models_mel_style.py

import copy
import math
import torch
from torch import nn
from torch.nn import functional as F
import os
import yaml
import commons
import modules
import attentions
import monotonic_align
import numpy as np
from mel_processing import mel_spectrogram_torch, spec_to_mel_torch, spectrogram_torch


from Attention import MultiHeadedAttention as BaseMultiHeadedAttention
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
from commons import init_weights, get_padding
from transformers import AlbertConfig, AlbertModel
from collections import OrderedDict
from text import sequence_to_text
import utils


log_dir = "configs"
config_path = os.path.join(log_dir, "vie_bert.yml")
plbert_config = yaml.safe_load(open(config_path))


# hps = utils.get_hparams()


class StochasticDurationPredictor(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 # it needs to be removed from future version.
    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.Log()
    self.flows = nn.ModuleList()
    self.flows.append(modules.ElementwiseAffine(2))
    for i in range(n_flows):
      self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
      self.flows.append(modules.Flip())

    self.post_pre = nn.Conv1d(1, filter_channels, 1)
    self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
    self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
    self.post_flows = nn.ModuleList()
    self.post_flows.append(modules.ElementwiseAffine(2))
    for i in range(4):
      self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
      self.post_flows.append(modules.Flip())

    self.pre = nn.Conv1d(in_channels, filter_channels, 1)
    self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
    self.convs = modules.DDSConv(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 # [b]
    else:
      flows = list(reversed(self.flows))
      flows = flows[:-2] + [flows[-1]] # remove a useless vflow
      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 DurationPredictor(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.LayerNorm(filter_channels)
    self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
    self.norm_2 = modules.LayerNorm(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


def length_to_mask(lengths):
    #print(lengths.max(),'final')
    mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
    mask = torch.gt(mask+1, lengths.unsqueeze(1))
    return mask

class TextEncoder(nn.Module):
  def __init__(self,
      n_vocab,
      out_channels,
      hidden_channels,
      filter_channels,
      n_heads,
      n_layers,
      kernel_size,
      p_dropout):
    super().__init__()
    
    self.out_channels = out_channels
    #self.hidden_channels = hidden_channels
    #self.p_dropout = p_dropout

    #self.emb = nn.Embedding(n_vocab, hidden_channels)
    #nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)

    self.encoder = attentions.Encoder(
      hidden_channels,
      filter_channels,
      n_heads,
      n_layers,
      kernel_size,
      p_dropout)
    
    albert_base_configuration = AlbertConfig(**plbert_config['model_params'])
    bert = AlbertModel(albert_base_configuration)
    

    
    
    # checkpoint = torch.load(log_dir + "/step_1000000"  + ".t7", map_location='cpu')
    # state_dict = checkpoint['net']

    checkpoint = torch.load(log_dir + "/bert_" + "5" + ".pt")
    state_dict = checkpoint
    new_state_dict = OrderedDict()
    for k, v in state_dict.items():
      name = k[7:] # remove `module.`
      if name.startswith('encoder.'):
          name = name[8:] # remove `encoder.`
          new_state_dict[name] = v
    #print(new_state_dict)
    bert.load_state_dict(new_state_dict, strict=False)
    
    
   # self.bert = bert.to('cuda') 
    self.bert = bert # no-gpu-inference
    
    # print(self.bert.pooler.weight.requires_grad)
    # print(self.bert.pooler.bias.requires_grad)
    # for param in self.bert.pooler.weight.parameters():
    #   param.requires_grad = True # or True
    
    # for param in self.bert.pooler.bias.parameters():
    #   param.requires_grad = True # or True
    
    self.linear = nn.Linear(plbert_config['model_params']['hidden_size'], hidden_channels)
    self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
    
    
   
