EDGAhab l1006986533 commited on
Commit
d1de79e
0 Parent(s):

Duplicate from l1006986533/vits-chinese

Browse files

Co-authored-by: liao <l1006986533@users.noreply.huggingface.co>

.gitattributes ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tflite filter=lfs diff=lfs merge=lfs -text
29
+ *.tgz filter=lfs diff=lfs merge=lfs -text
30
+ *.wasm filter=lfs diff=lfs merge=lfs -text
31
+ *.xz filter=lfs diff=lfs merge=lfs -text
32
+ *.zip filter=lfs diff=lfs merge=lfs -text
33
+ *.zst filter=lfs diff=lfs merge=lfs -text
34
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
.gitignore ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ .DS_Store
2
+ G_98000.pth
G_98000.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8229f5dfdc52704016fe1cd8a58d0c37ddabfe04266ea71a41fdce686fb30509
3
+ size 436704977
README.md ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: Vits Chinese
3
+ emoji: 🏃
4
+ colorFrom: red
5
+ colorTo: yellow
6
+ sdk: gradio
7
+ sdk_version: 3.9
8
+ app_file: app.py
9
+ pinned: false
10
+ duplicated_from: l1006986533/vits-chinese
11
+ ---
app.py ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+ import os
3
+ os.system('cd monotonic_align && python setup.py build_ext --inplace && cd ..')
4
+ import torch
5
+
6
+ import commons
7
+ import utils
8
+ from models import SynthesizerTrn
9
+ from text.symbols import symbols
10
+ from text import text_to_sequence
11
+
12
+ def get_text(text, hps):
13
+ text_norm = text_to_sequence(text, hps.data.text_cleaners)
14
+ if hps.data.add_blank:
15
+ text_norm = commons.intersperse(text_norm, 0)
16
+ text_norm = torch.LongTensor(text_norm)
17
+ return text_norm
18
+
19
+
20
+ hps = utils.get_hparams_from_file("configs/config.json")
21
+
22
+ net_g = SynthesizerTrn(
23
+ len(symbols),
24
+ hps.data.filter_length // 2 + 1,
25
+ hps.train.segment_size // hps.data.hop_length,
26
+ **hps.model)
27
+ _ = net_g.eval()
28
+
29
+ # _ = utils.load_checkpoint("logs/woman_csmsc/G_100000.pth", net_g, None)
30
+ _ = utils.load_checkpoint("G_98000.pth", net_g, None)
31
+
32
+ def vc_fn(input):
33
+ stn_tst = get_text(input, hps)
34
+ with torch.no_grad():
35
+ x_tst = stn_tst.unsqueeze(0)
36
+ x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
37
+
38
+ # x_tst = stn_tst.cpu().unsqueeze(0)
39
+ # x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).cpu()
40
+
41
+ audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=.667, noise_scale_w=0.8, length_scale=1)[0][0,0].data.cpu().float().numpy()
42
+ sampling_rate = 44100
43
+ return (sampling_rate, audio)
44
+
45
+ app = gr.Blocks()
46
+ with app:
47
+ with gr.Tabs():
48
+ with gr.TabItem("Basic"):
49
+ vc_input = gr.Textbox(label="Input Message")
50
+ vc_submit = gr.Button("Convert", variant="primary")
51
+ vc_output = gr.Audio(label="Output Audio")
52
+ vc_submit.click(vc_fn, [ vc_input], [vc_output])
53
+
54
+ 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/config.json ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 2000,
5
+ "seed": 1234,
6
+ "epochs": 20000,
7
+ "learning_rate": 2e-4,
8
+ "betas": [0.8, 0.99],
9
+ "eps": 1e-9,
10
+ "batch_size": 24,
11
+ "fp16_run": true,
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":"test/dsq.txt.cleaned",
21
+ "validation_files":"test/dsq_val.txt.cleaned",
22
+ "text_cleaners":["chinese_cleaners1"],
23
+ "max_wav_value": 32768.0,
24
+ "sampling_rate": 44100,
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": 0,
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
+ }
52
+ }
models.py ADDED
@@ -0,0 +1,534 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
157
+
158
+ self.encoder = attentions.Encoder(
159
+ hidden_channels,
160
+ filter_channels,
161
+ n_heads,
162
+ n_layers,
163
+ kernel_size,
164
+ p_dropout)
165
+ self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
166
+
167
+ def forward(self, x, x_lengths):
168
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
169
+ x = torch.transpose(x, 1, -1) # [b, h, t]
