xj commited on
Commit
29dc912
1 Parent(s): 978a56f

[feat] clone repo

Browse files
.gitignore ADDED
@@ -0,0 +1,382 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## Ignore Visual Studio temporary files, build results, and
2
+ ## files generated by popular Visual Studio add-ons.
3
+ ##
4
+ ## Get latest from https://github.com/github/gitignore/blob/master/VisualStudio.gitignore
5
+
6
+ # User-specific files
7
+ *.rsuser
8
+ *.suo
9
+ *.user
10
+ *.userosscache
11
+ *.sln.docstates
12
+
13
+ # User-specific files (MonoDevelop/Xamarin Studio)
14
+ *.userprefs
15
+
16
+ # Mono auto generated files
17
+ mono_crash.*
18
+
19
+ # Build results
20
+ [Dd]ebug/
21
+ [Dd]ebugPublic/
22
+ [Rr]elease/
23
+ [Rr]eleases/
24
+ x64/
25
+ x86/
26
+ [Ww][Ii][Nn]32/
27
+ [Aa][Rr][Mm]/
28
+ [Aa][Rr][Mm]64/
29
+ bld/
30
+ [Bb]in/
31
+ [Oo]bj/
32
+ [Oo]ut/
33
+ [Ll]og/
34
+ [Ll]ogs/
35
+
36
+ # Visual Studio 2015/2017 cache/options directory
37
+ .vs/
38
+ # Uncomment if you have tasks that create the project's static files in wwwroot
39
+ #wwwroot/
40
+
41
+ # Visual Studio 2017 auto generated files
42
+ Generated\ Files/
43
+
44
+ # MSTest test Results
45
+ [Tt]est[Rr]esult*/
46
+ [Bb]uild[Ll]og.*
47
+
48
+ # NUnit
49
+ *.VisualState.xml
50
+ TestResult.xml
51
+ nunit-*.xml
52
+
53
+ # Build Results of an ATL Project
54
+ [Dd]ebugPS/
55
+ [Rr]eleasePS/
56
+ dlldata.c
57
+
58
+ # Benchmark Results
59
+ BenchmarkDotNet.Artifacts/
60
+
61
+ # .NET Core
62
+ project.lock.json
63
+ project.fragment.lock.json
64
+ artifacts/
65
+
66
+ # ASP.NET Scaffolding
67
+ ScaffoldingReadMe.txt
68
+
69
+ # StyleCop
70
+ StyleCopReport.xml
71
+
72
+ # Files built by Visual Studio
73
+ *_i.c
74
+ *_p.c
75
+ *_h.h
76
+ *.ilk
77
+ *.meta
78
+ *.obj
79
+ *.iobj
80
+ *.pch
81
+ *.pdb
82
+ *.ipdb
83
+ *.pgc
84
+ *.pgd
85
+ *.rsp
86
+ *.sbr
87
+ *.tlb
88
+ *.tli
89
+ *.tlh
90
+ *.tmp
91
+ *.tmp_proj
92
+ *_wpftmp.csproj
93
+ *.log
94
+ *.vspscc
95
+ *.vssscc
96
+ .builds
97
+ *.pidb
98
+ *.svclog
99
+ *.scc
100
+
101
+ # Chutzpah Test files
102
+ _Chutzpah*
103
+
104
+ # Visual C++ cache files
105
+ ipch/
106
+ *.aps
107
+ *.ncb
108
+ *.opendb
109
+ *.opensdf
110
+ *.sdf
111
+ *.cachefile
112
+ *.VC.db
113
+ *.VC.VC.opendb
114
+
115
+ # Visual Studio profiler
116
+ *.psess
117
+ *.vsp
118
+ *.vspx
119
+ *.sap
120
+
121
+ # Visual Studio Trace Files
122
+ *.e2e
123
+
124
+ # TFS 2012 Local Workspace
125
+ $tf/
126
+
127
+ # Guidance Automation Toolkit
128
+ *.gpState
129
+
130
+ # ReSharper is a .NET coding add-in
131
+ _ReSharper*/
132
+ *.[Rr]e[Ss]harper
133
+ *.DotSettings.user
134
+
135
+ # TeamCity is a build add-in
136
+ _TeamCity*
137
+
138
+ # DotCover is a Code Coverage Tool
139
+ *.dotCover
140
+
141
+ # AxoCover is a Code Coverage Tool
142
+ .axoCover/*
143
+ !.axoCover/settings.json
144
+
145
+ # Coverlet is a free, cross platform Code Coverage Tool
146
+ coverage*.json
147
+ coverage*.xml
148
+ coverage*.info
149
+
150
+ # Visual Studio code coverage results
151
+ *.coverage
152
+ *.coveragexml
153
+
154
+ # NCrunch
155
+ _NCrunch_*
156
+ .*crunch*.local.xml
157
+ nCrunchTemp_*
158
+
159
+ # MightyMoose
160
+ *.mm.*
161
+ AutoTest.Net/
162
+
163
+ # Web workbench (sass)
164
+ .sass-cache/
165
+
166
+ # Installshield output folder
167
+ [Ee]xpress/
168
+
169
+ # DocProject is a documentation generator add-in
170
+ DocProject/buildhelp/
171
+ DocProject/Help/*.HxT
172
+ DocProject/Help/*.HxC
173
+ DocProject/Help/*.hhc
174
+ DocProject/Help/*.hhk
175
+ DocProject/Help/*.hhp
176
+ DocProject/Help/Html2
177
+ DocProject/Help/html
178
+
179
+ # Click-Once directory
180
+ publish/
181
+
182
+ # Publish Web Output
183
+ *.[Pp]ublish.xml
184
+ *.azurePubxml
185
+ # Note: Comment the next line if you want to checkin your web deploy settings,
186
+ # but database connection strings (with potential passwords) will be unencrypted
187
+ *.pubxml
188
+ *.publishproj
189
+
190
+ # Microsoft Azure Web App publish settings. Comment the next line if you want to
191
+ # checkin your Azure Web App publish settings, but sensitive information contained
192
+ # in these scripts will be unencrypted
193
+ PublishScripts/
194
+
195
+ # NuGet Packages
196
+ *.nupkg
197
+ # NuGet Symbol Packages
198
+ *.snupkg
199
+ # The packages folder can be ignored because of Package Restore
200
+ **/[Pp]ackages/*
201
+ # except build/, which is used as an MSBuild target.
202
+ !**/[Pp]ackages/build/
203
+ # Uncomment if necessary however generally it will be regenerated when needed
204
+ #!**/[Pp]ackages/repositories.config
205
+ # NuGet v3's project.json files produces more ignorable files
206
+ *.nuget.props
207
+ *.nuget.targets
208
+
209
+ # Microsoft Azure Build Output
210
+ csx/
211
+ *.build.csdef
212
+
213
+ # Microsoft Azure Emulator
214
+ ecf/
215
+ rcf/
216
+
217
+ # Windows Store app package directories and files
218
+ AppPackages/
219
+ BundleArtifacts/
220
+ Package.StoreAssociation.xml
221
+ _pkginfo.txt
222
+ *.appx
223
+ *.appxbundle
224
+ *.appxupload
225
+
226
+ # Visual Studio cache files
227
+ # files ending in .cache can be ignored
228
+ *.[Cc]ache
229
+ # but keep track of directories ending in .cache
230
+ !?*.[Cc]ache/
231
+
232
+ # Others
233
+ ClientBin/
234
+ ~$*
235
+ *~
236
+ *.dbmdl
237
+ *.dbproj.schemaview
238
+ *.jfm
239
+ *.pfx
240
+ *.publishsettings
241
+ orleans.codegen.cs
242
+
243
+ # Including strong name files can present a security risk
244
+ # (https://github.com/github/gitignore/pull/2483#issue-259490424)
245
+ #*.snk
246
+
247
+ # Since there are multiple workflows, uncomment next line to ignore bower_components
248
+ # (https://github.com/github/gitignore/pull/1529#issuecomment-104372622)
249
+ #bower_components/
250
+
251
+ # RIA/Silverlight projects
252
+ Generated_Code/
253
+
254
+ # Backup & report files from converting an old project file
255
+ # to a newer Visual Studio version. Backup files are not needed,
256
+ # because we have git ;-)
257
+ _UpgradeReport_Files/
258
+ Backup*/
259
+ UpgradeLog*.XML
260
+ UpgradeLog*.htm
261
+ ServiceFabricBackup/
262
+ *.rptproj.bak
263
+
264
+ # SQL Server files
265
+ *.mdf
266
+ *.ldf
267
+ *.ndf
268
+
269
+ # Business Intelligence projects
270
+ *.rdl.data
271
+ *.bim.layout
272
+ *.bim_*.settings
273
+ *.rptproj.rsuser
274
+ *- [Bb]ackup.rdl
275
+ *- [Bb]ackup ([0-9]).rdl
276
+ *- [Bb]ackup ([0-9][0-9]).rdl
277
+
278
+ # Microsoft Fakes
279
+ FakesAssemblies/
280
+
281
+ # GhostDoc plugin setting file
282
+ *.GhostDoc.xml
283
+
284
+ # Node.js Tools for Visual Studio
285
+ .ntvs_analysis.dat
286
+ node_modules/
287
+
288
+ # Visual Studio 6 build log
289
+ *.plg
290
+
291
+ # Visual Studio 6 workspace options file
292
+ *.opt
293
+
294
+ # Visual Studio 6 auto-generated workspace file (contains which files were open etc.)
295
+ *.vbw
296
+
297
+ # Visual Studio LightSwitch build output
298
+ **/*.HTMLClient/GeneratedArtifacts
299
+ **/*.DesktopClient/GeneratedArtifacts
300
+ **/*.DesktopClient/ModelManifest.xml
301
+ **/*.Server/GeneratedArtifacts
302
+ **/*.Server/ModelManifest.xml
303
+ _Pvt_Extensions
304
+
305
+ # Paket dependency manager
306
+ .paket/paket.exe
307
+ paket-files/
308
+
309
+ # FAKE - F# Make
310
+ .fake/
311
+
312
+ # CodeRush personal settings
313
+ .cr/personal
314
+
315
+ # Python Tools for Visual Studio (PTVS)
316
+ __pycache__/
317
+
318
+
319
+ # Cake - Uncomment if you are using it
320
+ # tools/**
321
+ # !tools/packages.config
322
+
323
+ # Tabs Studio
324
+ *.tss
325
+
326
+ # Telerik's JustMock configuration file
327
+ *.jmconfig
328
+
329
+ # BizTalk build output
330
+ *.btp.cs
331
+ *.btm.cs
332
+ *.odx.cs
333
+ *.xsd.cs
334
+
335
+ # OpenCover UI analysis results
336
+ OpenCover/
337
+
338
+ # Azure Stream Analytics local run output
339
+ ASALocalRun/
340
+
341
+ # MSBuild Binary and Structured Log
342
+ *.binlog
343
+
344
+ # NVidia Nsight GPU debugger configuration file
345
+ *.nvuser
346
+
347
+ # MFractors (Xamarin productivity tool) working folder
348
+ .mfractor/
349
+
350
+ # Local History for Visual Studio
351
+ .localhistory/
352
+
353
+ # BeatPulse healthcheck temp database
354
+ healthchecksdb
355
+
356
+ # Backup folder for Package Reference Convert tool in Visual Studio 2017
357
+ MigrationBackup/
358
+
359
+ # Ionide (cross platform F# VS Code tools) working folder
360
+ .ionide/
361
+
362
+ # Fody - auto-generated XML schema
363
+ FodyWeavers.xsd
364
+
365
+ # build
366
+ build
367
+ monotonic_align/core.c
368
+ *.o
369
+ *.so
370
+ *.dll
371
+
372
+ # data
373
+ /config.json
374
+ /*.pth
375
+ *.wav
376
+ /monotonic_align/monotonic_align
377
+ /resources
378
+ /MoeGoe.spec
379
+ /dist/MoeGoe
380
+ /dist
381
+
382
+ .idea
README copy.md ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ title: ' vits-uma-genshin-honkai'
4
+ sdk: gradio
5
+ sdk_version: 3.7
6
+ emoji: 🐨
7
+ colorTo: yellow
8
+ pinned: false
9
+ app_file: app.py
10
+ ---
app.py ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ import time
3
+ import os
4
+ import gradio as gr
5
+ import utils
6
+ import argparse
7
+ import commons
8
+ from models import SynthesizerTrn
9
+ from text import text_to_sequence
10
+ import torch
11
+ from torch import no_grad, LongTensor
12
+ import webbrowser
13
+ import logging
14
+ import gradio.processing_utils as gr_processing_utils
15
+ logging.getLogger('numba').setLevel(logging.WARNING)
16
+ limitation = os.getenv("SYSTEM") == "spaces" # limit text and audio length in huggingface spaces
17
+
18
+ audio_postprocess_ori = gr.Audio.postprocess
19
+ def audio_postprocess(self, y):
20
+ data = audio_postprocess_ori(self, y)
21
+ if data is None:
22
+ return None
23
+ return gr_processing_utils.encode_url_or_file_to_base64(data["name"])
24
+ gr.Audio.postprocess = audio_postprocess
25
+
26
+ def get_text(text, hps):
27
+ text_norm, clean_text = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
28
+ if hps.data.add_blank:
29
+ text_norm = commons.intersperse(text_norm, 0)
30
+ text_norm = LongTensor(text_norm)
31
+ return text_norm, clean_text
32
+
33
+ def vits(text, language, speaker_id, noise_scale, noise_scale_w, length_scale):
34
+ start = time.perf_counter()
35
+ if not len(text):
36
+ return "输入文本不能为空!", None, None
37
+ text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
38
+ if len(text) > 100 and limitation:
39
+ return f"输入文字过长!{len(text)}>100", None, None
40
+ if language == 0:
41
+ text = f"[ZH]{text}[ZH]"
42
+ elif language == 1:
43
+ text = f"[JA]{text}[JA]"
44
+ else:
45
+ text = f"{text}"
46
+ stn_tst, clean_text = get_text(text, hps_ms)
47
+ with no_grad():
48
+ x_tst = stn_tst.unsqueeze(0).to(device)
49
+ x_tst_lengths = LongTensor([stn_tst.size(0)]).to(device)
50
+ speaker_id = LongTensor([speaker_id]).to(device)
51
+ audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=speaker_id, noise_scale=noise_scale, noise_scale_w=noise_scale_w,
52
+ length_scale=length_scale)[0][0, 0].data.cpu().float().numpy()
53
+
54
+ return "生成成功!", (22050, audio), f"生成耗时 {round(time.perf_counter()-start, 2)} s"
55
+
56
+ def search_speaker(search_value):
57
+ for s in speakers:
58
+ if search_value == s:
59
+ return s
60
+ for s in speakers:
61
+ if search_value in s:
62
+ return s
63
+
64
+ def change_lang(language):
65
+ if language == 0:
66
+ return 0.6, 0.668, 1.2
67
+ else:
68
+ return 0.6, 0.668, 1.1
69
+
70
+ download_audio_js = """
71
+ () =>{{
72
+ let root = document.querySelector("body > gradio-app");
73
+ if (root.shadowRoot != null)
74
+ root = root.shadowRoot;
75
+ let audio = root.querySelector("#tts-audio").querySelector("audio");
76
+ let text = root.querySelector("#input-text").querySelector("textarea");
77
+ if (audio == undefined)
78
+ return;
79
+ text = text.value;
80
+ if (text == undefined)
81
+ text = Math.floor(Math.random()*100000000);
82
+ audio = audio.src;
83
+ let oA = document.createElement("a");
84
+ oA.download = text.substr(0, 20)+'.wav';
85
+ oA.href = audio;
86
+ document.body.appendChild(oA);
87
+ oA.click();
88
+ oA.remove();
89
+ }}
90
+ """
91
+
92
+ if __name__ == '__main__':
93
+ parser = argparse.ArgumentParser()
94
+ parser.add_argument('--device', type=str, default='cpu')
95
+ parser.add_argument('--api', action="store_true", default=False)
96
+ parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
97
+ parser.add_argument("--colab", action="store_true", default=False, help="share gradio app")
98
+ args = parser.parse_args()
99
+ device = torch.device(args.device)
100
+
101
+ hps_ms = utils.get_hparams_from_file(r'./model/config.json')
102
+ net_g_ms = SynthesizerTrn(
103
+ len(hps_ms.symbols),
104
+ hps_ms.data.filter_length // 2 + 1,
105
+ hps_ms.train.segment_size // hps_ms.data.hop_length,
106
+ n_speakers=hps_ms.data.n_speakers,
107
+ **hps_ms.model)
108
+ _ = net_g_ms.eval().to(device)
109
+ speakers = hps_ms.speakers
110
+ model, optimizer, learning_rate, epochs = utils.load_checkpoint(r'./model/G_953000.pth', net_g_ms, None)
111
+
112
+ with gr.Blocks() as app:
113
+ gr.Markdown(
114
+ "# <center> VITS语音在线合成demo\n"
115
+ )
116
+
117
+ with gr.Tabs():
118
+ with gr.TabItem("vits"):
119
+ with gr.Row():
120
+ with gr.Column():
121
+ input_text = gr.Textbox(label="Text (100 words limitation) " if limitation else "Text", lines=5, value="今天晚上吃啥好呢。", elem_id=f"input-text")
122
