innnky commited on
Commit
e3a7134
1 Parent(s): 2bf872d
G_49200.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da931601660595a5ab841b25668faf97ee66e10d3ae5da03f69c5e61f28476fd
3
+ size 479164585
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2021 Jaehyeon Kim
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.
all_emotions.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48e81667f1fc4ce2b2eaed80fadd0871e1ddfc8933767915954c39ac854d5724
3
+ size 22356096
app.py ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import torch
3
+ import commons
4
+ import utils
5
+ from models import SynthesizerTrn
6
+ from text.symbols import symbols
7
+ from text import text_to_sequence
8
+ import numpy as np
9
+
10
+
11
+ def get_text(text, hps):
12
+ text_norm = text_to_sequence(text, hps.data.text_cleaners)
13
+ if hps.data.add_blank:
14
+ text_norm = commons.intersperse(text_norm, 0)
15
+ text_norm = torch.LongTensor(text_norm)
16
+ return text_norm
17
+ hps = utils.get_hparams_from_file("./configs/vtubers.json")
18
+ net_g = SynthesizerTrn(
19
+ len(symbols),
20
+ hps.data.filter_length // 2 + 1,
21
+ hps.train.segment_size // hps.data.hop_length,
22
+ n_speakers=hps.data.n_speakers,
23
+ **hps.model)
24
+ _ = net_g.eval()
25
+
26
+ _ = utils.load_checkpoint("./G_49200.pth", net_g, None)
27
+ all_emotions = np.load("all_emotions.npy")
28
+ emotion_dict = {
29
+ "小声": 2077,
30
+ "激动": 111,
31
+ "平静1": 434,
32
+ "平静2": 3554
33
+ }
34
+ import random
35
+ def tts(txt, emotion):
36
+ stn_tst = get_text(txt, hps)
37
+ randsample = None
38
+ with torch.no_grad():
39
+ x_tst = stn_tst.unsqueeze(0)
40
+ x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
41
+ sid = torch.LongTensor([0])
42
+ if type(emotion) ==int:
43
+ emo = torch.FloatTensor(all_emotions[emotion]).unsqueeze(0)
44
+ elif emotion == "random":
45
+ emo = torch.randn([1,1024])
46
+ elif emotion == "random_sample":
47
+ randint = random.randint(0, all_emotions.shape[0])
48
+ emo = torch.FloatTensor(all_emotions[randint]).unsqueeze(0)
49
+ randsample = randint
50
+ elif emotion.endswith("wav"):
51
+ import emotion_extract
52
+ emo = torch.FloatTensor(emotion_extract.extract_wav(emotion))
53
+ else:
54
+ emo = torch.FloatTensor(all_emotions[emotion_dict[emotion]]).unsqueeze(0)
55
+
56
+ audio = net_g.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=0.667, noise_scale_w=0.8, length_scale=1, emo=emo)[0][0,0].data.float().numpy()
57
+ return audio, randsample
58
+
59
+
60
+ def tts1(text, emotion):
61
+ if len(text) > 150:
62
+ return "Error: Text is too long", None
63
+ audio, _ = tts(text, emotion)
64
+ return "Success", (hps.data.sampling_rate, audio)
65
+
66
+ def tts2(text):
67
+ if len(text) > 150:
68
+ return "Error: Text is too long", None
69
+ audio, randsample = tts(text, "random_sample")
70
+
71
+ return str(randsample), (hps.data.sampling_rate, audio)
72
+
73
+ def tts3(text, sample):
74
+ if len(text) > 150:
75
+ return "Error: Text is too long", None
76
+ try:
77
+ audio, _ = tts(text, int(sample))
78
+ return "Success", (hps.data.sampling_rate, audio)
79
+ except:
80
+ return "输入参数不为整数或其他错误"
81
+
82
+ app = gr.Blocks()
83
+ with app:
84
+ with gr.Tabs():
85
+ with gr.TabItem("使用预制情感合成"):
86
+ tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
87
+ tts_input2 = gr.Dropdown(label="情感", choices=list(emotion_dict.keys()), value="平静1")
88
+ tts_submit = gr.Button("合成音频", variant="primary")
89
+ tts_output1 = gr.Textbox(label="Message")
90
+ tts_output2 = gr.Audio(label="Output")
91
+ tts_submit.click(tts1, [tts_input1, tts_input2], [tts_output1, tts_output2])
92
+ with gr.TabItem("随机抽取训练集样本作为情感参数"):
93
+ tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
94
+ tts_submit = gr.Button("合成音频", variant="primary")
95
+ tts_output1 = gr.Textbox(label="随机样本id(可用于第三个tab中合成)")
96
+ tts_output2 = gr.Audio(label="Output")
97
+ tts_submit.click(tts2, [tts_input1], [tts_output1, tts_output2])
98
+
99
+ with gr.TabItem("使用情感样本id作为情感参数"):
100
+
101
+ tts_input1 = gr.TextArea(label="日语文本", value="こんにちは。私わあやちねねです。")
102
+ tts_input2 = gr.Number(label="情感样本id", value=2004)
103
+ tts_submit = gr.Button("合成音频", variant="primary")
104
+ tts_output1 = gr.Textbox(label="Message")
105
+ tts_output2 = gr.Audio(label="Output")
106
+ tts_submit.click(tts3, [tts_input1, tts_input2], [tts_output1, tts_output2])
107
+
108
+ with gr.TabItem("使用参考音频作为情感参数"):
109
+ tts_input1 = gr.TextArea(label="text", value="暂未实现")
110
+
111
+ app.launch()
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 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
8
+ def init_weights(m, mean=0.0, std=0.01):
9
+ classname = m.__class__.__name__
10
+ if classname.find("Conv") != -1:
11
+ m.weight.data.normal_(mean, std)
12
+
13
+
14
+ def get_padding(kernel_size, dilation=1):
15
+ return int((kernel_size*dilation - dilation)/2)
16
+
17
+
18
+ def convert_pad_shape(pad_shape):
19
+ l = pad_shape[::-1]
20
+ pad_shape = [item for sublist in l for item in sublist]
21
+ return pad_shape
22
+
23
+
24
+ def intersperse(lst, item):
25
+ result = [item] * (len(lst) * 2 + 1)
26
+ result[1::2] = lst
27
+ return result
28
+
29
+
30
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
31
+ """KL(P||Q)"""
32
+ kl = (logs_q - logs_p) - 0.5
33
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
34
+ return kl
35
+
36
+
37
+ def rand_gumbel(shape):
38
+ """Sample from the Gumbel distribution, protect from overflows."""
39
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
40
+ return -torch.log(-torch.log(uniform_samples))
41
+
42
+
43
+ def rand_gumbel_like(x):
44
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
45
+ return g
46
+
47
+
48
+ def slice_segments(x, ids_str, segment_size=4):
49
+ ret = torch.zeros_like(x[:, :, :segment_size])
50
+ for i in range(x.size(0)):
51
+ idx_str = ids_str[i]
52
+ idx_end = idx_str + segment_size
53
+ ret[i] = x[i, :, idx_str:idx_end]
54
+ return ret
55
+
56
+
57
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
58
+ b, d, t = x.size()
59
+ if x_lengths is None:
60
+ x_lengths = t
61
+ ids_str_max = x_lengths - segment_size + 1
62
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
63
+ ret = slice_segments(x, ids_str, segment_size)
64
+ return ret, ids_str
65
+
66
+
67
+ def get_timing_signal_1d(
68
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
69
+ position = torch.arange(length, dtype=torch.float)
70
+ num_timescales = channels // 2
71
+ log_timescale_increment = (
72
+ math.log(float(max_timescale) / float(min_timescale)) /
73
+ (num_timescales - 1))
74
+ inv_timescales = min_timescale * torch.exp(
75
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
76
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
77
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
78
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
79
+ signal = signal.view(1, channels, length)
80
+ return signal
81
+
82
+
83
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
84
+ b, channels, length = x.size()
85
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
86
+ return x + signal.to(dtype=x.dtype, device=x.device)
87
+
88
+
89
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
90
+ b, channels, length = x.size()
91
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
92
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
93
+
94
+
95
+ def subsequent_mask(length):
96
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
97
+ return mask
98
+
99
+
100
+ @torch.jit.script
101
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
102
+ n_channels_int = n_channels[0]
103
+ in_act = input_a + input_b
104
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
105
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
106
+ acts = t_act * s_act
107
+ return acts
108
+
109
+
110
+ def convert_pad_shape(pad_shape):
111
+ l = pad_shape[::-1]
112
+ pad_shape = [item for sublist in l for item in sublist]
113
+ return pad_shape
114
+
115
+
116
+ def shift_1d(x):
117
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
118
+ return x
119
+
120
+
121
+ def sequence_mask(length, max_length=None):
122
+ if max_length is None:
123
+ max_length = length.max()
124
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
125
+ return x.unsqueeze(0) < length.unsqueeze(1)
126
+
127
+
128
+ def generate_path(duration, mask):
129
+ """
130
+ duration: [b, 1, t_x]
131
+ mask: [b, 1, t_y, t_x]
132
+ """
133
+ device = duration.device
134
+
135
+ b, _, t_y, t_x = mask.shape
136
+ cum_duration = torch.cumsum(duration, -1)
137
+
138
+ cum_duration_flat = cum_duration.view(b * t_x)
139
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
140
+ path = path.view(b, t_x, t_y)
141
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
142
+ path = path.unsqueeze(1).transpose(2,3) * mask
143
+ return path
144
+
145
+
146
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
147
+ if isinstance(parameters, torch.Tensor):
148
+ parameters = [parameters]
149
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
150
+ norm_type = float(norm_type)
151
+ if clip_value is not None:
152
+ clip_value = float(clip_value)
153
+
154
+ total_norm = 0
155
+ for p in parameters:
156
+ param_norm = p.grad.data.norm(norm_type)
157
+ total_norm += param_norm.item() ** norm_type
158
+ if clip_value is not None:
159
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
160
+ total_norm = total_norm ** (1. / norm_type)
161
+ return total_norm
configs/vtubers.json ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 400,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 2e-4,
8
+ "betas": [0.8, 0.99],
9
+ "eps": 1e-9,
10
+ "batch_size": 8,
11
+ "fp16_run": false,
12
+ "lr_decay": 0.999875,
13
+ "segment_size": 8192,
14
+ "init_lr_ratio": 1,
15
+ "warmup_epochs": 0,
16
+ "c_mel": 45,
17
+ "c_kl": 1.0
18
+ },
19
+ "data": {
20
+ "training_files":"filelists/train.txt.cleaned",
21
+ "validation_files":"filelists/val.txt.cleaned",
22
+ "text_cleaners":["japanese_cleaners"],
23
+ "max_wav_value": 32768.0,
24
+ "sampling_rate": 22050,
25
+ "filter_length": 1024,
26
+ "hop_length": 256,
27
+ "win_length": 1024,
28
+ "n_mel_channels": 80,
29
+ "mel_fmin": 0.0,
30
+ "mel_fmax": null,
31
+ "add_blank": true,
32
+ "n_speakers": 7,
33
+ "cleaned_text": true
34
+ },
35
+ "model": {
36
+ "inter_channels": 192,
37
+ "hidden_channels": 192,
38
+ "filter_channels": 768,
39
+ "n_heads": 2,
40
+ "n_layers": 6,
41
+ "kernel_size": 3,
42
+ "p_dropout": 0.1,
43
+ "resblock": "1",
44
+ "resblock_kernel_sizes": [3,7,11],
45
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
46
+ "upsample_rates": [8,8,2,2],
47
+ "upsample_initial_channel": 512,
48
+ "upsample_kernel_sizes": [16,16,4,4],
49
+ "n_layers_q": 3,
50
+ "use_spectral_norm": false,
51
+ "gin_channels": 256
52
+ }
53
+ }
data_utils.py ADDED
@@ -0,0 +1,261 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
10
+ from utils import load_wav_to_torch, load_filepaths_and_text
11
+ from text import text_to_sequence, cleaned_text_to_sequence
12
+
13
+ """Multi speaker version"""
14
+ class TextAudioSpeakerLoader(torch.utils.data.Dataset):
15
+ """
16
+ 1) loads audio, speaker_id, text pairs
17
+ 2) normalizes text and converts them to sequences of integers
18
+ 3) computes spectrograms from audio files.
19
+ """
20
+ def __init__(self, audiopaths_sid_text, hparams):
21
+ self.audiopaths_sid_text = load_filepaths_and_text(audiopaths_sid_text)
22
+ self.text_cleaners = hparams.text_cleaners
23
+ self.max_wav_value = hparams.max_wav_value
24
+ self.sampling_rate = hparams.sampling_rate
25
+ self.filter_length = hparams.filter_length
26
+ self.hop_length = hparams.hop_length
27
+ self.win_length = hparams.win_length
28
+ self.sampling_rate = hparams.sampling_rate
29
+
30
+ self.cleaned_text = getattr(hparams, "cleaned_text", False)
31
+
32
+ self.add_blank = hparams.add_blank
33
+ self.min_text_len = getattr(hparams, "min_text_len", 1)
34
+ self.max_text_len = getattr(hparams, "max_text_len", 190)
35
+
36
+ random.seed(1234)
37
+ random.shuffle(self.audiopaths_sid_text)
38
+ self._filter()
39
+
40
+ def _filter(self):
41
+ """
42
+ Filter text & store spec lengths
43
+ """
44
+ # Store spectrogram lengths for Bucketing
45
+ # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
46
+ # spec_length = wav_length // hop_length
47
+
48
+ audiopaths_sid_text_new = []
49
+ lengths = []
50
+ for audiopath, sid, text in self.audiopaths_sid_text:
51
+ if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
52
+ audiopaths_sid_text_new.append([audiopath, sid, text])
53
+ lengths.append(os.path.getsize(audiopath) // (2 * self.hop_length))
54
+ self.audiopaths_sid_text = audiopaths_sid_text_new
55
+ self.lengths = lengths
56
+
57
+ def get_audio_text_speaker_pair(self, audiopath_sid_text):
58
+ # separate filename, speaker_id and text
59
+ audiopath, sid, text = audiopath_sid_text[0], audiopath_sid_text[1], audiopath_sid_text[2]
60
+ text = self.get_text(text)
61
+ spec, wav = self.get_audio(audiopath)
62
+ sid = self.get_sid(sid)
63
+ emo = torch.FloatTensor(np.load(audiopath+".emo.npy"))
64
+ return (text, spec, wav, sid, emo)
65
+
66
+ def get_audio(self, filename):
67
+ audio, sampling_rate = load_wav_to_torch(filename)
68
+ if sampling_rate != self.sampling_rate:
69
+ raise ValueError("{} {} SR doesn't match target {} SR".format(
70
+ sampling_rate, self.sampling_rate))
71
+ audio_norm = audio / self.max_wav_value
72
+ audio_norm = audio_norm.unsqueeze(0)
73
+ spec_filename = filename.replace(".wav", ".spec.pt")
74
+ if os.path.exists(spec_filename):
75
+ spec = torch.load(spec_filename)
76
+ else:
77
+ spec = spectrogram_torch(audio_norm, self.filter_length,
78
+ self.sampling_rate, self.hop_length, self.win_length,
79
+ center=False)
80
+ spec = torch.squeeze(spec, 0)
81
+ torch.save(spec, spec_filename)
82
+ return spec, audio_norm
83
+
84
+ def get_text(self, text):
85
+ if self.cleaned_text:
86
+ text_norm = cleaned_text_to_sequence(text)
87
+ else:
88
+ text_norm = text_to_sequence(text, self.text_cleaners)
89
+ if self.add_blank:
90
+ text_norm = commons.intersperse(text_norm, 0)
91
+ text_norm = torch.LongTensor(text_norm)
92
+ return text_norm
93
+
94
+ def get_sid(self, sid):
95
+ sid = torch.LongTensor([int(sid)])
96
+ return sid
97
+
98
+ def __getitem__(self, index):
99
+ return self.get_audio_text_speaker_pair(self.audiopaths_sid_text[index])
100
+
101
+ def __len__(self):
102
+ return len(self.audiopaths_sid_text)
103
+
104
+
105
+ class TextAudioSpeakerCollate():
106
+ """ Zero-pads model inputs and targets
107
+ """
108
+ def __init__(self, return_ids=False):
109
+ self.return_ids = return_ids
110
+
111
+ def __call__(self, batch):
112
+ """Collate's training batch from normalized text, audio and speaker identities
113
+ PARAMS
114
+ ------
115
+ batch: [text_normalized, spec_normalized, wav_normalized, sid]
116
+ """
117
+ # Right zero-pad all one-hot text sequences to max input length
118
+ _, ids_sorted_decreasing = torch.sort(
119
+ torch.LongTensor([x[1].size(1) for x in batch]),
120
+ dim=0, descending=True)
121
+
122
+ max_text_len = max([len(x[0]) for x in batch])
123
+ max_spec_len = max([x[1].size(1) for x in batch])
124
+ max_wav_len = max([x[2].size(1) for x in batch])
125
+
126
+ text_lengths = torch.LongTensor(len(batch))
127
+ spec_lengths = torch.LongTensor(len(batch))
128
+ wav_lengths = torch.LongTensor(len(batch))
129
+ sid = torch.LongTensor(len(batch))
130
+
131
+ text_padded = torch.LongTensor(len(batch), max_text_len)
132
+ spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
133
+ wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
134
+ emo = torch.FloatTensor(len(batch), 1024)
135
+
136
+ text_padded.zero_()
137
+ spec_padded.zero_()
138
+ wav_padded.zero_()
139
+ emo.zero_()
140
+
141
+ for i in range(len(ids_sorted_decreasing)):
142
+ row = batch[ids_sorted_decreasing[i]]
143
+
144
+ text = row[0]
145
+ text_padded[i, :text.size(0)] = text
146
+ text_lengths[i] = text.size(0)
147
+
148
+ spec = row[1]
149
+ spec_padded[i, :, :spec.size(1)] = spec
150
+ spec_lengths[i] = spec.size(1)
151
+
152
+ wav = row[2]
153
+ wav_padded[i, :, :wav.size(1)] = wav
154
+ wav_lengths[i] = wav.size(1)
155
+
156
+ sid[i] = row[3]
157
+
158
+ emo[i, :] = row[4]
159
+
160
+ if self.return_ids:
161
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid, ids_sorted_decreasing
162
+ return text_padded, text_lengths, spec_padded, spec_lengths, wav_padded, wav_lengths, sid,emo
163
+
164
+
165
+ class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
166
+ """
167
+ Maintain similar input lengths in a batch.
168
+ Length groups are specified by boundaries.
169
+ Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
170
+
171
+ It removes samples which are not included in the boundaries.
