Commit
•
2b3664e
1
Parent(s):
ed8b240
init
Browse files- .gitignore +382 -0
- app.py +156 -0
- attentions.py +547 -0
- commons.py +164 -0
- configs/finetune_speaker.json +202 -0
- configs/modified_finetune_speaker.json +321 -0
- configs/uma_trilingual.json +54 -0
- mel_processing.py +112 -0
- models.py +725 -0
- models/G_9700.pth +3 -0
- models/config.json +321 -0
- modules.py +390 -0
- monotonic_align/__init__.py +19 -0
- monotonic_align/core.c +0 -0
- text/LICENSE +19 -0
- text/__init__.py +60 -0
- text/cantonese.py +59 -0
- text/cleaners.py +134 -0
- text/english.py +188 -0
- text/japanese.py +153 -0
- text/korean.py +210 -0
- text/mandarin.py +326 -0
- text/ngu_dialect.py +30 -0
- text/sanskrit.py +62 -0
- text/shanghainese.py +64 -0
- text/symbols.py +76 -0
- text/thai.py +44 -0
- transforms.py +193 -0
- utils.py +486 -0
.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
|
app.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
from torch import no_grad, LongTensor
|
5 |
+
import argparse
|
6 |
+
import commons
|
7 |
+
from mel_processing import spectrogram_torch
|
8 |
+
import utils
|
9 |
+
from models import SynthesizerTrn
|
10 |
+
import gradio as gr
|
11 |
+
import librosa
|
12 |
+
import webbrowser
|
13 |
+
|
14 |
+
from text import text_to_sequence, _clean_text
|
15 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
16 |
+
language_marks = {
|
17 |
+
"Japanese": "",
|
18 |
+
"日本語": "[JA]",
|
19 |
+
"简体中文": "[ZH]",
|
20 |
+
"English": "[EN]",
|
21 |
+
"Mix": "",
|
22 |
+
}
|
23 |
+
lang = ['日本語', '简体中文', 'English', 'Mix']
|
24 |
+
|
25 |
+
|
26 |
+
def get_text(text, hps, is_symbol):
|
27 |
+
text_norm = text_to_sequence(
|
28 |
+
text, hps.symbols, [] if is_symbol else hps.data.text_cleaners)
|
29 |
+
if hps.data.add_blank:
|
30 |
+
text_norm = commons.intersperse(text_norm, 0)
|
31 |
+
text_norm = LongTensor(text_norm)
|
32 |
+
return text_norm
|
33 |
+
|
34 |
+
|
35 |
+
def create_tts_fn(model, hps, speaker_ids):
|
36 |
+
def tts_fn(text, speaker, language, speed):
|
37 |
+
if language is not None:
|
38 |
+
text = language_marks[language] + text + language_marks[language]
|
39 |
+
speaker_id = speaker_ids[speaker]
|
40 |
+
stn_tst = get_text(text, hps, False)
|
41 |
+
with no_grad():
|
42 |
+
x_tst = stn_tst.unsqueeze(0).to(device)
|
43 |
+
x_tst_lengths = LongTensor([stn_tst.size(0)]).to(device)
|
44 |
+
sid = LongTensor([speaker_id]).to(device)
|
45 |
+
audio = model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8,
|
46 |
+
length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy()
|
47 |
+
del stn_tst, x_tst, x_tst_lengths, sid
|
48 |
+
return "Success", (hps.data.sampling_rate, audio)
|
49 |
+
|
50 |
+
return tts_fn
|
51 |
+
|
52 |
+
|
53 |
+
def create_vc_fn(model, hps, speaker_ids):
|
54 |
+
def vc_fn(original_speaker, target_speaker, record_audio, upload_audio):
|
55 |
+
input_audio = record_audio if record_audio is not None else upload_audio
|
56 |
+
if input_audio is None:
|
57 |
+
return "You need to record or upload an audio", None
|
58 |
+
sampling_rate, audio = input_audio
|
59 |
+
original_speaker_id = speaker_ids[original_speaker]
|
60 |
+
target_speaker_id = speaker_ids[target_speaker]
|
61 |
+
|
62 |
+
audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
|
63 |
+
if len(audio.shape) > 1:
|
64 |
+
audio = librosa.to_mono(audio.transpose(1, 0))
|
65 |
+
if sampling_rate != hps.data.sampling_rate:
|
66 |
+
audio = librosa.resample(
|
67 |
+
audio, orig_sr=sampling_rate, target_sr=hps.data.sampling_rate)
|
68 |
+
with no_grad():
|
69 |
+
y = torch.FloatTensor(audio)
|
70 |
+
y = y / max(-y.min(), y.max()) / 0.99
|
71 |
+
y = y.to(device)
|
72 |
+
y = y.unsqueeze(0)
|
73 |
+
spec = spectrogram_torch(y, hps.data.filter_length,
|
74 |
+
hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,
|
75 |
+
center=False).to(device)
|
76 |
+
spec_lengths = LongTensor([spec.size(-1)]).to(device)
|
77 |
+
sid_src = LongTensor([original_speaker_id]).to(device)
|
78 |
+
sid_tgt = LongTensor([target_speaker_id]).to(device)
|
79 |
+
audio = model.voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt)[0][
|
80 |
+
0, 0].data.cpu().float().numpy()
|
81 |
+
del y, spec, spec_lengths, sid_src, sid_tgt
|
82 |
+
return "Success", (hps.data.sampling_rate, audio)
|
83 |
+
|
84 |
+
return vc_fn
|
85 |
+
|
86 |
+
|
87 |
+
if __name__ == "__main__":
|
88 |
+
parser = argparse.ArgumentParser()
|
89 |
+
parser.add_argument("--model_dir", default="./models/G_9700.pth",
|
90 |
+
help="directory to your fine-tuned model")
|
91 |
+
parser.add_argument("--config_dir", default="./configs/modified_finetune_speaker.json",
|
92 |
+
help="directory to your model config file")
|
93 |
+
parser.add_argument("--share", default=False,
|
94 |
+
help="make link public (used in colab)")
|
95 |
+
|
96 |
+
args = parser.parse_args()
|
97 |
+
hps = utils.get_hparams_from_file(args.config_dir)
|
98 |
+
|
99 |
+
net_g = SynthesizerTrn(
|
100 |
+
len(hps.symbols),
|
101 |
+
hps.data.filter_length // 2 + 1,
|
102 |
+
hps.train.segment_size // hps.data.hop_length,
|
103 |
+
n_speakers=hps.data.n_speakers,
|
104 |
+
**hps.model).to(device)
|
105 |
+
_ = net_g.eval()
|
106 |
+
|
107 |
+
_ = utils.load_checkpoint(args.model_dir, net_g, None)
|
108 |
+
speaker_ids = hps.speakers
|
109 |
+
speakers = list(hps.speakers.keys())
|
110 |
+
tts_fn = create_tts_fn(net_g, hps, speaker_ids)
|
111 |
+
vc_fn = create_vc_fn(net_g, hps, speaker_ids)
|
112 |
+
app = gr.Blocks()
|
113 |
+
with app:
|
114 |
+
with gr.Tab("Text-to-Speech"):
|
115 |
+
with gr.Row():
|
116 |
+
with gr.Column():
|
117 |
+
textbox = gr.TextArea(label="Text",
|
118 |
+
placeholder="Type your sentence here",
|
119 |
+
value="新たなキャラを解放できるようになったようですね。", elem_id=f"tts-input")
|
120 |
+
# select character
|
121 |
+
char_dropdown = gr.Dropdown(
|
122 |
+
choices=speakers, value=speakers[0], label='character')
|
123 |
+
language_dropdown = gr.Dropdown(
|
124 |
+
choices=lang, value=lang[0], label='language')
|
125 |
+
duration_slider = gr.Slider(minimum=0.1, maximum=5, value=1, step=0.1,
|
126 |
+
label='速度 Speed')
|
127 |
+
with gr.Column():
|
128 |
+
text_output = gr.Textbox(label="Message")
|
129 |
+
audio_output = gr.Audio(
|
130 |
+
label="Output Audio", elem_id="tts-audio")
|
131 |
+
btn = gr.Button("Generate!")
|
132 |
+
btn.click(tts_fn,
|
133 |
+
inputs=[textbox, char_dropdown,
|
134 |
+
language_dropdown, duration_slider, ],
|
135 |
+
outputs=[text_output, audio_output])
|
136 |
+
with gr.Tab("Voice Conversion"):
|
137 |
+
gr.Markdown("""
|
138 |
+
录制或上传声音,并选择要转换的音色。
|
139 |
+
""")
|
140 |
+
with gr.Column():
|
141 |
+
record_audio = gr.Audio(
|
142 |
+
label="record your voice", source="microphone")
|
143 |
+
upload_audio = gr.Audio(
|
144 |
+
label="or upload audio here", source="upload")
|
145 |
+
source_speaker = gr.Dropdown(
|
146 |
+
choices=speakers, value=speakers[0], label="source speaker")
|
147 |
+
target_speaker = gr.Dropdown(
|
148 |
+
choices=speakers, value=speakers[0], label="target speaker")
|
149 |
+
with gr.Column():
|
150 |
+
message_box = gr.Textbox(label="Message")
|
151 |
+
converted_audio = gr.Audio(label='converted audio')
|
152 |
+
btn = gr.Button("Convert!")
|
153 |
+
btn.click(vc_fn, inputs=[source_speaker, target_speaker, record_audio, upload_audio],
|
154 |
+
outputs=[message_box, converted_audio])
|
155 |
+
webbrowser.open("http://127.0.0.1:7899")
|
156 |
+
app.launch(share=args.share, server_port=7899)
|
attentions.py
ADDED
@@ -0,0 +1,547 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
from torch import nn
|
6 |
+
from torch.nn import functional as F
|
7 |
+
|
8 |
+
import commons
|
9 |
+
import modules
|
10 |
+
from modules import LayerNorm
|
11 |
+
|
12 |
+
from loralib_tmp import layers
|
13 |
+
import loralib as lora
|
14 |
+
|
15 |
+
|
16 |
+
class Encoder(nn.Module):
|
17 |
+
def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
|
18 |
+
super().__init__()
|
19 |
+
self.hidden_channels = hidden_channels
|
20 |
+
self.filter_channels = filter_channels
|
21 |
+
self.n_heads = n_heads
|
22 |
+
self.n_layers = n_layers
|
23 |
+
self.kernel_size = kernel_size
|
24 |
+
self.p_dropout = p_dropout
|
25 |
+
self.window_size = window_size
|
26 |
+
|
27 |
+
self.drop = nn.Dropout(p_dropout)
|
28 |
+
self.attn_layers = nn.ModuleList()
|
29 |
+
self.norm_layers_1 = nn.ModuleList()
|
30 |
+
self.ffn_layers = nn.ModuleList()
|
31 |
+
self.norm_layers_2 = nn.ModuleList()
|
32 |
+
for i in range(self.n_layers):
|
33 |
+
self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
|
34 |
+
self.norm_layers_1.append(LayerNorm(hidden_channels))
|
35 |
+
self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
|
36 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
37 |
+
|
38 |
+
def forward(self, x, x_mask):
|
39 |
+
attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
40 |
+
x = x * x_mask
|
41 |
+
for i in range(self.n_layers):
|
42 |
+
y = self.attn_layers[i](x, x, attn_mask)
|
43 |
+
y = self.drop(y)
|
44 |
+
x = self.norm_layers_1[i](x + y)
|
45 |
+
|
46 |
+
y = self.ffn_layers[i](x, x_mask)
|
47 |
+
y = self.drop(y)
|
48 |
+
x = self.norm_layers_2[i](x + y)
|
49 |
+
x = x * x_mask
|
50 |
+
return x
|
51 |
+
|
52 |
+
class Encoder_lora(nn.Module):
|
53 |
+
def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
|
54 |
+
super().__init__()
|
55 |
+
self.hidden_channels = hidden_channels
|
56 |
+
self.filter_channels = filter_channels
|
57 |
+
self.n_heads = n_heads
|
58 |
+
self.n_layers = n_layers
|
59 |
+
self.kernel_size = kernel_size
|
60 |
+
self.p_dropout = p_dropout
|
61 |
+
self.window_size = window_size
|
62 |
+
|
63 |
+
self.drop = nn.Dropout(p_dropout)
|
64 |
+
self.attn_layers = nn.ModuleList()
|
65 |
+
self.norm_layers_1 = nn.ModuleList()
|
66 |
+
self.ffn_layers = nn.ModuleList()
|
67 |
+
self.norm_layers_2 = nn.ModuleList()
|
68 |
+
for i in range(self.n_layers):
|
69 |
+
self.attn_layers.append(MultiHeadAttention_lora(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
|
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))
|
72 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
73 |
+
|
74 |
+
def forward(self, x, x_mask):
|
75 |
+
attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
76 |
+
x = x * x_mask
|
77 |
+
for i in range(self.n_layers):
|
78 |
+
y = self.attn_layers[i](x, x, attn_mask)
|
79 |
+
y = self.drop(y)
|
80 |
+
x = self.norm_layers_1[i](x + y)
|
81 |
+
|
82 |
+
y = self.ffn_layers[i](x, x_mask)
|
83 |
+
y = self.drop(y)
|
84 |
+
x = self.norm_layers_2[i](x + y)
|
85 |
+
x = x * x_mask
|
86 |
+
return x
|
87 |
+
|
88 |
+
class Decoder(nn.Module):
|
89 |
+
def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
|
90 |
+
super().__init__()
|
91 |
+
self.hidden_channels = hidden_channels
|
92 |
+
self.filter_channels = filter_channels
|
93 |
+
self.n_heads = n_heads
|
94 |
+
self.n_layers = n_layers
|
95 |
+
self.kernel_size = kernel_size
|
96 |
+
self.p_dropout = p_dropout
|
97 |
+
self.proximal_bias = proximal_bias
|
98 |
+
self.proximal_init = proximal_init
|
99 |
+
|
100 |
+
self.drop = nn.Dropout(p_dropout)
|
101 |
+
self.self_attn_layers = nn.ModuleList()
|
102 |
+
self.norm_layers_0 = nn.ModuleList()
|
103 |
+
self.encdec_attn_layers = nn.ModuleList()
|
104 |
+
self.norm_layers_1 = nn.ModuleList()
|
105 |
+
self.ffn_layers = nn.ModuleList()
|
106 |
+
self.norm_layers_2 = nn.ModuleList()
|
107 |
+
for i in range(self.n_layers):
|
108 |
+
self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
|
109 |
+
self.norm_layers_0.append(LayerNorm(hidden_channels))
|
110 |
+
self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
|
111 |
+
self.norm_layers_1.append(LayerNorm(hidden_channels))
|
112 |
+
self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
|
113 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
114 |
+
|
115 |
+
def forward(self, x, x_mask, h, h_mask):
|
116 |
+
"""
|
117 |
+
x: decoder input
|
118 |
+
h: encoder output
|
119 |
+
"""
|
120 |
+
self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
|
121 |
+
encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
122 |
+
x = x * x_mask
|
123 |
+
for i in range(self.n_layers):
|
124 |
+
y = self.self_attn_layers[i](x, x, self_attn_mask)
|
125 |
+
y = self.drop(y)
|
126 |
+
x = self.norm_layers_0[i](x + y)
|
127 |
+
|
128 |
+
y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
|
129 |
+
y = self.drop(y)
|
130 |
+
x = self.norm_layers_1[i](x + y)
|
131 |
+
|
132 |
+
y = self.ffn_layers[i](x, x_mask)
|
133 |
+
y = self.drop(y)
|
134 |
+
x = self.norm_layers_2[i](x + y)
|
135 |
+
x = x * x_mask
|
136 |
+
return x
|
137 |
+
|
138 |
+
# class Decoder_lora(nn.Module):
|
139 |
+
# def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
|
140 |
+
# super().__init__()
|
141 |
+
# self.hidden_channels = hidden_channels
|
142 |
+
# self.filter_channels = filter_channels
|
143 |
+
# self.n_heads = n_heads
|
144 |
+
# self.n_layers = n_layers
|
145 |
+
# self.kernel_size = kernel_size
|
146 |
+
# self.p_dropout = p_dropout
|
147 |
+
# self.proximal_bias = proximal_bias
|
148 |
+
# self.proximal_init = proximal_init
|
149 |
+
|
150 |
+
# self.drop = nn.Dropout(p_dropout)
|
151 |
+
# self.self_attn_layers = nn.ModuleList()
|
152 |
+
# self.norm_layers_0 = nn.ModuleList()
|
153 |
+
# self.encdec_attn_layers = nn.ModuleList()
|
154 |
+
# self.norm_layers_1 = nn.ModuleList()
|
155 |
+
# self.ffn_layers = nn.ModuleList()
|
156 |
+
# self.norm_layers_2 = nn.ModuleList()
|
157 |
+
# for i in range(self.n_layers):
|
158 |
+
# self.self_attn_layers.append(MultiHeadAttention_lora(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
|
159 |
+
# self.norm_layers_0.append(LayerNorm(hidden_channels))
|
160 |
+
# self.encdec_attn_layers.append(MultiHeadAttention_lora(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
|
161 |
+
# self.norm_layers_1.append(LayerNorm(hidden_channels))
|
162 |
+
# self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
|
163 |
+
# self.norm_layers_2.append(LayerNorm(hidden_channels))
|
164 |
+
|
165 |
+
# def forward(self, x, x_mask, h, h_mask):
|
166 |
+
# """
|
167 |
+
# x: decoder input
|
168 |
+
# h: encoder output
|
169 |
+
# """
|
170 |
+
# self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
|
171 |
+
# encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
172 |
+
# x = x * x_mask
|
173 |
+
# for i in range(self.n_layers):
|
174 |
+
# y = self.self_attn_layers[i](x, x, self_attn_mask)
|
175 |
+
# y = self.drop(y)
|
176 |
+
# x = self.norm_layers_0[i](x + y)
|
177 |
+
|
178 |
+
# y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
|
179 |
+
# y = self.drop(y)
|
180 |
+
# x = self.norm_layers_1[i](x + y)
|
181 |
+
|
182 |
+
# y = self.ffn_layers[i](x, x_mask)
|
183 |
+
# y = self.drop(y)
|
184 |
+
# x = self.norm_layers_2[i](x + y)
|
185 |
+
# x = x * x_mask
|
186 |
+
# return x
|
187 |
+
|
188 |
+
class MultiHeadAttention_lora(nn.Module):
|
189 |
+
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):
|
190 |
+
super().__init__()
|
191 |
+
assert channels % n_heads == 0
|
192 |
+
|
193 |
+
self.channels = channels
|
194 |
+
self.out_channels = out_channels
|
195 |
+
self.n_heads = n_heads
|
196 |
+
self.p_dropout = p_dropout
|
197 |
+
self.window_size = window_size
|
198 |
+
self.heads_share = heads_share
|
199 |
+
self.block_length = block_length
|
200 |
+
self.proximal_bias = proximal_bias
|
201 |
+
self.proximal_init = proximal_init
|
202 |
+
self.attn = None
|
203 |
+
|
204 |
+
self.k_channels = channels // n_heads
|
205 |
+
self.conv_q = nn.Conv1d(channels, channels, 1)
|
206 |
+
self.conv_k = nn.Conv1d(channels, channels, 1)
|
207 |
+
self.conv_v = nn.Conv1d(channels, channels, 1)
|
208 |
+
self.conv_o = nn.Conv1d(channels, out_channels, 1)
|
209 |
+
self.drop = nn.Dropout(p_dropout)
|
210 |
+
|
211 |
+
if window_size is not None:
|
212 |
+
n_heads_rel = 1 if heads_share else n_heads
|
213 |
+
rel_stddev = self.k_channels**-0.5
|
214 |
+
self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
|
215 |
+
self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
|
216 |
+
|
217 |
+
nn.init.xavier_uniform_(self.conv_q.weight)
|
218 |
+
nn.init.xavier_uniform_(self.conv_k.weight)
|
219 |
+
nn.init.xavier_uniform_(self.conv_v.weight)
|
220 |
+
if proximal_init:
|
221 |
+
with torch.no_grad():
|
222 |
+
self.conv_k.weight.copy_(self.conv_q.weight)
|
223 |
+
self.conv_k.bias.copy_(self.conv_q.bias)
|
224 |
+
|
225 |
+
def forward(self, x, c, attn_mask=None):
|
226 |
+
q = self.conv_q(x)
|
227 |
+
k = self.conv_k(c)
|
228 |
+
v = self.conv_v(c)
|
229 |
+
|
230 |
+
x, self.attn = self.attention(q, k, v, mask=attn_mask)
|
231 |
+
x = self.conv_o(x)
|
232 |
+
return x
|
233 |
+
|
234 |
+
def attention(self, query, key, value, mask=None):
|
235 |
+
# reshape [b, d, t] -> [b, n_h, t, d_k]
|
236 |
+
b, d, t_s, t_t = (*key.size(), query.size(2))
|
237 |
+
query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
|
238 |
+
key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
239 |
+
value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
240 |
+
|
241 |
+
scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
|
242 |
+
if self.window_size is not None:
|
243 |
+
assert t_s == t_t, "Relative attention is only available for self-attention."