+ lang = gr.Dropdown(label="Language", choices=["中文", "日语", "中日混合(中文用[ZH][ZH]包裹起来,日文用[JA][JA]包裹起来)"],
123
+ type="index", value="中文")
124
+ btn = gr.Button(value="Submit")
125
+ with gr.Row():
126
+ search = gr.Textbox(label="Search Speaker", lines=1)
127
+ btn2 = gr.Button(value="Search")
128
+ sid = gr.Dropdown(label="Speaker", choices=speakers, type="index", value=speakers[228])
129
+ with gr.Row():
130
+ ns = gr.Slider(label="noise_scale(控制感情变化程度)", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
131
+ nsw = gr.Slider(label="noise_scale_w(控制音素发音长度)", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
132
+ ls = gr.Slider(label="length_scale(控制整体语速)", minimum=0.1, maximum=2.0, step=0.1, value=1.2, interactive=True)
133
+ with gr.Column():
134
+ o1 = gr.Textbox(label="Output Message")
135
+ o2 = gr.Audio(label="Output Audio", elem_id=f"tts-audio")
136
+ o3 = gr.Textbox(label="Extra Info")
137
+ download = gr.Button("Download Audio")
138
+ btn.click(vits, inputs=[input_text, lang, sid, ns, nsw, ls], outputs=[o1, o2, o3])
139
+ download.click(None, [], [], _js=download_audio_js.format())
140
+ btn2.click(search_speaker, inputs=[search], outputs=[sid])
141
+ lang.change(change_lang, inputs=[lang], outputs=[ns, nsw, ls])
142
+ with gr.TabItem("可用人物一览"):
143
+ gr.Radio(label="Speaker", choices=speakers, interactive=False, type="index")
144
+ if args.colab:
145
+ webbrowser.open("http://127.0.0.1:7860")
146
+ app.queue(concurrency_count=1, api_open=args.api).launch(share=args.share)
attentions.py ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ from modules import LayerNorm
8
+
9
+
10
+ class Encoder(nn.Module):
11
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
12
+ super().__init__()
13
+ self.hidden_channels = hidden_channels
14
+ self.filter_channels = filter_channels
15
+ self.n_heads = n_heads
16
+ self.n_layers = n_layers
17
+ self.kernel_size = kernel_size
18
+ self.p_dropout = p_dropout
19
+ self.window_size = window_size
20
+
21
+ self.drop = nn.Dropout(p_dropout)
22
+ self.attn_layers = nn.ModuleList()
23
+ self.norm_layers_1 = nn.ModuleList()
24
+ self.ffn_layers = nn.ModuleList()
25
+ self.norm_layers_2 = nn.ModuleList()
26
+ for i in range(self.n_layers):
27
+ self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
28
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
29
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
30
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
31
+
32
+ def forward(self, x, x_mask):
33
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
34
+ x = x * x_mask
35
+ for i in range(self.n_layers):
36
+ y = self.attn_layers[i](x, x, attn_mask)
37
+ y = self.drop(y)
38
+ x = self.norm_layers_1[i](x + y)
39
+
40
+ y = self.ffn_layers[i](x, x_mask)
41
+ y = self.drop(y)
42
+ x = self.norm_layers_2[i](x + y)
43
+ x = x * x_mask
44
+ return x
45
+
46
+
47
+ class Decoder(nn.Module):
48
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
49
+ super().__init__()
50
+ self.hidden_channels = hidden_channels
51
+ self.filter_channels = filter_channels
52
+ self.n_heads = n_heads
53
+ self.n_layers = n_layers
54
+ self.kernel_size = kernel_size
55
+ self.p_dropout = p_dropout
56
+ self.proximal_bias = proximal_bias
57
+ self.proximal_init = proximal_init
58
+
59
+ self.drop = nn.Dropout(p_dropout)
60
+ self.self_attn_layers = nn.ModuleList()
61
+ self.norm_layers_0 = nn.ModuleList()
62
+ self.encdec_attn_layers = nn.ModuleList()
63
+ self.norm_layers_1 = nn.ModuleList()
64
+ self.ffn_layers = nn.ModuleList()
65
+ self.norm_layers_2 = nn.ModuleList()
66
+ for i in range(self.n_layers):
67
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
68
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
69
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
70
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
71
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
72
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
73
+
74
+ def forward(self, x, x_mask, h, h_mask):
75
+ """
76
+ x: decoder input
77
+ h: encoder output
78
+ """
79
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
80
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
81
+ x = x * x_mask
82
+ for i in range(self.n_layers):
83
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
84
+ y = self.drop(y)
85
+ x = self.norm_layers_0[i](x + y)
86
+
87
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
88
+ y = self.drop(y)
89
+ x = self.norm_layers_1[i](x + y)
90
+
91
+ y = self.ffn_layers[i](x, x_mask)
92
+ y = self.drop(y)
93
+ x = self.norm_layers_2[i](x + y)
94
+ x = x * x_mask
95
+ return x
96
+
97
+
98
+ class MultiHeadAttention(nn.Module):
99
+ def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
100
+ super().__init__()
101
+ assert channels % n_heads == 0
102
+
103
+ self.channels = channels
104
+ self.out_channels = out_channels
105
+ self.n_heads = n_heads
106
+ self.p_dropout = p_dropout
107
+ self.window_size = window_size
108
+ self.heads_share = heads_share
109
+ self.block_length = block_length
110
+ self.proximal_bias = proximal_bias
111
+ self.proximal_init = proximal_init
112
+ self.attn = None
113
+
114
+ self.k_channels = channels // n_heads
115
+ self.conv_q = nn.Conv1d(channels, channels, 1)
116
+ self.conv_k = nn.Conv1d(channels, channels, 1)
117
+ self.conv_v = nn.Conv1d(channels, channels, 1)
118
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
119
+ self.drop = nn.Dropout(p_dropout)
120
+
121
+ if window_size is not None:
122
+ n_heads_rel = 1 if heads_share else n_heads
123
+ rel_stddev = self.k_channels**-0.5
124
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
125
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
126
+
127
+ nn.init.xavier_uniform_(self.conv_q.weight)
128
+ nn.init.xavier_uniform_(self.conv_k.weight)
129
+ nn.init.xavier_uniform_(self.conv_v.weight)
130
+ if proximal_init:
131
+ with torch.no_grad():
132
+ self.conv_k.weight.copy_(self.conv_q.weight)
133
+ self.conv_k.bias.copy_(self.conv_q.bias)
134
+
135
+ def forward(self, x, c, attn_mask=None):
136
+ q = self.conv_q(x)
137
+ k = self.conv_k(c)
138
+ v = self.conv_v(c)
139
+
140
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
141
+
142
+ x = self.conv_o(x)
143
+ return x
144
+
145
+ def attention(self, query, key, value, mask=None):
146
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
147
+ b, d, t_s, t_t = (*key.size(), query.size(2))
148
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
149
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
150
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
151
+
152
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
153
+ if self.window_size is not None:
154
+ assert t_s == t_t, "Relative attention is only available for self-attention."
155
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
156
+ rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
157
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
158
+ scores = scores + scores_local
159
+ if self.proximal_bias:
160
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
161
+ scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
162
+ if mask is not None:
163
+ scores = scores.masked_fill(mask == 0, -1e4)
164
+ if self.block_length is not None:
165
+ assert t_s == t_t, "Local attention is only available for self-attention."
166
+ block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
167
+ scores = scores.masked_fill(block_mask == 0, -1e4)
168
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
169
+ p_attn = self.drop(p_attn)
170
+ output = torch.matmul(p_attn, value)
171
+ if self.window_size is not None:
172
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
173
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
174
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
175
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
176
+ return output, p_attn
177
+
178
+ def _matmul_with_relative_values(self, x, y):
179
+ """
180
+ x: [b, h, l, m]
181
+ y: [h or 1, m, d]
182
+ ret: [b, h, l, d]
183
+ """
184
+ ret = torch.matmul(x, y.unsqueeze(0))
185
+ return ret
186
+
187
+ def _matmul_with_relative_keys(self, x, y):
188
+ """
189
+ x: [b, h, l, d]
190
+ y: [h or 1, m, d]
191
+ ret: [b, h, l, m]
192
+ """
193
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
194
+ return ret
195
+
196
+ def _get_relative_embeddings(self, relative_embeddings, length):
197
+ max_relative_position = 2 * self.window_size + 1
198
+ # Pad first before slice to avoid using cond ops.
199
+ pad_length = max(length - (self.window_size + 1), 0)
200
+ slice_start_position = max((self.window_size + 1) - length, 0)
201
+ slice_end_position = slice_start_position + 2 * length - 1
202
+ if pad_length > 0:
203
+ padded_relative_embeddings = F.pad(
204
+ relative_embeddings,
205
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
206
+ else:
207
+ padded_relative_embeddings = relative_embeddings
208
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
209
+ return used_relative_embeddings
210
+
211
+ def _relative_position_to_absolute_position(self, x):
212
+ """
213
+ x: [b, h, l, 2*l-1]
214
+ ret: [b, h, l, l]
215
+ """
216
+ batch, heads, length, _ = x.size()
217
+ # Concat columns of pad to shift from relative to absolute indexing.
218
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
219
+
220
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
221
+ x_flat = x.view([batch, heads, length * 2 * length])
222
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
223
+
224
+ # Reshape and slice out the padded elements.
225
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
226
+ return x_final
227
+
228
+ def _absolute_position_to_relative_position(self, x):
229
+ """
230
+ x: [b, h, l, l]
231
+ ret: [b, h, l, 2*l-1]
232
+ """
233
+ batch, heads, length, _ = x.size()
234
+ # padd along column
235
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
236
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
237
+ # add 0's in the beginning that will skew the elements after reshape
238
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
239
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
240
+ return x_final
241
+
242
+ def _attention_bias_proximal(self, length):
243
+ """Bias for self-attention to encourage attention to close positions.
244
+ Args:
245
+ length: an integer scalar.
246
+ Returns:
247
+ a Tensor with shape [1, 1, length, length]
248
+ """
249
+ r = torch.arange(length, dtype=torch.float32)
250
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
251
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
252
+
253
+
254
+ class FFN(nn.Module):
255
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
256
+ super().__init__()
257
+ self.in_channels = in_channels
258
+ self.out_channels = out_channels
259
+ self.filter_channels = filter_channels
260
+ self.kernel_size = kernel_size
261
+ self.p_dropout = p_dropout
262
+ self.activation = activation
263
+ self.causal = causal
264
+
265
+ if causal:
266
+ self.padding = self._causal_padding
267
+ else:
268
+ self.padding = self._same_padding
269
+
270
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
271
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
272
+ self.drop = nn.Dropout(p_dropout)
273
+
274
+ def forward(self, x, x_mask):
275
+ x = self.conv_1(self.padding(x * x_mask))
276
+ if self.activation == "gelu":
277
+ x = x * torch.sigmoid(1.702 * x)
278
+ else:
279
+ x = torch.relu(x)
280
+ x = self.drop(x)
281
+ x = self.conv_2(self.padding(x * x_mask))
282
+ return x * x_mask
283
+
284
+ def _causal_padding(self, x):
285
+ if self.kernel_size == 1:
286
+ return x
287
+ pad_l = self.kernel_size - 1
288
+ pad_r = 0
289
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
290
+ x = F.pad(x, commons.convert_pad_shape(padding))
291
+ return x
292
+
293
+ def _same_padding(self, x):
294
+ if self.kernel_size == 1:
295
+ return x
296
+ pad_l = (self.kernel_size - 1) // 2
297
+ pad_r = self.kernel_size // 2
298
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
299
+ x = F.pad(x, commons.convert_pad_shape(padding))
300
+ return x
commons.py ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch.nn import functional as F
4
+ import torch.jit
5
+
6
+
7
+ def script_method(fn, _rcb=None):
8
+ return fn
9
+
10
+
11
+ def script(obj, optimize=True, _frames_up=0, _rcb=None):
12
+ return obj
13
+
14
+
15
+ torch.jit.script_method = script_method
16
+ torch.jit.script = script
17
+
18
+
19
+ def init_weights(m, mean=0.0, std=0.01):
20
+ classname = m.__class__.__name__
21
+ if classname.find("Conv") != -1:
22
+ m.weight.data.normal_(mean, std)
23
+
24
+
25
+ def get_padding(kernel_size, dilation=1):
26
+ return int((kernel_size*dilation - dilation)/2)
27
+
28
+
29
+ def convert_pad_shape(pad_shape):
30
+ l = pad_shape[::-1]
31
+ pad_shape = [item for sublist in l for item in sublist]
32
+ return pad_shape
33
+
34
+
35
+ def intersperse(lst, item):
36
+ result = [item] * (len(lst) * 2 + 1)
37
+ result[1::2] = lst
38
+ return result
39
+
40
+
41
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
42
+ """KL(P||Q)"""
43
+ kl = (logs_q - logs_p) - 0.5
44
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
45
+ return kl
46
+
47
+
48
+ def rand_gumbel(shape):
49
+ """Sample from the Gumbel distribution, protect from overflows."""