172
+ Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
173
+ """
174
+ def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
175
+ super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
176
+ self.lengths = dataset.lengths
177
+ self.batch_size = batch_size
178
+ self.boundaries = boundaries
179
+
180
+ self.buckets, self.num_samples_per_bucket = self._create_buckets()
181
+ self.total_size = sum(self.num_samples_per_bucket)
182
+ self.num_samples = self.total_size // self.num_replicas
183
+
184
+ def _create_buckets(self):
185
+ buckets = [[] for _ in range(len(self.boundaries) - 1)]
186
+ for i in range(len(self.lengths)):
187
+ length = self.lengths[i]
188
+ idx_bucket = self._bisect(length)
189
+ if idx_bucket != -1:
190
+ buckets[idx_bucket].append(i)
191
+
192
+ for i in range(len(buckets) - 1, 0, -1):
193
+ if len(buckets[i]) == 0:
194
+ buckets.pop(i)
195
+ self.boundaries.pop(i+1)
196
+
197
+ num_samples_per_bucket = []
198
+ for i in range(len(buckets)):
199
+ len_bucket = len(buckets[i])
200
+ total_batch_size = self.num_replicas * self.batch_size
201
+ rem = (total_batch_size - (len_bucket % total_batch_size)) % total_batch_size
202
+ num_samples_per_bucket.append(len_bucket + rem)
203
+ return buckets, num_samples_per_bucket
204
+
205
+ def __iter__(self):
206
+ # deterministically shuffle based on epoch
207
+ g = torch.Generator()
208
+ g.manual_seed(self.epoch)
209
+
210
+ indices = []
211
+ if self.shuffle:
212
+ for bucket in self.buckets:
213
+ indices.append(torch.randperm(len(bucket), generator=g).tolist())
214
+ else:
215
+ for bucket in self.buckets:
216
+ indices.append(list(range(len(bucket))))
217
+
218
+ batches = []
219
+ for i in range(len(self.buckets)):
220
+ bucket = self.buckets[i]
221
+ len_bucket = len(bucket)
222
+ ids_bucket = indices[i]
223
+ num_samples_bucket = self.num_samples_per_bucket[i]
224
+
225
+ # add extra samples to make it evenly divisible
226
+ rem = num_samples_bucket - len_bucket
227
+ ids_bucket = ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[:(rem % len_bucket)]
228
+
229
+ # subsample
230
+ ids_bucket = ids_bucket[self.rank::self.num_replicas]
231
+
232
+ # batching
233
+ for j in range(len(ids_bucket) // self.batch_size):
234
+ batch = [bucket[idx] for idx in ids_bucket[j*self.batch_size:(j+1)*self.batch_size]]
235
+ batches.append(batch)
236
+
237
+ if self.shuffle:
238
+ batch_ids = torch.randperm(len(batches), generator=g).tolist()
239
+ batches = [batches[i] for i in batch_ids]
240
+ self.batches = batches
241
+
242
+ assert len(self.batches) * self.batch_size == self.num_samples
243
+ return iter(self.batches)
244
+
245
+ def _bisect(self, x, lo=0, hi=None):
246
+ if hi is None:
247
+ hi = len(self.boundaries) - 1
248
+
249
+ if hi > lo:
250
+ mid = (hi + lo) // 2
251
+ if self.boundaries[mid] < x and x <= self.boundaries[mid+1]:
252
+ return mid
253
+ elif x <= self.boundaries[mid]:
254
+ return self._bisect(x, lo, mid)
255
+ else:
256
+ return self._bisect(x, mid + 1, hi)
257
+ else:
258
+ return -1
259
+
260
+ def __len__(self):
261
+ return self.num_samples // self.batch_size
emotion_extract.py ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from transformers import Wav2Vec2Processor
4
+ from transformers.models.wav2vec2.modeling_wav2vec2 import (
5
+ Wav2Vec2Model,
6
+ Wav2Vec2PreTrainedModel,
7
+ )
8
+ import os
9
+ import librosa
10
+ import numpy as np
11
+
12
+
13
+ class RegressionHead(nn.Module):
14
+ r"""Classification head."""
15
+
16
+ def __init__(self, config):
17
+ super().__init__()
18
+
19
+ self.dense = nn.Linear(config.hidden_size, config.hidden_size)
20
+ self.dropout = nn.Dropout(config.final_dropout)
21
+ self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
22
+
23
+ def forward(self, features, **kwargs):
24
+ x = features
25
+ x = self.dropout(x)
26
+ x = self.dense(x)
27
+ x = torch.tanh(x)
28
+ x = self.dropout(x)
29
+ x = self.out_proj(x)
30
+
31
+ return x
32
+
33
+
34
+ class EmotionModel(Wav2Vec2PreTrainedModel):
35
+ r"""Speech emotion classifier."""
36
+
37
+ def __init__(self, config):
38
+ super().__init__(config)
39
+
40
+ self.config = config
41
+ self.wav2vec2 = Wav2Vec2Model(config)
42
+ self.classifier = RegressionHead(config)
43
+ self.init_weights()
44
+
45
+ def forward(
46
+ self,
47
+ input_values,
48
+ ):
49
+ outputs = self.wav2vec2(input_values)
50
+ hidden_states = outputs[0]
51
+ hidden_states = torch.mean(hidden_states, dim=1)
52
+ logits = self.classifier(hidden_states)
53
+
54
+ return hidden_states, logits
55
+
56
+
57
+ # load model from hub
58
+ device = 'cuda' if torch.cuda.is_available() else "cpu"
59
+ model_name = 'audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim'
60
+ processor = Wav2Vec2Processor.from_pretrained(model_name)
61
+ model = EmotionModel.from_pretrained(model_name).to(device)
62
+
63
+
64
+ def process_func(
65
+ x: np.ndarray,
66
+ sampling_rate: int,
67
+ embeddings: bool = False,
68
+ ) -> np.ndarray:
69
+ r"""Predict emotions or extract embeddings from raw audio signal."""
70
+
71
+ # run through processor to normalize signal
72
+ # always returns a batch, so we just get the first entry
73
+ # then we put it on the device
74
+ y = processor(x, sampling_rate=sampling_rate)
75
+ y = y['input_values'][0]
76
+ y = torch.from_numpy(y).to(device)
77
+
78
+ # run through model
79
+ with torch.no_grad():
80
+ y = model(y)[0 if embeddings else 1]
81
+
82
+ # convert to numpy
83
+ y = y.detach().cpu().numpy()
84
+
85
+ return y
86
+ #
87
+ #
88
+ # def disp(rootpath, wavname):
89
+ # wav, sr = librosa.load(f"{rootpath}/{wavname}", 16000)
90
+ # display(ipd.Audio(wav, rate=sr))
91
+
92
+ rootpath = "dataset/nene"
93
+ embs = []
94
+ wavnames = []
95
+ def extract_dir(path):
96
+ rootpath = path
97
+ for idx, wavname in enumerate(os.listdir(rootpath)):
98
+ wav, sr =librosa.load(f"{rootpath}/{wavname}", 16000)
99
+ emb = process_func(np.expand_dims(wav, 0), sr, embeddings=True)
100
+ embs.append(emb)
101
+ wavnames.append(wavname)
102
+ np.save(f"{rootpath}/{wavname}.emo.npy", emb.squeeze(0))
103
+ print(idx, wavname)
104
+
105
+ def extract_wav(path):
106
+ wav, sr = librosa.load(path, 16000)
107
+ emb = process_func(np.expand_dims(wav, 0), sr, embeddings=True)
108
+ return emb
109
+
110
+ if __name__ == '__main__':
111
+ for spk in ["serena", "koni", "nyaru","shanoa", "mana"]:
112
+ extract_dir(f"dataset/{spk}")
filelists/train.txt ADDED
The diff for this file is too large to render. See raw diff
 
filelists/train.txt.cleaned ADDED
The diff for this file is too large to render. See raw diff
 
filelists/val.txt ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset/nene/nen001_001.wav|0|はい?呼びました?
2
+ dataset/nene/nen001_002.wav|0|驚かせたならごめんなさい。通りかかったときにちょうど名前が聞こえてきたので
3
+ dataset/nene/nen001_003.wav|0|私に何か用ですか?
4
+ dataset/nene/nen001_004.wav|0|いえ、少し気になっただけですから、別に怒っているわけじゃないです。気にしないで下さい
5
+ dataset/nene/nen001_005.wav|0|それより仮屋さん、例の件ですが――
6
+ dataset/nene/nen001_006.wav|0|喜んでもらえたならなによりです
7
+ dataset/nene/nen001_007.wav|0|また何かあったら、いつでも部室に来て下さい
8
+ dataset/nene/nen001_008.wav|0|大したことじゃないので。それより体調の方はもういいんですか?
9
+ dataset/nene/nen001_009.wav|0|それはよかったです。他に困ったことはありますか?
10
+ dataset/nene/nen001_010.wav|0|何かあったらいつでも話して下さい。学院のことじゃなく、私事に関することでも何でも
11
+ dataset/nene/nen001_011.wav|0|はい、いつでもどうぞ
12
+ dataset/nene/nen001_012.wav|0|保科君も
13
+ dataset/nene/nen001_013.wav|0|もし何か困ったことがあれば、力になりますから
14
+ dataset/nene/nen001_014.wav|0|そうですか?私には、なにか悩み事があるように見えたりしましたけど……
15
+ dataset/nene/nen001_015.wav|0|――なんて、言ってみただけですから、深い意味はありませんよ
16
+ dataset/nene/nen001_016.wav|0|いえ、そういうわけじゃなくって……ただ何となく。そんな気がしただけですから
17
+ dataset/nene/nen001_017.wav|0|あ、いえ、絶対に秘密というわけじゃないので気にしないで下さい
18
+ dataset/nene/nen001_018.wav|0|好奇心だけで来られると困るので、本当に悩んでる人以外には、広めないようにお願いしているだけです
19
+ dataset/nene/nen001_019.wav|0|そんな仰々しい話じゃなく……私は、オカ研に所属しているんです
20
+ dataset/nene/nen001_020.wav|0|はい。そのオカ研です
21
+ dataset/nene/nen001_021.wav|0|そうですよ。実は私たちが入学する前からあった部活なんです。今は部員がいなくて、所属してるのは私一人ですが
22
+ dataset/nene/nen001_022.wav|0|なので最近は、学生会の方からは結構つつかれてて……活動も、発表などを意欲的にしているわけじゃありませんからね
23
+ dataset/nene/nen001_023.wav|0|いえ、私は占いを。オカルトと言っても幅は広いので。それに部には私しかいませんから、結構好きにできるんです
24
+ dataset/nene/nen001_024.wav|0|さすがに白蛇占いはできませんよ
25
+ dataset/nene/nen001_025.wav|0|一応。知ってはいますけど、私にできるのはタロットぐらいです。あくまで趣味程度ですから
26
+ dataset/nene/nen001_026.wav|0|あくまで占いの延長線上のものですから。人生相談なんていうほど大層な物じゃありません
27
+ dataset/nene/nen001_027.wav|0|でも……もし保科君も嫌いでなければ、いつでも部室に来て下さい
28
+ dataset/nene/nen001_028.wav|0|あ、はい。すぐに行きます
29
+ dataset/nene/nen001_029.wav|0|それじゃあ私はこれで
30
+ dataset/nene/nen001_030.wav|0|あの、先生
31
+ dataset/nene/nen001_031.wav|0|はい。気付いたらこんな時間になってしまって
32
+ dataset/nene/nen001_032.wav|0|図書室は……もう誰もいないんですか?
33
+ dataset/nene/nen001_033.wav|0|その前にちょっと調べ物をさせて欲しいんですが、鍵を貸してもらえませんか?
34
+ dataset/nene/nen001_034.wav|0|タロットカードのことで少々。時間はかかりません。5分……は無理かも……でも20分もあれば終わりますから……お願いします
35
+ dataset/nene/nen001_035.wav|0|わかりました。ありがとう……ございます
36
+ dataset/nene/nen001_036.wav|0|は、はい………気をつけます……ハァ、ハァ……
37
+ dataset/nene/nen001_037.wav|0|……ハァ……ハァ……
38
+ dataset/nene/nen001_038.wav|0|……大丈夫、ですよね……んっ……
39
+ dataset/nene/nen001_039.wav|0|……ハァ……ハァ……
40
+ dataset/nene/nen001_040.wav|0|んっ、んん……
41
+ dataset/nene/nen001_041.wav|0|あ……あっ、んっ……んんン、んッ、んッ、んぅぅッ……
42
+ dataset/nene/nen001_042.wav|0|んっ、ふぅぅ……はぁ、はぁ、あっ、ああぁぁ……ふぁぁぁ……
43
+ dataset/nene/nen001_043.wav|0|はぁ、はぁぁ……ん、ん、んっ……はぁぁ……ぁ、ぁ、ぁ……ぁぁ……ンッ……!
44
+ dataset/nene/nen001_044.wav|0|あぁ、もう……こ、んなの、最低ですッ……はぁ、はぁ……学院内で、おっ、オナニー……をするなんてっ、ん、んんっ
45
+ dataset/nene/nen001_045.wav|0|あっ、ああぁぁ……はぁ、はぁ、はぁぁぁ……ん、んンッッ
46
+ dataset/nene/nen001_046.wav|0|はぁ、はぁ……私、��んてこと……あっ、あぁぁ……でも、気持ちよくて止まりません……ん、んんッ、ふぁ、あ、あぁぁぁ……
47
+ dataset/nene/nen001_047.wav|0|うっ……あ、あ、あっ、あっ、ひあっ、ん、んはっ……はっ、はっ、んんぅぅッ
48
+ dataset/nene/nen001_048.wav|0|はぁ……はぁ……んっ、んんん……んぁ、んぁ、あっ……んっ、ぃぃぃッ
49
+ dataset/nene/nen001_049.wav|0|んっ、んっ、んくっ……ひっ、あっ、ぁっ、ぁっ、んんーーッ……
50
+ dataset/nene/nen001_050.wav|0|はぁぁぁ……はぁ、はぁ……んんっ、んっ、んんん、んぁ……んぁ、ぁ、ぁ、ぁぁぁぁ……ッ!
51
+ dataset/nene/nen001_051.wav|0|だめ、早く、しないと……先生が、戻って、きます……こんなところ、見られたら……んっ、んっ、んぁ、んぁ、あ、あ、あ、あ、あ……
52
+ dataset/nene/nen001_052.wav|0|あ、あ、ん、んんんッ……はっ、はぁ、はぁ、あぁぁもう、本当に、最低ですっ……
53
+ dataset/nene/nen001_053.wav|0|んッ……んはぁッ、はぁ、はぁ……ンン、んぁ、んぁぁ……あ、あ、あぁぁ……
54
+ dataset/nene/nen001_054.wav|0|はぁー、はぁーーぁぁ、気持ちいい……あっ、あぃ、あぃっ、いい……本当に……んん
55
+ dataset/nene/nen001_055.wav|0|早く……こんなこと、早く……先生が、また様子を……見に、きたりしない、内に……はぁ、はぁ、はぁ、はぁ
56
+ dataset/nene/nen001_056.wav|0|はぁ、はぁ、はぁっ、はぁぁぁ……ぁ、ぁ、ぁ……ぁぁ……ぁぁんッ、んっ、んーッ
57
+ dataset/nene/nen001_057.wav|0|んんッ、ん、んはぁぁぁーー……はぁ、はぁ、はぁァン、あッ、あん、あぁぁンッ
58
+ dataset/nene/nen001_058.wav|0|はぁ、はぁ、はぁぁ……ヨダレ、でちゃう……じゅる……んぁっ、はぁ、はぁぁ……はっ、はっ、はっ……じゅる
59
+ dataset/nene/nen001_059.wav|0|こんな、に、気持ちいいなんて、最低です……本当、図書室でオナニーなんて、最低すぎます……でも、でも
60
+ dataset/nene/nen001_060.wav|0|あああぁぁ……今は、止められなくて……じゅる……はぁ、はぁぁ……あぁぁぁあぁ……
61
+ dataset/nene/nen001_061.wav|0|はぁ、はぁ、はぁ……早く……あっ、あっ、あぁぁ……ぁぁぁ、早く、早く
62
+ dataset/nene/nen001_062.wav|0|早く、イきたい、イきたい……このままイきたいぃ……はぁ、はぁ、はぁ、ぁぁ、ぁっ、ぁっ、あっ、あっ、あぁッ!
63
+ dataset/nene/nen001_063.wav|0|あっ、あっ、ああぁぁ……痺れて、きたぁ……だめ、イきそう……んっ、だめじゃなくて、あっ、あっ、早く、このまま……早く早く早く
filelists/val.txt.cleaned ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dataset/nene/nen001_001.wav|0|ha↓i? yo↑bima↓ʃIta?
2
+ dataset/nene/nen001_002.wav|0|o↑doro↓kasetanara go↑meNnasa↓i. to↑orikaka↓Qta to↓kini ʧo↑odo na↑maega kI↑koete ki↓tanode.
3
+ dataset/nene/nen001_003.wav|0|wa↑taʃini na↑nika↓yoodesUka?
4
+ dataset/nene/nen001_004.wav|0|i↓e, sU↑ko↓ʃI ki↑ni na↓QtadakedesUkara, be↑tsuni o↑ko↓Qte i↑ru wa↓kejanaidesU. ki↑ni ʃi↑na↓ide ku↑dasa↓i.
5
+ dataset/nene/nen001_005.wav|0|so↑reyo↓ri ka↑riyasaN, re↓eno ke↓NdesUga----
6
+ dataset/nene/nen001_006.wav|0|yo↑roko↓Nde mo↑raeta↓nara na↓niyoridesU.
7
+ dataset/nene/nen001_007.wav|0|ma↑ta na↓nika a↓Qtara, i↓tsudemo bu↑ʃItsuni ki↓te ku↑dasa↓i.
8
+ dataset/nene/nen001_008.wav|0|ta↓iʃIta ko↑to↓janainode. so↑reyo↓ri ta↑iʧoono ho↓owa mo↓o i↓i N↓desUka?
9
+ dataset/nene/nen001_009.wav|0|so↑rewa yo↓kaQtadesU. ta↓ni ko↑ma↓Qta ko↑to↓wa a↑rima↓sUka?
10
+ dataset/nene/nen001_010.wav|0|na↓nika a↓Qtara i↓tsudemo ha↑na↓ʃIte ku↑dasa↓i. ga↑kuiNno ko↑to↓janaku, ʃi↓jini ka↑Nsu↓ru ko↑to↓demo na↓nidemo.
11
+ dataset/nene/nen001_011.wav|0|ha↓i, i↓tsudemo do↓ozo.
12
+ dataset/nene/nen001_012.wav|0|ho↓ʃinakuNmo.
13
+ dataset/nene/nen001_013.wav|0|mo↓ʃi na↓nika ko↑ma↓Qta ko↑to↓ga a↑re↓ba, ʧI↑kara↓ni na↑rima↓sUkara.