|
244 |
+
key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
|
245 |
+
rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
|
246 |
+
scores_local = self._relative_position_to_absolute_position(rel_logits)
|
247 |
+
scores = scores + scores_local
|
248 |
+
if self.proximal_bias:
|
249 |
+
assert t_s == t_t, "Proximal bias is only available for self-attention."
|
250 |
+
scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
|
251 |
+
if mask is not None:
|
252 |
+
scores = scores.masked_fill(mask == 0, -1e4)
|
253 |
+
if self.block_length is not None:
|
254 |
+
assert t_s == t_t, "Local attention is only available for self-attention."
|
255 |
+
block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
|
256 |
+
scores = scores.masked_fill(block_mask == 0, -1e4)
|
257 |
+
p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
|
258 |
+
p_attn = self.drop(p_attn)
|
259 |
+
output = torch.matmul(p_attn, value)
|
260 |
+
if self.window_size is not None:
|
261 |
+
relative_weights = self._absolute_position_to_relative_position(p_attn)
|
262 |
+
value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
|
263 |
+
output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
|
264 |
+
output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
|
265 |
+
return output, p_attn
|
266 |
+
|
267 |
+
def _matmul_with_relative_values(self, x, y):
|
268 |
+
"""
|
269 |
+
x: [b, h, l, m]
|
270 |
+
y: [h or 1, m, d]
|
271 |
+
ret: [b, h, l, d]
|
272 |
+
"""
|
273 |
+
ret = torch.matmul(x, y.unsqueeze(0))
|
274 |
+
return ret
|
275 |
+
|
276 |
+
def _matmul_with_relative_keys(self, x, y):
|
277 |
+
"""
|
278 |
+
x: [b, h, l, d]
|
279 |
+
y: [h or 1, m, d]
|
280 |
+
ret: [b, h, l, m]
|
281 |
+
"""
|
282 |
+
ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
|
283 |
+
return ret
|
284 |
+
|
285 |
+
def _get_relative_embeddings(self, relative_embeddings, length):
|
286 |
+
max_relative_position = 2 * self.window_size + 1
|
287 |
+
# Pad first before slice to avoid using cond ops.
|
288 |
+
pad_length = max(length - (self.window_size + 1), 0)
|
289 |
+
slice_start_position = max((self.window_size + 1) - length, 0)
|
290 |
+
slice_end_position = slice_start_position + 2 * length - 1
|
291 |
+
if pad_length > 0:
|
292 |
+
padded_relative_embeddings = F.pad(
|
293 |
+
relative_embeddings,
|
294 |
+
commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
|
295 |
+
else:
|
296 |
+
padded_relative_embeddings = relative_embeddings
|
297 |
+
used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
|
298 |
+
return used_relative_embeddings
|
299 |
+
|
300 |
+
def _relative_position_to_absolute_position(self, x):
|
301 |
+
"""
|
302 |
+
x: [b, h, l, 2*l-1]
|
303 |
+
ret: [b, h, l, l]
|
304 |
+
"""
|
305 |
+
batch, heads, length, _ = x.size()
|
306 |
+
# Concat columns of pad to shift from relative to absolute indexing.
|
307 |
+
x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
|
308 |
+
|
309 |
+
# Concat extra elements so to add up to shape (len+1, 2*len-1).
|
310 |
+
x_flat = x.view([batch, heads, length * 2 * length])
|
311 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
|
312 |
+
|
313 |
+
# Reshape and slice out the padded elements.
|
314 |
+
x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
|
315 |
+
return x_final
|
316 |
+
|
317 |
+
def _absolute_position_to_relative_position(self, x):
|
318 |
+
"""
|
319 |
+
x: [b, h, l, l]
|
320 |
+
ret: [b, h, l, 2*l-1]
|
321 |
+
"""
|
322 |
+
batch, heads, length, _ = x.size()
|
323 |
+
# padd along column
|
324 |
+
x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
|
325 |
+
x_flat = x.view([batch, heads, length**2 + length*(length -1)])
|
326 |
+
# add 0's in the beginning that will skew the elements after reshape
|
327 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
|
328 |
+
x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
|
329 |
+
return x_final
|
330 |
+
|
331 |
+
def _attention_bias_proximal(self, length):
|
332 |
+
"""Bias for self-attention to encourage attention to close positions.
|
333 |
+
Args:
|
334 |
+
length: an integer scalar.
|
335 |
+
Returns:
|
336 |
+
a Tensor with shape [1, 1, length, length]
|
337 |
+
"""
|
338 |
+
r = torch.arange(length, dtype=torch.float32)
|
339 |
+
diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
|
340 |
+
return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
|
341 |
+
|
342 |
+
# 注意改回去
|
343 |
+
class MultiHeadAttention(nn.Module):
|
344 |
+
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):
|
345 |
+
super().__init__()
|
346 |
+
assert channels % n_heads == 0
|
347 |
+
|
348 |
+
self.channels = channels
|
349 |
+
self.out_channels = out_channels
|
350 |
+
self.n_heads = n_heads
|
351 |
+
self.p_dropout = p_dropout
|
352 |
+
self.window_size = window_size
|
353 |
+
self.heads_share = heads_share
|
354 |
+
self.block_length = block_length
|
355 |
+
self.proximal_bias = proximal_bias
|
356 |
+
self.proximal_init = proximal_init
|
357 |
+
self.attn = None
|
358 |
+
|
359 |
+
self.k_channels = channels // n_heads
|
360 |
+
self.conv_q = nn.Conv1d(channels, channels, 1)
|
361 |
+
self.conv_k = nn.Conv1d(channels, channels, 1)
|
362 |
+
self.conv_v = nn.Conv1d(channels, channels, 1)
|
363 |
+
self.conv_o = nn.Conv1d(channels, out_channels, 1)
|
364 |
+
self.drop = nn.Dropout(p_dropout)
|
365 |
+
|
366 |
+
if window_size is not None:
|
367 |
+
n_heads_rel = 1 if heads_share else n_heads
|
368 |
+
rel_stddev = self.k_channels**-0.5
|
369 |
+
self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
|
370 |
+
self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
|
371 |
+
|
372 |
+
nn.init.xavier_uniform_(self.conv_q.weight)
|
373 |
+
nn.init.xavier_uniform_(self.conv_k.weight)
|
374 |
+
nn.init.xavier_uniform_(self.conv_v.weight)
|
375 |
+
if proximal_init:
|
376 |
+
with torch.no_grad():
|
377 |
+
self.conv_k.weight.copy_(self.conv_q.weight)
|
378 |
+
self.conv_k.bias.copy_(self.conv_q.bias)
|
379 |
+
|
380 |
+
def forward(self, x, c, attn_mask=None):
|
381 |
+
q = self.conv_q(x)
|
382 |
+
k = self.conv_k(c)
|
383 |
+
v = self.conv_v(c)
|
384 |
+
|
385 |
+
x, self.attn = self.attention(q, k, v, mask=attn_mask)
|
386 |
+
|
387 |
+
x = self.conv_o(x)
|
388 |
+
return x
|
389 |
+
|
390 |
+
def attention(self, query, key, value, mask=None):
|
391 |
+
# reshape [b, d, t] -> [b, n_h, t, d_k]
|
392 |
+
b, d, t_s, t_t = (*key.size(), query.size(2))
|
393 |
+
query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
|
394 |
+
key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
395 |
+
value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
396 |
+
|
397 |
+
scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
|
398 |
+
if self.window_size is not None:
|
399 |
+
assert t_s == t_t, "Relative attention is only available for self-attention."
|
400 |
+
key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
|
401 |
+
rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
|
402 |
+
scores_local = self._relative_position_to_absolute_position(rel_logits)
|
403 |
+
scores = scores + scores_local
|
404 |
+
if self.proximal_bias:
|
405 |
+
assert t_s == t_t, "Proximal bias is only available for self-attention."
|
406 |
+
scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
|
407 |
+
if mask is not None:
|
408 |
+
scores = scores.masked_fill(mask == 0, -1e4)
|
409 |
+
if self.block_length is not None:
|
410 |
+
assert t_s == t_t, "Local attention is only available for self-attention."
|
411 |
+
block_mask = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
|
412 |
+
scores = scores.masked_fill(block_mask == 0, -1e4)
|
413 |
+
p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
|
414 |
+
p_attn = self.drop(p_attn)
|
415 |
+
output = torch.matmul(p_attn, value)
|
416 |
+
if self.window_size is not None:
|
417 |
+
relative_weights = self._absolute_position_to_relative_position(p_attn)
|
418 |
+
value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
|
419 |
+
output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
|
420 |
+
output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
|
421 |
+
return output, p_attn
|
422 |
+
|
423 |
+
def _matmul_with_relative_values(self, x, y):
|
424 |
+
"""
|
425 |
+
x: [b, h, l, m]
|
426 |
+
y: [h or 1, m, d]
|
427 |
+
ret: [b, h, l, d]
|
428 |
+
"""
|
429 |
+
ret = torch.matmul(x, y.unsqueeze(0))
|
430 |
+
return ret
|
431 |
+
|
432 |
+
def _matmul_with_relative_keys(self, x, y):
|
433 |
+
"""
|
434 |
+
x: [b, h, l, d]
|
435 |
+
y: [h or 1, m, d]
|
436 |
+
ret: [b, h, l, m]
|
437 |
+
"""
|
438 |
+
ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
|
439 |
+
return ret
|
440 |
+
|
441 |
+
def _get_relative_embeddings(self, relative_embeddings, length):
|
442 |
+
max_relative_position = 2 * self.window_size + 1
|
443 |
+
# Pad first before slice to avoid using cond ops.
|
444 |
+
pad_length = max(length - (self.window_size + 1), 0)
|
445 |
+
slice_start_position = max((self.window_size + 1) - length, 0)
|
446 |
+
slice_end_position = slice_start_position + 2 * length - 1
|
447 |
+
if pad_length > 0:
|
448 |
+
padded_relative_embeddings = F.pad(
|
449 |
+
relative_embeddings,
|
450 |
+
commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
|
451 |
+
else:
|
452 |
+
padded_relative_embeddings = relative_embeddings
|
453 |
+
used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
|
454 |
+
return used_relative_embeddings
|
455 |
+
|
456 |
+
def _relative_position_to_absolute_position(self, x):
|
457 |
+
"""
|
458 |
+
x: [b, h, l, 2*l-1]
|
459 |
+
ret: [b, h, l, l]
|
460 |
+
"""
|
461 |
+
batch, heads, length, _ = x.size()
|
462 |
+
# Concat columns of pad to shift from relative to absolute indexing.
|
463 |
+
x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
|
464 |
+
|
465 |
+
# Concat extra elements so to add up to shape (len+1, 2*len-1).
|
466 |
+
x_flat = x.view([batch, heads, length * 2 * length])
|
467 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
|
468 |
+
|
469 |
+
# Reshape and slice out the padded elements.
|
470 |
+
x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
|
471 |
+
return x_final
|
472 |
+
|
473 |
+
def _absolute_position_to_relative_position(self, x):
|
474 |
+
"""
|
475 |
+
x: [b, h, l, l]
|
476 |
+
ret: [b, h, l, 2*l-1]
|
477 |
+
"""
|
478 |
+
batch, heads, length, _ = x.size()
|
479 |
+
# padd along column
|
480 |
+
x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
|
481 |
+
x_flat = x.view([batch, heads, length**2 + length*(length -1)])
|
482 |
+
# add 0's in the beginning that will skew the elements after reshape
|
483 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
|
484 |
+
x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
|
485 |
+
return x_final
|
486 |
+
|
487 |
+
def _attention_bias_proximal(self, length):
|
488 |
+
"""Bias for self-attention to encourage attention to close positions.
|
489 |
+
Args:
|
490 |
+
length: an integer scalar.
|
491 |
+
Returns:
|
492 |
+
a Tensor with shape [1, 1, length, length]
|
493 |
+
"""
|
494 |
+
r = torch.arange(length, dtype=torch.float32)
|
495 |
+
diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
|
496 |
+
return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
|
497 |
+
|
498 |
+
|
499 |
+
class FFN(nn.Module):
|
500 |
+
def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
|
501 |
+
super().__init__()
|
502 |
+
self.in_channels = in_channels
|
503 |
+
self.out_channels = out_channels
|
504 |
+
self.filter_channels = filter_channels
|
505 |
+
self.kernel_size = kernel_size
|
506 |
+
self.p_dropout = p_dropout
|
507 |
+
self.activation = activation
|
508 |
+
self.causal = causal
|
509 |
+
|
510 |
+
if causal:
|
511 |
+
self.padding = self._causal_padding
|
512 |
+
else:
|
513 |
+
self.padding = self._same_padding
|
514 |
+
|
515 |
+
self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
|
516 |
+
self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
|
517 |
+
# self.conv_1 = layers.Conv1d(in_channels, filter_channels, kernel_size, r = 4, lora_alpha = 16, lora_dropout = 0.05)
|
518 |
+
# self.conv_2 = layers.Conv1d(filter_channels, out_channels, kernel_size, r = 4, lora_alpha = 16, lora_dropout = 0.05)
|
519 |
+
self.drop = nn.Dropout(p_dropout)
|
520 |
+
|
521 |
+
def forward(self, x, x_mask):
|
522 |
+
x = self.conv_1(self.padding(x * x_mask))
|
523 |
+
if self.activation == "gelu":
|
524 |
+
x = x * torch.sigmoid(1.702 * x)
|
525 |
+
else:
|
526 |
+
x = torch.relu(x)
|
527 |
+
x = self.drop(x)
|
528 |
+
x = self.conv_2(self.padding(x * x_mask))
|
529 |
+
return x * x_mask
|
530 |
+
|
531 |
+
def _causal_padding(self, x):
|
532 |
+
if self.kernel_size == 1:
|
533 |
+
return x
|
534 |
+
pad_l = self.kernel_size - 1
|
535 |
+
pad_r = 0
|
536 |
+
padding = [[0, 0], [0, 0], [pad_l, pad_r]]
|
537 |
+
x = F.pad(x, commons.convert_pad_shape(padding))
|
538 |
+
return x
|
539 |
+
|
540 |
+
def _same_padding(self, x):
|
541 |
+
if self.kernel_size == 1:
|
542 |
+
return x
|
543 |
+
pad_l = (self.kernel_size - 1) // 2
|
544 |
+
pad_r = self.kernel_size // 2
|
545 |
+
padding = [[0, 0], [0, 0], [pad_l, pad_r]]
|
546 |
+
x = F.pad(x, commons.convert_pad_shape(padding))
|
547 |
+
return x
|
commons.py
ADDED
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import functional as F
|
6 |
+
|
7 |
+
|
8 |
+
def init_weights(m, mean=0.0, std=0.01):
|
9 |
+
classname = m.__class__.__name__
|
10 |
+
if classname.find("Conv") != -1:
|
11 |
+
m.weight.data.normal_(mean, std)
|
12 |
+
|
13 |
+
|
14 |
+
def get_padding(kernel_size, dilation=1):
|
15 |
+
return int((kernel_size*dilation - dilation)/2)
|
16 |
+
|
17 |
+
|
18 |
+
def convert_pad_shape(pad_shape):
|
19 |
+
l = pad_shape[::-1]
|
20 |
+
pad_shape = [item for sublist in l for item in sublist]
|
21 |
+
return pad_shape
|
22 |
+
|
23 |
+
|
24 |
+
def intersperse(lst, item):
|
25 |
+
result = [item] * (len(lst) * 2 + 1)
|
26 |
+
result[1::2] = lst
|
27 |
+
return result
|
28 |
+
|
29 |
+
|
30 |
+
def kl_divergence(m_p, logs_p, m_q, logs_q):
|
31 |
+
"""KL(P||Q)"""
|
32 |
+
kl = (logs_q - logs_p) - 0.5
|
33 |
+
kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
|
34 |
+
return kl
|
35 |
+
|
36 |
+
|
37 |
+
def rand_gumbel(shape):
|
38 |
+
"""Sample from the Gumbel distribution, protect from overflows."""