170
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
171
+
172
+ x = self.encoder(x * x_mask, x_mask)
173
+ stats = self.proj(x) * x_mask
174
+
175
+ m, logs = torch.split(stats, self.out_channels, dim=1)
176
+ return x, m, logs, x_mask
177
+
178
+
179
+ class ResidualCouplingBlock(nn.Module):
180
+ def __init__(self,
181
+ channels,
182
+ hidden_channels,
183
+ kernel_size,
184
+ dilation_rate,
185
+ n_layers,
186
+ n_flows=4,
187
+ gin_channels=0):
188
+ super().__init__()
189
+ self.channels = channels
190
+ self.hidden_channels = hidden_channels
191
+ self.kernel_size = kernel_size
192
+ self.dilation_rate = dilation_rate
193
+ self.n_layers = n_layers
194
+ self.n_flows = n_flows
195
+ self.gin_channels = gin_channels
196
+
197
+ self.flows = nn.ModuleList()
198
+ for i in range(n_flows):
199
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
200
+ self.flows.append(modules.Flip())
201
+
202
+ def forward(self, x, x_mask, g=None, reverse=False):
203
+ if not reverse:
204
+ for flow in self.flows:
205
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
206
+ else:
207
+ for flow in reversed(self.flows):
208
+ x = flow(x, x_mask, g=g, reverse=reverse)
209
+ return x
210
+
211
+
212
+ class PosteriorEncoder(nn.Module):
213
+ def __init__(self,
214
+ in_channels,
215
+ out_channels,
216
+ hidden_channels,
217
+ kernel_size,
218
+ dilation_rate,
219
+ n_layers,
220
+ gin_channels=0):
221
+ super().__init__()
222
+ self.in_channels = in_channels
223
+ self.out_channels = out_channels
224
+ self.hidden_channels = hidden_channels
225
+ self.kernel_size = kernel_size
226
+ self.dilation_rate = dilation_rate
227
+ self.n_layers = n_layers
228
+ self.gin_channels = gin_channels
229
+
230
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
231
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
232
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
233
+
234
+ def forward(self, x, x_lengths, g=None):
235
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
236
+ x = self.pre(x) * x_mask
237
+ x = self.enc(x, x_mask, g=g)
238
+ stats = self.proj(x) * x_mask
239
+ m, logs = torch.split(stats, self.out_channels, dim=1)
240
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
241
+ return z, m, logs, x_mask
242
+
243
+
244
+ class Generator(torch.nn.Module):
245
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
246
+ super(Generator, self).__init__()
247
+ self.num_kernels = len(resblock_kernel_sizes)
248
+ self.num_upsamples = len(upsample_rates)
249
+ self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
250
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
251
+
252
+ self.ups = nn.ModuleList()
253
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
254
+ self.ups.append(weight_norm(
255
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
256
+ k, u, padding=(k-u)//2)))
257
+
258
+ self.resblocks = nn.ModuleList()
259
+ for i in range(len(self.ups)):
260
+ ch = upsample_initial_channel//(2**(i+1))
261
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
262
+ self.resblocks.append(resblock(ch, k, d))
263
+
264
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
265
+ self.ups.apply(init_weights)
266
+
267
+ if gin_channels != 0:
268
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
269
+
270
+ def forward(self, x, g=None):
271
+ x = self.conv_pre(x)
272
+ if g is not None:
273
+ x = x + self.cond(g)
274
+
275
+ for i in range(self.num_upsamples):
276
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
277
+ x = self.ups[i](x)
278
+ xs = None
279
+ for j in range(self.num_kernels):
280
+ if xs is None:
281
+ xs = self.resblocks[i*self.num_kernels+j](x)
282
+ else:
283
+ xs += self.resblocks[i*self.num_kernels+j](x)
284
+ x = xs / self.num_kernels
285
+ x = F.leaky_relu(x)
286
+ x = self.conv_post(x)
287
+ x = torch.tanh(x)
288
+
289
+ return x
290
+
291
+ def remove_weight_norm(self):
292
+ print('Removing weight norm...')
