reha commited on
Commit
5d50ffe
1 Parent(s): e700840

Upload 30 files

Browse files
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2021 Jingyi Li
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
add_speaker.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ from tqdm import tqdm
4
+ from random import shuffle
5
+ import json
6
+
7
+
8
+ if __name__ == "__main__":
9
+ parser = argparse.ArgumentParser()
10
+ parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list")
11
+ parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list")
12
+ parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list")
13
+ parser.add_argument("--source_dir", type=str, default="./dataset/32k", help="path to source dir")
14
+ args = parser.parse_args()
15
+
16
+ previous_config = json.load(open("configs/config.json", "rb"))
17
+
18
+ train = []
19
+ val = []
20
+ test = []
21
+ idx = 0
22
+ spk_dict = previous_config["spk"]
23
+ spk_id = max([i for i in spk_dict.values()]) + 1
24
+ for speaker in tqdm(os.listdir(args.source_dir)):
25
+ if speaker not in spk_dict.keys():
26
+ spk_dict[speaker] = spk_id
27
+ spk_id += 1
28
+ wavs = [os.path.join(args.source_dir, speaker, i)for i in os.listdir(os.path.join(args.source_dir, speaker))]
29
+ wavs = [i for i in wavs if i.endswith("wav")]
30
+ shuffle(wavs)
31
+ train += wavs[2:-10]
32
+ val += wavs[:2]
33
+ test += wavs[-10:]
34
+
35
+ assert previous_config["model"]["n_speakers"] > len(spk_dict.keys())
36
+ shuffle(train)
37
+ shuffle(val)
38
+ shuffle(test)
39
+
40
+ print("Writing", args.train_list)
41
+ with open(args.train_list, "w") as f:
42
+ for fname in tqdm(train):
43
+ wavpath = fname
44
+ f.write(wavpath + "\n")
45
+
46
+ print("Writing", args.val_list)
47
+ with open(args.val_list, "w") as f:
48
+ for fname in tqdm(val):
49
+ wavpath = fname
50
+ f.write(wavpath + "\n")
51
+
52
+ print("Writing", args.test_list)
53
+ with open(args.test_list, "w") as f:
54
+ for fname in tqdm(test):
55
+ wavpath = fname
56
+ f.write(wavpath + "\n")
57
+
58
+ previous_config["spk"] = spk_dict
59
+
60
+ print("Writing configs/config.json")
61
+ with open("configs/config.json", "w") as f:
62
+ json.dump(previous_config, f, indent=2)
attentions.py ADDED
@@ -0,0 +1,303 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import torch
5
+ from torch import nn
6
+ from torch.nn import functional as F
7
+
8
+ import commons
9
+ import modules
10
+ from modules import LayerNorm
11
+
12
+
13
+ class Encoder(nn.Module):
14
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
15
+ super().__init__()
16
+ self.hidden_channels = hidden_channels
17
+ self.filter_channels = filter_channels
18
+ self.n_heads = n_heads
19
+ self.n_layers = n_layers
20
+ self.kernel_size = kernel_size
21
+ self.p_dropout = p_dropout
22
+ self.window_size = window_size
23
+
24
+ self.drop = nn.Dropout(p_dropout)
25
+ self.attn_layers = nn.ModuleList()
26
+ self.norm_layers_1 = nn.ModuleList()
27
+ self.ffn_layers = nn.ModuleList()
28
+ self.norm_layers_2 = nn.ModuleList()
29
+ for i in range(self.n_layers):
30
+ self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
31
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
32
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
33
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
34
+
35
+ def forward(self, x, x_mask):
36
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
37
+ x = x * x_mask
38
+ for i in range(self.n_layers):
39
+ y = self.attn_layers[i](x, x, attn_mask)
40
+ y = self.drop(y)
41
+ x = self.norm_layers_1[i](x + y)
42
+
43
+ y = self.ffn_layers[i](x, x_mask)
44
+ y = self.drop(y)
45
+ x = self.norm_layers_2[i](x + y)
46
+ x = x * x_mask
47
+ return x
48
+
49
+
50
+ class Decoder(nn.Module):
51
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
52
+ super().__init__()
53
+ self.hidden_channels = hidden_channels
54
+ self.filter_channels = filter_channels
55
+ self.n_heads = n_heads
56
+ self.n_layers = n_layers
57
+ self.kernel_size = kernel_size
58
+ self.p_dropout = p_dropout
59
+ self.proximal_bias = proximal_bias
60
+ self.proximal_init = proximal_init
61
+
62
+ self.drop = nn.Dropout(p_dropout)
63
+ self.self_attn_layers = nn.ModuleList()
64
+ self.norm_layers_0 = nn.ModuleList()
65
+ self.encdec_attn_layers = nn.ModuleList()
66
+ self.norm_layers_1 = nn.ModuleList()
67
+ self.ffn_layers = nn.ModuleList()
68
+ self.norm_layers_2 = nn.ModuleList()
69
+ for i in range(self.n_layers):
70
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
71
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
72
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
73
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
74
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
75
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
76
+
77
+ def forward(self, x, x_mask, h, h_mask):
78
+ """
79
+ x: decoder input
80
+ h: encoder output
81
+ """
82
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
83
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
84
+ x = x * x_mask
85
+ for i in range(self.n_layers):
86
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
87
+ y = self.drop(y)
88
+ x = self.norm_layers_0[i](x + y)
89
+
90
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
91
+ y = self.drop(y)
92
+ x = self.norm_layers_1[i](x + y)
93
+
94
+ y = self.ffn_layers[i](x, x_mask)
95
+ y = self.drop(y)
96
+ x = self.norm_layers_2[i](x + y)
97
+ x = x * x_mask
98
+ return x
99
+
100
+
101
+ class MultiHeadAttention(nn.Module):
102
+ 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):
103
+ super().__init__()
104
+ assert channels % n_heads == 0
105
+
106
+ self.channels = channels
107
+ self.out_channels = out_channels
108
+ self.n_heads = n_heads
109
+ self.p_dropout = p_dropout
110
+ self.window_size = window_size
111
+ self.heads_share = heads_share
112
+ self.block_length = block_length
113
+ self.proximal_bias = proximal_bias
114
+ self.proximal_init = proximal_init
115
+ self.attn = None
116
+
117
+ self.k_channels = channels // n_heads
118
+ self.conv_q = nn.Conv1d(channels, channels, 1)
119
+ self.conv_k = nn.Conv1d(channels, channels, 1)
120
+ self.conv_v = nn.Conv1d(channels, channels, 1)
121
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
122
+ self.drop = nn.Dropout(p_dropout)
123
+
124
+ if window_size is not None:
125
+ n_heads_rel = 1 if heads_share else n_heads
126
+ rel_stddev = self.k_channels**-0.5
127
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
128
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
129
+
130
+ nn.init.xavier_uniform_(self.conv_q.weight)
131
+ nn.init.xavier_uniform_(self.conv_k.weight)
132
+ nn.init.xavier_uniform_(self.conv_v.weight)
133
+ if proximal_init:
134
+ with torch.no_grad():
135
+ self.conv_k.weight.copy_(self.conv_q.weight)
136
+ self.conv_k.bias.copy_(self.conv_q.bias)
137
+
138
+ def forward(self, x, c, attn_mask=None):
139
+ q = self.conv_q(x)
140
+ k = self.conv_k(c)
141
+ v = self.conv_v(c)
142
+
143
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
144
+
145
+ x = self.conv_o(x)
146
+ return x
147
+
148
+ def attention(self, query, key, value, mask=None):
149
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
150
+ b, d, t_s, t_t = (*key.size(), query.size(2))
151
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
152
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
153
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
154
+
155
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
156
+ if self.window_size is not None:
157
+ assert t_s == t_t, "Relative attention is only available for self-attention."
158
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
159
+ rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
160
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
161
+ scores = scores + scores_local
162
+ if self.proximal_bias:
163
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
164
+ scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
165
+ if mask is not None:
166
+ scores = scores.masked_fill(mask == 0, -1e4)
167
+ if self.block_length is not None:
168
+ assert t_s == t_t, "Local attention is only available for self-attention."
169
+ block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
170
+ scores = scores.masked_fill(block_mask == 0, -1e4)
171
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
172
+ p_attn = self.drop(p_attn)
173
+ output = torch.matmul(p_attn, value)
174
+ if self.window_size is not None:
175
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
176
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
177
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
178
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
179
+ return output, p_attn
180
+
181
+ def _matmul_with_relative_values(self, x, y):
182
+ """
183
+ x: [b, h, l, m]
184
+ y: [h or 1, m, d]
185
+ ret: [b, h, l, d]
186
+ """
187
+ ret = torch.matmul(x, y.unsqueeze(0))
188
+ return ret
189
+
190
+ def _matmul_with_relative_keys(self, x, y):
191
+ """
192
+ x: [b, h, l, d]
193
+ y: [h or 1, m, d]
194
+ ret: [b, h, l, m]
195
+ """
196
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
197
+ return ret
198
+
199
+ def _get_relative_embeddings(self, relative_embeddings, length):
200
+ max_relative_position = 2 * self.window_size + 1
201
+ # Pad first before slice to avoid using cond ops.
202
+ pad_length = max(length - (self.window_size + 1), 0)
203
+ slice_start_position = max((self.window_size + 1) - length, 0)
204
+ slice_end_position = slice_start_position + 2 * length - 1
205
+ if pad_length > 0:
206
+ padded_relative_embeddings = F.pad(
207
+ relative_embeddings,
208
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
209
+ else:
210
+ padded_relative_embeddings = relative_embeddings
211
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
212
+ return used_relative_embeddings
213
+
214
+ def _relative_position_to_absolute_position(self, x):
215
+ """
216
+ x: [b, h, l, 2*l-1]
217
+ ret: [b, h, l, l]
218
+ """
219
+ batch, heads, length, _ = x.size()
220
+ # Concat columns of pad to shift from relative to absolute indexing.
221
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
222
+
223
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
224
+ x_flat = x.view([batch, heads, length * 2 * length])
225
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
226
+
227
+ # Reshape and slice out the padded elements.
228
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
229
+ return x_final
230
+
231
+ def _absolute_position_to_relative_position(self, x):
232
+ """
233
+ x: [b, h, l, l]
234
+ ret: [b, h, l, 2*l-1]
235
+ """
236
+ batch, heads, length, _ = x.size()
237
+ # padd along column
238
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
239
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
240
+ # add 0's in the beginning that will skew the elements after reshape
241
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
242
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
243
+ return x_final
244
+
245
+ def _attention_bias_proximal(self, length):
246
+ """Bias for self-attention to encourage attention to close positions.
247
+ Args:
248
+ length: an integer scalar.
249
+ Returns:
250
+ a Tensor with shape [1, 1, length, length]
251
+ """
252
+ r = torch.arange(length, dtype=torch.float32)
253
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
254
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
255
+
256
+
257
+ class FFN(nn.Module):
258
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
259
+ super().__init__()
260
+ self.in_channels = in_channels
261
+ self.out_channels = out_channels
262
+ self.filter_channels = filter_channels
263
+ self.kernel_size = kernel_size
264
+ self.p_dropout = p_dropout
265
+ self.activation = activation
266
+ self.causal = causal
267
+
268
+ if causal:
269
+ self.padding = self._causal_padding
270
+ else:
271
+ self.padding = self._same_padding
272
+
273
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
274
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
275
+ self.drop = nn.Dropout(p_dropout)
276
+
277
+ def forward(self, x, x_mask):
278
+ x = self.conv_1(self.padding(x * x_mask))
279
+ if self.activation == "gelu":
280
+ x = x * torch.sigmoid(1.702 * x)
281
+ else:
282
+ x = torch.relu(x)
283
+ x = self.drop(x)
284
+ x = self.conv_2(self.padding(x * x_mask))
285
+ return x * x_mask
286
+
287
+ def _causal_padding(self, x):
288
+ if self.kernel_size == 1:
289
+ return x
290
+ pad_l = self.kernel_size - 1
291
+ pad_r = 0
292
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
293
+ x = F.pad(x, commons.convert_pad_shape(padding))
294
+ return x
295
+
296
+ def _same_padding(self, x):
297
+ if self.kernel_size == 1:
298
+ return x
299
+ pad_l = (self.kernel_size - 1) // 2
300
+ pad_r = self.kernel_size // 2
301
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
302
+ x = F.pad(x, commons.convert_pad_shape(padding))
303
+ return x
commons.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import numpy as np
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ def slice_pitch_segments(x, ids_str, segment_size=4):
8
+ ret = torch.zeros_like(x[:, :segment_size])
9
+ for i in range(x.size(0)):
10
+ idx_str = ids_str[i]
11
+ idx_end = idx_str + segment_size
12
+ ret[i] = x[i, idx_str:idx_end]
13
+ return ret
14
+
15
+ def rand_slice_segments_with_pitch(x, pitch, x_lengths=None, segment_size=4):
16
+ b, d, t = x.size()
17
+ if x_lengths is None:
18
+ x_lengths = t
19
+ ids_str_max = x_lengths - segment_size + 1
20
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
21
+ ret = slice_segments(x, ids_str, segment_size)
22
+ ret_pitch = slice_pitch_segments(pitch, ids_str, segment_size)
23
+ return ret, ret_pitch, ids_str
24
+
25
+ def init_weights(m, mean=0.0, std=0.01):
26
+ classname = m.__class__.__name__
27
+ if classname.find("Conv") != -1:
28
+ m.weight.data.normal_(mean, std)
29
+
30
+
31
+ def get_padding(kernel_size, dilation=1):
32
+ return int((kernel_size*dilation - dilation)/2)
33
+
34
+
35
+ def convert_pad_shape(pad_shape):
36
+ l = pad_shape[::-1]
37
+ pad_shape = [item for sublist in l for item in sublist]
38
+ return pad_shape
39
+
40
+
41
+ def intersperse(lst, item):
42
+ result = [item] * (len(lst) * 2 + 1)
43
+ result[1::2] = lst
44
+ return result
45
+
46
+
47
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
48
+ """KL(P||Q)"""
49
+ kl = (logs_q - logs_p) - 0.5
50
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
51
+ return kl
52
+
53
+
54
+ def rand_gumbel(shape):
55
+ """Sample from the Gumbel distribution, protect from overflows."""
