DD0101/VITS

1import copy2import math3import torch4from torch import nn5from torch.nn import functional as F6 7import commons8import modules9import attentions10import monotonic_align11 12from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d13from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm14from commons import init_weights, get_padding15 16 17class StochasticDurationPredictor(nn.Module):18  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):19    super().__init__()20    filter_channels = in_channels # it needs to be removed from future version.21    self.in_channels = in_channels22    self.filter_channels = filter_channels23    self.kernel_size = kernel_size24    self.p_dropout = p_dropout25    self.n_flows = n_flows26    self.gin_channels = gin_channels27 28    self.log_flow = modules.Log()29    self.flows = nn.ModuleList()30    self.flows.append(modules.ElementwiseAffine(2))31    for i in range(n_flows):32      self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))33      self.flows.append(modules.Flip())34 35    self.post_pre = nn.Conv1d(1, filter_channels, 1)36    self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)37    self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)38    self.post_flows = nn.ModuleList()39    self.post_flows.append(modules.ElementwiseAffine(2))40    for i in range(4):41      self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))42      self.post_flows.append(modules.Flip())43 44    self.pre = nn.Conv1d(in_channels, filter_channels, 1)45    self.proj = nn.Conv1d(filter_channels, filter_channels, 1)46    self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)47    if gin_channels != 0:48      self.cond = nn.Conv1d(gin_channels, filter_channels, 1)49 50  def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):51    x = torch.detach(x)52    x = self.pre(x)53    if g is not None:54      g = torch.detach(g)55      x = x + self.cond(g)56    x = self.convs(x, x_mask)57    x = self.proj(x) * x_mask58 59    if not reverse:60      flows = self.flows61      assert w is not None62 63      logdet_tot_q = 0 64      h_w = self.post_pre(w)65      h_w = self.post_convs(h_w, x_mask)66      h_w = self.post_proj(h_w) * x_mask67      e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask68      z_q = e_q69      for flow in self.post_flows:70        z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))71        logdet_tot_q += logdet_q72      z_u, z1 = torch.split(z_q, [1, 1], 1) 73      u = torch.sigmoid(z_u) * x_mask74      z0 = (w - u) * x_mask75      logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])76      logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q77 78      logdet_tot = 079      z0, logdet = self.log_flow(z0, x_mask)80      logdet_tot += logdet81      z = torch.cat([z0, z1], 1)82      for flow in flows:83        z, logdet = flow(z, x_mask, g=x, reverse=reverse)84        logdet_tot = logdet_tot + logdet85      nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot86      return nll + logq # [b]87    else:88      flows = list(reversed(self.flows))89      flows = flows[:-2] + [flows[-1]] # remove a useless vflow90      z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale91      for flow in flows:92        z = flow(z, x_mask, g=x, reverse=reverse)93      z0, z1 = torch.split(z, [1, 1], 1)94      logw = z095      return logw96 97 98class DurationPredictor(nn.Module):99  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):100    super().__init__()101 102    self.in_channels = in_channels103    self.filter_channels = filter_channels104    self.kernel_size = kernel_size105    self.p_dropout = p_dropout106    self.gin_channels = gin_channels107 108    self.drop = nn.Dropout(p_dropout)109    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)110    self.norm_1 = modules.LayerNorm(filter_channels)111    self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)112    self.norm_2 = modules.LayerNorm(filter_channels)113    self.proj = nn.Conv1d(filter_channels, 1, 1)114 115    if gin_channels != 0:116      self.cond = nn.Conv1d(gin_channels, in_channels, 1)117 118  def forward(self, x, x_mask, g=None):119    x = torch.detach(x)120    if g is not None:121      g = torch.detach(g)122      x = x + self.cond(g)123    x = self.conv_1(x * x_mask)124    x = torch.relu(x)125    x = self.norm_1(x)126    x = self.drop(x)127    x = self.conv_2(x * x_mask)128    x = torch.relu(x)129    x = self.norm_2(x)130    x = self.drop(x)131    x = self.proj(x * x_mask)132    return x * x_mask133 134 135class TextEncoder(nn.Module):136  def __init__(self,137      n_vocab,138      out_channels,139      hidden_channels,140      filter_channels,141      n_heads,142      n_layers,143      kernel_size,144      p_dropout):145    super().__init__()146    self.n_vocab = n_vocab147    self.out_channels = out_channels148    self.hidden_channels = hidden_channels149    self.filter_channels = filter_channels150    self.n_heads = n_heads151    self.n_layers = n_layers152    self.kernel_size = kernel_size153    self.p_dropout = p_dropout154 155    self.emb = nn.Embedding(n_vocab, hidden_channels)156    nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)157 158    self.encoder = attentions.Encoder(159      hidden_channels,160      filter_channels,161      n_heads,162      n_layers,163      kernel_size,164      p_dropout)165    self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)166 167  def forward(self, x, x_lengths):168    x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]169    x = torch.transpose(x, 1, -1) # [b, h, t]170    x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)171 172    x = self.encoder(x * x_mask, x_mask)173    stats = self.proj(x) * x_mask174 175    m, logs = torch.split(stats, self.out_channels, dim=1)176    return x, m, logs, x_mask177 178 179class ResidualCouplingBlock(nn.Module):180  def __init__(self,181      channels,182      hidden_channels,183      kernel_size,184      dilation_rate,185      n_layers,186      n_flows=4,187      gin_channels=0):188    super().