293
+ for l in self.ups:
294
+ remove_weight_norm(l)
295
+ for l in self.resblocks:
296
+ l.remove_weight_norm()
297
+
298
+
299
+ class DiscriminatorP(torch.nn.Module):
300
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
301
+ super(DiscriminatorP, self).__init__()
302
+ self.period = period
303
+ self.use_spectral_norm = use_spectral_norm
304
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
305
+ self.convs = nn.ModuleList([
306
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
307
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
308
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
309
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
310
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
311
+ ])
312
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
313
+
314
+ def forward(self, x):
315
+ fmap = []
316
+
317
+ # 1d to 2d
318
+ b, c, t = x.shape
319
+ if t % self.period != 0: # pad first
320
+ n_pad = self.period - (t % self.period)
321
+ x = F.pad(x, (0, n_pad), "reflect")
322
+ t = t + n_pad
323
+ x = x.view(b, c, t // self.period, self.period)
324
+
325
+ for l in self.convs:
326
+ x = l(x)
327
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
328
+ fmap.append(x)
329
+ x = self.conv_post(x)
330
+ fmap.append(x)
331
+ x = torch.flatten(x, 1, -1)
332
+
333
+ return x, fmap
334
+
335
+
336
+ class DiscriminatorS(torch.nn.Module):
337
+ def __init__(self, use_spectral_norm=False):
338
+ super(DiscriminatorS, self).__init__()
339
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
340
+ self.convs = nn.ModuleList([
341
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
342
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
343
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
344
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
345
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
346
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
347
+ ])
348
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
349
+
350
+ def forward(self, x):
351
+ fmap = []
352
+
353
+ for l in self.convs:
354
+ x = l(x)
355
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
356
+ fmap.append(x)
357
+ x = self.conv_post(x)
358
+ fmap.append(x)
359
+ x = torch.flatten(x, 1, -1)
360
+
361
+ return x, fmap
362
+
363
+
364
+ class MultiPeriodDiscriminator(torch.nn.Module):
365
+ def __init__(self, use_spectral_norm=False):
366
+ super(MultiPeriodDiscriminator, self).__init__()
367
+ periods = [2,3,5,7,11]
368
+
369
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
370
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
371
+ self.discriminators = nn.ModuleList(discs)
372
+
373
+ def forward(self, y, y_hat):
374
+ y_d_rs = []
375
+ y_d_gs = []
376
+ fmap_rs = []
377
+ fmap_gs = []
378
+ for i, d in enumerate(self.discriminators):
379
+ y_d_r, fmap_r = d(y)
380
+ y_d_g, fmap_g = d(y_hat)
381
+ y_d_rs.append(y_d_r)
382
+ y_d_gs.append(y_d_g)
383
+ fmap_rs.append(fmap_r)
384
+ fmap_gs.append(fmap_g)
385
+
386
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
387
+
388
+
389
+
390
+ class SynthesizerTrn(nn.Module):
391
+ """
392
+ Synthesizer for Training
393
+ """
394
+
395
+ def __init__(self,
396
+ n_vocab,
397
+ spec_channels,
398
+ segment_size,
399
+ inter_channels,
400
+ hidden_channels,
401
+ filter_channels,
402
+ n_heads,
403
+ n_layers,
404
+ kernel_size,
405
+ p_dropout,
406
+ resblock,
407
+ resblock_kernel_sizes,
408
+ resblock_dilation_sizes,
409
+ upsample_rates,
410
+ upsample_initial_channel,
411
+ upsample_kernel_sizes,
412
+ n_speakers=0,
413
+ gin_channels=0,
414
+ use_sdp=True,
415
+ **kwargs):
416
+
417
+ super().__init__()
418
+ self.n_vocab = n_vocab
419
+ self.spec_channels = spec_channels
420
+ self.inter_channels = inter_channels
421
+ self.hidden_channels = hidden_channels
422
+ self.filter_channels = filter_channels
423