50
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
51
+ return -torch.log(-torch.log(uniform_samples))
52
+
53
+
54
+ def rand_gumbel_like(x):
55
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
56
+ return g
57
+
58
+
59
+ def slice_segments(x, ids_str, segment_size=4):
60
+ ret = torch.zeros_like(x[:, :, :segment_size])
61
+ for i in range(x.size(0)):
62
+ idx_str = ids_str[i]
63
+ idx_end = idx_str + segment_size
64
+ ret[i] = x[i, :, idx_str:idx_end]
65
+ return ret
66
+
67
+
68
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
69
+ b, d, t = x.size()
70
+ if x_lengths is None:
71
+ x_lengths = t
72
+ ids_str_max = x_lengths - segment_size + 1
73
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
74
+ ret = slice_segments(x, ids_str, segment_size)
75
+ return ret, ids_str
76
+
77
+
78
+ def get_timing_signal_1d(
79
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
80
+ position = torch.arange(length, dtype=torch.float)
81
+ num_timescales = channels // 2
82
+ log_timescale_increment = (
83
+ math.log(float(max_timescale) / float(min_timescale)) /
84
+ (num_timescales - 1))
85
+ inv_timescales = min_timescale * torch.exp(
86
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
87
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
88
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
89
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
90
+ signal = signal.view(1, channels, length)
91
+ return signal
92
+
93
+
94
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
95
+ b, channels, length = x.size()
96
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
97
+ return x + signal.to(dtype=x.dtype, device=x.device)
98
+
99
+
100
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
101
+ b, channels, length = x.size()
102
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
103
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
104
+
105
+
106
+ def subsequent_mask(length):
107
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
108
+ return mask
109
+
110
+
111
+ @torch.jit.script
112
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
113
+ n_channels_int = n_channels[0]
114
+ in_act = input_a + input_b
115
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
116
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
117
+ acts = t_act * s_act
118
+ return acts
119
+
120
+
121
+ def convert_pad_shape(pad_shape):
122
+ l = pad_shape[::-1]
123
+ pad_shape = [item for sublist in l for item in sublist]
124
+ return pad_shape
125
+
126
+
127
+ def shift_1d(x):
128
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
129
+ return x
130
+
131
+
132
+ def sequence_mask(length, max_length=None):
133
+ if max_length is None:
134
+ max_length = length.max()
135
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
136
+ return x.unsqueeze(0) < length.unsqueeze(1)
137
+
138
+
139
+ def generate_path(duration, mask):
140
+ """
141
+ duration: [b, 1, t_x]
142
+ mask: [b, 1, t_y, t_x]
143
+ """
144
+ device = duration.device
145
+
146
+ b, _, t_y, t_x = mask.shape
147
+ cum_duration = torch.cumsum(duration, -1)
148
+
149
+ cum_duration_flat = cum_duration.view(b * t_x)
150
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
151
+ path = path.view(b, t_x, t_y)
152
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
153
+ path = path.unsqueeze(1).transpose(2,3) * mask
154
+ return path
155
+
156
+
157
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
158
+ if isinstance(parameters, torch.Tensor):
159
+ parameters = [parameters]
160
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
161
+ norm_type = float(norm_type)
162
+ if clip_value is not None:
163
+ clip_value = float(clip_value)
164
+
165
+ total_norm = 0
166
+ for p in parameters:
167
+ param_norm = p.grad.data.norm(norm_type)
168
+ total_norm += param_norm.item() ** norm_type
169
+ if clip_value is not None:
170
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
171
+ total_norm = total_norm ** (1. / norm_type)
172
+ return total_norm
mel_processing.py ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.utils.data
3
+ from librosa.filters import mel as librosa_mel_fn
4
+
5
+ MAX_WAV_VALUE = 32768.0
6
+
7
+
8
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
9
+ """
10
+ PARAMS
11
+ ------
12
+ C: compression factor
13
+ """
14
+ return torch.log(torch.clamp(x, min=clip_val) * C)
15
+
16
+
17
+ def dynamic_range_decompression_torch(x, C=1):
18
+ """
19
+ PARAMS
20
+ ------
21
+ C: compression factor used to compress
22
+ """
23
+ return torch.exp(x) / C
24
+
25
+
26
+ def spectral_normalize_torch(magnitudes):
27
+ output = dynamic_range_compression_torch(magnitudes)
28
+ return output
29
+
30
+
31
+ def spectral_de_normalize_torch(magnitudes):
32
+ output = dynamic_range_decompression_torch(magnitudes)
33
+ return output
34
+
35
+
36
+ mel_basis = {}
37
+ hann_window = {}
38
+
39
+
40
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
41
+ if torch.min(y) < -1.:
42
+ print('min value is ', torch.min(y))
43
+ if torch.max(y) > 1.:
44
+ print('max value is ', torch.max(y))
45
+
46
+ global hann_window
47
+ dtype_device = str(y.dtype) + '_' + str(y.device)
48
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
49
+ if wnsize_dtype_device not in hann_window:
50
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
51
+
52
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
53
+ y = y.squeeze(1)
54
+
55
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
56
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
57
+
58
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
59
+ return spec
60
+
61
+
62
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
63
+ global mel_basis
64
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
65
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
66
+ if fmax_dtype_device not in mel_basis:
67
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
68
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
69
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
70
+ spec = spectral_normalize_torch(spec)
71
+ return spec
72
+
73
+
74
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
75
+ if torch.min(y) < -1.:
76
+ print('min value is ', torch.min(y))
77
+ if torch.max(y) > 1.:
78
+ print('max value is ', torch.max(y))
79
+
80
+ global mel_basis, hann_window
81
+ dtype_device = str(y.dtype) + '_' + str(y.device)
82
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
83
+ wnsize_dtype_device = str(win_size) + '_' + 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=y.dtype, device=y.device)
87
+ if wnsize_dtype_device not in hann_window:
88
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
89
+
90
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
91
+ y = y.squeeze(1)
92
+
93
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
94
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
95
+
96
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
97
+
98
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
99
+ spec = spectral_normalize_torch(spec)
100
+
101
+ return spec
model/G_953000.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c54396a7a9027952e4d72fceb7e1da1f003d108837138927c2054a33eda0292
3
+ size 479276657
model/config.json ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "train": {
3
+ "log_interval": 200,
4
+ "eval_interval": 1000,
5
+ "seed": 1234,
6
+ "epochs": 10000,
7
+ "learning_rate": 2e-4,
8
+ "betas": [0.8, 0.99],
9
+ "eps": 1e-9,
10
+ "batch_size": 64,
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":"filelists/uma_genshin_genshinjp_bh3_train.txt.cleaned",
21
+ "validation_files":"filelists/uma_genshin_genshinjp_bh3_val.txt.cleaned",
22
+ "text_cleaners":["zh_ja_mixture_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": 804,
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
+ "speakers": ["\u7279\u522b\u5468", "\u65e0\u58f0\u94c3\u9e7f", "\u4e1c\u6d77\u5e1d\u7687\uff08\u5e1d\u5b9d\uff0c\u5e1d\u738b\uff09", "\u4e38\u5584\u65af\u57fa", "\u5bcc\u58eb\u5947\u8ff9", "\u5c0f\u6817\u5e3d", "\u9ec4\u91d1\u8239", "\u4f0f\u7279\u52a0", "\u5927\u548c\u8d64\u9aa5", "\u5927\u6811\u5feb\u8f66", "\u8349\u4e0a\u98de", "\u83f1\u4e9a\u9a6c\u900a", "\u76ee\u767d\u9ea6\u6606", "\u795e\u9e70", "\u597d\u6b4c\u5267", "\u6210\u7530\u767d\u4ec1", "\u9c81\u9053\u592b\u8c61\u5f81\uff08\u7687\u5e1d\uff09", "\u6c14\u69fd", "\u7231\u4e3d\u6570\u7801", "\u661f\u4e91\u5929\u7a7a", "\u7389\u85fb\u5341\u5b57", "\u7f8e\u5999\u59ff\u52bf", "\u7435\u7436\u6668\u5149", "\u6469\u8036\u91cd\u70ae", "\u66fc\u57ce\u8336\u5ea7", "\u7f8e\u6d66\u6ce2\u65c1", "\u76ee\u767d\u8d56\u6069", "\u83f1\u66d9", "\u96ea\u4e2d\u7f8e\u4eba", "\u7c73\u6d74", "\u827e\u5c3c\u65af\u98ce\u795e", "\u7231\u4e3d\u901f\u5b50\uff08\u7231\u4e3d\u5feb\u5b50\uff09", "\u7231\u6155\u7ec7\u59ec", "\u7a3b\u8377\u4e00", "\u80dc\u5229\u5956\u5238", "\u7a7a\u4e2d\u795e\u5bab", "\u8363\u8fdb\u95ea\u8000", "\u771f\u673a\u4f36", "\u5ddd\u4e0a\u516c\u4e3b", "\u9ec4\u91d1\u57ce\uff08\u9ec4\u91d1\u57ce\u5e02\uff09", "\u6a31\u82b1\u8fdb\u738b", "\u91c7\u73e0", "\u65b0\u5149\u98ce", "\u4e1c\u5546\u53d8\u9769", "\u8d85\u7ea7\u5c0f\u6d77\u6e7e", "\u9192\u76ee\u98de\u9e70\uff08\u5bc4\u5bc4\u5b50\uff09", "\u8352\u6f20\u82f1\u96c4", "\u4e1c\u701b\u4f50\u6566", "\u4e2d\u5c71\u5e86\u5178", "\u6210\u7530\u5927\u8fdb", "\u897f\u91ce\u82b1", "\u6625\u4e3d\uff08\u4e4c\u62c9\u62c9\uff09", "\u9752\u7af9\u56de\u5fc6", "\u5fae\u5149\u98de\u9a79", "\u7f8e\u4e3d\u5468\u65e5", "\u5f85\u517c\u798f\u6765", "mr cb\uff08cb\u5148\u751f\uff09", "\u540d\u5c06\u6012\u6d9b\uff08\u540d\u5c06\u6237\u4ec1\uff09", "\u76ee\u767d\u591a\u4f2f", "\u4f18\u79c0\u7d20\u8d28", "\u5e1d\u738b\u5149\u8f89", "\u5f85\u517c\u8bd7\u6b4c\u5267", "\u751f\u91ce\u72c4\u675c\u65af", "\u76ee\u767d\u5584\u4fe1", "\u5927\u62d3\u592a\u9633\u795e", "\u53cc\u6da1\u8f6e\uff08\u4e24\u7acb\u76f4\uff0c\u4e24\u55b7\u5c04\uff0c\u4e8c\u9505\u5934\uff0c\u9006\u55b7\u5c04\uff09", "\u91cc\u89c1\u5149\u94bb\uff08\u8428\u6258\u8bfa\u91d1\u521a\u77f3\uff09", "\u5317\u90e8\u7384\u9a79", "\u6a31\u82b1\u5343\u4ee3\u738b", "\u5929\u72fc\u661f\u8c61\u5f81", "\u76ee\u767d\u963f\u5c14\u4e39", "\u516b\u91cd\u65e0\u654c", "\u9e64\u4e38\u521a\u5fd7", "\u76ee\u767d\u5149\u660e", "\u6210\u7530\u62dc\u4ec1\uff08\u6210\u7530\u8def\uff09", "\u4e5f\u6587\u6444\u8f89", "\u5c0f\u6797\u5386\u5947", "\u5317\u6e2f\u706b\u5c71", "\u5947\u9510\u9a8f", "\u82e6\u6da9\u7cd6\u971c", "\u5c0f\u5c0f\u8695\u8327", "\u9a8f\u5ddd\u624b\u7eb2\uff08\u7eff\u5e3d\u6076\u9b54\uff09", "\u79cb\u5ddd\u5f25\u751f\uff08\u5c0f\u5c0f\u7406\u4e8b\u957f\uff09", "\u4e59\u540d\u53f2\u60a6\u5b50\uff08\u4e59\u540d\u8bb0\u8005\uff09", "\u6850\u751f\u9662\u8475", "\u5b89\u5fc3\u6cfd\u523a\u523a\u7f8e", "\u6a2b\u672c\u7406\u5b50", "\u795e\u91cc\u7eeb\u534e\uff08\u9f9f\u9f9f\uff09", "\u7434", "\u7a7a\uff08\u7a7a\u54e5\uff09", "\u4e3d\u838e", "\u8367\uff08\u8367\u59b9\uff09", "\u82ad\u82ad\u62c9", "\u51ef\u4e9a", "\u8fea\u5362\u514b", "\u96f7\u6cfd", "\u5b89\u67cf", "\u6e29\u8fea", "\u9999\u83f1", "\u5317\u6597", "\u884c\u79cb", "\u9b48", "\u51dd\u5149", "\u53ef\u8389", "\u949f\u79bb", "\u83f2\u8c22\u5c14\uff08\u7687\u5973\uff09", "\u73ed\u5c3c\u7279", "\u8fbe\u8fbe\u5229\u4e9a\uff08\u516c\u5b50\uff09", "\u8bfa\u827e\u5c14\uff08\u5973\u4ec6\uff09", "\u4e03\u4e03", "\u91cd\u4e91", "\u7518\u96e8\uff08\u6930\u7f8a\uff09", "\u963f\u8d1d\u591a", "\u8fea\u5965\u5a1c\uff08\u732b\u732b\uff09", "\u83ab\u5a1c", "\u523b\u6674", "\u7802\u7cd6", "\u8f9b\u7131", "\u7f57\u838e\u8389\u4e9a", "\u80e1\u6843", "\u67ab\u539f\u4e07\u53f6\uff08\u4e07\u53f6\uff09", "\u70df\u7eef", "\u5bb5\u5bab", "\u6258\u9a6c", "\u4f18\u83c8", "\u96f7\u7535\u5c06\u519b\uff08\u96f7\u795e\uff09", "\u65e9\u67da", "\u73ca\u745a\u5bab\u5fc3\u6d77\uff08\u5fc3\u6d77\uff0c\u6263\u6263\u7c73\uff09", "\u4e94\u90ce", "\u4e5d\u6761\u88df\u7f57", "\u8352\u6cf7\u4e00\u6597\uff08\u4e00\u6597\uff09", "\u57c3\u6d1b\u4f0a", "\u7533\u9e64", "\u516b\u91cd\u795e\u5b50\uff08\u795e\u5b50\uff09", "\u795e\u91cc\u7eeb\u4eba\uff08\u7eeb\u4eba\uff09", "\u591c\u5170", "\u4e45\u5c90\u5fcd", "\u9e7f\u91ce\u82d1\u5e73\u85cf", "\u63d0\u7eb3\u91cc", "\u67ef\u83b1", "\u591a\u8389", "\u4e91\u5807", "\u7eb3\u897f\u59b2\uff08\u8349\u795e\uff09", "\u6df1\u6e0a\u4f7f\u5f92", "\u59ae\u9732", "\u8d5b\u8bfa", "\u503a\u52a1\u5904\u7406\u4eba", "\u574e\u8482\u4e1d", "\u771f\u5f13\u5feb\u8f66", "\u79cb\u4eba", "\u671b\u65cf", "\u827e\u5c14\u83f2", "\u827e\u8389\u4e1d", "\u827e\u4f26", "\u963f\u6d1b\u74e6", "\u5929\u91ce", "\u5929\u76ee\u5341\u4e94", "\u611a\u4eba\u4f17-\u5b89\u5fb7\u70c8", "\u5b89\u987a", "\u5b89\u897f", "\u8475", "\u9752\u6728", "\u8352\u5ddd\u5e78\u6b21", "\u8352\u8c37", "\u6709\u6cfd", "\u6d45\u5ddd", "\u9ebb\u7f8e", "\u51dd\u5149