56
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
57
+ return -torch.log(-torch.log(uniform_samples))
58
+
59
+
60
+ def rand_gumbel_like(x):
61
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
62
+ return g
63
+
64
+
65
+ def slice_segments(x, ids_str, segment_size=4):
66
+ ret = torch.zeros_like(x[:, :, :segment_size])
67
+ for i in range(x.size(0)):
68
+ idx_str = ids_str[i]
69
+ idx_end = idx_str + segment_size
70
+ ret[i] = x[i, :, idx_str:idx_end]
71
+ return ret
72
+
73
+
74
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
75
+ b, d, t = x.size()
76
+ if x_lengths is None:
77
+ x_lengths = t
78
+ ids_str_max = x_lengths - segment_size + 1
79
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
80
+ ret = slice_segments(x, ids_str, segment_size)
81
+ return ret, ids_str
82
+
83
+
84
+ def rand_spec_segments(x, x_lengths=None, segment_size=4):
85
+ b, d, t = x.size()
86
+ if x_lengths is None:
87
+ x_lengths = t
88
+ ids_str_max = x_lengths - segment_size
89
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
90
+ ret = slice_segments(x, ids_str, segment_size)
91
+ return ret, ids_str
92
+
93
+
94
+ def get_timing_signal_1d(
95
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
96
+ position = torch.arange(length, dtype=torch.float)
97
+ num_timescales = channels // 2
98
+ log_timescale_increment = (
99
+ math.log(float(max_timescale) / float(min_timescale)) /
100
+ (num_timescales - 1))
101
+ inv_timescales = min_timescale * torch.exp(
102
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
103
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
104
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
105
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
106
+ signal = signal.view(1, channels, length)
107
+ return signal
108
+
109
+
110
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
111
+ b, channels, length = x.size()
112
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
113
+ return x + signal.to(dtype=x.dtype, device=x.device)
114
+
115
+
116
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
117
+ b, channels, length = x.size()
118
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
119
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
120
+
121
+
122
+ def subsequent_mask(length):
123
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
124
+ return mask
125
+
126
+
127
+ @torch.jit.script
128
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
129
+ n_channels_int = n_channels[0]
130
+ in_act = input_a + input_b
131
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
132
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
133
+ acts = t_act * s_act
134
+ return acts
135
+
136
+
137
+ def convert_pad_shape(pad_shape):
138
+ l = pad_shape[::-1]
139
+ pad_shape = [item for sublist in l for item in sublist]
140
+ return pad_shape
141
+
142
+
143
+ def shift_1d(x):
144
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
145
+ return x
146
+
147
+
148
+ def sequence_mask(length, max_length=None):
149
+ if max_length is None:
150
+ max_length = length.max()
151
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
152
+ return x.unsqueeze(0) < length.unsqueeze(1)
153
+
154
+
155
+ def generate_path(duration, mask):
156
+ """
157
+ duration: [b, 1, t_x]
158
+ mask: [b, 1, t_y, t_x]
159
+ """
160
+ device = duration.device
161
+
162
+ b, _, t_y, t_x = mask.shape
163
+ cum_duration = torch.cumsum(duration, -1)
164
+
165
+ cum_duration_flat = cum_duration.view(b * t_x)
166
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
167
+ path = path.view(b, t_x, t_y)
168
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
169
+ path = path.unsqueeze(1).transpose(2,3) * mask
170
+ return path
171
+
172
+
173
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
174
+ if isinstance(parameters, torch.Tensor):
175
+ parameters = [parameters]
176
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
177
+ norm_type = float(norm_type)
178
+ if clip_value is not None:
179
+ clip_value = float(clip_value)
180
+
181
+ total_norm = 0
182
+ for p in parameters:
183
+ param_norm = p.grad.data.norm(norm_type)
184
+ total_norm += param_norm.item() ** norm_type
185
+ if clip_value is not None:
186
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
187
+ total_norm = total_norm ** (1. / norm_type)
188
+ return total_norm
data_utils.py ADDED
@@ -0,0 +1,152 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import time
2
+ import os
3
+ import random
4
+ import numpy as np
5
+ import torch
6
+ import torch.utils.data
7
+
8
+ import commons
9
+ from mel_processing import spectrogram_torch, spec_to_mel_torch
10
+ from utils import load_wav_to_torch, load_filepaths_and_text, transform
11
+
12
+ # import h5py
13
+
14
+
15
+ """Multi speaker version"""
16
+
17
+
18
+ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
19
+ """
20
+ 1) loads audio, speaker_id, text pairs
21
+ 2) normalizes text and converts them to sequences of integers
22
+ 3) computes spectrograms from audio files.
23
+ """
24
+
25
+ def __init__(self, audiopaths, hparams):
26
+ self.audiopaths = load_filepaths_and_text(audiopaths)
27
+ self.max_wav_value = hparams.data.max_wav_value
28
+ self.sampling_rate = hparams.data.sampling_rate
29
+ self.filter_length = hparams.data.filter_length
30
+ self.hop_length = hparams.data.hop_length
31
+ self.win_length = hparams.data.win_length
32
+ self.sampling_rate = hparams.data.sampling_rate
33
+ self.use_sr = hparams.train.use_sr
34
+ self.spec_len = hparams.train.max_speclen
35
+ self.spk_map = hparams.spk
36
+
37
+ random.seed(1234)
38
+ random.shuffle(self.audiopaths)
39
+
40
+ def get_audio(self, filename):
41
+ audio, sampling_rate = load_wav_to_torch(filename)
42
+ if sampling_rate != self.sampling_rate:
43
+ raise ValueError("{} SR doesn't match target {} SR".format(
44
+ sampling_rate, self.sampling_rate))
45
+ audio_norm = audio / self.max_wav_value
46
+ audio_norm = audio_norm.unsqueeze(0)
47
+ spec_filename = filename.replace(".wav", ".spec.pt")
48
+ if os.path.exists(spec_filename):
49
+ spec = torch.load(spec_filename)
50
+ else:
51
+ spec = spectrogram_torch(audio_norm, self.filter_length,
52
+ self.sampling_rate, self.hop_length, self.win_length,
53
+ center=False)
54
+ spec = torch.squeeze(spec, 0)
55
+ torch.save(spec, spec_filename)
56
+
57
+ spk = filename.split(os.sep)[-2]
58
+ spk = torch.LongTensor([self.spk_map[spk]])
59
+
60
+ c = torch.load(filename + ".soft.pt").squeeze(0)
61
+ c = torch.repeat_interleave(c, repeats=2, dim=1)
62
+
63
+ f0 = np.load(filename + ".f0.npy")
64
+ f0 = torch.FloatTensor(f0)
65
+ lmin = min(c.size(-1), spec.size(-1), f0.shape[0])
66
+ assert abs(c.size(-1) - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape, filename)
67
+ assert abs(lmin - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape)
68
+ assert abs(lmin - c.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape)
69
+ spec, c, f0 = spec[:, :lmin], c[:, :lmin], f0[:lmin]
70
+ audio_norm = audio_norm[:, :lmin * self.hop_length]
71
+ _spec, _c, _audio_norm, _f0 = spec, c, audio_norm, f0
72
+ while spec.size(-1) < self.spec_len:
73
+ spec = torch.cat((spec, _spec), -1)
74
+ c = torch.cat((c, _c), -1)
75
+ f0 = torch.cat((f0, _f0), -1)
76
+ audio_norm = torch.cat((audio_norm, _audio_norm), -1)
77
+ start = random.randint(0, spec.size(-1) - self.spec_len)
78
+ end = start + self.spec_len
79
+ spec = spec[:, start:end]
80
+ c = c[:, start:end]
81
+ f0 = f0[start:end]
82
+ audio_norm = audio_norm[:, start * self.hop_length:end * self.hop_length]
83
+
84
+ return c, f0, spec, audio_norm, spk
85
+
86
+ def __getitem__(self, index):
87
+ return self.get_audio(self.audiopaths[index][0])
88
+
89
+ def __len__(self):
90
+ return len(self.audiopaths)
91
+
92
+
93
+ class EvalDataLoader(torch.utils.data.Dataset):
94
+ """
95
+ 1) loads audio, speaker_id, text pairs
96
+ 2) normalizes text and converts them to sequences of integers
97
+ 3) computes spectrograms from audio files.
98
+ """
99
+
100
+ def __init__(self, audiopaths, hparams):
101
+ self.audiopaths = load_filepaths_and_text(audiopaths)
102
+ self.max_wav_value = hparams.data.max_wav_value
103
+ self.sampling_rate = hparams.data.sampling_rate
104
+ self.filter_length = hparams.data.filter_length
105
+ self.hop_length = hparams.data.hop_length
106
+ self.win_length = hparams.data.win_length
107
+ self.sampling_rate = hparams.data.sampling_rate
108
+ self.use_sr = hparams.train.use_sr
109
+ self.audiopaths = self.audiopaths[:5]
110
+ self.spk_map = hparams.spk
111
+
112
+
113
+ def get_audio(self, filename):
114
+ audio, sampling_rate = load_wav_to_torch(filename)
115
+ if sampling_rate != self.sampling_rate:
116
+ raise ValueError("{} SR doesn't match target {} SR".format(
117
+ sampling_rate, self.sampling_rate))
118
+ audio_norm = audio / self.max_wav_value
119
+ audio_norm = audio_norm.unsqueeze(0)
120
+ spec_filename = filename.replace(".wav", ".spec.pt")
121
+ if os.path.exists(spec_filename):
122
+ spec = torch.load(spec_filename)
123
+ else:
124
+ spec = spectrogram_torch(audio_norm, self.filter_length,
125
+ self.sampling_rate, self.hop_length, self.win_length,
126
+ center=False)
127
+ spec = torch.squeeze(spec, 0)
128
+ torch.save(spec, spec_filename)
129
+
130
+ spk = filename.split(os.sep)[-2]
131
+ spk = torch.LongTensor([self.spk_map[spk]])
132
+
133
+ c = torch.load(filename + ".soft.pt").squeeze(0)
134
+
135
+ c = torch.repeat_interleave(c, repeats=2, dim=1)
136
+
137
+ f0 = np.load(filename + ".f0.npy")
138
+ f0 = torch.FloatTensor(f0)
139
+ lmin = min(c.size(-1), spec.size(-1), f0.shape[0])
140
+ assert abs(c.size(-1) - spec.size(-1)) < 4, (c.size(-1), spec.size(-1), f0.shape)
141
+ assert abs(f0.shape[0] - spec.shape[-1]) < 4, (c.size(-1), spec.size(-1), f0.shape)
142
+ spec, c, f0 = spec[:, :lmin], c[:, :lmin], f0[:lmin]
143
+ audio_norm = audio_norm[:, :lmin * self.hop_length]
144
+
145
+ return c, f0, spec, audio_norm, spk
146
+
147
+ def __getitem__(self, index):
148
+ return self.get_audio(self.audiopaths[index][0])
149
+
150
+ def __len__(self):
151
+ return len(self.audiopaths)
152
+
flask_api.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import logging
3
+
4
+ import soundfile
5
+ import torch
6
+ import torchaudio
7
+ from flask import Flask, request, send_file
8
+ from flask_cors import CORS
9
+
10
+ from inference.infer_tool import Svc, RealTimeVC
11
+
12
+ app = Flask(__name__)
13
+
14
+ CORS(app)
15
+
16
+ logging.getLogger('numba').setLevel(logging.WARNING)
17
+
18
+
19
+ @app.route("/voiceChangeModel", methods=["POST"])
20
+ def voice_change_model():
21
+ request_form = request.form
22
+ wave_file = request.files.get("sample", None)
23
+ # 变调信息
24
+ f_pitch_change = float(request_form.get("fPitchChange", 0))
25
+ # DAW所需的采样率
26
+ daw_sample = int(float(request_form.get("sampleRate", 0)))
27
+ speaker_id = int(float(request_form.get("sSpeakId", 0)))
28
+ # http获得wav文件并转换
29
+ input_wav_path = io.BytesIO(wave_file.read())
30
+
31
+ # 模型推理
32
+ if raw_infer:
33
+ out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
34
+ tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample)
35
+ else:
36
+ out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path)
37
+ tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample)
38
+ # 返回音频
39
+ out_wav_path = io.BytesIO()
40
+ soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav")
41
+ out_wav_path.seek(0)
42
+ return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)
43
+
44
+
45
+ if __name__ == '__main__':
46
+ # 启用则为直接切片合成,False为交叉淡化方式
47
+ # vst插件调整0.3-0.5s切片时间可以降低延迟,直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音
48
+ # 自行选择能接受的方法,或将vst最大切片时间调整为1s,此处设为Ture,延迟大音质稳定一些
49
+ raw_infer = True
50
+ # 每个模型和config是唯一对应的
51
+ model_name = "logs/32k/G_174000-Copy1.pth"
52
+ config_name = "configs/config.json"
53
+ svc_model = Svc(model_name, config_name)
54
+ svc = RealTimeVC()
55
+ # 此处与vst插件对应,不建议更改
56
+ app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)
inference_main.py ADDED
@@ -0,0 +1,65 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import logging
3
+ import time
4
+ from pathlib import Path
5
+
6
+ import librosa
7
+ import numpy as np
8
+ import soundfile
9
+
10
+ from inference import infer_tool
11
+ from inference import slicer
12
+ from inference.infer_tool import Svc
13
+
14
+ logging.getLogger('numba').setLevel(logging.WARNING)
15
+ chunks_dict = infer_tool.read_temp("inference/chunks_temp.json")
16
+
17
+ model_path = "logs/48k/G_174000-Copy1.pth"
18
+ config_path = "configs/config.json"
19
+ svc_model = Svc(model_path, config_path)
20
+ infer_tool.mkdir(["raw", "results"])
21
+
22
+ # 支持多个wav文件,放在raw文件夹下
23
+ clean_names = ["君の知らない物語-src"]
24
+ trans = [-5] # 音高调整,支持正负(半音)
25
+ spk_list = ['yunhao'] # 每次同时合成多语者音色
26
+ slice_db = -40 # 默认-40,嘈杂的音频可以-30,干声保留呼吸可以-50
27
+ wav_format = 'flac' # 音频输出格式
28
+
29
+ infer_tool.fill_a_to_b(trans, clean_names)
30
+ for clean_name, tran in zip(clean_names, trans):
31
+ raw_audio_path = f"raw/{clean_name}"
32
+ if "." not in raw_audio_path:
33
+ raw_audio_path += ".wav"
34
+ infer_tool.format_wav(raw_audio_path)
35
+ wav_path = Path(raw_audio_path).with_suffix('.wav')
36
+ audio, sr = librosa.load(wav_path, mono=True, sr=None)
37
+ wav_hash = infer_tool.get_md5(audio)
38
+ if wav_hash in chunks_dict.keys():
39
+ print("load chunks from temp")
40
+ chunks = chunks_dict[wav_hash]["chunks"]
41
+ else:
42
+ chunks = slicer.cut(wav_path, db_thresh=slice_db)
43
+ print(chunks)
44
+ chunks_dict[wav_hash] = {"chunks": chunks, "time": int(time.time())}
45
+ infer_tool.write_temp("inference/chunks_temp.json", chunks_dict)
46
+ audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
47
+
48
+ for spk in spk_list:
49
+ audio = []
50
+ for (slice_tag, data) in audio_data:
51
+ print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
52
+ length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample))
53
+ raw_path = io.BytesIO()
54
+ soundfile.write(raw_path, data, audio_sr, format="wav")
55
+ raw_path.seek(0)
56
+ if slice_tag:
57
+ print('jump empty segment')
58
+ _audio = np.zeros(length)
59
+ else:
60
+ out_audio, out_sr = svc_model.infer(spk, tran, raw_path)
61
+ _audio = out_audio.cpu().numpy()
62
+ audio.extend(list(_audio))
63
+
64
+ res_path = f'./results/{clean_name}_{tran}key_{spk}.{wav_format}'
65
+ soundfile.write(res_path, audio, svc_model.target_sample, format=wav_format)
logs/32k/G_98000.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5dfe7d808cfe6fa2e424de0b47b2f7544bbae9b97ceace5bb1a892138447e89a
3
+ size 699505437
logs/32k/model in here.txt ADDED
File without changes
losses.py ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import commons
5
+
6
+
7
+ def feature_loss(fmap_r, fmap_g):
8
+ loss = 0
9
+ for dr, dg in zip(fmap_r, fmap_g):
10
+ for rl, gl in zip(dr, dg):
11
+ rl = rl.float().detach()
12
+ gl = gl.float()
13
+ loss += torch.mean(torch.abs(rl - gl))
14
+
15
+ return loss * 2
16
+
17
+
18
+ def discriminator_loss(disc_real_outputs, disc_generated_outputs):
19
+ loss = 0
20
+ r_losses = []
21
+ g_losses = []
22
+ for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
23
+ dr = dr.float()
24
+ dg = dg.float()
25
+ r_loss = torch.mean((1-dr)**2)
26
+ g_loss = torch.mean(dg**2)
27
+ loss += (r_loss + g_loss)
28
+ r_losses.append(r_loss.item())
29
+ g_losses.append(g_loss.item())
30
+
31
+ return loss, r_losses, g_losses
32
+
33
+
34
+ def generator_loss(disc_outputs):
35
+ loss = 0
36
+ gen_losses = []
37
+ for dg in disc_outputs:
38
+ dg = dg.float()
39
+ l = torch.mean((1-dg)**2)
40
+ gen_losses.append(l)
41
+ loss += l
42
+
43
+ return loss, gen_losses
44
+
45
+
46
+ def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
47
+ """
48
+ z_p, logs_q: [b, h, t_t]
49
+ m_p, logs_p: [b, h, t_t]
50
+ """
51
+ z_p = z_p.float()
52
+ logs_q = logs_q.float()
53
+ m_p = m_p.float()
54
+ logs_p = logs_p.float()
55
+ z_mask = z_mask.float()
56
+ #print(logs_p)
57
+ kl = logs_p - logs_q - 0.5
58
+ kl += 0.5 * ((z_p - m_p)**2) * torch.exp(-2. * logs_p)
59
+ kl = torch.sum(kl * z_mask)
60
+ l = kl / torch.sum(z_mask)
61
+ return l
