Upload 14 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+monotonic_align/build/temp.linux-x86_64-3.6/core.o filter=lfs diff=lfs merge=lfs -text

attentions.py

@@ -0,0 +1,303 @@
+import copy
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+import commons
+import modules
+from modules import LayerNorm
+   
+
+class Encoder(nn.Module):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
+    super().__init__()
+    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.window_size = window_size
+
+    self.drop = nn.Dropout(p_dropout)
+    self.attn_layers = nn.ModuleList()
+    self.norm_layers_1 = nn.ModuleList()
+    self.ffn_layers = nn.ModuleList()
+    self.norm_layers_2 = nn.ModuleList()
+    for i in range(self.n_layers):
+      self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask):
+    attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.attn_layers[i](x, x, attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class Decoder(nn.Module):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
+    super().__init__()
+    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.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+
+    self.drop = nn.Dropout(p_dropout)
+    self.self_attn_layers = nn.ModuleList()
+    self.norm_layers_0 = nn.ModuleList()
+    self.encdec_attn_layers = nn.ModuleList()
+    self.norm_layers_1 = nn.ModuleList()
+    self.ffn_layers = nn.ModuleList()
+    self.norm_layers_2 = nn.ModuleList()
+    for i in range(self.n_layers):
+      self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
+      self.norm_layers_0.append(LayerNorm(hidden_channels))
+      self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask, h, h_mask):
+    """
+    x: decoder input
+    h: encoder output
+    """
+    self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
+    encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.self_attn_layers[i](x, x, self_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_0[i](x + y)
+
+      y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+      
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class MultiHeadAttention(nn.Module):
+  def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
+    super().__init__()
+    assert channels % n_heads == 0
+
+    self.channels = channels
+    self.out_channels = out_channels
+    self.n_heads = n_heads
+    self.p_dropout = p_dropout
+    self.window_size = window_size
+    self.heads_share = heads_share
+    self.block_length = block_length
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+    self.attn = None
+
+    self.k_channels = channels // n_heads
+    self.conv_q = nn.Conv1d(channels, channels, 1)
+    self.conv_k = nn.Conv1d(channels, channels, 1)
+    self.conv_v = nn.Conv1d(channels, channels, 1)
+    self.conv_o = nn.Conv1d(channels, out_channels, 1)
+    self.drop = nn.Dropout(p_dropout)
+
+    if window_size is not None:
+      n_heads_rel = 1 if heads_share else n_heads
+      rel_stddev = self.k_channels**-0.5
+      self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+      self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
+
+    nn.init.xavier_uniform_(self.conv_q.weight)
+    nn.init.xavier_uniform_(self.conv_k.weight)
+    nn.init.xavier_uniform_(self.conv_v.weight)
+    if proximal_init:
+      with torch.no_grad():
+        self.conv_k.weight.copy_(self.conv_q.weight)
+        self.conv_k.bias.copy_(self.conv_q.bias)
+      
+  def forward(self, x, c, attn_mask=None):
+    q = self.conv_q(x)
+    k = self.conv_k(c)
+    v = self.conv_v(c)
+    
+    x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+    x = self.conv_o(x)
+    return x
+
+  def attention(self, query, key, value, mask=None):
+    # reshape [b, d, t] -> [b, n_h, t, d_k]
+    b, d, t_s, t_t = (*key.size(), query.size(2))
+    query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+    key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+    value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+    scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+    if self.window_size is not None:
+      assert t_s == t_t, "Relative attention is only available for self-attention."
+      key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+      rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
+      scores_local = self._relative_position_to_absolute_position(rel_logits)
+      scores = scores + scores_local
+    if self.proximal_bias:
+      assert t_s == t_t, "Proximal bias is only available for self-attention."
+      scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
+    if mask is not None:
+      scores = scores.masked_fill(mask == 0, -1e4)
+      if self.block_length is not None:
+        assert t_s == t_t, "Local attention is only available for self-attention."
+        block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
+        scores = scores.masked_fill(block_mask == 0, -1e4)
+    p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
+    p_attn = self.drop(p_attn)
+    output = torch.matmul(p_attn, value)
+    if self.window_size is not None:
+      relative_weights = self._absolute_position_to_relative_position(p_attn)
+      value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+      output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
+    output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
+    return output, p_attn
+
+  def _matmul_with_relative_values(self, x, y):
+    """
+    x: [b, h, l, m]
+    y: [h or 1, m, d]
+    ret: [b, h, l, d]
+    """
+    ret = torch.matmul(x, y.unsqueeze(0))
+    return ret
+
+  def _matmul_with_relative_keys(self, x, y):
+    """
+    x: [b, h, l, d]
+    y: [h or 1, m, d]
+    ret: [b, h, l, m]
+    """
+    ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+    return ret
+
+  def _get_relative_embeddings(self, relative_embeddings, length):
+    max_relative_position = 2 * self.window_size + 1
+    # Pad first before slice to avoid using cond ops.
+    pad_length = max(length - (self.window_size + 1), 0)
+    slice_start_position = max((self.window_size + 1) - length, 0)
+    slice_end_position = slice_start_position + 2 * length - 1
+    if pad_length > 0:
+      padded_relative_embeddings = F.pad(
+          relative_embeddings,
+          commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
+    else:
+      padded_relative_embeddings = relative_embeddings
+    used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
+    return used_relative_embeddings
+
+  def _relative_position_to_absolute_position(self, x):
+    """
+    x: [b, h, l, 2*l-1]
+    ret: [b, h, l, l]
+    """
+    batch, heads, length, _ = x.size()
+    # Concat columns of pad to shift from relative to absolute indexing.
+    x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
+
+    # Concat extra elements so to add up to shape (len+1, 2*len-1).
+    x_flat = x.view([batch, heads, length * 2 * length])
+    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
+
+    # Reshape and slice out the padded elements.
+    x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
+    return x_final
+
+  def _absolute_position_to_relative_position(self, x):
+    """
+    x: [b, h, l, l]
+    ret: [b, h, l, 2*l-1]
+    """
+    batch, heads, length, _ = x.size()
+    # padd along column
+    x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
+    x_flat = x.view([batch, heads, length**2 + length*(length -1)])
+    # add 0's in the beginning that will skew the elements after reshape
+    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+    x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
+    return x_final
+
+  def _attention_bias_proximal(self, length):
+    """Bias for self-attention to encourage attention to close positions.
+    Args:
+      length: an integer scalar.
+    Returns:
+      a Tensor with shape [1, 1, length, length]
+    """
+    r = torch.arange(length, dtype=torch.float32)
+    diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+    return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+  def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
+    super().__init__()
+    self.in_channels = in_channels
+    self.out_channels = out_channels
+    self.filter_channels = filter_channels
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.activation = activation
+    self.causal = causal
+
+    if causal:
+      self.padding = self._causal_padding
+    else:
+      self.padding = self._same_padding
+
+    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+    self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+    self.drop = nn.Dropout(p_dropout)
+
+  def forward(self, x, x_mask):
+    x = self.conv_1(self.padding(x * x_mask))
+    if self.activation == "gelu":
+      x = x * torch.sigmoid(1.702 * x)
+    else:
+      x = torch.relu(x)
+    x = self.drop(x)
+    x = self.conv_2(self.padding(x * x_mask))
+    return x * x_mask
+  
+  def _causal_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = self.kernel_size - 1
+    pad_r = 0
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    x = F.pad(x, commons.convert_pad_shape(padding))
+    return x
+
+  def _same_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = (self.kernel_size - 1) // 2
+    pad_r = self.kernel_size // 2
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    x = F.pad(x, commons.convert_pad_shape(padding))
+    return x