\u52a9\u624b", "\u963f\u6258", "\u7afa\u5b50", "\u767e\u8bc6", "\u767e\u95fb", "\u767e\u6653", "\u767d\u672f", "\u8d1d\u96c5\u7279\u4e3d\u5947", "\u4e3d\u5854", "\u5931\u843d\u8ff7\u8fed", "\u7f2d\u4e71\u661f\u68d8", "\u4f0a\u7538", "\u4f0f\u7279\u52a0\u5973\u5b69", "\u72c2\u70ed\u84dd\u8c03", "\u8389\u8389\u5a05", "\u841d\u838e\u8389\u5a05", "\u516b\u91cd\u6a31", "\u516b\u91cd\u971e", "\u5361\u83b2", "\u7b2c\u516d\u591c\u60f3\u66f2", "\u5361\u841d\u5c14", "\u59ec\u5b50", "\u6781\u5730\u6218\u5203", "\u5e03\u6d1b\u59ae\u5a05", "\u6b21\u751f\u94f6\u7ffc", "\u7406\u4e4b\u5f8b\u8005%26\u5e0c\u513f", "\u7406\u4e4b\u5f8b\u8005", "\u8ff7\u57ce\u9a87\u5154", "\u5e0c\u513f", "\u9b47\u591c\u661f\u6e0a", "\u9ed1\u5e0c\u513f", "\u5e15\u6735\u83f2\u8389\u4e1d", "\u4e0d\u706d\u661f\u951a", "\u5929\u5143\u9a91\u82f1", "\u5e7d\u5170\u9edb\u5c14", "\u6d3e\u8499bh3", "\u7231\u9171", "\u7eef\u7389\u4e38", "\u5fb7\u4e3d\u838e", "\u6708\u4e0b\u521d\u62e5", "\u6714\u591c\u89c2\u661f", "\u66ae\u5149\u9a91\u58eb", "\u683c\u857e\u4fee", "\u7559\u4e91\u501f\u98ce\u771f\u541b", "\u6885\u6bd4\u4e4c\u65af", "\u4eff\u72b9\u5927", "\u514b\u83b1\u56e0", "\u5723\u5251\u5e7d\u5170\u9edb\u5c14", "\u5996\u7cbe\u7231\u8389", "\u7279\u65af\u62c9zero", "\u82cd\u7384", "\u82e5\u6c34", "\u897f\u7433", "\u6234\u56e0\u65af\u96f7\u5e03", "\u8d1d\u62c9", "\u8d64\u9e22", "\u9547\u9b42\u6b4c", "\u6e21\u9e26", "\u4eba\u4e4b\u5f8b\u8005", "\u7231\u8389\u5e0c\u96c5", "\u5929\u7a79\u6e38\u4fa0", "\u742a\u4e9a\u5a1c", "\u7a7a\u4e4b\u5f8b\u8005", "\u85aa\u708e\u4e4b\u5f8b\u8005", "\u4e91\u58a8\u4e39\u5fc3", "\u7b26\u534e", "\u8bc6\u4e4b\u5f8b\u8005", "\u7279\u74e6\u6797", "\u7ef4\u5c14\u8587", "\u82bd\u8863", "\u96f7\u4e4b\u5f8b\u8005", "\u65ad\u7f6a\u5f71\u821e", "\u963f\u6ce2\u5c3c\u4e9a", "\u698e\u672c", "\u5384\u5c3c\u65af\u7279", "\u6076\u9f99", "\u8303\u4e8c\u7237", "\u6cd5\u62c9", "\u611a\u4eba\u4f17\u58eb\u5175", "\u611a\u4eba\u4f17\u58eb\u5175a", "\u611a\u4eba\u4f17\u58eb\u5175b", "\u611a\u4eba\u4f17\u58eb\u5175c", "\u611a\u4eba\u4f17a", "\u611a\u4eba\u4f17b", "\u98de\u98de", "\u83f2\u5229\u514b\u65af", "\u5973\u6027\u8ddf\u968f\u8005", "\u9022\u5ca9", "\u6446\u6e21\u4eba", "\u72c2\u8e81\u7684\u7537\u4eba", "\u5965\u5179", "\u8299\u841d\u62c9", "\u8ddf\u968f\u8005", "\u871c\u6c41\u751f\u7269", "\u9ec4\u9ebb\u5b50", "\u6e0a\u4e0a", "\u85e4\u6728", "\u6df1\u89c1", "\u798f\u672c", "\u8299\u84c9", "\u53e4\u6cfd", "\u53e4\u7530", "\u53e4\u5c71", "\u53e4\u8c37\u6607", "\u5085\u4e09\u513f", "\u9ad8\u8001\u516d", "\u77ff\u5de5\u5192", "\u5143\u592a", "\u5fb7\u5b89\u516c", "\u8302\u624d\u516c", "\u6770\u62c9\u5fb7", "\u845b\u7f57\u4e3d", "\u91d1\u5ffd\u5f8b", "\u516c\u4fca", "\u9505\u5df4", "\u6b4c\u5fb7", "\u963f\u8c6a", "\u72d7\u4e09\u513f", "\u845b\u745e\u4e1d", "\u82e5\u5fc3", "\u963f\u5c71\u5a46", "\u602a\u9e1f", "\u5e7f\u7af9", "\u89c2\u6d77", "\u5173\u5b8f", "\u871c\u6c41\u536b\u5175", "\u5b88\u536b1", "\u50b2\u6162\u7684\u5b88\u536b", "\u5bb3\u6015\u7684\u5b88\u536b", "\u8d35\u5b89", "\u76d6\u4f0a", "\u963f\u521b", "\u54c8\u592b\u4e39", "\u65e5\u8bed\u963f\u8d1d\u591a\uff08\u91ce\u5c9b\u5065\u513f\uff09", "\u65e5\u8bed\u57c3\u6d1b\u4f0a\uff08\u9ad8\u57a3\u5f69\u9633\uff09", "\u65e5\u8bed\u5b89\u67cf\uff08\u77f3\u89c1\u821e\u83dc\u9999\uff09", "\u65e5\u8bed\u795e\u91cc\u7eeb\u534e\uff08\u65e9\u89c1\u6c99\u7ec7\uff09", "\u65e5\u8bed\u795e\u91cc\u7eeb\u4eba\uff08\u77f3\u7530\u5f70\uff09", "\u65e5\u8bed\u767d\u672f\uff08\u6e38\u4f50\u6d69\u4e8c\uff09", "\u65e5\u8bed\u82ad\u82ad\u62c9\uff08\u9b3c\u5934\u660e\u91cc\uff09", "\u65e5\u8bed\u5317\u6597\uff08\u5c0f\u6e05\u6c34\u4e9a\u7f8e\uff09", "\u65e5\u8bed\u73ed\u5c3c\u7279\uff08\u9022\u5742\u826f\u592a\uff09", "\u65e5\u8bed\u574e\u8482\u4e1d\uff08\u67da\u6728\u51c9\u9999\uff09", "\u65e5\u8bed\u91cd\u4e91\uff08\u9f50\u85e4\u58ee\u9a6c\uff09", "\u65e5\u8bed\u67ef\u83b1\uff08\u524d\u5ddd\u51c9\u5b50\uff09", "\u65e5\u8bed\u8d5b\u8bfa\uff08\u5165\u91ce\u81ea\u7531\uff09", "\u65e5\u8bed\u6234\u56e0\u65af\u96f7\u5e03\uff08\u6d25\u7530\u5065\u6b21\u90ce\uff09", "\u65e5\u8bed\u8fea\u5362\u514b\uff08\u5c0f\u91ce\u8d24\u7ae0\uff09", "\u65e5\u8bed\u8fea\u5965\u5a1c\uff08\u4e95\u6cfd\u8bd7\u7ec7\uff09", "\u65e5\u8bed\u591a\u8389\uff08\u91d1\u7530\u670b\u5b50\uff09", "\u65e5\u8bed\u4f18\u83c8\uff08\u4f50\u85e4\u5229\u5948\uff09", "\u65e5\u8bed\u83f2\u8c22\u5c14\uff08\u5185\u7530\u771f\u793c\uff09", "\u65e5\u8bed\u7518\u96e8\uff08\u4e0a\u7530\u4e3d\u5948\uff09", "\u65e5\u8bed\uff08\u7560\u4e2d\u7950\uff09", "\u65e5\u8bed\u9e7f\u91ce\u9662\u5e73\u85cf\uff08\u4e95\u53e3\u7950\u4e00\uff09", "\u65e5\u8bed\u7a7a\uff08\u5800\u6c5f\u77ac\uff09", "\u65e5\u8bed\u8367\uff08\u60a0\u6728\u78a7\uff09", "\u65e5\u8bed\u80e1\u6843\uff08\u9ad8\u6865\u674e\u4f9d\uff09", "\u65e5\u8bed\u4e00\u6597\uff08\u897f\u5ddd\u8d35\u6559\uff09", "\u65e5\u8bed\u51ef\u4e9a\uff08\u9e1f\u6d77\u6d69\u8f85\uff09", "\u65e5\u8bed\u4e07\u53f6\uff08\u5c9b\u5d0e\u4fe1\u957f\uff09", "\u65e5\u8bed\u523b\u6674\uff08\u559c\u591a\u6751\u82f1\u68a8\uff09", "\u65e5\u8bed\u53ef\u8389\uff08\u4e45\u91ce\u7f8e\u54b2\uff09", "\u65e5\u8bed\u5fc3\u6d77\uff08\u4e09\u68ee\u94c3\u5b50\uff09", "\u65e5\u8bed\u4e5d\u6761\u88df\u7f57\uff08\u6fd1\u6237\u9ebb\u6c99\u7f8e\uff09", "\u65e5\u8bed\u4e3d\u838e\uff08\u7530\u4e2d\u7406\u60e0\uff09", "\u65e5\u8bed\u83ab\u5a1c\uff08\u5c0f\u539f\u597d\u7f8e\uff09", "\u65e5\u8bed\u7eb3\u897f\u59b2\uff08\u7530\u6751\u7531\u52a0\u8389\uff09", "\u65e5\u8bed\u59ae\u9732\uff08\u91d1\u5143\u5bff\u5b50\uff09", "\u65e5\u8bed\u51dd\u5149\uff08\u5927\u539f\u6c99\u8036\u9999\uff09", "\u65e5\u8bed\u8bfa\u827e\u5c14\uff08\u9ad8\u5c3e\u594f\u97f3\uff09", "\u65e5\u8bed\u5965\u5179\uff08\u589e\u8c37\u5eb7\u7eaa\uff09", "\u65e5\u8bed\u6d3e\u8499\uff08\u53e4\u8d3a\u8475\uff09", "\u65e5\u8bed\u7434\uff08\u658b\u85e4\u5343\u548c\uff09", "\u65e5\u8bed\u4e03\u4e03\uff08\u7530\u6751\u7531\u52a0\u8389\uff09", "\u65e5\u8bed\u96f7\u7535\u5c06\u519b\uff08\u6cfd\u57ce\u7f8e\u96ea\uff09", "\u65e5\u8bed\u96f7\u6cfd\uff08\u5185\u5c71\u6602\u8f89\uff09", "\u65e5\u8bed\u7f57\u838e\u8389\u4e9a\uff08\u52a0\u9688\u4e9a\u8863\uff09", "\u65e5\u8bed\u65e9\u67da\uff08\u6d32\u5d0e\u7eeb\uff09", "\u65e5\u8bed\u6563\u5175\uff08\u67ff\u539f\u5f7b\u4e5f\uff09", "\u65e5\u8bed\u7533\u9e64\uff08\u5ddd\u6f84\u7eeb\u5b50\uff09", "\u65e5\u8bed\u4e45\u5c90\u5fcd\uff08\u6c34\u6865\u9999\u7ec7\uff09", "\u65e5\u8bed\u5973\u58eb\uff08\u5e84\u5b50\u88d5\u8863\uff09", "\u65e5\u8bed\u7802\u7cd6\uff08\u85e4\u7530\u831c\uff09", "\u65e5\u8bed\u8fbe\u8fbe\u5229\u4e9a\uff08\u6728\u6751\u826f\u5e73\uff09", "\u65e5\u8bed\u6258\u9a6c\uff08\u68ee\u7530\u6210\u4e00\uff09", "\u65e5\u8bed\u63d0\u7eb3\u91cc\uff08\u5c0f\u6797\u6c99\u82d7\uff09", "\u65e5\u8bed\u6e29\u8fea\uff08\u6751\u6fd1\u6b65\uff09", "\u65e5\u8bed\u9999\u83f1\uff08\u5c0f\u6cfd\u4e9a\u674e\uff09", "\u65e5\u8bed\u9b48\uff08\u677e\u5188\u796f\u4e1e\uff09", "\u65e5\u8bed\u884c\u79cb\uff08\u7686\u5ddd\u7eaf\u5b50\uff09", "\u65e5\u8bed\u8f9b\u7131\uff08\u9ad8\u6865\u667a\u79cb\uff09", "\u65e5\u8bed\u516b\u91cd\u795e\u5b50\uff08\u4f50\u4ed3\u7eeb\u97f3\uff09", "\u65e5\u8bed\u70df\u7eef\uff08\u82b1\u5b88\u7531\u7f8e\u91cc\uff09", "\u65e5\u8bed\u591c\u5170\uff08\u8fdc\u85e4\u7eeb\uff09", "\u65e5\u8bed\u5bb5\u5bab\uff08\u690d\u7530\u4f73\u5948\uff09", "\u65e5\u8bed\u4e91\u5807\uff08\u5c0f\u5ca9\u4e95\u5c0f\u9e1f\uff09", "\u65e5\u8bed\u949f\u79bb\uff08\u524d\u91ce\u667a\u662d\uff09", "\u6770\u514b", "\u963f\u5409", "\u6c5f\u821f", "\u9274\u79cb", "\u5609\u4e49", "\u7eaa\u82b3", "\u666f\u6f84", "\u7ecf\u7eb6", "\u666f\u660e", "\u664b\u4f18", "\u963f\u9e20", "\u9152\u5ba2", "\u4e54\u5c14", "\u4e54\u745f\u592b", "\u7ea6\u987f", "\u4e54\u4f0a\u65af", "\u5c45\u5b89", "\u541b\u541b", "\u987a\u5409", "\u7eaf\u4e5f", "\u91cd\u4f50", "\u5927\u5c9b\u7eaf\u5e73", "\u84b2\u6cfd", "\u52d8\u89e3\u7531\u5c0f\u8def\u5065\u4e09\u90ce", "\u67ab", "\u67ab\u539f\u4e49\u5e86", "\u836b\u5c71", "\u7532\u6590\u7530\u9f8d\u99ac", "\u6d77\u6597", "\u60df\u795e\u6674\u4e4b\u4ecb", "\u9e7f\u91ce\u5948\u5948", "\u5361\u7435\u8389\u4e9a", "\u51ef\u745f\u7433", "\u52a0\u85e4\u4fe1\u609f", "\u52a0\u85e4\u6d0b\u5e73", "\u80dc\u5bb6", "\u8305\u847a\u4e00\u5e86", "\u548c\u662d", "\u4e00\u6b63", "\u4e00\u9053", "\u6842\u4e00", "\u5e86\u6b21\u90ce", "\u963f\u8d24", "\u5065\u53f8", "\u5065\u6b21\u90ce", "\u5065\u4e09\u90ce", "\u5929\u7406", "\u6740\u624ba", "\u6740\u624bb", "\u6728\u5357\u674f\u5948", "\u6728\u6751", "\u56fd\u738b", "\u6728\u4e0b", "\u5317\u6751", "\u6e05\u60e0", "\u6e05\u4eba", "\u514b\u5217\u95e8\u7279", "\u9a91\u58eb", "\u5c0f\u6797", "\u5c0f\u6625", "\u5eb7\u62c9\u5fb7", "\u5927\u8089\u4e38", "\u7434\u7f8e", "\u5b8f\u4e00", "\u5eb7\u4ecb", "\u5e78\u5fb7", "\u9ad8\u5584", "\u68a2", "\u514b\u7f57\u7d22", "\u4e45\u4fdd", "\u4e5d\u6761\u9570\u6cbb", "\u4e45\u6728\u7530", "\u6606\u94a7", "\u83ca\u5730\u541b", "\u4e45\u5229\u987b", "\u9ed1\u7530", "\u9ed1\u6cfd\u4eac\u4e4b\u4ecb", "\u54cd\u592a", "\u5c9a\u59d0", "\u5170\u6eaa", "\u6f9c\u9633", "\u52b3\u4f26\u65af", "\u4e50\u660e", "\u83b1\u8bfa", "\u83b2", "\u826f\u5b50", "\u674e\u5f53", "\u674e\u4e01", "\u5c0f\u4e50", "\u7075", "\u5c0f\u73b2", "\u7433\u7405a", "\u7433\u7405b", "\u5c0f\u5f6c", "\u5c0f\u5fb7", "\u5c0f\u697d", "\u5c0f\u9f99", "\u5c0f\u5434", "\u5c0f\u5434\u7684\u8bb0\u5fc6", "\u7406\u6b63", "\u963f\u9f99", "\u5362\u5361", "\u6d1b\u6210", "\u7f57\u5de7", "\u5317\u98ce\u72fc", "\u5362\u6b63", "\u840d\u59e5\u59e5", "\u524d\u7530", "\u771f\u663c", "\u9ebb\u7eaa", "\u771f", "\u611a\u4eba\u4f17-\u9a6c\u514b\u897f\u59c6", "\u5973\u6027a", "\u5973\u6027b", "\u5973\u6027a\u7684\u8ddf\u968f\u8005", "\u963f\u5b88", "\u739b\u683c\u4e3d\u7279", "\u771f\u7406", "\u739b\u4e54\u4e3d", "\u739b\u6587", "\u6b63\u80dc", "\u660c\u4fe1", "\u5c06\u53f8", "\u6b63\u4eba", "\u8def\u7237", "\u8001\u7ae0", "\u677e\u7530", "\u677e\u672c", "\u677e\u6d66", "\u677e\u5742", "\u8001\u5b5f", "\u5b5f\u4e39", "\u5546\u4eba\u968f\u4ece", "\u4f20\u4ee4\u5175", "\u7c73\u6b47\u5c14", "\u5fa1\u8206\u6e90\u4e00\u90ce", "\u5fa1\u8206\u6e90\u6b21\u90ce", "\u5343\u5ca9\u519b\u6559\u5934", "\u5343\u5ca9\u519b\u58eb\u5175", "\u660e\u535a", "\u660e\u4fca", "\u7f8e\u94c3", "\u7f8e\u548c", "\u963f\u5e78", "\u524a\u6708\u7b51\u9633\u771f\u541b", "\u94b1\u773c\u513f", "\u68ee\u5f66", "\u5143\u52a9", "\u7406\u6c34\u53e0\u5c71\u771f\u541b", "\u7406\u6c34\u758a\u5c71\u771f\u541b", "\u6731\u8001\u677f", "\u6728\u6728", "\u6751\u4e0a", "\u6751\u7530", "\u6c38\u91ce", "\u957f\u91ce\u539f\u9f99\u4e4b\u4ecb", "\u957f\u6fd1", "\u4e2d\u91ce\u5fd7\u4e43", "\u83dc\u83dc\u5b50", "\u6960\u6960", "\u6210\u6fd1", "\u963f\u5185", "\u5b81\u7984", "\u725b\u5fd7", "\u4fe1\u535a", "\u4f38\u592b", "\u91ce\u65b9", "\u8bfa\u62c9", "\u7eaa\u9999", "\u8bfa\u66fc", "\u4fee\u5973", "\u7eaf\u6c34\u7cbe\u7075", "\u5c0f\u5ddd", "\u5c0f\u4ed3\u6faa", "\u5188\u6797", "\u5188\u5d0e\u7ed8\u91cc\u9999", "\u5188\u5d0e\u9646\u6597", "\u5965\u62c9\u592b", "\u8001\u79d1", "\u9b3c\u5a46\u5a46", "\u5c0f\u91ce\u5bfa", "\u5927\u6cb3\u539f\u4e94\u53f3\u536b\u95e8", "\u5927\u4e45\u4fdd\u5927\u4ecb", "\u5927\u68ee", "\u5927\u52a9", "\u5965\u7279", "\u6d3e\u8499", "\u6d3e\u84992", "\u75c5\u4ebaa", "\u75c5\u4ebab", "\u5df4\u987f", "\u6d3e\u6069", "\u670b\u4e49", "\u56f4\u89c2\u7fa4\u4f17", "\u56f4\u89c2\u7fa4\u4f17a", "\u56f4\u89c2\u7fa4\u4f17b", "\u56f4\u89c2\u7fa4\u4f17c", "\u56f4\u89c2\u7fa4\u4f17d", "\u56f4\u89c2\u7fa4\u4f17e", "\u94dc\u96c0", "\u963f\u80a5", "\u5174\u53d4", "\u8001\u5468\u53d4", "\u516c\u4e3b", "\u5f7c\u5f97", "\u4e7e\u5b50", "\u828a\u828a", "\u4e7e\u73ae", "\u7eee\u547d", "\u675e\u5e73", "\u79cb\u6708", "\u6606\u6069", "\u96f7\u7535\u5f71", "\u5170\u9053\u5c14", "\u96f7\u8499\u5fb7", "\u5192\u5931\u7684\u5e15\u62c9\u5fb7", "\u4f36\u4e00", "\u73b2\u82b1", "\u963f\u4ec1", "\u5bb6\u81e3\u4eec", "\u68a8\u7ed8", "\u8363\u6c5f", "\u620e\u4e16", "\u6d6a\u4eba", "\u7f57\u4f0a\u65af", "\u5982\u610f", "\u51c9\u5b50", "\u5f69\u9999", "\u9152\u4e95", "\u5742\u672c", "\u6714\u6b21\u90ce", "\u6b66\u58eba", "\u6b66\u58ebb", "\u6b66\u58ebc", "\u6b66\u58ebd", "\u73ca\u745a", "\u4e09\u7530", "\u838e\u62c9", "\u7b39\u91ce", "\u806a\u7f8e", "\u806a", "\u5c0f\u767e\u5408", "\u6563\u5175", "\u5bb3\u6015\u7684\u5c0f\u5218", "\u8212\u4f2f\u7279", "\u8212\u8328", "\u6d77\u9f99", "\u4e16\u5b50", "\u8c22\u5c14\u76d6", "\u5bb6\u4e01", "\u5546\u534e", "\u6c99\u5bc5", "\u963f\u5347", "\u67f4\u7530", "\u963f\u8302", "\u5f0f\u5927\u5c06", "\u6e05\u6c34", "\u5fd7\u6751\u52d8\u5175\u536b", "\u65b0\u4e4b\u4e1e", "\u5fd7\u7ec7", "\u77f3\u5934", "\u8bd7\u7fbd", "\u8bd7\u7b60", "\u77f3\u58ee", "\u7fd4\u592a", "\u6b63\u4e8c", "\u5468\u5e73", "\u8212\u6768", "\u9f50\u683c\u8299\u4e3d\u96c5", "\u5973\u58eb", "\u601d\u52e4", "\u516d\u6307\u4e54\u745f", "\u611a\u4eba\u4f17\u5c0f\u5175d", "\u611a\u4eba\u4f17\u5c0f\u5175a", "\u611a\u4eba\u4f17\u5c0f\u5175b", "\u611a\u4eba\u4f17\u5c0f\u5175c", "\u5434\u8001\u4e94", "\u5434\u8001\u4e8c", "\u6ed1\u5934\u9b3c", "\u8a00\u7b11", "\u5434\u8001\u4e03", "\u58eb\u5175h", "\u58eb\u5175i", "\u58eb\u5175a", "\u58eb\u5175b", "\u58eb\u5175c", "\u58eb\u5175d", "\u58eb\u5175e", "\u58eb\u5175f", "\u58eb\u5175g", "\u594f\u592a", "\u65af\u5766\u5229", "\u6387\u661f\u652b\u8fb0\u5929\u541b", "\u5c0f\u5934", "\u5927\u6b66", "\u9676\u4e49\u9686", "\u6749\u672c", "\u82cf\u897f", "\u5acc\u7591\u4ebaa", "\u5acc\u7591\u4ebab", "\u5acc\u7591\u4ebac", "\u5acc\u7591\u4ebad", "\u65af\u4e07", "\u5251\u5ba2a", "\u5251\u5ba2b", "\u963f\u4e8c", "\u5fe0\u80dc", "\u5fe0\u592b", "\u963f\u656c", "\u5b5d\u5229", "\u9e70\u53f8\u8fdb", "\u9ad8\u5c71", "\u4e5d\u6761\u5b5d\u884c", "\u6bc5", "\u7af9\u5185", "\u62d3\u771f", "\u5353\u4e5f", "\u592a\u90ce\u4e38", "\u6cf0\u52d2", "\u624b\u5c9b", "\u54f2\u5e73", "\u54f2\u592b", "\u6258\u514b", "\u5927boss", "\u963f\u5f3a", "\u6258\u5c14\u5fb7\u62c9", "\u65c1\u89c2\u8005", "\u5929\u6210", "\u963f\u5927", "\u8482\u739b\u4e4c\u65af", "\u63d0\u7c73", "\u6237\u7530", "\u963f\u4e09", "\u4e00\u8d77\u7684\u4eba", "\u5fb7\u7530", "\u5fb7\u957f", "\u667a\u6811", "\u5229\u5f66", "\u80d6\u4e4e\u4e4e\u7684\u65c5\u884c\u8005", "\u85cf\u5b9d\u4ebaa", "\u85cf\u5b9d\u4ebab", "\u85cf\u5b9d\u4ebac", "\u85cf\u5b9d\u4ebad", "\u963f\u7947", "\u6052\u96c4", "\u9732\u5b50", "\u8bdd\u5267\u56e2\u56e2\u957f", "\u5185\u6751", "\u4e0a\u91ce", "\u4e0a\u6749", "\u8001\u6234", "\u8001\u9ad8", "\u8001\u8d3e", "\u8001\u58a8", "\u8001\u5b59", "\u5929\u67a2\u661f", "\u8001\u4e91", "\u6709\u4e50\u658b", "\u4e11\u96c4", "\u4e4c\u7ef4", "\u74e6\u4eac", "\u83f2\u5c14\u6208\u9edb\u7279", "\u7ef4\u591a\u5229\u4e9a", "\u8587\u5c14", "\u74e6\u683c\u7eb3", "\u963f\u5916", "\u4f8d\u5973", "\u74e6\u62c9", "\u671b\u96c5", "\u5b9b\u70df", "\u742c\u7389", "\u6218\u58eba", "\u6218\u58ebb", "\u6e21\u8fba", "\u6e21\u90e8", "\u963f\u4f1f", "\u6587\u749f", "\u6587\u6e0a", "\u97e6\u5c14\u7eb3", "\u738b\u6273\u624b", "\u6b66\u6c9b", "\u6653\u98de", "\u8f9b\u7a0b", "\u661f\u706b", "\u661f\u7a00", "\u8f9b\u79c0", "\u79c0\u534e", "\u963f\u65ed", "\u5f90\u5218\u5e08", "\u77e2\u90e8", "\u516b\u6728", "\u5c71\u4e0a", "\u963f\u9633", "\u989c\u7b11", "\u5eb7\u660e", "\u6cf0\u4e45", "\u5b89\u6b66", "\u77e2\u7530\u5e78\u559c", "\u77e2\u7530\u8f9b\u559c", "\u4e49\u575a", "\u83ba\u513f", "\u76c8\u4e30", "\u5b9c\u5e74", "\u94f6\u674f", "\u9038\u8f69", "\u6a2a\u5c71", "\u6c38\u8d35", "\u6c38\u4e1a", "\u5609\u4e45", "\u5409\u5ddd", "\u4e49\u9ad8", "\u7528\u9ad8", "\u9633\u592a", "\u5143\u84c9", "\u73a5\u8f89", "\u6bd3\u534e", "\u6709\u9999", "\u5e78\u4e5f", "\u7531\u771f", "\u7ed3\u83dc", "\u97f5\u5b81", "\u767e\u5408", "\u767e\u5408\u534e", "\u5c24\u82cf\u6ce2\u592b", "\u88d5\u5b50", "\u60a0\u7b56", "\u60a0\u4e5f", "\u4e8e\u5ae3", "\u67da\u5b50", "\u8001\u90d1", "\u6b63\u8302", "\u5fd7\u6210", "\u82b7\u5de7", "\u77e5\u6613", "\u652f\u652f", "\u5468\u826f", "\u73e0\u51fd", "\u795d\u660e", "\u795d\u6d9b"],
54
+ "symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "A", "E", "I", "N", "O", "Q", "U", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "r", "s", "t", "u", "v", "w", "y", "z", "\u0283", "\u02a7", "\u02a6", "\u026f", "\u0279", "\u0259", "\u0265", "\u207c", "\u02b0", "`", "\u2192", "\u2193", "\u2191", " "]
55
+ }
models.py ADDED
@@ -0,0 +1,533 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ import modules
8
+ import attentions
9
+ import monotonic_align
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+ from commons import init_weights, get_padding
14
+
15
+
16
+ class StochasticDurationPredictor(nn.Module):
17
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
18
+ super().__init__()
19
+ filter_channels = in_channels # it needs to be removed from future version.
20
+ self.in_channels = in_channels
21
+ self.filter_channels = filter_channels
22
+ self.kernel_size = kernel_size
23
+ self.p_dropout = p_dropout
24
+ self.n_flows = n_flows
25
+ self.gin_channels = gin_channels
26
+
27
+ self.log_flow = modules.Log()
28
+ self.flows = nn.ModuleList()
29
+ self.flows.append(modules.ElementwiseAffine(2))
30
+ for i in range(n_flows):
31
+ self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
32
+ self.flows.append(modules.Flip())
33
+
34
+ self.post_pre = nn.Conv1d(1, filter_channels, 1)
35
+ self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
36
+ self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
37
+ self.post_flows = nn.ModuleList()
38
+ self.post_flows.append(modules.ElementwiseAffine(2))
39
+ for i in range(4):
40
+ self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
41
+ self.post_flows.append(modules.Flip())
42
+
43
+ self.pre = nn.Conv1d(in_channels, filter_channels, 1)
44
+ self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
45
+ self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
46
+ if gin_channels != 0:
47
+ self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
48
+
49
+ def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
50
+ x = torch.detach(x)
51
+ x = self.pre(x)
52
+ if g is not None:
53
+ g = torch.detach(g)
54
+ x = x + self.cond(g)
55
+ x = self.convs(x, x_mask)
56
+ x = self.proj(x) * x_mask
57
+
58
+ if not reverse:
59
+ flows = self.flows
60
+ assert w is not None
61
+
62
+ logdet_tot_q = 0
63
+ h_w = self.post_pre(w)
64
+ h_w = self.post_convs(h_w, x_mask)
65
+ h_w = self.post_proj(h_w) * x_mask
66
+ e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
67
+ z_q = e_q
68
+ for flow in self.post_flows:
69
+ z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
70
+ logdet_tot_q += logdet_q
71
+ z_u, z1 = torch.split(z_q, [1, 1], 1)
72
+ u = torch.sigmoid(z_u) * x_mask
73
+ z0 = (w - u) * x_mask
74
+ logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
75
+ logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
76
+
77
+ logdet_tot = 0
78
+ z0, logdet = self.log_flow(z0, x_mask)
79
+ logdet_tot += logdet
80
+ z = torch.cat([z0, z1], 1)
81
+ for flow in flows:
82
+ z, logdet = flow(z, x_mask, g=x, reverse=reverse)
83
+ logdet_tot = logdet_tot + logdet
84
+ nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
85
+ return nll + logq # [b]
86
+ else:
87
+ flows = list(reversed(self.flows))
88
+ flows = flows[:-2] + [flows[-1]] # remove a useless vflow
89
+ z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
90
+ for flow in flows:
91
+ z = flow(z, x_mask, g=x, reverse=reverse)
92
+ z0, z1 = torch.split(z, [1, 1], 1)
93
+ logw = z0
94
+ return logw
95
+
96
+
97
+ class DurationPredictor(nn.Module):
98
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
99
+ super().__init__()
100
+
101
+ self.in_channels = in_channels
102
+ self.filter_channels = filter_channels
103
+ self.kernel_size = kernel_size
104
+ self.p_dropout = p_dropout
105
+ self.gin_channels = gin_channels
106
+
107
+ self.drop = nn.Dropout(p_dropout)
108
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
109
+ self.norm_1 = modules.LayerNorm(filter_channels)
110
+ self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
111
+ self.norm_2 = modules.LayerNorm(filter_channels)
112
+ self.proj = nn.Conv1d(filter_channels, 1, 1)
113
+
114
+ if gin_channels != 0:
115
+ self.cond = nn.Conv1d(gin_channels, in_channels, 1)
116
+
117
+ def forward(self, x, x_mask, g=None):
118
+ x = torch.detach(x)
119
+ if g is not None:
120
+ g = torch.detach(g)
121
+ x = x + self.cond(g)
122
+ x = self.conv_1(x * x_mask)
123
+ x = torch.relu(x)
124
+ x = self.norm_1(x)
125
+ x = self.drop(x)
126
+ x = self.conv_2(x * x_mask)
127
+ x = torch.relu(x)
128
+ x = self.norm_2(x)
129
+ x = self.drop(x)
130
+ x = self.proj(x * x_mask)
131
+ return x * x_mask
132
+
133
+
134
+ class TextEncoder(nn.Module):
135
+ def __init__(self,
136
+ n_vocab,
137
+ out_channels,
138
+ hidden_channels,
139
+ filter_channels,
140
+ n_heads,
141
+ n_layers,
142
+ kernel_size,
143
+ p_dropout):
144
+ super().__init__()
145
+ self.n_vocab = n_vocab
146
+ self.out_channels = out_channels
147
+ self.hidden_channels = hidden_channels
148
+ self.filter_channels = filter_channels
149
+ self.n_heads = n_heads
150
+ self.n_layers = n_layers
151
+ self.kernel_size = kernel_size
152
+ self.p_dropout = p_dropout
153
+
154
+ self.emb = nn.Embedding(n_vocab, hidden_channels)
155
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
156
+
157
+ self.encoder = attentions.Encoder(
158
+ hidden_channels,
159
+ filter_channels,
160
+ n_heads,
161
+ n_layers,
162
+ kernel_size,
163
+ p_dropout)
164
+ self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
165
+
166
+ def forward(self, x, x_lengths):
167
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
168
+ x = torch.transpose(x, 1, -1) # [b, h, t]
169
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
170
+
171
+ x = self.encoder(x * x_mask, x_mask)
172
+ stats = self.proj(x) * x_mask
173
+
174
+ m, logs = torch.split(stats, self.out_channels, dim=1)
175
+ return x, m, logs, x_mask
176
+
177
+
178
+ class ResidualCouplingBlock(nn.Module):
179
+ def __init__(self,
180
+ channels,
181
+ hidden_channels,
182
+ kernel_size,
183
+ dilation_rate,
184
+ n_layers,
185
+ n_flows=4,
186
+ gin_channels=0):
187
+ super().__init__()
188
+ self.channels = channels
189
+ self.hidden_channels = hidden_channels
190
+ self.kernel_size = kernel_size
191
+ self.dilation_rate = dilation_rate
192
+ self.n_layers = n_layers
193
+ self.n_flows = n_flows
194
+ self.gin_channels = gin_channels
195
+
196
+ self.flows = nn.ModuleList()
197
+ for i in range(n_flows):
198
+ self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
199
+ self.flows.append(modules.Flip())
200
+
201
+ def forward(self, x, x_mask, g=None, reverse=False):
202
+ if not reverse:
203
+ for flow in self.flows:
204
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
205
+ else:
206
+ for flow in reversed(self.flows):
207
+ x = flow(x, x_mask, g=g, reverse=reverse)
208
+ return x
209
+
210
+
211
+ class PosteriorEncoder(nn.Module):
212
+ def __init__(self,
213
+ in_channels,
214
+ out_channels,
215
+ hidden_channels,
216
+ kernel_size,
217
+ dilation_rate,
218
+ n_layers,
219
+ gin_channels=0):
220
+ super().__init__()
221
+ self.in_channels = in_channels
222
+ self.out_channels = out_channels
223
+ self.hidden_channels = hidden_channels
224
+ self.kernel_size = kernel_size
225
+ self.dilation_rate = dilation_rate
226
+ self.n_layers = n_layers
227
+ self.gin_channels = gin_channels
228
+
229
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
230
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
231
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
232
+
233
+ def forward(self, x, x_lengths, g=None):
234
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
235
+ x = self.pre(x) * x_mask
236
+ x = self.enc(x, x_mask, g=g)
237
+ stats = self.proj(x) * x_mask
238
+ m, logs = torch.split(stats, self.out_channels, dim=1)
239
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
240
+ return z, m, logs, x_mask
241
+
242
+
243
+ class Generator(torch.nn.Module):
244
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
245
+ super(Generator, self).__init__()
246
+ self.num_kernels = len(resblock_kernel_sizes)
247
+ self.num_upsamples = len(upsample_rates)
248
+ self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
249
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
250
+
251
+ self.ups = nn.ModuleList()
252
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
253
+ self.ups.append(weight_norm(
254
+ ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
255
+ k, u, padding=(k-u)//2)))
256
+
257
+ self.resblocks = nn.ModuleList()
258
+ for i in range(len(self.ups)):
259
+ ch = upsample_initial_channel//(2**(i+1))
260
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
261
+ self.resblocks.append(resblock(ch, k, d))
262
+
263
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
264
+ self.ups.apply(init_weights)
265
+
266
+ if gin_channels != 0:
267
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
268
+
269
+ def forward(self, x, g=None):
270
+ x = self.conv_pre(x)
271
+ if g is not None:
272
+ x = x + self.cond(g)
273
+
274
+ for i in range(self.num_upsamples):
275
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
276
+ x = self.ups[i](x)
277
+ xs = None
278
+ for j in range(self.num_kernels):
279
+ if xs is None:
280
+ xs = self.resblocks[i*self.num_kernels+j](x)
281
+ else:
282
+ xs += self.resblocks[i*self.num_kernels+j](x)
283
+ x = xs / self.num_kernels
284
+ x = F.leaky_relu(x)
285
+ x = self.conv_post(x)
286
+ x = torch.tanh(x)
287
+
288
+ return x
289
+
290
+ def remove_weight_norm(self):
291
+ print('Removing weight norm...')