__init__()189    self.channels = channels190    self.hidden_channels = hidden_channels191    self.kernel_size = kernel_size192    self.dilation_rate = dilation_rate193    self.n_layers = n_layers194    self.n_flows = n_flows195    self.gin_channels = gin_channels196 197    self.flows = nn.ModuleList()198    for i in range(n_flows):199      self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))200      self.flows.append(modules.Flip())201 202  def forward(self, x, x_mask, g=None, reverse=False):203    if not reverse:204      for flow in self.flows:205        x, _ = flow(x, x_mask, g=g, reverse=reverse)206    else:207      for flow in reversed(self.flows):208        x = flow(x, x_mask, g=g, reverse=reverse)209    return x210 211 212class PosteriorEncoder(nn.Module):213  def __init__(self,214      in_channels,215      out_channels,216      hidden_channels,217      kernel_size,218      dilation_rate,219      n_layers,220      gin_channels=0):221    super().__init__()222    self.in_channels = in_channels223    self.out_channels = out_channels224    self.hidden_channels = hidden_channels225    self.kernel_size = kernel_size226    self.dilation_rate = dilation_rate227    self.n_layers = n_layers228    self.gin_channels = gin_channels229 230    self.pre = nn.Conv1d(in_channels, hidden_channels, 1)231    self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)232    self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)233 234  def forward(self, x, x_lengths, g=None):235    x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)236    x = self.pre(x) * x_mask237    x = self.enc(x, x_mask, g=g)238    stats = self.proj(x) * x_mask239    m, logs = torch.split(stats, self.out_channels, dim=1)240    z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask241    return z, m, logs, x_mask242 243 244class Generator(torch.nn.Module):245    def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):246        super(Generator, self).__init__()247        self.num_kernels = len(resblock_kernel_sizes)248        self.num_upsamples = len(upsample_rates)249        self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)250        resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2251 252        self.ups = nn.ModuleList()253        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):254            self.ups.append(weight_norm(255                ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),256                                k, u, padding=(k-u)//2)))257 258        self.resblocks = nn.ModuleList()259        for i in range(len(self.ups)):260            ch = upsample_initial_channel//(2**(i+1))261            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):262                self.resblocks.append(resblock(ch, k, d))263 264        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)265        self.ups.apply(init_weights)266 267        if gin_channels != 0:268            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)269 270    def forward(self, x, g=None):271        x = self.conv_pre(x)272        if g is not None:273          x = x + self.cond(g)274 275        for i in range(self.num_upsamples):276            x = F.leaky_relu(x, modules.LRELU_SLOPE)277            x = self.ups[i](x)278            xs = None279            for j in range(self.num_kernels):280                if xs is None:281                    xs = self.resblocks[i*self.num_kernels+j](x)282                else:283                    xs += self.resblocks[i*self.num_kernels+j](x)284            x = xs / self.num_kernels285        x = F.leaky_relu(x)286        x = self.conv_post(x)287        x = torch.tanh(x)288 289        return x290 291    def remove_weight_norm(self):292        print('Removing weight norm...')293        for l in self.ups:294            remove_weight_norm(l)295        for l in self.resblocks:296            l.remove_weight_norm()297 298 299class DiscriminatorP(torch.nn.Module):300    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):301        super(DiscriminatorP, self).__init__()302        self.period = period303        self.use_spectral_norm = use_spectral_norm304        norm_f = weight_norm if use_spectral_norm == False else spectral_norm305        self.convs = nn.ModuleList([306            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),307            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),308            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),309            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),310            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),311        ])312        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))313 314    def forward(self, x):315        fmap = []316 317        # 1d to 2d318        b, c, t = x.shape319        if t % self.period != 0: # pad first320            n_pad = self.period - (t % self.period)321            x = F.pad(x, (0, n_pad), "reflect")322            t = t + n_pad323        x = x.view(b, c, t // self.period, self.period)324 325        for l in self.convs:326            x = l(x)327            x = F.leaky_relu(x, modules.LRELU_SLOPE)328            fmap.append(x)329        x = self.conv_post(x)330        fmap.append(x)331        x = torch.flatten(x, 1, -1)332 333        return x, fmap334 335 336class DiscriminatorS(torch.nn.Module):337    def __init__(self, use_spectral_norm=False):338        super(DiscriminatorS, self).__init__()339        norm_f = weight_norm if use_spectral_norm == False else spectral_norm340        self.convs = nn.ModuleList([341            norm_f(Conv1d(1, 16, 15, 1, padding=7)),342            norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),343            norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),344            norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),345            norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),346            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),347        ])348        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))349 350    def forward(self, x):351        fmap = []352 353        for l in self.convs:354            x = l(x)355            x = F.leaky_relu(x, modules.LRELU_SLOPE)356            fmap.append(x)357        x = self.conv_post(x)358        fmap.append(x)359        x = torch.flatten(x, 1, -1)360 361        return x, fmap362 363 364class MultiPeriodDiscriminator(torch.nn.Module):365    def __init__(self, use_spectral_norm=False):366        super(MultiPeriodDiscriminator, self).