data_utils.py

@@ -0,0 +1,392 @@
+import time
+import os
+import random
+import numpy as np
+import torch
+import torch.utils.data
+
+import commons 
+from mel_processing import spectrogram_torch
+from utils import load_wav_to_torch, load_filepaths_and_text
+from text import text_to_sequence, cleaned_text_to_sequence
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+    """
+        1) loads audio, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.text_cleaners  = hparams.text_cleaners
+        self.max_wav_value  = hparams.max_wav_value
+        self.sampling_rate  = hparams.sampling_rate
+        self.filter_length  = hparams.filter_length 
+        self.hop_length     = hparams.hop_length 
+        self.win_length     = hparams.win_length
+        self.sampling_rate  = hparams.sampling_rate 
+
+        self.cleaned_text = getattr(hparams, "cleaned_text", False)
+
+        self.add_blank = hparams.add_blank
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 190)
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths_and_text)
+        self._filter()
+
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text])
+                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        audiopath, text = audiopath_and_text[0], audiopath_and_text[1]
+        text = self.get_text(text)
+        spec, wav = self.get_audio(audiopath)
+        return (text, spec, wav)
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} {} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
+            spec = spectrogram_torch(audio_norm, self.filter_length,
+                self.sampling_rate, self.hop_length, self.win_length,
+                center=False)
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename)
+        return spec, audio_norm
+
+    def get_text(self, text):
+        if self.cleaned_text:
+            text_norm = cleaned_text_to_sequence(text)
+        else:
+            text_norm = text_to_sequence(text, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(text_norm, 0)
+        text_norm = torch.LongTensor(text_norm)
+        return text_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate():
+    """ Zero-pads model inputs and targets
+    """
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[1].size(1) for x in batch]),
+            dim=0, descending=True)
+
+        max_text_len = max([len(x[0]) for x in batch])
+        max_spec_len = max([x[1].size(1) for x in batch])
+        max_wav_len = max([x[2].size(1) for x in batch])
+
+        text_lengths = torch.LongTensor(len(batch))
+        spec_lengths = torch.LongTensor(len(batch))
+        wav_lengths = torch.LongTensor(len(batch))
+
+        text_padded = torch.LongTensor(len(batch), max_text_len)
+        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
+        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
+        text_padded.zero_()
+        spec_padded.zero_()
+        wav_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            text = row[0]
+            text_padded[i, :text.size(0)] = text
+            text_lengths[i] = text.size(0)
+
+            spec = row[1]
+            spec_padded[i, :, :spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wav = row[2]
+            wav_padded[i, :, :wav.size(1)] = wav
+            wav_lengths[i] = wav.size(1)
+
+        if self.return_ids:
+            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, ids_sorted_decreasing
+        return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths
+
+
+"""Multi speaker version"""
+class TextAudioSpeakerLoader(torch.utils.data.Dataset):
+    """
+        1) loads audio, speaker_id, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+    def __init__(self, audiopaths_sid_text, hparams):
+        self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
+        self.text_cleaners = hparams.text_cleaners
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length  = hparams.filter_length
+        self.hop_length     = hparams.hop_length
+        self.win_length     = hparams.win_length
+        self.sampling_rate  = hparams.sampling_rate
+
+        self.cleaned_text = getattr(hparams, "cleaned_text", False)
+
+        self.add_blank = hparams.add_blank
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 190)
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths_sid_text)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+
+        audiopaths_sid_text_new = []
+        lengths = []
+        for audiopath, sid, text in self.audiopaths_sid_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_sid_text_new.append([audiopath, sid, text])
+                lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
+        self.audiopaths_sid_text = audiopaths_sid_text_new
+        self.lengths = lengths
+
+    def get_audio_text_speaker_pair(self, audiopath_sid_text):
+        # separate filename, speaker_id and text
+        audiopath, sid, text = audiopath_sid_text[0], audiopath_sid_text[1], audiopath_sid_text[2]
+        text = self.get_text(text)
+        spec, wav = self.get_audio(audiopath)
+        sid = self.get_sid(sid)
+        return (text, spec, wav, sid)