+ self.n_heads = n_heads
424
+ self.n_layers = n_layers
425
+ self.kernel_size = kernel_size
426
+ self.p_dropout = p_dropout
427
+ self.resblock = resblock
428
+ self.resblock_kernel_sizes = resblock_kernel_sizes
429
+ self.resblock_dilation_sizes = resblock_dilation_sizes
430
+ self.upsample_rates = upsample_rates
431
+ self.upsample_initial_channel = upsample_initial_channel
432
+ self.upsample_kernel_sizes = upsample_kernel_sizes
433
+ self.segment_size = segment_size
434
+ self.n_speakers = n_speakers
435
+ self.gin_channels = gin_channels
436
+
437
+ self.use_sdp = use_sdp
438
+
439
+ self.enc_p = TextEncoder(n_vocab,
440
+ inter_channels,
441
+ hidden_channels,
442
+ filter_channels,
443
+ n_heads,
444
+ n_layers,
445
+ kernel_size,
446
+ p_dropout)
447
+ self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
448
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
449
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
450
+
451
+ if use_sdp:
452
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
453
+ else:
454
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
455
+
456
+ if n_speakers > 1:
457
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
458
+
459
+ def forward(self, x, x_lengths, y, y_lengths, sid=None):
460
+
461
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
462
+ if self.n_speakers > 0:
463
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
464
+ else:
465
+ g = None
466
+
467
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
468
+ z_p = self.flow(z, y_mask, g=g)
469
+
470
+ with torch.no_grad():
471
+ # negative cross-entropy
472
+ s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
473
+ neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
474
+ 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]
475
+ 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]
476
+ neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
477
+ neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
478
+
479
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
480
+ attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
481
+
482
+ w = attn.sum(2)
483
+ if self.use_sdp:
484
+ l_length = self.dp(x, x_mask, w, g=g)
485
+ l_length = l_length / torch.sum(x_mask)
486
+ else:
487
+ logw_ = torch.log(w + 1e-6) * x_mask
488
+ logw = self.dp(x, x_mask, g=g)
489
+ l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
490
+
491
+ # expand prior
492
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
493
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
494
+
495
+ z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
496
+ o = self.dec(z_slice, g=g)
497
+ return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
498
+
499
+ def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
500
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
501
+ if self.n_speakers > 0:
502
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
503
+ else:
504
+ g = None
505
+
506
+ if self.use_sdp:
507
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
508
+ else:
509
+ logw = self.dp(x, x_mask, g=g)
510
+ w = torch.exp(logw) * x_mask * length_scale
511
+ w_ceil = torch.ceil(w)
512
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
513
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
514
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
515
+ attn = commons.generate_path(w_ceil, attn_mask)
516
+
517
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
518
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
519
+
520
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
521
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
522
+ o = self.dec((z * y_mask)[:,:,:max_len], g=g)
523
+ return o, attn, y_mask, (z, z_p, m_p, logs_p)
524
+
525
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
526
+ assert self.n_speakers > 0, "n_speakers have to be larger than 0."