292
+ for l in self.ups:
293
+ remove_weight_norm(l)
294
+ for l in self.resblocks:
295
+ l.remove_weight_norm()
296
+
297
+
298
+ class DiscriminatorP(torch.nn.Module):
299
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
300
+ super(DiscriminatorP, self).__init__()
301
+ self.period = period
302
+ self.use_spectral_norm = use_spectral_norm
303
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
304
+ self.convs = nn.ModuleList([
305
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
306
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
307
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
308
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
309
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
310
+ ])
311
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
312
+
313
+ def forward(self, x):
314
+ fmap = []
315
+
316
+ # 1d to 2d
317
+ b, c, t = x.shape
318
+ if t % self.period != 0: # pad first
319
+ n_pad = self.period - (t % self.period)
320
+ x = F.pad(x, (0, n_pad), "reflect")
321
+ t = t + n_pad
322
+ x = x.view(b, c, t // self.period, self.period)
323
+
324
+ for l in self.convs:
325
+ x = l(x)
326
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
327
+ fmap.append(x)
328
+ x = self.conv_post(x)
329
+ fmap.append(x)
330
+ x = torch.flatten(x, 1, -1)
331
+
332
+ return x, fmap
333
+
334
+
335
+ class DiscriminatorS(torch.nn.Module):
336
+ def __init__(self, use_spectral_norm=False):
337
+ super(DiscriminatorS, self).__init__()
338
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
339
+ self.convs = nn.ModuleList([
340
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
341
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
342
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
343
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
344
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
345
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
346
+ ])
347
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
348
+
349
+ def forward(self, x):
350
+ fmap = []
351
+
352
+ for l in self.convs:
353
+ x = l(x)
354
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
355
+ fmap.append(x)
356
+ x = self.conv_post(x)
357
+ fmap.append(x)
358
+ x = torch.flatten(x, 1, -1)
359
+
360
+ return x, fmap
361
+
362
+
363
+ class MultiPeriodDiscriminator(torch.nn.Module):
364
+ def __init__(self, use_spectral_norm=False):
365
+ super(MultiPeriodDiscriminator, self).__init__()
366
+ periods = [2,3,5,7,11]
367
+
368
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
369
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
370
+ self.discriminators = nn.ModuleList(discs)
371
+
372
+ def forward(self, y, y_hat):
373
+ y_d_rs = []
374
+ y_d_gs = []
375
+ fmap_rs = []
376
+ fmap_gs = []
377
+ for i, d in enumerate(self.discriminators):
378
+ y_d_r, fmap_r = d(y)
379
+ y_d_g, fmap_g = d(y_hat)
380
+ y_d_rs.append(y_d_r)
381
+ y_d_gs.append(y_d_g)
382
+ fmap_rs.append(fmap_r)
383
+ fmap_gs.append(fmap_g)
384
+
385
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
386
+
387
+
388
+
389
+ class SynthesizerTrn(nn.Module):
390
+ """
391
+ Synthesizer for Training
392
+ """
393
+
394
+ def __init__(self,
395
+ n_vocab,
396
+ spec_channels,
397
+ segment_size,
398
+ inter_channels,
399
+ hidden_channels,
400
+ filter_channels,
401
+ n_heads,
402
+ n_layers,
403
+ kernel_size,
404
+ p_dropout,
405
+ resblock,
406
+ resblock_kernel_sizes,
407
+ resblock_dilation_sizes,
408
+ upsample_rates,
409
+ upsample_initial_channel,
410
+ upsample_kernel_sizes,
411
+ n_speakers=0,
412
+ gin_channels=0,
413
+ use_sdp=True,
414
+ **kwargs):
415
+
416
+ super().__init__()
417
+ self.n_vocab = n_vocab
418
+ self.spec_channels = spec_channels
419
+ self.inter_channels = inter_channels
420
+ self.hidden_channels = hidden_channels
421
+ self.filter_channels = filter_channels
422
+ self.n_heads = n_heads
423
+ self.n_layers = n_layers
424
+ self.kernel_size = kernel_size
425
+ self.p_dropout = p_dropout
426
+ self.resblock = resblock
427
+ self.resblock_kernel_sizes = resblock_kernel_sizes
428
+ self.resblock_dilation_sizes = resblock_dilation_sizes
429
+ self.upsample_rates = upsample_rates
430
+ self.upsample_initial_channel = upsample_initial_channel
431
+ self.upsample_kernel_sizes = upsample_kernel_sizes
432
+ self.segment_size = segment_size
433
+ self.n_speakers = n_speakers
434
+ self.gin_channels = gin_channels
435
+
436
+ self.use_sdp = use_sdp
437
+
438
+ self.enc_p = TextEncoder(n_vocab,
439
+ inter_channels,
440
+ hidden_channels,
441
+ filter_channels,
442
+ n_heads,
443
+ n_layers,
444
+ kernel_size,
445
+ p_dropout)
446
+ self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
447
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
448
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
449
+
450
+ if use_sdp:
451
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
452
+ else:
453
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
454
+
455
+ if n_speakers > 1:
456
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
457
+
458
+ def forward(self, x, x_lengths, y, y_lengths, sid=None):
459
+
460
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
461
+ if self.n_speakers > 0:
462
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
463
+ else:
464
+ g = None
465
+
466
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
467
+ z_p = self.flow(z, y_mask, g=g)
468
+
469
+ with torch.no_grad():
470
+ # negative cross-entropy
471
+ s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
472
+ neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
473
+ 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]
474
+ 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]
475
+ neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
476
+ neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
477
+
478
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
479
+ attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
480
+
481
+ w = attn.sum(2)
482
+ if self.use_sdp:
483
+ l_length = self.dp(x, x_mask, w, g=g)
484
+ l_length = l_length / torch.sum(x_mask)
485
+ else:
486
+ logw_ = torch.log(w + 1e-6) * x_mask
487
+ logw = self.dp(x, x_mask, g=g)
488
+ l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
489
+
490
+ # expand prior
491
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
492
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
493
+
494
+ z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
495
+ o = self.dec(z_slice, g=g)
496
+ return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
497
+
498
+ def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
499
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
500
+ if self.n_speakers > 0:
501
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
502
+ else:
503
+ g = None
504
+
505
+ if self.use_sdp:
506
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
507
+ else:
508
+ logw = self.dp(x, x_mask, g=g)
509
+ w = torch.exp(logw) * x_mask * length_scale
510
+ w_ceil = torch.ceil(w)
511
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
512
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
513
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
514
+ attn = commons.generate_path(w_ceil, attn_mask)
515
+
516
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
517
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
518
+
519
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
520
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
521
+ o = self.dec((z * y_mask)[:,:,:max_len], g=g)
522
+ return o, attn, y_mask, (z, z_p, m_p, logs_p)
523
+
524
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
525
+ assert self.n_speakers > 0, "n_speakers have to be larger than 0."
526
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
527
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
528
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
529
+ z_p = self.flow(z, y_mask, g=g_src)
530
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
531
+ o_hat = self.dec(z_hat * y_mask, g=g_tgt)
532
+ return o_hat, y_mask, (z, z_p, z_hat)
533
+
modules.py ADDED
@@ -0,0 +1,388 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
8
+ from torch.nn.utils import weight_norm, remove_weight_norm
9
+
10
+ import commons
11
+ from commons import init_weights, get_padding
12
+ from transforms import piecewise_rational_quadratic_transform
13
+
14
+
15
+ LRELU_SLOPE = 0.1
16
+
17
+
18
+ class LayerNorm(nn.Module):
19
+ def __init__(self, channels, eps=1e-5):
20
+ super().__init__()
21
+ self.channels = channels
22
+ self.eps = eps
23
+
24
+ self.gamma = nn.Parameter(torch.ones(channels))
25
+ self.beta = nn.Parameter(torch.zeros(channels))
26
+
27
+ def forward(self, x):
28
+ x = x.transpose(1, -1)
29
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
30
+ return x.transpose(1, -1)
31
+
32
+
33
+ class ConvReluNorm(nn.Module):
34
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
35
+ super().__init__()
36
+ self.in_channels = in_channels
37
+ self.hidden_channels = hidden_channels
38
+ self.out_channels = out_channels
39
+ self.kernel_size = kernel_size
40
+ self.n_layers = n_layers
41
+ self.p_dropout = p_dropout
42
+ assert n_layers > 1, "Number of layers should be larger than 0."
43
+
44
+ self.conv_layers = nn.ModuleList()
45
+ self.norm_layers = nn.ModuleList()
46
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
47
+ self.norm_layers.append(LayerNorm(hidden_channels))
48
+ self.relu_drop = nn.Sequential(
49
+ nn.ReLU(),
50
+ nn.Dropout(p_dropout))
51
+ for _ in range(n_layers-1):
52
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
53
+ self.norm_layers.append(LayerNorm(hidden_channels))
54
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
55
+ self.proj.weight.data.zero_()
56
+ self.proj.bias.data.zero_()
57
+
58
+ def forward(self, x, x_mask):
59
+ x_org = x
60
+ for i in range(self.n_layers):
61
+ x = self.conv_layers[i](x * x_mask)
62
+ x = self.norm_layers[i](x)
63
+ x = self.relu_drop(x)
64
+ x = x_org + self.proj(x)
65
+ return x * x_mask
66
+
67
+
68
+ class DDSConv(nn.Module):
69
+ """
70
+ Dialted and Depth-Separable Convolution
71
+ """
72
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
73
+ super().__init__()
74
+ self.channels = channels
75
+ self.kernel_size = kernel_size
76
+ self.n_layers = n_layers
77
+ self.p_dropout = p_dropout
78
+
79
+ self.drop = nn.Dropout(p_dropout)
80
+ self.convs_sep = nn.ModuleList()
81
+ self.convs_1x1 = nn.ModuleList()
82
+ self.norms_1 = nn.ModuleList()
83
+ self.norms_2 = nn.ModuleList()
84
+ for i in range(n_layers):
85
+ dilation = kernel_size ** i
86
+ padding = (kernel_size * dilation - dilation) // 2
87
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
88
+ groups=channels, dilation=dilation, padding=padding
89
+ ))
90
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
91
+ self.norms_1.append(LayerNorm(channels))
92
+ self.norms_2.append(LayerNorm(channels))
93
+
94
+ def forward(self, x, x_mask, g=None):
95
+ if g is not None:
96
+ x = x + g
97
+ for i in range(self.n_layers):
98
+ y = self.convs_sep[i](x * x_mask)
99
+ y = self.norms_1[i](y)
100
+ y = F.gelu(y)
101
+ y = self.convs_1x1[i](y)
102
+ y = self.norms_2[i](y)
103
+ y = F.gelu(y)
104
+ y = self.drop(y)
105
+ x = x + y
106
+ return x * x_mask
107
+
108
+
109
+ class WN(torch.nn.Module):
110
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
111
+ super(WN, self).__init__()
112
+ assert(kernel_size % 2 == 1)
113
+ self.hidden_channels =hidden_channels
114
+ self.kernel_size = kernel_size,
115
+ self.dilation_rate = dilation_rate
116
+ self.n_layers = n_layers
117
+ self.gin_channels = gin_channels
118
+ self.p_dropout = p_dropout
119
+
120
+ self.in_layers = torch.nn.ModuleList()
121
+ self.res_skip_layers = torch.nn.ModuleList()
122
+ self.drop = nn.Dropout(p_dropout)
123
+
124
+ if gin_channels != 0:
125
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
126
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
127
+
128
+ for i in range(n_layers):
129
+ dilation = dilation_rate ** i
130
+ padding = int((kernel_size * dilation - dilation) / 2)
131
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
132
+ dilation=dilation, padding=padding)
133
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
134
+ self.in_layers.append(in_layer)
135
+
136
+ # last one is not necessary
137
+ if i < n_layers - 1:
138
+ res_skip_channels = 2 * hidden_channels
139
+ else:
140
+ res_skip_channels = hidden_channels
141
+
142
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
143
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
144
+ self.res_skip_layers.append(res_skip_layer)
145
+
146
+ def forward(self, x, x_mask, g=None, **kwargs):
147
+ output = torch.zeros_like(x)
148
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
149
+
150
+ if g is not None:
151
+ g = self.cond_layer(g)
152
+
153
+ for i in range(self.n_layers):
154
+ x_in = self.in_layers[i](x)
155
+ if g is not None:
156
+ cond_offset = i * 2 * self.hidden_channels
157
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
158
+ else:
159
+ g_l = torch.zeros_like(x_in)
160
+
161
+ acts = commons.fused_add_tanh_sigmoid_multiply(
162
+ x_in,
163
+ g_l,
164
+ n_channels_tensor)
165
+ acts = self.drop(acts)
166
+
167
+ res_skip_acts = self.res_skip_layers[i](acts)
168
+ if i < self.n_layers - 1:
169
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
170
+ x = (x + res_acts) * x_mask
171
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
172
+ else:
173
+ output = output + res_skip_acts
174
+ return output * x_mask
175
+
176
+ def remove_weight_norm(self):
177
+ if self.gin_channels != 0:
178
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
179
+ for l in self.in_layers:
180
+ torch.nn.utils.remove_weight_norm(l)
181
+ for l in self.res_skip_layers:
182
+ torch.nn.utils.remove_weight_norm(l)
183
+
184
+
185
+ class ResBlock1(torch.nn.Module):
186
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
187
+ super(ResBlock1, self).__init__()
188
+ self.convs1 = nn.ModuleList([
189
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
190
+ padding=get_padding(kernel_size, dilation[0]))),
191
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
192
+ padding=get_padding(kernel_size, dilation[1]))),
193
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
194
+ padding=get_padding(kernel_size, dilation[2])))
195
+ ])
196
+ self.convs1.apply(init_weights)
197
+
198
+ self.convs2 = nn.ModuleList([
199
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
200
+ padding=get_padding(kernel_size, 1))),
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
+ ])
206
+ self.convs2.apply(init_weights)
207
+
208
+ def forward(self, x, x_mask=None):
209
+ for c1, c2 in zip(self.convs1, self.convs2):
210
+ xt = F.leaky_relu(x, LRELU_SLOPE)
211
+ if x_mask is not None:
212
+ xt = xt * x_mask
213
+ xt = c1(xt)
214
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
215
+ if x_mask is not None:
216
+ xt = xt * x_mask
217
+ xt = c2(xt)
218
+ x = xt + x
219
+ if x_mask is not None:
220
+ x = x * x_mask
221
+ return x
222
+
223
+ def remove_weight_norm(self):
224
+ for l in self.convs1:
225
+ remove_weight_norm(l)
226
+ for l in self.convs2:
227
+ remove_weight_norm(l)
228
+
229
+
230
+ class ResBlock2(torch.nn.Module):
231
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
232
+ super(ResBlock2, self).__init__()
233
+ self.convs = nn.ModuleList([
234
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
235
+ padding=get_padding(kernel_size, dilation[0]))),
236
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
237
+ padding=get_padding(kernel_size, dilation[1])))
238
+ ])
239
+ self.convs.apply(init_weights)
240
+
241
+ def forward(self, x, x_mask=None):
242
+ for c in self.convs:
243
+ xt = F.leaky_relu(x, LRELU_SLOPE)
244
+ if x_mask is not None:
245
+ xt = xt * x_mask
246
+ xt = c(xt)
247
+ x = xt + x
248
+ if x_mask is not None:
249
+ x = x * x_mask
250
+ return x
251
+
252
+ def remove_weight_norm(self):
253
+ for l in self.convs:
254
+ remove_weight_norm(l)
255
+
256
+
257
+ class Log(nn.Module):
258
+ def forward(self, x, x_mask, reverse=False, **kwargs):
259
+ if not reverse:
260
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
261
+ logdet = torch.sum(-y, [1, 2])
262
+ return y, logdet
263
+ else:
264
+ x = torch.exp(x) * x_mask
265
+ return x
266
+
267
+
268
+ class Flip(nn.Module):
269
+ def forward(self, x, *args, reverse=False, **kwargs):
270
+ x = torch.flip(x, [1])
271
+ if not reverse:
272
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
273
+ return x, logdet
274
+ else:
275
+ return x
276
+
277
+
278
+ class ElementwiseAffine(nn.Module):
279
+ def __init__(self, channels):
280
+ super().__init__()
281
+ self.channels = channels
282
+ self.m = nn.Parameter(torch.zeros(channels,1))
283
+ self.logs = nn.Parameter(torch.zeros(channels,1))
284
+
285
+ def forward(self, x, x_mask, reverse=False, **kwargs):
286
+ if not reverse:
287
+ y = self.m + torch.exp(self.logs) * x
288
+ y = y * x_mask
289
+ logdet = torch.sum(self.logs * x_mask, [1,2])
290
+ return y, logdet
291
+ else:
292
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
293
+ return x
294
+
295
+
296
+ class ResidualCouplingLayer(nn.Module):
297
+ def __init__(self,
298
+ channels,
299
+ hidden_channels,
300
+ kernel_size,
301
+ dilation_rate,
302
+ n_layers,
303
+ p_dropout=0,
304
+ gin_channels=0,
305
+ mean_only=False):
306
+ assert channels % 2 == 0, "channels should be divisible by 2"
307
+ super().__init__()
308
+ self.channels = channels
309
+ self.hidden_channels = hidden_channels
310
+ self.kernel_size = kernel_size
311
+ self.dilation_rate = dilation_rate
312
+ self.n_layers = n_layers
313
+ self.half_channels = channels // 2
314
+ self.mean_only = mean_only
315
+
316
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
317
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
318
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
319
+ self.post.weight.data.zero_()
320
+ self.post.bias.data.zero_()
321
+
322
+ def forward(self, x, x_mask, g=None, reverse=False):
323
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
324
+ h = self.pre(x0) * x_mask
325
+ h = self.enc(h, x_mask, g=g)
326
+ stats = self.post(h) * x_mask
327
+ if not self.mean_only:
328
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
329
+ else:
330
+ m = stats
331
+ logs = torch.zeros_like(m)
332
+
333
+ if not reverse:
334
+ x1 = m + x1 * torch.exp(logs) * x_mask
335
+ x = torch.cat([x0, x1], 1)
336
+ logdet = torch.sum(logs, [1,2])
337
+ return x, logdet
338
+ else:
339
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
340
+ x = torch.cat([x0, x1], 1)
341
+ return x
342
+
343
+
344
+ class ConvFlow(nn.Module):
345
+ def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
346
+ super().__init__()
347
+ self.in_channels = in_channels
348
+ self.filter_channels = filter_channels
349
+ self.kernel_size = kernel_size
350
+ self.n_layers = n_layers
351
+ self.num_bins = num_bins
352
+ self.tail_bound = tail_bound
353
+ self.half_channels = in_channels // 2
354
+
355
+ self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
356
+ self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
357
+ self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
358
+ self.proj.weight.data.zero_()
359
+ self.proj.bias.data.zero_()
360
+
361
+ def forward(self, x, x_mask, g=None, reverse=False):
362
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
363
+ h = self.pre(x0)
364
+ h = self.convs(h, x_mask, g=g)
365
+ h = self.proj(h) * x_mask
366
+
367
+ b, c, t = x0.shape
368
+ h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
369
+
370
+ unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
371
+ unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
372
+ unnormalized_derivatives = h[..., 2 * self.num_bins:]
373
+
374
+ x1, logabsdet = piecewise_rational_quadratic_transform(x1,
375
+ unnormalized_widths,
376
+ unnormalized_heights,
377
+ unnormalized_derivatives,
378
+ inverse=reverse,
379
+ tails='linear',
380
+ tail_bound=self.tail_bound
381
+ )
382
+
383
+ x = torch.cat([x0, x1], 1) * x_mask
384
+ logdet = torch.sum(logabsdet * x_mask, [1,2])
385
+ if not reverse:
386
+ return x, logdet
387
+ else:
388
+ return x
monotonic_align/__init__.py ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from numpy import zeros, int32, float32
2
+ from torch import from_numpy
3
+
4
+ from .core import maximum_path_jit
5
+
6
+
7
+ def maximum_path(neg_cent, mask):
8
+ """ numba optimized version.