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} {} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            spec = torch.load(spec_filename)
+        else:
+            spec = spectrogram_torch(audio_norm, self.filter_length,
+                self.sampling_rate, self.hop_length, self.win_length,
+                center=False)
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename)
+        return spec, audio_norm
+
+    def get_text(self, text):
+        if self.cleaned_text:
+            text_norm = cleaned_text_to_sequence(text)
+        else:
+            text_norm = text_to_sequence(text, self.text_cleaners)
+        if self.add_blank:
+            text_norm = commons.intersperse(text_norm, 0)
+        text_norm = torch.LongTensor(text_norm)
+        return text_norm
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def __getitem__(self, index):
+        return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_sid_text)
+
+
+class TextAudioSpeakerCollate():
+    """ Zero-pads model inputs and targets
+    """
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text, audio and speaker identities
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized, sid]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[1].size(1) for x in batch]),
+            dim=0, descending=True)
+
+        max_text_len = max([len(x[0]) for x in batch])
+        max_spec_len = max([x[1].size(1) for x in batch])
+        max_wav_len = max([x[2].size(1) for x in batch])
+
+        text_lengths = torch.LongTensor(len(batch))
+        spec_lengths = torch.LongTensor(len(batch))
+        wav_lengths = torch.LongTensor(len(batch))
+        sid = torch.LongTensor(len(batch))
+
+        text_padded = torch.LongTensor(len(batch), max_text_len)
+        spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
+        wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
+        text_padded.zero_()
+        spec_padded.zero_()
+        wav_padded.zero_()
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            text = row[0]
+            text_padded[i, :text.size(0)] = text
+            text_lengths[i] = text.size(0)
+
+            spec = row[1]
+            spec_padded[i, :, :spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wav = row[2]
+            wav_padded[i, :, :wav.size(1)] = wav
+            wav_lengths[i] = wav.size(1)
+
+            sid[i] = row[3]
+
+        if self.return_ids:
+            return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing
+        return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    """
+    Maintain similar input lengths in a batch.
+    Length groups are specified by boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+  
+    It removes samples which are not included in the boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+    """
+    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+  
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+  
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1:
+                buckets[idx_bucket].append(i)
+  
+        for i in range(len(buckets) - 1, 0, -1):
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i+1)
+  
+        num_samples_per_bucket = []
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+        return buckets, num_samples_per_bucket
+  
+    def __iter__(self):
+      # deterministically shuffle based on epoch
+      g = torch.Generator()
+      g.manual_seed(self.epoch)
+  
+      indices = []
+      if self.shuffle:
+          for bucket in self.buckets:
+              indices.append(torch.randperm(len(bucket), generator=g).tolist())
+      else:
+          for bucket in self.buckets:
+              indices.append(list(range(len(bucket))))
+  
+      batches = []
+      for i in range(len(self.buckets)):
+          bucket = self.buckets[i]
+          len_bucket = len(bucket)
+          ids_bucket = indices[i]
+          num_samples_bucket = self.num_samples_per_bucket[i]
+  
+          # add extra samples to make it evenly divisible
+          rem = num_samples_bucket - len_bucket
+          ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
+  
+          # subsample
+          ids_bucket = ids_bucket[self.rank::self.num_replicas]
+  
+          # batching
+          for j in range(len(ids_bucket) // self.batch_size):
+              batch = [bucket[idx] for idx in ids_bucket[j*self.batch_size:(j+1)*self.batch_size]]
+              batches.append(batch)
+  
+      if self.shuffle:
+          batch_ids = torch.randperm(len(batches), generator=g).tolist()
+          batches = [batches[i] for i in batch_ids]
+      self.batches = batches
+  
+      assert len(self.batches) * self.batch_size == self.num_samples
+      return iter(self.batches)
+  
+    def _bisect(self, x, lo=0, hi=None):
+      if hi is None:
+          hi = len(self.boundaries) - 1
+  
+      if hi > lo:
+          mid = (hi + lo) // 2
+          if self.boundaries[mid] < x and x <= self.boundaries[mid+1]:
+              return mid
+          elif x <= self.boundaries[mid]:
+              return self._bisect(x, lo, mid)
+          else:
+              return self._bisect(x, mid + 1, hi)
+      else:
+          return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size