mel_processing.py ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import os
3
+ import random
4
+ import torch
5
+ from torch import nn
6
+ import torch.nn.functional as F
7
+ import torch.utils.data
8
+ import numpy as np
9
+ import librosa
10
+ import librosa.util as librosa_util
11
+ from librosa.util import normalize, pad_center, tiny
12
+ from scipy.signal import get_window
13
+ from scipy.io.wavfile import read
14
+ from librosa.filters import mel as librosa_mel_fn
15
+
16
+ MAX_WAV_VALUE = 32768.0
17
+
18
+
19
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
20
+ """
21
+ PARAMS
22
+ ------
23
+ C: compression factor
24
+ """
25
+ return torch.log(torch.clamp(x, min=clip_val) * C)
26
+
27
+
28
+ def dynamic_range_decompression_torch(x, C=1):
29
+ """
30
+ PARAMS
31
+ ------
32
+ C: compression factor used to compress
33
+ """
34
+ return torch.exp(x) / C
35
+
36
+
37
+ def spectral_normalize_torch(magnitudes):
38
+ output = dynamic_range_compression_torch(magnitudes)
39
+ return output
40
+
41
+
42
+ def spectral_de_normalize_torch(magnitudes):
43
+ output = dynamic_range_decompression_torch(magnitudes)
44
+ return output
45
+
46
+
47
+ mel_basis = {}
48
+ hann_window = {}
49
+
50
+
51
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
52
+ if torch.min(y) < -1.:
53
+ print('min value is ', torch.min(y))
54
+ if torch.max(y) > 1.:
55
+ print('max value is ', torch.max(y))
56
+
57
+ global hann_window
58
+ dtype_device = str(y.dtype) + '_' + str(y.device)
59
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
60
+ if wnsize_dtype_device not in hann_window:
61
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
62
+
63
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
64
+ y = y.squeeze(1)
65
+
66
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
67
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
68
+
69
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
70
+ return spec
71
+
72
+
73
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
74
+ global mel_basis
75
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
76
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
77
+ if fmax_dtype_device not in mel_basis:
78
+ mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
79
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
80
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
81
+ spec = spectral_normalize_torch(spec)
82
+ return spec
83
+
84
+
85
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
86
+ if torch.min(y) < -1.:
87
+ print('min value is ', torch.min(y))
88
+ if torch.max(y) > 1.:
89
+ print('max value is ', torch.max(y))
90
+
91
+ global mel_basis, hann_window
92
+ dtype_device = str(y.dtype) + '_' + str(y.device)
93
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
94
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
95
+ if fmax_dtype_device not in mel_basis:
96
+ mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
97
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
98
+ if wnsize_dtype_device not in hann_window:
99
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
100
+
101
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
102
+ y = y.squeeze(1)
103
+
104
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
105
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
106
+
107
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
108
+
109
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
110
+ spec = spectral_normalize_torch(spec)
111
+
112
+ return spec
models.py ADDED
@@ -0,0 +1,351 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import torch
4
+ from torch import nn
5
+ from torch.nn import functional as F
6
+
7
+ import attentions
8
+ import commons
9
+ import modules
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+ from commons import init_weights, get_padding
14
+ from vdecoder.hifigan.models import Generator
15
+ from utils import f0_to_coarse
16
+
17
+ class ResidualCouplingBlock(nn.Module):
18
+ def __init__(self,
19
+ channels,
20
+ hidden_channels,
21
+ kernel_size,
22
+ dilation_rate,
23
+ n_layers,
24
+ n_flows=4,
25
+ gin_channels=0):
26
+ super().__init__()
27
+ self.channels = channels
28
+ self.hidden_channels = hidden_channels
29
+ self.kernel_size = kernel_size
30
+ self.dilation_rate = dilation_rate
31
+ self.n_layers = n_layers
32
+ self.n_flows = n_flows
33
+ self.gin_channels = gin_channels
34
+
35
+ self.flows = nn.ModuleList()
36
+ for i in range(n_flows):
37
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
38
+ self.flows.append(modules.Flip())
39
+
40
+ def forward(self, x, x_mask, g=None, reverse=False):
41
+ if not reverse:
42
+ for flow in self.flows:
43
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
44
+ else:
45
+ for flow in reversed(self.flows):
46
+ x = flow(x, x_mask, g=g, reverse=reverse)
47
+ return x
48
+
49
+
50
+ class Encoder(nn.Module):
51
+ def __init__(self,
52
+ in_channels,
53
+ out_channels,
54
+ hidden_channels,
55
+ kernel_size,
56
+ dilation_rate,
57
+ n_layers,
58
+ gin_channels=0):
59
+ super().__init__()
60
+ self.in_channels = in_channels
61
+ self.out_channels = out_channels
62
+ self.hidden_channels = hidden_channels
63
+ self.kernel_size = kernel_size
64
+ self.dilation_rate = dilation_rate
65
+ self.n_layers = n_layers
66
+ self.gin_channels = gin_channels
67
+
68
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
69
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
70
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
71
+
72
+ def forward(self, x, x_lengths, g=None):
73
+ # print(x.shape,x_lengths.shape)
74
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
75
+ x = self.pre(x) * x_mask
76
+ x = self.enc(x, x_mask, g=g)
77
+ stats = self.proj(x) * x_mask
78
+ m, logs = torch.split(stats, self.out_channels, dim=1)
79
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
80
+ return z, m, logs, x_mask
81
+
82
+
83
+ class TextEncoder(nn.Module):
84
+ def __init__(self,
85
+ in_channels,
86
+ out_channels,
87
+ hidden_channels,
88
+ kernel_size,
89
+ dilation_rate,
90
+ n_layers,
91
+ gin_channels=0,
92
+ filter_channels=None,
93
+ n_heads=None,
94
+ p_dropout=None):
95
+ super().__init__()
96
+ self.in_channels = in_channels
97
+ self.out_channels = out_channels
98
+ self.hidden_channels = hidden_channels
99
+ self.kernel_size = kernel_size
100
+ self.dilation_rate = dilation_rate
101
+ self.n_layers = n_layers
102
+ self.gin_channels = gin_channels
103
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
104
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
105
+ self.f0_emb = nn.Embedding(256, hidden_channels)
106
+
107
+ self.enc_ = attentions.Encoder(
108
+ hidden_channels,
109
+ filter_channels,
110
+ n_heads,
111
+ n_layers,
112
+ kernel_size,
113
+ p_dropout)
114
+
115
+ def forward(self, x, x_lengths, f0=None):
116
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
117
+ x = self.pre(x) * x_mask
118
+ x = x + self.f0_emb(f0).transpose(1,2)
119
+ x = self.enc_(x * x_mask, x_mask)
120
+ stats = self.proj(x) * x_mask
121
+ m, logs = torch.split(stats, self.out_channels, dim=1)
122
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
123
+
124
+ return z, m, logs, x_mask
125
+
126
+
127
+
128
+ class DiscriminatorP(torch.nn.Module):
129
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
130
+ super(DiscriminatorP, self).__init__()
131
+ self.period = period
132
+ self.use_spectral_norm = use_spectral_norm
133
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
134
+ self.convs = nn.ModuleList([
135
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
136
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
137
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
138
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
139
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
140
+ ])
141
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
142
+
143
+ def forward(self, x):
144
+ fmap = []
145
+
146
+ # 1d to 2d
147
+ b, c, t = x.shape
148
+ if t % self.period != 0: # pad first
149
+ n_pad = self.period - (t % self.period)
150
+ x = F.pad(x, (0, n_pad), "reflect")
151
+ t = t + n_pad
152
+ x = x.view(b, c, t // self.period, self.period)
153
+
154
+ for l in self.convs:
155
+ x = l(x)
156
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
157
+ fmap.append(x)
158
+ x = self.conv_post(x)
159
+ fmap.append(x)
160
+ x = torch.flatten(x, 1, -1)
161
+
162
+ return x, fmap
163
+
164
+
165
+ class DiscriminatorS(torch.nn.Module):
166
+ def __init__(self, use_spectral_norm=False):
167
+ super(DiscriminatorS, self).__init__()
168
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
169
+ self.convs = nn.ModuleList([
170
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
171
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
172
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
173
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
174
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
175
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
176
+ ])
177
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
178
+
179
+ def forward(self, x):
180
+ fmap = []
181
+
182
+ for l in self.convs:
183
+ x = l(x)
184
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
185
+ fmap.append(x)
186
+ x = self.conv_post(x)
187
+ fmap.append(x)
188
+ x = torch.flatten(x, 1, -1)
189
+
190
+ return x, fmap
191
+
192
+
193
+ class MultiPeriodDiscriminator(torch.nn.Module):
194
+ def __init__(self, use_spectral_norm=False):
195
+ super(MultiPeriodDiscriminator, self).__init__()
196
+ periods = [2,3,5,7,11]
197
+
198
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
199
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
200
+ self.discriminators = nn.ModuleList(discs)
201
+
202
+ def forward(self, y, y_hat):
203
+ y_d_rs = []
204
+ y_d_gs = []
205
+ fmap_rs = []
206
+ fmap_gs = []
207
+ for i, d in enumerate(self.discriminators):
208
+ y_d_r, fmap_r = d(y)
209
+ y_d_g, fmap_g = d(y_hat)
210
+ y_d_rs.append(y_d_r)
211
+ y_d_gs.append(y_d_g)
212
+ fmap_rs.append(fmap_r)
213
+ fmap_gs.append(fmap_g)
214
+
215
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
216
+
217
+
218
+ class SpeakerEncoder(torch.nn.Module):
219
+ def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256):
220
+ super(SpeakerEncoder, self).__init__()
221
+ self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers, batch_first=True)
222
+ self.linear = nn.Linear(model_hidden_size, model_embedding_size)
223
+ self.relu = nn.ReLU()
224
+
225
+ def forward(self, mels):
226
+ self.lstm.flatten_parameters()
227
+ _, (hidden, _) = self.lstm(mels)
228
+ embeds_raw = self.relu(self.linear(hidden[-1]))
229
+ return embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
230
+
231
+ def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
232
+ mel_slices = []
233
+ for i in range(0, total_frames-partial_frames, partial_hop):
234
+ mel_range = torch.arange(i, i+partial_frames)
235
+ mel_slices.append(mel_range)
236
+
237
+ return mel_slices
238
+
239
+ def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
240
+ mel_len = mel.size(1)
241
+ last_mel = mel[:,-partial_frames:]
242
+
243
+ if mel_len > partial_frames:
244
+ mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop)
245
+ mels = list(mel[:,s] for s in mel_slices)
246
+ mels.append(last_mel)
247
+ mels = torch.stack(tuple(mels), 0).squeeze(1)
248
+
249
+ with torch.no_grad():
250
+ partial_embeds = self(mels)
251
+ embed = torch.mean(partial_embeds, axis=0).unsqueeze(0)
252
+ #embed = embed / torch.linalg.norm(embed, 2)
253
+ else:
254
+ with torch.no_grad():
255
+ embed = self(last_mel)
256
+
257
+ return embed
258
+
259
+
260
+ class SynthesizerTrn(nn.Module):
261
+ """
262
+ Synthesizer for Training
263
+ """
264
+
265
+ def __init__(self,
266
+ spec_channels,
267
+ segment_size,
268
+ inter_channels,
269
+ hidden_channels,
270
+ filter_channels,
271
+ n_heads,
272
+ n_layers,
273
+ kernel_size,
274
+ p_dropout,
275
+ resblock,
276
+ resblock_kernel_sizes,
277
+ resblock_dilation_sizes,
278
+ upsample_rates,
279
+ upsample_initial_channel,
280
+ upsample_kernel_sizes,
281
+ gin_channels,
282
+ ssl_dim,
283
+ n_speakers,
284
+ **kwargs):
285
+
286
+ super().__init__()
287
+ self.spec_channels = spec_channels
288
+ self.inter_channels = inter_channels
289
+ self.hidden_channels = hidden_channels
290
+ self.filter_channels = filter_channels
291
+ self.n_heads = n_heads
292
+ self.n_layers = n_layers
293
+ self.kernel_size = kernel_size
294
+ self.p_dropout = p_dropout
295
+ self.resblock = resblock
296
+ self.resblock_kernel_sizes = resblock_kernel_sizes
297
+ self.resblock_dilation_sizes = resblock_dilation_sizes
298
+ self.upsample_rates = upsample_rates
299
+ self.upsample_initial_channel = upsample_initial_channel
300
+ self.upsample_kernel_sizes = upsample_kernel_sizes
301
+ self.segment_size = segment_size
302
+ self.gin_channels = gin_channels
303
+ self.ssl_dim = ssl_dim
304
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
305
+
306
+ self.enc_p_ = TextEncoder(ssl_dim, inter_channels, hidden_channels, 5, 1, 16,0, filter_channels, n_heads, p_dropout)
307
+ hps = {
308
+ "sampling_rate": 32000,
309
+ "inter_channels": 192,
310
+ "resblock": "1",
311
+ "resblock_kernel_sizes": [3, 7, 11],
312
+ "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
313
+ "upsample_rates": [10, 8, 2, 2],
314
+ "upsample_initial_channel": 512,
315
+ "upsample_kernel_sizes": [16, 16, 4, 4],
316
+ "gin_channels": 256,
317
+ }
318
+ self.dec = Generator(h=hps)
319
+ self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
320
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
321
+
322
+ def forward(self, c, f0, spec, g=None, mel=None, c_lengths=None, spec_lengths=None):
323
+ if c_lengths == None:
324
+ c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
325
+ if spec_lengths == None:
326
+ spec_lengths = (torch.ones(spec.size(0)) * spec.size(-1)).to(spec.device)
327
+
328
+ g = self.emb_g(g).transpose(1,2)
329
+
330
+ z_ptemp, m_p, logs_p, _ = self.enc_p_(c, c_lengths, f0=f0_to_coarse(f0))
331
+ z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g)
332
+
333
+ z_p = self.flow(z, spec_mask, g=g)
334
+ z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(z, f0, spec_lengths, self.segment_size)
335
+
336
+ # o = self.dec(z_slice, g=g)
337
+ o = self.dec(z_slice, g=g, f0=pitch_slice)
338
+
339
+ return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
340
+
341
+ def infer(self, c, f0, g=None, mel=None, c_lengths=None):
342
+ if c_lengths == None:
343
+ c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
344
+ g = self.emb_g(g).transpose(1,2)
345
+
346
+ z_p, m_p, logs_p, c_mask = self.enc_p_(c, c_lengths, f0=f0_to_coarse(f0))
347
+ z = self.flow(z_p, c_mask, g=g, reverse=True)
348
+
349
+ o = self.dec(z * c_mask, g=g, f0=f0)
350
+
351
+ return o
modules.py ADDED
@@ -0,0 +1,342 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import math
3
+ import numpy as np
4
+ import scipy
5
+ import torch
6
+ from torch import nn
7
+ from torch.nn import functional as F
8
+
9
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
10
+ from torch.nn.utils import weight_norm, remove_weight_norm
11
+
12
+ import commons
13
+ from commons import init_weights, get_padding
14
+
15
+
16
+ LRELU_SLOPE = 0.1
17
+
18
+
19
+ class LayerNorm(nn.Module):
20
+ def __init__(self, channels, eps=1e-5):
21
+ super().__init__()
22
+ self.channels = channels
23
+ self.eps = eps
24
+
25
+ self.gamma = nn.Parameter(torch.ones(channels))
26
+ self.beta = nn.Parameter(torch.zeros(channels))
27
+
28
+ def forward(self, x):
29
+ x = x.transpose(1, -1)
30
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
31
+ return x.transpose(1, -1)
32
+
33
+
34
+ class ConvReluNorm(nn.Module):
35
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
36
+ super().__init__()
37
+ self.in_channels = in_channels
38
+ self.hidden_channels = hidden_channels
39
+ self.out_channels = out_channels
40
+ self.kernel_size = kernel_size
41
+ self.n_layers = n_layers
42
+ self.p_dropout = p_dropout
43
+ assert n_layers > 1, "Number of layers should be larger than 0."
