fighter-programmer/voicegen

1import math2import torch3from torch import nn4from torch.nn import functional as F5 6import commons7from modules import LayerNorm8   9 10class Encoder(nn.Module):11  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):12    super().__init__()13    self.hidden_channels = hidden_channels14    self.filter_channels = filter_channels15    self.n_heads = n_heads16    self.n_layers = n_layers17    self.kernel_size = kernel_size18    self.p_dropout = p_dropout19    self.window_size = window_size20 21    self.drop = nn.Dropout(p_dropout)22    self.attn_layers = nn.ModuleList()23    self.norm_layers_1 = nn.ModuleList()24    self.ffn_layers = nn.ModuleList()25    self.norm_layers_2 = nn.ModuleList()26    for i in range(self.n_layers):27      self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))28      self.norm_layers_1.append(LayerNorm(hidden_channels))29      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))30      self.norm_layers_2.append(LayerNorm(hidden_channels))31 32  def forward(self, x, x_mask):33    attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)34    x = x * x_mask35    for i in range(self.n_layers):36      y = self.attn_layers[i](x, x, attn_mask)37      y = self.drop(y)38      x = self.norm_layers_1[i](x + y)39 40      y = self.ffn_layers[i](x, x_mask)41      y = self.drop(y)42      x = self.norm_layers_2[i](x + y)43    x = x * x_mask44    return x45 46 47class Decoder(nn.Module):48  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):49    super().__init__()50    self.hidden_channels = hidden_channels51    self.filter_channels = filter_channels52    self.n_heads = n_heads53    self.n_layers = n_layers54    self.kernel_size = kernel_size55    self.p_dropout = p_dropout56    self.proximal_bias = proximal_bias57    self.proximal_init = proximal_init58 59    self.drop = nn.Dropout(p_dropout)60    self.self_attn_layers = nn.ModuleList()61    self.norm_layers_0 = nn.ModuleList()62    self.encdec_attn_layers = nn.ModuleList()63    self.norm_layers_1 = nn.ModuleList()64    self.ffn_layers = nn.ModuleList()65    self.norm_layers_2 = nn.ModuleList()66    for i in range(self.n_layers):67      self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))68      self.norm_layers_0.append(LayerNorm(hidden_channels))69      self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))70      self.norm_layers_1.append(LayerNorm(hidden_channels))71      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))72      self.norm_layers_2.append(LayerNorm(hidden_channels))73 74  def forward(self, x, x_mask, h, h_mask):75    """76    x: decoder input77    h: encoder output78    """79    self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)80    encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)81    x = x * x_mask82    for i in range(self.n_layers):83      y = self.self_attn_layers[i](x, x, self_attn_mask)84      y = self.drop(y)85      x = self.norm_layers_0[i](x + y)86 87      y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)88      y = self.drop(y)89      x = self.norm_layers_1[i](x + y)90      91      y = self.ffn_layers[i](x, x_mask)92      y = self.drop(y)93      x = self.norm_layers_2[i](x + y)94    x = x * x_mask95    return x96 97 98class MultiHeadAttention(nn.Module):99  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):100    super().__init__()101    assert channels % n_heads == 0102 103    self.channels = channels104    self.out_channels = out_channels105    self.n_heads = n_heads106    self.p_dropout = p_dropout107    self.window_size = window_size108    self.heads_share = heads_share109    self.block_length = block_length110    self.proximal_bias = proximal_bias111    self.proximal_init = proximal_init112    self.attn = None113 114    self.k_channels = channels // n_heads115    self.conv_q = nn.Conv1d(channels, channels, 1)116    self.conv_k = nn.Conv1d(channels, channels, 1)117    self.conv_v = nn.Conv1d(channels, channels, 1)118    self.conv_o = nn.Conv1d(channels, out_channels, 1)119    self.drop = nn.Dropout(p_dropout)120 121    if window_size is not None:122      n_heads_rel = 1 if heads_share else n_heads123      rel_stddev = self.k_channels**-0.5124      self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)125      self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)126 127    nn.init.xavier_uniform_(self.conv_q.weight)128    nn.init.xavier_uniform_(self.conv_k.weight)129    nn.init.xavier_uniform_(self.conv_v.weight)130    if proximal_init:131      with torch.no_grad():132        self.conv_k.weight.copy_(self.conv_q.weight)133        self.conv_k.bias.copy_(self.conv_q.bias)134      135  def forward(self, x, c, attn_mask=None):136    q = self.conv_q(x)137    k = self.conv_k(c)138    v = self.conv_v(c)139    140    x, self.attn = self.attention(q, k, v, mask=attn_mask)141 142    x = self.conv_o(x)143    return x144 145  def attention(self, query, key, value, mask=None):146    # reshape [b, d, t] -> [b, n_h, t, d_k]147    b, d, t_s, t_t = (*key.size(), query.size(2))148    query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)149    key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)150    value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)151 152    scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))153    if self.window_size is not None:154      assert t_s == t_t, "Relative attention is only available for self-attention."155      key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)156      rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)157      scores_local = self._relative_position_to_absolute_position(rel_logits)158      scores = scores + scores_local159    if self.proximal_bias:160      assert t_s == t_t, "Proximal bias is only available for self-attention."161      scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)162    if mask is not None:163      scores = scores.masked_fill(mask == 0, -1e4)164      if self.block_length is not None:165        assert t_s == t_t, "Local attention is only available for self-attention."166        block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)167        scores = scores.masked_fill(block_mask == 0, -1e4)168    p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]169    p_attn = self.drop(p_attn)170    output = torch.matmul(p_attn, value)171    if self.window_size is not None:172      relative_weights = self._absolute_position_to_relative_position(p_attn)173      value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)174      output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)175    output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]176    return output, p_attn177 178  def _matmul_with_relative_values(self, x, y):179    """180    x: [b, h, l, m]181    y: [h or 1, m, d]182    ret: [b, h, l, d]183    """184    ret = torch.matmul(x, y.unsqueeze(0))185    return ret186 187  def _matmul_with_relative_keys(self, x, y):188    """189    x: [b, h, l, d]190    y: [h or 1, m, d]191    ret: [b, h, l, m]192    """193    ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))194    return ret195 196  def _get_relative_embeddings(self, relative_embeddings, length):197    max_relative_position = 2 * self.window_size + 1198    # Pad first before slice to avoid using cond ops.199    pad_length = max(length - (self.window_size + 1), 0)200    slice_start_position = max((self.window_size + 1) - length, 0)201    slice_end_position = slice_start_position + 2 * length - 1202    if pad_length > 0:203      padded_relative_embeddings = F.pad(204          relative_embeddings,205          commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))206    else:207      padded_relative_embeddings = relative_embeddings208    used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]209    return used_relative_embeddings210 211  def _relative_position_to_absolute_position(self, x):212    """213    x: [b, h, l, 2*l-1]214    ret: [b, h, l, l]215    """216    batch, heads, length, _ = x.size()217    # Concat columns of pad to shift from relative to absolute indexing.218    x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))219 220    # Concat extra elements so to add up to shape (len+1, 2*len-1).221    x_flat = x.view([batch, heads, length * 2 * length])222    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))223 224    # Reshape and slice out the padded elements.225    x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]226    return x_final227 228  def _absolute_position_to_relative_position(self, x):229    """230    x: [b, h, l, l]231    ret: [b, h, l, 2*l-1]232    """233    batch, heads, length, _ = x.size()234    # padd along column235    x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))236    x_flat = x.view([batch, heads, length**2 + length*(length -1)])237    # add 0's in the beginning that will skew the elements after reshape238    x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))239    x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]240    return x_final241 242  def _attention_bias_proximal(self, length):243    """Bias for self-attention to encourage attention to close positions.244    Args:245      length: an integer scalar.246    Returns:247      a Tensor with shape [1, 1, length, length]248    """249    r = torch.arange(length, dtype=torch.float32)250    diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)251    return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)252 253 254class FFN(nn.Module):255  def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):256    super().__init__()257    self.in_channels = in_channels258    self.out_channels = out_channels259    self.filter_channels = filter_channels260    self.kernel_size = kernel_size261    self.p_dropout = p_dropout262    self.activation = activation263    self.causal = causal264 265    if causal:266      self.padding = self._causal_padding267    else:268      self.padding = self._same_padding269 270    self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)271    self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)272    self.drop = nn.Dropout(p_dropout)273 274  def forward(self, x, x_mask):275    x = self.conv_1(self.padding(x * x_mask))276    if self.activation == "gelu":277      x = x * torch.sigmoid(1.702 * x)278    else:279      x = torch.relu(x)280    x = self.drop(x)281    x = self.conv_2(self.padding(x * x_mask))282    return x * x_mask283  284  def _causal_padding(self, x):285    if self.kernel_size == 1:286      return x287    pad_l = self.kernel_size - 1288    pad_r = 0289    padding = [[0, 0], [0, 0], [pad_l, pad_r]]290    x = F.pad(x, commons.convert_pad_shape(padding))291    return x292 293  def _same_padding(self, x):294    if self.kernel_size == 1:295      return x296    pad_l = (self.kernel_size - 1) // 2297    pad_r = self.kernel_size // 2298    padding = [[0, 0], [0, 0], [pad_l, pad_r]]299    x = F.pad(x, commons.convert_pad_shape(padding))300    return x301