mel_processing.py

@@ -0,0 +1,112 @@
+import math
+import os
+import random
+import torch
+from torch import nn
+import torch.nn.functional as F
+import torch.utils.data
+import numpy as np
+import librosa
+import librosa.util as librosa_util
+from librosa.util import normalize, pad_center, tiny
+from scipy.signal import get_window
+from scipy.io.wavfile import read
+from librosa.filters import mel as librosa_mel_fn
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.device)
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+    global mel_basis
+    dtype_device = str(spec.dtype) + '_' + str(spec.device)
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
+    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+    return spec
+
+
+def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global mel_basis, hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.device)
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+
+    spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec

melspec_gen.py

@@ -0,0 +1,61 @@
+from scipy.io.wavfile import read
+import torch
+import numpy as np
+import os
+from multiprocessing import Pool
+#from tqdm import tqdm
+
+# Change here
+base="/mnt/beegfs/home/espinosa/wallon/wallon_ms_02/p02/"
+
+hann_window = {}
+def load_wav_to_torch(full_path):
+    sampling_rate, data = read(full_path)
+    # data, sampling_rate = librosa.load(full_path)
+    return torch.FloatTensor(data.astype(np.float32)), sampling_rate
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.device)
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+    return spec
+
+def get_audio(filename):
+    max_wave_length = 32768.0
+    filter_length = 1024
+    hop_length = 256
+    win_length = 1024
+    audio, sampling_rate = load_wav_to_torch(filename)
+    audio_norm = audio / max_wave_length
+    audio_norm = audio_norm.unsqueeze(0)
+    spec_filename = filename.replace(".wav", ".spec.pt")
+    spec = spectrogram_torch(audio_norm, filter_length,
+        sampling_rate, hop_length, win_length,
+        center=False)
+    spec = torch.squeeze(spec, 0)
+    torch.save(spec, spec_filename)
+
+if __name__=="__main__":
+    waves = []
+    for wav_name in os.listdir(base):
+        wav_path = os.path.join(base, wav_name)
+        if wav_path.endswith(".wav"):
+            waves.append(wav_path)
+            
+    for wav_path in waves:
+        get_audio(wav_path)

monotonic_align/__init__.py

@@ -0,0 +1,19 @@
+import numpy as np
+import torch
+from .monotonic_align.core import maximum_path_c
+
+
+def maximum_path(neg_cent, mask):
+  """ Cython optimized version.
+  neg_cent: [b, t_t, t_s]
+  mask: [b, t_t, t_s]
+  """
+  device = neg_cent.device
+  dtype = neg_cent.dtype
+  neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
+  path = np.zeros(neg_cent.shape, dtype=np.int32)
+
+  t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
+  t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
+  maximum_path_c(path, neg_cent, t_t_max, t_s_max)
+  return torch.from_numpy(path).to(device=device, dtype=dtype)

monotonic_align/build/temp.linux-x86_64-3.6/core.o

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cf087f75fefa878bf0c2823e9481756bf83d24988a4ba54f9e2bbb9ddfdf9f8
+size 2165448

