bacow/GPT-SoVITS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

gpt_sovits_v3

Browse Source 
gpt_sovits_v3
This commit is contained in:
RVC-Boss
2025-02-11 21:07:26 +08:00
committed by GitHub
parent fa42d26d0e
commit 25cb1bf400
2 changed files with 474 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

250
GPT_SoVITS/module/models.py
View File

@@ -12,7 +12,7 @@ from torch.nn import functional as F
from module import commons
from module import modules
from module import attentions
from f5_tts.model import DiT
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
from module.commons import init_weights, get_padding
@@ -22,7 +22,7 @@ from module.quantize import ResidualVectorQuantizer
from text import symbols as symbols_v1
from text import symbols2 as symbols_v2
from torch.cuda.amp import autocast
import contextlib
import contextlib,random
class StochasticDurationPredictor(nn.Module):
@@ -371,6 +371,37 @@ class PosteriorEncoder(nn.Module):
return z, m, logs, x_mask
class Encoder(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, 1)
def forward(self, x, x_lengths, g=None):
if(g!=None):
g = g.detach()
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
x = self.pre(x) * x_mask
x = self.enc(x, x_mask, g=g)
stats = self.proj(x) * x_mask
return stats, x_mask
class WNEncoder(nn.Module):
def __init__(
self,
@@ -1028,3 +1059,218 @@ class SynthesizerTrn(nn.Module):
ssl = self.ssl_proj(x)
quantized, codes, commit_loss, quantized_list = self.quantizer(ssl)
return codes.transpose(0, 1)
class CFM(torch.nn.Module):
def __init__(
self,
in_channels,dit
):
super().__init__()
self.sigma_min = 1e-6
self.estimator = dit
self.in_channels = in_channels
self.criterion = torch.nn.MSELoss()
@torch.inference_mode()
def inference(self, mu, x_lens, prompt, n_timesteps, temperature=1.0, inference_cfg_rate=0):
"""Forward diffusion"""
B, T = mu.size(0), mu.size(1)
x = torch.randn([B, self.in_channels, T], device=mu.device,dtype=mu.dtype) * temperature
prompt_len = prompt.size(-1)
prompt_x = torch.zeros_like(x,dtype=mu.dtype)
prompt_x[..., :prompt_len] = prompt[..., :prompt_len]
x[..., :prompt_len] = 0
mu=mu.transpose(2,1)
t = 0
d = 1 / n_timesteps
for j in range(n_timesteps):
t_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * t
d_tensor = torch.ones(x.shape[0], device=x.device,dtype=mu.dtype) * d
# v_pred = model(x, t_tensor, d_tensor, **extra_args)
v_pred = self.estimator(x, prompt_x, x_lens, t_tensor,d_tensor, mu,drop_audio_cond=False,drop_text=False).transpose(2, 1)
if inference_cfg_rate>1e-5:
neg = self.estimator(x, prompt_x, x_lens, t_tensor, d_tensor, mu, drop_audio_cond=True, drop_text=True).transpose(2, 1)
v_pred=v_pred+(v_pred-neg)*inference_cfg_rate
x = x + d * v_pred
t = t + d
x[:, :, :prompt_len] = 0
return x
def forward(self, x1, x_lens, prompt_lens, mu):
b, _, t = x1.shape
# random timestep
t = torch.rand([b], device=mu.device, dtype=x1.dtype)
x0 = torch.randn_like(x1,device=mu.device)
vt = x1 - x0
xt = x0 + t[:, None, None] * vt
dt = torch.zeros_like(t,device=mu.device)
prompt = torch.zeros_like(x1)
for bib in range(b):
prompt[bib, :, :prompt_lens[bib]] = x1[bib, :, :prompt_lens[bib]]
xt[bib, :, :prompt_lens[bib]] = 0
gailv=0.3# if ttime()>1736250488 else 0.1
if random.random() < gailv:
base = torch.randint(2, 8, (t.shape[0],), device=mu.device)
d = 1/torch.pow(2, base)
d_input = d.clone()
d_input[d_input < 1e-2] = 0
# with torch.no_grad():
v_pred_1 = self.estimator(xt, prompt, x_lens, t, d_input, mu).transpose(2, 1).detach()
# v_pred_1 = self.diffusion(xt, t, d_input, cond=conditioning).detach()
x_mid = xt + d[:, None, None] * v_pred_1
# v_pred_2 = self.diffusion(x_mid, t+d, d_input, cond=conditioning).detach()
v_pred_2 = self.estimator(x_mid, prompt, x_lens, t+d, d_input, mu).transpose(2, 1).detach()
vt = (v_pred_1 + v_pred_2) / 2
vt = vt.detach()
dt = 2*d
vt_pred = self.estimator(xt, prompt, x_lens, t,dt, mu).transpose(2,1)
loss = 0
# print(45555555,estimator_out.shape,u.shape,x_lens,prompt_lens)#45555555 torch.Size([7, 465, 100]) torch.Size([7, 100, 465]) tensor([461, 461, 451, 451, 442, 442, 442], device='cuda:0') tensor([ 96, 93, 185, 59, 244, 262, 294], device='cuda:0')
for bib in range(b):
loss += self.criterion(vt_pred[bib, :, prompt_lens[bib]:x_lens[bib]], vt[bib, :, prompt_lens[bib]:x_lens[bib]])