44
+
45
+ self.conv_layers = nn.ModuleList()
46
+ self.norm_layers = nn.ModuleList()
47
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
48
+ self.norm_layers.append(LayerNorm(hidden_channels))
49
+ self.relu_drop = nn.Sequential(
50
+ nn.ReLU(),
51
+ nn.Dropout(p_dropout))
52
+ for _ in range(n_layers-1):
53
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
54
+ self.norm_layers.append(LayerNorm(hidden_channels))
55
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
56
+ self.proj.weight.data.zero_()
57
+ self.proj.bias.data.zero_()
58
+
59
+ def forward(self, x, x_mask):
60
+ x_org = x
61
+ for i in range(self.n_layers):
62
+ x = self.conv_layers[i](x * x_mask)
63
+ x = self.norm_layers[i](x)
64
+ x = self.relu_drop(x)
65
+ x = x_org + self.proj(x)
66
+ return x * x_mask
67
+
68
+
69
+ class DDSConv(nn.Module):
70
+ """
71
+ Dialted and Depth-Separable Convolution
72
+ """
73
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
74
+ super().__init__()
75
+ self.channels = channels
76
+ self.kernel_size = kernel_size
77
+ self.n_layers = n_layers
78
+ self.p_dropout = p_dropout
79
+
80
+ self.drop = nn.Dropout(p_dropout)
81
+ self.convs_sep = nn.ModuleList()
82
+ self.convs_1x1 = nn.ModuleList()
83
+ self.norms_1 = nn.ModuleList()
84
+ self.norms_2 = nn.ModuleList()
85
+ for i in range(n_layers):
86
+ dilation = kernel_size ** i
87
+ padding = (kernel_size * dilation - dilation) // 2
88
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
89
+ groups=channels, dilation=dilation, padding=padding
90
+ ))
91
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
92
+ self.norms_1.append(LayerNorm(channels))
93
+ self.norms_2.append(LayerNorm(channels))
94
+
95
+ def forward(self, x, x_mask, g=None):
96
+ if g is not None:
97
+ x = x + g
98
+ for i in range(self.n_layers):
99
+ y = self.convs_sep[i](x * x_mask)
100
+ y = self.norms_1[i](y)
101
+ y = F.gelu(y)
102
+ y = self.convs_1x1[i](y)
103
+ y = self.norms_2[i](y)
104
+ y = F.gelu(y)
105
+ y = self.drop(y)
106
+ x = x + y
107
+ return x * x_mask
108
+
109
+
110
+ class WN(torch.nn.Module):
111
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
112
+ super(WN, self).__init__()
113
+ assert(kernel_size % 2 == 1)
114
+ self.hidden_channels =hidden_channels
115
+ self.kernel_size = kernel_size,
116
+ self.dilation_rate = dilation_rate
117
+ self.n_layers = n_layers
118
+ self.gin_channels = gin_channels
119
+ self.p_dropout = p_dropout
120
+
121
+ self.in_layers = torch.nn.ModuleList()
122
+ self.res_skip_layers = torch.nn.ModuleList()
123
+ self.drop = nn.Dropout(p_dropout)
124
+
125
+ if gin_channels != 0:
126
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
127
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
128
+
129
+ for i in range(n_layers):
130
+ dilation = dilation_rate ** i
131
+ padding = int((kernel_size * dilation - dilation) / 2)
132
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
133
+ dilation=dilation, padding=padding)
134
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
135
+ self.in_layers.append(in_layer)
136
+
137
+ # last one is not necessary
138
+ if i < n_layers - 1:
139
+ res_skip_channels = 2 * hidden_channels
140
+ else:
141
+ res_skip_channels = hidden_channels
142
+
143
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
144
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
145
+ self.res_skip_layers.append(res_skip_layer)
146
+
147
+ def forward(self, x, x_mask, g=None, **kwargs):
148
+ output = torch.zeros_like(x)
149
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
150
+
151
+ if g is not None:
152
+ g = self.cond_layer(g)
153
+
154
+ for i in range(self.n_layers):
155
+ x_in = self.in_layers[i](x)
156
+ if g is not None:
157
+ cond_offset = i * 2 * self.hidden_channels
158
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
159
+ else:
160
+ g_l = torch.zeros_like(x_in)
161
+
162
+ acts = commons.fused_add_tanh_sigmoid_multiply(
163
+ x_in,
164
+ g_l,
165
+ n_channels_tensor)
166
+ acts = self.drop(acts)
167
+
168
+ res_skip_acts = self.res_skip_layers[i](acts)
169
+ if i < self.n_layers - 1:
170
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
171
+ x = (x + res_acts) * x_mask
172
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
173
+ else:
174
+ output = output + res_skip_acts
175
+ return output * x_mask
176
+
177
+ def remove_weight_norm(self):
178
+ if self.gin_channels != 0:
179
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
180
+ for l in self.in_layers:
181
+ torch.nn.utils.remove_weight_norm(l)
182
+ for l in self.res_skip_layers:
183
+ torch.nn.utils.remove_weight_norm(l)
184
+
185
+
186
+ class ResBlock1(torch.nn.Module):
187
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
188
+ super(ResBlock1, self).__init__()
189
+ self.convs1 = nn.ModuleList([
190
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
191
+ padding=get_padding(kernel_size, dilation[0]))),
192
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
193
+ padding=get_padding(kernel_size, dilation[1]))),
194
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
195
+ padding=get_padding(kernel_size, dilation[2])))
196
+ ])
197
+ self.convs1.apply(init_weights)
198
+
199
+ self.convs2 = nn.ModuleList([
200
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
201
+ padding=get_padding(kernel_size, 1))),
202
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
203
+ padding=get_padding(kernel_size, 1))),
204
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
205
+ padding=get_padding(kernel_size, 1)))
206
+ ])
207
+ self.convs2.apply(init_weights)
208
+
209
+ def forward(self, x, x_mask=None):
210
+ for c1, c2 in zip(self.convs1, self.convs2):
211
+ xt = F.leaky_relu(x, LRELU_SLOPE)
212
+ if x_mask is not None:
213
+ xt = xt * x_mask
214
+ xt = c1(xt)
215
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
216
+ if x_mask is not None:
217
+ xt = xt * x_mask
218
+ xt = c2(xt)
219
+ x = xt + x
220
+ if x_mask is not None:
221
+ x = x * x_mask
222
+ return x
223
+
224
+ def remove_weight_norm(self):
225
+ for l in self.convs1:
226
+ remove_weight_norm(l)
227
+ for l in self.convs2:
228
+ remove_weight_norm(l)
229
+
230
+
231
+ class ResBlock2(torch.nn.Module):
232
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
233
+ super(ResBlock2, self).__init__()
234
+ self.convs = nn.ModuleList([
235
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
236
+ padding=get_padding(kernel_size, dilation[0]))),
237
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
238
+ padding=get_padding(kernel_size, dilation[1])))
239
+ ])
240
+ self.convs.apply(init_weights)
241
+
242
+ def forward(self, x, x_mask=None):
243
+ for c in self.convs:
244
+ xt = F.leaky_relu(x, LRELU_SLOPE)
245
+ if x_mask is not None:
246
+ xt = xt * x_mask
247
+ xt = c(xt)
248
+ x = xt + x
249
+ if x_mask is not None:
250
+ x = x * x_mask
251
+ return x
252
+
253
+ def remove_weight_norm(self):
254
+ for l in self.convs:
255
+ remove_weight_norm(l)
256
+
257
+
258
+ class Log(nn.Module):
259
+ def forward(self, x, x_mask, reverse=False, **kwargs):
260
+ if not reverse:
261
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
262
+ logdet = torch.sum(-y, [1, 2])
263
+ return y, logdet
264
+ else:
265
+ x = torch.exp(x) * x_mask
266
+ return x
267
+
268
+
269
+ class Flip(nn.Module):
270
+ def forward(self, x, *args, reverse=False, **kwargs):
271
+ x = torch.flip(x, [1])
272
+ if not reverse:
273
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
274
+ return x, logdet
275
+ else:
276
+ return x
277
+
278
+
279
+ class ElementwiseAffine(nn.Module):
280
+ def __init__(self, channels):
281
+ super().__init__()
282
+ self.channels = channels
283
+ self.m = nn.Parameter(torch.zeros(channels,1))
284
+ self.logs = nn.Parameter(torch.zeros(channels,1))
285
+
286
+ def forward(self, x, x_mask, reverse=False, **kwargs):
287
+ if not reverse:
288
+ y = self.m + torch.exp(self.logs) * x
289
+ y = y * x_mask
290
+ logdet = torch.sum(self.logs * x_mask, [1,2])
291
+ return y, logdet
292
+ else:
293
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
294
+ return x
295
+
296
+
297
+ class ResidualCouplingLayer(nn.Module):
298
+ def __init__(self,
299
+ channels,
300
+ hidden_channels,
301
+ kernel_size,
302
+ dilation_rate,
303
+ n_layers,
304
+ p_dropout=0,
305
+ gin_channels=0,
306
+ mean_only=False):
307
+ assert channels % 2 == 0, "channels should be divisible by 2"
308
+ super().__init__()
309
+ self.channels = channels
310
+ self.hidden_channels = hidden_channels
311
+ self.kernel_size = kernel_size
312
+ self.dilation_rate = dilation_rate
313
+ self.n_layers = n_layers
314
+ self.half_channels = channels // 2
315
+ self.mean_only = mean_only
316
+
317
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
318
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
319
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
320
+ self.post.weight.data.zero_()
321
+ self.post.bias.data.zero_()
322
+
323
+ def forward(self, x, x_mask, g=None, reverse=False):
324
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
325
+ h = self.pre(x0) * x_mask
326
+ h = self.enc(h, x_mask, g=g)
327
+ stats = self.post(h) * x_mask
328
+ if not self.mean_only:
329
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
330
+ else:
331
+ m = stats
332
+ logs = torch.zeros_like(m)
333
+
334
+ if not reverse:
335
+ x1 = m + x1 * torch.exp(logs) * x_mask
336
+ x = torch.cat([x0, x1], 1)
337
+ logdet = torch.sum(logs, [1,2])
338
+ return x, logdet
339
+ else:
340
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
341
+ x = torch.cat([x0, x1], 1)
342
+ return x
preprocess_flist_config.py ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ from tqdm import tqdm
4
+ from random import shuffle
5
+ import json
6
+ config_template = {
7
+ "train": {
8
+ "log_interval": 200,
9
+ "eval_interval": 1000,
10
+ "seed": 1234,
11
+ "epochs": 10000,
12
+ "learning_rate": 2e-4,
13
+ "betas": [0.8, 0.99],
14
+ "eps": 1e-9,
15
+ "batch_size": 12,
16
+ "fp16_run": False,
17
+ "lr_decay": 0.999875,
18
+ "segment_size": 17920,
19
+ "init_lr_ratio": 1,
20
+ "warmup_epochs": 0,
21
+ "c_mel": 45,
22
+ "c_kl": 1.0,
23
+ "use_sr": True,
24
+ "max_speclen": 384,
25
+ "port": "8001"
26
+ },
27
+ "data": {
28
+ "training_files":"filelists/train.txt",
29
+ "validation_files":"filelists/val.txt",
30
+ "max_wav_value": 32768.0,
31
+ "sampling_rate": 32000,
32
+ "filter_length": 1280,
33
+ "hop_length": 320,
34
+ "win_length": 1280,
35
+ "n_mel_channels": 80,
36
+ "mel_fmin": 0.0,
37
+ "mel_fmax": None
38
+ },
39
+ "model": {
40
+ "inter_channels": 192,
41
+ "hidden_channels": 192,
42
+ "filter_channels": 768,
43
+ "n_heads": 2,
44
+ "n_layers": 6,
45
+ "kernel_size": 3,
46
+ "p_dropout": 0.1,