527
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
528
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
529
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
530
+ z_p = self.flow(z, y_mask, g=g_src)
531
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
532
+ o_hat = self.dec(z_hat * y_mask, g=g_tgt)
533
+ return o_hat, y_mask, (z, z_p, z_hat)
534
+
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.c ADDED
The diff for this file is too large to render. See raw diff
 
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
+ )
requirements.txt ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ torch
2
+ scipy
3
+ Cython
4
+ unidecode
5
+ pypinyin
6
+ phonemizer
text/__init__.py ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+ from text import cleaners
3
+ from text.symbols import symbols,symbols_zh
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
+ chinese_mode = True
11
+ if chinese_mode:
12
+ _symbol_to_id = {s: i for i, s in enumerate(symbols_zh)}
13
+ _id_to_symbol = {i: s for i, s in enumerate(symbols_zh)}
14
+ else:
15
+ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
16
+ _id_to_symbol = {i: s for i, s in enumerate(symbols)}
17
+
18
+ def text_to_sequence(text, cleaner_names, ):
19
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
20
+ Args:
21
+ text: string to convert to a sequence
22
+ cleaner_names: names of the cleaner functions to run the text through
23
+ Returns:
24
+ List of integers corresponding to the symbols in the text
25
+ '''
26
+ sequence = []
27
+
28
+ clean_text = _clean_text(text, cleaner_names)
29
+ for symbol in clean_text:
30
+ symbol_id = _symbol_to_id[symbol]
31
+ sequence += [symbol_id]
32
+ return sequence
33
+
34
+
35
+ def cleaned_text_to_sequence(cleaned_text, chinese_mode=True):
36
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
37
+ Args:
38
+ text: string to convert to a sequence
39
+ Returns:
40
+ List of integers corresponding to the symbols in the text
41
+ '''
42
+ # if chinese_mode:
43
+ # sequence = [_symbol_to_id_zh[symbol] for symbol in cleaned_text]
44
+ # else:
45
+ sequence = [_symbol_to_id[symbol] for symbol in cleaned_text]
46
+ return sequence
47
+
48
+
49
+ def sequence_to_text(sequence):
50
+ '''Converts a sequence of IDs back to a string'''
51
+ result = ''
52
+ for symbol_id in sequence:
53
+ s = _id_to_symbol[symbol_id]
54
+ result += s
55
+ return result
56
+
57
+
58
+ def _clean_text(text, cleaner_names):
59
+ for name in cleaner_names:
60
+ cleaner = getattr(cleaners, name)
61
+ if not cleaner:
62
+ raise Exception('Unknown cleaner: %s' % name)
63
+ text = cleaner(text)
64
+ return text
text/cleaners.py ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+
3
+ '''
4
+ Cleaners are transformations that run over the input text at both training and eval time.
5
+
6
+ Cleaners can be selected by passing a comma-delimited list of cleaner names as the "cleaners"
7
+ hyperparameter. Some cleaners are English-specific. You'll typically want to use:
8
+ 1. "english_cleaners" for English text
9
+ 2. "transliteration_cleaners" for non-English text that can be transliterated to ASCII using
10
+ the Unidecode library (https://pypi.python.org/pypi/Unidecode)
11
+ 3. "basic_cleaners" if you do not want to transliterate (in this case, you should also update
12
+ the symbols in symbols.py to match your data).
13
+ '''
14
+
15
+ import re
16
+ from unidecode import unidecode
17
+ from phonemizer import phonemize
18
+ from pypinyin import Style, pinyin
19
+ from pypinyin.style._utils import get_finals, get_initials
20
+ # Regular expression matching whitespace:
21
+ _whitespace_re = re.compile(r'\s+')
22
+
23
+ # List of (regular expression, replacement) pairs for abbreviations:
24
+ _abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [
25
+ ('mrs', 'misess'),
26
+ ('mr', 'mister'),
27
+ ('dr', 'doctor'),
28
+ ('st', 'saint'),
29
+ ('co', 'company'),
30
+ ('jr', 'junior'),
31
+ ('maj', 'major'),
32
+ ('gen', 'general'),
33
+ ('drs', 'doctors'),
34
+ ('rev', 'reverend'),
35
+ ('lt', 'lieutenant'),
36
+ ('hon', 'honorable'),
37
+ ('sgt', 'sergeant'),
38
+ ('capt', 'captain'),
39
+ ('esq', 'esquire'),
40
+ ('ltd', 'limited'),
41
+ ('col', 'colonel'),
42
+ ('ft', 'fort'),
43
+ ]]
44
+
45
+
46
+ def expand_abbreviations(text):
47
+ for regex, replacement in _abbreviations:
48
+ text = re.sub(regex, replacement, text)
49
+ return text
50
+
51
+
52
+ def expand_numbers(text):
53
+ return normalize_numbers(text)
54
+
55
+
56
+ def lowercase(text):
57
+ return text.lower()
58
+
59
+
60
+ def collapse_whitespace(text):
61
+ return re.sub(_whitespace_re, ' ', text)
62
+
63
+
64
+ def convert_to_ascii(text):
65
+ return unidecode(text)
66
+
67
+
68
+ def basic_cleaners(text):
69
+ '''Basic pipeline that lowercases and collapses whitespace without transliteration.'''
