goathead777/Zero_Shot_Inference

1import math2import numpy as np3import torch4from torch import nn5from torch.nn import functional as F6 7from torch.nn import Conv1d8from torch.nn.utils import weight_norm, remove_weight_norm9 10from module import commons11from module.commons import init_weights, get_padding12from module.transforms import piecewise_rational_quadratic_transform13import torch.distributions as D14 15 16LRELU_SLOPE = 0.117 18 19class LayerNorm(nn.Module):20    def __init__(self, channels, eps=1e-5):21        super().__init__()22        self.channels = channels23        self.eps = eps24 25        self.gamma = nn.Parameter(torch.ones(channels))26        self.beta = nn.Parameter(torch.zeros(channels))27 28    def forward(self, x):29        x = x.transpose(1, -1)30        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)31        return x.transpose(1, -1)32 33 34class ConvReluNorm(nn.Module):35    def __init__(36        self,37        in_channels,38        hidden_channels,39        out_channels,40        kernel_size,41        n_layers,42        p_dropout,43    ):44        super().__init__()45        self.in_channels = in_channels46        self.hidden_channels = hidden_channels47        self.out_channels = out_channels48        self.kernel_size = kernel_size49        self.n_layers = n_layers50        self.p_dropout = p_dropout51        assert n_layers > 1, "Number of layers should be larger than 0."52 53        self.conv_layers = nn.ModuleList()54        self.norm_layers = nn.ModuleList()55        self.conv_layers.append(56            nn.Conv1d(57                in_channels, hidden_channels, kernel_size, padding=kernel_size // 258            )59        )60        self.norm_layers.append(LayerNorm(hidden_channels))61        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))62        for _ in range(n_layers - 1):63            self.conv_layers.append(64                nn.Conv1d(65                    hidden_channels,66                    hidden_channels,67                    kernel_size,68                    padding=kernel_size // 2,69                )70            )71            self.norm_layers.append(LayerNorm(hidden_channels))72        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)73        self.proj.weight.data.zero_()74        self.proj.bias.data.zero_()75 76    def forward(self, x, x_mask):77        x_org = x78        for i in range(self.n_layers):79            x = self.conv_layers[i](x * x_mask)80            x = self.norm_layers[i](x)81            x = self.relu_drop(x)82        x = x_org + self.proj(x)83        return x * x_mask84 85 86class DDSConv(nn.Module):87    """88    Dialted and Depth-Separable Convolution89    """90 91    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):92        super().__init__()93        self.channels = channels94        self.kernel_size = kernel_size95        self.n_layers = n_layers96        self.p_dropout = p_dropout97 98        self.drop = nn.Dropout(p_dropout)99        self.convs_sep = nn.ModuleList()100        self.convs_1x1 = nn.ModuleList()101        self.norms_1 = nn.ModuleList()102        self.norms_2 = nn.ModuleList()103        for i in range(n_layers):104            dilation = kernel_size**i105            padding = (kernel_size * dilation - dilation) // 2106            self.convs_sep.append(107                nn.Conv1d(108                    channels,109                    channels,110                    kernel_size,111                    groups=channels,112                    dilation=dilation,113                    padding=padding,114                )115            )116            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))117            self.norms_1.append(LayerNorm(channels))118            self.norms_2.append(LayerNorm(channels))119 120    def forward(self, x, x_mask, g=None):121        if g is not None:122            x = x + g123        for i in range(self.n_layers):124            y = self.convs_sep[i](x * x_mask)125            y = self.norms_1[i](y)126            y = F.gelu(y)127            y = self.convs_1x1[i](y)128            y = self.norms_2[i](y)129            y = F.gelu(y)130            y = self.drop(y)131            x = x + y132        return x * x_mask133 134 135class WN(torch.nn.Module):136    def __init__(137        self,138        hidden_channels,139        kernel_size,140        dilation_rate,141        n_layers,142        gin_channels=0,143        p_dropout=0,144    ):145        super(WN, self).