14
+ dataset/nene/nen001_014.wav|0|so↑odesU↓ka? wa↑taʃiniwa, na↓nika na↑yami↓gotoga a↓ru yo↓oni mi↑eta↓ri ʃi↑ma↓ʃItakedo......
15
+ dataset/nene/nen001_015.wav|0|---- na↓Nte, i↑Qte mi↓tadakedesUkara, fU↑ka↓i i↑miwaarimase↓Nyo.
16
+ dataset/nene/nen001_016.wav|0|i↓e, so↑oyuu wa↓kejanakuQte...... ta↓da na↑Ntona↓ku. so↑Nna ki↑ga ʃI↑ta↓dakedesUkara.
17
+ dataset/nene/nen001_017.wav|0|a, i↓e, ze↑Qtaini hi↑mitsUto i↑u wa↓kejanainode ki↑ni ʃi↑na↓ide ku↑dasa↓i.
18
+ dataset/nene/nen001_018.wav|0|ko↑okIʃiNdakede ko↑rare↓ruto ko↑ma↓runode, ho↑Ntooni na↑ya↓Nde ru↑niNi↓gainiwa, hi↑romenai yo↓oni o↑negai ʃI↑te i↑rudakedesU.
19
+ dataset/nene/nen001_019.wav|0|so↑Nna gyo↑ogyooʃi↓i ha↑naʃi↓janaku...... wa↑taʃiwa, o↑kakeNni ʃo↑zoku ʃI↑te i↑ru N↓desU.
20
+ dataset/nene/nen001_020.wav|0|ha↓i. so↑no o↑kakeNde↓sU.
21
+ dataset/nene/nen001_021.wav|0|so↑odesUyo. ji↑tsu↓wa wa↑taʃi↓taʧiga nyu↑ugakU su↑ru ma↓ekara a↓Qta bu↑katsuna N↓desU. i↓mawa bu↑iNga i↑nakUte, ʃo↑zoku ʃI↑te↓ru no↑wa wa↑taʃI hi↑to↓ridesUga.
22
+ dataset/nene/nen001_022.wav|0|na↑node sa↑ikiNwa, ga↑kUsee↓kaino ho↓okarawa ke↓Qkoo tsU↑tsu↓karetete...... ka↑tsudoomo, ha↑Qpyoona↓doo i↑yoku↓tekini ʃI↑te i↑ru wa↑kejaarimase↓Nkarane.
23
+ dataset/nene/nen001_023.wav|0|i↓e, wa↑taʃiwa u↑ranaio. o↑karutoto i↑Qtemo ha↑bawa hi↑ro↓inode. so↑reni bu↓niwa wa↑taʃIʃi↓ka i↑mase↓Nkara, ke↓Qkoo sU↑ki↓ni de↑ki↓ru N↓desU.
24
+ dataset/nene/nen001_024.wav|0|sa↑sugani ʃi↑rohebiu↓ranaiwa de↑kimase↓Nyo.
25
+ dataset/nene/nen001_025.wav|0|i↑ʧioo. ʃi↑Qte ha↑ima↓sUkedo, wa↑taʃini de↑ki↓ru no↑wa ta↑roQtogu↓raidesU. a↑ku↓made ʃu↑mite↓edodesUkara.
26
+ dataset/nene/nen001_026.wav|0|a↑ku↓made u↑ranaino e↑NʧooseNjoono mo↑no↓desUkara. ji↑Nseeso↓odaNnaNte i↑uhodo ta↓isoona mo↑nojaarimase↓N.
27
+ dataset/nene/nen001_027.wav|0|de↓mo...... mo↓ʃI ho↓ʃinakuNmo ki↑raidenakere↓ba, i↓tsudemo bu↑ʃItsuni ki↓te ku↑dasa↓i.
28
+ dataset/nene/nen001_028.wav|0|a, ha↓i. su↓guni i↑kima↓sU.
29
+ dataset/nene/nen001_029.wav|0|so↑reja↓a wa↑taʃiwa ko↑rede.
30
+ dataset/nene/nen001_030.wav|0|a↑no, se↑Nse↓e.
31
+ dataset/nene/nen001_031.wav|0|ha↓i. ki↑zu↓itara ko↑Nna ji↑kaNni na↓Qte ʃi↑ma↓Qte.
32
+ dataset/nene/nen001_032.wav|0|to↑ʃo↓ʃItsuwa...... mo↓o da↓remo i↑nai N↓desUka?
33
+ dataset/nene/nen001_033.wav|0|so↑no ma↓eni ʧo↓Qto ʃi↑rabebutsuo sa↑setehoʃii N↓desUga, ka↑gi↓o ka↑ʃIte mo↑raemase↓Nka?
34
+ dataset/nene/nen001_034.wav|0|ta↑roQtoka↓adono ko↑to↓de ʃo↓oʃoo. ji↑kaNwa ka↑karimase↓N. go↓fuN...... w a mu↓rikamo...... de↓mo ni↑juQ↓puNmo a↑re↓ba o↑warima↓sUkara...... o↑negai ʃi↑ma↓sU.
35
+ dataset/nene/nen001_035.wav|0|wa↑karima↓ʃIta. a↑ri↓gatoo...... go↑zaima↓sU.
36
+ dataset/nene/nen001_036.wav|0|w a, ha↓i......... ki↑o tsU↑kema↓sU...... ha↓a, ha↓a......
37
+ dataset/nene/nen001_037.wav|0|...... ha↓a...... ha↓a......
38
+ dataset/nene/nen001_038.wav|0|...... da↑ijo↓obu, de↓sUyone...... N↓Q......
39
+ dataset/nene/nen001_039.wav|0|...... ha↓a...... ha↓a......
40
+ dataset/nene/nen001_040.wav|0|N↓Q, N↓N......
41
+ dataset/nene/nen001_041.wav|0|a...... a↓Q, N↓Q...... N↓N N, N↓Q, N↓Q, N↓uuQ......
42
+ dataset/nene/nen001_042.wav|0|N↓Q, fu↓uu...... ha↓a, ha↓a, a↓Q, a↓aaa...... fa↓aa......
43
+ dataset/nene/nen001_043.wav|0|ha↓a, ha↓aa...... N, N, N↓Q...... ha↓aa...... a, a, a...... a↓a...... N↓Q......!
44
+ dataset/nene/nen001_044.wav|0|a↓a, mo↓o...... k o, N↑na n o, sa↑iteede su↓Q...... ha↓a, ha↓a...... ga↑kuiN↓naide, o↑Q, o↓nanii...... o su↑ru↓naNteQ, N, N↓NQ.
45
+ dataset/nene/nen001_045.wav|0|a↓Q, a↓aaa...... ha↓a, ha↓a, ha↓aaa...... N, N↓NQQ.
46
+ dataset/nene/nen001_046.wav|0|ha↓a, ha↓a...... wa↑taʃi, na↓Nte ko↑to...... a↓Q, a↓aa...... de↓mo, ki↑moʧiyo↓kUte to↑marimase↓N...... N, N↓NQ, f a, a, a↓aaa......
47
+ dataset/nene/nen001_047.wav|0|u↓Q...... a, a, a↓Q, a↓Q, hi↓aQ, N, N↓haQ...... ha↓Q, ha↓Q, N↓NuuQ.
48
+ dataset/nene/nen001_048.wav|0|ha↓a...... ha↓a...... N↓Q, N↓NN...... N↓a, N↓a, a↓Q...... N↓Q, i↓iiQ.
49
+ dataset/nene/nen001_049.wav|0|N↓Q, N↓Q, N ku↓Q...... hi↓Q, a↓Q, a↓Q, a↓Q, N↓NNNQ......
50
+ dataset/nene/nen001_050.wav|0|ha↓aaa...... ha↓a, ha↓a...... N↓NQ, N↓Q, N↓NN, N↓a...... N↓a, a, a, a↓aaa...... Q!
51
+ dataset/nene/nen001_051.wav|0|da↑me, ha↓yaku, ʃi↑naito...... se↑Nse↓ega, mo↑do↓Qte, ki↑ma↓sU...... ko↑Nna to↑koro, mi↓raretara...... N↓Q, N↓Q, N↓a, N↓a, a, a, a, a, a......
52
+ dataset/nene/nen001_052.wav|0|a, a, N, N↓NN Q...... ha↓Q, ha↓a, ha↓a, a↓aa mo↓o, ho↑Ntooni, sa↑iteede su↓Q......
53
+ dataset/nene/nen001_053.wav|0|N↓Q...... N↓haaQ, ha↓a, ha↓a...... N↓N, N↓a, N↓aa...... a, a, a↓aa......
54
+ dataset/nene/nen001_054.wav|0|ha↓aa, ha↓aaaaa, ki↑moʧii↓i...... a↓Q, a↓i, a↓iQ, i↓i...... ho↑Ntooni...... N↓N.
55
+ dataset/nene/nen001_055.wav|0|ha↓yaku...... ko↑Nna ko↑to, ha↓yaku...... se↑Nse↓ega, ma↑ta yo↑osuo...... mi↑ni, kI↑tariʃi↓nai, u↑ʧini...... ha↓a, ha↓a, ha↓a, ha↓a.
56
+ dataset/nene/nen001_056.wav|0|ha↓a, ha↓a, ha↓aQ, ha↓aaa...... a, a, a...... a↓a...... a↓aNQ, N↓Q, N↓NQ.
57
+ dataset/nene/nen001_057.wav|0|N↓NQ, N, N↓haaaaaa...... ha↓a, ha↓a, ha↓aaN, a↓Q, a↑N, a↓aaNQ.
58
+ dataset/nene/nen001_058.wav|0|ha↓a, ha↓a, ha↓aa...... yo↑dare, de↑ʧau...... ju↑ru...... N↓aQ, ha↓a, ha↓aa...... ha↓Q, ha↓Q, ha↓Q...... ju↑ru.
59
+ dataset/nene/nen001_059.wav|0|ko↑Nna, n i, ki↑moʧii↓inaNte, sa↑iteede↓sU...... ho↑Ntoo, to↑ʃo↓ʃItsude o↓naniinaNte, sa↑itee su↑gima↓sU...... de↓mo, de↓mo.
60
+ dataset/nene/nen001_060.wav|0|a↑a↓aaa...... i↓mawa, to↑merarenakUte...... ju↑ru...... ha↓a, ha↓aa...... a↑aaa↓aa......
61
+ dataset/nene/nen001_061.wav|0|ha↓a, ha↓a, ha↓a...... ha↓yaku...... a↓Q, a↓Q, a↓aa...... a↓aa, ha↓yaku, ha↓yaku.
62
+ dataset/nene/nen001_062.wav|0|ha↓yaku, i↑ki↓tai, i↑ki↓tai...... ko↑no ma↑maiki↓taii...... ha↓a, ha↓a, ha↓a, a↓a, a↓Q, a↓Q, a↓Q, a↓Q, a↓aQ!
63
+ dataset/nene/nen001_063.wav|0|a↓Q, a↓Q, a↓aaa...... ʃi↑bire↓te, ki↓ta a...... da↑me, i↓kIsou...... N↓Q, da↑me↓janakUte, a↓Q, a↓Q, ha↓yaku, ko↑no ma↑ma...... ha↓yakUhayakUhayaku.
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
+
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 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
10
+ import logging
11
+
12
+ numba_logger = logging.getLogger('numba')
13
+ numba_logger.setLevel(logging.WARNING)
14
+ import warnings
15
+ warnings.filterwarnings('ignore')
16
+ import librosa
17
+ import librosa.util as librosa_util
18
+ from librosa.util import normalize, pad_center, tiny
19
+ from scipy.signal import get_window
20
+ from scipy.io.wavfile import read
21
+ from librosa.filters import mel as librosa_mel_fn
22
+
23
+ MAX_WAV_VALUE = 32768.0
24
+
25
+
26
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
27
+ """
28
+ PARAMS
29
+ ------
30
+ C: compression factor
31
+ """
32
+ return torch.log(torch.clamp(x, min=clip_val) * C)
33
+
34
+
35
+ def dynamic_range_decompression_torch(x, C=1):
36
+ """
37
+ PARAMS
38
+ ------
39
+ C: compression factor used to compress
40
+ """
41
+ return torch.exp(x) / C
42
+
43
+
44
+ def spectral_normalize_torch(magnitudes):
45
+ output = dynamic_range_compression_torch(magnitudes)
46
+ return output
47
+
48
+
49
+ def spectral_de_normalize_torch(magnitudes):
50
+ output = dynamic_range_decompression_torch(magnitudes)
51
+ return output
52
+
53
+
54
+ mel_basis = {}
55
+ hann_window = {}
56
+
57
+
58
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
59
+ if torch.min(y) < -1.:
60
+ print('min value is ', torch.min(y))
61
+ if torch.max(y) > 1.:
62
+ print('max value is ', torch.max(y))
63
+
64
+ global hann_window
65
+ dtype_device = str(y.dtype) + '_' + str(y.device)
66
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
67
+ if wnsize_dtype_device not in hann_window:
68
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
69
+
70
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
71
+ y = y.squeeze(1)
72
+
73
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
74
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
75
+
76
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
77
+ return spec
78
+
79
+
80
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
81
+ global mel_basis
82
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
83
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
84
+ if fmax_dtype_device not in mel_basis:
85
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
86
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
87
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
88
+ spec = spectral_normalize_torch(spec)
89
+ return spec
90
+
91
+
92
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
93
+ if torch.min(y) < -1.:
94
+ print('min value is ', torch.min(y))
95
+ if torch.max(y) > 1.:
96
+ print('max value is ', torch.max(y))
97
+
98
+ global mel_basis, hann_window
99
+ dtype_device = str(y.dtype) + '_' + str(y.device)
100
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
101
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
102
+ if fmax_dtype_device not in mel_basis:
103
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
104
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
105
+ if wnsize_dtype_device not in hann_window:
106
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
107
+
108
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
109
+ y = y.squeeze(1)
110
+
111
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
112
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
113
+
114
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
115
+
116
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
117
+ spec = spectral_normalize_torch(spec)
118
+
119
+ return spec
models.py ADDED
@@ -0,0 +1,537 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 commons
8
+ import modules
9
+ import attentions
10
+ # import monotonic_align
11
+
12
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
13
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
14
+ from commons import init_weights, get_padding
15
+
16
+
17
+ class StochasticDurationPredictor(nn.Module):
18
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
19
+ super().__init__()
20
+ filter_channels = in_channels # it needs to be removed from future version.