loss /= b
return loss#, estimator_out + (1 - self.sigma_min) * z
class SynthesizerTrnV3(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,
upsample_rates,
upsample_initial_channel,
upsample_kernel_sizes,
n_speakers=0,
gin_channels=0,
use_sdp=True,
semantic_frame_rate=None,
freeze_quantizer=None,
**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.upsample_rates = upsample_rates
self.upsample_initial_channel = upsample_initial_channel
self.upsample_kernel_sizes = upsample_kernel_sizes
self.segment_size = segment_size
self.n_speakers = n_speakers
self.gin_channels = gin_channels
self.model_dim=512
self.use_sdp = use_sdp
self.enc_p = TextEncoder(inter_channels,hidden_channels,filter_channels,n_heads,n_layers,kernel_size,p_dropout)
# self.ref_enc = modules.MelStyleEncoder(spec_channels, style_vector_dim=gin_channels)###回滚。。。
self.ref_enc = modules.MelStyleEncoder(704, style_vector_dim=gin_channels)###回滚。。。
# self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates,
# upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
# self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16,
# gin_channels=gin_channels)
# self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
ssl_dim = 768
assert semantic_frame_rate in ['25hz', "50hz"]
self.semantic_frame_rate = semantic_frame_rate
if semantic_frame_rate == '25hz':
self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 2, stride=2)
else:
self.ssl_proj = nn.Conv1d(ssl_dim, ssl_dim, 1, stride=1)
self.quantizer = ResidualVectorQuantizer(
dimension=ssl_dim,
n_q=1,
bins=1024
)
self.freeze_quantizer=freeze_quantizer
inter_channels2=512
self.bridge=nn.Sequential(
nn.Conv1d(inter_channels, inter_channels2, 1, stride=1),
nn.LeakyReLU()
)
self.wns1=Encoder(inter_channels2, inter_channels2, inter_channels2, 5, 1, 8,gin_channels=gin_channels)
self.linear_mel=nn.Conv1d(inter_channels2,100,1,stride=1)
self.cfm = CFM(100,DiT(**dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=inter_channels2, conv_layers=4)),)#text_dim is condition feature dim
def forward(self, ssl, y, mel,ssl_lengths,y_lengths, text, text_lengths,mel_lengths):#ssl_lengths no need now
with autocast(enabled=False):
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(y.dtype)
ge = self.ref_enc(y[:,:704] * y_mask, y_mask)
maybe_no_grad = torch.no_grad() if self.freeze_quantizer else contextlib.nullcontext()
with maybe_no_grad:
if self.freeze_quantizer:
self.ssl_proj.eval()#
self.quantizer.eval()
ssl = self.ssl_proj(ssl)
quantized, codes, commit_loss, quantized_list = self.quantizer(
ssl, layers=[0]
)
with maybe_no_grad:
quantized = F.interpolate(quantized, scale_factor=2, mode="nearest")##BCT
x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge)
fea=self.bridge(x)
fea = F.interpolate(fea, scale_factor=1.875, mode="nearest")##BCT
fea, y_mask_ = self.wns1(fea, mel_lengths, ge)###如果1min微调没问题就不需要微操学习率了
B=ssl.shape[0]
prompt_len_max = mel_lengths*2/3
prompt_len = (torch.rand([B], device=fea.device) * prompt_len_max).floor().to(dtype=torch.long)
minn=min(mel.shape[-1],fea.shape[-1])
mel=mel[:,:,:minn]
fea=fea[:,:,:minn]
cfm_loss= self.cfm(mel, mel_lengths, prompt_len, fea)
return cfm_loss
@torch.no_grad()
def decode_encp(self, codes,text, refer,ge=None):
# print(2333333,refer.shape)
# ge=None
if(ge==None):
refer_lengths = torch.LongTensor([refer.size(2)]).to(refer.device)
refer_mask = torch.unsqueeze(commons.sequence_mask(refer_lengths, refer.size(2)), 1).to(refer.dtype)
ge = self.ref_enc(refer[:,:704] * refer_mask, refer_mask)
y_lengths = torch.LongTensor([int(codes.size(2)*2)]).to(codes.device)
y_lengths1 = torch.LongTensor([int(codes.size(2)*2.5*1.5)]).to(codes.device)
text_lengths = torch.LongTensor([text.size(-1)]).to(text.device)
quantized = self.quantizer.decode(codes)
if self.semantic_frame_rate == '25hz':
quantized = F.interpolate(quantized, scale_factor=2, mode="nearest")##BCT
x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge)
fea=self.bridge(x)
fea = F.interpolate(fea, scale_factor=1.875, mode="nearest")##BCT
####more wn paramter to learn mel
fea, y_mask_ = self.wns1(fea, y_lengths1, ge)
return fea,ge
def extract_latent(self, x):
ssl = self.ssl_proj(x)
quantized, codes, commit_loss, quantized_list = self.quantizer(ssl)
return codes.transpose(0,1)