import math import torch from torch import nn from torch.nn import functional as F from tiny_tts.nn import commons from tiny_tts.nn import modules from tiny_tts.nn import attentions from torch.nn import Conv1d, ConvTranspose1d from torch.nn.utils import weight_norm, remove_weight_norm from tiny_tts.nn.commons import initialize_weights, compute_padding import tiny_tts.alignment as alignment class AttentionFlowBlock(nn.Module): def __init__( self, channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, n_flows=4, gin_channels=0, share_parameter=False, ): super().__init__() self.channels = channels self.hidden_channels = hidden_channels self.kernel_size = kernel_size self.n_layers = n_layers self.n_flows = n_flows self.gin_channels = gin_channels self.flows = nn.ModuleList() self.wn = ( attentions.FeedForward( hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, isflow=True, gin_channels=self.gin_channels, ) if share_parameter else None ) for i in range(n_flows): self.flows.append( modules.TransformerCouplingLayer( channels, hidden_channels, kernel_size, n_layers, n_heads, p_dropout, filter_channels, mean_only=True, wn_sharing_parameter=self.wn, gin_channels=self.gin_channels, ) ) self.flows.append(modules.FlipTransform()) def forward(self, x, x_mask, g=None, reverse=False): 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 VariationalDurationModel(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 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.LogTransform() self.flows = nn.ModuleList() self.flows.append(modules.AffineCoupling(2)) for i in range(n_flows): self.flows.append( modules.ConvolutionalFlow(2, filter_channels, kernel_size, n_layers=3) ) self.flows.append(modules.FlipTransform()) self.post_pre = nn.Conv1d(1, filter_channels, 1) self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1) self.post_convs = modules.DepthwiseSepConv( filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout ) self.post_flows = nn.ModuleList() self.post_flows.append(modules.AffineCoupling(2)) for i in range(4): self.post_flows.append( modules.ConvolutionalFlow(2, filter_channels, kernel_size, n_layers=3) ) self.post_flows.append(modules.FlipTransform()) self.pre = nn.Conv1d(in_channels, filter_channels, 1) self.proj = nn.Conv1d(filter_channels, filter_channels, 1) self.convs = modules.DepthwiseSepConv( 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 else: flows = list(reversed(self.flows)) flows = flows[:-2] + [flows[-1]] 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 DurationEstimator(nn.Module): def __init__( self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0 ): super().__init__() self.in_channels = in_channels self.filter_channels = filter_channels self.kernel_size = kernel_size self.p_dropout = p_dropout self.gin_channels = gin_channels self.drop = nn.Dropout(p_dropout) self.conv_1 = nn.Conv1d( in_channels, filter_channels, kernel_size, padding=kernel_size // 2 ) self.norm_1 = modules.ChannelNorm(filter_channels) self.conv_2 = nn.Conv1d( filter_channels, filter_channels, kernel_size, padding=kernel_size // 2 ) self.norm_2 = modules.ChannelNorm(filter_channels) self.proj = nn.Conv1d(filter_channels, 1, 1) if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, in_channels, 1) def forward(self, x, x_mask, g=None): x = torch.detach(x) if g is not None: g = torch.detach(g) x = x + self.cond(g) x = self.conv_1(x * x_mask) x = torch.relu(x) x = self.norm_1(x) x = self.drop(x) x = self.conv_2(x * x_mask) x = torch.relu(x) x = self.norm_2(x) x = self.drop(x) x = self.proj(x * x_mask) return x * x_mask class PhonemeEncoder(nn.Module): def __init__( self, n_vocab, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, gin_channels=0, num_languages=None, num_tones=None, ): super().__init__() if num_languages is None: from tiny_tts.text import num_languages if num_tones is None: from tiny_tts.text import num_tones 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.gin_channels = gin_channels self.emb = nn.Embedding(n_vocab, hidden_channels) nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5) self.tone_emb = nn.Embedding(num_tones, hidden_channels) nn.init.normal_(self.tone_emb.weight, 0.0, hidden_channels**-0.5) self.language_emb = nn.Embedding(num_languages, hidden_channels) nn.init.normal_(self.language_emb.weight, 0.0, hidden_channels**-0.5) self.bert_proj = nn.Conv1d(1024, hidden_channels, 1) self.ja_bert_proj = nn.Conv1d(768, hidden_channels, 1) self.encoder = attentions.TransformerBlock( hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, gin_channels=self.gin_channels, ) self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1) def forward(self, x, x_lengths, tone, language, bert, ja_bert, g=None): bert_emb = self.bert_proj(bert).transpose(1, 2) ja_bert_emb = self.ja_bert_proj(ja_bert).transpose(1, 2) x = ( self.emb(x) + self.tone_emb(tone) + self.language_emb(language) + bert_emb + ja_bert_emb ) * math.sqrt( self.hidden_channels ) x = torch.transpose(x, 1, -1) x_mask = torch.unsqueeze(commons.create_length_mask(x_lengths, x.size(2)), 1).to( x.dtype ) x = self.encoder(x * x_mask, x_mask, g=g) stats = self.proj(x) * x_mask m, logs = torch.split(stats, self.out_channels, dim=1) return x, m, logs, x_mask class FlowBlock(nn.Module): def __init__( self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, 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.flows = nn.ModuleList() for i in range(n_flows): self.flows.append( modules.FlowCouplingLayer( channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True, ) ) self.flows.append(modules.FlipTransform()) def forward(self, x, x_mask, g=None, reverse=False): 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 LatentEncoder(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.WaveNet( 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_lengths, g=None, tau=1.0): x_mask = torch.unsqueeze(commons.create_length_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 m, logs = torch.split(stats, self.out_channels, dim=1) z = (m + torch.randn_like(m) * tau * torch.exp(logs)) * x_mask return z, m, logs, x_mask class WaveformDecoder(torch.nn.Module): def __init__( self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0, ): super(WaveformDecoder, self).