21
+ self.in_channels = in_channels
22
+ self.filter_channels = filter_channels
23
+ self.kernel_size = kernel_size
24
+ self.p_dropout = p_dropout
25
+ self.n_flows = n_flows
26
+ self.gin_channels = gin_channels
27
+
28
+ self.log_flow = modules.Log()
29
+ self.flows = nn.ModuleList()
30
+ self.flows.append(modules.ElementwiseAffine(2))
31
+ for i in range(n_flows):
32
+ self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
33
+ self.flows.append(modules.Flip())
34
+
35
+ self.post_pre = nn.Conv1d(1, filter_channels, 1)
36
+ self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
37
+ self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
38
+ self.post_flows = nn.ModuleList()
39
+ self.post_flows.append(modules.ElementwiseAffine(2))
40
+ for i in range(4):
41
+ self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
42
+ self.post_flows.append(modules.Flip())
43
+
44
+ self.pre = nn.Conv1d(in_channels, filter_channels, 1)
45
+ self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
46
+ self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
47
+ if gin_channels != 0:
48
+ self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
49
+
50
+ def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
51
+ x = torch.detach(x)
52
+ x = self.pre(x)
53
+ if g is not None:
54
+ g = torch.detach(g)
55
+ x = x + self.cond(g)
56
+ x = self.convs(x, x_mask)
57
+ x = self.proj(x) * x_mask
58
+
59
+ if not reverse:
60
+ flows = self.flows
61
+ assert w is not None
62
+
63
+ logdet_tot_q = 0
64
+ h_w = self.post_pre(w)
65
+ h_w = self.post_convs(h_w, x_mask)
66
+ h_w = self.post_proj(h_w) * x_mask
67
+ e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
68
+ z_q = e_q
69
+ for flow in self.post_flows:
70
+ z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
71
+ logdet_tot_q += logdet_q
72
+ z_u, z1 = torch.split(z_q, [1, 1], 1)
73
+ u = torch.sigmoid(z_u) * x_mask
74
+ z0 = (w - u) * x_mask
75
+ logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
76
+ logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
77
+
78
+ logdet_tot = 0
79
+ z0, logdet = self.log_flow(z0, x_mask)
80
+ logdet_tot += logdet
81
+ z = torch.cat([z0, z1], 1)
82
+ for flow in flows:
83
+ z, logdet = flow(z, x_mask, g=x, reverse=reverse)
84
+ logdet_tot = logdet_tot + logdet
85
+ nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
86
+ return nll + logq # [b]
87
+ else:
88
+ flows = list(reversed(self.flows))
89
+ flows = flows[:-2] + [flows[-1]] # remove a useless vflow
90
+ z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
91
+ for flow in flows:
92
+ z = flow(z, x_mask, g=x, reverse=reverse)
93
+ z0, z1 = torch.split(z, [1, 1], 1)
94
+ logw = z0
95
+ return logw
96
+
97
+
98
+ class DurationPredictor(nn.Module):
99
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
100
+ super().__init__()
101
+
102
+ self.in_channels = in_channels
103
+ self.filter_channels = filter_channels
104
+ self.kernel_size = kernel_size
105
+ self.p_dropout = p_dropout
106
+ self.gin_channels = gin_channels
107
+
108
+ self.drop = nn.Dropout(p_dropout)
109
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
110
+ self.norm_1 = modules.LayerNorm(filter_channels)
111
+ self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
112
+ self.norm_2 = modules.LayerNorm(filter_channels)
113
+ self.proj = nn.Conv1d(filter_channels, 1, 1)
114
+
115
+ if gin_channels != 0:
116
+ self.cond = nn.Conv1d(gin_channels, in_channels, 1)
117
+
118
+ def forward(self, x, x_mask, g=None):
119
+ x = torch.detach(x)
120
+ if g is not None:
121
+ g = torch.detach(g)
122
+ x = x + self.cond(g)
123
+ x = self.conv_1(x * x_mask)
124
+ x = torch.relu(x)
125
+ x = self.norm_1(x)
126
+ x = self.drop(x)
127
+ x = self.conv_2(x * x_mask)
128
+ x = torch.relu(x)
129
+ x = self.norm_2(x)
130
+ x = self.drop(x)
131
+ x = self.proj(x * x_mask)
132
+ return x * x_mask
133
+
134
+
135
+ class TextEncoder(nn.Module):
136
+ def __init__(self,
137
+ n_vocab,
138
+ out_channels,
139
+ hidden_channels,
140
+ filter_channels,
141
+ n_heads,
142
+ n_layers,
143
+ kernel_size,
144
+ p_dropout):
145
+ super().__init__()
146
+ self.n_vocab = n_vocab
147
+ self.out_channels = out_channels
148
+ self.hidden_channels = hidden_channels
149
+ self.filter_channels = filter_channels
150
+ self.n_heads = n_heads
151
+ self.n_layers = n_layers
152
+ self.kernel_size = kernel_size
153
+ self.p_dropout = p_dropout
154
+
155
+ self.emb = nn.Embedding(n_vocab, hidden_channels)
156
+ self.emo_proj = nn.Linear(1024, hidden_channels)
157
+
158
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
159
+
160
+ self.encoder = attentions.Encoder(
161
+ hidden_channels,
162
+ filter_channels,
163
+ n_heads,
164
+ n_layers,
165
+ kernel_size,
166
+ p_dropout)
167
+ self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
168
+
169
+ def forward(self, x, x_lengths, emo):
170
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
171
+ x = x + self.emo_proj(emo.unsqueeze(1))
172
+ x = torch.transpose(x, 1, -1) # [b, h, t]
173
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
174
+
175
+ x = self.encoder(x * x_mask, x_mask)
176
+ stats = self.proj(x) * x_mask
177
+
178
+ m, logs = torch.split(stats, self.out_channels, dim=1)
179
+ return x, m, logs, x_mask
180
+
181
+
182
+ class ResidualCouplingBlock(nn.Module):
183
+ def __init__(self,
184
+ channels,
185
+ hidden_channels,
186
+ kernel_size,
187
+ dilation_rate,
188
+ n_layers,
189
+ n_flows=4,
190
+ gin_channels=0):
191
+ super().__init__()
192
+ self.channels = channels
193
+ self.hidden_channels = hidden_channels
194
+ self.kernel_size = kernel_size
195
+ self.dilation_rate = dilation_rate
196
+ self.n_layers = n_layers
197
+ self.n_flows = n_flows
198
+ self.gin_channels = gin_channels
199
+
200
+ self.flows = nn.ModuleList()
201
+ for i in range(n_flows):
202
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
203
+ self.flows.append(modules.Flip())
204
+
205
+ def forward(self, x, x_mask, g=None, reverse=False):
206
+ if not reverse:
207
+ for flow in self.flows:
208
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
209
+ else:
210
+ for flow in reversed(self.flows):
211
+ x = flow(x, x_mask, g=g, reverse=reverse)
212
+ return x
213
+
214
+
215
+ class PosteriorEncoder(nn.Module):
216
+ def __init__(self,
217
+ in_channels,
218
+ out_channels,
219
+ hidden_channels,
220
+ kernel_size,
221
+ dilation_rate,
222
+ n_layers,
223
+ gin_channels=0):
224
+ super().__init__()
225
+ self.in_channels = in_channels
226
+ self.out_channels = out_channels
227
+ self.hidden_channels = hidden_channels
228
+ self.kernel_size = kernel_size
229
+ self.dilation_rate = dilation_rate
230
+ self.n_layers = n_layers
231
+ self.gin_channels = gin_channels
232
+
233
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
234
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
235
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
236
+
237
+ def forward(self, x, x_lengths, g=None):
238
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
239
+ x = self.pre(x) * x_mask
240
+ x = self.enc(x, x_mask, g=g)
241
+ stats = self.proj(x) * x_mask
242
+ m, logs = torch.split(stats, self.out_channels, dim=1)
243
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
244
+ return z, m, logs, x_mask
245
+
246
+
247
+ class Generator(torch.nn.Module):
248
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
249
+ super(Generator, self).__init__()
250
+ self.num_kernels = len(resblock_kernel_sizes)
251
+ self.num_upsamples = len(upsample_rates)
252
+ self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
253
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
254
+
255
+ self.ups = nn.ModuleList()
256
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
257
+ self.ups.append(weight_norm(
258
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
259
+ k, u, padding=(k-u)//2)))
260
+
261
+ self.resblocks = nn.ModuleList()
262
+ for i in range(len(self.ups)):
263
+ ch = upsample_initial_channel//(2**(i+1))
264
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
265
+ self.resblocks.append(resblock(ch, k, d))
266
+
267
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
268
+ self.ups.apply(init_weights)
269
+
270
+ if gin_channels != 0:
271
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
272
+
273
+ def forward(self, x, g=None):
274
+ x = self.conv_pre(x)
275
+ if g is not None:
276
+ x = x + self.cond(g)
277
+
278
+ for i in range(self.num_upsamples):
279
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
280
+ x = self.ups[i](x)
281
+ xs = None
282
+ for j in range(self.num_kernels):
283
+ if xs is None:
284
+ xs = self.resblocks[i*self.num_kernels+j](x)
285
+ else:
286
+ xs += self.resblocks[i*self.num_kernels+j](x)
287
+ x = xs / self.num_kernels
288
+ x = F.leaky_relu(x)
289
+ x = self.conv_post(x)
290
+ x = torch.tanh(x)
291
+
292
+ return x
293
+
294
+ def remove_weight_norm(self):
295
+ print('Removing weight norm...')
296
+ for l in self.ups:
297
+ remove_weight_norm(l)
298
+ for l in self.resblocks:
299
+ l.remove_weight_norm()
300
+
301
+
302
+ class DiscriminatorP(torch.nn.Module):
303
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
304
+ super(DiscriminatorP, self).__init__()
305
+ self.period = period
306
+ self.use_spectral_norm = use_spectral_norm
307
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
308
+ self.convs = nn.ModuleList([
309
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
310
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
311
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
312
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
313
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
314
+ ])
315
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
316
+
317
+ def forward(self, x):
318
+ fmap = []
319
+
320
+ # 1d to 2d
321
+ b, c, t = x.shape
322
+ if t % self.period != 0: # pad first
323
+ n_pad = self.period - (t % self.period)
324
+ x = F.pad(x, (0, n_pad), "reflect")
325
+ t = t + n_pad
326
+ x = x.view(b, c, t // self.period, self.period)
327
+
328
+ for l in self.convs:
329
+ x = l(x)
330
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
331
+ fmap.append(x)
332
+ x = self.conv_post(x)
333
+ fmap.append(x)
334
+ x = torch.flatten(x, 1, -1)
335
+
336
+ return x, fmap
337
+
338
+
339
+ class DiscriminatorS(torch.nn.Module):
340
+ def __init__(self, use_spectral_norm=False):
341
+ super(DiscriminatorS, self).__init__()
342
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
343
+ self.convs = nn.ModuleList([
344
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
345
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
346
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
347
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
348
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
349
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
350
+ ])
351
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
352
+
353
+ def forward(self, x):
354
+ fmap = []
355
+
356
+ for l in self.convs:
357
+ x = l(x)
358
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
359
+ fmap.append(x)
360
+ x = self.conv_post(x)
361
+ fmap.append(x)
362
+ x = torch.flatten(x, 1, -1)
363
+
364
+ return x, fmap
365
+
366
+
367
+ class MultiPeriodDiscriminator(torch.nn.Module):
368
+ def __init__(self, use_spectral_norm=False):
369
+ super(MultiPeriodDiscriminator, self).__init__()
370
+ periods = [2,3,5,7,11]
371
+
372
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
373
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
374
+ self.discriminators = nn.ModuleList(discs)
375
+
376
+ def forward(self, y, y_hat):
377
+ y_d_rs = []
378
+ y_d_gs = []
379
+ fmap_rs = []
380
+ fmap_gs = []
381
+ for i, d in enumerate(self.discriminators):
382
+ y_d_r, fmap_r = d(y)
383
+ y_d_g, fmap_g = d(y_hat)
384
+ y_d_rs.append(y_d_r)
385
+ y_d_gs.append(y_d_g)
386
+ fmap_rs.append(fmap_r)
387
+ fmap_gs.append(fmap_g)
388
+
389
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
390
+
391
+
392
+
393
+ class SynthesizerTrn(nn.Module):
394
+ """
395
+ Synthesizer for Training
396
+ """
397
+
398
+ def __init__(self,
399
+ n_vocab,
400
+ spec_channels,
401
+ segment_size,
402
+ inter_channels,
403
+ hidden_channels,
404
+ filter_channels,
405
+ n_heads,
406
+ n_layers,
407
+ kernel_size,
408
+ p_dropout,
409
+ resblock,
410
+ resblock_kernel_sizes,
411
+ resblock_dilation_sizes,
412
+ upsample_rates,
413
+ upsample_initial_channel,
414
+ upsample_kernel_sizes,
415
+ n_speakers=0,
416
+ gin_channels=0,
417
+ use_sdp=True,
418
+ **kwargs):
419
+
420
+ super().__init__()
421
+ self.n_vocab = n_vocab
422
+ self.spec_channels = spec_channels
423
+ self.inter_channels = inter_channels
424
+ self.hidden_channels = hidden_channels
425
+ self.filter_channels = filter_channels
426
+ self.n_heads = n_heads
427
+ self.n_layers = n_layers
428
+ self.kernel_size = kernel_size
429
+ self.p_dropout = p_dropout
430
+ self.resblock = resblock
431
+ self.resblock_kernel_sizes = resblock_kernel_sizes
432
+ self.resblock_dilation_sizes = resblock_dilation_sizes
433
+ self.upsample_rates = upsample_rates
434
+ self.upsample_initial_channel = upsample_initial_channel
435
+ self.upsample_kernel_sizes = upsample_kernel_sizes
436
+ self.segment_size = segment_size
437
+ self.n_speakers = n_speakers
438
+ self.gin_channels = gin_channels
439
+
440
+ self.use_sdp = use_sdp
441
+
442
+ self.enc_p = TextEncoder(n_vocab,
443
+ inter_channels,
444
+ hidden_channels,
445
+ filter_channels,
446
+ n_heads,
447
+ n_layers,
448
+ kernel_size,
449
+ p_dropout)
450
+ self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
451
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
452
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
453
+
454
+ if use_sdp:
455
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
456
+ else:
457
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
458
+
459
+ if n_speakers > 1:
460
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
461
+
462
+ def forward(self, x, x_lengths, y, y_lengths, sid=None, emo=None):
463
+
464
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, emo)
465
+ if self.n_speakers > 0:
466
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
467
+ else:
468
+ g = None
469
+
470
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
471
+ z_p = self.flow(z, y_mask, g=g)
472
+
473
+ with torch.no_grad():
474
+ # negative cross-entropy
475
+ s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
476
+ neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
477
+ neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2), s_p_sq_r) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
478
+ neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r)) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
479
+ neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
480
+ neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
481
+
482
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
483
+ attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
484
+
485
+ w = attn.sum(2)
486
+ if self.use_sdp:
487
+ l_length = self.dp(x, x_mask, w, g=g)
488
+ l_length = l_length / torch.sum(x_mask)
489
+ else:
490
+ logw_ = torch.log(w + 1e-6) * x_mask
491
+ logw = self.dp(x, x_mask, g=g)
492
+ l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
493
+
494
+ # expand prior
495
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
496
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
497
+
498
+ z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
499
+ o = self.dec(z_slice, g=g)
500
+ return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
501
+
502
+ def infer(self, x, x_lengths, sid=None, emo=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
503
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths,emo)
504
+ if self.n_speakers > 0:
505
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
506
+ else:
507
+ g = None
508
+
509
+ if self.use_sdp:
510
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
511
+ else:
512
+ logw = self.dp(x, x_mask, g=g)
513
+ w = torch.exp(logw) * x_mask * length_scale
514
+ w_ceil = torch.ceil(w)
515
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
516
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
517
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
518
+ attn = commons.generate_path(w_ceil, attn_mask)
519
+
520
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
521
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
522
+
523
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
524
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
525
+ o = self.dec((z * y_mask)[:,:,:max_len], g=g)
526
+ return o, attn, y_mask, (z, z_p, m_p, logs_p)
527
+
528
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
529
+ assert self.n_speakers > 0, "n_speakers have to be larger than 0."
530
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
531
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
532
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
533
+ z_p = self.flow(z, y_mask, g=g_src)
534
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
535
+ o_hat = self.dec(z_hat * y_mask, g=g_tgt)
536
+ return o_hat, y_mask, (z, z_p, z_hat)
537
+
modules.py ADDED
@@ -0,0 +1,390 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from transforms import piecewise_rational_quadratic_transform
15
+
16
+
17
+ LRELU_SLOPE = 0.1
18
+
19
+
20
+ class LayerNorm(nn.Module):
21
+ def __init__(self, channels, eps=1e-5):
22
+ super().__init__()
23
+ self.channels = channels
24
+ self.eps = eps
25
+
26
+ self.gamma = nn.Parameter(torch.ones(channels))
27
+ self.beta = nn.Parameter(torch.zeros(channels))
28
+
29
+ def forward(self, x):
30
+ x = x.transpose(1, -1)
31
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
32
+ return x.transpose(1, -1)
33
+
34
+
35
+ class ConvReluNorm(nn.Module):
36
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
37
+ super().__init__()
38
+ self.in_channels = in_channels
39
+ self.hidden_channels = hidden_channels
40
+ self.out_channels = out_channels
41
+ self.kernel_size = kernel_size
42
+ self.n_layers = n_layers
43
+ self.p_dropout = p_dropout
44
+ assert n_layers > 1, "Number of layers should be larger than 0."