__init__()146        assert kernel_size % 2 == 1147        self.hidden_channels = hidden_channels148        self.kernel_size = (kernel_size,)149        self.dilation_rate = dilation_rate150        self.n_layers = n_layers151        self.gin_channels = gin_channels152        self.p_dropout = p_dropout153 154        self.in_layers = torch.nn.ModuleList()155        self.res_skip_layers = torch.nn.ModuleList()156        self.drop = nn.Dropout(p_dropout)157 158        if gin_channels != 0:159            cond_layer = torch.nn.Conv1d(160                gin_channels, 2 * hidden_channels * n_layers, 1161            )162            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")163 164        for i in range(n_layers):165            dilation = dilation_rate**i166            padding = int((kernel_size * dilation - dilation) / 2)167            in_layer = torch.nn.Conv1d(168                hidden_channels,169                2 * hidden_channels,170                kernel_size,171                dilation=dilation,172                padding=padding,173            )174            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")175            self.in_layers.append(in_layer)176 177            # last one is not necessary178            if i < n_layers - 1:179                res_skip_channels = 2 * hidden_channels180            else:181                res_skip_channels = hidden_channels182 183            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)184            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")185            self.res_skip_layers.append(res_skip_layer)186 187    def forward(self, x, x_mask, g=None, **kwargs):188        output = torch.zeros_like(x)189        n_channels_tensor = torch.IntTensor([self.hidden_channels])190 191        if g is not None:192            g = self.cond_layer(g)193 194        for i in range(self.n_layers):195            x_in = self.in_layers[i](x)196            if g is not None:197                cond_offset = i * 2 * self.hidden_channels198                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]199            else:200                g_l = torch.zeros_like(x_in)201 202            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)203            acts = self.drop(acts)204 205            res_skip_acts = self.res_skip_layers[i](acts)206            if i < self.n_layers - 1:207                res_acts = res_skip_acts[:, : self.hidden_channels, :]208                x = (x + res_acts) * x_mask209                output = output + res_skip_acts[:, self.hidden_channels :, :]210            else:211                output = output + res_skip_acts212        return output * x_mask213 214    def remove_weight_norm(self):215        if self.gin_channels != 0:216            torch.nn.utils.remove_weight_norm(self.cond_layer)217        for l in self.in_layers:218            torch.nn.utils.remove_weight_norm(l)219        for l in self.res_skip_layers:220            torch.nn.utils.remove_weight_norm(l)221 222 223class ResBlock1(torch.nn.Module):224    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):225        super(ResBlock1, self).__init__()226        self.convs1 = nn.ModuleList(227            [228                weight_norm(229                    Conv1d(230                        channels,231                        channels,232                        kernel_size,233                        1,234                        dilation=dilation[0],235                        padding=get_padding(kernel_size, dilation[0]),236                    )237                ),238                weight_norm(239                    Conv1d(240                        channels,241                        channels,242                        kernel_size,243                        1,244                        dilation=dilation[1],245                        padding=get_padding(kernel_size, dilation[1]),246                    )247                ),248                weight_norm(249                    Conv1d(250                        channels,251                        channels,252                        kernel_size,253                        1,254                        dilation=dilation[2],255                        padding=get_padding(kernel_size, dilation[2]),256                    )257                ),258            ]259        )260        self.convs1.apply(init_weights)261 262        self.convs2 = nn.ModuleList(263            [264                weight_norm(265                    Conv1d(266                        channels,267                        channels,268                        kernel_size,269                        1,270                        dilation=1,271                        padding=get_padding(kernel_size, 1),272                    )273                ),274                weight_norm(275                    Conv1d(276                        channels,277                        channels,278                        kernel_size,279                        1,280                        dilation=1,281                        padding=get_padding(kernel_size, 1),282                    )283                ),284                weight_norm(285                    Conv1d(286                        channels,287                        channels,288                        kernel_size,289                        1,290                        dilation=1,291                        padding=get_padding(kernel_size, 1),292                    )293                ),294            ]295        )296        self.convs2.apply(init_weights)297 298    def forward(self, x, x_mask=None):299        for c1, c2 in zip(self.convs1, self.convs2):300            xt = F.leaky_relu(x, LRELU_SLOPE)301            if x_mask is not None:302                xt = xt * x_mask303            xt = c1(xt)304            xt = F.leaky_relu(xt, LRELU_SLOPE)305            if x_mask is not None:306                xt = xt * x_mask307            xt = c2(xt)308            x = xt + x309        if x_mask is not None:310            x = x * x_mask311        return x312 313    def remove_weight_norm(self):314        for l in self.convs1:315            remove_weight_norm(l)316        for l in self.convs2:317            remove_weight_norm(l)318 319 320class ResBlock2(torch.nn.Module):321    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):322        super(ResBlock2, self).__init__()323        self.convs = nn.ModuleList(324            [325                weight_norm(326                    Conv1d(327                        channels,328                        channels,329                        kernel_size,330                        1,331                        dilation=dilation[0],332                        padding=get_padding(kernel_size, dilation[0]),333                    )334                ),335                weight_norm(336                    Conv1d(337                        channels,338                        channels,339                        kernel_size,340                        1,341                        dilation=dilation[1],342                        padding=get_padding(kernel_size, dilation[1]),343                    )344                ),345            ]346        )347        self.convs.apply(init_weights)348 349    def forward(self, x, x_mask=None):350        for c in self.convs:351            xt = F.leaky_relu(x, LRELU_SLOPE)352            if x_mask is not None:353                xt = xt * x_mask354            xt = c(xt)355            x = xt + x356        if x_mask is not None:357            x = x * x_mask358        return x359 360    def remove_weight_norm(self):361        for l in self.convs:362            remove_weight_norm(l)363 364 365class Log(nn.Module):366    def forward(self, x, x_mask, reverse=False, **kwargs):367        if not reverse:368            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask369            logdet = torch.sum(-y, [1, 2])370            return y, logdet371        else:372            x = torch.exp(x) * x_mask373            return x374 375 376class Flip(nn.Module):377    def forward(self, x, *args, reverse=False, **kwargs):378        x = torch.flip(x, [1])379        if not reverse:380            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)381            return x, logdet382        else:383            return x384 385 386class ElementwiseAffine(nn.Module):387    def __init__(self, channels):388        super().