Kafke/Code-Realize-TTS

1import math2import torch3from torch import nn4from torch.nn import functional as F5 6import commons7import modules8import attentions9 10from torch.nn import Conv1d, ConvTranspose1d, Conv2d11from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm12from commons import init_weights, get_padding13 14 15class StochasticDurationPredictor(nn.Module):16  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):17    super().__init__()18    filter_channels = in_channels # it needs to be removed from future version.19    self.in_channels = in_channels20    self.filter_channels = filter_channels21    self.kernel_size = kernel_size22    self.p_dropout = p_dropout23    self.n_flows = n_flows24    self.gin_channels = gin_channels25 26    self.log_flow = modules.Log()27    self.flows = nn.ModuleList()28    self.flows.append(modules.ElementwiseAffine(2))29    for i in range(n_flows):30      self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))31      self.flows.append(modules.Flip())32 33    self.post_pre = nn.Conv1d(1, filter_channels, 1)34    self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)35    self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)36    self.post_flows = nn.ModuleList()37    self.post_flows.append(modules.ElementwiseAffine(2))38    for i in range(4):39      self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))40      self.post_flows.append(modules.Flip())41 42    self.pre = nn.Conv1d(in_channels, filter_channels, 1)43    self.proj = nn.Conv1d(filter_channels, filter_channels, 1)44    self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)45    if gin_channels != 0:46      self.cond = nn.Conv1d(gin_channels, filter_channels, 1)47 48  def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):49    x = torch.detach(x)50    x = self.pre(x)51    if g is not None:52      g = torch.detach(g)53      x = x + self.cond(g)54    x = self.convs(x, x_mask)55    x = self.proj(x) * x_mask56 57    if not reverse:58      flows = self.flows59      assert w is not None60 61      logdet_tot_q = 0 62      h_w = self.post_pre(w)63      h_w = self.post_convs(h_w, x_mask)64      h_w = self.post_proj(h_w) * x_mask65      e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask66      z_q = e_q67      for flow in self.post_flows:68        z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))69        logdet_tot_q += logdet_q70      z_u, z1 = torch.split(z_q, [1, 1], 1) 71      u = torch.sigmoid(z_u) * x_mask72      z0 = (w - u) * x_mask73      logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])74      logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q75 76      logdet_tot = 077      z0, logdet = self.log_flow(z0, x_mask)78      logdet_tot += logdet79      z = torch.cat([z0, z1], 1)80      for flow in flows:81        z, logdet = flow(z, x_mask, g=x, reverse=reverse)82        logdet_tot = logdet_tot + logdet83      nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot84      return nll + logq # [b]85    else:86      flows = list(reversed(self.flows))87      flows = flows[:-2] + [flows[-1]] # remove a useless vflow88      z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale89      for flow in flows:90        z = flow(z, x_mask, g=x, reverse=reverse)91      z0, z1 = torch.split(z, [1, 1], 1)92      logw = z093      return logw94 95 96class DurationPredictor(nn.Module):97  def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):98    super().__init__()99 100    self.in_channels = in_channels101    self.filter_channels = filter_channels102    self.kernel_size = kernel_size103    self.p_dropout = p_dropout104    self.gin_channels = gin_channels105 106    self.drop = nn.Dropout(p_dropout)107    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)108    self.norm_1 = modules.LayerNorm(filter_channels)109    self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)110    self.norm_2 = modules.LayerNorm(filter_channels)111    self.proj = nn.Conv1d(filter_channels, 1, 1)112 113    if gin_channels != 0:114      self.cond = nn.Conv1d(gin_channels, in_channels, 1)115 116  def forward(self, x, x_mask, g=None):117    x = torch.detach(x)118    if g is not None:119      g = torch.detach(g)120      x = x + self.cond(g)121    x = self.conv_1(x * x_mask)122    x = torch.relu(x)123    x = self.norm_1(x)124    x = self.drop(x)125    x = self.conv_2(x * x_mask)126    x = torch.relu(x)127    x = self.norm_2(x)128    x = self.drop(x)129    x = self.proj(x * x_mask)130    return x * x_mask131 132 133class TextEncoder(nn.Module):134  def __init__(self,135      n_vocab,136      out_channels,137      hidden_channels,138      filter_channels,139      n_heads,140      n_layers,141      kernel_size,142      p_dropout,143      emotion_embedding):144    super().