45
+
46
+ self.conv_layers = nn.ModuleList()
47
+ self.norm_layers = nn.ModuleList()
48
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
49
+ self.norm_layers.append(LayerNorm(hidden_channels))
50
+ self.relu_drop = nn.Sequential(
51
+ nn.ReLU(),
52
+ nn.Dropout(p_dropout))
53
+ for _ in range(n_layers-1):
54
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
55
+ self.norm_layers.append(LayerNorm(hidden_channels))
56
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
57
+ self.proj.weight.data.zero_()
58
+ self.proj.bias.data.zero_()
59
+
60
+ def forward(self, x, x_mask):
61
+ x_org = x
62
+ for i in range(self.n_layers):
63
+ x = self.conv_layers[i](x * x_mask)
64
+ x = self.norm_layers[i](x)
65
+ x = self.relu_drop(x)
66
+ x = x_org + self.proj(x)
67
+ return x * x_mask
68
+
69
+
70
+ class DDSConv(nn.Module):
71
+ """
72
+ Dialted and Depth-Separable Convolution
73
+ """
74
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
75
+ super().__init__()
76
+ self.channels = channels
77
+ self.kernel_size = kernel_size
78
+ self.n_layers = n_layers
79
+ self.p_dropout = p_dropout
80
+
81
+ self.drop = nn.Dropout(p_dropout)
82
+ self.convs_sep = nn.ModuleList()
83
+ self.convs_1x1 = nn.ModuleList()
84
+ self.norms_1 = nn.ModuleList()
85
+ self.norms_2 = nn.ModuleList()
86
+ for i in range(n_layers):
87
+ dilation = kernel_size ** i
88
+ padding = (kernel_size * dilation - dilation) // 2
89
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
90
+ groups=channels, dilation=dilation, padding=padding
91
+ ))
92
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
93
+ self.norms_1.append(LayerNorm(channels))
94
+ self.norms_2.append(LayerNorm(channels))
95
+
96
+ def forward(self, x, x_mask, g=None):
97
+ if g is not None:
98
+ x = x + g
99
+ for i in range(self.n_layers):
100
+ y = self.convs_sep[i](x * x_mask)
101
+ y = self.norms_1[i](y)
102
+ y = F.gelu(y)
103
+ y = self.convs_1x1[i](y)
104
+ y = self.norms_2[i](y)
105
+ y = F.gelu(y)
106
+ y = self.drop(y)
107
+ x = x + y
108
+ return x * x_mask
109
+
110
+
111
+ class WN(torch.nn.Module):
112
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
113
+ super(WN, self).__init__()
114
+ assert(kernel_size % 2 == 1)
115
+ self.hidden_channels =hidden_channels
116
+ self.kernel_size = kernel_size,
117
+ self.dilation_rate = dilation_rate
118
+ self.n_layers = n_layers
119
+ self.gin_channels = gin_channels
120
+ self.p_dropout = p_dropout
121
+
122
+ self.in_layers = torch.nn.ModuleList()
123
+ self.res_skip_layers = torch.nn.ModuleList()
124
+ self.drop = nn.Dropout(p_dropout)
125
+
126
+ if gin_channels != 0:
127
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
128
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
129
+
130
+ for i in range(n_layers):
131
+ dilation = dilation_rate ** i
132
+ padding = int((kernel_size * dilation - dilation) / 2)
133
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
134
+ dilation=dilation, padding=padding)
135
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
136
+ self.in_layers.append(in_layer)
137
+
138
+ # last one is not necessary
139
+ if i < n_layers - 1:
140
+ res_skip_channels = 2 * hidden_channels
141
+ else:
142
+ res_skip_channels = hidden_channels
143
+
144
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
145
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
146
+ self.res_skip_layers.append(res_skip_layer)
147
+
148
+ def forward(self, x, x_mask, g=None, **kwargs):
149
+ output = torch.zeros_like(x)
150
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
151
+
152
+ if g is not None:
153
+ g = self.cond_layer(g)
154
+
155
+ for i in range(self.n_layers):
156
+ x_in = self.in_layers[i](x)
157
+ if g is not None:
158
+ cond_offset = i * 2 * self.hidden_channels
159
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
160
+ else:
161
+ g_l = torch.zeros_like(x_in)
162
+
163
+ acts = commons.fused_add_tanh_sigmoid_multiply(
164
+ x_in,
165
+ g_l,
166
+ n_channels_tensor)
167
+ acts = self.drop(acts)
168
+
169
+ res_skip_acts = self.res_skip_layers[i](acts)
170
+ if i < self.n_layers - 1:
171
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
172
+ x = (x + res_acts) * x_mask
173
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
174
+ else:
175
+ output = output + res_skip_acts
176
+ return output * x_mask
177
+
178
+ def remove_weight_norm(self):
179
+ if self.gin_channels != 0:
180
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
181
+ for l in self.in_layers:
182
+ torch.nn.utils.remove_weight_norm(l)
183
+ for l in self.res_skip_layers:
184
+ torch.nn.utils.remove_weight_norm(l)
185
+
186
+
187
+ class ResBlock1(torch.nn.Module):
188
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
189
+ super(ResBlock1, self).__init__()
190
+ self.convs1 = nn.ModuleList([
191
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
192
+ padding=get_padding(kernel_size, dilation[0]))),
193
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
194
+ padding=get_padding(kernel_size, dilation[1]))),
195
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
196
+ padding=get_padding(kernel_size, dilation[2])))
197
+ ])
198
+ self.convs1.apply(init_weights)
199
+
200
+ self.convs2 = nn.ModuleList([
201
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
202
+ padding=get_padding(kernel_size, 1))),
203
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
204
+ padding=get_padding(kernel_size, 1))),
205
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
206
+ padding=get_padding(kernel_size, 1)))
207
+ ])
208
+ self.convs2.apply(init_weights)
209
+
210
+ def forward(self, x, x_mask=None):
211
+ for c1, c2 in zip(self.convs1, self.convs2):
212
+ xt = F.leaky_relu(x, LRELU_SLOPE)
213
+ if x_mask is not None:
214
+ xt = xt * x_mask
215
+ xt = c1(xt)
216
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
217
+ if x_mask is not None:
218
+ xt = xt * x_mask
219
+ xt = c2(xt)
220
+ x = xt + x
221
+ if x_mask is not None:
222
+ x = x * x_mask
223
+ return x
224
+
225
+ def remove_weight_norm(self):
226
+ for l in self.convs1:
227
+ remove_weight_norm(l)
228
+ for l in self.convs2:
229
+ remove_weight_norm(l)
230
+
231
+
232
+ class ResBlock2(torch.nn.Module):
233
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
234
+ super(ResBlock2, self).__init__()
235
+ self.convs = nn.ModuleList([
236
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
237
+ padding=get_padding(kernel_size, dilation[0]))),
238
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
239
+ padding=get_padding(kernel_size, dilation[1])))
240
+ ])
241
+ self.convs.apply(init_weights)
242
+
243
+ def forward(self, x, x_mask=None):
244
+ for c in self.convs:
245
+ xt = F.leaky_relu(x, LRELU_SLOPE)
246
+ if x_mask is not None:
247
+ xt = xt * x_mask
248
+ xt = c(xt)
249
+ x = xt + x
250
+ if x_mask is not None:
251
+ x = x * x_mask
252
+ return x
253
+
254
+ def remove_weight_norm(self):
255
+ for l in self.convs:
256
+ remove_weight_norm(l)
257
+
258
+
259
+ class Log(nn.Module):
260
+ def forward(self, x, x_mask, reverse=False, **kwargs):
261
+ if not reverse:
262
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
263
+ logdet = torch.sum(-y, [1, 2])
264
+ return y, logdet
265
+ else:
266
+ x = torch.exp(x) * x_mask
267
+ return x
268
+
269
+
270
+ class Flip(nn.Module):
271
+ def forward(self, x, *args, reverse=False, **kwargs):
272
+ x = torch.flip(x, [1])
273
+ if not reverse:
274
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
275
+ return x, logdet
276
+ else:
277
+ return x
278
+
279
+
280
+ class ElementwiseAffine(nn.Module):
281
+ def __init__(self, channels):
282
+ super().__init__()
283
+ self.channels = channels
284
+ self.m = nn.Parameter(torch.zeros(channels,1))
285
+ self.logs = nn.Parameter(torch.zeros(channels,1))
286
+
287
+ def forward(self, x, x_mask, reverse=False, **kwargs):
288
+ if not reverse:
289
+ y = self.m + torch.exp(self.logs) * x
290
+ y = y * x_mask
291
+ logdet = torch.sum(self.logs * x_mask, [1,2])
292
+ return y, logdet
293
+ else:
294
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
295
+ return x
296
+
297
+
298
+ class ResidualCouplingLayer(nn.Module):
299
+ def __init__(self,
300
+ channels,
301
+ hidden_channels,
302
+ kernel_size,
303
+ dilation_rate,
304
+ n_layers,
305
+ p_dropout=0,
306
+ gin_channels=0,
307
+ mean_only=False):
308
+ assert channels % 2 == 0, "channels should be divisible by 2"
309
+ super().__init__()
310
+ self.channels = channels
311
+ self.hidden_channels = hidden_channels
312
+ self.kernel_size = kernel_size
313
+ self.dilation_rate = dilation_rate
314
+ self.n_layers = n_layers
315
+ self.half_channels = channels // 2
316
+ self.mean_only = mean_only
317
+
318
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
319
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
320
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
321
+ self.post.weight.data.zero_()
322
+ self.post.bias.data.zero_()
323
+
324
+ def forward(self, x, x_mask, g=None, reverse=False):
325
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
326
+ h = self.pre(x0) * x_mask
327
+ h = self.enc(h, x_mask, g=g)
328
+ stats = self.post(h) * x_mask
329
+ if not self.mean_only:
330
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
331
+ else:
332
+ m = stats
333
+ logs = torch.zeros_like(m)
334
+
335
+ if not reverse:
336
+ x1 = m + x1 * torch.exp(logs) * x_mask
337
+ x = torch.cat([x0, x1], 1)
338
+ logdet = torch.sum(logs, [1,2])
339
+ return x, logdet
340
+ else:
341
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
342
+ x = torch.cat([x0, x1], 1)
343
+ return x
344
+
345
+
346
+ class ConvFlow(nn.Module):
347
+ def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
348
+ super().__init__()
349
+ self.in_channels = in_channels
350
+ self.filter_channels = filter_channels
351
+ self.kernel_size = kernel_size
352
+ self.n_layers = n_layers
353
+ self.num_bins = num_bins
354
+ self.tail_bound = tail_bound
355
+ self.half_channels = in_channels // 2
356
+
357
+ self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
358
+ self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
359
+ self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
360
+ self.proj.weight.data.zero_()
361
+ self.proj.bias.data.zero_()
362
+
363
+ def forward(self, x, x_mask, g=None, reverse=False):
364
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
365
+ h = self.pre(x0)
366
+ h = self.convs(h, x_mask, g=g)
367
+ h = self.proj(h) * x_mask
368
+
369
+ b, c, t = x0.shape
370
+ h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
371
+
372
+ unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
373
+ unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
374
+ unnormalized_derivatives = h[..., 2 * self.num_bins:]
375
+
376
+ x1, logabsdet = piecewise_rational_quadratic_transform(x1,
377
+ unnormalized_widths,
378
+ unnormalized_heights,
379
+ unnormalized_derivatives,
380
+ inverse=reverse,
381
+ tails='linear',
382
+ tail_bound=self.tail_bound
383
+ )
384
+
385
+ x = torch.cat([x0, x1], 1) * x_mask
386
+ logdet = torch.sum(logabsdet * x_mask, [1,2])
387
+ if not reverse:
388
+ return x, logdet
389
+ else:
390
+ return x
monotonic_align/__init__.py ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import torch
3
+ from .monotonic_align.core import maximum_path_c
4
+
5
+
6
+ def maximum_path(neg_cent, mask):
7
+ """ Cython optimized version.
8
+ neg_cent: [b, t_t, t_s]
9
+ mask: [b, t_t, t_s]
10
+ """
11
+ device = neg_cent.device
12
+ dtype = neg_cent.dtype
13
+ neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
14
+ path = np.zeros(neg_cent.shape, dtype=np.int32)
15
+
16
+ t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
17
+ t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
18
+ maximum_path_c(path, neg_cent, t_t_max, t_s_max)
19
+ return torch.from_numpy(path).to(device=device, dtype=dtype)
monotonic_align/core.pyx ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cimport cython
2
+ from cython.parallel import prange
3
+
4
+
5
+ @cython.boundscheck(False)
6
+ @cython.wraparound(False)
7
+ cdef void maximum_path_each(int[:,::1] path, float[:,::1] value, int t_y, int t_x, float max_neg_val=-1e9) nogil:
8
+ cdef int x
9
+ cdef int y
10
+ cdef float v_prev
11
+ cdef float v_cur
12
+ cdef float tmp
13
+ cdef int index = t_x - 1
14
+
15
+ for y in range(t_y):
16
+ for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
17
+ if x == y:
18
+ v_cur = max_neg_val
19
+ else:
20
+ v_cur = value[y-1, x]
21
+ if x == 0:
22
+ if y == 0:
23
+ v_prev = 0.
24
+ else:
25
+ v_prev = max_neg_val
26
+ else:
27
+ v_prev = value[y-1, x-1]
28
+ value[y, x] += max(v_prev, v_cur)
29
+
30
+ for y in range(t_y - 1, -1, -1):
31
+ path[y, index] = 1
32
+ if index != 0 and (index == y or value[y-1, index] < value[y-1, index-1]):
33
+ index = index - 1
34
+
35
+
36
+ @cython.boundscheck(False)
37
+ @cython.wraparound(False)
38
+ cpdef void maximum_path_c(int[:,:,::1] paths, float[:,:,::1] values, int[::1] t_ys, int[::1] t_xs) nogil:
39
+ cdef int b = paths.shape[0]
40
+ cdef int i
41
+ for i in prange(b, nogil=True):
42
+ maximum_path_each(paths[i], values[i], t_ys[i], t_xs[i])
monotonic_align/setup.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ from distutils.core import setup
2
+ from Cython.Build import cythonize
3
+ import numpy
4
+
5
+ setup(
6
+ name = 'monotonic_align',
7
+ ext_modules = cythonize("core.pyx"),
8
+ include_dirs=[numpy.get_include()]
9
+ )
preprocess.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import text
3
+ from utils import load_filepaths_and_text
4
+
5
+ if __name__ == '__main__':
6
+ parser = argparse.ArgumentParser()
7
+ parser.add_argument("--out_extension", default="cleaned")
8
+ parser.add_argument("--text_index", default=1, type=int)
9
+ parser.add_argument("--filelists", nargs="+", default=["filelists/ljs_audio_text_val_filelist.txt", "filelists/ljs_audio_text_test_filelist.txt"])
10
+ parser.add_argument("--text_cleaners", nargs="+", default=["english_cleaners2"])
11
+
12
+ args = parser.parse_args()
13
+
14
+
15
+ for filelist in args.filelists:
16
+ print("START:", filelist)
17
+ filepaths_and_text = load_filepaths_and_text(filelist)
18
+ for i in range(len(filepaths_and_text)):
19
+ original_text = filepaths_and_text[i][args.text_index]
20
+ cleaned_text = text._clean_text(original_text, args.text_cleaners)
21
+ filepaths_and_text[i][args.text_index] = cleaned_text
22
+
23
+ new_filelist = filelist + "." + args.out_extension
24
+ with open(new_filelist, "w", encoding="utf-8") as f:
25
+ f.writelines(["|".join(x) + "\n" for x in filepaths_and_text])
requirements.txt ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Cython
2
+ librosa
3
+ matplotlib
4
+ numpy
5
+ phonemizer
6
+ scipy
7
+ tensorboard
8
+ torch
9
+ torchvision
10
+ Unidecode
11
+ jieba
12
+ cn2an
13
+ pypinyin
resources/fig_1a.png ADDED
resources/fig_1b.png ADDED
resources/training.png ADDED
text/LICENSE ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (c) 2017 Keith Ito
2
+
3
+ Permission is hereby granted, free of charge, to any person obtaining a copy
4
+ of this software and associated documentation files (the "Software"), to deal
5
+ in the Software without restriction, including without limitation the rights
6
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
7
+ copies of the Software, and to permit persons to whom the Software is
8
+ furnished to do so, subject to the following conditions:
9
+
10
+ The above copyright notice and this permission notice shall be included in
11
+ all copies or substantial portions of the Software.
12
+
13
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
16
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
17
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
18
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
19
+ THE SOFTWARE.
text/__init__.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+ from text import cleaners
3
+ from text.symbols import symbols
4
+
5
+
6
+ # Mappings from symbol to numeric ID and vice versa:
7
+ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
8
+ _id_to_symbol = {i: s for i, s in enumerate(symbols)}
9
+
10
+
11
+ def text_to_sequence(text, cleaner_names):
12
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
13
+ Args:
14
+ text: string to convert to a sequence
15
+ cleaner_names: names of the cleaner functions to run the text through
16
+ Returns:
17
+ List of integers corresponding to the symbols in the text
18
+ '''
19
+ sequence = []
20
+
21
+ clean_text = _clean_text(text, cleaner_names)
22
+ for symbol in clean_text:
23
+ if symbol not in _symbol_to_id.keys():
24
+ continue
25
+ symbol_id = _symbol_to_id[symbol]
26
+ sequence += [symbol_id]
27
+ return sequence
28
+
29
+
30
+ def cleaned_text_to_sequence(cleaned_text):
31
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
32
+ Args:
33
+ text: string to convert to a sequence
34
+ Returns:
35
+ List of integers corresponding to the symbols in the text
36
+ '''
37
+ sequence = [_symbol_to_id[symbol] for symbol in cleaned_text if symbol in _symbol_to_id.keys()]
38
+ return sequence
39
+
40
+
41
+ def sequence_to_text(sequence):
42
+ '''Converts a sequence of IDs back to a string'''
43
+ result = ''
44
+ for symbol_id in sequence:
45
+ s = _id_to_symbol[symbol_id]
46
+ result += s
47
+ return result
48
+
49
+
50
+ def _clean_text(text, cleaner_names):
51
+ for name in cleaner_names:
52
+ cleaner = getattr(cleaners, name)
53
+ if not cleaner:
54
+ raise Exception('Unknown cleaner: %s' % name)
55
+ text = cleaner(text)
56
+ return text
text/cleaners.py ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ try:
3
+ from text.japanese import japanese_to_romaji_with_accent, japanese_to_ipa, japanese_to_ipa2, japanese_to_ipa3
4
+ from text.mandarin import number_to_chinese, chinese_to_bopomofo, latin_to_bopomofo, chinese_to_romaji, chinese_to_lazy_ipa, chinese_to_ipa, chinese_to_ipa2
5
+ except:
6
+ pass
7
+ # from text.sanskrit import devanagari_to_ipa
8
+ # from text.english import english_to_lazy_ipa, english_to_ipa2, english_to_lazy_ipa2
9
+ # from text.thai import num_to_thai, latin_to_thai
10
+ # from text.shanghainese import shanghainese_to_ipa
11
+ # from text.cantonese import cantonese_to_ipa
12
+ # from text.ngu_dialect import ngu_dialect_to_ipa
13
+
14
+
15
+ def japanese_cleaners(text):
16
+ text = japanese_to_romaji_with_accent(text)
17
+ if re.match('[A-Za-z]', text[-1]):
18
+ text += '.'
19
+ return text
20
+
21
+
22
+ def japanese_cleaners2(text):
23
+ return japanese_cleaners(text).replace('ts', 'ʦ').replace('...', '…')
24
+
25
+
26
+ def korean_cleaners(text):
27
+ '''Pipeline for Korean text'''
28
+ text = latin_to_hangul(text)
29
+ text = number_to_hangul(text)
30
+ text = divide_hangul(text)
31
+ if re.match('[\u3131-\u3163]', text[-1]):
32
+ text += '.'
33
+ return text
34
+
35
+
36
+ def chinese_cleaners(text):
37
+ '''Pipeline for Chinese text'''
38
+ text = number_to_chinese(text)
39
+ text = chinese_to_bopomofo(text)
40
+ text = latin_to_bopomofo(text)
41
+ if re.match('[ˉˊˇˋ˙]', text[-1]):
42
+ text += '。'
43
+ return text
44
+
45
+
46
+ def zh_ja_mixture_cleaners(text):
47
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
48
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
49
+ for chinese_text in chinese_texts:
50
+ cleaned_text = chinese_to_romaji(chinese_text[4:-4])
51
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
52
+ for japanese_text in japanese_texts:
53
+ cleaned_text = japanese_to_romaji_with_accent(
54
+ japanese_text[4:-4]).replace('ts', 'ʦ').replace('u', 'ɯ').replace('...', '…')
55
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
56
+ text = text[:-1]
57
+ if re.match('[A-Za-zɯɹəɥ→↓↑]', text[-1]):
58
+ text += '.'
59
+ return text
60
+
61
+
62
+ def sanskrit_cleaners(text):
63
+ text = text.replace('॥', '।').replace('ॐ', 'ओम्')
64
+ if text[-1] != '।':
65
+ text += ' ।'
66
+ return text
67
+
68
+
69
+ def cjks_cleaners(text):
70
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
71
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
72
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
73
+ sanskrit_texts = re.findall(r'\[SA\].*?\[SA\]', text)
74
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
75
+ for chinese_text in chinese_texts:
76
+ cleaned_text = chinese_to_lazy_ipa(chinese_text[4:-4])
77
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
78
+ for japanese_text in japanese_texts:
79
+ cleaned_text = japanese_to_ipa(japanese_text[4:-4])
80
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
81
+ for korean_text in korean_texts:
82
+ cleaned_text = korean_to_lazy_ipa(korean_text[4:-4])
83
+ text = text.replace(korean_text, cleaned_text+' ', 1)
84
+ for sanskrit_text in sanskrit_texts:
85
+ cleaned_text = devanagari_to_ipa(sanskrit_text[4:-4])
86
+ text = text.replace(sanskrit_text, cleaned_text+' ', 1)
87
+ for english_text in english_texts:
88
+ cleaned_text = english_to_lazy_ipa(english_text[4:-4])
89
+ text = text.replace(english_text, cleaned_text+' ', 1)
90
+ text = text[:-1]
91
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
92
+ text += '.'
93
+ return text
94
+
95
+
96
+ def cjke_cleaners(text):
97
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
98
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
99
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
100
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
101
+ for chinese_text in chinese_texts:
102
+ cleaned_text = chinese_to_lazy_ipa(chinese_text[4:-4])
103
+ cleaned_text = cleaned_text.replace(
104
+ 'ʧ', 'tʃ').replace('ʦ', 'ts').replace('ɥan', 'ɥæn')
105
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
106
+ for japanese_text in japanese_texts:
107
+ cleaned_text = japanese_to_ipa(japanese_text[4:-4])
108
+ cleaned_text = cleaned_text.replace('ʧ', 'tʃ').replace(
109
+ 'ʦ', 'ts').replace('ɥan', 'ɥæn').replace('ʥ', 'dz')
110
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
111
+ for korean_text in korean_texts:
112
+ cleaned_text = korean_to_ipa(korean_text[4:-4])
113
+ text = text.replace(korean_text, cleaned_text+' ', 1)
114
+ for english_text in english_texts:
115
+ cleaned_text = english_to_ipa2(english_text[4:-4])
116
+ cleaned_text = cleaned_text.replace('ɑ', 'a').replace(
117
+ 'ɔ', 'o').replace('ɛ', 'e').replace('ɪ', 'i').replace('ʊ', 'u')
118
+ text = text.replace(english_text, cleaned_text+' ', 1)
119
+ text = text[:-1]
120
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
121
+ text += '.'