__init__()389        self.channels = channels390        self.m = nn.Parameter(torch.zeros(channels, 1))391        self.logs = nn.Parameter(torch.zeros(channels, 1))392 393    def forward(self, x, x_mask, reverse=False, **kwargs):394        if not reverse:395            y = self.m + torch.exp(self.logs) * x396            y = y * x_mask397            logdet = torch.sum(self.logs * x_mask, [1, 2])398            return y, logdet399        else:400            x = (x - self.m) * torch.exp(-self.logs) * x_mask401            return x402 403 404class ResidualCouplingLayer(nn.Module):405    def __init__(406        self,407        channels,408        hidden_channels,409        kernel_size,410        dilation_rate,411        n_layers,412        p_dropout=0,413        gin_channels=0,414        mean_only=False,415    ):416        assert channels % 2 == 0, "channels should be divisible by 2"417        super().__init__()418        self.channels = channels419        self.hidden_channels = hidden_channels420        self.kernel_size = kernel_size421        self.dilation_rate = dilation_rate422        self.n_layers = n_layers423        self.half_channels = channels // 2424        self.mean_only = mean_only425 426        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)427        self.enc = WN(428            hidden_channels,429            kernel_size,430            dilation_rate,431            n_layers,432            p_dropout=p_dropout,433            gin_channels=gin_channels,434        )435        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)436        self.post.weight.data.zero_()437        self.post.bias.data.zero_()438 439    def forward(self, x, x_mask, g=None, reverse=False):440        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)441        h = self.pre(x0) * x_mask442        h = self.enc(h, x_mask, g=g)443        stats = self.post(h) * x_mask444        if not self.mean_only:445            m, logs = torch.split(stats, [self.half_channels] * 2, 1)446        else:447            m = stats448            logs = torch.zeros_like(m)449 450        if not reverse:451            x1 = m + x1 * torch.exp(logs) * x_mask452            x = torch.cat([x0, x1], 1)453            logdet = torch.sum(logs, [1, 2])454            return x, logdet455        else:456            x1 = (x1 - m) * torch.exp(-logs) * x_mask457            x = torch.cat([x0, x1], 1)458            return x459 460 461class ConvFlow(nn.Module):462    def __init__(463        self,464        in_channels,465        filter_channels,466        kernel_size,467        n_layers,468        num_bins=10,469        tail_bound=5.0,470    ):471        super().__init__()472        self.in_channels = in_channels473        self.filter_channels = filter_channels474        self.kernel_size = kernel_size475        self.n_layers = n_layers476        self.num_bins = num_bins477        self.tail_bound = tail_bound478        self.half_channels = in_channels // 2479 480        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)481        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)482        self.proj = nn.Conv1d(483            filter_channels, self.half_channels * (num_bins * 3 - 1), 1484        )485        self.proj.weight.data.zero_()486        self.proj.bias.data.zero_()487 488    def forward(self, x, x_mask, g=None, reverse=False):489        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)490        h = self.pre(x0)491        h = self.convs(h, x_mask, g=g)492        h = self.proj(h) * x_mask493 494        b, c, t = x0.shape495        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]496 497        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)498        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(499            self.filter_channels500        )501        unnormalized_derivatives = h[..., 2 * self.num_bins :]502 503        x1, logabsdet = piecewise_rational_quadratic_transform(504            x1,505            unnormalized_widths,506            unnormalized_heights,507            unnormalized_derivatives,508            inverse=reverse,509            tails="linear",510            tail_bound=self.tail_bound,511        )512 513        x = torch.cat([x0, x1], 1) * x_mask514        logdet = torch.sum(logabsdet * x_mask, [1, 2])515        if not reverse:516            return x, logdet517        else:518            return x519 520 521class LinearNorm(nn.Module):522    def __init__(523        self,524        in_channels,525        out_channels,526        bias=True,527        spectral_norm=False,528    ):529        super(LinearNorm, self).