monotonic_align/core.pyx

@@ -0,0 +1,42 @@
+cimport cython
+from cython.parallel import prange
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
+  cdef int x
+  cdef int y
+  cdef float v_prev
+  cdef float v_cur
+  cdef float tmp
+  cdef int index = t_x - 1
+
+  for y in range(t_y):
+    for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+      if x == y:
+        v_cur = max_neg_val
+      else:
+        v_cur = value[y-1, x]
+      if x == 0:
+        if y == 0:
+          v_prev = 0.
+        else:
+          v_prev = max_neg_val
+      else:
+        v_prev = value[y-1, x-1]
+      value[y, x] += max(v_prev, v_cur)
+
+  for y in range(t_y - 1, -1, -1):
+    path[y, index] = 1
+    if index != 0 and (index == y or value[y-1, index] < value[y-1, index-1]):
+      index = index - 1
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
+  cdef int b = paths.shape[0]
+  cdef int i
+  for i in prange(b, nogil=True):
+    maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])

monotonic_align/setup.py

@@ -0,0 +1,9 @@
+from distutils.core import setup
+from Cython.Build import cythonize
+import numpy
+
+setup(
+  name = 'monotonic_align',
+  ext_modules = cythonize("core.pyx"),
+  include_dirs=[numpy.get_include()]
+)

preprocess.py

@@ -0,0 +1,25 @@
+import argparse
+import text
+from utils import load_filepaths_and_text
+
+if __name__ == '__main__':
+  parser = argparse.ArgumentParser()
+  parser.add_argument("--out_extension", default="cleaned")
+  parser.add_argument("--text_index", default=1, type=int)
+  parser.add_argument("--filelists", nargs="+", default=["filelists/ljs_audio_text_val_filelist.txt", "filelists/ljs_audio_text_test_filelist.txt"])
+  parser.add_argument("--text_cleaners", nargs="+", default=["basic_cleaners"])
+
+  args = parser.parse_args()
+    
+
+  for filelist in args.filelists:
+    print("START:", filelist)
+    filepaths_and_text = load_filepaths_and_text(filelist)
+    for i in range(len(filepaths_and_text)):
+      original_text = filepaths_and_text[i][args.text_index]
+      cleaned_text = text._clean_text(original_text, args.text_cleaners)
+      filepaths_and_text[i][args.text_index] = cleaned_text
+
+    new_filelist = filelist + "." + args.out_extension
+    with open(new_filelist, "w", encoding="utf-8") as f:
+      f.writelines(["|".join(x) + "\n" for x in filepaths_and_text])