122
+ return text
123
+
124
+
125
+ def cjke_cleaners2(text):
126
+ chinese_texts = re.findall(r'\[ZH\].*?\[ZH\]', text)
127
+ japanese_texts = re.findall(r'\[JA\].*?\[JA\]', text)
128
+ korean_texts = re.findall(r'\[KO\].*?\[KO\]', text)
129
+ english_texts = re.findall(r'\[EN\].*?\[EN\]', text)
130
+ for chinese_text in chinese_texts:
131
+ cleaned_text = chinese_to_ipa(chinese_text[4:-4])
132
+ text = text.replace(chinese_text, cleaned_text+' ', 1)
133
+ for japanese_text in japanese_texts:
134
+ cleaned_text = japanese_to_ipa2(japanese_text[4:-4])
135
+ text = text.replace(japanese_text, cleaned_text+' ', 1)
136
+ for korean_text in korean_texts:
137
+ cleaned_text = korean_to_ipa(korean_text[4:-4])
138
+ text = text.replace(korean_text, cleaned_text+' ', 1)
139
+ for english_text in english_texts:
140
+ cleaned_text = english_to_ipa2(english_text[4:-4])
141
+ text = text.replace(english_text, cleaned_text+' ', 1)
142
+ text = text[:-1]
143
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
144
+ text += '.'
145
+ return text
146
+
147
+
148
+ def thai_cleaners(text):
149
+ text = num_to_thai(text)
150
+ text = latin_to_thai(text)
151
+ return text
152
+
153
+
154
+ def shanghainese_cleaners(text):
155
+ text = shanghainese_to_ipa(text)
156
+ if re.match(r'[^\.,!\?\-…~]', text[-1]):
157
+ text += '.'
158
+ return text
159
+
160
+
161
+ def chinese_dialect_cleaners(text):
162
+ text = re.sub(r'\[MD\](.*?)\[MD\]',
163
+ lambda x: chinese_to_ipa2(x.group(1))+' ', text)
164
+ text = re.sub(r'\[TW\](.*?)\[TW\]',
165
+ lambda x: chinese_to_ipa2(x.group(1), True)+' ', text)
166
+ text = re.sub(r'\[JA\](.*?)\[JA\]',
167
+ lambda x: japanese_to_ipa3(x.group(1)).replace('Q', 'ʔ')+' ', text)
168
+ text = re.sub(r'\[SH\](.*?)\[SH\]', lambda x: shanghainese_to_ipa(x.group(1)).replace('1', '˥˧').replace('5',
169
+ '˧˧˦').replace('6', '˩˩˧').replace('7', '˥').replace('8', '˩˨').replace('ᴀ', 'ɐ').replace('ᴇ', 'e')+' ', text)
170
+ text = re.sub(r'\[GD\](.*?)\[GD\]',
171
+ lambda x: cantonese_to_ipa(x.group(1))+' ', text)
172
+ text = re.sub(r'\[EN\](.*?)\[EN\]',
173
+ lambda x: english_to_lazy_ipa2(x.group(1))+' ', text)
174
+ text = re.sub(r'\[([A-Z]{2})\](.*?)\[\1\]', lambda x: ngu_dialect_to_ipa(x.group(2), x.group(
175
+ 1)).replace('ʣ', 'dz').replace('ʥ', 'dʑ').replace('ʦ', 'ts').replace('ʨ', 'tɕ')+' ', text)
176
+ text = re.sub(r'\s+$', '', text)
177
+ text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
178
+ return text
text/japanese.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import re
2
+ from unidecode import unidecode
3
+ import pyopenjtalk
4
+
5
+
6
+ # Regular expression matching Japanese without punctuation marks:
7
+ _japanese_characters = re.compile(
8
+ r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
9
+
10
+ # Regular expression matching non-Japanese characters or punctuation marks:
11
+ _japanese_marks = re.compile(
12
+ r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
13
+
14
+ # List of (symbol, Japanese) pairs for marks:
15
+ _symbols_to_japanese = [(re.compile('%s' % x[0]), x[1]) for x in [
16
+ ('%', 'パーセント')
17
+ ]]
18
+
19
+ # List of (romaji, ipa) pairs for marks:
20
+ _romaji_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
21
+ ('ts', 'ʦ'),
22
+ ('u', 'ɯ'),
23
+ ('j', 'ʥ'),
24
+ ('y', 'j'),
25
+ ('ni', 'n^i'),
26
+ ('nj', 'n^'),
27
+ ('hi', 'çi'),
28
+ ('hj', 'ç'),
29
+ ('f', 'ɸ'),
30
+ ('I', 'i*'),
31
+ ('U', 'ɯ*'),
32
+ ('r', 'ɾ')
33
+ ]]
34
+
35
+ # List of (romaji, ipa2) pairs for marks:
36
+ _romaji_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
37
+ ('u', 'ɯ'),
38
+ ('ʧ', 'tʃ'),
39
+ ('j', 'dʑ'),
40
+ ('y', 'j'),
41
+ ('ni', 'n^i'),
42
+ ('nj', 'n^'),
43
+ ('hi', 'çi'),
44
+ ('hj', 'ç'),
45
+ ('f', 'ɸ'),
46
+ ('I', 'i*'),
47
+ ('U', 'ɯ*'),
48
+ ('r', 'ɾ')
49
+ ]]
50
+
51
+ # List of (consonant, sokuon) pairs:
52
+ _real_sokuon = [(re.compile('%s' % x[0]), x[1]) for x in [
53
+ (r'Q([↑↓]*[kg])', r'k#\1'),
54
+ (r'Q([↑↓]*[tdjʧ])', r't#\1'),
55
+ (r'Q([↑↓]*[sʃ])', r's\1'),
56
+ (r'Q([↑↓]*[pb])', r'p#\1')
57
+ ]]
58
+
59
+ # List of (consonant, hatsuon) pairs:
60
+ _real_hatsuon = [(re.compile('%s' % x[0]), x[1]) for x in [
61
+ (r'N([↑↓]*[pbm])', r'm\1'),
62
+ (r'N([↑↓]*[ʧʥj])', r'n^\1'),
63
+ (r'N([↑↓]*[tdn])', r'n\1'),
64
+ (r'N([↑↓]*[kg])', r'ŋ\1')
65
+ ]]
66
+
67
+
68
+ def symbols_to_japanese(text):
69
+ for regex, replacement in _symbols_to_japanese:
70
+ text = re.sub(regex, replacement, text)
71
+ return text
72
+
73
+
74
+ def japanese_to_romaji_with_accent(text):
75
+ '''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
76
+ text = symbols_to_japanese(text)
77
+ sentences = re.split(_japanese_marks, text)
78
+ marks = re.findall(_japanese_marks, text)
79
+ text = ''
80
+ for i, sentence in enumerate(sentences):
81
+ if re.match(_japanese_characters, sentence):
82
+ if text != '':
83
+ text += ' '
84
+ labels = pyopenjtalk.extract_fullcontext(sentence)
85
+ for n, label in enumerate(labels):
86
+ phoneme = re.search(r'\-([^\+]*)\+', label).group(1)
87
+ if phoneme not in ['sil', 'pau']:
88
+ text += phoneme.replace('ch', 'ʧ').replace('sh',
89
+ 'ʃ').replace('cl', 'Q')
90
+ else:
91
+ continue
92
+ # n_moras = int(re.search(r'/F:(\d+)_', label).group(1))
93
+ a1 = int(re.search(r"/A:(\-?[0-9]+)\+", label).group(1))
94
+ a2 = int(re.search(r"\+(\d+)\+", label).group(1))
95
+ a3 = int(re.search(r"\+(\d+)/", label).group(1))
96
+ if re.search(r'\-([^\+]*)\+', labels[n + 1]).group(1) in ['sil', 'pau']:
97
+ a2_next = -1
98
+ else:
99
+ a2_next = int(
100
+ re.search(r"\+(\d+)\+", labels[n + 1]).group(1))
101
+ # Accent phrase boundary
102
+ if a3 == 1 and a2_next == 1:
103
+ text += ' '
104
+ # Falling
105
+ elif a1 == 0 and a2_next == a2 + 1:
106
+ text += '↓'
107
+ # Rising
108
+ elif a2 == 1 and a2_next == 2:
109
+ text += '↑'
110
+ if i < len(marks):
111
+ text += unidecode(marks[i]).replace(' ', '')
112
+ return text
113
+
114
+
115
+ def get_real_sokuon(text):
116
+ for regex, replacement in _real_sokuon:
117
+ text = re.sub(regex, replacement, text)
118
+ return text
119
+
120
+
121
+ def get_real_hatsuon(text):
122
+ for regex, replacement in _real_hatsuon:
123
+ text = re.sub(regex, replacement, text)
124
+ return text
125
+
126
+
127
+ def japanese_to_ipa(text):
128
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
129
+ text = re.sub(
130
+ r'([aiueo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
131
+ text = get_real_sokuon(text)
132
+ text = get_real_hatsuon(text)
133
+ for regex, replacement in _romaji_to_ipa:
134
+ text = re.sub(regex, replacement, text)
135
+ return text
136
+
137
+
138
+ def japanese_to_ipa2(text):
139
+ text = japanese_to_romaji_with_accent(text).replace('...', '…')
140
+ text = get_real_sokuon(text)
141
+ text = get_real_hatsuon(text)
142
+ for regex, replacement in _romaji_to_ipa2:
143
+ text = re.sub(regex, replacement, text)
144
+ return text
145
+
146
+
147
+ def japanese_to_ipa3(text):
148
+ text = japanese_to_ipa2(text).replace('n^', 'ȵ').replace(
149
+ 'ʃ', 'ɕ').replace('*', '\u0325').replace('#', '\u031a')
150
+ text = re.sub(
151
+ r'([aiɯeo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
152
+ text = re.sub(r'((?:^|\s)(?:ts|tɕ|[kpt]))', r'\1ʰ', text)
153
+ return text
text/mandarin.py ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+ import re
4
+ from pypinyin import lazy_pinyin, BOPOMOFO
5
+ import jieba
6
+ import cn2an
7
+
8
+
9
+ # List of (Latin alphabet, bopomofo) pairs:
10
+ _latin_to_bopomofo = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
11
+ ('a', 'ㄟˉ'),
12
+ ('b', 'ㄅㄧˋ'),
13
+ ('c', 'ㄙㄧˉ'),
14
+ ('d', 'ㄉㄧˋ'),
15
+ ('e', 'ㄧˋ'),
16
+ ('f', 'ㄝˊㄈㄨˋ'),
17
+ ('g', 'ㄐㄧˋ'),
18
+ ('h', 'ㄝˇㄑㄩˋ'),
19
+ ('i', 'ㄞˋ'),
20
+ ('j', 'ㄐㄟˋ'),
21
+ ('k', 'ㄎㄟˋ'),
22
+ ('l', 'ㄝˊㄛˋ'),
23
+ ('m', 'ㄝˊㄇㄨˋ'),
24
+ ('n', 'ㄣˉ'),
25
+ ('o', 'ㄡˉ'),
26
+ ('p', 'ㄆㄧˉ'),
27
+ ('q', 'ㄎㄧㄡˉ'),
28
+ ('r', 'ㄚˋ'),
29
+ ('s', 'ㄝˊㄙˋ'),
30
+ ('t', 'ㄊㄧˋ'),
31
+ ('u', 'ㄧㄡˉ'),
32
+ ('v', 'ㄨㄧˉ'),
33
+ ('w', 'ㄉㄚˋㄅㄨˋㄌㄧㄡˋ'),
34
+ ('x', 'ㄝˉㄎㄨˋㄙˋ'),
35
+ ('y', 'ㄨㄞˋ'),
36
+ ('z', 'ㄗㄟˋ')
37
+ ]]
38
+
39
+ # List of (bopomofo, romaji) pairs:
40
+ _bopomofo_to_romaji = [(re.compile('%s' % x[0]), x[1]) for x in [
41
+ ('ㄅㄛ', 'p⁼wo'),
42
+ ('ㄆㄛ', 'pʰwo'),
43
+ ('ㄇㄛ', 'mwo'),
44
+ ('ㄈㄛ', 'fwo'),
45
+ ('ㄅ', 'p⁼'),
46
+ ('ㄆ', 'pʰ'),
47
+ ('ㄇ', 'm'),
48
+ ('ㄈ', 'f'),
49
+ ('ㄉ', 't⁼'),
50
+ ('ㄊ', 'tʰ'),
51
+ ('ㄋ', 'n'),
52
+ ('ㄌ', 'l'),
53
+ ('ㄍ', 'k⁼'),
54
+ ('ㄎ', 'kʰ'),
55
+ ('ㄏ', 'h'),
56
+ ('ㄐ', 'ʧ⁼'),
57
+ ('ㄑ', 'ʧʰ'),
58
+ ('ㄒ', 'ʃ'),
59
+ ('ㄓ', 'ʦ`⁼'),
60
+ ('ㄔ', 'ʦ`ʰ'),
61
+ ('ㄕ', 's`'),
62
+ ('ㄖ', 'ɹ`'),
63
+ ('ㄗ', 'ʦ⁼'),
64
+ ('ㄘ', 'ʦʰ'),
65
+ ('ㄙ', 's'),
66
+ ('ㄚ', 'a'),
67
+ ('ㄛ', 'o'),
68
+ ('ㄜ', 'ə'),
69
+ ('ㄝ', 'e'),
70
+ ('ㄞ', 'ai'),
71
+ ('ㄟ', 'ei'),
72
+ ('ㄠ', 'au'),
73
+ ('ㄡ', 'ou'),
74
+ ('ㄧㄢ', 'yeNN'),
75
+ ('ㄢ', 'aNN'),
76
+ ('ㄧㄣ', 'iNN'),
77
+ ('ㄣ', 'əNN'),
78
+ ('ㄤ', 'aNg'),
79
+ ('ㄧㄥ', 'iNg'),
80
+ ('ㄨㄥ', 'uNg'),
81
+ ('ㄩㄥ', 'yuNg'),
82
+ ('ㄥ', 'əNg'),
83
+ ('ㄦ', 'əɻ'),
84
+ ('ㄧ', 'i'),
85
+ ('ㄨ', 'u'),
86
+ ('ㄩ', 'ɥ'),
87
+ ('ˉ', '→'),
88
+ ('ˊ', '↑'),
89
+ ('ˇ', '↓↑'),
90
+ ('ˋ', '↓'),
91
+ ('˙', ''),
92
+ (',', ','),
93
+ ('。', '.'),
94
+ ('!', '!'),
95
+ ('?', '?'),
96
+ ('—', '-')
97
+ ]]
98
+
99
+ # List of (romaji, ipa) pairs:
100
+ _romaji_to_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
101
+ ('ʃy', 'ʃ'),
102
+ ('ʧʰy', 'ʧʰ'),
103
+ ('ʧ⁼y', 'ʧ⁼'),
104
+ ('NN', 'n'),
105
+ ('Ng', 'ŋ'),
106
+ ('y', 'j'),
107
+ ('h', 'x')
108
+ ]]
109
+
110
+ # List of (bopomofo, ipa) pairs:
111
+ _bopomofo_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
112
+ ('ㄅㄛ', 'p⁼wo'),
113
+ ('ㄆㄛ', 'pʰwo'),
114
+ ('ㄇㄛ', 'mwo'),
115
+ ('ㄈㄛ', 'fwo'),
116
+ ('ㄅ', 'p⁼'),
117
+ ('ㄆ', 'pʰ'),
118
+ ('ㄇ', 'm'),
119
+ ('ㄈ', 'f'),
120
+ ('ㄉ', 't⁼'),
121
+ ('ㄊ', 'tʰ'),
122
+ ('ㄋ', 'n'),
123
+ ('ㄌ', 'l'),
124
+ ('ㄍ', 'k⁼'),
125
+ ('ㄎ', 'kʰ'),
126
+ ('ㄏ', 'x'),
127
+ ('ㄐ', 'tʃ⁼'),
128
+ ('ㄑ', 'tʃʰ'),
129
+ ('ㄒ', 'ʃ'),
130
+ ('ㄓ', 'ts`⁼'),
131
+ ('ㄔ', 'ts`ʰ'),
132
+ ('ㄕ', 's`'),
133
+ ('ㄖ', 'ɹ`'),
134
+ ('ㄗ', 'ts⁼'),
135
+ ('ㄘ', 'tsʰ'),
136
+ ('ㄙ', 's'),
137
+ ('ㄚ', 'a'),
138
+ ('ㄛ', 'o'),
139
+ ('ㄜ', 'ə'),
140
+ ('ㄝ', 'ɛ'),
141
+ ('ㄞ', 'aɪ'),
142
+ ('ㄟ', 'eɪ'),
143
+ ('ㄠ', 'ɑʊ'),
144
+ ('ㄡ', 'oʊ'),
145
+ ('ㄧㄢ', 'jɛn'),
146
+ ('ㄩㄢ', 'ɥæn'),
147
+ ('ㄢ', 'an'),
148
+ ('ㄧㄣ', 'in'),
149
+ ('ㄩㄣ', 'ɥn'),
150
+ ('ㄣ', 'ən'),
151
+ ('ㄤ', 'ɑŋ'),
152
+ ('ㄧㄥ', 'iŋ'),
153
+ ('ㄨㄥ', 'ʊŋ'),
154
+ ('ㄩㄥ', 'jʊŋ'),
155
+ ('ㄥ', 'əŋ'),
156
+ ('ㄦ', 'əɻ'),
157
+ ('ㄧ', 'i'),
158
+ ('ㄨ', 'u'),
159
+ ('ㄩ', 'ɥ'),
160
+ ('ˉ', '→'),
161
+ ('ˊ', '↑'),
162
+ ('ˇ', '↓↑'),
163
+ ('ˋ', '↓'),
164
+ ('˙', ''),
165
+ (',', ','),
166
+ ('。', '.'),
167
+ ('!', '!'),
168
+ ('?', '?'),
169
+ ('—', '-')
170
+ ]]
171
+
172
+ # List of (bopomofo, ipa2) pairs:
173
+ _bopomofo_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
174
+ ('ㄅㄛ', 'pwo'),
175
+ ('ㄆㄛ', 'pʰwo'),
176
+ ('ㄇㄛ', 'mwo'),
177
+ ('ㄈㄛ', 'fwo'),
178
+ ('ㄅ', 'p'),
179
+ ('ㄆ', 'pʰ'),
180
+ ('ㄇ', 'm'),
181
+ ('ㄈ', 'f'),
182
+ ('ㄉ', 't'),
183
+ ('ㄊ', 'tʰ'),
184
+ ('ㄋ', 'n'),
185
+ ('ㄌ', 'l'),
186
+ ('ㄍ', 'k'),
187
+ ('ㄎ', 'kʰ'),
188
+ ('ㄏ', 'h'),
189
+ ('ㄐ', 'tɕ'),
190
+ ('ㄑ', 'tɕʰ'),
191
+ ('ㄒ', 'ɕ'),
192
+ ('ㄓ', 'tʂ'),
193
+ ('ㄔ', 'tʂʰ'),
194
+ ('ㄕ', 'ʂ'),
195
+ ('ㄖ', 'ɻ'),
196
+ ('ㄗ', 'ts'),
197
+ ('ㄘ', 'tsʰ'),
198
+ ('ㄙ', 's'),
199
+ ('ㄚ', 'a'),
200
+ ('ㄛ', 'o'),
201
+ ('ㄜ', 'ɤ'),
202
+ ('ㄝ', 'ɛ'),
203
+ ('ㄞ', 'aɪ'),
204
+ ('ㄟ', 'eɪ'),
205
+ ('ㄠ', 'ɑʊ'),
206
+ ('ㄡ', 'oʊ'),
207
+ ('ㄧㄢ', 'jɛn'),
208
+ ('ㄩㄢ', 'yæn'),
209
+ ('ㄢ', 'an'),
210
+ ('ㄧㄣ', 'in'),
211
+ ('ㄩㄣ', 'yn'),
212
+ ('ㄣ', 'ən'),
213
+ ('ㄤ', 'ɑŋ'),
214
+ ('ㄧㄥ', 'iŋ'),
215
+ ('ㄨㄥ', 'ʊŋ'),
216
+ ('ㄩㄥ', 'jʊŋ'),
217
+ ('ㄥ', 'ɤŋ'),
218
+ ('ㄦ', 'əɻ'),
219
+ ('ㄧ', 'i'),
220
+ ('ㄨ', 'u'),
221
+ ('ㄩ', 'y'),
222
+ ('ˉ', '˥'),
223
+ ('ˊ', '˧˥'),
224
+ ('ˇ', '˨˩˦'),
225
+ ('ˋ', '˥˩'),
226
+ ('˙', ''),
227
+ (',', ','),
228
+ ('。', '.'),
229
+ ('!', '!'),
230
+ ('?', '?'),
231
+ ('—', '-')
232
+ ]]
233
+
234
+
235
+ def number_to_chinese(text):
236
+ numbers = re.findall(r'\d+(?:\.?\d+)?', text)
237
+ for number in numbers:
238
+ text = text.replace(number, cn2an.an2cn(number), 1)
239
+ return text
240
+
241
+
242
+ def chinese_to_bopomofo(text, taiwanese=False):
243
+ text = text.replace('、', ',').replace(';', ',').replace(':', ',')
244
+ words = jieba.lcut(text, cut_all=False)
245
+ text = ''
246
+ for word in words:
247
+ bopomofos = lazy_pinyin(word, BOPOMOFO)
248
+ if not re.search('[\u4e00-\u9fff]', word):
249
+ text += word
250
+ continue
251
+ for i in range(len(bopomofos)):
252
+ bopomofos[i] = re.sub(r'([\u3105-\u3129])$', r'\1ˉ', bopomofos[i])
253
+ if text != '':
254
+ text += ' '
255
+ if taiwanese:
256
+ text += '#'+'#'.join(bopomofos)
257
+ else:
258
+ text += ''.join(bopomofos)
259
+ return text
260
+
261
+
262
+ def latin_to_bopomofo(text):
263
+ for regex, replacement in _latin_to_bopomofo:
264
+ text = re.sub(regex, replacement, text)
265
+ return text
266
+
267
+
268
+ def bopomofo_to_romaji(text):
269
+ for regex, replacement in _bopomofo_to_romaji:
270
+ text = re.sub(regex, replacement, text)
271
+ return text
272
+
273
+
274
+ def bopomofo_to_ipa(text):
275
+ for regex, replacement in _bopomofo_to_ipa:
276
+ text = re.sub(regex, replacement, text)
277
+ return text
278
+
279
+
280
+ def bopomofo_to_ipa2(text):
281
+ for regex, replacement in _bopomofo_to_ipa2:
282
+ text = re.sub(regex, replacement, text)
283
+ return text
284
+
285
+
286
+ def chinese_to_romaji(text):
287
+ text = number_to_chinese(text)
288
+ text = chinese_to_bopomofo(text)
289
+ text = latin_to_bopomofo(text)
290
+ text = bopomofo_to_romaji(text)
291
+ text = re.sub('i([aoe])', r'y\1', text)
292
+ text = re.sub('u([aoəe])', r'w\1', text)
293
+ text = re.sub('([ʦsɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
294
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
295
+ text = re.sub('([ʦs][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
296
+ return text
297
+
298
+
299
+ def chinese_to_lazy_ipa(text):
300
+ text = chinese_to_romaji(text)
301
+ for regex, replacement in _romaji_to_ipa:
302
+ text = re.sub(regex, replacement, text)
303
+ return text
304
+
305
+
306
+ def chinese_to_ipa(text):
307
+ text = number_to_chinese(text)
308
+ text = chinese_to_bopomofo(text)
309
+ text = latin_to_bopomofo(text)
310
+ text = bopomofo_to_ipa(text)
311
+ text = re.sub('i([aoe])', r'j\1', text)
312
+ text = re.sub('u([aoəe])', r'w\1', text)
313
+ text = re.sub('([sɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
314
+ r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
315
+ text = re.sub('([s][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
316
+ return text
317
+
318
+
319
+ def chinese_to_ipa2(text, taiwanese=False):
320
+ text = number_to_chinese(text)
321
+ text = chinese_to_bopomofo(text, taiwanese)
322
+ text = latin_to_bopomofo(text)
323
+ text = bopomofo_to_ipa2(text)
324
+ text = re.sub(r'i([aoe])', r'j\1', text)
325
+ text = re.sub(r'u([aoəe])', r'w\1', text)
326
+ text = re.sub(r'([ʂɹ]ʰ?)([˩˨˧˦˥ ]+|$)', r'\1ʅ\2', text)
327
+ text = re.sub(r'(sʰ?)([˩˨˧˦˥ ]+|$)', r'\1ɿ\2', text)
328
+ return text
text/symbols.py ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ '''
2
+ Defines the set of symbols used in text input to the model.