70
+ text = lowercase(text)
71
+ text = collapse_whitespace(text)
72
+ return text
73
+
74
+
75
+ def transliteration_cleaners(text):
76
+ '''Pipeline for non-English text that transliterates to ASCII.'''
77
+ text = convert_to_ascii(text)
78
+ text = lowercase(text)
79
+ text = collapse_whitespace(text)
80
+ return text
81
+
82
+
83
+ def english_cleaners(text):
84
+ '''Pipeline for English text, including abbreviation expansion.'''
85
+ text = convert_to_ascii(text)
86
+ text = lowercase(text)
87
+ text = expand_abbreviations(text)
88
+ phonemes = phonemize(text, language='en-us', backend='espeak', strip=True)
89
+ phonemes = collapse_whitespace(phonemes)
90
+ return phonemes
91
+
92
+
93
+ def english_cleaners2(text):
94
+ '''Pipeline for English text, including abbreviation expansion. + punctuation + stress'''
95
+ text = convert_to_ascii(text)
96
+ text = lowercase(text)
97
+ text = expand_abbreviations(text)
98
+ phonemes = phonemize(text, language='en-us', backend='espeak', strip=True, preserve_punctuation=True, with_stress=True)
99
+ phonemes = collapse_whitespace(phonemes)
100
+ return phonemes
101
+
102
+
103
+
104
+
105
+ def chinese_cleaners1(text):
106
+ from pypinyin import Style, pinyin
107
+
108
+ phones = [phone[0] for phone in pinyin(text, style=Style.TONE3)]
109
+ return ' '.join(phones)
110
+
111
+
112
+ def chinese_cleaners2(text):
113
+ phones = [
114
+ p
115
+ for phone in pinyin(text, style=Style.TONE3)
116
+ for p in [
117
+ get_initials(phone[0], strict=True),
118
+ get_finals(phone[0][:-1], strict=True) + phone[0][-1]
119
+ if phone[0][-1].isdigit()
120
+ else get_finals(phone[0], strict=True)
121
+ if phone[0][-1].isalnum()
122
+ else phone[0],
123
+ ]
124
+ # Remove the case of individual tones as a phoneme
125
+ if len(p) != 0 and not p.isdigit()
126
+ ]
127
+ return phones
128
+ # return phonemes
129
+
130
+ if __name__ == '__main__':
131
+ res = chinese_cleaners2('这是语音测试!')
132
+ print(res)
133
+ res = chinese_cleaners1('"第一,南京不是发展的不行,是大家对他期望很高,')
134
+ print(res)
135
+
136
+
137
+ res = english_cleaners2('this is a club test for one train.GDP')
138
+ print(res)
text/symbols.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ from https://github.com/keithito/tacotron """
2
+
3
+ '''
4
+ Defines the set of symbols used in text input to the model.