    

  def forward(self, x, x_lengths):
    #print(x, x_lengths, 'test2')
    
    attention_mask = length_to_mask(torch.Tensor(x_lengths))
    #print(len(x[0]), len(attention_mask[0]), 'test3')
    #print((~attention_mask).int())
    # print(self.bert(x, attention_mask=(~attention_mask).int()))
    x = self.bert(x, attention_mask=(~attention_mask).int()).last_hidden_state # [b, t, h1]
  

    
    x = self.linear(x)
    #x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
    x = torch.transpose(x, 1, -1) # [b, h, t]
    
    #x_mask = torch.gt(torch.arange(torch.Tensor(x_lengths).max()).unsqueeze(0).expand(torch.Tensor(x_lengths).shape[0], -1).type_as(torch.Tensor(x_lengths))+1, torch.Tensor(x_lengths).unsqueeze(1)).int()
    x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
    #print(x_mask)

    x = self.encoder(x * x_mask, x_mask)
    stats = self.proj(x) * x_mask

    m, logs = torch.split(stats, self.out_channels, dim=1)
    return x, m, logs, x_mask


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 PosteriorEncoder(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.WN(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):
    x_mask = torch.unsqueeze(commons.sequence_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) * torch.exp(logs)) * x_mask
    return z, m, logs, x_mask


class Generator(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(Generator, 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.ResBlock1 if resblock == '1' else modules.ResBlock2

        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(init_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):
        print('Removing weight norm...')
        for l in self.ups:
            remove_weight_norm(l)
        for l in self.resblocks:
            l.remove_weight_norm()


class DiscriminatorP(torch.nn.Module):
    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
        super(DiscriminatorP, self).__init__()
        self.period = period
        self.use_spectral_norm = use_spectral_norm
        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
        self.convs = nn.ModuleList([
            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
        ])
        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))

    def forward(self, x):
        fmap = []

        # 1d to 2d
        b, c, t = x.shape
        if t % self.period != 0: # pad first
            n_pad = self.period - (t % self.period)
            x = F.pad(x, (0, n_pad), "reflect")
            t = t + n_pad
        x = x.view(b, c, t // self.period, self.period)

        for l in self.convs:
            x = l(x)
            x = F.leaky_relu(x, modules.LRELU_SLOPE)
            fmap.append(x)
        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)

        return x, fmap


class DiscriminatorS(torch.nn.Module):
    def __init__(self, use_spectral_norm=False):
        super(DiscriminatorS, self).__init__()
        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
        self.convs = nn.ModuleList([
            norm_f(Conv1d(1, 16, 15, 1, padding=7)),
            norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
            norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
            norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
            norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
        ])
        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))

    def forward(self, x):
        fmap = []

        for l in self.convs:
            x = l(x)
            x = F.leaky_relu(x, modules.LRELU_SLOPE)
            fmap.append(x)
        x = self.conv_post(x)
        fmap.append(x)
        x = torch.flatten(x, 1, -1)

        return x, fmap


class MultiPeriodDiscriminator(torch.nn.Module):
    def __init__(self, use_spectral_norm=False):
        super(MultiPeriodDiscriminator, self).__init__()
        periods = [2,3,5,7,11]

        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
        discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
        
        self.discriminators = nn.ModuleList(discs)
        

    def forward(self, y, y_hat):
        y_d_rs = []
        y_d_gs = []
        fmap_rs = []
        fmap_gs = []
        for i, d in enumerate(self.discriminators):
            y_d_r, fmap_r = d(y)
            y_d_g, fmap_g = d(y_hat)
            y_d_rs.append(y_d_r)
            y_d_gs.append(y_d_g)
            fmap_rs.append(fmap_r)
            fmap_gs.append(fmap_g)
            
        return y_d_rs, y_d_gs, fmap_rs, fmap_gs



class SynthesizerTrn(nn.Module):
  """
  Synthesizer for Training
  """