3
+ '''
4
+
5
+ _pad = '_'
6
+ _punctuation = ',.!?-'
7
+ _letters = 'AEINOQUabdefghijkmnoprstuvwyzʃʧ↓↑ '
8
+
9
+ '''
10
+ # japanese_cleaners2
11
+ _pad = '_'
12
+ _punctuation = ',.!?-~…'
13
+ _letters = 'AEINOQUabdefghijkmnoprstuvwyzʃʧʦ↓↑ '
14
+ '''
15
+
16
+ '''# korean_cleaners
17
+ _pad = '_'
18
+ _punctuation = ',.!?…~'
19
+ _letters = 'ㄱㄴㄷㄹㅁㅂㅅㅇㅈㅊㅋㅌㅍㅎㄲㄸㅃㅆㅉㅏㅓㅗㅜㅡㅣㅐㅔ '
20
+ '''
21
+
22
+ '''# chinese_cleaners
23
+ _pad = '_'
24
+ _punctuation = ',。!?—…'
25
+ _letters = 'ㄅㄆㄇㄈㄉㄊㄋㄌㄍㄎㄏㄐㄑㄒㄓㄔㄕㄖㄗㄘㄙㄚㄛㄜㄝㄞㄟㄠㄡㄢㄣㄤㄥㄦㄧㄨㄩˉˊˇˋ˙ '
26
+ '''
27
+
28
+
29
+ '''# sanskrit_cleaners
30
+ _pad = '_'
31
+ _punctuation = '।'
32
+ _letters = 'ँंःअआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलळवशषसहऽािीुूृॄेैोौ्ॠॢ '
33
+ '''
34
+
35
+ '''# cjks_cleaners
36
+ _pad = '_'
37
+ _punctuation = ',.!?-~…'
38
+ _letters = 'NQabdefghijklmnopstuvwxyzʃʧʥʦɯɹəɥçɸɾβŋɦː⁼ʰ`^#*=→↓↑ '
39
+ '''
40
+
41
+ '''# thai_cleaners
42
+ _pad = '_'
43
+ _punctuation = '.!? '
44
+ _letters = 'กขฃคฆงจฉชซฌญฎฏฐฑฒณดตถทธนบปผฝพฟภมยรฤลวศษสหฬอฮฯะัาำิีึืุูเแโใไๅๆ็่้๊๋์'
45
+ '''
46
+
47
+ '''# cjke_cleaners2
48
+ _pad = '_'
49
+ _punctuation = ',.!?-~…'
50
+ _letters = 'NQabdefghijklmnopstuvwxyzɑæʃʑçɯɪɔɛɹðəɫɥɸʊɾʒθβŋɦ⁼ʰ`^#*=ˈˌ→↓↑ '
51
+ '''
52
+
53
+ '''# shanghainese_cleaners
54
+ _pad = '_'
55
+ _punctuation = ',.!?…'
56
+ _letters = 'abdfghiklmnopstuvyzøŋȵɑɔɕəɤɦɪɿʑʔʰ̩̃ᴀᴇ15678 '
57
+ '''
58
+
59
+ '''# chinese_dialect_cleaners
60
+ _pad = '_'
61
+ _punctuation = ',.!?~…─'
62
+ _letters = '#Nabdefghijklmnoprstuvwxyzæçøŋœȵɐɑɒɓɔɕɗɘəɚɛɜɣɤɦɪɭɯɵɷɸɻɾɿʂʅʊʋʌʏʑʔʦʮʰʷˀː˥˦˧˨˩̥̩̃̚αᴀᴇ↑↓∅ⱼ '
63
+ '''
64
+
65
+ # Export all symbols:
66
+ symbols = [_pad] + list(_punctuation) + list(_letters)
67
+
68
+ # Special symbol ids
69
+ SPACE_ID = symbols.index(" ")
train_ms.py ADDED
@@ -0,0 +1,296 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import json
3
+ import argparse
4
+ import itertools
5
+ import math
6
+ import torch
7
+ from torch import nn, optim
8
+ from torch.nn import functional as F
9
+ from torch.utils.data import DataLoader
10
+ from torch.utils.tensorboard import SummaryWriter
11
+ import torch.multiprocessing as mp
12
+ import torch.distributed as dist
13
+ from torch.nn.parallel import DistributedDataParallel as DDP
14
+ from torch.cuda.amp import autocast, GradScaler
15
+
16
+ import commons
17
+ import utils
18
+ from data_utils import (
19
+ TextAudioSpeakerLoader,
20
+ TextAudioSpeakerCollate,
21
+ DistributedBucketSampler
22
+ )
23
+ from models import (
24
+ SynthesizerTrn,
25
+ MultiPeriodDiscriminator,
26
+ )
27
+ from losses import (
28
+ generator_loss,
29
+ discriminator_loss,
30
+ feature_loss,
31
+ kl_loss
32
+ )
33
+ from mel_processing import mel_spectrogram_torch, spec_to_mel_torch
34
+ from text.symbols import symbols
35
+
36
+
37
+ torch.backends.cudnn.benchmark = True
38
+ global_step = 0
39
+
40
+
41
+ def main():
42
+ """Assume Single Node Multi GPUs Training Only"""
43
+ assert torch.cuda.is_available(), "CPU training is not allowed."
44
+
45
+ n_gpus = torch.cuda.device_count()
46
+ os.environ['MASTER_ADDR'] = 'localhost'
47
+ os.environ['MASTER_PORT'] = '8899'
48
+
49
+ hps = utils.get_hparams()
50
+ mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
51
+
52
+
53
+ def run(rank, n_gpus, hps):
54
+ global global_step
55
+ if rank == 0:
56
+ logger = utils.get_logger(hps.model_dir)
57
+ print(hps) or logger.info(hps)
58
+ utils.check_git_hash(hps.model_dir)
59
+ writer = SummaryWriter(log_dir=hps.model_dir)
60
+ writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
61
+
62
+ dist.init_process_group(backend='nccl', init_method='env://', world_size=n_gpus, rank=rank)
63
+ torch.manual_seed(hps.train.seed)
64
+ torch.cuda.set_device(rank)
65
+
66
+ train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps.data)
67
+ train_sampler = DistributedBucketSampler(
68
+ train_dataset,
69
+ hps.train.batch_size,
70
+ [32,300,400,500,600,700,800,900,1000],
71
+ num_replicas=n_gpus,
72
+ rank=rank,
73
+ shuffle=True)
74
+ collate_fn = TextAudioSpeakerCollate()
75
+ train_loader = DataLoader(train_dataset, num_workers=8, shuffle=False, pin_memory=True,
76
+ collate_fn=collate_fn, batch_sampler=train_sampler)
77
+ if rank == 0:
78
+ eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps.data)
79
+ eval_loader = DataLoader(eval_dataset, num_workers=8, shuffle=False,
80
+ batch_size=hps.train.batch_size, pin_memory=True,
81
+ drop_last=False, collate_fn=collate_fn)
82
+
83
+ net_g = SynthesizerTrn(
84
+ len(symbols),
85
+ hps.data.filter_length // 2 + 1,
86
+ hps.train.segment_size // hps.data.hop_length,
87
+ n_speakers=hps.data.n_speakers,
88
+ **hps.model).cuda(rank)
89
+ net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
90
+ optim_g = torch.optim.AdamW(
91
+ net_g.parameters(),
92
+ hps.train.learning_rate,
93
+ betas=hps.train.betas,
94
+ eps=hps.train.eps)
95
+ optim_d = torch.optim.AdamW(
96
+ net_d.parameters(),
97
+ hps.train.learning_rate,
98
+ betas=hps.train.betas,
99
+ eps=hps.train.eps)
100
+ net_g = DDP(net_g, device_ids=[rank])
101
+ net_d = DDP(net_d, device_ids=[rank])
102
+
103
+ try:
104
+ _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g)
105
+ _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d)
106
+ global_step = (epoch_str - 1) * len(train_loader)
107
+ except:
108
+ epoch_str = 1
109
+ global_step = 0
110
+
111
+ scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
112
+ scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str-2)
113
+
114
+ scaler = GradScaler(enabled=hps.train.fp16_run)
115
+
116
+ for epoch in range(epoch_str, hps.train.epochs + 1):
117
+ if rank==0:
118
+ train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, eval_loader], logger, [writer, writer_eval])
119
+ else:
120
+ train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler, [train_loader, None], None, None)
121
+ scheduler_g.step()
122
+ scheduler_d.step()
123
+
124
+
125
+ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
126
+ net_g, net_d = nets
127
+ optim_g, optim_d = optims
128
+ scheduler_g, scheduler_d = schedulers
129
+ train_loader, eval_loader = loaders
130
+ if writers is not None:
131
+ writer, writer_eval = writers
132
+
133
+ train_loader.batch_sampler.set_epoch(epoch)
134
+ global global_step
135
+
136
+ net_g.train()
137
+ net_d.train()
138
+ for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers, emo) in enumerate(train_loader):
139
+ x, x_lengths = x.cuda(rank, non_blocking=True), x_lengths.cuda(rank, non_blocking=True)
140
+ spec, spec_lengths = spec.cuda(rank, non_blocking=True), spec_lengths.cuda(rank, non_blocking=True)
141
+ y, y_lengths = y.cuda(rank, non_blocking=True), y_lengths.cuda(rank, non_blocking=True)
142
+ speakers = speakers.cuda(rank, non_blocking=True)
143
+ emo = emo.cuda(rank, non_blocking=True)
144
+
145
+ with autocast(enabled=hps.train.fp16_run):
146
+ y_hat, l_length, attn, ids_slice, x_mask, z_mask,\
147
+ (z, z_p, m_p, logs_p, m_q, logs_q) = net_g(x, x_lengths, spec, spec_lengths, speakers, emo)
148
+
149
+ mel = spec_to_mel_torch(
150
+ spec,
151
+ hps.data.filter_length,
152
+ hps.data.n_mel_channels,
153
+ hps.data.sampling_rate,
154
+ hps.data.mel_fmin,
155
+ hps.data.mel_fmax)
156
+ y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
157
+ y_hat_mel = mel_spectrogram_torch(
158
+ y_hat.squeeze(1),
159
+ hps.data.filter_length,
160
+ hps.data.n_mel_channels,
161
+ hps.data.sampling_rate,
162
+ hps.data.hop_length,
163
+ hps.data.win_length,
164
+ hps.data.mel_fmin,
165
+ hps.data.mel_fmax
166
+ )
167
+
168
+ y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size) # slice
169
+
170
+ # Discriminator
171
+ y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
172
+ with autocast(enabled=False):
173
+ loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
174
+ loss_disc_all = loss_disc
175
+ optim_d.zero_grad()
176
+ scaler.scale(loss_disc_all).backward()
177
+ scaler.unscale_(optim_d)
178
+ grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
179
+ scaler.step(optim_d)
180
+
181
+ with autocast(enabled=hps.train.fp16_run):
182
+ # Generator
183
+ y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
184
+ with autocast(enabled=False):
185
+ loss_dur = torch.sum(l_length.float())
186
+ loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
187
+ loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
188
+
189
+ loss_fm = feature_loss(fmap_r, fmap_g)
190
+ loss_gen, losses_gen = generator_loss(y_d_hat_g)
191
+ loss_gen_all = loss_gen + loss_fm + loss_mel + loss_dur + loss_kl
192
+ optim_g.zero_grad()
193
+ scaler.scale(loss_gen_all.float()).backward()
194
+ scaler.unscale_(optim_g)
195
+ grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
196
+ scaler.step(optim_g)
197
+ scaler.update()
198
+
199
+ if rank==0:
200
+ if global_step % hps.train.log_interval == 0:
201
+ lr = optim_g.param_groups[0]['lr']
202
+ losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_dur, loss_kl]
203
+ logger.info('Train Epoch: {} [{:.0f}%]'.format(
204
+ epoch,
205
+ 100. * batch_idx / len(train_loader)))
206
+ print([x.item() for x in losses] + [global_step, lr]) or logger.info([x.item() for x in losses] + [global_step, lr])
207
+
208
+ scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr, "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
209
+ scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/dur": loss_dur, "loss/g/kl": loss_kl})
210
+
211
+ scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
212
+ scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
213
+ scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
214
+ image_dict = {
215
+ "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
216
+ "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()),
217
+ "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy()),
218
+ "all/attn": utils.plot_alignment_to_numpy(attn[0,0].data.cpu().numpy())
219
+ }
220
+ utils.summarize(
221
+ writer=writer,
222
+ global_step=global_step,
223
+ images=image_dict,
224
+ scalars=scalar_dict)
225
+
226
+ if global_step % hps.train.eval_interval == 0:
227
+ evaluate(hps, net_g, eval_loader, writer_eval)
228
+ utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
229
+ utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch, os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
230
+ global_step += 1
231
+
232
+ if rank == 0:
233
+ print('====> Epoch: {}'.format(epoch)) or logger.info('====> Epoch: {}'.format(epoch))
234
+
235
+
236
+ def evaluate(hps, generator, eval_loader, writer_eval):
237
+ generator.eval()
238
+ with torch.no_grad():
239
+ for batch_idx, (x, x_lengths, spec, spec_lengths, y, y_lengths, speakers, emo) in enumerate(eval_loader):
240
+ x, x_lengths = x.cuda(0), x_lengths.cuda(0)
241
+ spec, spec_lengths = spec.cuda(0), spec_lengths.cuda(0)
242
+ y, y_lengths = y.cuda(0), y_lengths.cuda(0)
243
+ speakers = speakers.cuda(0)
244
+ emo = emo.cuda(0)
245
+ # remove else
246
+ x = x[:1]
247
+ x_lengths = x_lengths[:1]
248
+ spec = spec[:1]
249
+ spec_lengths = spec_lengths[:1]
250
+ y = y[:1]
251
+ y_lengths = y_lengths[:1]
252
+ speakers = speakers[:1]
253
+ emo = emo[:1]
254
+ break
255
+ y_hat, attn, mask, *_ = generator.module.infer(x, x_lengths, speakers,emo, max_len=1000)
256
+ y_hat_lengths = mask.sum([1,2]).long() * hps.data.hop_length
257
+
258
+ mel = spec_to_mel_torch(
259
+ spec,
260
+ hps.data.filter_length,
261
+ hps.data.n_mel_channels,
262
+ hps.data.sampling_rate,
263
+ hps.data.mel_fmin,
264
+ hps.data.mel_fmax)
265
+ y_hat_mel = mel_spectrogram_torch(
266
+ y_hat.squeeze(1).float(),
267
+ hps.data.filter_length,
268
+ hps.data.n_mel_channels,
269
+ hps.data.sampling_rate,
270
+ hps.data.hop_length,
271
+ hps.data.win_length,
272
+ hps.data.mel_fmin,
273
+ hps.data.mel_fmax
274
+ )
275
+ image_dict = {
276
+ "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy())
277
+ }
278
+ audio_dict = {
279
+ "gen/audio": y_hat[0,:,:y_hat_lengths[0]]
280
+ }
281
+ if global_step == 0:
282
+ image_dict.update({"gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())})
283
+ audio_dict.update({"gt/audio": y[0,:,:y_lengths[0]]})
284
+
285
+ utils.summarize(
286
+ writer=writer_eval,
287
+ global_step=global_step,
288
+ images=image_dict,
289
+ audios=audio_dict,
290
+ audio_sampling_rate=hps.data.sampling_rate
291
+ )
292
+ generator.train()
293
+
294
+
295
+ if __name__ == "__main__":
296
+ main()
transforms.py ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import numpy as np
5
+
6
+
7
+ DEFAULT_MIN_BIN_WIDTH = 1e-3
8
+ DEFAULT_MIN_BIN_HEIGHT = 1e-3
9
+ DEFAULT_MIN_DERIVATIVE = 1e-3
10
+
11
+
12
+ def piecewise_rational_quadratic_transform(inputs,
13
+ unnormalized_widths,
14
+ unnormalized_heights,
15
+ unnormalized_derivatives,
16
+ inverse=False,
17
+ tails=None,
18
+ tail_bound=1.,
19
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
20
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
21
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
22
+
23
+ if tails is None:
24
+ spline_fn = rational_quadratic_spline
25
+ spline_kwargs = {}
26
+ else:
27
+ spline_fn = unconstrained_rational_quadratic_spline
28
+ spline_kwargs = {
29
+ 'tails': tails,
30
+ 'tail_bound': tail_bound
31
+ }
32
+
33
+ outputs, logabsdet = spline_fn(
34
+ inputs=inputs,
35
+ unnormalized_widths=unnormalized_widths,
36
+ unnormalized_heights=unnormalized_heights,
37
+ unnormalized_derivatives=unnormalized_derivatives,
38
+ inverse=inverse,
39
+ min_bin_width=min_bin_width,
40
+ min_bin_height=min_bin_height,
41
+ min_derivative=min_derivative,
42
+ **spline_kwargs
43
+ )
44
+ return outputs, logabsdet
45
+
46
+
47
+ def searchsorted(bin_locations, inputs, eps=1e-6):
48
+ bin_locations[..., -1] += eps
49
+ return torch.sum(
50
+ inputs[..., None] >= bin_locations,
51
+ dim=-1
52
+ ) - 1
53
+
54
+
55
+ def unconstrained_rational_quadratic_spline(inputs,
56
+ unnormalized_widths,
57
+ unnormalized_heights,
58
+ unnormalized_derivatives,