__init__()145    self.n_vocab = n_vocab146    self.out_channels = out_channels147    self.hidden_channels = hidden_channels148    self.filter_channels = filter_channels149    self.n_heads = n_heads150    self.n_layers = n_layers151    self.kernel_size = kernel_size152    self.p_dropout = p_dropout153    self.emotion_embedding = emotion_embedding154    155    if self.n_vocab!=0:156      self.emb = nn.Embedding(n_vocab, hidden_channels)157      if emotion_embedding:158        self.emotion_emb = nn.Linear(1024, hidden_channels)159      nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)160 161    self.encoder = attentions.Encoder(162      hidden_channels,163      filter_channels,164      n_heads,165      n_layers,166      kernel_size,167      p_dropout)168    self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)169 170  def forward(self, x, x_lengths, emotion_embedding=None):171    if self.n_vocab!=0:172      x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]173    if emotion_embedding is not None:174      x = x + self.emotion_emb(emotion_embedding.unsqueeze(1))175    x = torch.transpose(x, 1, -1) # [b, h, t]176    x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)177 178    x = self.encoder(x * x_mask, x_mask)179    stats = self.proj(x) * x_mask180 181    m, logs = torch.split(stats, self.out_channels, dim=1)182    return x, m, logs, x_mask183 184 185class ResidualCouplingBlock(nn.Module):186  def __init__(self,187      channels,188      hidden_channels,189      kernel_size,190      dilation_rate,191      n_layers,192      n_flows=4,193      gin_channels=0):194    super().__init__()195    self.channels = channels196    self.hidden_channels = hidden_channels197    self.kernel_size = kernel_size198    self.dilation_rate = dilation_rate199    self.n_layers = n_layers200    self.n_flows = n_flows201    self.gin_channels = gin_channels202 203    self.flows = nn.ModuleList()204    for i in range(n_flows):205      self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))206      self.flows.append(modules.Flip())207 208  def forward(self, x, x_mask, g=None, reverse=False):209    if not reverse:210      for flow in self.flows:211        x, _ = flow(x, x_mask, g=g, reverse=reverse)212    else:213      for flow in reversed(self.flows):214        x = flow(x, x_mask, g=g, reverse=reverse)215    return x216 217 218class PosteriorEncoder(nn.Module):219  def __init__(self,220      in_channels,221      out_channels,222      hidden_channels,223      kernel_size,224      dilation_rate,225      n_layers,226      gin_channels=0):227    super().__init__()228    self.in_channels = in_channels229    self.out_channels = out_channels230    self.hidden_channels = hidden_channels231    self.kernel_size = kernel_size232    self.dilation_rate = dilation_rate233    self.n_layers = n_layers234    self.gin_channels = gin_channels235 236    self.pre = nn.Conv1d(in_channels, hidden_channels, 1)237    self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)238    self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)239 240  def forward(self, x, x_lengths, g=None):241    x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)242    x = self.pre(x) * x_mask243    x = self.enc(x, x_mask, g=g)244    stats = self.proj(x) * x_mask245    m, logs = torch.split(stats, self.out_channels, dim=1)246    z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask247    return z, m, logs, x_mask248 249 250class Generator(torch.nn.Module):251    def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):252        super(Generator, self).__init__()253        self.num_kernels = len(resblock_kernel_sizes)254        self.num_upsamples = len(upsample_rates)255        self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)256        resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2257 258        self.ups = nn.ModuleList()259        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):260            self.ups.append(weight_norm(261                ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),262                                k, u, padding=(k-u)//2)))263 264        self.resblocks = nn.ModuleList()265        for i in range(len(self.ups)):266            ch = upsample_initial_channel//(2**(i+1))267            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):268                self.resblocks.append(resblock(ch, k, d))269 270        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)271        self.ups.apply(init_weights)272 273        if gin_channels != 0:274            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)275 276    def forward(self, x, g=None):277        x = self.conv_pre(x)278        