  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=0,
    gin_channels=0,
    use_sdp=True,
    **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.use_sdp = use_sdp

    self.enc_p = TextEncoder(n_vocab,
        inter_channels,
        hidden_channels,
        filter_channels,
        n_heads,
        n_layers,
        kernel_size,
        p_dropout)
    self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
    self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
    self.style_encoder = StyleEmbedding()
    if use_sdp:
      self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
    else:
      self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)

    if n_speakers > 1:
      self.emb_g = nn.Embedding(n_speakers, gin_channels)

  def forward(self, x, x_lengths, mel, y, y_lengths, sid=None):
    '''
    set g = None for posterior enc, sdp(dp), vocoder except flow
    '''
    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)

    
    if self.n_speakers > 0:
      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
    else:
      g = None
    #* g: (8,256,1)
    # z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
    z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=None)
    #* y_mask:(8,1,262)
    # z_p = self.flow(z, y_mask, g=g)
    #* Zero-shot
    style_vector = self.style_encoder(mel.transpose(1,2), torch.tensor(np.full((mel.shape[0]), mel.shape[2])))
    z_p = self.flow(z, y_mask, g=style_vector.unsqueeze(-1))

    with torch.no_grad():
      # negative cross-entropy
      s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
      neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
      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]
      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]
      neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
      neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4

      attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
      attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
    
    w = attn.sum(2)
    if self.use_sdp:
      # l_length = self.dp(x, x_mask, w, g=g)
      l_length = self.dp(x, x_mask, w, g=None)
      l_length = l_length / torch.sum(x_mask)
    else:
      logw_ = torch.log(w + 1e-6) * x_mask
      # logw = self.dp(x, x_mask, g=g)
      logw = self.dp(x, x_mask, g=None)
      l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging 

    # expand prior
    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_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
    # o = self.dec(z_slice, g=g)
    o = self.dec(z_slice, g=None)
    return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
#   def forward(self, mel_src, mel_tgt, y, y_lengths, y_ref, y_lengths_ref, sid=None):
        
#         style_vector_src = self.style_encoder(mel_src.transpose(1,2), torch.tensor(np.full((mel_src.shape[0]), mel_src.shape[2])))
#         style_vector_ref = self.style_encoder(mel_tgt.transpose(1,2), torch.tensor(np.full((mel_tgt.shape[0]), mel_tgt.shape[2])))
#         ## SRC
#         z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=None)
#         z_p = self.flow(z, y_mask, g=style_vector_src.unsqueeze(-1))
#         z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
        
#         ## REF
#         z_ref, z_q_ref, logs_q_ref, y_mask_ref = self.enc_q(y_ref, y_lengths_ref, g=None)
#         z_p_ref = self.flow(z_ref, y_mask_ref, g=style_vector_ref.unsqueeze(-1))
#         z_slice_ref, ids_slice_ref = commons.rand_slice_segments(z_ref, y_lengths_ref, self.segment_size)
        
        
#         o = self.dec(z_slice, g=None)
#         o_ref = self.dec(z_slice_ref, g=None)
        
        
#         ## Style reconstruction
#         z_vc = self.flow(z_p, y_mask, g=style_vector_ref.unsqueeze(-1), reverse=True)
#         o_hat_vc = self.dec(z_vc * y_mask, g=None)
#         o_hat_vc_mel = mel_spectrogram_torch(
#           o_hat_vc.squeeze(1), 
#           hps.data.filter_length, 
#           hps.data.n_mel_channels, 
#           hps.data.sampling_rate, 
#           hps.data.hop_length, 
#           hps.data.win_length, 
#           hps.data.mel_fmin, 
#           hps.data.mel_fmax
#        )
# #         spec_vc = spectrogram_torch(y_hat_vc, hps.data.filter_length,
# #                 hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,
# #                 center=False)
        
# #         spec_vc = torch.squeeze(spec_vc, 0)
        
# #         mel_vc = spec_to_mel_torch(
# #           spec_vc, 
# #           hps.data.filter_length, 
# #           hps.data.n_mel_channels, 
# #           hps.data.sampling_rate,
# #           hps.data.mel_fmin, 
# #           hps.data.mel_fmax)
        