train.py

@@ -0,0 +1,290 @@
+import os
+import json
+import argparse
+import itertools
+import math
+import torch
+from torch import nn, optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+import torch.multiprocessing as mp
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.cuda.amp import autocast, GradScaler
+
+import commons
+import utils
+from data_utils import (
+  TextAudioLoader,
+  TextAudioCollate,
+  DistributedBucketSampler
+)
+from models import (
+  SynthesizerTrn,
+  MultiPeriodDiscriminator,
+)
+from losses import (
+  generator_loss,
+  discriminator_loss,
+  feature_loss,
+  kl_loss
+)
+from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from text.symbols import symbols
+
+
+torch.backends.cudnn.benchmark = True
+global_step = 0
+
+
+def main():
+  """Assume Single Node Multi GPUs Training Only"""
+  assert torch.cuda.is_available(), "CPU training is not allowed."
+
+  n_gpus = torch.cuda.device_count()
+  os.environ['MASTER_ADDR'] = 'localhost'
+  os.environ['MASTER_PORT'] = '80000'
+
+  hps = utils.get_hparams()
+  mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
+
+
+def run(rank, n_gpus, hps):
+  global global_step
+  if rank == 0:
+    logger = utils.get_logger(hps.model_dir)
+    logger.info(hps)
+    utils.check_git_hash(hps.model_dir)
+    writer = SummaryWriter(log_dir=hps.model_dir)
+    writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+  dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
+  torch.manual_seed(hps.train.seed)
+  torch.cuda.set_device(rank)
+
+  train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
+  train_sampler = DistributedBucketSampler(
+      train_dataset,
+      hps.train.batch_size,
+      [32,300,400,500,600,700,800,900,1000],
+      num_replicas=n_gpus,
+      rank=rank,
+      shuffle=True)
+  collate_fn = TextAudioCollate()
+  train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
+      collate_fn=collate_fn, batch_sampler=train_sampler)
+  if rank == 0:
+    eval_dataset = TextAudioLoader(hps.data.validation_files, hps.data)
+    eval_loader = DataLoader(eval_dataset, num_workers=8, shuffle=False,
+        batch_size=hps.train.batch_size, pin_memory=True,
+        drop_last=False, collate_fn=collate_fn)
+
+  net_g = SynthesizerTrn(
+      len(symbols),
+      hps.data.filter_length // 2 + 1,
+      hps.train.segment_size // hps.data.hop_length,
+      **hps.model).cuda(rank)
+  net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
+  optim_g = torch.optim.AdamW(
+      net_g.parameters(), 
+      hps.train.learning_rate, 
+      betas=hps.train.betas, 
+      eps=hps.train.eps)
+  optim_d = torch.optim.AdamW(
+      net_d.parameters(),
+      hps.train.learning_rate, 
+      betas=hps.train.betas, 
+      eps=hps.train.eps)
+  net_g = DDP(net_g, device_ids=[rank])
+  net_d = DDP(net_d, device_ids=[rank])
+
+  try:
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d)
+    global_step = (epoch_str - 1) * len(train_loader)
+  except:
+    epoch_str = 1
+    global_step = 0
+
+  scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
+  scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
+
+  scaler = GradScaler(enabled=hps.train.fp16_run)
+
+  for epoch in range(epoch_str, hps.train.epochs + 1):
+    if rank==0:
+      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, eval_loader], logger, [writer, writer_eval])
+    else:
+      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, None], None, None)
+    scheduler_g.step()
+    scheduler_d.step()
+
+
+def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
+  net_g, net_d = nets
+  optim_g, optim_d = optims
+  scheduler_g, scheduler_d = schedulers
+  train_loader, eval_loader = loaders
+  if writers is not None:
+    writer, writer_eval = writers
+
+  train_loader.batch_sampler.set_epoch(epoch)
+  global global_step
+
+  net_g.train()
+  net_d.train()
+  for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths) in enumerate(train_loader):
+    x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
+    spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
+    y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
+
+    with autocast(enabled=hps.train.fp16_run):
+      y_hat, l_length, attn, ids_slice, x_mask, z_mask,\
+      (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(x, x_lengths, spec, spec_lengths)
+
+      mel = spec_to_mel_torch(
+          spec, 
+          hps.data.filter_length, 
+          hps.data.n_mel_channels, 
+          hps.data.sampling_rate,
+          hps.data.mel_fmin, 
+          hps.data.mel_fmax)
+      y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
+      y_hat_mel = mel_spectrogram_torch(
+          y_hat.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
+      )
+
+      y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice 
+
+      # Discriminator
+      y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
+      with autocast(enabled=False):
+        loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
+        loss_disc_all = loss_disc
+    optim_d.zero_grad()
+    scaler.scale(loss_disc_all).backward()
+    scaler.unscale_(optim_d)
+    grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
+    scaler.step(optim_d)
+
+    with autocast(enabled=hps.train.fp16_run):
+      # Generator
+      y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
+      with autocast(enabled=False):
+        loss_dur = torch.sum(l_length.float())
+        loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
+        loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
+
+        loss_fm = feature_loss(fmap_r, fmap_g)
+        loss_gen, losses_gen = generator_loss(y_d_hat_g)
+        loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
+    optim_g.zero_grad()
+    scaler.scale(loss_gen_all).backward()
+    scaler.unscale_(optim_g)
+    grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
+    scaler.step(optim_g)
+    scaler.update()
+
+    if rank==0:
+      if global_step % hps.train.log_interval == 0:
+        lr = optim_g.param_groups[0]['lr']
+        losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
+        logger.info('Train Epoch: {} [{:.0f}%]'.format(
+          epoch,
+          100. * batch_idx / len(train_loader)))
+        logger.info([x.item() for x in losses] + [global_step, lr])
+        
+        scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
+        scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "loss/g/kl": loss_kl})
+
+        scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
+        scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
+        scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
+        image_dict = { 
+            "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
+            "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), 
+            "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
+            "all/attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy())
+        }
+        utils.summarize(
+          writer=writer,
+          global_step=global_step, 
+          images=image_dict,
+          scalars=scalar_dict)
+
+      if global_step % hps.train.eval_interval == 0:
+        evaluate(hps, net_g, eval_loader, writer_eval)
+        utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
+        utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
+    global_step += 1
+  
+  if rank == 0:
+    logger.info('====> Epoch: {}'.format(epoch))
+
+ 
+def evaluate(hps, generator, eval_loader, writer_eval):
+    generator.eval()
+    with torch.no_grad():
+      for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths) in enumerate(eval_loader):
+        x, x_lengths = x.cuda(0), x_lengths.cuda(0)
+        spec, spec_lengths = spec.cuda(0), spec_lengths.cuda(0)
+        y, y_lengths = y.cuda(0), y_lengths.cuda(0)
+
+        # remove else
+        x = x[:1]
+        x_lengths = x_lengths[:1]
+        spec = spec[:1]
+        spec_lengths = spec_lengths[:1]
+        y = y[:1]
+        y_lengths = y_lengths[:1]
+        break
+      y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, max_len=1000)
+      y_hat_lengths = mask.sum([1,2]).long() * hps.data.hop_length
+
+      mel = spec_to_mel_torch(
+        spec, 
+        hps.data.filter_length, 
+        hps.data.n_mel_channels, 
+        hps.data.sampling_rate,
+        hps.data.mel_fmin, 
+        hps.data.mel_fmax)
+      y_hat_mel = mel_spectrogram_torch(
+        y_hat.squeeze(1).float(),
+        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
+      )
+    image_dict = {
+      "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
+    }
+    audio_dict = {
+      "gen/audio": y_hat[0,:,:y_hat_lengths[0]]
+    }
+    if global_step == 0:
+      image_dict.update({"gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
+      audio_dict.update({"gt/audio": y[0,:,:y_lengths[0]]})
+
+    utils.summarize(
+      writer=writer_eval,
+      global_step=global_step, 
+      images=image_dict,
+      audios=audio_dict,
+      audio_sampling_rate=hps.data.sampling_rate
+    )
+    generator.train()
+
+                           
+if __name__ == "__main__":
+  main()