|
39 |
+
uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
|
40 |
+
return -torch.log(-torch.log(uniform_samples))
|
41 |
+
|
42 |
+
|
43 |
+
def rand_gumbel_like(x):
|
44 |
+
g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
|
45 |
+
return g
|
46 |
+
|
47 |
+
|
48 |
+
def slice_segments(x, ids_str, segment_size=4):
|
49 |
+
ret = torch.zeros_like(x[:, :, :segment_size])
|
50 |
+
for i in range(x.size(0)):
|
51 |
+
idx_str = ids_str[i]
|
52 |
+
idx_end = idx_str + segment_size
|
53 |
+
try:
|
54 |
+
ret[i] = x[i, :, idx_str:idx_end]
|
55 |
+
except RuntimeError:
|
56 |
+
print("?")
|
57 |
+
return ret
|
58 |
+
|
59 |
+
|
60 |
+
def rand_slice_segments(x, x_lengths=None, segment_size=4):
|
61 |
+
b, d, t = x.size()
|
62 |
+
if x_lengths is None:
|
63 |
+
x_lengths = t
|
64 |
+
ids_str_max = x_lengths - segment_size + 1
|
65 |
+
ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
|
66 |
+
ret = slice_segments(x, ids_str, segment_size)
|
67 |
+
return ret, ids_str
|
68 |
+
|
69 |
+
|
70 |
+
def get_timing_signal_1d(
|
71 |
+
length, channels, min_timescale=1.0, max_timescale=1.0e4):
|
72 |
+
position = torch.arange(length, dtype=torch.float)
|
73 |
+
num_timescales = channels // 2
|
74 |
+
log_timescale_increment = (
|
75 |
+
math.log(float(max_timescale) / float(min_timescale)) /
|
76 |
+
(num_timescales - 1))
|
77 |
+
inv_timescales = min_timescale * torch.exp(
|
78 |
+
torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
|
79 |
+
scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
|
80 |
+
signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
|
81 |
+
signal = F.pad(signal, [0, 0, 0, channels % 2])
|
82 |
+
signal = signal.view(1, channels, length)
|
83 |
+
return signal
|
84 |
+
|
85 |
+
|
86 |
+
def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
|
87 |
+
b, channels, length = x.size()
|
88 |
+
signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
|
89 |
+
return x + signal.to(dtype=x.dtype, device=x.device)
|
90 |
+
|
91 |
+
|
92 |
+
def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
|
93 |
+
b, channels, length = x.size()
|
94 |
+
signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
|
95 |
+
return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
|
96 |
+
|
97 |
+
|
98 |
+
def subsequent_mask(length):
|
99 |
+
mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
|
100 |
+
return mask
|
101 |
+
|
102 |
+
|
103 |
+
@torch.jit.script
|
104 |
+
def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
|
105 |
+
n_channels_int = n_channels[0]
|
106 |
+
in_act = input_a + input_b
|
107 |
+
t_act = torch.tanh(in_act[:, :n_channels_int, :])
|
108 |
+
s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
|
109 |
+
acts = t_act * s_act
|
110 |
+
return acts
|
111 |
+
|
112 |
+
|
113 |
+
def convert_pad_shape(pad_shape):
|
114 |
+
l = pad_shape[::-1]
|
115 |
+
pad_shape = [item for sublist in l for item in sublist]
|
116 |
+
return pad_shape
|
117 |
+
|
118 |
+
|
119 |
+
def shift_1d(x):
|
120 |
+
x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
|
121 |
+
return x
|
122 |
+
|
123 |
+
|
124 |
+
def sequence_mask(length, max_length=None):
|
125 |
+
if max_length is None:
|
126 |
+
max_length = length.max()
|
127 |
+
x = torch.arange(max_length, dtype=length.dtype, device=length.device)
|
128 |
+
return x.unsqueeze(0) < length.unsqueeze(1)
|
129 |
+
|
130 |
+
|
131 |
+
def generate_path(duration, mask):
|
132 |
+
"""
|
133 |
+
duration: [b, 1, t_x]
|
134 |
+
mask: [b, 1, t_y, t_x]
|
135 |
+
"""
|
136 |
+
device = duration.device
|
137 |
+
|
138 |
+
b, _, t_y, t_x = mask.shape
|
139 |
+
cum_duration = torch.cumsum(duration, -1)
|
140 |
+
|
141 |
+
cum_duration_flat = cum_duration.view(b * t_x)
|
142 |
+
path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
|
143 |
+
path = path.view(b, t_x, t_y)
|
144 |
+
path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
|
145 |
+
path = path.unsqueeze(1).transpose(2,3) * mask
|
146 |
+
return path
|
147 |
+
|
148 |
+
|
149 |
+
def clip_grad_value_(parameters, clip_value, norm_type=2):
|
150 |
+
if isinstance(parameters, torch.Tensor):
|
151 |
+
parameters = [parameters]
|
152 |
+
parameters = list(filter(lambda p: p.grad is not None, parameters))
|
153 |
+
norm_type = float(norm_type)
|
154 |
+
if clip_value is not None:
|
155 |
+
clip_value = float(clip_value)
|
156 |
+
|
157 |
+
total_norm = 0
|
158 |
+
for p in parameters:
|
159 |
+
param_norm = p.grad.data.norm(norm_type)
|
160 |
+
total_norm += param_norm.item() ** norm_type
|
161 |
+
if clip_value is not None:
|
162 |
+
p.grad.data.clamp_(min=-clip_value, max=clip_value)
|
163 |
+
total_norm = total_norm ** (1. / norm_type)
|
164 |
+
return total_norm
|
configs/finetune_speaker.json
ADDED
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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": 16,
|
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":"../CH_JA_EN_mix_voice/clipped_3_vits_trilingual_annotations.train.txt.cleaned",
|
21 |
+
"validation_files":"../CH_JA_EN_mix_voice/clipped_3_vits_trilingual_annotations.val.txt.cleaned",
|
22 |
+
"text_cleaners":["cjke_cleaners2"],
|
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": 999,
|
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 |
+
"symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "N", "Q", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "s", "t", "u", "v", "w", "x", "y", "z", "\u0251", "\u00e6", "\u0283", "\u0291", "\u00e7", "\u026f", "\u026a", "\u0254", "\u025b", "\u0279", "\u00f0", "\u0259", "\u026b", "\u0265", "\u0278", "\u028a", "\u027e", "\u0292", "\u03b8", "\u03b2", "\u014b", "\u0266", "\u207c", "\u02b0", "`", "^", "#", "*", "=", "\u02c8", "\u02cc", "\u2192", "\u2193", "\u2191", " "],
|
54 |
+
"speakers": {"特别周 Special Week (Umamusume Pretty Derby)": 0,
|
55 |
+
"无声铃鹿 Silence Suzuka (Umamusume Pretty Derby)": 1,
|
56 |
+
"东海帝王 Tokai Teio (Umamusume Pretty Derby)": 2,
|
57 |
+
"丸善斯基 Maruzensky (Umamusume Pretty Derby)": 3,
|
58 |
+
"富士奇迹 Fuji Kiseki (Umamusume Pretty Derby)": 4,
|
59 |
+
"小栗帽 Oguri Cap (Umamusume Pretty Derby)": 5,
|
60 |
+
"黄金船 Gold Ship (Umamusume Pretty Derby)": 6,
|
61 |
+
"伏特加 Vodka (Umamusume Pretty Derby)": 7,
|
62 |
+
"大和赤骥 Daiwa Scarlet (Umamusume Pretty Derby)": 8,
|
63 |
+
"大树快车 Taiki Shuttle (Umamusume Pretty Derby)": 9,
|
64 |
+
"草上飞 Grass Wonder (Umamusume Pretty Derby)": 10,
|
65 |
+
"菱亚马逊 Hishi Amazon (Umamusume Pretty Derby)": 11,
|
66 |
+
"目白麦昆 Mejiro Mcqueen (Umamusume Pretty Derby)": 12,
|
67 |
+
"神鹰 El Condor Pasa (Umamusume Pretty Derby)": 13,
|
68 |
+
"好歌剧 T.M. Opera O (Umamusume Pretty Derby)": 14,
|
69 |
+
"成田白仁 Narita Brian (Umamusume Pretty Derby)": 15,
|
70 |
+
"鲁道夫象征 Symboli Rudolf (Umamusume Pretty Derby)": 16,
|
71 |
+
"气槽 Air Groove (Umamusume Pretty Derby)": 17,
|
72 |
+
"爱丽数码 Agnes Digital (Umamusume Pretty Derby)": 18,
|
73 |
+
"青云天空 Seiun Sky (Umamusume Pretty Derby)": 19,
|
74 |
+
"玉藻十字 Tamamo Cross (Umamusume Pretty Derby)": 20,
|
75 |
+
"美妙姿势 Fine Motion (Umamusume Pretty Derby)": 21,
|
76 |
+
"琵琶晨光 Biwa Hayahide (Umamusume Pretty Derby)": 22,
|
77 |
+
"重炮 Mayano Topgun (Umamusume Pretty Derby)": 23,
|
78 |
+
"曼城茶座 Manhattan Cafe (Umamusume Pretty Derby)": 24,
|
79 |
+
"美普波旁 Mihono Bourbon (Umamusume Pretty Derby)": 25,
|
80 |
+
"目白雷恩 Mejiro Ryan (Umamusume Pretty Derby)": 26,
|
81 |
+
"雪之美人 Yukino Bijin (Umamusume Pretty Derby)": 28,
|
82 |
+
"米浴 Rice Shower (Umamusume Pretty Derby)": 29,
|
83 |
+
"艾尼斯风神 Ines Fujin (Umamusume Pretty Derby)": 30,
|
84 |
+
"爱丽速子 Agnes Tachyon (Umamusume Pretty Derby)": 31,
|
85 |
+
"爱慕织姬 Admire Vega (Umamusume Pretty Derby)": 32,
|
86 |
+
"稻荷一 Inari One (Umamusume Pretty Derby)": 33,
|
87 |
+
"胜利奖券 Winning Ticket (Umamusume Pretty Derby)": 34,
|
88 |
+
"空中神宫 Air Shakur (Umamusume Pretty Derby)": 35,
|
89 |
+
"荣进闪耀 Eishin Flash (Umamusume Pretty Derby)": 36,
|
90 |
+
"真机伶 Curren Chan (Umamusume Pretty Derby)": 37,
|
91 |
+
"川上公主 Kawakami Princess (Umamusume Pretty Derby)": 38,
|
92 |
+
"黄金城市 Gold City (Umamusume Pretty Derby)": 39,
|
93 |
+
"樱花进王 Sakura Bakushin O (Umamusume Pretty Derby)": 40,
|
94 |
+
"采珠 Seeking the Pearl (Umamusume Pretty Derby)": 41,
|
95 |
+
"新光风 Shinko Windy (Umamusume Pretty Derby)": 42,
|
96 |
+
"东商变革 Sweep Tosho (Umamusume Pretty Derby)": 43,
|
97 |
+
"超级小溪 Super Creek (Umamusume Pretty Derby)": 44,
|
98 |
+
"醒目飞鹰 Smart Falcon (Umamusume Pretty Derby)": 45,
|
99 |
+
"荒漠英雄 Zenno Rob Roy (Umamusume Pretty Derby)": 46,
|
100 |
+
"东瀛佐敦 Tosen Jordan (Umamusume Pretty Derby)": 47,
|
101 |
+
"中山庆典 Nakayama Festa (Umamusume Pretty Derby)": 48,
|
102 |
+
"成田大进 Narita Taishin (Umamusume Pretty Derby)": 49,
|
103 |
+
"西野花 Nishino Flower (Umamusume Pretty Derby)": 50,
|
104 |
+
"春乌拉拉 Haru Urara (Umamusume Pretty Derby)": 51,
|
105 |
+
"青竹回忆 Bamboo Memory (Umamusume Pretty Derby)": 52,
|
106 |
+
"待兼福来 Matikane Fukukitaru (Umamusume Pretty Derby)": 55,
|
107 |
+
"名将怒涛 Meisho Doto (Umamusume Pretty Derby)": 57,
|
108 |
+
"目白多伯 Mejiro Dober (Umamusume Pretty Derby)": 58,
|
109 |
+
"优秀素质 Nice Nature (Umamusume Pretty Derby)": 59,
|
110 |
+
"帝王光环 King Halo (Umamusume Pretty Derby)": 60,
|
111 |
+
"待兼诗歌剧 Matikane Tannhauser (Umamusume Pretty Derby)": 61,
|
112 |
+
"生野狄杜斯 Ikuno Dictus (Umamusume Pretty Derby)": 62,
|
113 |
+
"目白善信 Mejiro Palmer (Umamusume Pretty Derby)": 63,
|
114 |
+
"大拓太阳神 Daitaku Helios (Umamusume Pretty Derby)": 64,
|
115 |
+
"双涡轮 Twin Turbo (Umamusume Pretty Derby)": 65,
|
116 |
+
"里见光钻 Satono Diamond (Umamusume Pretty Derby)": 66,
|
117 |
+
"北部玄驹 Kitasan Black (Umamusume Pretty Derby)": 67,
|
118 |
+
"樱花千代王 Sakura Chiyono O (Umamusume Pretty Derby)": 68,
|
119 |
+
"天狼星象征 Sirius Symboli (Umamusume Pretty Derby)": 69,
|
120 |
+
"目白阿尔丹 Mejiro Ardan (Umamusume Pretty Derby)": 70,
|
121 |
+
"八重无敌 Yaeno Muteki (Umamusume Pretty Derby)": 71,
|
122 |
+
"鹤丸刚志 Tsurumaru Tsuyoshi (Umamusume Pretty Derby)": 72,
|
123 |
+
"目白光明 Mejiro Bright (Umamusume Pretty Derby)": 73,
|
124 |
+
"樱花桂冠 Sakura Laurel (Umamusume Pretty Derby)": 74,
|
125 |
+
"成田路 Narita Top Road (Umamusume Pretty Derby)": 75,
|
126 |
+
"也文摄辉 Yamanin Zephyr (Umamusume Pretty Derby)": 76,
|
127 |
+
"真弓快车 Aston Machan (Umamusume Pretty Derby)": 80,
|
128 |
+
"骏川手纲 Hayakawa Tazuna (Umamusume Pretty Derby)": 81,
|
129 |
+
"小林历奇 Kopano Rickey (Umamusume Pretty Derby)": 83,
|
130 |
+
"奇锐骏 Wonder Acute (Umamusume Pretty Derby)": 85,
|
131 |
+
"秋川理事长 President Akikawa (Umamusume Pretty Derby)": 86,
|
132 |
+
"綾地 寧々 Ayachi Nene (Sanoba Witch)": 87,
|
133 |
+
"因幡 めぐる Inaba Meguru (Sanoba Witch)": 88,
|
134 |
+
"椎葉 紬 Shiiba Tsumugi (Sanoba Witch)": 89,
|
135 |
+
"仮屋 和奏 Kariya Wakama (Sanoba Witch)": 90,
|
136 |
+
"戸隠 憧子 Togakushi Touko (Sanoba Witch)": 91,
|
137 |
+
"九条裟罗 Kujou Sara (Genshin Impact)": 92,
|
138 |
+
"芭芭拉 Barbara (Genshin Impact)": 93,
|
139 |
+
"派蒙 Paimon (Genshin Impact)": 94,
|
140 |
+
"荒泷一斗 Arataki Itto (Genshin Impact)": 96,
|
141 |
+
"早柚 Sayu (Genshin Impact)": 97,
|
142 |
+
"香菱 Xiangling (Genshin Impact)": 98,
|
143 |
+
"神里绫华 Kamisato Ayaka (Genshin Impact)": 99,
|
144 |
+
"重云 Chongyun (Genshin Impact)": 100,
|
145 |
+
"流浪者 Wanderer (Genshin Impact)": 102,
|
146 |
+
"优菈 Eula (Genshin Impact)": 103,
|
147 |
+
"凝光 Ningguang (Genshin Impact)": 105,
|
148 |
+
"钟离 Zhongli (Genshin Impact)": 106,
|
149 |
+
"雷电将军 Raiden Shogun (Genshin Impact)": 107,
|
150 |
+
"枫原万叶 Kaedehara Kazuha (Genshin Impact)": 108,
|
151 |
+
"赛诺 Cyno (Genshin Impact)": 109,
|
152 |
+
"诺艾尔 Noelle (Genshin Impact)": 112,
|
153 |
+
"八重神子 Yae Miko (Genshin Impact)": 113,
|
154 |
+
"凯亚 Kaeya (Genshin Impact)": 114,
|
155 |
+
"魈 Xiao (Genshin Impact)": 115,
|
156 |
+
"托马 Thoma (Genshin Impact)": 116,
|
157 |
+
"可莉 Klee (Genshin Impact)": 117,
|
158 |
+
"迪卢克 Diluc (Genshin Impact)": 120,
|
159 |
+
"夜兰 Yelan (Genshin Impact)": 121,
|
160 |
+
"鹿野院平藏 Shikanoin Heizou (Genshin Impact)": 123,
|
161 |
+
"辛焱 Xinyan (Genshin Impact)": 124,
|
162 |
+
"丽莎 Lisa (Genshin Impact)": 125,
|
163 |
+
"云堇 Yun Jin (Genshin Impact)": 126,
|
164 |
+
"坎蒂丝 Candace (Genshin Impact)": 127,
|
165 |
+
"罗莎莉亚 Rosaria (Genshin Impact)": 128,
|
166 |
+
"北斗 Beidou (Genshin Impact)": 129,
|
167 |
+
"珊瑚宫心海 Sangonomiya Kokomi (Genshin Impact)": 132,
|
168 |
+
"烟绯 Yanfei (Genshin Impact)": 133,
|
169 |
+
"久岐忍 Kuki Shinobu (Genshin Impact)": 136,
|
170 |
+
"宵宫 Yoimiya (Genshin Impact)": 139,
|
171 |
+
"安柏 Amber (Genshin Impact)": 143,
|
172 |
+
"迪奥娜 Diona (Genshin Impact)": 144,
|
173 |
+
"班尼特 Bennett (Genshin Impact)": 146,
|
174 |
+
"雷泽 Razor (Genshin Impact)": 147,
|
175 |
+
"阿贝多 Albedo (Genshin Impact)": 151,
|
176 |
+
"温迪 Venti (Genshin Impact)": 152,
|
177 |
+
"空 Player Male (Genshin Impact)": 153,
|
178 |
+
