: tp.Tuple[int, int], dilation: tp.Tuple[int, int] = (1, 1)):
return (((kernel_size[0] - 1) * dilation[0]) // 2, ((kernel_size[1] - 1) * dilation[1]) // 2)
class Wav2vec2(torch.nn.Module):
def __init__(self, layer=7):
"""we use the intermediate features of xls-r-300m.
More specifically, we used the output from the 12th layer of the 24-layer transformer encoder.
"""
super().__init__()
# self.wav2vec2 = transformers.Wav2Vec2ForPreTraining.from_pretrained("facebook/wav2vec2-xls-r-300m")
self.wav2vec2 = transformers.Wav2Vec2ForPreTraining.from_pretrained("facebook/mms-300m")
for param in self.wav2vec2.parameters():
param.requires_grad = False
param.grad = None
self.wav2vec2.eval()
self.feature_layer = layer
@torch.no_grad()
def forward(self, x):
"""
Args:
x: torch.Tensor of shape (B x t)
Returns:
y: torch.Tensor of shape(B x C x t)
"""
outputs = self.wav2vec2(x.squeeze(1), output_hidden_states=True)
y = outputs.hidden_states[self.feature_layer]
y = y.permute((0, 2, 1))
return y
class TextEncoder(nn.Module):
def __init__(self,
n_vocab,
out_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout):
super().__init__()
self.n_vocab = n_vocab
self.out_channels = out_channels
self.hidden_channels = hidden_channels
self.filter_channels = filter_channels
self.n_heads = n_heads
self.n_layers = n_layers
self.kernel_size = kernel_size
self.p_dropout = p_dropout
self.emb = nn.Embedding(n_vocab, hidden_channels)
nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
self.cond = nn.Conv1d(256, hidden_channels, 1)
self.encoder = attentions.Encoder(
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout)
self.encoder2 = attentions.Encoder(
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout)
self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
def forward(self, x, x_lengths, g):
x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
x = torch.transpose(x, 1, -1) # [b, h, t]
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
x = self.encoder(x * x_mask, x_mask)
x = x + self.cond(g)
x = self.encoder2(x * x_mask, x_mask)
stats = self.proj(x) * x_mask
m, logs = torch.split(stats, self.out_channels, dim=1)
return x, m, logs, x_mask
class ResidualCouplingBlock_Transformer(nn.Module):
def __init__(self,
channels,
hidden_channels,
kernel_size,
dilation_rate,
n_layers=3,
n_flows=4,
gin_channels=0):
super().__init__()
self.channels = channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
self.dilation_rate = dilation_rate
self.n_layers = n_layers
self.n_flows = n_flows
self.gin_channels = gin_channels
self.cond_block = torch.nn.Sequential(torch.nn.Linear(gin_channels, 4 * hidden_channels),
nn.SiLU(), torch.nn.Linear(4 * hidden_channels, hidden_channels))
self.flows = nn.ModuleList()
for i in range(n_flows):
self.flows.append(modules.ResidualCouplingLayer_Transformer_simple(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True, attention_head=4))
self.flows.append(modules.Flip())
def forward(self, x, x_mask, g=None, reverse=False):
g = self.cond_block(g.squeeze(2))
if not reverse:
for flow in self.flows:
x, _ = flow(x, x_mask, g=g, reverse=reverse)
else:
for flow in reversed(self.flows):
x = flow(x, x_mask, g=g, reverse=reverse)
return x
class PosteriorEncoder(nn.Module):
def __init__(self,
in_channels,
out_channels,
hidden_channels,
kernel_size,
dilation_rate,
n_layers,
gin_channels=0):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
self.dilation_rate = dilation_rate
self.n_layers = n_layers
self.gin_channels = gin_channels
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
def forward(self, x, x_mask, g=None):
x = self.pre(x) * x_mask
x = self.enc(x, x_mask, g=g)
stats = self.proj(x) * x_mask
m, logs = torch.split(stats, self.out_channels, dim=1)
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
return z, m, logs
class StochasticDurationPredictor(nn.Module):
def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
super().__init__()
filter_channels = in_channels # it needs to be removed from future version.