train_ms.py

@@ -0,0 +1,294 @@
+import os
+import json
+import argparse
+import itertools
+import math
+import torch
+from torch import nn, optim
+from torch.nn import functional as F
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+import torch.multiprocessing as mp
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.cuda.amp import autocast, GradScaler
+
+import commons
+import utils
+from data_utils import (
+  TextAudioSpeakerLoader,
+  TextAudioSpeakerCollate,
+  DistributedBucketSampler
+)
+from models import (
+  SynthesizerTrn,
+  MultiPeriodDiscriminator,
+)
+from losses import (
+  generator_loss,
+  discriminator_loss,
+  feature_loss,
+  kl_loss
+)
+from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from text.symbols import symbols
+
+
+torch.backends.cudnn.benchmark = True
+global_step = 0
+
+
+def main():
+  """Assume Single Node Multi GPUs Training Only"""
+  assert torch.cuda.is_available(), "CPU training is not allowed."
+
+  n_gpus = torch.cuda.device_count()
+  os.environ['MASTER_ADDR'] = 'localhost'
+  os.environ['MASTER_PORT'] = '80000'
+
+  hps = utils.get_hparams()
+  mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
+
+
+def run(rank, n_gpus, hps):
+  global global_step
+  if rank == 0:
+    logger = utils.get_logger(hps.model_dir)
+    logger.info(hps)
+    utils.check_git_hash(hps.model_dir)
+    writer = SummaryWriter(log_dir=hps.model_dir)
+    writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+  dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
+  torch.manual_seed(hps.train.seed)
+  torch.cuda.set_device(rank)
+
+  train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
+  train_sampler = DistributedBucketSampler(
+      train_dataset,
+      hps.train.batch_size,
+      [32,300,400,500,600,700,800,900,1000],
+      num_replicas=n_gpus,
+      rank=rank,
+      shuffle=True)
+  collate_fn = TextAudioSpeakerCollate()
+  train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
+      collate_fn=collate_fn, batch_sampler=train_sampler)
+  if rank == 0:
+    eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
+    eval_loader = DataLoader(eval_dataset, num_workers=8, shuffle=False,
+        batch_size=hps.train.batch_size, pin_memory=True,
+        drop_last=False, collate_fn=collate_fn)
+
+  net_g = SynthesizerTrn(
+      len(symbols),
+      hps.data.filter_length // 2 + 1,
+      hps.train.segment_size // hps.data.hop_length,
+      n_speakers=hps.data.n_speakers,
+      **hps.model).cuda(rank)
+  net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
+  optim_g = torch.optim.AdamW(
+      net_g.parameters(), 
+      hps.train.learning_rate, 
+      betas=hps.train.betas, 
+      eps=hps.train.eps)
+  optim_d = torch.optim.AdamW(
+      net_d.parameters(),
+      hps.train.learning_rate, 
+      betas=hps.train.betas, 
+      eps=hps.train.eps)
+  net_g = DDP(net_g, device_ids=[rank])
+  net_d = DDP(net_d, device_ids=[rank])
+
+  try:
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
+    _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d)
+    global_step = (epoch_str - 1) * len(train_loader)
+  except:
+    epoch_str = 1
+    global_step = 0
+
+  scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
+  scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
+
+  scaler = GradScaler(enabled=hps.train.fp16_run)
+
+  for epoch in range(epoch_str, hps.train.epochs + 1):
+    if rank==0:
+      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, eval_loader], logger, [writer, writer_eval])
+    else:
+      train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, None], None, None)
+    scheduler_g.step()
+    scheduler_d.step()
+
+
+def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
+  net_g, net_d = nets
+  optim_g, optim_d = optims
+  scheduler_g, scheduler_d = schedulers
+  train_loader, eval_loader = loaders
+  if writers is not None:
+    writer, writer_eval = writers
+
+  train_loader.batch_sampler.set_epoch(epoch)
+  global global_step
+
+  net_g.train()
+  net_d.train()
+  for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers) in enumerate(train_loader):
+    x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
+    spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
+    y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
+    speakers = speakers.cuda(rank, non_blocking=True)
+
+    with autocast(enabled=hps.train.fp16_run):
+      y_hat, l_length, attn, ids_slice, x_mask, z_mask,\
+      (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(x, x_lengths, spec, spec_lengths, speakers)
+
+      mel = spec_to_mel_torch(
+          spec, 
+          hps.data.filter_length, 
+          hps.data.n_mel_channels, 
+          hps.data.sampling_rate,
+          hps.data.mel_fmin, 