47
+ "resblock": "1",
48
+ "resblock_kernel_sizes": [3,7,11],
49
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
50
+ "upsample_rates": [10,8,2,2],
51
+ "upsample_initial_channel": 512,
52
+ "upsample_kernel_sizes": [16,16,4,4],
53
+ "n_layers_q": 3,
54
+ "use_spectral_norm": False,
55
+ "gin_channels": 256,
56
+ "ssl_dim": 256,
57
+ "n_speakers": 0,
58
+ },
59
+ "spk":{
60
+ "nen": 0,
61
+ "paimon": 1,
62
+ "yunhao": 2
63
+ }
64
+ }
65
+
66
+
67
+ if __name__ == "__main__":
68
+ parser = argparse.ArgumentParser()
69
+ parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list")
70
+ parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list")
71
+ parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list")
72
+ parser.add_argument("--source_dir", type=str, default="./dataset/32k", help="path to source dir")
73
+ args = parser.parse_args()
74
+
75
+ train = []
76
+ val = []
77
+ test = []
78
+ idx = 0
79
+ spk_dict = {}
80
+ spk_id = 0
81
+ for speaker in tqdm(os.listdir(args.source_dir)):
82
+ spk_dict[speaker] = spk_id
83
+ spk_id += 1
84
+ wavs = [os.path.join(args.source_dir, speaker, i)for i in os.listdir(os.path.join(args.source_dir, speaker))]
85
+ wavs = [i for i in wavs if i.endswith("wav")]
86
+ shuffle(wavs)
87
+ train += wavs[2:-10]
88
+ val += wavs[:2]
89
+ test += wavs[-10:]
90
+ n_speakers = len(spk_dict.keys())*2
91
+ shuffle(train)
92
+ shuffle(val)
93
+ shuffle(test)
94
+
95
+ print("Writing", args.train_list)
96
+ with open(args.train_list, "w") as f:
97
+ for fname in tqdm(train):
98
+ wavpath = fname
99
+ f.write(wavpath + "\n")
100
+
101
+ print("Writing", args.val_list)
102
+ with open(args.val_list, "w") as f:
103
+ for fname in tqdm(val):
104
+ wavpath = fname
105
+ f.write(wavpath + "\n")
106
+
107
+ print("Writing", args.test_list)
108
+ with open(args.test_list, "w") as f:
109
+ for fname in tqdm(test):
110
+ wavpath = fname
111
+ f.write(wavpath + "\n")
112
+
113
+ config_template["model"]["n_speakers"] = n_speakers
114
+ config_template["spk"] = spk_dict
115
+ print("Writing configs/config.json")
116
+ with open("configs/config.json", "w") as f:
117
+ json.dump(config_template, f, indent=2)
preprocess_hubert_f0.py ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+
4
+ import torch
5
+ import json
6
+ from glob import glob
7
+
8
+ from pyworld import pyworld
9
+ from tqdm import tqdm
10
+ from scipy.io import wavfile
11
+
12
+ import utils
13
+ from mel_processing import mel_spectrogram_torch
14
+ #import h5py
15
+ import logging
16
+ logging.getLogger('numba').setLevel(logging.WARNING)
17
+
18
+ import parselmouth
19
+ import librosa
20
+ import numpy as np
21
+
22
+
23
+ def get_f0(path,p_len=None, f0_up_key=0):
24
+ x, _ = librosa.load(path, 32000)
25
+ if p_len is None:
26
+ p_len = x.shape[0]//320
27
+ else:
28
+ assert abs(p_len-x.shape[0]//320) < 3, (path, p_len, x.shape)
29
+ time_step = 320 / 32000 * 1000
30
+ f0_min = 50
31
+ f0_max = 1100
32
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
33
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
34
+
35
+ f0 = parselmouth.Sound(x, 32000).to_pitch_ac(
36
+ time_step=time_step / 1000, voicing_threshold=0.6,
37
+ pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
38
+
39
+ pad_size=(p_len - len(f0) + 1) // 2
40
+ if(pad_size>0 or p_len - len(f0) - pad_size>0):
41
+ f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
42
+
43
+ f0bak = f0.copy()
44
+ f0 *= pow(2, f0_up_key / 12)
45
+ f0_mel = 1127 * np.log(1 + f0 / 700)
46
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
47
+ f0_mel[f0_mel <= 1] = 1
48
+ f0_mel[f0_mel > 255] = 255
49
+ f0_coarse = np.rint(f0_mel).astype(np.int)
50
+ return f0_coarse, f0bak
51
+
52
+ def resize2d(x, target_len):
53
+ source = np.array(x)
54
+ source[source<0.001] = np.nan
55
+ target = np.interp(np.arange(0, len(source)*target_len, len(source))/ target_len, np.arange(0, len(source)), source)
56
+ res = np.nan_to_num(target)
57
+ return res
58
+
59
+ def compute_f0(path, c_len):
60
+ x, sr = librosa.load(path, sr=32000)
61
+ f0, t = pyworld.dio(
62
+ x.astype(np.double),
63
+ fs=sr,
64
+ f0_ceil=800,
65
+ frame_period=1000 * 320 / sr,
66
+ )
67
+ f0 = pyworld.stonemask(x.astype(np.double), f0, t, 32000)
68
+ for index, pitch in enumerate(f0):
69
+ f0[index] = round(pitch, 1)
70
+ assert abs(c_len - x.shape[0]//320) < 3, (c_len, f0.shape)
71
+
72
+ return None, resize2d(f0, c_len)
73
+
74
+
75
+ def process(filename):
76
+ print(filename)
77
+ save_name = filename+".soft.pt"
78
+ if not os.path.exists(save_name):
79
+ devive = torch.device("cuda" if torch.cuda.is_available() else "cpu")
80
+ wav, _ = librosa.load(filename, sr=16000)
81
+ wav = torch.from_numpy(wav).unsqueeze(0).to(devive)
82
+ c = utils.get_hubert_content(hmodel, wav)
83
+ torch.save(c.cpu(), save_name)
84
+ else:
85
+ c = torch.load(save_name)
86
+ f0path = filename+".f0.npy"
87
+ if not os.path.exists(f0path):
88
+ cf0, f0 = compute_f0(filename, c.shape[-1] * 2)
89
+ np.save(f0path, f0)
90
+
91
+
92
+
93
+ if __name__ == "__main__":
94
+ parser = argparse.ArgumentParser()
95
+ parser.add_argument("--in_dir", type=str, default="dataset/32k", help="path to input dir")
96
+ args = parser.parse_args()
97
+
98
+ print("Loading hubert for content...")
99
+ hmodel = utils.get_hubert_model(0 if torch.cuda.is_available() else None)
100
+ print("Loaded hubert.")
101
+
102
+ filenames = glob(f'{args.in_dir}/*/*.wav', recursive=True)#[:10]
103
+
104
+ for filename in tqdm(filenames):
105
+ process(filename)
106
+
requirements.txt ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ numpy
2
+ playsound
3
+ pydub
4
+ pyworld
5
+ requests
6
+ scipy
7
+ sounddevice
8
+ SoundFile
9
+ starlette
10
+ torch
11
+ torchaudio
12
+ tqdm
13
+ scikit-maad
14
+ praat-parselmouth
15
+ librosa
16
+ torchvision
resample.py ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+ import librosa
4
+ import numpy as np
5
+ from multiprocessing import Pool, cpu_count
6
+ from scipy.io import wavfile
7
+ from tqdm import tqdm
8
+
9
+
10
+ def process(item):
11
+ spkdir, wav_name, args = item
12
+ # speaker 's5', 'p280', 'p315' are excluded,
13
+ speaker = spkdir.split(os.sep)[-1]
14
+ wav_path = os.path.join(args.in_dir, speaker, wav_name)
15
+ if os.path.exists(wav_path) and '.wav' in wav_path:
16
+ os.makedirs(os.path.join(args.out_dir2, speaker), exist_ok=True)
17
+ wav, sr = librosa.load(wav_path, None)
18
+ wav, _ = librosa.effects.trim(wav, top_db=20)
19
+ peak = np.abs(wav).max()
20
+ if peak > 1.0:
21
+ wav = 0.98 * wav / peak
22
+ wav2 = librosa.resample(wav, orig_sr=sr, target_sr=args.sr2)
23
+ save_name = wav_name
24
+ save_path2 = os.path.join(args.out_dir2, speaker, save_name)
25
+ wavfile.write(
26
+ save_path2,
27
+ args.sr2,
28
+ (wav2 * np.iinfo(np.int16).max).astype(np.int16)
29
+ )
30
+
31
+
32
+
33
+ if __name__ == "__main__":
34
+ parser = argparse.ArgumentParser()
35
+ parser.add_argument("--sr2", type=int, default=32000, help="sampling rate")
36
+ parser.add_argument("--in_dir", type=str, default="./dataset_raw", help="path to source dir")
37
+ parser.add_argument("--out_dir2", type=str, default="./dataset/32k", help="path to target dir")
38
+ args = parser.parse_args()
39
+ processs = cpu_count()-2 if cpu_count() >4 else 1
40
+ pool = Pool(processes=processs)
41
+
42
+ for speaker in os.listdir(args.in_dir):
43
+ spk_dir = os.path.join(args.in_dir, speaker)
44
+ if os.path.isdir(spk_dir):
45
+ print(spk_dir)
46
+ for _ in tqdm(pool.imap_unordered(process, [(spk_dir, i, args) for i in os.listdir(spk_dir) if i.endswith("wav")])):
47
+ pass
spec_gen.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from data_utils import TextAudioSpeakerLoader, EvalDataLoader
2
+ import json
3
+ from tqdm import tqdm
4
+
5
+ from utils import HParams
6
+
7
+ config_path = 'configs/config.json'
8
+ with open(config_path, "r") as f:
9
+ data = f.read()
10
+ config = json.loads(data)
11
+ hps = HParams(**config)
12
+
13
+ train_dataset = TextAudioSpeakerLoader("filelists/train.txt", hps)
14
+ test_dataset = TextAudioSpeakerLoader("filelists/test.txt", hps)
15
+ eval_dataset = TextAudioSpeakerLoader("filelists/val.txt", hps)
16
+
17
+ for _ in tqdm(train_dataset):
18
+ pass
19
+ for _ in tqdm(eval_dataset):
20
+ pass
21
+ for _ in tqdm(test_dataset):
22
+ pass
terms.md ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 在使用此模型前请阅读以下协议,本协议修改自MasterSatori
2
+
3
+ AI粘连科技模型使用协议
4
+
5
+ 【前言】AI粘连科技模型所有者及训练者海龙王kokopelli@bilibili(以下也称“我”)希望通过《AI粘连科技模模型使用协议》(以下简称“本协议”)向您说明您在使用AI粘连科技模模型时应当履行的责任及使用范围。
6
+
7
+ 【特别提示】在使用AI粘连科技模模型前,请您务必仔细阅读并透彻理解本协议,在确认充分理解并同意后再开始使用。
8
+
9
+ ​ 本协议将帮助您了解以下内容:
10
+
11
+ ​ 一、免责声明
12
+
13
+ ​ 二、您在非个人使用场合时使用AI粘连科技模型应当做的事
14
+
15
+ ​ 三、AI粘连科技的使用范围
16
+
17
+ ​ 四、如何联系我
18
+
19
+ ​ (一) 免责声明:
20
+
21
+ ​ 您因使用AI粘连科技对其它任何实体(个人/企业)所造成的任何损失由您自身承担,您因使用AI粘连科技模型所产生的一切法律风险及法律纠纷由您自身承担。
22
+
23
+ ​ (二) 您在非个人使用场合时使用AI粘连科技模型应当做的事:
24
+
25
+ ​ 1、注明soVITS项目作者:Rcell
26
+
27
+ ​ 2、注明我(可选):海龙王kokopelli@bilibili
28
+
29
+ ​ (三) AI粘连科技模型的使用范围:
30
+
31
+ ​ 1、您可以使用的范围:
32
+
33
+ ​ (1) 个人使用
34
+
35
+ ​ (2) 将产生的音频用于投稿(投稿内容不得包含“您不可使用的范围”中的内容)
36
+
37
+ ​ (3) 符合投稿平台和当地法律的二创内容
38
+
39
+ ​ (4) 使用本软件必须注明作品使用了AI
40
+
41
+ ​ 2、您不可使用的范围:
42
+
43
+ ​ (1) 商业使用
44
+
45
+ ​ (2) 假冒本人
46
+
47
+ ​ (3) 当作变声器等使用
48
+
49
+ ​ (4) 将AI粘连科技模型再次上传
50
+
51
+ ​ (5) 低创内容(合成的音频中有过多的爆音或电音属于“低创内容”)
52
+
53
+ ​ (6) 敏感内容(包括但不限于:政治、低俗、色情、暴力等)
54
+
55
+ ​ 3、补充内容:
56
+
57
+ ​ 在其他未被提及的场合使用AI草莓猫taffy模型及其所产生的数据时您应当征求我的意见.海龙王kokopelli@bilibili。
train.py ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ logging.getLogger('matplotlib').setLevel(logging.WARNING)
3
+ import os
4
+ import json
5
+ import argparse
6
+ import itertools
7
+ import math
8
+ import torch
9
+ from torch import nn, optim
10
+ from torch.nn import functional as F
11
+ from torch.utils.data import DataLoader
12
+ from torch.utils.tensorboard import SummaryWriter
13
+ import torch.multiprocessing as mp
14
+ import torch.distributed as dist
15
+ from torch.nn.parallel import DistributedDataParallel as DDP
16
+ from torch.cuda.amp import autocast, GradScaler
17
+
18
+ import commons
19
+ import utils
20
+ from data_utils import TextAudioSpeakerLoader, EvalDataLoader
21
+ from models import (
22
+ SynthesizerTrn,
23
+ MultiPeriodDiscriminator,
24
+ )
25
+ from losses import (
26
+ kl_loss,
27
+ generator_loss, discriminator_loss, feature_loss
28
+ )
29
+
30
+ from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
31
+
32
+ torch.backends.cudnn.benchmark = True
33
+ global_step = 0
34
+
35
+
36
+ # os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'
37
+
38
+
39
+ def main():
40
+ """Assume Single Node Multi GPUs Training Only"""
41
+ assert torch.cuda.is_available(), "CPU training is not allowed."