self.in_channels = in_channels
self.filter_channels = filter_channels
self.kernel_size = kernel_size
self.p_dropout = p_dropout
self.n_flows = n_flows
self.gin_channels = gin_channels
self.log_flow = modules.Log()
self.flows = nn.ModuleList()
self.flows.append(modules.ElementwiseAffine(2))
for i in range(n_flows):
self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
self.flows.append(modules.Flip())
self.post_pre = nn.Conv1d(1, filter_channels, 1)
self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
self.post_flows = nn.ModuleList()
self.post_flows.append(modules.ElementwiseAffine(2))
for i in range(4):
self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
self.post_flows.append(modules.Flip())
self.pre = nn.Conv1d(in_channels, filter_channels, 1)
self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
if gin_channels != 0:
self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
x = torch.detach(x)
x = self.pre(x)
if g is not None:
g = torch.detach(g)
x = x + self.cond(g)
x = self.convs(x, x_mask)
x = self.proj(x) * x_mask
if not reverse:
flows = self.flows
assert w is not None
logdet_tot_q = 0
h_w = self.post_pre(w)
h_w = self.post_convs(h_w, x_mask)
h_w = self.post_proj(h_w) * x_mask
e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
z_q = e_q
for flow in self.post_flows:
z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
logdet_tot_q += logdet_q
z_u, z1 = torch.split(z_q, [1, 1], 1)
u = torch.sigmoid(z_u) * x_mask
z0 = (w - u) * x_mask
logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
logdet_tot = 0
z0, logdet = self.log_flow(z0, x_mask)
logdet_tot += logdet
z = torch.cat([z0, z1], 1)
for flow in flows:
z, logdet = flow(z, x_mask, g=x, reverse=reverse)
logdet_tot = logdet_tot + logdet
nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
return nll + logq # [b]
else:
flows = list(reversed(self.flows))
flows = flows[:-2] + [flows[-1]] # remove a useless vflow
z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
for flow in flows:
z = flow(z, x_mask, g=x, reverse=reverse)
z0, z1 = torch.split(z, [1, 1], 1)
logw = z0
return logw
class W2VDecoder(nn.Module):
def __init__(self,
in_channels,
hidden_channels,
kernel_size,
dilation_rate,
n_layers,
output_size=1024,
gin_channels=0,
p_dropout=0):
super().__init__()
self.in_channels = in_channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
self.dilation_rate = dilation_rate
self.n_layers = n_layers
self.gin_channels = gin_channels
self.p_dropout = p_dropout
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, p_dropout=p_dropout)
self.proj = nn.Conv1d(hidden_channels, output_size, 1)
def forward(self, x, x_mask, g=None):
x = self.pre(x * x_mask) * x_mask
x = self.enc(x, x_mask, g=g)
x = self.proj(x) * x_mask
return x
class PitchPredictor(nn.Module):
def __init__(self):
super().__init__()
resblock_kernel_sizes = [3,5,7]
upsample_rates = [2,2]
initial_channel = 1024
upsample_initial_channel = 256
upsample_kernel_sizes = [4,4]
resblock_dilation_sizes = [[1,3,5], [1,3,5], [1,3,5]]
self.num_kernels = len(resblock_kernel_sizes)
self.num_upsamples = len(upsample_rates)
self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
resblock = modules.ResBlock1
self.ups = nn.ModuleList()
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
self.ups.append(weight_norm(
ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
k, u, padding=(k-u)//2)))
self.resblocks = nn.ModuleList()
for i in range(len(self.ups)):
ch = upsample_initial_channel//(2**(i+1))
for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
self.resblocks.append(resblock(ch, k, d))
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
self.ups.apply(init_weights)