9
+ neg_cent: [b, t_t, t_s]
10
+ mask: [b, t_t, t_s]
11
+ """
12
+ device = neg_cent.device
13
+ dtype = neg_cent.dtype
14
+ neg_cent = neg_cent.data.cpu().numpy().astype(float32)
15
+ path = zeros(neg_cent.shape, dtype=int32)
16
+
17
+ t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
18
+ t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
19
+ maximum_path_jit(path, neg_cent, t_t_max, t_s_max)
20
+ return from_numpy(path).to(device=device, dtype=dtype)
monotonic_align/core.py ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numba
2
+
3
+
4
+ @numba.jit(numba.void(numba.int32[:, :, ::1], numba.float32[:, :, ::1], numba.int32[::1], numba.int32[::1]),
5
+ nopython=True, nogil=True)
6
+ def maximum_path_jit(paths, values, t_ys, t_xs):
7
+ b = paths.shape[0]
8
+ max_neg_val = -1e9
9
+ for i in range(int(b)):
10
+ path = paths[i]
11
+ value = values[i]
12
+ t_y = t_ys[i]
13
+ t_x = t_xs[i]
14
+
15
+ v_prev = v_cur = 0.0
16
+ index = t_x - 1
17
+
18
+ for y in range(t_y):
19
+ for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
20
+ if x == y:
21
+ v_cur = max_neg_val
22
+ else:
23
+ v_cur = value[y - 1, x]
24
+ if x == 0:
25
+ if y == 0:
26
+ v_prev = 0.
27
+ else:
28
+ v_prev = max_neg_val
29
+ else:
30
+ v_prev = value[y - 1, x - 1]
31
+ value[y, x] += max(v_prev, v_cur)
32
+
33
+ for y in range(t_y - 1, -1, -1):
34
+ path[y, index] = 1
35
+ if index != 0 and (index == y or value[y - 1, index] < value[y - 1, index - 1]):
36
+ index = index - 1
requirements.txt ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Cython
2
+ librosa
3
+ matplotlib
4
+ numpy
5
+ phonemizer
6
+ scipy
7
+ tensorboard
8
+ torch
9
+ torchvision
10
+ Unidecode
11
+ pyopenjtalk
12
+ ffmpeg
13
+ jamo
14
+ cn2an
15
+ gradio
16
+ pypinyin
17
+ jieba
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 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, symbols, 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
+ _symbol_to_id = {s: i for i, s in enumerate(symbols)}
20
+ sequence = []
21
+
22
+ clean_text = _clean_text(text, cleaner_names)
23
+ for symbol in clean_text:
24
+ if symbol not in _symbol_to_id.keys():
25
+ continue
26
+ symbol_id = _symbol_to_id[symbol]
27
+ sequence += [symbol_id]
28
+ return sequence, clean_text
29
+
30
+
31
+ def cleaned_text_to_sequence(cleaned_text):
32
+ '''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
33
+ Args:
34
+ text: string to convert to a sequence
35
+ Returns:
36
+ List of integers corresponding to the symbols in the text
37
+ '''
38
+ sequence = [_symbol_to_id[symbol] for symbol in cleaned_text if symbol in _symbol_to_id.keys()]
39
+ return sequence
40
+
41
+
42
+ def sequence_to_text(sequence):
43
+ '''Converts a sequence of IDs back to a string'''
44
+ result = ''
45
+ for symbol_id in sequence:
46
+ s = _id_to_symbol[symbol_id]
47
+ result += s
48
+ return result
49
+
50
+
51
+ def _clean_text(text, cleaner_names):
52
+ for name in cleaner_names:
53
+ cleaner = getattr(cleaners, name)
54
+ if not cleaner:
55
+ raise Exception('Unknown cleaner: %s' % name)
56
+ text = cleaner(text)
57
+ return text
text/cleaners.py ADDED
@@ -0,0 +1,475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ import pyopenjtalk
18
+ from jamo import h2j, j2hcj
19
+ from pypinyin import lazy_pinyin, BOPOMOFO
20
+ import jieba, cn2an
21
+
22
+
23
+ # This is a list of Korean classifiers preceded by pure Korean numerals.
24
+ _korean_classifiers = '군데 권 개 그루 닢 대 두 마리 모 모금 뭇 발 발짝 방 번 벌 보루 살 수 술 시 쌈 움큼 정 짝 채 척 첩 축 켤레 톨 통'
25
+
26
+ # Regular expression matching whitespace:
27
+ _whitespace_re = re.compile(r'\s+')
28
+
29
+ # Regular expression matching Japanese without punctuation marks:
30
+ _japanese_characters = re.compile(r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
31
+
32
+ # Regular expression matching non-Japanese characters or punctuation marks:
33
+ _japanese_marks = re.compile(r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
34
+
35
+ # List of (regular expression, replacement) pairs for abbreviations:
36
+ _abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [
37
+ ('mrs', 'misess'),
38
+ ('mr', 'mister'),
39
+ ('dr', 'doctor'),
40
+ ('st', 'saint'),
41
+ ('co', 'company'),
42
+ ('jr', 'junior'),
43
+ ('maj', 'major'),
44
+ ('gen', 'general'),
45
+ ('drs', 'doctors'),
46
+ ('rev', 'reverend'),
47
+ ('lt', 'lieutenant'),
48
+ ('hon', 'honorable'),
49
+ ('sgt', 'sergeant'),
50
+ ('capt', 'captain'),
51
+ ('esq', 'esquire'),
52
+ ('ltd', 'limited'),
53
+ ('col', 'colonel'),
54
+ ('ft', 'fort'),
55
+ ]]
56
+
57
+ # List of (hangul, hangul divided) pairs:
58
+ _hangul_divided = [(re.compile('%s' % x[0]), x[1]) for x in [
59
+ ('ㄳ', 'ㄱㅅ'),
60
+ ('ㄵ', 'ㄴㅈ'),
61
+ ('ㄶ', 'ㄴㅎ'),
62
+ ('ㄺ', 'ㄹㄱ'),
63
+ ('ㄻ', 'ㄹㅁ'),
64
+ ('ㄼ', 'ㄹㅂ'),
65
+ ('ㄽ', 'ㄹㅅ'),
66
+ ('ㄾ', 'ㄹㅌ'),
67
+ ('ㄿ', 'ㄹㅍ'),
68
+ ('ㅀ', 'ㄹㅎ'),
69
+ ('ㅄ', 'ㅂㅅ'),
70
+ ('ㅘ', 'ㅗㅏ'),
71
+ ('ㅙ', 'ㅗㅐ'),
72
+ ('ㅚ', 'ㅗㅣ'),
73
+ ('ㅝ', 'ㅜㅓ'),
74
+ ('ㅞ', 'ㅜㅔ'),
75
+ ('ㅟ', 'ㅜㅣ'),
76
+ ('ㅢ', 'ㅡㅣ'),
77
+ ('ㅑ', 'ㅣㅏ'),
78
+ ('ㅒ', 'ㅣㅐ'),
79
+ ('ㅕ', 'ㅣㅓ'),
80
+ ('ㅖ', 'ㅣㅔ'),
81
+ ('ㅛ', 'ㅣㅗ'),
82
+ ('ㅠ', 'ㅣㅜ')
83
+ ]]
84
+
85
+ # List of (Latin alphabet, hangul) pairs:
86
+ _latin_to_hangul = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
87
+ ('a', '에이'),
88
+ ('b', '비'),
89
+ ('c', '시'),
90
+ ('d', '디'),
91
+ ('e', '이'),
92
+ ('f', '에프'),
93
+ ('g', '지'),
94
+ ('h', '에이치'),
95
+ ('i', '아이'),
96
+ ('j', '제이'),
97
+ ('k', '케이'),
98
+ ('l', '엘'),
99
+ ('m', '엠'),
100
+ ('n', '엔'),
101
+ ('o', '오'),
102
+ ('p', '피'),
103
+ ('q', '큐'),
104
+ ('r', '아르'),
105
+ ('s', '에스'),
106
+ ('t', '티'),
107
+ ('u', '유'),
108
+ ('v', '브이'),
109
+ ('w', '더블유'),
110
+ ('x', '엑스'),
111
+ ('y', '와이'),
112
+ ('z', '제트')
113
+ ]]
114
+
115
+ # List of (Latin alphabet, bopomofo) pairs:
116
+ _latin_to_bopomofo = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
117
+ ('a', 'ㄟˉ'),
118
+ ('b', 'ㄅㄧˋ'),
119
+ ('c', 'ㄙㄧˉ'),
120
+ ('d', 'ㄉㄧˋ'),
121
+ ('e', 'ㄧˋ'),
122
+ ('f', 'ㄝˊㄈㄨˋ'),
123
+ ('g', 'ㄐㄧˋ'),
124
+ ('h', 'ㄝˇㄑㄩˋ'),
125
+ ('i', 'ㄞˋ'),
126
+ ('j', 'ㄐㄟˋ'),
127
+ ('k', 'ㄎㄟˋ'),
128
+ ('l', 'ㄝˊㄛˋ'),
129
+ ('m', 'ㄝˊㄇㄨˋ'),
130
+ ('n', 'ㄣˉ'),
131
+ ('o', 'ㄡˉ'),
132
+ ('p', 'ㄆㄧˉ'),
133
+ ('q', 'ㄎㄧㄡˉ'),
134
+ ('r', 'ㄚˋ'),
135
+ ('s', 'ㄝˊㄙˋ'),
136
+ ('t', 'ㄊㄧˋ'),
137
+ ('u', 'ㄧㄡˉ'),
138
+ ('v', 'ㄨㄧˉ'),
139
+ ('w', 'ㄉㄚˋㄅㄨˋㄌㄧㄡˋ'),
140
+ ('x', 'ㄝˉㄎㄨˋㄙˋ'),
141
+ ('y', 'ㄨㄞˋ'),
142
+ ('z', 'ㄗㄟˋ')
143
+ ]]
144
+
145
+
146
+ # List of (bopomofo, romaji) pairs:
147
+ _bopomofo_to_romaji = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
148
+ ('ㄅㄛ', 'p⁼wo'),
149
+ ('ㄆㄛ', 'pʰwo'),
150
+ ('ㄇㄛ', 'mwo'),
151
+ ('ㄈㄛ', 'fwo'),
152
+ ('ㄅ', 'p⁼'),
153
+ ('ㄆ', 'pʰ'),
154
+ ('ㄇ', 'm'),
155
+ ('ㄈ', 'f'),
156
+ ('ㄉ', 't⁼'),
157
+ ('ㄊ', 'tʰ'),
158
+ ('ㄋ', 'n'),
159
+ ('ㄌ', 'l'),
160
+ ('ㄍ', 'k⁼'),
161
+ ('ㄎ', 'kʰ'),
162
+ ('ㄏ', 'h'),
163
+ ('ㄐ', 'ʧ⁼'),
164
+ ('ㄑ', 'ʧʰ'),
165
+ ('ㄒ', 'ʃ'),
166
+ ('ㄓ', 'ʦ`⁼'),
167
+ ('ㄔ', 'ʦ`ʰ'),
168
+ ('ㄕ', 's`'),
169
+ ('ㄖ', 'ɹ`'),
170
+ ('ㄗ', 'ʦ⁼'),
171
+ ('ㄘ', 'ʦʰ'),
172
+ ('ㄙ', 's'),
173
+ ('ㄚ', 'a'),
174
+ ('ㄛ', 'o'),
175
+ ('ㄜ', 'ə'),
176
+ ('ㄝ', 'e'),
177
+ ('ㄞ', 'ai'),
178
+ ('ㄟ', 'ei'),
179
+ ('ㄠ', 'au'),
180
+ ('ㄡ', 'ou'),
181
+ ('ㄧㄢ', 'yeNN'),
182
+ ('ㄢ', 'aNN'),
183
+ ('ㄧㄣ', 'iNN'),
184
+ ('ㄣ', 'əNN'),
185
+ ('ㄤ', 'aNg'),
186
+ ('ㄧㄥ', 'iNg'),
187
+ ('ㄨㄥ', 'uNg'),
188
+ ('ㄩㄥ', 'yuNg'),
189
+ ('ㄥ', 'əNg'),
190
+ ('ㄦ', 'əɻ'),
191
+ ('ㄧ', 'i'),
192
+ ('ㄨ', 'u'),
193
+ ('ㄩ', 'ɥ'),
194
+ ('ˉ', '→'),
195
+ ('ˊ', '↑'),
196
+ ('ˇ', '↓↑'),
197
+ ('ˋ', '↓'),
198
+ ('˙', ''),
199
+ (',', ','),
200
+ ('。', '.'),
201
+ ('!', '!'),
202
+ ('?', '?'),
203
+ ('—', '-')
204
+ ]]
205
+
206
+
207
+ def expand_abbreviations(text):
208
+ for regex, replacement in _abbreviations:
209
+ text = re.sub(regex, replacement, text)
210
+ return text
211
+
212
+
213
+ def lowercase(text):
214
+ return text.lower()
215
+
216
+
217
+ def collapse_whitespace(text):
218
+ return re.sub(_whitespace_re, ' ', text)
219
+
220
+
221
+ def convert_to_ascii(text):
222
+ return unidecode(text)
223
+
224
+
225
+ def japanese_to_romaji_with_accent(text):
226
+ '''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
227
+ sentences = re.split(_japanese_marks, text)
228
+ marks = re.findall(_japanese_marks, text)
229
+ text = ''
230
+ for i, sentence in enumerate(sentences):
231
+ if re.match(_japanese_characters, sentence):
232
+ if text!='':
233
+ text+=' '
234
+ labels = pyopenjtalk.extract_fullcontext(sentence)
235
+ for n, label in enumerate(labels):
236
+ phoneme = re.search(r'\-([^\+]*)\+', label).group(1)
237
+ if phoneme not in ['sil','pau']:
238
+ text += phoneme.replace('ch','ʧ').replace('sh','ʃ').replace('cl','Q')
239
+ else:
240
+ continue
241
+ n_moras = int(re.search(r'/F:(\d+)_', label).group(1))
242
+ a1 = int(re.search(r"/A:(\-?[0-9]+)\+", label).group(1))
243
+ a2 = int(re.search(r"\+(\d+)\+", label).group(1))
244
+ a3 = int(re.search(r"\+(\d+)/", label).group(1))
245
+ if re.search(r'\-([^\+]*)\+', labels[n + 1]).group(1) in ['sil','pau']:
246
+ a2_next=-1
247
+ else:
248
+ a2_next = int(re.search(r"\+(\d+)\+", labels[n + 1]).group(1))
249
+ # Accent phrase boundary
250
+ if a3 == 1 and a2_next == 1:
251
+ text += ' '
252
+ # Falling
253
+ elif a1 == 0 and a2_next == a2 + 1 and a2 != n_moras:
254
+ text += '↓'
255
+ # Rising
256
+ elif a2 == 1 and a2_next == 2:
257
+ text += '↑'
258
+ if i<len(marks):
259
+ text += unidecode(marks[i]).replace(' ','')
260
+ return text
261
+
262
+
263
+ def latin_to_hangul(text):
264
+ for regex, replacement in _latin_to_hangul:
265
+ text = re.sub(regex, replacement, text)
266
+ return text
267
+
268
+
269
+ def divide_hangul(text):
270
+ for regex, replacement in _hangul_divided:
271
+ text = re.sub(regex, replacement, text)
272
+ return text
273
+
274
+
275
+ def hangul_number(num, sino=True):
276
+ '''Reference https://github.com/Kyubyong/g2pK'''
277
+ num = re.sub(',', '', num)
278
+
279
+ if num == '0':
280
+ return '영'
281
+ if not sino and num == '20':
282
+ return '스무'
283
+
284
+ digits = '123456789'
285
+ names = '일이삼사오육칠팔구'
286
+ digit2name = {d: n for d, n in zip(digits, names)}
287
+
288
+ modifiers = '한 두 세 네 다섯 여섯 일곱 여덟 아홉'
289
+ decimals = '열 스물 서른 마흔 쉰 예순 일흔 여든 아흔'
290
+ digit2mod = {d: mod for d, mod in zip(digits, modifiers.split())}
291
+ digit2dec = {d: dec for d, dec in zip(digits, decimals.split())}
292
+
293
+ spelledout = []
294
+ for i, digit in enumerate(num):
295
+ i = len(num) - i - 1
296
+ if sino:
297
+ if i == 0:
298
+ name = digit2name.get(digit, '')
299
+ elif i == 1:
300
+ name = digit2name.get(digit, '') + '십'
301
+ name = name.replace('일십', '십')
302
+ else:
303
+ if i == 0:
304
+ name = digit2mod.get(digit, '')
305
+ elif i == 1:
306
+ name = digit2dec.get(digit, '')
307
+ if digit == '0':
308
+ if i % 4 == 0:
309
+ last_three = spelledout[-min(3, len(spelledout)):]
310
+ if ''.join(last_three) == '':
311
+ spelledout.append('')
312
+ continue
313
+ else:
314
+ spelledout.append('')
315
+ continue
316
+ if i == 2:
317
+ name = digit2name.get(digit, '') + '백'
318
+ name = name.replace('일백', '백')
319
+ elif i == 3:
320
+ name = digit2name.get(digit, '') + '천'
321
+ name = name.replace('일천', '천')
322
+ elif i == 4:
323
+ name = digit2name.get(digit, '') + '만'
324
+ name = name.replace('일만', '만')
325
+ elif i == 5:
326
+ name = digit2name.get(digit, '') + '십'
327
+ name = name.replace('일십', '십')
328
+ elif i == 6:
329
+ name = digit2name.get(digit, '') + '백'
330
+ name = name.replace('일백', '백')
331
+ elif i == 7:
332
+ name = digit2name.get(digit, '') + '천'
333
+ name = name.replace('일천', '천')
334
+ elif i == 8:
335
+ name = digit2name.get(digit, '') + '억'
336
+ elif i == 9:
337
+ name = digit2name.get(digit, '') + '십'
338
+ elif i == 10:
339
+ name = digit2name.get(digit, '') + '백'
340
+ elif i == 11:
341
+ name = digit2name.get(digit, '') + '천'
342
+ elif i == 12:
343
+ name = digit2name.get(digit, '') + '조'
344
+ elif i == 13:
345
+ name = digit2name.get(digit, '') + '십'
346
+ elif i == 14:
347
+ name = digit2name.get(digit, '') + '백'
348
+ elif i == 15:
349
+ name = digit2name.get(digit, '') + '천'
350
+ spelledout.append(name)
351
+ return ''.join(elem for elem in spelledout)
352
+
353
+
354
+ def number_to_hangul(text):
355
+ '''Reference https://github.com/Kyubyong/g2pK'''
356
+ tokens = set(re.findall(r'(\d[\d,]*)([\uac00-\ud71f]+)', text))
357
+ for token in tokens:
358
+ num, classifier = token
359
+ if classifier[:2] in _korean_classifiers or classifier[0] in _korean_classifiers:
360
+ spelledout = hangul_number(num, sino=False)
361
+ else:
362
+ spelledout = hangul_number(num, sino=True)
363
+ text = text.replace(f'{num}{classifier}', f'{spelledout}{classifier}')
364
+ # digit by digit for remaining digits
365
+ digits = '0123456789'
366
+ names = '영일이삼사오육칠팔구'
367
+ for d, n in zip(digits, names):
368
+ text = text.replace(d, n)
369
+ return text
370
+
371
+
372
+ def number_to_chinese(text):
373
+ numbers = re.findall(r'\d+(?:\.?\d+)?', text)
374
+ for number in numbers:
375
+ text = text.replace(number, cn2an.an2cn(number),1)
376
+ return text
377
+
378
+
379
+ def chinese_to_bopomofo(text):
380
+ text=text.replace('、',',').replace(';',',').replace(':',',')
381
+ words=jieba.lcut(text,cut_all=False)
382
+ text=''
383
+ for word in words:
384
+ bopomofos=lazy_pinyin(word,BOPOMOFO)
385
+ if not re.search('[\u4e00-\u9fff]',word):
386
+ text+=word
387
+ continue
388
+ for i in range(len(bopomofos)):
389
+ if re.match('[\u3105-\u3129]',bopomofos[i][-1]):
390
+ bopomofos[i]+='ˉ'
391
+ if text!='':
392
+ text+=' '
393
+ text+=''.join(bopomofos)
394
+ return text
395
+
396
+
397
+ def latin_to_bopomofo(text):
398
+ for regex, replacement in _latin_to_bopomofo:
399
+ text = re.sub(regex, replacement, text)
400
+ return text
401
+
402
+
403
+ def bopomofo_to_romaji(text):
404
+ for regex, replacement in _bopomofo_to_romaji:
405
+ text = re.sub(regex, replacement, text)
406
+ return text
407
+
408
+
409
+ def basic_cleaners(text):
410
+ '''Basic pipeline that lowercases and collapses whitespace without transliteration.'''