42
+ hps = utils.get_hparams()
43
+
44
+ n_gpus = torch.cuda.device_count()
45
+ os.environ['MASTER_ADDR'] = 'localhost'
46
+ os.environ['MASTER_PORT'] = hps.train.port
47
+
48
+ mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
49
+
50
+
51
+ def run(rank, n_gpus, hps):
52
+ global global_step
53
+ if rank == 0:
54
+ logger = utils.get_logger(hps.model_dir)
55
+ logger.info(hps)
56
+ utils.check_git_hash(hps.model_dir)
57
+ writer = SummaryWriter(log_dir=hps.model_dir)
58
+ writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
59
+
60
+ dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
61
+ torch.manual_seed(hps.train.seed)
62
+ torch.cuda.set_device(rank)
63
+
64
+ train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps)
65
+ train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
66
+ batch_size=hps.train.batch_size)
67
+ if rank == 0:
68
+ eval_dataset = EvalDataLoader(hps.data.validation_files, hps)
69
+ eval_loader = DataLoader(eval_dataset, num_workers=1, shuffle=False,
70
+ batch_size=1, pin_memory=False,
71
+ drop_last=False)
72
+
73
+ net_g = SynthesizerTrn(
74
+ hps.data.filter_length // 2 + 1,
75
+ hps.train.segment_size // hps.data.hop_length,
76
+ **hps.model).cuda(rank)
77
+ net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
78
+ optim_g = torch.optim.AdamW(
79
+ net_g.parameters(),
80
+ hps.train.learning_rate,
81
+ betas=hps.train.betas,
82
+ eps=hps.train.eps)
83
+ optim_d = torch.optim.AdamW(
84
+ net_d.parameters(),
85
+ hps.train.learning_rate,
86
+ betas=hps.train.betas,
87
+ eps=hps.train.eps)
88
+ net_g = DDP(net_g, device_ids=[rank]) # , find_unused_parameters=True)
89
+ net_d = DDP(net_d, device_ids=[rank])
90
+
91
+ try:
92
+ _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
93
+ optim_g)
94
+ _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
95
+ optim_d)
96
+ global_step = (epoch_str - 1) * len(train_loader)
97
+ except:
98
+ epoch_str = 1
99
+ global_step = 0
100
+
101
+ scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
102
+ scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
103
+
104
+ scaler = GradScaler(enabled=hps.train.fp16_run)
105
+
106
+ for epoch in range(epoch_str, hps.train.epochs + 1):
107
+ if rank == 0:
108
+ train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
109
+ [train_loader, eval_loader], logger, [writer, writer_eval])
110
+ else:
111
+ train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
112
+ [train_loader, None], None, None)
113
+ scheduler_g.step()
114
+ scheduler_d.step()
115
+
116
+
117
+ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
118
+ net_g, net_d = nets
119
+ optim_g, optim_d = optims
120
+ scheduler_g, scheduler_d = schedulers
121
+ train_loader, eval_loader = loaders
122
+ if writers is not None:
123
+ writer, writer_eval = writers
124
+
125
+ # train_loader.batch_sampler.set_epoch(epoch)
126
+ global global_step
127
+
128
+ net_g.train()
129
+ net_d.train()
130
+ for batch_idx, items in enumerate(train_loader):
131
+ c, f0, spec, y, spk = items
132
+ g = spk.cuda(rank, non_blocking=True)
133
+ spec, y = spec.cuda(rank, non_blocking=True), y.cuda(rank, non_blocking=True)
134
+ c = c.cuda(rank, non_blocking=True)
135
+ f0 = f0.cuda(rank, non_blocking=True)
136
+ mel = spec_to_mel_torch(
137
+ spec,
138
+ hps.data.filter_length,
139
+ hps.data.n_mel_channels,
140
+ hps.data.sampling_rate,
141
+ hps.data.mel_fmin,
142
+ hps.data.mel_fmax)
143
+
144
+ with autocast(enabled=hps.train.fp16_run):
145
+ y_hat, ids_slice, z_mask, \
146
+ (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(c, f0, spec, g=g, mel=mel)
147
+
148
+ y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
149
+ y_hat_mel = mel_spectrogram_torch(
150
+ y_hat.squeeze(1),
151
+ hps.data.filter_length,
152
+ hps.data.n_mel_channels,
153
+ hps.data.sampling_rate,
154
+ hps.data.hop_length,
155
+ hps.data.win_length,
156
+ hps.data.mel_fmin,
157
+ hps.data.mel_fmax
158
+ )
159
+ y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice
160
+
161
+ # Discriminator
162
+ y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
163
+
164
+ with autocast(enabled=False):
165
+ loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
166
+ loss_disc_all = loss_disc
167
+
168
+ optim_d.zero_grad()
169
+ scaler.scale(loss_disc_all).backward()
170
+ scaler.unscale_(optim_d)
171
+ grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
172
+ scaler.step(optim_d)
173
+
174
+ with autocast(enabled=hps.train.fp16_run):
175
+ # Generator
176
+ y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
177
+ with autocast(enabled=False):
178
+ loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
179
+ loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
180
+ loss_fm = feature_loss(fmap_r, fmap_g)
181
+ loss_gen, losses_gen = generator_loss(y_d_hat_g)
182
+ loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
183
+ optim_g.zero_grad()
184
+ scaler.scale(loss_gen_all).backward()
185
+ scaler.unscale_(optim_g)
186
+ grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
187
+ scaler.step(optim_g)
188
+ scaler.update()
189
+
190
+ if rank == 0:
191
+ if global_step % hps.train.log_interval == 0:
192
+ lr = optim_g.param_groups[0]['lr']
193
+ losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_kl]
194
+ logger.info('Train Epoch: {} [{:.0f}%]'.format(
195
+ epoch,
196
+ 100. * batch_idx / len(train_loader)))
197
+ logger.info([x.item() for x in losses] + [global_step, lr])
198
+
199
+ scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr,
200
+ "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
201
+ scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl})
202
+
203
+ scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
204
+ scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
205
+ scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
206
+ image_dict = {
207
+ "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
208
+ "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()),
209
+ "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
210
+ }
211
+
212
+ utils.summarize(
213
+ writer=writer,
214
+ global_step=global_step,
215
+ images=image_dict,
216
+ scalars=scalar_dict
217
+ )
218
+
219
+ if global_step % hps.train.eval_interval == 0:
220
+ evaluate(hps, net_g, eval_loader, writer_eval)
221
+ utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch,
222
+ os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
223
+ utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch,
224
+ os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
225
+ global_step += 1
226
+
227
+ if rank == 0:
228
+ logger.info('====> Epoch: {}'.format(epoch))
229
+
230
+
231
+ def evaluate(hps, generator, eval_loader, writer_eval):
232
+ generator.eval()
233
+ image_dict = {}
234
+ audio_dict = {}
235
+ with torch.no_grad():
236
+ for batch_idx, items in enumerate(eval_loader):
237
+ c, f0, spec, y, spk = items
238
+ g = spk[:1].cuda(0)
239
+ spec, y = spec[:1].cuda(0), y[:1].cuda(0)
240
+ c = c[:1].cuda(0)
241
+ f0 = f0[:1].cuda(0)
242
+ mel = spec_to_mel_torch(
243
+ spec,
244
+ hps.data.filter_length,
245
+ hps.data.n_mel_channels,
246
+ hps.data.sampling_rate,
247
+ hps.data.mel_fmin,
248
+ hps.data.mel_fmax)
249
+ y_hat = generator.module.infer(c, f0, g=g, mel=mel)
250
+
251
+ y_hat_mel = mel_spectrogram_torch(
252
+ y_hat.squeeze(1).float(),
253
+ hps.data.filter_length,
254
+ hps.data.n_mel_channels,
255
+ hps.data.sampling_rate,
256
+ hps.data.hop_length,
257
+ hps.data.win_length,
258
+ hps.data.mel_fmin,
259
+ hps.data.mel_fmax
260
+ )
261
+
262
+ audio_dict.update({
263
+ f"gen/audio_{batch_idx}": y_hat[0],
264
+ f"gt/audio_{batch_idx}": y[0]
265
+ })
266
+ image_dict.update({
267
+ f"gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy()),
268
+ "gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())
269
+ })
270
+ utils.summarize(
271
+ writer=writer_eval,
272
+ global_step=global_step,
273
+ images=image_dict,
274
+ audios=audio_dict,
275
+ audio_sampling_rate=hps.data.sampling_rate
276
+ )
277
+ generator.train()
278
+
279
+
280
+ if __name__ == "__main__":
281
+ main()
utils.py ADDED
@@ -0,0 +1,338 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import glob
3
+ import sys
4
+ import argparse
5
+ import logging
6
+ import json
7
+ import subprocess
8
+
9
+ import librosa
10
+ import numpy as np
11
+ import torchaudio
12
+ from scipy.io.wavfile import read
13
+ import torch
14
+ import torchvision
15
+ from torch.nn import functional as F
16
+ from commons import sequence_mask
17
+ from hubert import hubert_model
18
+ MATPLOTLIB_FLAG = False
19
+
20
+ logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
21
+ logger = logging
22
+
23
+ f0_bin = 256
24
+ f0_max = 1100.0
25
+ f0_min = 50.0
26
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
27
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
28
+
29
+ def f0_to_coarse(f0):
30
+ is_torch = isinstance(f0, torch.Tensor)
31
+ f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700)
32
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (f0_mel_max - f0_mel_min) + 1
33
+
34
+ f0_mel[f0_mel <= 1] = 1
35
+ f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1
36
+ f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int)
37
+ assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (f0_coarse.max(), f0_coarse.min())
38
+ return f0_coarse
39
+
40
+
41
+ def get_hubert_model(rank=None):
42
+
43
+ hubert_soft = hubert_model.hubert_soft("hubert/hubert-soft-0d54a1f4.pt")
44
+ if rank is not None:
45
+ hubert_soft = hubert_soft.cuda(rank)
46
+ return hubert_soft
47
+
48
+ def get_hubert_content(hmodel, y=None, path=None):
49
+ if path is not None:
50
+ source, sr = torchaudio.load(path)
51
+ source = torchaudio.functional.resample(source, sr, 16000)
52
+ if len(source.shape) == 2 and source.shape[1] >= 2:
53
+ source = torch.mean(source, dim=0).unsqueeze(0)
54
+ else:
55
+ source = y
56
+ source = source.unsqueeze(0)
57
+ with torch.inference_mode():
58
+ units = hmodel.units(source)
59
+ return units.transpose(1,2)
60
+
61
+
62
+ def get_content(cmodel, y):
63
+ with torch.no_grad():
64
+ c = cmodel.extract_features(y.squeeze(1))[0]
65
+ c = c.transpose(1, 2)
66
+ return c
67
+
68
+
69
+
70
+ def transform(mel, height): # 68-92
71
+ #r = np.random.random()
72
+ #rate = r * 0.3 + 0.85 # 0.85-1.15
73
+ #height = int(mel.size(-2) * rate)
74
+ tgt = torchvision.transforms.functional.resize(mel, (height, mel.size(-1)))
75
+ if height >= mel.size(-2):
76
+ return tgt[:, :mel.size(-2), :]
77
+ else:
78
+ silence = tgt[:,-1:,:].repeat(1,mel.size(-2)-height,1)
79
+ silence += torch.randn_like(silence) / 10
80
+ return torch.cat((tgt, silence), 1)
81
+
82
+
83
+ def stretch(mel, width): # 0.5-2
84
+ return torchvision.transforms.functional.resize(mel, (mel.size(-2), width))
85
+
86
+
87
+ def load_checkpoint(checkpoint_path, model, optimizer=None):
88
+ assert os.path.isfile(checkpoint_path)
89
+ checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
90
+ iteration = checkpoint_dict['iteration']
91
+ learning_rate = checkpoint_dict['learning_rate']
92
+ if iteration is None:
93
+ iteration = 1
94
+ if learning_rate is None:
95
+ learning_rate = 0.0002
96
+ if optimizer is not None and checkpoint_dict['optimizer'] is not None:
97
+ optimizer.load_state_dict(checkpoint_dict['optimizer'])
98
+ saved_state_dict = checkpoint_dict['model']
99
+ if hasattr(model, 'module'):
100
+ state_dict = model.module.state_dict()
101
+ else:
102
+ state_dict = model.state_dict()
103
+ new_state_dict= {}
104
+ for k, v in state_dict.items():
105
+ try:
106
+ new_state_dict[k] = saved_state_dict[k]
107
+ except:
108
+ logger.info("%s is not in the checkpoint" % k)
109
+ new_state_dict[k] = v
110
+ if hasattr(model, 'module'):
111
+ model.module.load_state_dict(new_state_dict)
112
+ else:
113
+ model.load_state_dict(new_state_dict)
114
+ logger.info("Loaded checkpoint '{}' (iteration {})" .format(
115
+ checkpoint_path, iteration))
116
+ return model, optimizer, learning_rate, iteration
117
+
118
+
119
+ def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
120
+ # ckptname = checkpoint_path.split(os.sep)[-1]
121
+ # newest_step = int(ckptname.split(".")[0].split("_")[1])
122
+ # val_steps = 2000
123
+ # last_ckptname = checkpoint_path.replace(str(newest_step), str(newest_step - val_steps*3))
124
+ # if newest_step >= val_steps*3:
125
+ # os.system(f"rm {last_ckptname}")
126
+ logger.info("Saving model and optimizer state at iteration {} to {}".format(
127
+ iteration, checkpoint_path))
128
+ if hasattr(model, 'module'):
129
+ state_dict = model.module.state_dict()
130
+ else:
131
+ state_dict = model.state_dict()
132
+ torch.save({'model': state_dict,
133
+ 'iteration': iteration,
134
+ 'optimizer': optimizer.state_dict(),
135
+ 'learning_rate': learning_rate}, checkpoint_path)
136
+
137
+
138
+ def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
139
+ for k, v in scalars.items():
140
+ writer.add_scalar(k, v, global_step)
141
+ for k, v in histograms.items():
142
+ writer.add_histogram(k, v, global_step)
143
+ for k, v in images.items():
144
+ writer.add_image(k, v, global_step, dataformats='HWC')
145
+ for k, v in audios.items():
146
+ writer.add_audio(k, v, global_step, audio_sampling_rate)
147
+
148
+
149
+ def latest_checkpoint_path(dir_path, regex="G_*.pth"):
150
+ f_list = glob.glob(os.path.join(dir_path, regex))
151
+ f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
152
+ x = f_list[-1]
153
+ print(x)
154
+ return x
155
+
156
+
157
+ def plot_spectrogram_to_numpy(spectrogram):
158
+ global MATPLOTLIB_FLAG
159
+ if not MATPLOTLIB_FLAG:
160
+ import matplotlib
161
+ matplotlib.use("Agg")
162
+ MATPLOTLIB_FLAG = True
163
+ mpl_logger = logging.getLogger('matplotlib')
164
+ mpl_logger.setLevel(logging.WARNING)
165
+ import matplotlib.pylab as plt
166
+ import numpy as np
167
+
168
+ fig, ax = plt.subplots(figsize=(10,2))
169
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
170
+ interpolation='none')
171
+ plt.colorbar(im, ax=ax)
172
+ plt.xlabel("Frames")
173
+ plt.ylabel("Channels")
174
+ plt.tight_layout()
175
+
176
+ fig.canvas.draw()
177
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
178
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
179
+ plt.close()
180
+ return data
181
+
182
+
183
+ def plot_alignment_to_numpy(alignment, info=None):
184
+ global MATPLOTLIB_FLAG
185
+ if not MATPLOTLIB_FLAG:
186
+ import matplotlib
187
+ matplotlib.use("Agg")
188
+ MATPLOTLIB_FLAG = True
189
+ mpl_logger = logging.getLogger('matplotlib')
190
+ mpl_logger.setLevel(logging.WARNING)
191
+ import matplotlib.pylab as plt
192
+ import numpy as np
193
+
194
+ fig, ax = plt.subplots(figsize=(6, 4))
195
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
196
+ interpolation='none')
197
+ fig.colorbar(im, ax=ax)
198
+ xlabel = 'Decoder timestep'
199
+ if info is not None:
200
+ xlabel += '\n\n' + info
201
+ plt.xlabel(xlabel)
202
+ plt.ylabel('Encoder timestep')
203
+ plt.tight_layout()
204
+
205
+ fig.canvas.draw()
206
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
207
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
208
+ plt.close()
209
+ return data
210
+
211
+
212
+ def load_wav_to_torch(full_path):
213
+ sampling_rate, data = read(full_path)
214
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
215
+
216
+
217
+ def load_filepaths_and_text(filename, split="|"):
218
+ with open(filename, encoding='utf-8') as f:
219
+ filepaths_and_text = [line.strip().split(split) for line in f]
220
+ return filepaths_and_text
221
+
222
+
223
+ def get_hparams(init=True):
224
+ parser = argparse.ArgumentParser()
225
+ parser.add_argument('-c', '--config', type=str, default="./configs/base.json",
226
+ help='JSON file for configuration')
227
+ parser.add_argument('-m', '--model', type=str, required=True,
228
+ help='Model name')
229
+
230
+ args = parser.parse_args()
231
+ model_dir = os.path.join("./logs", args.model)
232
+
233
+ if not os.path.exists(model_dir):
234
+ os.makedirs(model_dir)
235
+
236
+ config_path = args.config
237
+ config_save_path = os.path.join(model_dir, "config.json")
238
+ if init:
239
+ with open(config_path, "r") as f:
240