__init__()530        self.fc = nn.Linear(in_channels, out_channels, bias)531 532        if spectral_norm:533            self.fc = nn.utils.spectral_norm(self.fc)534 535    def forward(self, input):536        out = self.fc(input)537        return out538 539 540class Mish(nn.Module):541    def __init__(self):542        super(Mish, self).__init__()543 544    def forward(self, x):545        return x * torch.tanh(F.softplus(x))546 547 548class Conv1dGLU(nn.Module):549    """550    Conv1d + GLU(Gated Linear Unit) with residual connection.551    For GLU refer to https://arxiv.org/abs/1612.08083 paper.552    """553 554    def __init__(self, in_channels, out_channels, kernel_size, dropout):555        super(Conv1dGLU, self).__init__()556        self.out_channels = out_channels557        self.conv1 = ConvNorm(in_channels, 2 * out_channels, kernel_size=kernel_size)558        self.dropout = nn.Dropout(dropout)559 560    def forward(self, x):561        residual = x562        x = self.conv1(x)563        x1, x2 = torch.split(x, split_size_or_sections=self.out_channels, dim=1)564        x = x1 * torch.sigmoid(x2)565        x = residual + self.dropout(x)566        return x567 568 569class ConvNorm(nn.Module):570    def __init__(571        self,572        in_channels,573        out_channels,574        kernel_size=1,575        stride=1,576        padding=None,577        dilation=1,578        bias=True,579        spectral_norm=False,580    ):581        super(ConvNorm, self).__init__()582 583        if padding is None:584            assert kernel_size % 2 == 1585            padding = int(dilation * (kernel_size - 1) / 2)586 587        self.conv = torch.nn.Conv1d(588            in_channels,589            out_channels,590            kernel_size=kernel_size,591            stride=stride,592            padding=padding,593            dilation=dilation,594            bias=bias,595        )596 597        if spectral_norm:598            self.conv = nn.utils.spectral_norm(self.conv)599 600    def forward(self, input):601        out = self.conv(input)602        return out603 604 605class MultiHeadAttention(nn.Module):606    """Multi-Head Attention module"""607 608    def __init__(self, n_head, d_model, d_k, d_v, dropout=0.0, spectral_norm=False):609        super().__init__()610 611        self.n_head = n_head612        self.d_k = d_k613        self.d_v = d_v614 615        self.w_qs = nn.Linear(d_model, n_head * d_k)616        self.w_ks = nn.Linear(d_model, n_head * d_k)617        self.w_vs = nn.Linear(d_model, n_head * d_v)618 619        self.attention = ScaledDotProductAttention(620            temperature=np.power(d_model, 0.5), dropout=dropout621        )622 623        self.fc = nn.Linear(n_head * d_v, d_model)624        self.dropout = nn.Dropout(dropout)625 626        if spectral_norm:627            self.w_qs = nn.utils.spectral_norm(self.w_qs)628            self.w_ks = nn.utils.spectral_norm(self.w_ks)629            self.w_vs = nn.utils.spectral_norm(self.w_vs)630            self.fc = nn.utils.spectral_norm(self.fc)631 632    def forward(self, x, mask=None):633        d_k, d_v, n_head = self.d_k, self.d_v, self.n_head634        sz_b, len_x, _ = x.size()635 636        residual = x637 638        q = self.w_qs(x).view(sz_b, len_x, n_head, d_k)639        k = self.w_ks(x).view(sz_b, len_x, n_head, d_k)640        v = self.w_vs(x).view(sz_b, len_x, n_head, d_v)641        q = q.permute(2, 0, 1, 3).contiguous().view(-1, len_x, d_k)  # (n*b) x lq x dk642        k = k.permute(2, 0, 1, 3).contiguous().view(-1, len_x, d_k)  # (n*b) x lk x dk643        v = v.permute(2, 0, 1, 3).contiguous().view(-1, len_x, d_v)  # (n*b) x lv x dv644 645        if mask is not None:646            slf_mask = mask.repeat(n_head, 1, 1)  # (n*b) x .. x ..647        else:648            slf_mask = None649        output, attn = self.attention(q, k, v, mask=slf_mask)650 651        output = output.view(n_head, sz_b, len_x, d_v)652        output = (653            output.permute(1, 2, 0, 3).contiguous().view(sz_b, len_x, -1)654        )  # b x lq x (n*dv)655 656        output = self.fc(output)657 658        output = self.dropout(output) + residual659        return output, attn660 661 662class ScaledDotProductAttention(nn.Module):663    """Scaled Dot-Product Attention"""664 665    def __init__(self, temperature, dropout):666        super().