5
+ '''
6
+ _pad = '_'
7
+ _punctuation = ';:,.!?¡¿—…"«»“” '
8
+
9
+ _punctuation_zh = ';:,。!?-“”《》、 '
10
+ _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
11
+
12
+ _numbers = '123450'
13
+ _others = ''
14
+
15
+ _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
16
+
17
+
18
+ # Export all symbols:
19
+ symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
20
+
21
+ symbols_zh = [_pad] + list(_punctuation_zh) + list(_letters) + list(_numbers)
22
+
23
+ # Special symbol ids
24
+ SPACE_ID = symbols.index(" ")
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 +1,259 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ 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
+ return model, optimizer, learning_rate, iteration
44
+
45
+
46
+ def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
47
+ logger.info("Saving model and optimizer state at iteration {} to {}".format(
48
+ iteration, checkpoint_path))
49
+ if hasattr(model, 'module'):
50
+ state_dict = model.module.state_dict()
51
+ else:
52
+ state_dict = model.state_dict()
53
+ torch.save({'model': state_dict,
54
+ 'iteration': iteration,
55
+ 'optimizer': optimizer.state_dict(),
56
+ 'learning_rate': learning_rate}, checkpoint_path)
57
+
58
+
59
+ def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
60
+ for k, v in scalars.items():
61
+ writer.add_scalar(k, v, global_step)
62
+ for k, v in histograms.items():
63
+ writer.add_histogram(k, v, global_step)
64
+ for k, v in images.items():
65
+ writer.add_image(k, v, global_step, dataformats='HWC')
66
+ for k, v in audios.items():
67
+ writer.add_audio(k, v, global_step, audio_sampling_rate)
68
+
69
+
70
+ def latest_checkpoint_path(dir_path, regex="G_*.pth"):
71
+ f_list = glob.glob(os.path.join(dir_path, regex))
72
+ f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
73
+ x = f_list[-1]
74
+ print(x)
75
+ return x
76
+
77
+
78
+ def plot_spectrogram_to_numpy(spectrogram):
79
+ global MATPLOTLIB_FLAG
80
+ if not MATPLOTLIB_FLAG:
81
+ import matplotlib
82
+ matplotlib.use("Agg")
83
+ MATPLOTLIB_FLAG = True
84
+ mpl_logger = logging.getLogger('matplotlib')
85
+ mpl_logger.setLevel(logging.WARNING)
86
+ import matplotlib.pylab as plt
87
+ import numpy as np
88
+
89
+ fig, ax = plt.subplots(figsize=(10,2))
90
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
91
+ interpolation='none')
92
+ plt.colorbar(im, ax=ax)
93
+ plt.xlabel("Frames")
94
+ plt.ylabel("Channels")
95
+ plt.tight_layout()
96
+
97
+ fig.canvas.draw()
98
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
99
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
100
+ plt.close()
101
+ return data
102
+
103
+
104
+ def plot_alignment_to_numpy(alignment, info=None):
105
+ global MATPLOTLIB_FLAG
106
+ if not MATPLOTLIB_FLAG:
107
+ import matplotlib
108
+ matplotlib.use("Agg")
109
+ MATPLOTLIB_FLAG = True
110
+ mpl_logger = logging.getLogger('matplotlib')
111
+ mpl_logger.setLevel(logging.WARNING)
112
+ import matplotlib.pylab as plt
113
+ import numpy as np
114
+
115
+ fig, ax = plt.subplots(figsize=(6, 4))
116
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
117
+ interpolation='none')
118
+ fig.colorbar(im, ax=ax)
119
+ xlabel = 'Decoder timestep'
120
+ if info is not None:
121
+ xlabel += '\n\n' + info
122
+ plt.xlabel(xlabel)
123
+ plt.ylabel('Encoder timestep')
124
+ plt.tight_layout()
125
+
126
+ fig.canvas.draw()
127
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
128
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
129
+ plt.close()
130
+ return data
131
+
132
+
133
+ def load_wav_to_torch(full_path):
134
+ sampling_rate, data = read(full_path)
