so-vits-svc/onnxexport/model_onnx_speaker_mix.py

import copy
import math
import torch
from torch import nn
from torch.nn import functional as F

import modules.attentions as attentions
import modules.commons as commons
import modules.modules as modules

from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm

import utils
from modules.commons import init_weights, get_padding
from utils import f0_to_coarse

class ResidualCouplingBlock(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.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers,
                                              gin_channels=gin_channels, mean_only=True))
            self.flows.append(modules.Flip())

    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 TextEncoder(nn.Module):
    def __init__(self,
                 out_channels,
                 hidden_channels,
                 kernel_size,
                 n_layers,
                 gin_channels=0,
                 filter_channels=None,
                 n_heads=None,
                 p_dropout=None):
        super().__init__()
        self.out_channels = out_channels
        self.hidden_channels = hidden_channels
        self.kernel_size = kernel_size
        self.n_layers = n_layers
        self.gin_channels = gin_channels
        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
        self.f0_emb = nn.Embedding(256, hidden_channels)

        self.enc_ = attentions.Encoder(
            hidden_channels,
            filter_channels,
            n_heads,
            n_layers,
            kernel_size,
            p_dropout)

    def forward(self, x, x_mask, f0=None, z=None):
        x = x + self.f0_emb(f0).transpose(1, 2)
        x = self.enc_(x * x_mask, x_mask)
        stats = self.proj(x) * x_mask
        m, logs = torch.split(stats, self.out_channels, dim=1)
        z = (m + z * torch.exp(logs)) * x_mask

        return z, m, logs, x_mask


class F0Decoder(nn.Module):
    def __init__(self,
                 out_channels,
                 hidden_channels,
                 filter_channels,
                 n_heads,
                 n_layers,
                 kernel_size,
                 p_dropout,
                 spk_channels=0):
        super().__init__()
        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.spk_channels = spk_channels

        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
        self.decoder = attentions.FFT(
            hidden_channels,
            filter_channels,
            n_heads,
            n_layers,
            kernel_size,
            p_dropout)
        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)

    def forward(self, x, norm_f0, x_mask, spk_emb=None):
        x = torch.detach(x)
        if (spk_emb is not None):
            x = x + self.cond(spk_emb)
        x += self.f0_prenet(norm_f0)
        x = self.prenet(x) * x_mask
        x = self.decoder(x * x_mask, x_mask)
        x = self.proj(x) * x_mask
        return x


class SynthesizerTrn(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,
                 gin_channels,
                 ssl_dim,
                 n_speakers,
                 sampling_rate=44100,
                 vol_embedding=False,
                 vocoder_name = "nsf-hifigan",
                 **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.gin_channels = gin_channels
        self.ssl_dim = ssl_dim
        self.vol_embedding = vol_embedding
        self.emb_g = nn.Embedding(n_speakers, gin_channels)
        if vol_embedding:
           self.emb_vol = nn.Linear(1, hidden_channels)

        self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)

        self.enc_p = TextEncoder(
            inter_channels,
            hidden_channels,
            filter_channels=filter_channels,
            n_heads=n_heads,
            n_layers=n_layers,
            kernel_size=kernel_size,
            p_dropout=p_dropout
        )
        hps = {
            "sampling_rate": sampling_rate,
            "inter_channels": inter_channels,
            "resblock": resblock,
            "resblock_kernel_sizes": resblock_kernel_sizes,
            "resblock_dilation_sizes": resblock_dilation_sizes,
            "upsample_rates": upsample_rates,
            "upsample_initial_channel": upsample_initial_channel,
            "upsample_kernel_sizes": upsample_kernel_sizes,
            "gin_channels": gin_channels,
        }
        
        if vocoder_name == "nsf-hifigan":
            from vdecoder.hifigan.models import Generator
            self.dec = Generator(h=hps)
        elif vocoder_name == "nsf-snake-hifigan":
            from vdecoder.hifiganwithsnake.models import Generator
            self.dec = Generator(h=hps)
        else:
            print("[?] Unkown vocoder: use default(nsf-hifigan)")
            from vdecoder.hifigan.models import Generator
            self.dec = Generator(h=hps)

        self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
        self.f0_decoder = F0Decoder(
            1,
            hidden_channels,
            filter_channels,
            n_heads,
            n_layers,
            kernel_size,
            p_dropout,
            spk_channels=gin_channels
        )
        self.emb_uv = nn.Embedding(2, hidden_channels)
        self.predict_f0 = False
        self.speaker_map = []
        self.export_mix = False

    def export_chara_mix(self, speakers_mix):
        self.speaker_map = torch.zeros((len(speakers_mix), 1, 1, self.gin_channels))
        i = 0
        for key in speakers_mix.keys():
            spkidx = speakers_mix[key]
            self.speaker_map[i] = self.emb_g(torch.LongTensor([[spkidx]]))
            i = i + 1
        self.speaker_map = self.speaker_map.unsqueeze(0)
        self.export_mix = True

    def forward(self, c, f0, mel2ph, uv, noise=None, g=None, vol = None):
        decoder_inp = F.pad(c, [0, 0, 1, 0])
        mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, c.shape[-1]])
        c = torch.gather(decoder_inp, 1, mel2ph_).transpose(1, 2)  # [B, T, H]

        if self.export_mix:   # [N, S]  *  [S, B, 1, H]
            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
            g = g * self.speaker_map  # [N, S, B, 1, H]
            g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
            g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
        else:
            if g.dim() == 1:
                g = g.unsqueeze(0)
            g = self.emb_g(g).transpose(1, 2)
        
        x_mask = torch.unsqueeze(torch.ones_like(f0), 1).to(c.dtype)
        # vol proj
        vol = self.emb_vol(vol[:,:,None]).transpose(1,2) if vol!=None and self.vol_embedding else 0
        
        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2) + vol
        
        z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), z=noise)
        z = self.flow(z_p, c_mask, g=g, reverse=True)
        o = self.dec(z * c_mask, g=g, f0=f0)
        return o