"神里绫人 Kamisato Ayato (Genshin Impact)": 154,
|
179 |
+
"琴 Jean (Genshin Impact)": 155,
|
180 |
+
"艾尔海森 Alhaitham (Genshin Impact)": 156,
|
181 |
+
"莫娜 Mona (Genshin Impact)": 157,
|
182 |
+
"妮露 Nilou (Genshin Impact)": 159,
|
183 |
+
"胡桃 Hu Tao (Genshin Impact)": 160,
|
184 |
+
"甘雨 Ganyu (Genshin Impact)": 161,
|
185 |
+
"纳西妲 Nahida (Genshin Impact)": 162,
|
186 |
+
"刻晴 Keqing (Genshin Impact)": 165,
|
187 |
+
"荧 Player Female (Genshin Impact)": 169,
|
188 |
+
"埃洛伊 Aloy (Genshin Impact)": 179,
|
189 |
+
"柯莱 Collei (Genshin Impact)": 182,
|
190 |
+
"多莉 Dori (Genshin Impact)": 184,
|
191 |
+
"提纳里 Tighnari (Genshin Impact)": 186,
|
192 |
+
"砂糖 Sucrose (Genshin Impact)": 188,
|
193 |
+
"行秋 Xingqiu (Genshin Impact)": 190,
|
194 |
+
"奥兹 Oz (Genshin Impact)": 193,
|
195 |
+
"五郎 Gorou (Genshin Impact)": 198,
|
196 |
+
"达达利亚 Tartalia (Genshin Impact)": 202,
|
197 |
+
"七七 Qiqi (Genshin Impact)": 207,
|
198 |
+
"申鹤 Shenhe (Genshin Impact)": 217,
|
199 |
+
"莱依拉 Layla (Genshin Impact)": 228,
|
200 |
+
"菲谢尔 Fishl (Genshin Impact)": 230
|
201 |
+
}
|
202 |
+
}
|
configs/modified_finetune_speaker.json
ADDED
@@ -0,0 +1,321 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train": {
|
3 |
+
"log_interval": 10,
|
4 |
+
"eval_interval": 100,
|
5 |
+
"seed": 1234,
|
6 |
+
"epochs": 10000,
|
7 |
+
"learning_rate": 0.0002,
|
8 |
+
"betas": [
|
9 |
+
0.8,
|
10 |
+
0.99
|
11 |
+
],
|
12 |
+
"eps": 1e-09,
|
13 |
+
"batch_size": 16,
|
14 |
+
"fp16_run": true,
|
15 |
+
"lr_decay": 0.999875,
|
16 |
+
"segment_size": 8192,
|
17 |
+
"init_lr_ratio": 1,
|
18 |
+
"warmup_epochs": 0,
|
19 |
+
"c_mel": 45,
|
20 |
+
"c_kl": 1.0
|
21 |
+
},
|
22 |
+
"data": {
|
23 |
+
"training_files": "final_annotation_train.txt",
|
24 |
+
"validation_files": "final_annotation_val.txt",
|
25 |
+
"text_cleaners": [
|
26 |
+
"cjke_cleaners2"
|
27 |
+
],
|
28 |
+
"max_wav_value": 32768.0,
|
29 |
+
"sampling_rate": 22050,
|
30 |
+
"filter_length": 1024,
|
31 |
+
"hop_length": 256,
|
32 |
+
"win_length": 1024,
|
33 |
+
"n_mel_channels": 80,
|
34 |
+
"mel_fmin": 0.0,
|
35 |
+
"mel_fmax": null,
|
36 |
+
"add_blank": true,
|
37 |
+
"n_speakers": 161,
|
38 |
+
"cleaned_text": true
|
39 |
+
},
|
40 |
+
"model": {
|
41 |
+
"inter_channels": 192,
|
42 |
+
"hidden_channels": 192,
|
43 |
+
"filter_channels": 768,
|
44 |
+
"n_heads": 2,
|
45 |
+
"n_layers": 6,
|
46 |
+
"kernel_size": 3,
|
47 |
+
"p_dropout": 0.1,
|
48 |
+
"resblock": "1",
|
49 |
+
"resblock_kernel_sizes": [
|
50 |
+
3,
|
51 |
+
7,
|
52 |
+
11
|
53 |
+
],
|
54 |
+
"resblock_dilation_sizes": [
|
55 |
+
[
|
56 |
+
1,
|
57 |
+
3,
|
58 |
+
5
|
59 |
+
],
|
60 |
+
[
|
61 |
+
1,
|
62 |
+
3,
|
63 |
+
5
|
64 |
+
],
|
65 |
+
[
|
66 |
+
1,
|
67 |
+
3,
|
68 |
+
5
|
69 |
+
]
|
70 |
+
],
|
71 |
+
"upsample_rates": [
|
72 |
+
8,
|
73 |
+
8,
|
74 |
+
2,
|
75 |
+
2
|
76 |
+
],
|
77 |
+
"upsample_initial_channel": 512,
|
78 |
+
"upsample_kernel_sizes": [
|
79 |
+
16,
|
80 |
+
16,
|
81 |
+
4,
|
82 |
+
4
|
83 |
+
],
|
84 |
+
"n_layers_q": 3,
|
85 |
+
"use_spectral_norm": false,
|
86 |
+
"gin_channels": 256
|
87 |
+
},
|
88 |
+
"symbols": [
|
89 |
+
"_",
|
90 |
+
",",
|
91 |
+
".",
|
92 |
+
"!",
|
93 |
+
"?",
|
94 |
+
"-",
|
95 |
+
"~",
|
96 |
+
"\u2026",
|
97 |
+
"N",
|
98 |
+
"Q",
|
99 |
+
"a",
|
100 |
+
"b",
|
101 |
+
"d",
|
102 |
+
"e",
|
103 |
+
"f",
|
104 |
+
"g",
|
105 |
+
"h",
|
106 |
+
"i",
|
107 |
+
"j",
|
108 |
+
"k",
|
109 |
+
"l",
|
110 |
+
"m",
|
111 |
+
"n",
|
112 |
+
"o",
|
113 |
+
"p",
|
114 |
+
"s",
|
115 |
+
"t",
|
116 |
+
"u",
|
117 |
+
"v",
|
118 |
+
"w",
|
119 |
+
"x",
|
120 |
+
"y",
|
121 |
+
"z",
|
122 |
+
"\u0251",
|
123 |
+
"\u00e6",
|
124 |
+
"\u0283",
|
125 |
+
"\u0291",
|
126 |
+
"\u00e7",
|
127 |
+
"\u026f",
|
128 |
+
"\u026a",
|
129 |
+
"\u0254",
|
130 |
+
"\u025b",
|
131 |
+
"\u0279",
|
132 |
+
"\u00f0",
|
133 |
+
"\u0259",
|
134 |
+
"\u026b",
|
135 |
+
"\u0265",
|
136 |
+
"\u0278",
|
137 |
+
"\u028a",
|
138 |
+
"\u027e",
|
139 |
+
"\u0292",
|
140 |
+
"\u03b8",
|
141 |
+
"\u03b2",
|
142 |
+
"\u014b",
|
143 |
+
"\u0266",
|
144 |
+
"\u207c",
|
145 |
+
"\u02b0",
|
146 |
+
"`",
|
147 |
+
"^",
|
148 |
+
"#",
|
149 |
+
"*",
|
150 |
+
"=",
|
151 |
+
"\u02c8",
|
152 |
+
"\u02cc",
|
153 |
+
"\u2192",
|
154 |
+
"\u2193",
|
155 |
+
"\u2191",
|
156 |
+
" "
|
157 |
+
],
|
158 |
+
"speakers": {
|
159 |
+
"\u83c8\u6bd4\u8389\u65af\u5854\uff08Overload\uff09": 0,
|
160 |
+
"\u94c3\u5948\uff08\u590f\u65e5\uff09": 1,
|
161 |
+
"\u674f\u5948": 2,
|
162 |
+
"\u96ea\u83f2": 3,
|
163 |
+
"\u53ef\u53ef\u841d": 4,
|
164 |
+
"\u83c8\u6bd4\u8389\u65af\u5854": 5,
|
165 |
+
"\u7eeb\u97f3\uff08\u5723\u8bde\u8282\uff09": 6,
|
166 |
+
"\u78a7\uff08\u63d2\u73ed\u751f\uff09": 7,
|
167 |
+
"\u5609\u591c": 8,
|
168 |
+
"\u955c\u534e\uff08\u4e07\u5723\u8282\uff09": 9,
|
169 |
+
"\u73e0\u5e0c": 10,
|
170 |
+
"\u671b\uff08\u590f\u65e5\uff09": 11,
|
171 |
+
"\u94c3\u8393": 12,
|
172 |
+
"\u771f\u7434\uff08\u590f\u65e5\uff09": 13,
|
173 |
+
"\u9759\u6d41\uff08\u60c5\u4eba\u8282\uff09": 14,
|
174 |
+
"\u54b2\u604b": 15,
|
175 |
+
"\u8389\u739b": 16,
|
176 |
+
"\u9999\u6f84\uff08\u590f\u65e5\uff09": 17,
|
177 |
+
"\u5343\u6b4c": 18,
|
178 |
+
"\u5fcd": 19,
|
179 |
+
"\u4f9d\u91cc": 20,
|
180 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u65b0\u5e74\uff09": 21,
|
181 |
+
"\u80e1\u6843": 22,
|
182 |
+
"\u667a\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 23,
|
183 |
+
"\u4f18\u8863": 24,
|
184 |
+
"\u601c\uff08\u4e07\u5723\u8282\uff09": 25,
|
185 |
+
"\u681e": 26,
|
186 |
+
"\u9999\u7ec7\uff08\u590f\u65e5\uff09": 27,
|
187 |
+
"\u7948\u68a8\uff08\u65f6\u95f4\u65c5\u884c\uff09": 28,
|
188 |
+
"\u54b2\u604b\uff08\u590f\u65e5\uff09": 29,
|
189 |
+
"\u94c3\u5948": 30,
|
190 |
+
"\u771f\u6b65\uff08\u590f\u65e5\uff09": 31,
|
191 |
+
"\u4e9a\u91cc\u838e": 32,
|
192 |
+
"\u955c\u534e": 33,
|
193 |
+
"\u672a\u594f\u5e0c": 34,
|
194 |
+
"\u4f0a\u8389\u4e9a\uff08\u5723\u8bde\u8282\uff09": 35,
|
195 |
+
"\u7f8e\u91cc": 36,
|
196 |
+
"\u4f3c\u4f3c\u82b1": 37,
|
197 |
+
"\u514b\u8389\u4e1d\u63d0\u5a1c\uff08\u5723\u8bde\u8282\uff09": 38,
|
198 |
+
"\u7f8e\u51ac\uff08\u590f\u65e5\uff09": 39,
|
199 |
+
"\u8389\u739b\uff08\u7070\u59d1\u5a18\uff09": 40,
|
200 |
+
"\u94c3\u8393\uff08\u590f\u65e5\uff09": 41,
|
201 |
+
"\u53e4\u857e\u96c5": 42,
|
202 |
+
"\u7f8e\u7f8e\uff08\u4e07\u5723\u8282\uff09": 43,
|
203 |
+
"\u5343\u6b4c\uff08\u5723\u8bde\u8282\uff09": 44,
|
204 |
+
"\u78a7": 45,
|
205 |
+
"\u96ea": 46,
|
206 |
+
"\u60e0\u7406\u5b50\uff08\u60c5\u4eba\u8282\uff09": 47,
|
207 |
+
"\u4f0a\u7eea": 48,
|
208 |
+
"\u62c9\u59c6": 49,
|
209 |
+
"\u4e03\u4e03\u9999\uff08\u590f\u65e5\uff09": 50,
|
210 |
+
"\u94c3\u8393\uff08\u65b0\u5e74\uff09": 51,
|
211 |
+
"\u94c3\uff08\u6e38\u9a91\u5175\uff09": 52,
|
212 |
+
"\u73af\u5948": 53,
|
213 |
+
"\u60e0\u7406\u5b50": 54,
|
214 |
+
"\u53ef\u53ef\u841d\uff08\u793c\u670d\uff09": 55,
|
215 |
+
"\u4f18\u8863\uff08\u65b0\u5e74\uff09": 56,
|
216 |
+
"\u671b\uff08\u5723\u8bde\u8282\uff09": 57,
|
217 |
+
"\u53ef\u53ef\u841d\uff08\u516c\u4e3b\uff09": 58,
|
218 |
+
"\u7eaf\uff08\u590f\u65e5\uff09": 59,
|
219 |
+
"\u73af\u5948\uff08\u632f\u8896\uff09": 60,
|
220 |
+
"\u54b2\u604b\uff08\u5723\u8bde\u8282\uff09": 61,
|
221 |
+
"\u9732\u5a1c": 62,
|
222 |
+
"\u9999\u6f84\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 63,
|
223 |
+
"\u59ec\u5854": 64,
|
224 |
+
"\u77db\u4f9d\u672a\uff08\u65b0\u5e74\uff09": 65,
|
225 |
+
"\u681e\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 66,
|
226 |
+
"\u771f\u6b65\uff08\u7070\u59d1\u5a18\uff09": 67,
|
227 |
+
"\u601c\uff08\u65b0\u5e74\uff09": 68,
|
228 |
+
"\u5343\u6b4c\uff08\u590f\u65e5\uff09": 69,
|
229 |
+
"\u7948\u68a8": 70,
|
230 |
+
"\u7531\u52a0\u8389": 71,
|
231 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u590f\u65e5\uff09": 72,
|
232 |
+
"\u6b65\u7f8e\uff08\u4ed9\u5883\uff09": 73,
|
233 |
+
"\u53ef\u53ef\u841d\uff08\u590f\u65e5\uff09": 74,
|
234 |
+
"\u7eeb\u97f3": 75,
|
235 |
+
"\u7531\u52a0\u8389\uff08\u5723\u8bde\u8282\uff09": 76,
|
236 |
+
"\u771f\u9633": 77,
|
237 |
+
"\u96f7\u59c6": 78,
|
238 |
+
"\u831c\u91cc": 79,
|
239 |
+
"\u94c3": 80,
|
240 |
+
"\u51db\uff08\u5076\u50cf\u5927\u5e08\uff09": 81,
|
241 |
+
"\u51ef\u9732": 82,
|
242 |
+
"\u514b\u7f57\u4f9d\uff08\u5723\u5b66\u796d\uff09": 83,
|
243 |
+
"\u79cb\u4e43": 84,
|
244 |
+
"\u514b\u8389\u4e1d\u63d0\u5a1c": 85,
|
245 |
+
"\u4f69\u53ef\u8389\u59c6": 86,
|
246 |
+
"\u9732": 87,
|
247 |
+
"\u83ab\u59ae\u5361": 88,
|
248 |
+
"\u7483\u4e43": 89,
|
249 |
+
"\u9759\u6d41\uff08\u590f\u65e5\uff09": 90,
|
250 |
+
"\u7eba\u5e0c\uff08\u4e07\u5723\u8282\uff09": 91,
|
251 |
+
"\u9999\u6f84": 92,
|
252 |
+
"\u7f8e\u54b2": 93,
|
253 |
+
"\u7f8e\u91cc\uff08\u590f\u65e5\uff09": 94,
|
254 |
+
"\u667a": 95,
|
255 |
+
"\u521d\u97f3": 96,
|
256 |
+
"\u7483\u4e43\uff08\u4ed9\u5883\uff09": 97,
|
257 |
+
"\u536f\u6708\uff08\u5076\u50cf\u5927\u5e08\uff09": 98,
|
258 |
+
"\u7a7a\u82b1": 99,
|
259 |
+
"\u73e0\u5e0c\uff08\u590f\u65e5\uff09": 100,
|
260 |
+
"\u7eaf": 101,
|
261 |
+
"\u7f8e\u7f8e": 102,
|
262 |
+
"\u5fcd\uff08\u4e07\u5723\u8282\uff09": 103,
|
263 |
+
"\u59ae\u4fac\uff08\u5927\u6c5f\u6237\uff09": 104,
|
264 |
+
"\u5343\u7231\u7460": 105,
|
265 |
+
"\u53ef\u53ef\u841d\uff08\u65b0\u5e74\uff09": 106,
|
266 |
+
"\u601c\uff08\u516c\u4e3b\uff09": 107,
|
267 |
+
"\u672a\u594f\u5e0c\u3001\u7f8e\u7f8e\u3001\u955c\u534e": 108,
|
268 |
+
"\u4f0a\u8389\u4e9a": 109,
|
269 |
+
"\u671b": 110,
|
270 |
+
"\u4f18\u8863\uff08\u516c\u4e3b\uff09": 111,
|
271 |
+
"\u51ef\u9732\uff08\u516c\u4e3b\uff09": 112,
|
272 |
+
"\u514b\u7f57\u4f9d": 113,
|
273 |
+
"\u79cb\u4e43\uff08\u5723\u8bde\u8282\uff09": 114,
|
274 |
+
"\u4f18\u8863\uff08\u793c\u670d\uff09": 115,
|
275 |
+
"\u674f\u5948\uff08\u590f\u65e5\uff09": 116,
|
276 |
+
"\u9759\u6d41": 117,
|
277 |
+
"\u672a\u594f\u5e0c\uff08\u4e07\u5723\u8282\uff09": 118,
|
278 |
+
"\u831c\u91cc\uff08\u5929\u4f7f\uff09": 119,
|
279 |
+
"\u5bab\u5b50": 120,
|
280 |
+
"\u65e5\u548c\u8389\uff08\u65b0\u5e74\uff09": 121,
|
281 |
+
"\u4f18\u59ae": 122,
|
282 |
+
"\u5b89": 123,
|
283 |
+
"\u6d41\u590f\uff08\u590f\u65e5\uff09": 124,
|
284 |
+
"\u7f8e\u51ac\uff08\u5de5\u4f5c\u670d\uff09": 125,
|
285 |
+
"\u7f8e\u51ac": 126,
|
286 |
+
"\u521d\u97f3\uff08\u590f\u65e5\uff09": 127,
|
287 |
+
"\u77db\u4f9d\u672a": 128,
|
288 |
+
"\u51ef\u9732\uff08\u590f\u65e5\uff09": 129,
|
289 |
+
"\u83ab\u59ae\u5361\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 130,
|
290 |
+
"\u9999\u7ec7": 131,
|
291 |
+
"\u5bab\u5b50\uff08\u4e07\u5723\u8282\uff09": 132,
|
292 |
+
"\u7f8e\u54b2\uff08\u4e07\u5723\u8282\uff09": 133,
|
293 |
+
"\u59ae\u4fac": 134,
|
294 |
+
"\u601c": 135,
|
295 |
+
"\u771f\u9633\uff08\u6e38\u9a91\u5175\uff09": 136,
|
296 |
+
"\u78a7\uff08\u5de5\u4f5c\u670d\uff09": 137,
|
297 |
+
"\u65e5\u548c\u8389": 138,
|
298 |
+
"\u4e03\u4e03\u9999": 139,
|
299 |
+
"\u771f\u7434": 140,
|
300 |
+
"\u6b65\u7f8e": 141,
|
301 |
+
"\u6df1\u6708": 142,
|
302 |
+
"\u65e5\u548c\u8389\uff08\u516c\u4e3b\uff09": 143,
|
303 |
+
"\u4f0a\u7eea\uff08\u590f\u65e5\uff09": 144,
|
304 |
+
"\u8309\u8389": 145,
|
305 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u516c\u4e3b\uff09": 146,
|
306 |
+
"\u7a7a\u82b1\uff08\u5927\u6c5f\u6237\uff09": 147,
|
307 |
+
"\u60e0\u7406\u5b50\uff08\u590f\u65e5\uff09": 148,
|
308 |
+
"\u51ef\u9732\uff08\u65b0\u5e74\uff09": 149,
|
309 |
+
"\u7231\u871c\u8389\u96c5": 150,
|
310 |
+
"\u4f9d\u91cc\uff08\u5929\u4f7f\uff09": 151,
|
311 |
+
"\u8309\u8389\uff08\u4e07\u5723\u8282\uff09": 152,
|
312 |
+
"\u771f\u7434\uff08\u7070\u59d1\u5a18\uff09": 153,
|
313 |
+
"\u7eba\u5e0c": 154,
|
314 |
+
"\u771f\u6b65": 155,
|
315 |
+
"\u6d41\u590f": 156,
|
316 |
+
"\u4f3c\u4f3c\u82b1\uff08\u65b0\u5e74\uff09": 157,
|
317 |
+
"\u672a\u592e\uff08\u5076\u50cf\u5927\u5e08\uff09": 158,
|
318 |
+
"\u80e1\u6843\uff08\u5723\u8bde\u8282\uff09": 159,
|
319 |
+
"\u5343\u7231\u7460\uff08\u5723\u5b66\u796d\uff09": 160
|
320 |
+
}
|
321 |
+
}
|
configs/uma_trilingual.json
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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": 16,
|
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":"../CH_JA_EN_mix_voice/clipped_3_vits_trilingual_annotations.train.txt.cleaned",
|
21 |
+
"validation_files":"../CH_JA_EN_mix_voice/clipped_3_vits_trilingual_annotations.val.txt.cleaned",
|
22 |
+
"text_cleaners":["cjke_cleaners2"],
|
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": 999,
|
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 |
+
