File size: 11,734 Bytes
bd603fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
import math
import torch
from torch import nn
from torch.nn import functional as F

import commons
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