__init__() 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.ConvResBlock if resblock == "1" else modules.ConvResBlockLight 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(initialize_weights) if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1) def forward(self, x, g=None): x = self.conv_pre(x) if g is not None: x = 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) x = self.conv_post(x) x = torch.tanh(x) return x def remove_weight_norm(self): for layer in self.ups: remove_weight_norm(layer) for layer in self.resblocks: layer.remove_weight_norm() class StyleEncoder(nn.Module): def __init__(self, spec_channels, gin_channels=0, layernorm=False): super().__init__() self.spec_channels = spec_channels ref_enc_filters = [32, 32, 64, 64, 128, 128] K = len(ref_enc_filters) filters = [1] + ref_enc_filters convs = [ weight_norm( nn.Conv2d( in_channels=filters[i], out_channels=filters[i + 1], kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), ) ) for i in range(K) ] self.convs = nn.ModuleList(convs) out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K) self.gru = nn.GRU( input_size=ref_enc_filters[-1] * out_channels, hidden_size=256 // 2, batch_first=True, ) self.proj = nn.Linear(128, gin_channels) if layernorm: self.layernorm = nn.LayerNorm(self.spec_channels) else: self.layernorm = None def forward(self, inputs, mask=None): N = inputs.size(0) out = inputs.view(N, 1, -1, self.spec_channels) if self.layernorm is not None: out = self.layernorm(out) for conv in self.convs: out = conv(out) out = F.relu(out) out = out.transpose(1, 2) T = out.size(1) N = out.size(0) out = out.contiguous().view(N, T, -1) self.gru.flatten_parameters() memory, out = self.gru(out) return self.proj(out.squeeze(0)) def calculate_channels(self, L, kernel_size, stride, pad, n_convs): for i in range(n_convs): L = (L - kernel_size + 2 * pad) // stride + 1 return L class VoiceSynthesizer(nn.Module): """Voice synthesis model for inference.""" def __init__( self, n_vocab, 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=256, gin_channels=256, use_sdp=True, n_flow_layer=4, n_layers_trans_flow=6, flow_share_parameter=False, use_transformer_flow=True, use_vc=False, num_languages=None, num_tones=None, norm_refenc=False, **kwargs ): super().__init__() self.n_vocab = n_vocab 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.n_layers_trans_flow = n_layers_trans_flow self.use_spk_conditioned_encoder = kwargs.get( "use_spk_conditioned_encoder", True ) self.use_sdp = use_sdp self.use_noise_scaled_mas = kwargs.get("use_noise_scaled_mas", False) self.mas_noise_scale_initial = kwargs.get("mas_noise_scale_initial", 0.01) self.noise_scale_delta = kwargs.get("noise_scale_delta", 2e-6) self.current_mas_noise_scale = self.mas_noise_scale_initial if self.use_spk_conditioned_encoder and gin_channels > 0: self.enc_gin_channels = gin_channels else: self.enc_gin_channels = 0 self.enc_p = PhonemeEncoder( n_vocab, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, gin_channels=self.enc_gin_channels, num_languages=num_languages, num_tones=num_tones, ) self.dec = WaveformDecoder( inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, ) self.enc_q = LatentEncoder( spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels, ) if use_transformer_flow: self.flow = AttentionFlowBlock( inter_channels, hidden_channels, filter_channels, n_heads, n_layers_trans_flow, 5, p_dropout, n_flow_layer, gin_channels=gin_channels, share_parameter=flow_share_parameter, ) else: self.flow = FlowBlock( inter_channels, hidden_channels, 5, 1, n_flow_layer, gin_channels=gin_channels, ) self.sdp = VariationalDurationModel( hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels ) self.dp = DurationEstimator( hidden_channels, 256, 3, 0.5, gin_channels=gin_channels ) if n_speakers > 0: self.emb_g = nn.Embedding(n_speakers, gin_channels) else: self.ref_enc = StyleEncoder(spec_channels, gin_channels, layernorm=norm_refenc) self.use_vc = use_vc def infer( self, x, x_lengths, sid, tone, language, bert, ja_bert, noise_scale=0.667, length_scale=1, noise_scale_w=0.8, max_len=None, sdp_ratio=0, y=None, g=None, ): if g is None: if self.n_speakers > 0: g = self.emb_g(sid).unsqueeze(-1) else: g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1) if self.use_vc: g_p = None else: g_p = g x, m_p, logs_p, x_mask = self.enc_p( x, x_lengths, tone, language, bert, ja_bert, g=g_p ) logw = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w) * ( sdp_ratio ) + self.dp(x, x_mask, g=g) * (1 - sdp_ratio) 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.create_length_mask(y_lengths, None), 1).to( x_mask.dtype ) attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1) attn = commons.compute_alignment_path(w_ceil, attn_mask) m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose( 1, 2 ) logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose( 1, 2 ) 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) o = self.dec((z * y_mask)[:, :, :max_len], g=g) return o, attn, y_mask, (z, z_p, m_p, logs_p)