"symbols": ["_", ",", ".", "!", "?", "-", "~", "\u2026", "N", "Q", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "s", "t", "u", "v", "w", "x", "y", "z", "\u0251", "\u00e6", "\u0283", "\u0291", "\u00e7", "\u026f", "\u026a", "\u0254", "\u025b", "\u0279", "\u00f0", "\u0259", "\u026b", "\u0265", "\u0278", "\u028a", "\u027e", "\u0292", "\u03b8", "\u03b2", "\u014b", "\u0266", "\u207c", "\u02b0", "`", "^", "#", "*", "=", "\u02c8", "\u02cc", "\u2192", "\u2193", "\u2191", " "]
|
54 |
+
}
|
mel_processing.py
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
import torch
|
5 |
+
from torch import nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
import torch.utils.data
|
8 |
+
import numpy as np
|
9 |
+
import librosa
|
10 |
+
import librosa.util as librosa_util
|
11 |
+
from librosa.util import normalize, pad_center, tiny
|
12 |
+
from scipy.signal import get_window
|
13 |
+
from scipy.io.wavfile import read
|
14 |
+
from librosa.filters import mel as librosa_mel_fn
|
15 |
+
|
16 |
+
MAX_WAV_VALUE = 32768.0
|
17 |
+
|
18 |
+
|
19 |
+
def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
|
20 |
+
"""
|
21 |
+
PARAMS
|
22 |
+
------
|
23 |
+
C: compression factor
|
24 |
+
"""
|
25 |
+
return torch.log(torch.clamp(x, min=clip_val) * C)
|
26 |
+
|
27 |
+
|
28 |
+
def dynamic_range_decompression_torch(x, C=1):
|
29 |
+
"""
|
30 |
+
PARAMS
|
31 |
+
------
|
32 |
+
C: compression factor used to compress
|
33 |
+
"""
|
34 |
+
return torch.exp(x) / C
|
35 |
+
|
36 |
+
|
37 |
+
def spectral_normalize_torch(magnitudes):
|
38 |
+
output = dynamic_range_compression_torch(magnitudes)
|
39 |
+
return output
|
40 |
+
|
41 |
+
|
42 |
+
def spectral_de_normalize_torch(magnitudes):
|
43 |
+
output = dynamic_range_decompression_torch(magnitudes)
|
44 |
+
return output
|
45 |
+
|
46 |
+
|
47 |
+
mel_basis = {}
|
48 |
+
hann_window = {}
|
49 |
+
|
50 |
+
|
51 |
+
def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
|
52 |
+
if torch.min(y) < -1.:
|
53 |
+
print('min value is ', torch.min(y))
|
54 |
+
if torch.max(y) > 1.:
|
55 |
+
print('max value is ', torch.max(y))
|
56 |
+
|
57 |
+
global hann_window
|
58 |
+
dtype_device = str(y.dtype) + '_' + str(y.device)
|
59 |
+
wnsize_dtype_device = str(win_size) + '_' + dtype_device
|
60 |
+
if wnsize_dtype_device not in hann_window:
|
61 |
+
hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
|
62 |
+
|
63 |
+
y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
|
64 |
+
y = y.squeeze(1)
|
65 |
+
|
66 |
+
spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
|
67 |
+
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
|
68 |
+
|
69 |
+
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
|
70 |
+
return spec
|
71 |
+
|
72 |
+
|
73 |
+
def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
|
74 |
+
global mel_basis
|
75 |
+
dtype_device = str(spec.dtype) + '_' + str(spec.device)
|
76 |
+
fmax_dtype_device = str(fmax) + '_' + dtype_device
|
77 |
+
if fmax_dtype_device not in mel_basis:
|
78 |
+
mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
|
79 |
+
mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
|
80 |
+
spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
|
81 |
+
spec = spectral_normalize_torch(spec)
|
82 |
+
return spec
|
83 |
+
|
84 |
+
|
85 |
+
def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
|
86 |
+
if torch.min(y) < -1.:
|
87 |
+
print('min value is ', torch.min(y))
|
88 |
+
if torch.max(y) > 1.:
|
89 |
+
print('max value is ', torch.max(y))
|
90 |
+
|
91 |
+
global mel_basis, hann_window
|
92 |
+
dtype_device = str(y.dtype) + '_' + str(y.device)
|
93 |
+
fmax_dtype_device = str(fmax) + '_' + dtype_device
|
94 |
+
wnsize_dtype_device = str(win_size) + '_' + dtype_device
|
95 |
+
if fmax_dtype_device not in mel_basis:
|
96 |
+
mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
|
97 |
+
mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
|
98 |
+
if wnsize_dtype_device not in hann_window:
|
99 |
+
hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
|
100 |
+
|
101 |
+
y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
|
102 |
+
y = y.squeeze(1)
|
103 |
+
|
104 |
+
spec = torch.stft(y.float(), n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
|
105 |
+
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
|
106 |
+
|
107 |
+
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
|
108 |
+
|
109 |
+
spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
|
110 |
+
spec = spectral_normalize_torch(spec)
|
111 |
+
|
112 |
+
return spec
|
models.py
ADDED
@@ -0,0 +1,725 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import functional as F
|
6 |
+
|
7 |
+
import commons
|
8 |
+
import modules
|
9 |
+
import attentions
|
10 |
+
import monotonic_align
|
11 |
+
|
12 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
13 |
+
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
|
14 |
+
from commons import init_weights, get_padding
|
15 |
+
|
16 |
+
from loralib_tmp import layers
|
17 |
+
import loralib as lora
|
18 |
+
|
19 |
+
|
20 |
+
class StochasticDurationPredictor(nn.Module):
|
21 |
+
def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
|
22 |
+
super().__init__()
|
23 |
+
filter_channels = in_channels # it needs to be removed from future version.
|
24 |
+
self.in_channels = in_channels
|
25 |
+
self.filter_channels = filter_channels
|
26 |
+
self.kernel_size = kernel_size
|
27 |
+
self.p_dropout = p_dropout
|
28 |
+
self.n_flows = n_flows
|
29 |
+
self.gin_channels = gin_channels
|
30 |
+
|
31 |
+
self.log_flow = modules.Log()
|
32 |
+
self.flows = nn.ModuleList()
|
33 |
+
self.flows.append(modules.ElementwiseAffine(2))
|
34 |
+
for i in range(n_flows):
|
35 |
+
self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
|
36 |
+
self.flows.append(modules.Flip())
|
37 |
+
|
38 |
+
self.post_pre = nn.Conv1d(1, filter_channels, 1)
|
39 |
+
self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
|
40 |
+
self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
|
41 |
+
self.post_flows = nn.ModuleList()
|
42 |
+
self.post_flows.append(modules.ElementwiseAffine(2))
|
43 |
+
for i in range(4):
|
44 |
+
self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
|
45 |
+
self.post_flows.append(modules.Flip())
|
46 |
+
|
47 |
+
self.pre = nn.Conv1d(in_channels, filter_channels, 1)
|
48 |
+
self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
|
49 |
+
self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
|
50 |
+
if gin_channels != 0:
|
51 |
+
self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
|
52 |
+
|
53 |
+
def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
|
54 |
+
x = torch.detach(x)
|
55 |
+
x = self.pre(x)
|
56 |
+
if g is not None:
|
57 |
+
g = torch.detach(g)
|
58 |
+
x = x + self.cond(g)
|
59 |
+
x = self.convs(x, x_mask)
|
60 |
+
x = self.proj(x) * x_mask
|
61 |
+
|
62 |
+
if not reverse:
|
63 |
+
flows = self.flows
|
64 |
+
assert w is not None
|
65 |
+
|
66 |
+
logdet_tot_q = 0
|
67 |
+
h_w = self.post_pre(w)
|
68 |
+
h_w = self.post_convs(h_w, x_mask)
|
69 |
+
h_w = self.post_proj(h_w) * x_mask
|
70 |
+
e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
|
71 |
+
z_q = e_q
|
72 |
+
for flow in self.post_flows:
|
73 |
+
z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
|
74 |
+
logdet_tot_q += logdet_q
|
75 |
+
z_u, z1 = torch.split(z_q, [1, 1], 1)
|
76 |
+
u = torch.sigmoid(z_u) * x_mask
|
77 |
+
z0 = (w - u) * x_mask
|
78 |
+
logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1,2])
|
79 |
+
logq = torch.sum(-0.5 * (math.log(2*math.pi) + (e_q**2)) * x_mask, [1,2]) - logdet_tot_q
|
80 |
+
|
81 |
+
logdet_tot = 0
|
82 |
+
z0, logdet = self.log_flow(z0, x_mask)
|
83 |
+
logdet_tot += logdet
|
84 |
+
z = torch.cat([z0, z1], 1)
|
85 |
+
for flow in flows:
|
86 |
+
z, logdet = flow(z, x_mask, g=x, reverse=reverse)
|
87 |
+
logdet_tot = logdet_tot + logdet
|
88 |
+
nll = torch.sum(0.5 * (math.log(2*math.pi) + (z**2)) * x_mask, [1,2]) - logdet_tot
|
89 |
+
return nll + logq # [b]
|
90 |
+
else:
|
91 |
+
flows = list(reversed(self.flows))
|
92 |
+
flows = flows[:-2] + [flows[-1]] # remove a useless vflow
|
93 |
+
z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
|
94 |
+
for flow in flows:
|
95 |
+
z = flow(z, x_mask, g=x, reverse=reverse)
|
96 |
+
z0, z1 = torch.split(z, [1, 1], 1)
|
97 |
+
logw = z0
|
98 |
+
return logw
|
99 |
+
|
100 |
+
|
101 |
+
class DurationPredictor(nn.Module):
|
102 |
+
def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
|
103 |
+
super().__init__()
|
104 |
+
|
105 |
+
self.in_channels = in_channels
|
106 |
+
self.filter_channels = filter_channels
|
107 |
+
self.kernel_size = kernel_size
|
108 |
+
self.p_dropout = p_dropout
|
109 |
+
self.gin_channels = gin_channels
|
110 |
+
|
111 |
+
self.drop = nn.Dropout(p_dropout)
|
112 |
+
self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size//2)
|
113 |
+
self.norm_1 = modules.LayerNorm(filter_channels)
|
114 |
+
self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size//2)
|
115 |
+
self.norm_2 = modules.LayerNorm(filter_channels)
|
116 |
+
self.proj = nn.Conv1d(filter_channels, 1, 1)
|
117 |
+
|
118 |
+
if gin_channels != 0:
|
119 |
+
self.cond = nn.Conv1d(gin_channels, in_channels, 1)
|
120 |
+
|
121 |
+
def forward(self, x, x_mask, g=None):
|
122 |
+
x = torch.detach(x)
|
123 |
+
if g is not None:
|
124 |
+
g = torch.detach(g)
|
125 |
+
x = x + self.cond(g)
|
126 |
+
x = self.conv_1(x * x_mask)
|
127 |
+
x = torch.relu(x)
|
128 |
+
x = self.norm_1(x)
|
129 |
+
x = self.drop(x)
|
130 |
+
x = self.conv_2(x * x_mask)
|
131 |
+
x = torch.relu(x)
|
132 |
+
x = self.norm_2(x)
|
133 |
+
x = self.drop(x)
|
134 |
+
x = self.proj(x * x_mask)
|
135 |
+
return x * x_mask
|
136 |
+
|
137 |
+
|
138 |
+
class TextEncoder(nn.Module):
|
139 |
+
def __init__(self,
|
140 |
+
n_vocab,
|
141 |
+
out_channels,
|
142 |
+
hidden_channels,
|
143 |
+
filter_channels,
|
144 |
+
n_heads,
|
145 |
+
n_layers,
|
146 |
+
kernel_size,
|
147 |
+
p_dropout):
|
148 |
+
super().__init__()
|
149 |
+
self.n_vocab = n_vocab
|
150 |
+
self.out_channels = out_channels
|
151 |
+
self.hidden_channels = hidden_channels
|
152 |
+
self.filter_channels = filter_channels
|
153 |
+
self.n_heads = n_heads
|
154 |
+
self.n_layers = n_layers
|
155 |
+
self.kernel_size = kernel_size
|
156 |
+
self.p_dropout = p_dropout
|
157 |
+
|
158 |
+
self.emb = nn.Embedding(n_vocab, hidden_channels)
|
159 |
+
nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
|
160 |
+
|
161 |
+
self.encoder = attentions.Encoder(
|
162 |
+
hidden_channels,
|
163 |
+
filter_channels,
|
164 |
+
n_heads,
|
165 |
+
n_layers,
|
166 |
+
kernel_size,
|
167 |
+
p_dropout)
|
168 |
+
self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
169 |
+
|
170 |
+
def forward(self, x, x_lengths):
|
171 |
+
x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
|
172 |
+
x = torch.transpose(x, 1, -1) # [b, h, t]
|
173 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
|
174 |
+
|
175 |
+
x = self.encoder(x * x_mask, x_mask)
|
176 |
+
stats = self.proj(x) * x_mask
|
177 |
+
|
178 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
179 |
+
return x, m, logs, x_mask
|
180 |
+
|
181 |
+
|
182 |
+
class TextEncoder_lora(nn.Module):
|
183 |
+
def __init__(self,
|
184 |
+
n_vocab,
|
185 |
+
out_channels,
|
186 |
+
hidden_channels,
|
187 |
+
filter_channels,
|
188 |
+
n_heads,
|
189 |
+
n_layers,
|
190 |
+
kernel_size,
|
191 |
+
p_dropout):
|
192 |
+
super().__init__()
|
193 |
+
self.n_vocab = n_vocab
|
194 |
+
self.out_channels = out_channels
|
195 |
+
self.hidden_channels = hidden_channels
|
196 |
+
self.filter_channels = filter_channels
|
197 |
+
self.n_heads = n_heads
|
198 |
+
self.n_layers = n_layers
|
199 |
+
self.kernel_size = kernel_size
|
200 |
+
self.p_dropout = p_dropout
|
201 |
+
|
202 |
+
self.emb = nn.Embedding(n_vocab, hidden_channels, r=4)
|
203 |
+
nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
|
204 |
+
|
205 |
+
self.encoder = attentions.Encoder_lora(
|
206 |
+
hidden_channels,
|
207 |
+
filter_channels,
|
208 |
+
n_heads,
|
209 |
+
n_layers,
|
210 |
+
kernel_size,
|
211 |
+
p_dropout)
|
212 |
+
self.proj= nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
213 |
+
|
214 |
+
def forward(self, x, x_lengths):
|
215 |
+
x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
|
216 |
+
x = torch.transpose(x, 1, -1) # [b, h, t]
|
217 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
|
218 |
+
|
219 |
+
x = self.encoder(x * x_mask, x_mask)
|
220 |
+
stats = self.proj(x) * x_mask
|
221 |
+
|
222 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
223 |
+
return x, m, logs, x_mask
|
224 |
+
|
225 |
+
class ResidualCouplingBlock(nn.Module):
|
226 |
+
def __init__(self,
|
227 |
+
channels,
|
228 |
+
hidden_channels,
|
229 |
+
kernel_size,
|
230 |
+
dilation_rate,
|
231 |
+
n_layers,
|
232 |