+          hps.data.mel_fmax)
+      y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
+      y_hat_mel = mel_spectrogram_torch(
+          y_hat.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
+      )
+
+      y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice 
+
+      # Discriminator
+      y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
+      with autocast(enabled=False):
+        loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
+        loss_disc_all = loss_disc
+    optim_d.zero_grad()
+    scaler.scale(loss_disc_all).backward()
+    scaler.unscale_(optim_d)
+    grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
+    scaler.step(optim_d)
+
+    with autocast(enabled=hps.train.fp16_run):
+      # Generator
+      y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
+      with autocast(enabled=False):
+        loss_dur = torch.sum(l_length.float())
+        loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
+        loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
+
+        loss_fm = feature_loss(fmap_r, fmap_g)
+        loss_gen, losses_gen = generator_loss(y_d_hat_g)
+        loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
+    optim_g.zero_grad()
+    scaler.scale(loss_gen_all).backward()
+    scaler.unscale_(optim_g)
+    grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
+    scaler.step(optim_g)
+    scaler.update()
+
+    if rank==0:
+      if global_step % hps.train.log_interval == 0:
+        lr = optim_g.param_groups[0]['lr']
+        losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
+        logger.info('Train Epoch: {} [{:.0f}%]'.format(
+          epoch,
+          100. * batch_idx / len(train_loader)))
+        logger.info([x.item() for x in losses] + [global_step, lr])
+        
+        scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
+        scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "loss/g/kl": loss_kl})
+
+        scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
+        scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
+        scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
+        image_dict = { 
+            "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
+            "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), 
+            "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
+            "all/attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy())
+        }
+        utils.summarize(
+          writer=writer,
+          global_step=global_step, 
+          images=image_dict,
+          scalars=scalar_dict)
+
+      if global_step % hps.train.eval_interval == 0:
+        evaluate(hps, net_g, eval_loader, writer_eval)
+        utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
+        utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
+    global_step += 1
+  
+  if rank == 0:
+    logger.info('====> Epoch: {}'.format(epoch))
+
+ 
+def evaluate(hps, generator, eval_loader, writer_eval):
+    generator.eval()
+    with torch.no_grad():
+      for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers) in enumerate(eval_loader):
+        x, x_lengths = x.cuda(0), x_lengths.cuda(0)
+        spec, spec_lengths = spec.cuda(0), spec_lengths.cuda(0)
+        y, y_lengths = y.cuda(0), y_lengths.cuda(0)
+        speakers = speakers.cuda(0)
+
+        # remove else
+        x = x[:1]
+        x_lengths = x_lengths[:1]
+        spec = spec[:1]
+        spec_lengths = spec_lengths[:1]
+        y = y[:1]
+        y_lengths = y_lengths[:1]
+        speakers = speakers[:1]
+        break
+      y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, speakers, max_len=1000)
+      y_hat_lengths = mask.sum([1,2]).long() * hps.data.hop_length
+
+      mel = spec_to_mel_torch(
+        spec, 
+        hps.data.filter_length, 
+        hps.data.n_mel_channels, 
+        hps.data.sampling_rate,
+        hps.data.mel_fmin, 
+        hps.data.mel_fmax)
+      y_hat_mel = mel_spectrogram_torch(
+        y_hat.squeeze(1).float(),
+        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
+      )
+    image_dict = {
+      "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
+    }
+    audio_dict = {
+      "gen/audio": y_hat[0,:,:y_hat_lengths[0]]
+    }
+    if global_step == 0:
+      image_dict.update({"gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
+      audio_dict.update({"gt/audio": y[0,:,:y_lengths[0]]})
+
+    utils.summarize(
+      writer=writer_eval,
+      global_step=global_step, 
+      images=image_dict,
+      audios=audio_dict,
+      audio_sampling_rate=hps.data.sampling_rate
+    )
+    generator.train()
+
+                           
+if __name__ == "__main__":
+  main()