if g is not None:279          x = x + self.cond(g)280 281        for i in range(self.num_upsamples):282            x = F.leaky_relu(x, modules.LRELU_SLOPE)283            x = self.ups[i](x)284            xs = None285            for j in range(self.num_kernels):286                if xs is None:287                    xs = self.resblocks[i*self.num_kernels+j](x)288                else:289                    xs += self.resblocks[i*self.num_kernels+j](x)290            x = xs / self.num_kernels291        x = F.leaky_relu(x)292        x = self.conv_post(x)293        x = torch.tanh(x)294 295        return x296 297    def remove_weight_norm(self):298        print('Removing weight norm...')299        for l in self.ups:300            remove_weight_norm(l)301        for l in self.resblocks:302            l.remove_weight_norm()303 304 305class DiscriminatorP(torch.nn.Module):306    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):307        super(DiscriminatorP, self).__init__()308        self.period = period309        self.use_spectral_norm = use_spectral_norm310        norm_f = weight_norm if use_spectral_norm == False else spectral_norm311        self.convs = nn.ModuleList([312            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),313            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),314            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),315            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),316            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),317        ])318        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))319 320    def forward(self, x):321        fmap = []322 323        # 1d to 2d324        b, c, t = x.shape325        if t % self.period != 0: # pad first326            n_pad = self.period - (t % self.period)327            x = F.pad(x, (0, n_pad), "reflect")328            t = t + n_pad329        x = x.view(b, c, t // self.period, self.period)330 331        for l in self.convs:332            x = l(x)333            x = F.leaky_relu(x, modules.LRELU_SLOPE)334            fmap.append(x)335        x = self.conv_post(x)336        fmap.append(x)337        x = torch.flatten(x, 1, -1)338 339        return x, fmap340 341 342class DiscriminatorS(torch.nn.Module):343    def __init__(self, use_spectral_norm=False):344        super(DiscriminatorS, self).__init__()345        norm_f = weight_norm if use_spectral_norm == False else spectral_norm346        self.convs = nn.ModuleList([347            norm_f(Conv1d(1, 16, 15, 1, padding=7)),348            norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),349            norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),350            norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),351            norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),352            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),353        ])354        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))355 356    def forward(self, x):357        fmap = []358 359        for l in self.convs:360            x = l(x)361            x = F.leaky_relu(x, modules.LRELU_SLOPE)362            fmap.append(x)363        x = self.conv_post(x)364        fmap.append(x)365        x = torch.flatten(x, 1, -1)366 367        return x, fmap368 369 370class MultiPeriodDiscriminator(torch.nn.Module):371    def __init__(self, use_spectral_norm=False):372        super(MultiPeriodDiscriminator, self).__init__()373        periods = [2,3,5,7,11]374 375        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]376        discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]377        self.discriminators = nn.ModuleList(discs)378 379    def forward(self, y, y_hat):380        y_d_rs = []381        y_d_gs = []382        fmap_rs = []383        fmap_gs = []384        for i, d in enumerate(self.discriminators):385            y_d_r, fmap_r = d(y)386            y_d_g, fmap_g = d(y_hat)387            y_d_rs.append(y_d_r)388            y_d_gs.append(y_d_g)389            fmap_rs.append(fmap_r)390            fmap_gs.append(fmap_g)391 392        return y_d_rs, y_d_gs, fmap_rs, fmap_gs393 394 395 396class SynthesizerTrn(nn.Module):397  """398  Synthesizer for Training399  """400 401  def __init__(self, 402    n_vocab,403    spec_channels,404    segment_size,405    inter_channels,406    hidden_channels,407    filter_channels,408    n_heads,409    n_layers,410    kernel_size,411    p_dropout,412    resblock, 413    resblock_kernel_sizes, 414    resblock_dilation_sizes, 415    upsample_rates, 416    upsample_initial_channel, 417    upsample_kernel_sizes,418    n_speakers=0,419    gin_channels=0,420    use_sdp=True,421    emotion_embedding=False,422    **kwargs):423 424    super().