+
n_flows=4,
|
233 |
+
gin_channels=0):
|
234 |
+
super().__init__()
|
235 |
+
self.channels = channels
|
236 |
+
self.hidden_channels = hidden_channels
|
237 |
+
self.kernel_size = kernel_size
|
238 |
+
self.dilation_rate = dilation_rate
|
239 |
+
self.n_layers = n_layers
|
240 |
+
self.n_flows = n_flows
|
241 |
+
self.gin_channels = gin_channels
|
242 |
+
|
243 |
+
self.flows = nn.ModuleList()
|
244 |
+
for i in range(n_flows):
|
245 |
+
self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
|
246 |
+
self.flows.append(modules.Flip())
|
247 |
+
|
248 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
249 |
+
if not reverse:
|
250 |
+
for flow in self.flows:
|
251 |
+
x, _ = flow(x, x_mask, g=g, reverse=reverse)
|
252 |
+
else:
|
253 |
+
for flow in reversed(self.flows):
|
254 |
+
x = flow(x, x_mask, g=g, reverse=reverse)
|
255 |
+
return x
|
256 |
+
|
257 |
+
|
258 |
+
class PosteriorEncoder(nn.Module):
|
259 |
+
def __init__(self,
|
260 |
+
in_channels,
|
261 |
+
out_channels,
|
262 |
+
hidden_channels,
|
263 |
+
kernel_size,
|
264 |
+
dilation_rate,
|
265 |
+
n_layers,
|
266 |
+
gin_channels=0):
|
267 |
+
super().__init__()
|
268 |
+
self.in_channels = in_channels
|
269 |
+
self.out_channels = out_channels
|
270 |
+
self.hidden_channels = hidden_channels
|
271 |
+
self.kernel_size = kernel_size
|
272 |
+
self.dilation_rate = dilation_rate
|
273 |
+
self.n_layers = n_layers
|
274 |
+
self.gin_channels = gin_channels
|
275 |
+
|
276 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
277 |
+
self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
|
278 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
279 |
+
|
280 |
+
def forward(self, x, x_lengths, g=None):
|
281 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
|
282 |
+
x = self.pre(x) * x_mask
|
283 |
+
x = self.enc(x, x_mask, g=g)
|
284 |
+
stats = self.proj(x) * x_mask
|
285 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
286 |
+
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
|
287 |
+
return z, m, logs, x_mask
|
288 |
+
|
289 |
+
|
290 |
+
class Generator(torch.nn.Module):
|
291 |
+
def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
|
292 |
+
super(Generator, self).__init__()
|
293 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
294 |
+
self.num_upsamples = len(upsample_rates)
|
295 |
+
self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
|
296 |
+
resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
|
297 |
+
|
298 |
+
self.ups = nn.ModuleList()
|
299 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
300 |
+
self.ups.append(weight_norm(
|
301 |
+
ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
|
302 |
+
k, u, padding=(k-u)//2)))
|
303 |
+
|
304 |
+
self.resblocks = nn.ModuleList()
|
305 |
+
for i in range(len(self.ups)):
|
306 |
+
ch = upsample_initial_channel//(2**(i+1))
|
307 |
+
for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
|
308 |
+
self.resblocks.append(resblock(ch, k, d))
|
309 |
+
|
310 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
311 |
+
self.ups.apply(init_weights)
|
312 |
+
|
313 |
+
if gin_channels != 0:
|
314 |
+
self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
315 |
+
|
316 |
+
def forward(self, x, g=None):
|
317 |
+
x = self.conv_pre(x)
|
318 |
+
if g is not None:
|
319 |
+
x = x + self.cond(g)
|
320 |
+
|
321 |
+
for i in range(self.num_upsamples):
|
322 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
323 |
+
x = self.ups[i](x)
|
324 |
+
xs = None
|
325 |
+
for j in range(self.num_kernels):
|
326 |
+
if xs is None:
|
327 |
+
xs = self.resblocks[i*self.num_kernels+j](x)
|
328 |
+
else:
|
329 |
+
xs += self.resblocks[i*self.num_kernels+j](x)
|
330 |
+
x = xs / self.num_kernels
|
331 |
+
x = F.leaky_relu(x)
|
332 |
+
x = self.conv_post(x)
|
333 |
+
x = torch.tanh(x)
|
334 |
+
|
335 |
+
return x
|
336 |
+
|
337 |
+
def remove_weight_norm(self):
|
338 |
+
print('Removing weight norm...')
|
339 |
+
for l in self.ups:
|
340 |
+
remove_weight_norm(l)
|
341 |
+
for l in self.resblocks:
|
342 |
+
l.remove_weight_norm()
|
343 |
+
|
344 |
+
|
345 |
+
class DiscriminatorP(torch.nn.Module):
|
346 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
347 |
+
super(DiscriminatorP, self).__init__()
|
348 |
+
self.period = period
|
349 |
+
self.use_spectral_norm = use_spectral_norm
|
350 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
351 |
+
self.convs = nn.ModuleList([
|
352 |
+
norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
353 |
+
norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
354 |
+
norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
355 |
+
norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
356 |
+
norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
|
357 |
+
])
|
358 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
359 |
+
|
360 |
+
def forward(self, x):
|
361 |
+
fmap = []
|
362 |
+
|
363 |
+
# 1d to 2d
|
364 |
+
b, c, t = x.shape
|
365 |
+
if t % self.period != 0: # pad first
|
366 |
+
n_pad = self.period - (t % self.period)
|
367 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
368 |
+
t = t + n_pad
|
369 |
+
x = x.view(b, c, t // self.period, self.period)
|
370 |
+
|
371 |
+
for l in self.convs:
|
372 |
+
x = l(x)
|
373 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
374 |
+
fmap.append(x)
|
375 |
+
x = self.conv_post(x)
|
376 |
+
fmap.append(x)
|
377 |
+
x = torch.flatten(x, 1, -1)
|
378 |
+
|
379 |
+
return x, fmap
|
380 |
+
|
381 |
+
|
382 |
+
class DiscriminatorS(torch.nn.Module):
|
383 |
+
def __init__(self, use_spectral_norm=False):
|
384 |
+
super(DiscriminatorS, self).__init__()
|
385 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
386 |
+
self.convs = nn.ModuleList([
|
387 |
+
norm_f(Conv1d(1, 16, 15, 1, padding=7)),
|
388 |
+
norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
|
389 |
+
norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
|
390 |
+
norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
|
391 |
+
norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
|
392 |
+
norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
|
393 |
+
])
|
394 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
395 |
+
|
396 |
+
def forward(self, x):
|
397 |
+
fmap = []
|
398 |
+
|
399 |
+
for l in self.convs:
|
400 |
+
x = l(x)
|
401 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
402 |
+
fmap.append(x)
|
403 |
+
x = self.conv_post(x)
|
404 |
+
fmap.append(x)
|
405 |
+
x = torch.flatten(x, 1, -1)
|
406 |
+
|
407 |
+
return x, fmap
|
408 |
+
|
409 |
+
|
410 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
411 |
+
def __init__(self, use_spectral_norm=False):
|
412 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
413 |
+
periods = [2,3,5,7,11]
|
414 |
+
|
415 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
416 |
+
discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
|
417 |
+
self.discriminators = nn.ModuleList(discs)
|
418 |
+
|
419 |
+
def forward(self, y, y_hat):
|
420 |
+
y_d_rs = []
|
421 |
+
y_d_gs = []
|
422 |
+
fmap_rs = []
|
423 |
+
fmap_gs = []
|
424 |
+
for i, d in enumerate(self.discriminators):
|
425 |
+
y_d_r, fmap_r = d(y)
|
426 |
+
y_d_g, fmap_g = d(y_hat)
|
427 |
+
y_d_rs.append(y_d_r)
|
428 |
+
y_d_gs.append(y_d_g)
|
429 |
+
fmap_rs.append(fmap_r)
|
430 |
+
fmap_gs.append(fmap_g)
|
431 |
+
|
432 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
433 |
+
|
434 |
+
|
435 |
+
|
436 |
+
class SynthesizerTrn(nn.Module):
|
437 |
+
"""
|
438 |
+
Synthesizer for Training
|
439 |
+
"""
|
440 |
+
|
441 |
+
def __init__(self,
|
442 |
+
n_vocab,
|
443 |
+
spec_channels,
|
444 |
+
segment_size,
|
445 |
+
inter_channels,
|
446 |
+
hidden_channels,
|
447 |
+
filter_channels,
|
448 |
+
n_heads,
|
449 |
+
n_layers,
|
450 |
+
kernel_size,
|
451 |
+
p_dropout,
|
452 |
+
resblock,
|
453 |
+
resblock_kernel_sizes,
|
454 |
+
resblock_dilation_sizes,
|
455 |
+
upsample_rates,
|
456 |
+
upsample_initial_channel,
|
457 |
+
upsample_kernel_sizes,
|
458 |
+
n_speakers=0,
|
459 |
+
gin_channels=0,
|
460 |
+
use_sdp=True,
|
461 |
+
**kwargs):
|
462 |
+
|
463 |
+
super().__init__()
|
464 |
+
self.n_vocab = n_vocab
|
465 |
+
self.spec_channels = spec_channels
|
466 |
+
self.inter_channels = inter_channels
|
467 |
+
self.hidden_channels = hidden_channels
|
468 |
+
self.filter_channels = filter_channels
|
469 |
+
self.n_heads = n_heads
|
470 |
+
self.n_layers = n_layers
|
471 |
+
self.kernel_size = kernel_size
|
472 |
+
self.p_dropout = p_dropout
|
473 |
+
self.resblock = resblock
|
474 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
475 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
476 |
+
self.upsample_rates = upsample_rates
|
477 |
+
self.upsample_initial_channel = upsample_initial_channel
|
478 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
479 |
+
self.segment_size = segment_size
|
480 |
+
self.n_speakers = n_speakers
|
481 |
+
self.gin_channels = gin_channels
|
482 |
+
|
483 |
+
self.use_sdp = use_sdp
|
484 |
+
|
485 |
+
self.enc_p = TextEncoder(n_vocab,
|
486 |
+
inter_channels,
|
487 |
+
hidden_channels,
|
488 |
+
filter_channels,
|
489 |
+
n_heads,
|
490 |
+
n_layers,
|
491 |
+
kernel_size,
|
492 |
+
p_dropout)
|
493 |
+
self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
|
494 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
495 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
|
496 |
+
|
497 |
+
if use_sdp:
|
498 |
+
self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
|
499 |
+
else:
|
500 |
+
self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
|
501 |
+
|
502 |
+
if n_speakers >= 1:
|
503 |
+
self.emb_g = nn.Embedding(n_speakers, gin_channels)
|
504 |
+
|
505 |
+
def forward(self, x, x_lengths, y, y_lengths, sid=None):
|
506 |
+
|
507 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
|
508 |
+
if self.n_speakers > 0:
|
509 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
510 |
+
else:
|
511 |
+
g = None
|
512 |
+
|
513 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
514 |
+
z_p = self.flow(z, y_mask, g=g)
|
515 |
+
|
516 |
+
with torch.no_grad():
|
517 |
+
# negative cross-entropy
|
518 |
+
s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
|
519 |
+
neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
|
520 |
+
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]
|
521 |
+
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]
|
522 |
+
neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
|
523 |
+
neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
|
524 |
+
|
525 |
+
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
|
526 |
+
attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
|
527 |
+
|
528 |
+
w = attn.sum(2)
|
529 |
+
if self.use_sdp:
|
530 |
+
l_length = self.dp(x, x_mask, w, g=g)
|
531 |
+
l_length = l_length / torch.sum(x_mask)
|
532 |
+
else:
|
533 |
+
logw_ = torch.log(w + 1e-6) * x_mask
|
534 |
+
logw = self.dp(x, x_mask, g=g)
|
535 |
+
l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
|
536 |
+
|
537 |
+
# expand prior
|
538 |
+
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
|
539 |
+
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
|
540 |
+
|
541 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
542 |
+
o = self.dec(z_slice, g=g)
|
543 |
+
return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
544 |
+
|
545 |
+
def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
|
546 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
|
547 |
+
if self.n_speakers > 0:
|
548 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
549 |
+
else:
|
550 |
+
g = None
|
551 |
+
|
552 |
+
if self.use_sdp:
|
553 |
+
logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
|
554 |
+
else:
|
555 |
+
logw = self.dp(x, x_mask, g=g)
|
556 |
+
w = torch.exp(logw) * x_mask * length_scale
|
557 |
+
w_ceil = torch.ceil(w)
|
558 |
+
y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
|
559 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
|
560 |
+
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
|
561 |
+
attn = commons.generate_path(w_ceil, attn_mask)
|
562 |
+
|
563 |
+
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
|
564 |
+
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
|
565 |
+
|
566 |
+
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
|
567 |
+
z = self.flow(z_p, y_mask, g=g, reverse=True)
|
568 |
+
o = self.dec((z * y_mask)[:,:,:max_len], g=g)
|
569 |
+
return o, attn, y_mask, (z, z_p, m_p, logs_p)
|
570 |
+
|
571 |
+
def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
|
572 |
+
assert self.n_speakers > 0, "n_speakers have to be larger than 0."