#         style_vector_vc = self.style_encoder(o_hat_vc_mel.transpose(1,2), torch.tensor(np.full((o_hat_vc_mel.shape[0]), o_hat_vc_mel.shape[2])))
        
        
#         return o, o_ref, ids_slice, ids_slice_ref, y_mask, y_mask_ref, (z, z_ref, z_p, z_p_ref, logs_q, logs_q_ref), style_vector_vc, style_vector_ref
  

  def infer(self, x, x_lengths, mel, mel_lengths = None, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
    #print(m_p.transpose(1,2).shape, m_p.transpose(1,2))
    if self.n_speakers > 0:
      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
    else:
      g = None

    if self.use_sdp:
      # logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
      logw = self.dp(x, x_mask, g=None, reverse=True, noise_scale=noise_scale_w)
    else:
      # logw = self.dp(x, x_mask, g=g)
      logw = self.dp(x, x_mask, g=None)
    w = torch.exp(logw) * x_mask * length_scale
    w_ceil = torch.ceil(w)
    
    #print(w_ceil)
    y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
    y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
    
    attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
    attn = commons.generate_path(w_ceil, attn_mask)

    m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
    
    #print(m_p.transpose(1,2).shape, m_p.transpose(1,2))
    logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']

    z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
    #print(z_p.transpose(1,2).shape, z_p.transpose(1,2))
    #* used for mel style encoder
    if mel_lengths is not None:
      style_mask = torch.unsqueeze(commons.sequence_mask(mel_lengths, mel.size(2)), 1).to(x.dtype)
      style_vector = self.style_encoder(mel.transpose(1,2), torch.tensor(np.full((mel.shape[0]), mel.shape[2])))
    else:
      style_vector = self.style_encoder(mel.transpose(1,2), torch.tensor(np.full((mel.shape[0]), mel.shape[2])))
    # z = self.flow(z_p, y_mask, g=g, reverse=True)

    
    z = self.flow(z_p, y_mask, g=style_vector.unsqueeze(-1), reverse=True)
    # o = self.dec((z * y_mask)[:,:,:max_len], g=g)
    o = self.dec((z * y_mask)[:,:,:max_len], g=None)
    return o, attn, y_mask, (z, z_p, m_p, logs_p)


  # def forward(self, spec_source_pattern, spec_lengths_source, mel_source, mel_ref):
  #       style_vector_src = self.style_encoder(mel_source.transpose(1,2), torch.tensor(np.full((mel_source.shape[0]), mel_source.shape[2])))
  #       style_vector_ref = self.style_encoder(mel_ref.transpose(1,2), torch.tensor(np.full((mel_ref.shape[0]), mel_ref.shape[2])))
  #       z, m_q, logs_q, y_mask = self.enc_q(spec_source_pattern, spec_lengths_source, g=None)
  #       z_p = self.flow(z, y_mask, g=style_vector_src.unsqueeze(-1))
  #       z_hat = self.flow(z_p, y_mask, g=style_vector_ref.unsqueeze(-1), reverse=True)
  #       o_hat = self.dec(z_hat * y_mask, g=None)
  #       z_slice, ids_slice = commons.rand_slice_segments(z, spec_lengths_source, self.segment_size)
  #       return o_hat, y_mask, (z, z_p, z_hat), style_vector_src, style_vector_ref, ids_slice

  def voice_conversion(self, spec_source_pattern, spec_lengths_source, mel_source, mel_ref):
        style_vector_src = self.style_encoder(mel_source.transpose(1,2), torch.tensor(np.full((mel_source.shape[0]), mel_source.shape[2])))
        style_vector_ref = self.style_encoder(mel_ref.transpose(1,2), torch.tensor(np.full((mel_ref.shape[0]), mel_ref.shape[2])))
        z, m_q, logs_q, y_mask = self.enc_q(spec_source_pattern, spec_lengths_source, g=None)
        z_p = self.flow(z, y_mask, g=style_vector_src.unsqueeze(-1))
        z_hat = self.flow(z_p, y_mask, g=style_vector_ref.unsqueeze(-1), reverse=True)
        o_hat = self.dec(z_hat * y_mask, g=None)
        return o_hat, y_mask, (z, z_p, z_hat)
    