+ data = f.read()
241
+ with open(config_save_path, "w") as f:
242
+ f.write(data)
243
+ else:
244
+ with open(config_save_path, "r") as f:
245
+ data = f.read()
246
+ config = json.loads(data)
247
+
248
+ hparams = HParams(**config)
249
+ hparams.model_dir = model_dir
250
+ return hparams
251
+
252
+
253
+ def get_hparams_from_dir(model_dir):
254
+ config_save_path = os.path.join(model_dir, "config.json")
255
+ with open(config_save_path, "r") as f:
256
+ data = f.read()
257
+ config = json.loads(data)
258
+
259
+ hparams =HParams(**config)
260
+ hparams.model_dir = model_dir
261
+ return hparams
262
+
263
+
264
+ def get_hparams_from_file(config_path):
265
+ with open(config_path, "r") as f:
266
+ data = f.read()
267
+ config = json.loads(data)
268
+
269
+ hparams =HParams(**config)
270
+ return hparams
271
+
272
+
273
+ def check_git_hash(model_dir):
274
+ source_dir = os.path.dirname(os.path.realpath(__file__))
275
+ if not os.path.exists(os.path.join(source_dir, ".git")):
276
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
277
+ source_dir
278
+ ))
279
+ return
280
+
281
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
282
+
283
+ path = os.path.join(model_dir, "githash")
284
+ if os.path.exists(path):
285
+ saved_hash = open(path).read()
286
+ if saved_hash != cur_hash:
287
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
288
+ saved_hash[:8], cur_hash[:8]))
289
+ else:
290
+ open(path, "w").write(cur_hash)
291
+
292
+
293
+ def get_logger(model_dir, filename="train.log"):
294
+ global logger
295
+ logger = logging.getLogger(os.path.basename(model_dir))
296
+ logger.setLevel(logging.DEBUG)
297
+
298
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
299
+ if not os.path.exists(model_dir):
300
+ os.makedirs(model_dir)
301
+ h = logging.FileHandler(os.path.join(model_dir, filename))
302
+ h.setLevel(logging.DEBUG)
303
+ h.setFormatter(formatter)
304
+ logger.addHandler(h)
305
+ return logger
306
+
307
+
308
+ class HParams():
309
+ def __init__(self, **kwargs):
310
+ for k, v in kwargs.items():
311
+ if type(v) == dict:
312
+ v = HParams(**v)
313
+ self[k] = v
314
+
315
+ def keys(self):
316
+ return self.__dict__.keys()
317
+
318
+ def items(self):
319
+ return self.__dict__.items()
320
+
321
+ def values(self):
322
+ return self.__dict__.values()
323
+
324
+ def __len__(self):
325
+ return len(self.__dict__)
326
+
327
+ def __getitem__(self, key):
328
+ return getattr(self, key)
329
+
330
+ def __setitem__(self, key, value):
331
+ return setattr(self, key, value)
332
+
333
+ def __contains__(self, key):
334
+ return key in self.__dict__
335
+
336
+ def __repr__(self):
337
+ return self.__dict__.__repr__()
338
+
vdecoder/__init__.py ADDED
File without changes
vdecoder/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (129 Bytes). View file
 
vdecoder/hifigan/__pycache__/env.cpython-310.pyc ADDED
Binary file (802 Bytes). View file
 
vdecoder/hifigan/__pycache__/models.cpython-310.pyc ADDED
Binary file (14.9 kB). View file
 
vdecoder/hifigan/__pycache__/utils.cpython-310.pyc ADDED
Binary file (2.34 kB). View file
 
vdecoder/hifigan/env.py ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import shutil
3
+
4
+
5
+ class AttrDict(dict):
6
+ def __init__(self, *args, **kwargs):
7
+ super(AttrDict, self).__init__(*args, **kwargs)
8
+ self.__dict__ = self
9
+
10
+
11
+ def build_env(config, config_name, path):
12
+ t_path = os.path.join(path, config_name)
13
+ if config != t_path:
14
+ os.makedirs(path, exist_ok=True)
15
+ shutil.copyfile(config, os.path.join(path, config_name))
vdecoder/hifigan/models.py ADDED
@@ -0,0 +1,500 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import json
3
+ from .env import AttrDict
4
+ import numpy as np
5
+ import torch
6
+ import torch.nn.functional as F
7
+ import torch.nn as nn
8
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
9
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
10
+ from .utils import init_weights, get_padding
11
+
12
+ LRELU_SLOPE = 0.1
13
+
14
+
15
+ def load_model(model_path, device='cuda'):
16
+ config_file = os.path.join(os.path.split(model_path)[0], 'config.json')
17
+ with open(config_file) as f:
18
+ data = f.read()
19
+
20
+ global h
21
+ json_config = json.loads(data)
22
+ h = AttrDict(json_config)
23
+
24
+ generator = Generator(h).to(device)
25
+
26
+ cp_dict = torch.load(model_path)
27
+ generator.load_state_dict(cp_dict['generator'])
28
+ generator.eval()
29
+ generator.remove_weight_norm()
30
+ del cp_dict
31
+ return generator, h
32
+
33
+
34
+ class ResBlock1(torch.nn.Module):
35
+ def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
36
+ super(ResBlock1, self).__init__()
37
+ self.h = h
38
+ self.convs1 = nn.ModuleList([
39
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
40
+ padding=get_padding(kernel_size, dilation[0]))),
41
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
42
+ padding=get_padding(kernel_size, dilation[1]))),
43
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
44
+ padding=get_padding(kernel_size, dilation[2])))
45
+ ])
46
+ self.convs1.apply(init_weights)
47
+
48
+ self.convs2 = nn.ModuleList([
49
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
50
+ padding=get_padding(kernel_size, 1))),
51
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
52
+ padding=get_padding(kernel_size, 1))),
53
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
54
+ padding=get_padding(kernel_size, 1)))
55
+ ])
56
+ self.convs2.apply(init_weights)
57
+
58
+ def forward(self, x):
59
+ for c1, c2 in zip(self.convs1, self.convs2):
60
+ xt = F.leaky_relu(x, LRELU_SLOPE)
61
+ xt = c1(xt)
62
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
63
+ xt = c2(xt)
64
+ x = xt + x
65
+ return x
66
+
67
+ def remove_weight_norm(self):
68
+ for l in self.convs1:
69
+ remove_weight_norm(l)
70
+ for l in self.convs2:
71
+ remove_weight_norm(l)
72
+
73
+
74
+ class ResBlock2(torch.nn.Module):
75
+ def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
76
+ super(ResBlock2, self).__init__()
77
+ self.h = h
78
+ self.convs = nn.ModuleList([
79
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
80
+ padding=get_padding(kernel_size, dilation[0]))),
81
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
82
+ padding=get_padding(kernel_size, dilation[1])))
83
+ ])
84
+ self.convs.apply(init_weights)
85
+
86
+ def forward(self, x):
87
+ for c in self.convs:
88
+ xt = F.leaky_relu(x, LRELU_SLOPE)
89
+ xt = c(xt)
90
+ x = xt + x
91
+ return x
92
+
93
+ def remove_weight_norm(self):
94
+ for l in self.convs:
95
+ remove_weight_norm(l)
96
+
97
+
98
+ class SineGen(torch.nn.Module):
99
+ """ Definition of sine generator
100
+ SineGen(samp_rate, harmonic_num = 0,
101
+ sine_amp = 0.1, noise_std = 0.003,
102
+ voiced_threshold = 0,
103
+ flag_for_pulse=False)
104
+ samp_rate: sampling rate in Hz
105
+ harmonic_num: number of harmonic overtones (default 0)
106
+ sine_amp: amplitude of sine-wavefrom (default 0.1)
107
+ noise_std: std of Gaussian noise (default 0.003)
108
+ voiced_thoreshold: F0 threshold for U/V classification (default 0)
109
+ flag_for_pulse: this SinGen is used inside PulseGen (default False)
110
+ Note: when flag_for_pulse is True, the first time step of a voiced
111
+ segment is always sin(np.pi) or cos(0)
112
+ """
113
+
114
+ def __init__(self, samp_rate, harmonic_num=0,
115
+ sine_amp=0.1, noise_std=0.003,
116
+ voiced_threshold=0,
117
+ flag_for_pulse=False):
118
+ super(SineGen, self).__init__()
119
+ self.sine_amp = sine_amp
120
+ self.noise_std = noise_std
121
+ self.harmonic_num = harmonic_num
122
+ self.dim = self.harmonic_num + 1
123
+ self.sampling_rate = samp_rate
124
+ self.voiced_threshold = voiced_threshold
125
+ self.flag_for_pulse = flag_for_pulse
126
+
127
+ def _f02uv(self, f0):
128
+ # generate uv signal
129
+ uv = (f0 > self.voiced_threshold).type(torch.float32)
130
+ return uv
131
+
132
+ def _f02sine(self, f0_values):
133
+ """ f0_values: (batchsize, length, dim)
134
+ where dim indicates fundamental tone and overtones
135
+ """
136
+ # convert to F0 in rad. The interger part n can be ignored
137
+ # because 2 * np.pi * n doesn't affect phase
138
+ rad_values = (f0_values / self.sampling_rate) % 1
139
+
140
+ # initial phase noise (no noise for fundamental component)
141
+ rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \
142
+ device=f0_values.device)
143
+ rand_ini[:, 0] = 0
144
+ rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
145
+
146
+ # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
147
+ if not self.flag_for_pulse:
148
+ # for normal case
149
+
150
+ # To prevent torch.cumsum numerical overflow,
151
+ # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1.
152
+ # Buffer tmp_over_one_idx indicates the time step to add -1.
153
+ # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi
154
+ tmp_over_one = torch.cumsum(rad_values, 1) % 1
155
+ tmp_over_one_idx = (torch.diff(tmp_over_one, dim=1)) < 0
156
+ cumsum_shift = torch.zeros_like(rad_values)
157
+ cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
158
+
159
+ sines = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1)
160
+ * 2 * np.pi)
161
+ else:
162
+ # If necessary, make sure that the first time step of every
163
+ # voiced segments is sin(pi) or cos(0)
164
+ # This is used for pulse-train generation
165
+
166
+ # identify the last time step in unvoiced segments
167
+ uv = self._f02uv(f0_values)
168
+ uv_1 = torch.roll(uv, shifts=-1, dims=1)
169
+ uv_1[:, -1, :] = 1
170
+ u_loc = (uv < 1) * (uv_1 > 0)
171
+
172
+ # get the instantanouse phase
173
+ tmp_cumsum = torch.cumsum(rad_values, dim=1)
174
+ # different batch needs to be processed differently
175
+ for idx in range(f0_values.shape[0]):
176
+ temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
177
+ temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
178
+ # stores the accumulation of i.phase within
179
+ # each voiced segments
180
+ tmp_cumsum[idx, :, :] = 0
181
+ tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
182
+
183
+ # rad_values - tmp_cumsum: remove the accumulation of i.phase
184
+ # within the previous voiced segment.
185
+ i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
186
+
187
+ # get the sines
188
+ sines = torch.cos(i_phase * 2 * np.pi)
189
+ return sines
190
+
191
+ def forward(self, f0):
192
+ """ sine_tensor, uv = forward(f0)
193
+ input F0: tensor(batchsize=1, length, dim=1)
194
+ f0 for unvoiced steps should be 0
195
+ output sine_tensor: tensor(batchsize=1, length, dim)
196
+ output uv: tensor(batchsize=1, length, 1)
197
+ """
198
+ with torch.no_grad():
199
+ f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim,
200
+ device=f0.device)
201
+ # fundamental component
202
+ fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
203
+
204
+ # generate sine waveforms
205
+ sine_waves = self._f02sine(fn) * self.sine_amp
206
+
207
+ # generate uv signal
208
+ # uv = torch.ones(f0.shape)
209
+ # uv = uv * (f0 > self.voiced_threshold)
210
+ uv = self._f02uv(f0)
211
+
212
+ # noise: for unvoiced should be similar to sine_amp
213
+ # std = self.sine_amp/3 -> max value ~ self.sine_amp
214
+ # . for voiced regions is self.noise_std
215
+ noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
216
+ noise = noise_amp * torch.randn_like(sine_waves)
217
+
218
+ # first: set the unvoiced part to 0 by uv
219
+ # then: additive noise
220
+ sine_waves = sine_waves * uv + noise
221
+ return sine_waves, uv, noise
222
+
223
+
224
+ class SourceModuleHnNSF(torch.nn.Module):
225
+ """ SourceModule for hn-nsf
226
+ SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
227
+ add_noise_std=0.003, voiced_threshod=0)
228
+ sampling_rate: sampling_rate in Hz
229
+ harmonic_num: number of harmonic above F0 (default: 0)
230
+ sine_amp: amplitude of sine source signal (default: 0.1)
231
+ add_noise_std: std of additive Gaussian noise (default: 0.003)
232
+ note that amplitude of noise in unvoiced is decided
233
+ by sine_amp
234
+ voiced_threshold: threhold to set U/V given F0 (default: 0)
235
+ Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
236
+ F0_sampled (batchsize, length, 1)
237
+ Sine_source (batchsize, length, 1)
238
+ noise_source (batchsize, length 1)
239
+ uv (batchsize, length, 1)
240
+ """
241
+
242
+ def __init__(self, sampling_rate, harmonic_num=0, sine_amp=0.1,
243
+ add_noise_std=0.003, voiced_threshod=0):
244
+ super(SourceModuleHnNSF, self).__init__()
245
+
246
+ self.sine_amp = sine_amp
247
+ self.noise_std = add_noise_std
248
+
249
+ # to produce sine waveforms
250
+ self.l_sin_gen = SineGen(sampling_rate, harmonic_num,
251
+ sine_amp, add_noise_std, voiced_threshod)
252
+
253
+ # to merge source harmonics into a single excitation
254
+ self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
255
+ self.l_tanh = torch.nn.Tanh()
256
+
257
+ def forward(self, x):
258
+ """
259
+ Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
260
+ F0_sampled (batchsize, length, 1)
261
+ Sine_source (batchsize, length, 1)
262
+ noise_source (batchsize, length 1)
263
+ """
264
+ # source for harmonic branch
265
+ sine_wavs, uv, _ = self.l_sin_gen(x)
266
+ sine_merge = self.l_tanh(self.l_linear(sine_wavs))
267
+
268
+ # source for noise branch, in the same shape as uv
269
+ noise = torch.randn_like(uv) * self.sine_amp / 3
270
+ return sine_merge, noise, uv
271
+
272
+
273
+ class Generator(torch.nn.Module):
274
+ def __init__(self, h):
275
+ super(Generator, self).__init__()
276
+ self.h = h
277
+
278
+ self.num_kernels = len(h["resblock_kernel_sizes"])
279
+ self.num_upsamples = len(h["upsample_rates"])
280
+ self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(h["upsample_rates"]))
281
+ self.m_source = SourceModuleHnNSF(
282
+ sampling_rate=h["sampling_rate"],
283
+ harmonic_num=8)
284
+ self.noise_convs = nn.ModuleList()
285
+ self.conv_pre = weight_norm(Conv1d(h["inter_channels"], h["upsample_initial_channel"], 7, 1, padding=3))
286
+ resblock = ResBlock1 if h["resblock"] == '1' else ResBlock2
287
+ self.ups = nn.ModuleList()
288
+ for i, (u, k) in enumerate(zip(h["upsample_rates"], h["upsample_kernel_sizes"])):
289
+ c_cur = h["upsample_initial_channel"] // (2 ** (i + 1))
290
+ self.ups.append(weight_norm(
291
+ ConvTranspose1d(h["upsample_initial_channel"] // (2 ** i), h["upsample_initial_channel"] // (2 ** (i + 1)),
292
+ k, u, padding=(k - u) // 2)))
293
+ if i + 1 < len(h["upsample_rates"]): #
294
+ stride_f0 = np.prod(h["upsample_rates"][i + 1:])
295
+ self.noise_convs.append(Conv1d(
296
+ 1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2))
297
+ else:
298
+ self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
299
+ self.resblocks = nn.ModuleList()
300
+ for i in range(len(self.ups)):
301
+ ch = h["upsample_initial_channel"] // (2 ** (i + 1))
302
+ for j, (k, d) in enumerate(zip(h["resblock_kernel_sizes"], h["resblock_dilation_sizes"])):
303
+ self.resblocks.append(resblock(h, ch, k, d))
304
+
305
+ self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
306
+ self.ups.apply(init_weights)
307
+ self.conv_post.apply(init_weights)
308
+ self.cond = nn.Conv1d(h['gin_channels'], h['upsample_initial_channel'], 1)
309
+
310
+ def forward(self, x, f0, g=None):
311
+ # print(1,x.shape,f0.shape,f0[:, None].shape)
312
+ f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2) # bs,n,t
313
+ # print(2,f0.shape)
314
+ har_source, noi_source, uv = self.m_source(f0)
315
+ har_source = har_source.transpose(1, 2)
316
+ x = self.conv_pre(x)
317
+ x = x + self.cond(g)
318
+ # print(124,x.shape,har_source.shape)
319
+ for i in range(self.num_upsamples):
320
+ x = F.leaky_relu(x, LRELU_SLOPE)
321
+ # print(3,x.shape)
322
+ x = self.ups[i](x)
323
+ x_source = self.noise_convs[i](har_source)
324
+ # print(4,x_source.shape,har_source.shape,x.shape)
325
+ x = x + x_source
326
+ xs = None
327
+ for j in range(self.num_kernels):
328
+ if xs is None:
329
+ xs = self.resblocks[i * self.num_kernels + j](x)
330
+ else:
331
+ xs += self.resblocks[i * self.num_kernels + j](x)
332
+ x = xs / self.num_kernels
333
+ x = F.leaky_relu(x)
334
+ x = self.conv_post(x)
335
+ x = torch.tanh(x)
336
+
337
+ return x
338
+
339
+ def remove_weight_norm(self):
340
+ print('Removing weight norm...')