__init__()667        self.temperature = temperature668        self.softmax = nn.Softmax(dim=2)669        self.dropout = nn.Dropout(dropout)670 671    def forward(self, q, k, v, mask=None):672        attn = torch.bmm(q, k.transpose(1, 2))673        attn = attn / self.temperature674 675        if mask is not None:676            attn = attn.masked_fill(mask, -np.inf)677 678        attn = self.softmax(attn)679        p_attn = self.dropout(attn)680 681        output = torch.bmm(p_attn, v)682        return output, attn683 684 685class MelStyleEncoder(nn.Module):686    """MelStyleEncoder"""687 688    def __init__(689        self,690        n_mel_channels=80,691        style_hidden=128,692        style_vector_dim=256,693        style_kernel_size=5,694        style_head=2,695        dropout=0.1,696    ):697        super(MelStyleEncoder, self).__init__()698        self.in_dim = n_mel_channels699        self.hidden_dim = style_hidden700        self.out_dim = style_vector_dim701        self.kernel_size = style_kernel_size702        self.n_head = style_head703        self.dropout = dropout704 705        self.spectral = nn.Sequential(706            LinearNorm(self.in_dim, self.hidden_dim),707            Mish(),708            nn.Dropout(self.dropout),709            LinearNorm(self.hidden_dim, self.hidden_dim),710            Mish(),711            nn.Dropout(self.dropout),712        )713 714        self.temporal = nn.Sequential(715            Conv1dGLU(self.hidden_dim, self.hidden_dim, self.kernel_size, self.dropout),716            Conv1dGLU(self.hidden_dim, self.hidden_dim, self.kernel_size, self.dropout),717        )718 719        self.slf_attn = MultiHeadAttention(720            self.n_head,721            self.hidden_dim,722            self.hidden_dim // self.n_head,723            self.hidden_dim // self.n_head,724            self.dropout,725        )726 727        self.fc = LinearNorm(self.hidden_dim, self.out_dim)728 729    def temporal_avg_pool(self, x, mask=None):730        if mask is None:731            out = torch.mean(x, dim=1)732        else:733            len_ = (~mask).sum(dim=1).unsqueeze(1)734            x = x.masked_fill(mask.unsqueeze(-1), 0)735            x = x.sum(dim=1)736            out = torch.div(x, len_)737        return out738 739    def forward(self, x, mask=None):740        x = x.transpose(1, 2)741        if mask is not None:742            mask = (mask.int() == 0).squeeze(1)743        max_len = x.shape[1]744        slf_attn_mask = (745            mask.unsqueeze(1).expand(-1, max_len, -1) if mask is not None else None746        )747 748        # spectral749        x = self.spectral(x)750        # temporal751        x = x.transpose(1, 2)752        x = self.temporal(x)753        x = x.transpose(1, 2)754        # self-attention755        if mask is not None:756            x = x.masked_fill(mask.unsqueeze(-1), 0)757        x, _ = self.slf_attn(x, mask=slf_attn_mask)758        # fc759        x = self.fc(x)760        # temoral average pooling761        w = self.temporal_avg_pool(x, mask=mask)762 763        return w.unsqueeze(-1)764 765 766class MelStyleEncoderVAE(nn.Module):767    def __init__(self, spec_channels, z_latent_dim, emb_dim):768        super().__init__()769        self.ref_encoder = MelStyleEncoder(spec_channels, style_vector_dim=emb_dim)770        self.fc1 = nn.Linear(emb_dim, z_latent_dim)771        self.fc2 = nn.Linear(emb_dim, z_latent_dim)772        self.fc3 = nn.Linear(z_latent_dim, emb_dim)773        self.z_latent_dim = z_latent_dim774 775    def reparameterize(self, mu, logvar):776        if self.training:777            std = torch.exp(0.5 * logvar)778            eps = torch.randn_like(std)779            return eps.mul(std).add_(mu)780        else:781            return mu782 783    def forward(self, inputs, mask=None):784        enc_out = self.ref_encoder(inputs.squeeze(-1), mask).squeeze(-1)785        mu = self.fc1(enc_out)786        logvar = self.fc2(enc_out)787        posterior = D.Normal(mu, torch.exp(logvar))788        kl_divergence = D.kl_divergence(789            posterior, D.Normal(torch.zeros_like(mu), torch.ones_like(logvar))790        )791        loss_kl = kl_divergence.mean()792 793        z = posterior.rsample()794        style_embed = self.fc3(z)795 796        return style_embed.unsqueeze(-1), loss_kl797 798    def infer(self, inputs=None, random_sample=False, manual_latent=None):799        if manual_latent is None:800            if random_sample:801                dev = next(self.parameters()).device802                posterior = D.Normal(803                    torch.zeros(1, self.z_latent_dim, device=dev),804                    torch.ones(1, self.z_latent_dim, device=dev),805                )806                z = posterior.rsample()807            else:808                enc_out = self.ref_encoder(inputs.transpose(1, 2))809                mu = self.fc1(enc_out)810                z = mu811        else:812            z = manual_latent813        style_embed = self.fc3(z)814        return style_embed.unsqueeze(-1), z815 816 817class ActNorm(nn.Module):818    def __init__(self, channels, ddi=False, **kwargs):819        super().