__init__()425    self.n_vocab = n_vocab426    self.spec_channels = spec_channels427    self.inter_channels = inter_channels428    self.hidden_channels = hidden_channels429    self.filter_channels = filter_channels430    self.n_heads = n_heads431    self.n_layers = n_layers432    self.kernel_size = kernel_size433    self.p_dropout = p_dropout434    self.resblock = resblock435    self.resblock_kernel_sizes = resblock_kernel_sizes436    self.resblock_dilation_sizes = resblock_dilation_sizes437    self.upsample_rates = upsample_rates438    self.upsample_initial_channel = upsample_initial_channel439    self.upsample_kernel_sizes = upsample_kernel_sizes440    self.segment_size = segment_size441    self.n_speakers = n_speakers442    self.gin_channels = gin_channels443 444    self.use_sdp = use_sdp445 446    self.enc_p = TextEncoder(n_vocab,447        inter_channels,448        hidden_channels,449        filter_channels,450        n_heads,451        n_layers,452        kernel_size,453        p_dropout,454        emotion_embedding)455    self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)456    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)457    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)458 459    if use_sdp:460      self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)461    else:462      self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)463 464    if n_speakers > 1:465      self.emb_g = nn.Embedding(n_speakers, gin_channels)466 467  def forward(self, x, x_lengths, y, y_lengths, sid=None):468 469    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)470    if self.n_speakers > 0:471      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]472    else:473      g = None474 475    z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)476    z_p = self.flow(z, y_mask, g=g)477 478    with torch.no_grad():479      # negative cross-entropy480      s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]481      neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]482      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]483      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]484      neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]485      neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4486 487      attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)488      attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()489 490    w = attn.sum(2)491    if self.use_sdp:492      l_length = self.dp(x, x_mask, w, g=g)493      l_length = l_length / torch.sum(x_mask)494    else:495      logw_ = torch.log(w + 1e-6) * x_mask496      logw = self.dp(x, x_mask, g=g)497      l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging 498 499    # expand prior500    m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)501    logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)502 503    z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)504    o = self.dec(z_slice, g=g)505    return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)506 507  def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None, emotion_embedding=None):508    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, emotion_embedding)509    if self.n_speakers > 0:510      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]511    else:512      g = None513 514    if self.use_sdp:515      logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)516    else:517      logw = self.dp(x, x_mask, g=g)518    w = torch.exp(logw) * x_mask * length_scale519    w_ceil = torch.ceil(w)520    y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()521    y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)522    attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)523    attn = commons.generate_path(w_ceil, attn_mask)524 525    m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']526    logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']527 528    z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale529    z = self.flow(z_p, y_mask, g=g, reverse=True)530    o = self.dec((z * y_mask)[:,:,:max_len], g=g)531    return o, attn, y_mask, (z, z_p, m_p, logs_p)532 533  def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):534    assert self.n_speakers > 0, "n_speakers have to be larger than 0."535    g_src = self.emb_g(sid_src).unsqueeze(-1)536    g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)537    z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)538    z_p = self.flow(z, y_mask, g=g_src)539    z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)540    o_hat = self.dec(z_hat * y_mask, g=g_tgt)541    return o_hat, y_mask, (z, z_p, z_hat)542 543