|
573 |
+
g_src = self.emb_g(sid_src).unsqueeze(-1)
|
574 |
+
g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
|
575 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
|
576 |
+
z_p = self.flow(z, y_mask, g=g_src)
|
577 |
+
z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
|
578 |
+
o_hat = self.dec(z_hat * y_mask, g=g_tgt)
|
579 |
+
return o_hat, y_mask, (z, z_p, z_hat)
|
580 |
+
|
581 |
+
|
582 |
+
class SynthesizerTrn_lora(nn.Module):
|
583 |
+
"""
|
584 |
+
Synthesizer for Training
|
585 |
+
"""
|
586 |
+
|
587 |
+
def __init__(self,
|
588 |
+
n_vocab,
|
589 |
+
spec_channels,
|
590 |
+
segment_size,
|
591 |
+
inter_channels,
|
592 |
+
hidden_channels,
|
593 |
+
filter_channels,
|
594 |
+
n_heads,
|
595 |
+
n_layers,
|
596 |
+
kernel_size,
|
597 |
+
p_dropout,
|
598 |
+
resblock,
|
599 |
+
resblock_kernel_sizes,
|
600 |
+
resblock_dilation_sizes,
|
601 |
+
upsample_rates,
|
602 |
+
upsample_initial_channel,
|
603 |
+
upsample_kernel_sizes,
|
604 |
+
n_speakers=0,
|
605 |
+
gin_channels=0,
|
606 |
+
use_sdp=True,
|
607 |
+
**kwargs):
|
608 |
+
|
609 |
+
super().__init__()
|
610 |
+
self.n_vocab = n_vocab
|
611 |
+
self.spec_channels = spec_channels
|
612 |
+
self.inter_channels = inter_channels
|
613 |
+
self.hidden_channels = hidden_channels
|
614 |
+
self.filter_channels = filter_channels
|
615 |
+
self.n_heads = n_heads
|
616 |
+
self.n_layers = n_layers
|
617 |
+
self.kernel_size = kernel_size
|
618 |
+
self.p_dropout = p_dropout
|
619 |
+
self.resblock = resblock
|
620 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
621 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
622 |
+
self.upsample_rates = upsample_rates
|
623 |
+
self.upsample_initial_channel = upsample_initial_channel
|
624 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
625 |
+
self.segment_size = segment_size
|
626 |
+
self.n_speakers = n_speakers
|
627 |
+
self.gin_channels = gin_channels
|
628 |
+
|
629 |
+
self.use_sdp = use_sdp
|
630 |
+
|
631 |
+
self.enc_p = TextEncoder_lora(n_vocab,
|
632 |
+
inter_channels,
|
633 |
+
hidden_channels,
|
634 |
+
filter_channels,
|
635 |
+
n_heads,
|
636 |
+
n_layers,
|
637 |
+
kernel_size,
|
638 |
+
p_dropout)
|
639 |
+
self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
|
640 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
641 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
|
642 |
+
|
643 |
+
if use_sdp:
|
644 |
+
self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
|
645 |
+
else:
|
646 |
+
self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
|
647 |
+
|
648 |
+
if n_speakers >= 1:
|
649 |
+
self.emb_g = nn.Embedding(n_speakers, gin_channels)
|
650 |
+
|
651 |
+
def forward(self, x, x_lengths, y, y_lengths, sid=None):
|
652 |
+
|
653 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
|
654 |
+
if self.n_speakers > 0:
|
655 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
656 |
+
else:
|
657 |
+
g = None
|
658 |
+
|
659 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
660 |
+
z_p = self.flow(z, y_mask, g=g)
|
661 |
+
|
662 |
+
with torch.no_grad():
|
663 |
+
# negative cross-entropy
|
664 |
+
s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
|
665 |
+
neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
|
666 |
+
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]
|
667 |
+
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]
|
668 |
+
neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
|
669 |
+
neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
|
670 |
+
|
671 |
+
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
|
672 |
+
attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
|
673 |
+
|
674 |
+
w = attn.sum(2)
|
675 |
+
if self.use_sdp:
|
676 |
+
l_length = self.dp(x, x_mask, w, g=g)
|
677 |
+
l_length = l_length / torch.sum(x_mask)
|
678 |
+
else:
|
679 |
+
logw_ = torch.log(w + 1e-6) * x_mask
|
680 |
+
logw = self.dp(x, x_mask, g=g)
|
681 |
+
l_length = torch.sum((logw - logw_)**2, [1,2]) / torch.sum(x_mask) # for averaging
|
682 |
+
|
683 |
+
# expand prior
|
684 |
+
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
|
685 |
+
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
|
686 |
+
|
687 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
688 |
+
o = self.dec(z_slice, g=g)
|
689 |
+
return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
690 |
+
|
691 |
+
def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
|
692 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
|
693 |
+
if self.n_speakers > 0:
|
694 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
695 |
+
else:
|
696 |
+
g = None
|
697 |
+
|
698 |
+
if self.use_sdp:
|
699 |
+
logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
|
700 |
+
else:
|
701 |
+
logw = self.dp(x, x_mask, g=g)
|
702 |
+
w = torch.exp(logw) * x_mask * length_scale
|
703 |
+
w_ceil = torch.ceil(w)
|
704 |
+
y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
|
705 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
|
706 |
+
attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
|
707 |
+
attn = commons.generate_path(w_ceil, attn_mask)
|
708 |
+
|
709 |
+
m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
|
710 |
+
logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
|
711 |
+
|
712 |
+
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
|
713 |
+
z = self.flow(z_p, y_mask, g=g, reverse=True)
|
714 |
+
o = self.dec((z * y_mask)[:,:,:max_len], g=g)
|
715 |
+
return o, attn, y_mask, (z, z_p, m_p, logs_p)
|
716 |
+
|
717 |
+
def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
|
718 |
+
assert self.n_speakers > 0, "n_speakers have to be larger than 0."
|
719 |
+
g_src = self.emb_g(sid_src).unsqueeze(-1)
|
720 |
+
g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
|
721 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
|
722 |
+
z_p = self.flow(z, y_mask, g=g_src)
|
723 |
+
z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
|
724 |
+
o_hat = self.dec(z_hat * y_mask, g=g_tgt)
|
725 |
+
return o_hat, y_mask, (z, z_p, z_hat)
|
models/G_9700.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5e250a75c0fb446825edcfba0b1ebc0872afff50143d66f3327a3eeb628bd874
|
3 |
+
size 159065393
|
models/config.json
ADDED
@@ -0,0 +1,321 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train": {
|
3 |
+
"log_interval": 10,
|
4 |
+
"eval_interval": 100,
|
5 |
+
"seed": 1234,
|
6 |
+
"epochs": 10000,
|
7 |
+
"learning_rate": 0.0002,
|
8 |
+
"betas": [
|
9 |
+
0.8,
|
10 |
+
0.99
|
11 |
+
],
|
12 |
+
"eps": 1e-09,
|
13 |
+
"batch_size": 16,
|
14 |
+
"fp16_run": true,
|
15 |
+
"lr_decay": 0.999875,
|
16 |
+
"segment_size": 8192,
|
17 |
+
"init_lr_ratio": 1,
|
18 |
+
"warmup_epochs": 0,
|
19 |
+
"c_mel": 45,
|
20 |
+
"c_kl": 1.0
|
21 |
+
},
|
22 |
+
"data": {
|
23 |
+
"training_files": "final_annotation_train.txt",
|
24 |
+
"validation_files": "final_annotation_val.txt",
|
25 |
+
"text_cleaners": [
|
26 |
+
"cjke_cleaners2"
|
27 |
+
],
|
28 |
+
"max_wav_value": 32768.0,
|
29 |
+
"sampling_rate": 22050,
|
30 |
+
"filter_length": 1024,
|
31 |
+
"hop_length": 256,
|
32 |
+
"win_length": 1024,
|
33 |
+
"n_mel_channels": 80,
|
34 |
+
"mel_fmin": 0.0,
|
35 |
+
"mel_fmax": null,
|
36 |
+
"add_blank": true,
|
37 |
+
"n_speakers": 161,
|
38 |
+
"cleaned_text": true
|
39 |
+
},
|
40 |
+
"model": {
|
41 |
+
"inter_channels": 192,
|
42 |
+
"hidden_channels": 192,
|
43 |
+
"filter_channels": 768,
|
44 |
+
"n_heads": 2,
|
45 |
+
"n_layers": 6,
|
46 |
+
"kernel_size": 3,
|
47 |
+
"p_dropout": 0.1,
|
48 |
+
"resblock": "1",
|
49 |
+
"resblock_kernel_sizes": [
|
50 |
+
3,
|
51 |
+
7,
|
52 |
+
11
|
53 |
+
],
|
54 |
+
"resblock_dilation_sizes": [
|
55 |
+
[
|
56 |
+
1,
|
57 |
+
3,
|
58 |
+
5
|
59 |
+
],
|
60 |
+
[
|
61 |
+
1,
|
62 |
+
3,
|
63 |
+
5
|
64 |
+
],
|
65 |
+
[
|
66 |
+
1,
|
67 |
+
3,
|
68 |
+
5
|
69 |
+
]
|
70 |
+
],
|
71 |
+
"upsample_rates": [
|
72 |
+
8,
|
73 |
+
8,
|
74 |
+
2,
|
75 |
+
2
|
76 |
+
],
|
77 |
+
"upsample_initial_channel": 512,
|
78 |
+
"upsample_kernel_sizes": [
|
79 |
+
16,
|
80 |
+
16,
|
81 |
+
4,
|
82 |
+
4
|
83 |
+
],
|
84 |
+
"n_layers_q": 3,
|
85 |
+
"use_spectral_norm": false,
|
86 |
+
"gin_channels": 256
|
87 |
+
},
|
88 |
+
"symbols": [
|
89 |
+
"_",
|
90 |
+
",",
|
91 |
+
".",
|
92 |
+
"!",
|
93 |
+
"?",
|
94 |
+
"-",
|
95 |
+
"~",
|
96 |
+
"\u2026",
|
97 |
+
"N",
|
98 |
+
"Q",
|
99 |
+
"a",
|
100 |
+
"b",
|
101 |
+
"d",
|
102 |
+
"e",
|
103 |
+
"f",
|
104 |
+
"g",
|
105 |
+
"h",
|
106 |
+
"i",
|
107 |
+
"j",
|
108 |
+
"k",
|
109 |
+
"l",
|
110 |
+
"m",
|
111 |
+
"n",
|
112 |
+
"o",
|
113 |
+
"p",
|
114 |
+
"s",
|
115 |
+
"t",
|
116 |
+
"u",
|
117 |
+
"v",
|
118 |
+
"w",
|
119 |
+
"x",
|
120 |
+
"y",
|
121 |
+
"z",
|
122 |
+
"\u0251",
|
123 |
+
"\u00e6",
|
124 |
+
"\u0283",
|
125 |
+
"\u0291",
|
126 |
+
"\u00e7",
|
127 |
+
"\u026f",
|
128 |
+
"\u026a",
|
129 |
+
"\u0254",
|
130 |
+
"\u025b",
|
131 |
+
"\u0279",
|
132 |
+
"\u00f0",
|
133 |
+
"\u0259",
|
134 |
+
"\u026b",
|
135 |
+
"\u0265",
|
136 |
+
"\u0278",
|
137 |
+
"\u028a",
|
138 |
+
"\u027e",
|
139 |
+
"\u0292",
|
140 |
+
"\u03b8",
|
141 |
+
"\u03b2",
|
142 |
+
"\u014b",
|
143 |
+
"\u0266",
|
144 |
+
"\u207c",
|
145 |
+
"\u02b0",
|
146 |
+
"`",
|
147 |
+
"^",
|
148 |
+
"#",
|
149 |
+
"*",
|
150 |
+
"=",
|
151 |
+
"\u02c8",
|
152 |
+
"\u02cc",
|
153 |
+
"\u2192",
|
154 |
+
"\u2193",
|
155 |
+
"\u2191",
|
156 |
+
" "
|
157 |
+
],
|
158 |
+
"speakers": {
|
159 |
+
"\u83c8\u6bd4\u8389\u65af\u5854\uff08Overload\uff09": 0,
|
160 |
+
"\u94c3\u5948\uff08\u590f\u65e5\uff09": 1,
|
161 |
+
"\u674f\u5948": 2,
|
162 |
+
"\u96ea\u83f2": 3,
|
163 |
+
"\u53ef\u53ef\u841d": 4,
|
164 |
+
"\u83c8\u6bd4\u8389\u65af\u5854": 5,
|
165 |
+
"\u7eeb\u97f3\uff08\u5723\u8bde\u8282\uff09": 6,
|
166 |
+
"\u78a7\uff08\u63d2\u73ed\u751f\uff09": 7,
|
167 |
+
"\u5609\u591c": 8,
|
168 |
+
"\u955c\u534e\uff08\u4e07\u5723\u8282\uff09": 9,
|
169 |
+
"\u73e0\u5e0c": 10,
|
170 |
+
"\u671b\uff08\u590f\u65e5\uff09": 11,
|
171 |
+
"\u94c3\u8393": 12,
|
172 |
+
"\u771f\u7434\uff08\u590f\u65e5\uff09": 13,
|
173 |
+
"\u9759\u6d41\uff08\u60c5\u4eba\u8282\uff09": 14,
|
174 |
+
"\u54b2\u604b": 15,
|
175 |
+
"\u8389\u739b": 16,
|
176 |
+
"\u9999\u6f84\uff08\u590f\u65e5\uff09": 17,
|
177 |
+
"\u5343\u6b4c": 18,
|
178 |
+
"\u5fcd": 19,
|
179 |
+
"\u4f9d\u91cc": 20,
|
180 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u65b0\u5e74\uff09": 21,
|
181 |
+
"\u80e1\u6843": 22,
|
182 |
+
"\u667a\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 23,
|
183 |
+
"\u4f18\u8863": 24,
|
184 |
+
"\u601c\uff08\u4e07\u5723\u8282\uff09": 25,
|
185 |
+
"\u681e": 26,
|
186 |
+
"\u9999\u7ec7\uff08\u590f\u65e5\uff09": 27,
|
187 |
+
"\u7948\u68a8\uff08\u65f6\u95f4\u65c5\u884c\uff09": 28,
|
188 |
+
"\u54b2\u604b\uff08\u590f\u65e5\uff09": 29,
|
189 |
+
"\u94c3\u5948": 30,
|
190 |
+
"\u771f\u6b65\uff08\u590f\u65e5\uff09": 31,
|
191 |
+
"\u4e9a\u91cc\u838e": 32,
|
192 |
+
"\u955c\u534e": 33,
|
193 |
+
"\u672a\u594f\u5e0c": 34,
|
194 |
+
"\u4f0a\u8389\u4e9a\uff08\u5723\u8bde\u8282\uff09": 35,
|
195 |
+
"\u7f8e\u91cc": 36,
|
196 |
+
"\u4f3c\u4f3c\u82b1": 37,
|
197 |
+
"\u514b\u8389\u4e1d\u63d0\u5a1c\uff08\u5723\u8bde\u8282\uff09": 38,
|
198 |
+
"\u7f8e\u51ac\uff08\u590f\u65e5\uff09": 39,
|
199 |
+
"\u8389\u739b\uff08\u7070\u59d1\u5a18\uff09": 40,
|
200 |
+
"\u94c3\u8393\uff08\u590f\u65e5\uff09": 41,
|
201 |
+
"\u53e4\u857e\u96c5": 42,
|
202 |
+
"\u7f8e\u7f8e\uff08\u4e07\u5723\u8282\uff09": 43,
|
203 |
+
"\u5343\u6b4c\uff08\u5723\u8bde\u8282\uff09": 44,
|
204 |
+
"\u78a7": 45,
|
205 |
+
"\u96ea": 46,
|
206 |
+
"\u60e0\u7406\u5b50\uff08\u60c5\u4eba\u8282\uff09": 47,
|
207 |
+
"\u4f0a\u7eea": 48,
|
208 |
+
"\u62c9\u59c6": 49,
|
209 |
+
"\u4e03\u4e03\u9999\uff08\u590f\u65e5\uff09": 50,
|
210 |
+
"\u94c3\u8393\uff08\u65b0\u5e74\uff09": 51,
|
211 |
+
"\u94c3\uff08\u6e38\u9a91\u5175\uff09": 52,
|
212 |
+
"\u73af\u5948": 53,
|
213 |
+
"\u60e0\u7406\u5b50": 54,
|
214 |
+
"\u53ef\u53ef\u841d\uff08\u793c\u670d\uff09": 55,
|
215 |
+
"\u4f18\u8863\uff08\u65b0\u5e74\uff09": 56,
|
216 |
+
"\u671b\uff08\u5723\u8bde\u8282\uff09": 57,
|
217 |
+
"\u53ef\u53ef\u841d\uff08\u516c\u4e3b\uff09": 58,
|
218 |
+
"\u7eaf\uff08\u590f\u65e5\uff09": 59,
|
219 |
+
"\u73af\u5948\uff08\u632f\u8896\uff09": 60,
|
220 |
+
"\u54b2\u604b\uff08\u5723\u8bde\u8282\uff09": 61,
|
221 |
+
"\u9732\u5a1c": 62,
|
222 |
+
"\u9999\u6f84\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 63,
|
223 |
+
"\u59ec\u5854": 64,
|
224 |
+
"\u77db\u4f9d\u672a\uff08\u65b0\u5e74\uff09": 65,
|
225 |
+
"\u681e\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 66,
|
226 |
+
"\u771f\u6b65\uff08\u7070\u59d1\u5a18\uff09": 67,
|
227 |
+
"\u601c\uff08\u65b0\u5e74\uff09": 68,
|
228 |
+
"\u5343\u6b4c\uff08\u590f\u65e5\uff09": 69,
|
229 |
+
"\u7948\u68a8": 70,
|
230 |
+
"\u7531\u52a0\u8389": 71,
|
231 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u590f\u65e5\uff09": 72,
|
232 |
+
"\u6b65\u7f8e\uff08\u4ed9\u5883\uff09": 73,