    
class MelStyleEncoder(nn.Module):
    ''' MelStyleEncoder '''
    def __init__(self, n_mel_channels=80,
          style_hidden=128,
          style_vector_dim=256,
          style_kernel_size=5,
          style_head=2,
          dropout=0.1):
        super(MelStyleEncoder, self).__init__()
        self.in_dim = n_mel_channels 
        self.hidden_dim = style_hidden
        self.out_dim = style_vector_dim
        self.kernel_size = style_kernel_size
        self.n_head = style_head
        self.dropout = dropout

        self.spectral = nn.Sequential(
            modules.LinearNorm(self.in_dim, self.hidden_dim),
            modules.Mish(),
            nn.Dropout(self.dropout),
            modules.LinearNorm(self.hidden_dim, self.hidden_dim),
            modules.Mish(),
            nn.Dropout(self.dropout)
        )

        self.temporal = nn.Sequential(
            modules.Conv1dGLU(self.hidden_dim, self.hidden_dim, self.kernel_size, self.dropout),
            modules.Conv1dGLU(self.hidden_dim, self.hidden_dim, self.kernel_size, self.dropout),
        )

        self.slf_attn = modules.MultiHeadAttention(self.n_head, self.hidden_dim, 
                                self.hidden_dim//self.n_head, self.hidden_dim//self.n_head, self.dropout) 

        self.fc = modules.LinearNorm(self.hidden_dim, self.out_dim)

    def temporal_avg_pool(self, x, mask=None):
        if mask is None:
            out = torch.mean(x, dim=1)
        else:
            len_ = (~mask).sum(dim=1).unsqueeze(1)
            x = x.masked_fill(mask.unsqueeze(-1), 0)
            x = x.sum(dim=1)
            out = torch.div(x, len_)
        return out

    def forward(self, x, mask=None):
        max_len = x.shape[1]
        slf_attn_mask = mask.unsqueeze(1).expand(-1, max_len, -1) if mask is not None else None
        
        # spectral
        x = self.spectral(x)
        # temporal
        x = x.transpose(1,2)
        x = self.temporal(x)
        x = x.transpose(1,2)
        # self-attention
        #print(x.shape, mask.shape)
        if mask is not None:
            x = x.masked_fill(mask.unsqueeze(-1), 0)
        x, _ = self.slf_attn(x, mask=slf_attn_mask)
        # fc
        x = self.fc(x)
        # temoral average pooling
        w = self.temporal_avg_pool(x, mask=mask)

        return w

    
class StyleEmbedding(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.gst = StyleEncoder()

    def forward(self,
                batch_of_spectrograms,
                batch_of_spectrogram_lengths,
                return_all_outs=False,
                return_only_refs=False):
        minimum_sequence_length = 812
        specs = list()

        for index, spec_length in enumerate(batch_of_spectrogram_lengths):
            spec = batch_of_spectrograms[index][:spec_length]
            # double the length at least once, then check
            spec = spec.repeat((2, 1))
            current_spec_length = len(spec)
            while current_spec_length < minimum_sequence_length:
                # make it longer
                spec = spec.repeat((2, 1))
                current_spec_length = len(spec)
            specs.append(spec[:812])

        spec_batch = torch.stack(specs, dim=0)
        return self.gst(speech=spec_batch,
                        return_all_outs=return_all_outs,
                        return_only_ref=return_only_refs)
    
class StyleEncoder(torch.nn.Module):
    def __init__(
        self,
        idim: int = 80,
        gst_tokens: int = 2000,
        gst_token_dim: int = 256,
        gst_heads: int = 8,
        conv_layers: int = 8,
        conv_chans_list=(32, 32, 64, 64, 128, 128, 256, 256),
        conv_kernel_size: int = 3,
        conv_stride: int = 2,
        gst_layers: int = 2,
        gst_units: int = 256,
    ):
        """Initialize global style encoder module."""
        super(StyleEncoder, self).__init__()