__init__()367        periods = [2,3,5,7,11]368 369        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]370        discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]371        self.discriminators = nn.ModuleList(discs)372 373    def forward(self, y, y_hat):374        y_d_rs = []375        y_d_gs = []376        fmap_rs = []377        fmap_gs = []378        for i, d in enumerate(self.discriminators):379            y_d_r, fmap_r = d(y)380            y_d_g, fmap_g = d(y_hat)381            y_d_rs.append(y_d_r)382            y_d_gs.append(y_d_g)383            fmap_rs.append(fmap_r)384            fmap_gs.append(fmap_g)385 386        return y_d_rs, y_d_gs, fmap_rs, fmap_gs387 388 389 390class SynthesizerTrn(nn.Module):391  """392  Synthesizer for Training393  """394 395  def __init__(self, 396    n_vocab,397    spec_channels,398    segment_size,399    inter_channels,400    hidden_channels,401    filter_channels,402    n_heads,403    n_layers,404    kernel_size,405    p_dropout,406    resblock, 407    resblock_kernel_sizes, 408    resblock_dilation_sizes, 409    upsample_rates, 410    upsample_initial_channel, 411    upsample_kernel_sizes,412    n_speakers=0,413    gin_channels=0,414    use_sdp=True,415    **kwargs):416 417    super().__init__()418    self.n_vocab = n_vocab419    self.spec_channels = spec_channels420    self.inter_channels = inter_channels421    self.hidden_channels = hidden_channels422    self.filter_channels = filter_channels423    self.n_heads = n_heads424    self.n_layers = n_layers425    self.kernel_size = kernel_size426    self.p_dropout = p_dropout427    self.resblock = resblock428    self.resblock_kernel_sizes = resblock_kernel_sizes429    self.resblock_dilation_sizes = resblock_dilation_sizes430    self.upsample_rates = upsample_rates431    self.upsample_initial_channel = upsample_initial_channel432    self.upsample_kernel_sizes = upsample_kernel_sizes433    self.segment_size = segment_size434    self.n_speakers = n_speakers435    self.gin_channels = gin_channels436 437    self.use_sdp = use_sdp438 439    self.enc_p = TextEncoder(n_vocab,440        inter_channels,441        hidden_channels,442        filter_channels,443        n_heads,444        n_layers,445        kernel_size,446        p_dropout)447    self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)448    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)449    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)450 451    if use_sdp:452      self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)453    else:454      self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)455 456    if n_speakers > 1:457      self.emb_g = nn.Embedding(n_speakers, gin_channels)458 459  def forward(self, x, x_lengths, y, y_lengths, sid=None):460 461    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)462    if self.n_speakers > 0:463      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]464    else:465      g = None466 467    z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)468    z_p = self.flow(z, y_mask, g=g)469 470    with torch.no_grad():471      # negative cross-entropy472      s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]473      neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]474      neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2), s_p_sq_r) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]475      neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r)) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]476      neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]477      neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4478 479      attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)480      attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()481 482    w = attn.sum(2)483    if self.use_sdp:484      l_length = self.dp(x, x_mask, w, g=g)485      l_length = l_length / torch.sum(x_mask)486    else:487      logw_ = torch.log(w + 1e-6) * x_mask488      logw = self.dp(x, x_mask, g=g)489      l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging 490 491    # expand prior492    m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)493    logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)494 495    z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)496    o = self.dec(z_slice, g=g)497    return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)498 499  def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):500    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)501    if self.n_speakers > 0:502      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]503    else:504      g = None505 506    if self.use_sdp:507      logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)508    else:509      logw = self.dp(x, x_mask, g=g)510    w = torch.exp(logw) * x_mask * length_scale511    w_ceil = torch.ceil(w)512    y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()513    y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)514    attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)515    attn = commons.generate_path(w_ceil, attn_mask)516 517    m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']518    logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']519 520    z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale521    z = self.flow(z_p, y_mask, g=g, reverse=True)522    o = self.dec((z * y_mask)[:,:,:max_len], g=g)523    return o, attn, y_mask, (z, z_p, m_p, logs_p)524 525  def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):526    assert self.n_speakers > 0, "n_speakers have to be larger than 0."527    g_src = self.emb_g(sid_src).unsqueeze(-1)528    g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)529    z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)530    z_p = self.flow(z, y_mask, g=g_src)531    z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)532    o_hat = self.dec(z_hat * y_mask, g=g_tgt)533    return o_hat, y_mask, (z, z_p, z_hat)534 535