341
+ for l in self.ups:
342
+ remove_weight_norm(l)
343
+ for l in self.resblocks:
344
+ l.remove_weight_norm()
345
+ remove_weight_norm(self.conv_pre)
346
+ remove_weight_norm(self.conv_post)
347
+
348
+
349
+ class DiscriminatorP(torch.nn.Module):
350
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
351
+ super(DiscriminatorP, self).__init__()
352
+ self.period = period
353
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
354
+ self.convs = nn.ModuleList([
355
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
356
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
357
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
358
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
359
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
360
+ ])
361
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
362
+
363
+ def forward(self, x):
364
+ fmap = []
365
+
366
+ # 1d to 2d
367
+ b, c, t = x.shape
368
+ if t % self.period != 0: # pad first
369
+ n_pad = self.period - (t % self.period)
370
+ x = F.pad(x, (0, n_pad), "reflect")
371
+ t = t + n_pad
372
+ x = x.view(b, c, t // self.period, self.period)
373
+
374
+ for l in self.convs:
375
+ x = l(x)
376
+ x = F.leaky_relu(x, LRELU_SLOPE)
377
+ fmap.append(x)
378
+ x = self.conv_post(x)
379
+ fmap.append(x)
380
+ x = torch.flatten(x, 1, -1)
381
+
382
+ return x, fmap
383
+
384
+
385
+ class MultiPeriodDiscriminator(torch.nn.Module):
386
+ def __init__(self, periods=None):
387
+ super(MultiPeriodDiscriminator, self).__init__()
388
+ self.periods = periods if periods is not None else [2, 3, 5, 7, 11]
389
+ self.discriminators = nn.ModuleList()
390
+ for period in self.periods:
391
+ self.discriminators.append(DiscriminatorP(period))
392
+
393
+ def forward(self, y, y_hat):
394
+ y_d_rs = []
395
+ y_d_gs = []
396
+ fmap_rs = []
397
+ fmap_gs = []
398
+ for i, d in enumerate(self.discriminators):
399
+ y_d_r, fmap_r = d(y)
400
+ y_d_g, fmap_g = d(y_hat)
401
+ y_d_rs.append(y_d_r)
402
+ fmap_rs.append(fmap_r)
403
+ y_d_gs.append(y_d_g)
404
+ fmap_gs.append(fmap_g)
405
+
406
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
407
+
408
+
409
+ class DiscriminatorS(torch.nn.Module):
410
+ def __init__(self, use_spectral_norm=False):
411
+ super(DiscriminatorS, self).__init__()
412
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
413
+ self.convs = nn.ModuleList([
414
+ norm_f(Conv1d(1, 128, 15, 1, padding=7)),
415
+ norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
416
+ norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
417
+ norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
418
+ norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
419
+ norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
420
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
421
+ ])
422
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
423
+
424
+ def forward(self, x):
425
+ fmap = []
426
+ for l in self.convs:
427
+ x = l(x)
428
+ x = F.leaky_relu(x, LRELU_SLOPE)
429
+ fmap.append(x)
430
+ x = self.conv_post(x)
431
+ fmap.append(x)
432
+ x = torch.flatten(x, 1, -1)
433
+
434
+ return x, fmap
435
+
436
+
437
+ class MultiScaleDiscriminator(torch.nn.Module):
438
+ def __init__(self):
439
+ super(MultiScaleDiscriminator, self).__init__()
440
+ self.discriminators = nn.ModuleList([
441
+ DiscriminatorS(use_spectral_norm=True),
442
+ DiscriminatorS(),
443
+ DiscriminatorS(),
444
+ ])
445
+ self.meanpools = nn.ModuleList([
446
+ AvgPool1d(4, 2, padding=2),
447
+ AvgPool1d(4, 2, padding=2)
448
+ ])
449
+
450
+ def forward(self, y, y_hat):
451
+ y_d_rs = []
452
+ y_d_gs = []
453
+ fmap_rs = []
454
+ fmap_gs = []
455
+ for i, d in enumerate(self.discriminators):
456
+ if i != 0:
457
+ y = self.meanpools[i - 1](y)
458
+ y_hat = self.meanpools[i - 1](y_hat)
459
+ y_d_r, fmap_r = d(y)
460
+ y_d_g, fmap_g = d(y_hat)
461
+ y_d_rs.append(y_d_r)
462
+ fmap_rs.append(fmap_r)
463
+ y_d_gs.append(y_d_g)
464
+ fmap_gs.append(fmap_g)
465
+
466
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
467
+
468
+
469
+ def feature_loss(fmap_r, fmap_g):
470
+ loss = 0
471
+ for dr, dg in zip(fmap_r, fmap_g):
472
+ for rl, gl in zip(dr, dg):
473
+ loss += torch.mean(torch.abs(rl - gl))
474
+
475
+ return loss * 2
476
+
477
+
478
+ def discriminator_loss(disc_real_outputs, disc_generated_outputs):
479
+ loss = 0
480
+ r_losses = []
481
+ g_losses = []
482
+ for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
483
+ r_loss = torch.mean((1 - dr) ** 2)
484
+ g_loss = torch.mean(dg ** 2)
485
+ loss += (r_loss + g_loss)
486
+ r_losses.append(r_loss.item())
487
+ g_losses.append(g_loss.item())
488
+
489
+ return loss, r_losses, g_losses
490
+
491
+
492
+ def generator_loss(disc_outputs):
493
+ loss = 0
494
+ gen_losses = []
495
+ for dg in disc_outputs:
496
+ l = torch.mean((1 - dg) ** 2)
497
+ gen_losses.append(l)
498
+ loss += l
499
+
500
+ return loss, gen_losses
vdecoder/hifigan/nvSTFT.py ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import os
3
+ os.environ["LRU_CACHE_CAPACITY"] = "3"
4
+ import random
5
+ import torch
6
+ import torch.utils.data
7
+ import numpy as np
8
+ import librosa
9
+ from librosa.util import normalize
10
+ from librosa.filters import mel as librosa_mel_fn
11
+ from scipy.io.wavfile import read
12
+ import soundfile as sf
13
+
14
+ def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False):
15
+ sampling_rate = None
16
+ try:
17
+ data, sampling_rate = sf.read(full_path, always_2d=True)# than soundfile.
18
+ except Exception as ex:
19
+ print(f"'{full_path}' failed to load.\nException:")
20
+ print(ex)
21
+ if return_empty_on_exception:
22
+ return [], sampling_rate or target_sr or 32000
23
+ else:
24
+ raise Exception(ex)
25
+
26
+ if len(data.shape) > 1:
27
+ data = data[:, 0]
28
+ assert len(data) > 2# check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension)
29
+
30
+ if np.issubdtype(data.dtype, np.integer): # if audio data is type int
31
+ max_mag = -np.iinfo(data.dtype).min # maximum magnitude = min possible value of intXX
32
+ else: # if audio data is type fp32
33
+ max_mag = max(np.amax(data), -np.amin(data))
34
+ max_mag = (2**31)+1 if max_mag > (2**15) else ((2**15)+1 if max_mag > 1.01 else 1.0) # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
35
+
36
+ data = torch.FloatTensor(data.astype(np.float32))/max_mag
37
+
38
+ if (torch.isinf(data) | torch.isnan(data)).any() and return_empty_on_exception:# resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except
39
+ return [], sampling_rate or target_sr or 32000
40
+ if target_sr is not None and sampling_rate != target_sr:
41
+ data = torch.from_numpy(librosa.core.resample(data.numpy(), orig_sr=sampling_rate, target_sr=target_sr))
42
+ sampling_rate = target_sr
43
+
44
+ return data, sampling_rate
45
+
46
+ def dynamic_range_compression(x, C=1, clip_val=1e-5):
47
+ return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
48
+
49
+ def dynamic_range_decompression(x, C=1):
50
+ return np.exp(x) / C
51
+
52
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
53
+ return torch.log(torch.clamp(x, min=clip_val) * C)
54
+
55
+ def dynamic_range_decompression_torch(x, C=1):
56
+ return torch.exp(x) / C
57
+
58
+ class STFT():
59
+ def __init__(self, sr=22050, n_mels=80, n_fft=1024, win_size=1024, hop_length=256, fmin=20, fmax=11025, clip_val=1e-5):
60
+ self.target_sr = sr
61
+
62
+ self.n_mels = n_mels
63
+ self.n_fft = n_fft
64
+ self.win_size = win_size
65
+ self.hop_length = hop_length
66
+ self.fmin = fmin
67
+ self.fmax = fmax
68
+ self.clip_val = clip_val
69
+ self.mel_basis = {}
70
+ self.hann_window = {}
71
+
72
+ def get_mel(self, y, center=False):
73
+ sampling_rate = self.target_sr
74
+ n_mels = self.n_mels
75
+ n_fft = self.n_fft
76
+ win_size = self.win_size
77
+ hop_length = self.hop_length
78
+ fmin = self.fmin
79
+ fmax = self.fmax
80
+ clip_val = self.clip_val
81
+
82
+ if torch.min(y) < -1.:
83
+ print('min value is ', torch.min(y))
84
+ if torch.max(y) > 1.:
85
+ print('max value is ', torch.max(y))
86
+
87
+ if fmax not in self.mel_basis:
88
+ mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax)
89
+ self.mel_basis[str(fmax)+'_'+str(y.device)] = torch.from_numpy(mel).float().to(y.device)
90
+ self.hann_window[str(y.device)] = torch.hann_window(self.win_size).to(y.device)
91
+
92
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_length)/2), int((n_fft-hop_length)/2)), mode='reflect')
93
+ y = y.squeeze(1)
94
+
95
+ spec = torch.stft(y, n_fft, hop_length=hop_length, win_length=win_size, window=self.hann_window[str(y.device)],
96
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
97
+ # print(111,spec)
98
+ spec = torch.sqrt(spec.pow(2).sum(-1)+(1e-9))
99
+ # print(222,spec)
100
+ spec = torch.matmul(self.mel_basis[str(fmax)+'_'+str(y.device)], spec)
101
+ # print(333,spec)
102
+ spec = dynamic_range_compression_torch(spec, clip_val=clip_val)
103
+ # print(444,spec)
104
+ return spec
105
+
106
+ def __call__(self, audiopath):
107
+ audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr)
108
+ spect = self.get_mel(audio.unsqueeze(0)).squeeze(0)
109
+ return spect
110
+
111
+ stft = STFT()
vdecoder/hifigan/utils.py ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import glob
2
+ import os
3
+ import matplotlib
4
+ import torch
5
+ from torch.nn.utils import weight_norm
6
+ matplotlib.use("Agg")
7
+ import matplotlib.pylab as plt
8
+
9
+
10
+ def plot_spectrogram(spectrogram):
11
+ fig, ax = plt.subplots(figsize=(10, 2))
12
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
13
+ interpolation='none')
14
+ plt.colorbar(im, ax=ax)
15
+
16
+ fig.canvas.draw()
17
+ plt.close()
18
+
19
+ return fig
20
+
21
+
22
+ def init_weights(m, mean=0.0, std=0.01):
23
+ classname = m.__class__.__name__
24
+ if classname.find("Conv") != -1:
25
+ m.weight.data.normal_(mean, std)
26
+
27
+
28
+ def apply_weight_norm(m):
29
+ classname = m.__class__.__name__
30
+ if classname.find("Conv") != -1:
31
+ weight_norm(m)
32
+
33
+
34
+ def get_padding(kernel_size, dilation=1):
35
+ return int((kernel_size*dilation - dilation)/2)
36
+
37
+
38
+ def load_checkpoint(filepath, device):
39
+ assert os.path.isfile(filepath)
40
+ print("Loading '{}'".format(filepath))
41
+ checkpoint_dict = torch.load(filepath, map_location=device)
42
+ print("Complete.")
43
+ return checkpoint_dict
44
+
45
+
46
+ def save_checkpoint(filepath, obj):
47
+ print("Saving checkpoint to {}".format(filepath))
48
+ torch.save(obj, filepath)
49
+ print("Complete.")
50
+
51
+
52
+ def del_old_checkpoints(cp_dir, prefix, n_models=2):
53
+ pattern = os.path.join(cp_dir, prefix + '????????')
54
+ cp_list = glob.glob(pattern) # get checkpoint paths
55
+ cp_list = sorted(cp_list)# sort by iter
56
+ if len(cp_list) > n_models: # if more than n_models models are found
57
+ for cp in cp_list[:-n_models]:# delete the oldest models other than lastest n_models
58
+ open(cp, 'w').close()# empty file contents
59
+ os.unlink(cp)# delete file (move to trash when using Colab)
60
+
61
+
62
+ def scan_checkpoint(cp_dir, prefix):
63
+ pattern = os.path.join(cp_dir, prefix + '????????')
64
+ cp_list = glob.glob(pattern)
65
+ if len(cp_list) == 0:
66
+ return None
67
+ return sorted(cp_list)[-1]
68
+