        self.num_tokens = gst_tokens
        self.ref_enc = ReferenceEncoder(idim=idim,
                                        conv_layers=conv_layers,
                                        conv_chans_list=conv_chans_list,
                                        conv_kernel_size=conv_kernel_size,
                                        conv_stride=conv_stride,
                                        gst_layers=gst_layers,
                                        gst_units=gst_units, )
        self.stl = StyleTokenLayer(ref_embed_dim=gst_units,
                                   gst_tokens=gst_tokens,
                                   gst_token_dim=gst_token_dim,
                                   gst_heads=gst_heads, )
        
        self.ref_mel = MelStyleEncoder(n_mel_channels = idim)

    def forward(self, speech, return_all_outs=False, return_only_ref=False):
        ref_mels = self.ref_mel(speech)
        ref_embs = self.ref_enc(speech)
        if return_only_ref and not return_all_outs:
            return ref_embs
        style_embs = self.stl(ref_embs)

        if return_all_outs:
            if return_only_ref:
                return ref_embs, [ref_embs] + [style_embs]
            return style_embs, [ref_embs] + [style_embs]
        
        #print(style_embs.shape, ref_mels.shape, ref_embs.shape)
        return style_embs + ref_mels

    def calculate_ada4_regularization_loss(self):
        losses = list()
        for emb1_index in range(self.num_tokens):
            for emb2_index in range(emb1_index + 1, self.num_tokens):
                if emb1_index != emb2_index:
                    losses.append(torch.nn.functional.cosine_similarity(self.stl.gst_embs[emb1_index],
                                                                        self.stl.gst_embs[emb2_index], dim=0))
        return sum(losses)


class ReferenceEncoder(torch.nn.Module):

    def __init__(
        self,
        idim=80,
        conv_layers: int = 6,
        conv_chans_list=(32, 32, 64, 64, 128, 128),
        conv_kernel_size: int = 3,
        conv_stride: int = 2,
        gst_layers: int = 1,
        gst_units: int = 128,
    ):
        """Initialize reference encoder module."""
        super(ReferenceEncoder, self).__init__()

        # check hyperparameters are valid
        assert conv_kernel_size % 2 == 1, "kernel size must be odd."
        assert (
                len(conv_chans_list) == conv_layers), "the number of conv layers and length of channels list must be the same."

        convs = []
        padding = (conv_kernel_size - 1) // 2
        for i in range(conv_layers):
            conv_in_chans = 1 if i == 0 else conv_chans_list[i - 1]
            conv_out_chans = conv_chans_list[i]
            convs += [torch.nn.Conv2d(conv_in_chans,
                                      conv_out_chans,
                                      kernel_size=conv_kernel_size,
                                      stride=conv_stride,
                                      padding=padding,
                                      # Do not use bias due to the following batch norm
                                      bias=False, ),
                      torch.nn.BatchNorm2d(conv_out_chans),
                      torch.nn.ReLU(inplace=True), ]
        self.convs = torch.nn.Sequential(*convs)

        self.conv_layers = conv_layers
        self.kernel_size = conv_kernel_size
        self.stride = conv_stride
        self.padding = padding