135
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
136
+
137
+
138
+ def load_filepaths_and_text(filename, split="|"):
139
+ with open(filename, encoding='utf-8') as f:
140
+ filepaths_and_text = [line.strip().split(split) for line in f]
141
+ return filepaths_and_text
142
+
143
+
144
+ def get_hparams(init=True):
145
+ parser = argparse.ArgumentParser()
146
+ parser.add_argument('-c', '--config', type=str, default="./configs/woman_csmsc.json",
147
+ help='JSON file for configuration')
148
+ parser.add_argument('-m', '--model', type=str, default="./woman_csmsc",
149
+ # required=True,
150
+ help='Model name')
151
+
152
+ args = parser.parse_args()
153
+ model_dir = os.path.join("./logs", args.model)
154
+
155
+ if not os.path.exists(model_dir):
156
+ os.makedirs(model_dir)
157
+
158
+ config_path = args.config
159
+ config_save_path = os.path.join(model_dir, "config.json")
160
+ if init:
161
+ with open(config_path, "r") as f:
162
+ data = f.read()
163
+ with open(config_save_path, "w") as f:
164
+ f.write(data)
165
+ else:
166
+ with open(config_save_path, "r") as f:
167
+ data = f.read()
168
+ config = json.loads(data)
169
+
170
+ hparams = HParams(**config)
171
+ hparams.model_dir = model_dir
172
+ return hparams
173
+
174
+
175
+ def get_hparams_from_dir(model_dir):
176
+ config_save_path = os.path.join(model_dir, "config.json")
177
+ with open(config_save_path, "r") as f:
178
+ data = f.read()
179
+ config = json.loads(data)
180
+
181
+ hparams =HParams(**config)
182
+ hparams.model_dir = model_dir
183
+ return hparams
184
+
185
+
186
+ def get_hparams_from_file(config_path):
187
+ with open(config_path, "r") as f:
188
+ data = f.read()
189
+ config = json.loads(data)
190
+
191
+ hparams =HParams(**config)
192
+ return hparams
193
+
194
+
195
+ def check_git_hash(model_dir):
196
+ source_dir = os.path.dirname(os.path.realpath(__file__))
197
+ if not os.path.exists(os.path.join(source_dir, ".git")):
198
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
199
+ source_dir
200
+ ))
201
+ return
202
+
203
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
204
+
205
+ path = os.path.join(model_dir, "githash")
206
+ if os.path.exists(path):
207
+ saved_hash = open(path).read()
208
+ if saved_hash != cur_hash:
209
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
210
+ saved_hash[:8], cur_hash[:8]))
211
+ else:
212
+ open(path, "w").write(cur_hash)
213
+
214
+
215
+ def get_logger(model_dir, filename="train.log"):
216
+ global logger
217
+ logger = logging.getLogger(os.path.basename(model_dir))
218
+ logger.setLevel(logging.DEBUG)
219
+
220
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
221
+ if not os.path.exists(model_dir):
222
+ os.makedirs(model_dir)
223
+ h = logging.FileHandler(os.path.join(model_dir, filename))
224
+ h.setLevel(logging.DEBUG)
225
+ h.setFormatter(formatter)
226
+ logger.addHandler(h)
227
+ return logger
228
+
229
+
230
+ class HParams():
231
+ def __init__(self, **kwargs):
232
+ for k, v in kwargs.items():
233
+ if type(v) == dict:
234
+ v = HParams(**v)
235
+ self[k] = v
236
+
237
+ def keys(self):
238
+ return self.__dict__.keys()
239
+
240
+ def items(self):
241
+ return self.__dict__.items()
242
+
243
+ def values(self):
244
+ return self.__dict__.values()
245
+
246
+ def __len__(self):
247
+ return len(self.__dict__)
248
+
249
+ def __getitem__(self, key):
250
+ return getattr(self, key)
251
+
252
+ def __setitem__(self, key, value):
253
+ return setattr(self, key, value)
254
+
255
+ def __contains__(self, key):
256
+ return key in self.__dict__
257
+
258
+ def __repr__(self):
259
+ return self.__dict__.__repr__()