411
+ text = lowercase(text)
412
+ text = collapse_whitespace(text)
413
+ return text
414
+
415
+
416
+ def transliteration_cleaners(text):
417
+ '''Pipeline for non-English text that transliterates to ASCII.'''
418
+ text = convert_to_ascii(text)
419
+ text = lowercase(text)
420
+ text = collapse_whitespace(text)
421
+ return text
422
+
423
+
424
+ def japanese_cleaners(text):
425
+ text=japanese_to_romaji_with_accent(text)
426
+ if re.match('[A-Za-z]',text[-1]):
427
+ text += '.'
428
+ return text
429
+
430
+
431
+ def japanese_cleaners2(text):
432
+ return japanese_cleaners(text).replace('ts','ʦ').replace('...','…')
433
+
434
+
435
+ def korean_cleaners(text):
436
+ '''Pipeline for Korean text'''
437
+ text = latin_to_hangul(text)
438
+ text = number_to_hangul(text)
439
+ text = j2hcj(h2j(text))
440
+ text = divide_hangul(text)
441
+ if re.match('[\u3131-\u3163]',text[-1]):
442
+ text += '.'
443
+ return text
444
+
445
+
446
+ def chinese_cleaners(text):
447
+ '''Pipeline for Chinese text'''
448
+ text=number_to_chinese(text)
449
+ text=chinese_to_bopomofo(text)
450
+ text=latin_to_bopomofo(text)
451
+ if re.match('[ˉˊˇˋ˙]',text[-1]):
452
+ text += '。'
453
+ return text
454
+
455
+
456
+ def zh_ja_mixture_cleaners(text):
457
+ chinese_texts=re.findall(r'\[ZH\].*?\[ZH\]',text)
458
+ japanese_texts=re.findall(r'\[JA\].*?\[JA\]',text)
459
+ for chinese_text in chinese_texts:
460
+ cleaned_text=number_to_chinese(chinese_text[4:-4])
461
+ cleaned_text=chinese_to_bopomofo(cleaned_text)
462
+ cleaned_text=latin_to_bopomofo(cleaned_text)
463
+ cleaned_text=bopomofo_to_romaji(cleaned_text)
464
+ cleaned_text=re.sub('i[aoe]',lambda x:'y'+x.group(0)[1:],cleaned_text)
465
+ cleaned_text=re.sub('u[aoəe]',lambda x:'w'+x.group(0)[1:],cleaned_text)
466
+ cleaned_text=re.sub('([ʦsɹ]`[⁼ʰ]?)([→↓↑]+)',lambda x:x.group(1)+'ɹ`'+x.group(2),cleaned_text).replace('ɻ','ɹ`')
467
+ cleaned_text=re.sub('([ʦs][⁼ʰ]?)([→↓↑]+)',lambda x:x.group(1)+'ɹ'+x.group(2),cleaned_text)
468
+ text = text.replace(chinese_text,cleaned_text+' ',1)
469
+ for japanese_text in japanese_texts:
470
+ cleaned_text=japanese_to_romaji_with_accent(japanese_text[4:-4]).replace('ts','ʦ').replace('u','ɯ').replace('...','…')
471
+ text = text.replace(japanese_text,cleaned_text+' ',1)
472
+ text=text[:-1]
473
+ if re.match('[A-Za-zɯɹəɥ→↓↑]',text[-1]):
474
+ text += '.'
475
+ return text
text/symbols.py ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ '''
2
+ Defines the set of symbols used in text input to the model.
3
+ '''
4
+
5
+ '''# japanese_cleaners
6
+ _pad = '_'
7
+ _punctuation = ',.!?-'
8
+ _letters = 'AEINOQUabdefghijkmnoprstuvwyzʃʧ↓↑ '
9
+ '''
10
+
11
+ '''# japanese_cleaners2
12
+ _pad = '_'
13
+ _punctuation = ',.!?-~…'
14
+ _letters = 'AEINOQUabdefghijkmnoprstuvwyzʃʧʦ↓↑ '
15
+ '''
16
+
17
+ '''# korean_cleaners
18
+ _pad = '_'
19
+ _punctuation = ',.!?…~'
20
+ _letters = 'ㄱㄴㄷㄹㅁㅂㅅㅇㅈㅊㅋㅌㅍㅎㄲㄸㅃㅆㅉㅏㅓㅗㅜㅡㅣㅐㅔ '
21
+ '''
22
+
23
+ '''# chinese_cleaners
24
+ _pad = '_'
25
+ _punctuation = ',。!?—…'
26
+ _letters = 'ㄅㄆㄇㄈㄉㄊㄋㄌㄍㄎㄏㄐㄑㄒㄓㄔㄕㄖㄗㄘㄙㄚㄛㄜㄝㄞㄟㄠㄡㄢㄣㄤㄥㄦㄧㄨㄩˉˊˇˋ˙ '
27
+ '''
28
+
29
+ # zh_ja_mixture_cleaners
30
+ _pad = '_'
31
+ _punctuation = ',.!?-~…'
32
+ _letters = 'AEINOQUabdefghijklmnoprstuvwyzʃʧʦɯɹəɥ⁼ʰ`→↓↑ '
33
+
34
+
35
+ # Export all symbols:
36
+ symbols = [_pad] + list(_punctuation) + list(_letters)
37
+
38
+ # Special symbol ids
39
+ 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 +1,225 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import sys
3
+ import argparse
4
+ import logging
5
+ import json
6
+ import subprocess
7
+ import numpy as np
8
+ import librosa
9
+ import torch
10
+
11
+ MATPLOTLIB_FLAG = False
12
+
13
+ logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
14
+ logger = logging
15
+
16
+
17
+ def load_checkpoint(checkpoint_path, model, optimizer=None):
18
+ assert os.path.isfile(checkpoint_path)
19
+ checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
20
+ iteration = checkpoint_dict['iteration']
21
+ learning_rate = checkpoint_dict['learning_rate']
22
+ if optimizer is not None:
23
+ optimizer.load_state_dict(checkpoint_dict['optimizer'])
24
+ saved_state_dict = checkpoint_dict['model']
25
+ if hasattr(model, 'module'):
26
+ state_dict = model.module.state_dict()
27
+ else:
28
+ state_dict = model.state_dict()
29
+ new_state_dict= {}
30
+ for k, v in state_dict.items():
31
+ try:
32
+ new_state_dict[k] = saved_state_dict[k]
33
+ except:
34
+ logger.info("%s is not in the checkpoint" % k)
35
+ new_state_dict[k] = v
36
+ if hasattr(model, 'module'):
37
+ model.module.load_state_dict(new_state_dict)
38
+ else:
39
+ model.load_state_dict(new_state_dict)
40
+ logger.info("Loaded checkpoint '{}' (iteration {})" .format(
41
+ checkpoint_path, iteration))
42
+ return model, optimizer, learning_rate, iteration
43
+
44
+
45
+ def plot_spectrogram_to_numpy(spectrogram):
46
+ global MATPLOTLIB_FLAG
47
+ if not MATPLOTLIB_FLAG:
48
+ import matplotlib
49
+ matplotlib.use("Agg")
50
+ MATPLOTLIB_FLAG = True
51
+ mpl_logger = logging.getLogger('matplotlib')
52
+ mpl_logger.setLevel(logging.WARNING)
53
+ import matplotlib.pylab as plt
54
+ import numpy as np
55
+
56
+ fig, ax = plt.subplots(figsize=(10,2))
57
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
58
+ interpolation='none')
59
+ plt.colorbar(im, ax=ax)
60
+ plt.xlabel("Frames")
61
+ plt.ylabel("Channels")
62
+ plt.tight_layout()
63
+
64
+ fig.canvas.draw()
65
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
66
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
67
+ plt.close()
68
+ return data
69
+
70
+
71
+ def plot_alignment_to_numpy(alignment, info=None):
72
+ global MATPLOTLIB_FLAG
73
+ if not MATPLOTLIB_FLAG:
74
+ import matplotlib
75
+ matplotlib.use("Agg")
76
+ MATPLOTLIB_FLAG = True
77
+ mpl_logger = logging.getLogger('matplotlib')
78
+ mpl_logger.setLevel(logging.WARNING)
79
+ import matplotlib.pylab as plt
80
+ import numpy as np
81
+
82
+ fig, ax = plt.subplots(figsize=(6, 4))
83
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
84
+ interpolation='none')
85
+ fig.colorbar(im, ax=ax)
86
+ xlabel = 'Decoder timestep'
87
+ if info is not None:
88
+ xlabel += '\n\n' + info
89
+ plt.xlabel(xlabel)
90
+ plt.ylabel('Encoder timestep')
91
+ plt.tight_layout()
92
+
93
+ fig.canvas.draw()
94
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
95
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
96
+ plt.close()
97
+ return data
98
+
99
+
100
+ def load_audio_to_torch(full_path, target_sampling_rate):
101
+ audio, sampling_rate = librosa.load(full_path, sr=target_sampling_rate, mono=True)
102
+ return torch.FloatTensor(audio.astype(np.float32))
103
+
104
+
105
+ def load_filepaths_and_text(filename, split="|"):
106
+ with open(filename, encoding='utf-8') as f:
107
+ filepaths_and_text = [line.strip().split(split) for line in f]
108
+ return filepaths_and_text
109
+
110
+
111
+ def get_hparams(init=True):
112
+ parser = argparse.ArgumentParser()
113
+ parser.add_argument('-c', '--config', type=str, default="./configs/base.json",
114
+ help='JSON file for configuration')
115
+ parser.add_argument('-m', '--model', type=str, required=True,
116
+ help='Model name')
117
+
118
+ args = parser.parse_args()
119
+ model_dir = os.path.join("./logs", args.model)
120
+
121
+ if not os.path.exists(model_dir):
122
+ os.makedirs(model_dir)
123
+
124
+ config_path = args.config
125
+ config_save_path = os.path.join(model_dir, "config.json")
126
+ if init:
127
+ with open(config_path, "r") as f:
128
+ data = f.read()
129
+ with open(config_save_path, "w") as f:
130
+ f.write(data)
131
+ else:
132
+ with open(config_save_path, "r") as f:
133
+ data = f.read()
134
+ config = json.loads(data)
135
+
136
+ hparams = HParams(**config)
137
+ hparams.model_dir = model_dir
138
+ return hparams
139
+
140
+
141
+ def get_hparams_from_dir(model_dir):
142
+ config_save_path = os.path.join(model_dir, "config.json")
143
+ with open(config_save_path, "r") as f:
144
+ data = f.read()
145
+ config = json.loads(data)
146
+
147
+ hparams =HParams(**config)
148
+ hparams.model_dir = model_dir
149
+ return hparams
150
+
151
+
152
+ def get_hparams_from_file(config_path):
153
+ with open(config_path, "r") as f:
154
+ data = f.read()
155
+ config = json.loads(data)
156
+
157
+ hparams =HParams(**config)
158
+ return hparams
159
+
160
+
161
+ def check_git_hash(model_dir):
162
+ source_dir = os.path.dirname(os.path.realpath(__file__))
163
+ if not os.path.exists(os.path.join(source_dir, ".git")):
164
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
165
+ source_dir
166
+ ))
167
+ return
168
+
169
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
170
+
171
+ path = os.path.join(model_dir, "githash")
172
+ if os.path.exists(path):
173
+ saved_hash = open(path).read()
174
+ if saved_hash != cur_hash:
175
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
176
+ saved_hash[:8], cur_hash[:8]))
177
+ else:
178
+ open(path, "w").write(cur_hash)
179
+
180
+
181
+ def get_logger(model_dir, filename="train.log"):
182
+ global logger
183
+ logger = logging.getLogger(os.path.basename(model_dir))
184
+ logger.setLevel(logging.DEBUG)
185
+
186
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
187
+ if not os.path.exists(model_dir):
188
+ os.makedirs(model_dir)
189
+ h = logging.FileHandler(os.path.join(model_dir, filename))
190
+ h.setLevel(logging.DEBUG)
191
+ h.setFormatter(formatter)
192
+ logger.addHandler(h)
193
+ return logger
194
+
195
+
196
+ class HParams():
197
+ def __init__(self, **kwargs):
198
+ for k, v in kwargs.items():
199
+ if type(v) == dict:
200
+ v = HParams(**v)
201
+ self[k] = v
202
+
203
+ def keys(self):
204
+ return self.__dict__.keys()
205
+
206
+ def items(self):
207
+ return self.__dict__.items()
208
+
209
+ def values(self):
210
+ return self.__dict__.values()
211
+
212
+ def __len__(self):
213
+ return len(self.__dict__)
214
+
215
+ def __getitem__(self, key):
216
+ return getattr(self, key)
217
+
218
+ def __setitem__(self, key, value):
219
+ return setattr(self, key, value)
220
+
221
+ def __contains__(self, key):
222
+ return key in self.__dict__
223
+
224
+ def __repr__(self):
225
+ return self.__dict__.__repr__()