        # get the number of GRU input units
        gst_in_units = idim
        for i in range(conv_layers):
            gst_in_units = (gst_in_units - conv_kernel_size + 2 * padding) // conv_stride + 1
        gst_in_units *= conv_out_chans
        self.gst = torch.nn.GRU(gst_in_units, gst_units, gst_layers, batch_first=True)

    def forward(self, speech):
        """Calculate forward propagation.
        Args:
            speech (Tensor): Batch of padded target features (B, Lmax, idim).
        Returns:
            Tensor: Reference embedding (B, gst_units)
        """
        batch_size = speech.size(0)
        xs = speech.unsqueeze(1)  # (B, 1, Lmax, idim)
        hs = self.convs(xs).transpose(1, 2)  # (B, Lmax', conv_out_chans, idim')
        time_length = hs.size(1)
        hs = hs.contiguous().view(batch_size, time_length, -1)  # (B, Lmax', gst_units)
        self.gst.flatten_parameters()
        # pack_padded_sequence(hs, speech_lens, enforce_sorted=False, batch_first=True)
        _, ref_embs = self.gst(hs)  # (gst_layers, batch_size, gst_units)
        ref_embs = ref_embs[-1]  # (batch_size, gst_units)

        return ref_embs


class StyleTokenLayer(torch.nn.Module):
    """Style token layer module.
    This module is style token layer introduced in `Style Tokens: Unsupervised Style
    Modeling, Control and Transfer in End-to-End Speech Synthesis`.
    .. _`Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End
        Speech Synthesis`: https://arxiv.org/abs/1803.09017
    Args:
        ref_embed_dim (int, optional): Dimension of the input reference embedding.
        gst_tokens (int, optional): The number of GST embeddings.
        gst_token_dim (int, optional): Dimension of each GST embedding.
        gst_heads (int, optional): The number of heads in GST multihead attention.
        dropout_rate (float, optional): Dropout rate in multi-head attention.
    """

    def __init__(
        self,
        ref_embed_dim: int = 128,
        gst_tokens: int = 10,
        gst_token_dim: int = 128,
        gst_heads: int = 4,
        dropout_rate: float = 0.0,
    ):
        """Initialize style token layer module."""
        super(StyleTokenLayer, self).__init__()

        gst_embs = torch.randn(gst_tokens, gst_token_dim // gst_heads)
        self.register_parameter("gst_embs", torch.nn.Parameter(gst_embs))
        self.mha = MultiHeadedAttention(q_dim=ref_embed_dim,
                                        k_dim=gst_token_dim // gst_heads,
                                        v_dim=gst_token_dim // gst_heads,
                                        n_head=gst_heads,
                                        n_feat=gst_token_dim,
                                        dropout_rate=dropout_rate, )

    def forward(self, ref_embs):
        """Calculate forward propagation.
        Args:
            ref_embs (Tensor): Reference embeddings (B, ref_embed_dim).
        Returns:
            Tensor: Style token embeddings (B, gst_token_dim).
        """
        batch_size = ref_embs.size(0)
        # (num_tokens, token_dim) -> (batch_size, num_tokens, token_dim)
        gst_embs = torch.tanh(self.gst_embs).unsqueeze(0).expand(batch_size, -1, -1)
        # NOTE(kan-bayashi): Shoule we apply Tanh?
        ref_embs = ref_embs.unsqueeze(1)  # (batch_size, 1 ,ref_embed_dim)
        style_embs = self.mha(ref_embs, gst_embs, gst_embs, None)

        return style_embs.squeeze(1)


class MultiHeadedAttention(BaseMultiHeadedAttention):
    """Multi head attention module with different input dimension."""

    def __init__(self, q_dim, k_dim, v_dim, n_head, n_feat, dropout_rate=0.0):
        """Initialize multi head attention module."""
        # NOTE(kan-bayashi): Do not use super().__init__() here since we want to
        #   overwrite BaseMultiHeadedAttention.__init__() method.
        torch.nn.Module.__init__(self)
        assert n_feat % n_head == 0
        # We assume d_v always equals d_k
        self.d_k = n_feat // n_head
        self.h = n_head
        self.linear_q = torch.nn.Linear(q_dim, n_feat)
        self.linear_k = torch.nn.Linear(k_dim, n_feat)
        self.linear_v = torch.nn.Linear(v_dim, n_feat)
        self.linear_out = torch.nn.Linear(n_feat, n_feat)
        self.attn = None
        self.dropout = torch.nn.Dropout(p=dropout_rate)