__init__()820        self.channels = channels821        self.initialized = not ddi822 823        self.logs = nn.Parameter(torch.zeros(1, channels, 1))824        self.bias = nn.Parameter(torch.zeros(1, channels, 1))825 826    def forward(self, x, x_mask=None, g=None, reverse=False, **kwargs):827        if x_mask is None:828            x_mask = torch.ones(x.size(0), 1, x.size(2)).to(829                device=x.device, dtype=x.dtype830            )831        x_len = torch.sum(x_mask, [1, 2])832        if not self.initialized:833            self.initialize(x, x_mask)834            self.initialized = True835 836        if reverse:837            z = (x - self.bias) * torch.exp(-self.logs) * x_mask838            logdet = None839            return z840        else:841            z = (self.bias + torch.exp(self.logs) * x) * x_mask842            logdet = torch.sum(self.logs) * x_len  # [b]843            return z, logdet844 845    def store_inverse(self):846        pass847 848    def set_ddi(self, ddi):849        self.initialized = not ddi850 851    def initialize(self, x, x_mask):852        with torch.no_grad():853            denom = torch.sum(x_mask, [0, 2])854            m = torch.sum(x * x_mask, [0, 2]) / denom855            m_sq = torch.sum(x * x * x_mask, [0, 2]) / denom856            v = m_sq - (m**2)857            logs = 0.5 * torch.log(torch.clamp_min(v, 1e-6))858 859            bias_init = (860                (-m * torch.exp(-logs)).view(*self.bias.shape).to(dtype=self.bias.dtype)861            )862            logs_init = (-logs).view(*self.logs.shape).to(dtype=self.logs.dtype)863 864            self.bias.data.copy_(bias_init)865            self.logs.data.copy_(logs_init)866 867 868class InvConvNear(nn.Module):869    def __init__(self, channels, n_split=4, no_jacobian=False, **kwargs):870        super().__init__()871        assert n_split % 2 == 0872        self.channels = channels873        self.n_split = n_split874        self.no_jacobian = no_jacobian875 876        w_init = torch.linalg.qr(877            torch.FloatTensor(self.n_split, self.n_split).normal_()878        )[0]879        if torch.det(w_init) < 0:880            w_init[:, 0] = -1 * w_init[:, 0]881        self.weight = nn.Parameter(w_init)882 883    def forward(self, x, x_mask=None, g=None, reverse=False, **kwargs):884        b, c, t = x.size()885        assert c % self.n_split == 0886        if x_mask is None:887            x_mask = 1888            x_len = torch.ones((b,), dtype=x.dtype, device=x.device) * t889        else:890            x_len = torch.sum(x_mask, [1, 2])891 892        x = x.view(b, 2, c // self.n_split, self.n_split // 2, t)893        x = (894            x.permute(0, 1, 3, 2, 4)895            .contiguous()896            .view(b, self.n_split, c // self.n_split, t)897        )898 899        if reverse:900            if hasattr(self, "weight_inv"):901                weight = self.weight_inv902            else:903                weight = torch.inverse(self.weight.float()).to(dtype=self.weight.dtype)904            logdet = None905        else:906            weight = self.weight907            if self.no_jacobian:908                logdet = 0909            else:910                logdet = torch.logdet(self.weight) * (c / self.n_split) * x_len  # [b]911 912        weight = weight.view(self.n_split, self.n_split, 1, 1)913        z = F.conv2d(x, weight)914 915        z = z.view(b, 2, self.n_split // 2, c // self.n_split, t)916        z = z.permute(0, 1, 3, 2, 4).contiguous().view(b, c, t) * x_mask917        if reverse:918            return z919        else:920            return z, logdet921 922    def store_inverse(self):923        self.weight_inv = torch.inverse(self.weight.float()).to(dtype=self.weight.dtype)924