59
+ inverse=False,
60
+ tails='linear',
61
+ tail_bound=1.,
62
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
63
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
64
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
65
+ inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
66
+ outside_interval_mask = ~inside_interval_mask
67
+
68
+ outputs = torch.zeros_like(inputs)
69
+ logabsdet = torch.zeros_like(inputs)
70
+
71
+ if tails == 'linear':
72
+ unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
73
+ constant = np.log(np.exp(1 - min_derivative) - 1)
74
+ unnormalized_derivatives[..., 0] = constant
75
+ unnormalized_derivatives[..., -1] = constant
76
+
77
+ outputs[outside_interval_mask] = inputs[outside_interval_mask]
78
+ logabsdet[outside_interval_mask] = 0
79
+ else:
80
+ raise RuntimeError('{} tails are not implemented.'.format(tails))
81
+
82
+ outputs[inside_interval_mask], logabsdet[inside_interval_mask] = rational_quadratic_spline(
83
+ inputs=inputs[inside_interval_mask],
84
+ unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
85
+ unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
86
+ unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
87
+ inverse=inverse,
88
+ left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
89
+ min_bin_width=min_bin_width,
90
+ min_bin_height=min_bin_height,
91
+ min_derivative=min_derivative
92
+ )
93
+
94
+ return outputs, logabsdet
95
+
96
+ def rational_quadratic_spline(inputs,
97
+ unnormalized_widths,
98
+ unnormalized_heights,
99
+ unnormalized_derivatives,
100
+ inverse=False,
101
+ left=0., right=1., bottom=0., top=1.,
102
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
103
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
104
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
105
+ if torch.min(inputs) < left or torch.max(inputs) > right:
106
+ raise ValueError('Input to a transform is not within its domain')
107
+
108
+ num_bins = unnormalized_widths.shape[-1]
109
+
110
+ if min_bin_width * num_bins > 1.0:
111
+ raise ValueError('Minimal bin width too large for the number of bins')
112
+ if min_bin_height * num_bins > 1.0:
113
+ raise ValueError('Minimal bin height too large for the number of bins')
114
+
115
+ widths = F.softmax(unnormalized_widths, dim=-1)
116
+ widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
117
+ cumwidths = torch.cumsum(widths, dim=-1)
118
+ cumwidths = F.pad(cumwidths, pad=(1, 0), mode='constant', value=0.0)
119
+ cumwidths = (right - left) * cumwidths + left
120
+ cumwidths[..., 0] = left
121
+ cumwidths[..., -1] = right
122
+ widths = cumwidths[..., 1:] - cumwidths[..., :-1]
123
+
124
+ derivatives = min_derivative + F.softplus(unnormalized_derivatives)
125
+
126
+ heights = F.softmax(unnormalized_heights, dim=-1)
127
+ heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
128
+ cumheights = torch.cumsum(heights, dim=-1)
129
+ cumheights = F.pad(cumheights, pad=(1, 0), mode='constant', value=0.0)
130
+ cumheights = (top - bottom) * cumheights + bottom
131
+ cumheights[..., 0] = bottom
132
+ cumheights[..., -1] = top
133
+ heights = cumheights[..., 1:] - cumheights[..., :-1]
134
+
135
+ if inverse:
136
+ bin_idx = searchsorted(cumheights, inputs)[..., None]
137
+ else:
138
+ bin_idx = searchsorted(cumwidths, inputs)[..., None]
139
+
140
+ input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
141
+ input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
142
+
143
+ input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
144
+ delta = heights / widths
145
+ input_delta = delta.gather(-1, bin_idx)[..., 0]
146
+
147
+ input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
148
+ input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
149
+
150
+ input_heights = heights.gather(-1, bin_idx)[..., 0]
151
+
152
+ if inverse:
153
+ a = (((inputs - input_cumheights) * (input_derivatives
154
+ + input_derivatives_plus_one
155
+ - 2 * input_delta)
156
+ + input_heights * (input_delta - input_derivatives)))
157
+ b = (input_heights * input_derivatives
158
+ - (inputs - input_cumheights) * (input_derivatives
159
+ + input_derivatives_plus_one
160
+ - 2 * input_delta))
161
+ c = - input_delta * (inputs - input_cumheights)
162
+
163
+ discriminant = b.pow(2) - 4 * a * c
164
+ assert (discriminant >= 0).all()
165
+
166
+ root = (2 * c) / (-b - torch.sqrt(discriminant))
167
+ outputs = root * input_bin_widths + input_cumwidths
168
+
169
+ theta_one_minus_theta = root * (1 - root)
170
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
171
+ * theta_one_minus_theta)
172
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)
173
+ + 2 * input_delta * theta_one_minus_theta
174
+ + input_derivatives * (1 - root).pow(2))
175
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
176
+
177
+ return outputs, -logabsdet
178
+ else:
179
+ theta = (inputs - input_cumwidths) / input_bin_widths
180
+ theta_one_minus_theta = theta * (1 - theta)
181
+
182
+ numerator = input_heights * (input_delta * theta.pow(2)
183
+ + input_derivatives * theta_one_minus_theta)
184
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
185
+ * theta_one_minus_theta)
186
+ outputs = input_cumheights + numerator / denominator
187
+
188
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2)
189
+ + 2 * input_delta * theta_one_minus_theta
190
+ + input_derivatives * (1 - theta).pow(2))
191
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
192
+
193
+ return outputs, logabsdet
utils.py ADDED
@@ -0,0 +1,267 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import glob
3
+ import sys
4
+ import argparse
5
+ import logging
6
+ import json
7
+ import subprocess
8
+ import numpy as np
9
+ from scipy.io.wavfile import read
10
+ import torch
11
+
12
+ MATPLOTLIB_FLAG = False
13
+
14
+ logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
15
+ logger = logging
16
+
17
+
18
+ def load_checkpoint(checkpoint_path, model, optimizer=None):
19
+ assert os.path.isfile(checkpoint_path)
20
+ checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
21
+ iteration = checkpoint_dict['iteration']
22
+ learning_rate = checkpoint_dict['learning_rate']
23
+ if optimizer is not None:
24
+ optimizer.load_state_dict(checkpoint_dict['optimizer'])
25
+ saved_state_dict = checkpoint_dict['model']
26
+ if hasattr(model, 'module'):
27
+ state_dict = model.module.state_dict()
28
+ else:
29
+ state_dict = model.state_dict()
30
+ new_state_dict= {}
31
+ for k, v in state_dict.items():
32
+ try:
33
+ new_state_dict[k] = saved_state_dict[k]
34
+ except:
35
+ print("%s is not in the checkpoint" % k) or logger.info("%s is not in the checkpoint" % k)
36
+ new_state_dict[k] = v
37
+ if hasattr(model, 'module'):
38
+ model.module.load_state_dict(new_state_dict)
39
+ else:
40
+ model.load_state_dict(new_state_dict)
41
+ logger.info("Loaded checkpoint '{}' (iteration {})" .format(
42
+ checkpoint_path, iteration))
43
+ print("Loaded checkpoint '{}' (iteration {}) " .format(
44
+ checkpoint_path, iteration))
45
+ return model, optimizer, learning_rate, iteration
46
+
47
+
48
+ def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
49
+ ckptname = checkpoint_path.split("/")[-1]
50
+ newest_step = int(ckptname.split(".")[0].split("_")[1])
51
+ last_ckptname = checkpoint_path.replace(str(newest_step), str(newest_step-1200))
52
+ if newest_step >= 1200:
53
+ os.system(f"rm {last_ckptname}")
54
+ logger.info("Saving model and optimizer state at iteration {} to {}".format(
55
+ iteration, checkpoint_path))
56
+ print("Saving model and optimizer state at iteration {} to {}".format(
57
+ iteration, checkpoint_path))
58
+ if hasattr(model, 'module'):
59
+ state_dict = model.module.state_dict()
60
+ else:
61
+ state_dict = model.state_dict()
62
+ torch.save({'model': state_dict,
63
+ 'iteration': iteration,
64
+ 'optimizer': optimizer.state_dict(),
65
+ 'learning_rate': learning_rate}, checkpoint_path)
66
+
67
+
68
+ def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
69
+ for k, v in scalars.items():
70
+ writer.add_scalar(k, v, global_step)
71
+ for k, v in histograms.items():
72
+ writer.add_histogram(k, v, global_step)
73
+ for k, v in images.items():
74
+ writer.add_image(k, v, global_step, dataformats='HWC')
75
+ for k, v in audios.items():
76
+ writer.add_audio(k, v, global_step, audio_sampling_rate)
77
+
78
+
79
+ def latest_checkpoint_path(dir_path, regex="G_*.pth"):
80
+ f_list = glob.glob(os.path.join(dir_path, regex))
81
+ f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
82
+ x = f_list[-1]
83
+ print(x)
84
+ return x
85
+
86
+
87
+ def plot_spectrogram_to_numpy(spectrogram):
88
+ global MATPLOTLIB_FLAG
89
+ if not MATPLOTLIB_FLAG:
90
+ import matplotlib
91
+ matplotlib.use("Agg")
92
+ MATPLOTLIB_FLAG = True
93
+ mpl_logger = logging.getLogger('matplotlib')
94
+ mpl_logger.setLevel(logging.WARNING)
95
+ import matplotlib.pylab as plt
96
+ import numpy as np
97
+
98
+ fig, ax = plt.subplots(figsize=(10,2))
99
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
100
+ interpolation='none')
101
+ plt.colorbar(im, ax=ax)
102
+ plt.xlabel("Frames")
103
+ plt.ylabel("Channels")
104
+ plt.tight_layout()
105
+
106
+ fig.canvas.draw()
107
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
108
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
109
+ plt.close()
110
+ return data
111
+
112
+
113
+ def plot_alignment_to_numpy(alignment, info=None):
114
+ global MATPLOTLIB_FLAG
115
+ if not MATPLOTLIB_FLAG:
116
+ import matplotlib
117
+ matplotlib.use("Agg")
118
+ MATPLOTLIB_FLAG = True
119
+ mpl_logger = logging.getLogger('matplotlib')
120
+ mpl_logger.setLevel(logging.WARNING)
121
+ import matplotlib.pylab as plt
122
+ import numpy as np
123
+
124
+ fig, ax = plt.subplots(figsize=(6, 4))
125
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
126
+ interpolation='none')
127
+ fig.colorbar(im, ax=ax)
128
+ xlabel = 'Decoder timestep'
129
+ if info is not None:
130
+ xlabel += '\n\n' + info
131
+ plt.xlabel(xlabel)
132
+ plt.ylabel('Encoder timestep')
133
+ plt.tight_layout()
134
+
135
+ fig.canvas.draw()
136
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
137
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
138
+ plt.close()
139
+ return data
140
+
141
+
142
+ def load_wav_to_torch(full_path):
143
+ sampling_rate, data = read(full_path)
144
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
145
+
146
+
147
+ def load_filepaths_and_text(filename, split="|"):
148
+ with open(filename, encoding='utf-8') as f:
149
+ filepaths_and_text = [line.strip().split(split) for line in f]
150
+ return filepaths_and_text
151
+
152
+
153
+ def get_hparams(init=True):
154
+ parser = argparse.ArgumentParser()
155
+ parser.add_argument('-c', '--config', type=str, default="./configs/base.json",
156
+ help='JSON file for configuration')
157
+ parser.add_argument('-m', '--model', type=str, required=True,
158
+ help='Model name')
159
+
160
+ args = parser.parse_args()
161
+ model_dir = os.path.join("./logs", args.model)
162
+
163
+ if not os.path.exists(model_dir):
164
+ os.makedirs(model_dir)
165
+
166
+ config_path = args.config
167
+ config_save_path = os.path.join(model_dir, "config.json")
168
+ if init:
169
+ with open(config_path, "r") as f:
170
+ data = f.read()
171
+ with open(config_save_path, "w") as f:
172
+ f.write(data)
173
+ else:
174
+ with open(config_save_path, "r") as f:
175
+ data = f.read()
176
+ config = json.loads(data)
177
+
178
+ hparams = HParams(**config)
179
+ hparams.model_dir = model_dir
180
+ return hparams
181
+
182
+
183
+ def get_hparams_from_dir(model_dir):
184
+ config_save_path = os.path.join(model_dir, "config.json")
185
+ with open(config_save_path, "r") as f:
186
+ data = f.read()
187
+ config = json.loads(data)
188
+
189
+ hparams =HParams(**config)
190
+ hparams.model_dir = model_dir
191
+ return hparams
192
+
193
+
194
+ def get_hparams_from_file(config_path):
195
+ with open(config_path, "r") as f:
196
+ data = f.read()
197
+ config = json.loads(data)
198
+
199
+ hparams =HParams(**config)
200
+ return hparams
201
+
202
+
203
+ def check_git_hash(model_dir):
204
+ source_dir = os.path.dirname(os.path.realpath(__file__))
205
+ if not os.path.exists(os.path.join(source_dir, ".git")):
206
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
207
+ source_dir
208
+ ))
209
+ return
210
+
211
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
212
+
213
+ path = os.path.join(model_dir, "githash")
214
+ if os.path.exists(path):
215
+ saved_hash = open(path).read()
216
+ if saved_hash != cur_hash:
217
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
218
+ saved_hash[:8], cur_hash[:8]))
219
+ else:
220
+ open(path, "w").write(cur_hash)
221
+
222
+
223
+ def get_logger(model_dir, filename="train.log"):
224
+ global logger
225
+ logger = logging.getLogger(os.path.basename(model_dir))
226
+ logger.setLevel(logging.DEBUG)
227
+
228
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
229
+ if not os.path.exists(model_dir):
230
+ os.makedirs(model_dir)
231
+ h = logging.FileHandler(os.path.join(model_dir, filename))
232
+ h.setLevel(logging.DEBUG)
233
+ h.setFormatter(formatter)
234
+ logger.addHandler(h)
235
+ return logger
236
+
237
+
238
+ class HParams():
239
+ def __init__(self, **kwargs):
240
+ for k, v in kwargs.items():
241
+ if type(v) == dict:
242
+ v = HParams(**v)
243
+ self[k] = v
244
+
245
+ def keys(self):
246
+ return self.__dict__.keys()
247
+
248
+ def items(self):
249
+ return self.__dict__.items()
250
+
251
+ def values(self):
252
+ return self.__dict__.values()
253
+
254
+ def __len__(self):
255
+ return len(self.__dict__)
256
+
257
+ def __getitem__(self, key):
258
+ return getattr(self, key)
259
+
260
+ def __setitem__(self, key, value):
261
+ return setattr(self, key, value)
262
+
263
+ def __contains__(self, key):
264
+ return key in self.__dict__
265
+
266
+ def __repr__(self):
267
+ return self.__dict__.__repr__()