self.cond = Conv1d(256, upsample_initial_channel, 1)
def forward(self, x, g):
x = self.conv_pre(x) + self.cond(g)
for i in range(self.num_upsamples):
x = F.leaky_relu(x, modules.LRELU_SLOPE)
x = self.ups[i](x)
xs = None
for j in range(self.num_kernels):
if xs is None:
xs = self.resblocks[i*self.num_kernels+j](x)
else:
xs += self.resblocks[i*self.num_kernels+j](x)
x = xs / self.num_kernels
x = F.leaky_relu(x)
## Predictor
x = self.conv_post(x)
return x
class SynthesizerTrn(nn.Module):
"""
Synthesizer for Training
"""
def __init__(self,
spec_channels,
segment_size,
inter_channels,
hidden_channels,
filter_channels,
n_heads,
n_layers,
kernel_size,
p_dropout,
resblock,
resblock_kernel_sizes,
resblock_dilation_sizes,
gin_channels=256,
prosody_size=20,
cfg=False,
**kwargs):
super().__init__()
self.spec_channels = spec_channels
self.inter_channels = inter_channels
self.hidden_channels = hidden_channels
self.filter_channels = filter_channels
self.n_heads = n_heads
self.n_layers = n_layers
self.kernel_size = kernel_size
self.p_dropout = p_dropout
self.resblock = resblock
self.resblock_kernel_sizes = resblock_kernel_sizes
self.resblock_dilation_sizes = resblock_dilation_sizes
self.segment_size = segment_size
self.mel_size = prosody_size
self.enc_q = PosteriorEncoder(1024, inter_channels, hidden_channels, 5, 1, 16, gin_channels=256)
self.enc_p = TextEncoder(178, out_channels=inter_channels, hidden_channels=inter_channels, filter_channels=inter_channels*4,
n_heads=4, n_layers=3, kernel_size=9, p_dropout=0.2)
self.flow = ResidualCouplingBlock_Transformer(inter_channels, hidden_channels, 5, 1, 3, gin_channels=256)
self.w2v_decoder = W2VDecoder(inter_channels, inter_channels*2, 5, 1, 8, output_size=1024, p_dropout=0.1, gin_channels=256)
self.emb_g = StyleEncoder(in_dim=80, hidden_dim=256, out_dim=256)
self.dp = StochasticDurationPredictor(inter_channels, inter_channels, 3, 0.5, 4, gin_channels=256)
self.pp = PitchPredictor()
self.phoneme_classifier = Conv1d(inter_channels, 178, 1, bias=False)
@torch.no_grad()
def infer(self, x, x_lengths, y_mel, y_length, noise_scale=1, noise_scale_w=1, length_scale=1):
y_mask = torch.unsqueeze(commons.sequence_mask(y_length, y_mel.size(2)), 1).to(y_mel.dtype)
# Speaker embedding from mel (Style Encoder)
g = self.emb_g(y_mel, y_mask).unsqueeze(-1)
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, g=g)
logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
w = torch.exp(logw) * x_mask * length_scale
w_ceil = torch.ceil(w)
y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
attn = commons.generate_path(w_ceil, attn_mask)
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
z = self.flow(z_p, y_mask, g=g, reverse=True)
w2v = self.w2v_decoder(z, y_mask, g=g)
pitch = self.pp(w2v, g)
return w2v, pitch
@torch.no_grad()
def infer_noise_control(self, x, x_lengths, y_mel, y_length, noise_scale=0.333, noise_scale_w=1, length_scale=1, denoise_ratio = 0):
y_mask = torch.unsqueeze(commons.sequence_mask(y_length, y_mel.size(2)), 1).to(y_mel.dtype)
# Speaker embedding from mel (Style Encoder)
g = self.emb_g(y_mel, y_mask).unsqueeze(-1)
g_org, g_denoise = g[:1, :, :], g[1:, :, :]
g = (1-denoise_ratio)*g_org + denoise_ratio*g_denoise
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, g=g)
logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
w = torch.exp(logw) * x_mask * length_scale
w_ceil = torch.ceil(w)
y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
attn = commons.generate_path(w_ceil, attn_mask)
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
z = self.flow(z_p, y_mask, g=g, reverse=True)
w2v = self.w2v_decoder(z, y_mask, g=g)
pitch = self.pp(w2v, g)
return w2v, pitch