|
233 |
+
"\u53ef\u53ef\u841d\uff08\u590f\u65e5\uff09": 74,
|
234 |
+
"\u7eeb\u97f3": 75,
|
235 |
+
"\u7531\u52a0\u8389\uff08\u5723\u8bde\u8282\uff09": 76,
|
236 |
+
"\u771f\u9633": 77,
|
237 |
+
"\u96f7\u59c6": 78,
|
238 |
+
"\u831c\u91cc": 79,
|
239 |
+
"\u94c3": 80,
|
240 |
+
"\u51db\uff08\u5076\u50cf\u5927\u5e08\uff09": 81,
|
241 |
+
"\u51ef\u9732": 82,
|
242 |
+
"\u514b\u7f57\u4f9d\uff08\u5723\u5b66\u796d\uff09": 83,
|
243 |
+
"\u79cb\u4e43": 84,
|
244 |
+
"\u514b\u8389\u4e1d\u63d0\u5a1c": 85,
|
245 |
+
"\u4f69\u53ef\u8389\u59c6": 86,
|
246 |
+
"\u9732": 87,
|
247 |
+
"\u83ab\u59ae\u5361": 88,
|
248 |
+
"\u7483\u4e43": 89,
|
249 |
+
"\u9759\u6d41\uff08\u590f\u65e5\uff09": 90,
|
250 |
+
"\u7eba\u5e0c\uff08\u4e07\u5723\u8282\uff09": 91,
|
251 |
+
"\u9999\u6f84": 92,
|
252 |
+
"\u7f8e\u54b2": 93,
|
253 |
+
"\u7f8e\u91cc\uff08\u590f\u65e5\uff09": 94,
|
254 |
+
"\u667a": 95,
|
255 |
+
"\u521d\u97f3": 96,
|
256 |
+
"\u7483\u4e43\uff08\u4ed9\u5883\uff09": 97,
|
257 |
+
"\u536f\u6708\uff08\u5076\u50cf\u5927\u5e08\uff09": 98,
|
258 |
+
"\u7a7a\u82b1": 99,
|
259 |
+
"\u73e0\u5e0c\uff08\u590f\u65e5\uff09": 100,
|
260 |
+
"\u7eaf": 101,
|
261 |
+
"\u7f8e\u7f8e": 102,
|
262 |
+
"\u5fcd\uff08\u4e07\u5723\u8282\uff09": 103,
|
263 |
+
"\u59ae\u4fac\uff08\u5927\u6c5f\u6237\uff09": 104,
|
264 |
+
"\u5343\u7231\u7460": 105,
|
265 |
+
"\u53ef\u53ef\u841d\uff08\u65b0\u5e74\uff09": 106,
|
266 |
+
"\u601c\uff08\u516c\u4e3b\uff09": 107,
|
267 |
+
"\u672a\u594f\u5e0c\u3001\u7f8e\u7f8e\u3001\u955c\u534e": 108,
|
268 |
+
"\u4f0a\u8389\u4e9a": 109,
|
269 |
+
"\u671b": 110,
|
270 |
+
"\u4f18\u8863\uff08\u516c\u4e3b\uff09": 111,
|
271 |
+
"\u51ef\u9732\uff08\u516c\u4e3b\uff09": 112,
|
272 |
+
"\u514b\u7f57\u4f9d": 113,
|
273 |
+
"\u79cb\u4e43\uff08\u5723\u8bde\u8282\uff09": 114,
|
274 |
+
"\u4f18\u8863\uff08\u793c\u670d\uff09": 115,
|
275 |
+
"\u674f\u5948\uff08\u590f\u65e5\uff09": 116,
|
276 |
+
"\u9759\u6d41": 117,
|
277 |
+
"\u672a\u594f\u5e0c\uff08\u4e07\u5723\u8282\uff09": 118,
|
278 |
+
"\u831c\u91cc\uff08\u5929\u4f7f\uff09": 119,
|
279 |
+
"\u5bab\u5b50": 120,
|
280 |
+
"\u65e5\u548c\u8389\uff08\u65b0\u5e74\uff09": 121,
|
281 |
+
"\u4f18\u59ae": 122,
|
282 |
+
"\u5b89": 123,
|
283 |
+
"\u6d41\u590f\uff08\u590f\u65e5\uff09": 124,
|
284 |
+
"\u7f8e\u51ac\uff08\u5de5\u4f5c\u670d\uff09": 125,
|
285 |
+
"\u7f8e\u51ac": 126,
|
286 |
+
"\u521d\u97f3\uff08\u590f\u65e5\uff09": 127,
|
287 |
+
"\u77db\u4f9d\u672a": 128,
|
288 |
+
"\u51ef\u9732\uff08\u590f\u65e5\uff09": 129,
|
289 |
+
"\u83ab\u59ae\u5361\uff08\u9b54\u6cd5\u5c11\u5973\uff09": 130,
|
290 |
+
"\u9999\u7ec7": 131,
|
291 |
+
"\u5bab\u5b50\uff08\u4e07\u5723\u8282\uff09": 132,
|
292 |
+
"\u7f8e\u54b2\uff08\u4e07\u5723\u8282\uff09": 133,
|
293 |
+
"\u59ae\u4fac": 134,
|
294 |
+
"\u601c": 135,
|
295 |
+
"\u771f\u9633\uff08\u6e38\u9a91\u5175\uff09": 136,
|
296 |
+
"\u78a7\uff08\u5de5\u4f5c\u670d\uff09": 137,
|
297 |
+
"\u65e5\u548c\u8389": 138,
|
298 |
+
"\u4e03\u4e03\u9999": 139,
|
299 |
+
"\u771f\u7434": 140,
|
300 |
+
"\u6b65\u7f8e": 141,
|
301 |
+
"\u6df1\u6708": 142,
|
302 |
+
"\u65e5\u548c\u8389\uff08\u516c\u4e3b\uff09": 143,
|
303 |
+
"\u4f0a\u7eea\uff08\u590f\u65e5\uff09": 144,
|
304 |
+
"\u8309\u8389": 145,
|
305 |
+
"\u4f69\u53ef\u8389\u59c6\uff08\u516c\u4e3b\uff09": 146,
|
306 |
+
"\u7a7a\u82b1\uff08\u5927\u6c5f\u6237\uff09": 147,
|
307 |
+
"\u60e0\u7406\u5b50\uff08\u590f\u65e5\uff09": 148,
|
308 |
+
"\u51ef\u9732\uff08\u65b0\u5e74\uff09": 149,
|
309 |
+
"\u7231\u871c\u8389\u96c5": 150,
|
310 |
+
"\u4f9d\u91cc\uff08\u5929\u4f7f\uff09": 151,
|
311 |
+
"\u8309\u8389\uff08\u4e07\u5723\u8282\uff09": 152,
|
312 |
+
"\u771f\u7434\uff08\u7070\u59d1\u5a18\uff09": 153,
|
313 |
+
"\u7eba\u5e0c": 154,
|
314 |
+
"\u771f\u6b65": 155,
|
315 |
+
"\u6d41\u590f": 156,
|
316 |
+
"\u4f3c\u4f3c\u82b1\uff08\u65b0\u5e74\uff09": 157,
|
317 |
+
"\u672a\u592e\uff08\u5076\u50cf\u5927\u5e08\uff09": 158,
|
318 |
+
"\u80e1\u6843\uff08\u5723\u8bde\u8282\uff09": 159,
|
319 |
+
"\u5343\u7231\u7460\uff08\u5723\u5b66\u796d\uff09": 160
|
320 |
+
}
|
321 |
+
}
|
modules.py
ADDED
@@ -0,0 +1,390 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import scipy
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
from torch.nn import functional as F
|
8 |
+
|
9 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
10 |
+
from torch.nn.utils import weight_norm, remove_weight_norm
|
11 |
+
|
12 |
+
import commons
|
13 |
+
from commons import init_weights, get_padding
|
14 |
+
from transforms import piecewise_rational_quadratic_transform
|
15 |
+
|
16 |
+
|
17 |
+
LRELU_SLOPE = 0.1
|
18 |
+
|
19 |
+
|
20 |
+
class LayerNorm(nn.Module):
|
21 |
+
def __init__(self, channels, eps=1e-5):
|
22 |
+
super().__init__()
|
23 |
+
self.channels = channels
|
24 |
+
self.eps = eps
|
25 |
+
|
26 |
+
self.gamma = nn.Parameter(torch.ones(channels))
|
27 |
+
self.beta = nn.Parameter(torch.zeros(channels))
|
28 |
+
|
29 |
+
def forward(self, x):
|
30 |
+
x = x.transpose(1, -1)
|
31 |
+
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
|
32 |
+
return x.transpose(1, -1)
|
33 |
+
|
34 |
+
|
35 |
+
class ConvReluNorm(nn.Module):
|
36 |
+
def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
|
37 |
+
super().__init__()
|
38 |
+
self.in_channels = in_channels
|
39 |
+
self.hidden_channels = hidden_channels
|
40 |
+
self.out_channels = out_channels
|
41 |
+
self.kernel_size = kernel_size
|
42 |
+
self.n_layers = n_layers
|
43 |
+
self.p_dropout = p_dropout
|
44 |
+
assert n_layers > 1, "Number of layers should be larger than 0."
|
45 |
+
|
46 |
+
self.conv_layers = nn.ModuleList()
|
47 |
+
self.norm_layers = nn.ModuleList()
|
48 |
+
self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
|
49 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
50 |
+
self.relu_drop = nn.Sequential(
|
51 |
+
nn.ReLU(),
|
52 |
+
nn.Dropout(p_dropout))
|
53 |
+
for _ in range(n_layers-1):
|
54 |
+
self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
|
55 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
56 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
|
57 |
+
self.proj.weight.data.zero_()
|
58 |
+
self.proj.bias.data.zero_()
|
59 |
+
|
60 |
+
def forward(self, x, x_mask):
|
61 |
+
x_org = x
|
62 |
+
for i in range(self.n_layers):
|
63 |
+
x = self.conv_layers[i](x * x_mask)
|
64 |
+
x = self.norm_layers[i](x)
|
65 |
+
x = self.relu_drop(x)
|
66 |
+
x = x_org + self.proj(x)
|
67 |
+
return x * x_mask
|
68 |
+
|
69 |
+
|
70 |
+
class DDSConv(nn.Module):
|
71 |
+
"""
|
72 |
+
Dialted and Depth-Separable Convolution
|
73 |
+
"""
|
74 |
+
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
|
75 |
+
super().__init__()
|
76 |
+
self.channels = channels
|
77 |
+
self.kernel_size = kernel_size
|
78 |
+
self.n_layers = n_layers
|
79 |
+
self.p_dropout = p_dropout
|
80 |
+
|
81 |
+
self.drop = nn.Dropout(p_dropout)
|
82 |
+
self.convs_sep = nn.ModuleList()
|
83 |
+
self.convs_1x1 = nn.ModuleList()
|
84 |
+
self.norms_1 = nn.ModuleList()
|
85 |
+
self.norms_2 = nn.ModuleList()
|
86 |
+
for i in range(n_layers):
|
87 |
+
dilation = kernel_size ** i
|
88 |
+
padding = (kernel_size * dilation - dilation) // 2
|
89 |
+
self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
|
90 |
+
groups=channels, dilation=dilation, padding=padding
|
91 |
+
))
|
92 |
+
self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
|
93 |
+
self.norms_1.append(LayerNorm(channels))
|
94 |
+
self.norms_2.append(LayerNorm(channels))
|
95 |
+
|
96 |
+
def forward(self, x, x_mask, g=None):
|
97 |
+
if g is not None:
|
98 |
+
x = x + g
|
99 |
+
for i in range(self.n_layers):
|
100 |
+
y = self.convs_sep[i](x * x_mask)
|
101 |
+
y = self.norms_1[i](y)
|
102 |
+
y = F.gelu(y)
|
103 |
+
y = self.convs_1x1[i](y)
|
104 |
+
y = self.norms_2[i](y)
|
105 |
+
y = F.gelu(y)
|
106 |
+
y = self.drop(y)
|
107 |
+
x = x + y
|
108 |
+
return x * x_mask
|
109 |
+
|
110 |
+
|
111 |
+
class WN(torch.nn.Module):
|
112 |
+
def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
|
113 |
+
super(WN, self).__init__()
|
114 |
+
assert(kernel_size % 2 == 1)
|
115 |
+
self.hidden_channels =hidden_channels
|
116 |
+
self.kernel_size = kernel_size,
|
117 |
+
self.dilation_rate = dilation_rate
|
118 |
+
self.n_layers = n_layers
|
119 |
+
self.gin_channels = gin_channels
|
120 |
+
self.p_dropout = p_dropout
|
121 |
+
|
122 |
+
self.in_layers = torch.nn.ModuleList()
|
123 |
+
self.res_skip_layers = torch.nn.ModuleList()
|
124 |
+
self.drop = nn.Dropout(p_dropout)
|
125 |
+
|
126 |
+
if gin_channels != 0:
|
127 |
+
cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
|
128 |
+
self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
|
129 |
+
|
130 |
+
for i in range(n_layers):
|
131 |
+
dilation = dilation_rate ** i
|
132 |
+
padding = int((kernel_size * dilation - dilation) / 2)
|
133 |
+
in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
|
134 |
+
dilation=dilation, padding=padding)
|
135 |
+
in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
|
136 |
+
self.in_layers.append(in_layer)
|
137 |
+
|
138 |
+
# last one is not necessary
|
139 |
+
if i < n_layers - 1:
|
140 |
+
res_skip_channels = 2 * hidden_channels
|
141 |
+
else:
|
142 |
+
res_skip_channels = hidden_channels
|
143 |
+
|
144 |
+
res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
|
145 |
+
res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
|
146 |
+
self.res_skip_layers.append(res_skip_layer)
|
147 |
+
|
148 |
+
def forward(self, x, x_mask, g=None, **kwargs):
|
149 |
+
output = torch.zeros_like(x)
|
150 |
+
n_channels_tensor = torch.IntTensor([self.hidden_channels])
|
151 |
+
|
152 |
+
if g is not None:
|
153 |
+
g = self.cond_layer(g)
|
154 |
+
|
155 |
+
for i in range(self.n_layers):
|
156 |
+
x_in = self.in_layers[i](x)
|
157 |
+
if g is not None:
|
158 |
+
cond_offset = i * 2 * self.hidden_channels
|
159 |
+
g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
|
160 |
+
else:
|
161 |
+
g_l = torch.zeros_like(x_in)
|
162 |
+
|
163 |
+
acts = commons.fused_add_tanh_sigmoid_multiply(
|
164 |
+
x_in,
|
165 |
+
g_l,
|
166 |
+
n_channels_tensor)
|
167 |
+
acts = self.drop(acts)
|
168 |
+
|
169 |
+
res_skip_acts = self.res_skip_layers[i](acts)
|
170 |
+
if i < self.n_layers - 1:
|
171 |
+
res_acts = res_skip_acts[:,:self.hidden_channels,:]
|
172 |
+
x = (x + res_acts) * x_mask
|
173 |
+
output = output + res_skip_acts[:,self.hidden_channels:,:]
|
174 |
+
else:
|
175 |
+
output = output + res_skip_acts
|
176 |
+
return output * x_mask
|
177 |
+
|
178 |
+
def remove_weight_norm(self):
|
179 |
+
if self.gin_channels != 0:
|
180 |
+
torch.nn.utils.remove_weight_norm(self.cond_layer)
|
181 |
+
for l in self.in_layers:
|
182 |
+
torch.nn.utils.remove_weight_norm(l)
|
183 |
+
for l in self.res_skip_layers:
|
184 |
+
torch.nn.utils.remove_weight_norm(l)
|
185 |
+
|
186 |
+
|
187 |
+
class ResBlock1(torch.nn.Module):
|
188 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
|
189 |
+
super(ResBlock1, self).__init__()
|
190 |
+
self.convs1 = nn.ModuleList([
|
191 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
|
192 |
+
padding=get_padding(kernel_size, dilation[0]))),
|
193 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
|
194 |
+
padding=get_padding(kernel_size, dilation[1]))),
|
195 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
|
196 |
+
padding=get_padding(kernel_size, dilation[2])))
|
197 |
+
])
|
198 |
+
self.convs1.apply(init_weights)
|
199 |
+
|
200 |
+
self.convs2 = nn.ModuleList([
|
201 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
202 |
+
padding=get_padding(kernel_size, 1))),
|
203 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
204 |
+
padding=get_padding(kernel_size, 1))),
|
205 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
|
206 |
+
padding=get_padding(kernel_size, 1)))
|
207 |
+
])
|
208 |
+
self.convs2.apply(init_weights)
|
209 |
+
|
210 |
+
def forward(self, x, x_mask=None):
|
211 |
+
for c1, c2 in zip(self.convs1, self.convs2):
|
212 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
213 |
+
if x_mask is not None:
|
214 |
+
xt = xt * x_mask
|
215 |
+
xt = c1(xt)
|
216 |
+
xt = F.leaky_relu(xt, LRELU_SLOPE)
|
217 |
+
if x_mask is not None:
|
218 |
+
xt = xt * x_mask
|
219 |
+
xt = c2(xt)
|
220 |
+
x = xt + x
|
221 |
+
if x_mask is not None:
|
222 |
+
x = x * x_mask
|
223 |
+
return x
|
224 |
+
|
225 |
+
def remove_weight_norm(self):
|
226 |
+
for l in self.convs1:
|
227 |
+
remove_weight_norm(l)
|
228 |
+
for l in self.convs2:
|
229 |
+
remove_weight_norm(l)
|
230 |
+
|
231 |
+
|
232 |
+
class ResBlock2(torch.nn.Module):
|
233 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
|
234 |
+
super(ResBlock2, self).__init__()
|
235 |
+
self.convs = nn.ModuleList([
|
236 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
|
237 |
+
padding=get_padding(kernel_size, dilation[0]))),
|
238 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
|
239 |
+
padding=get_padding(kernel_size, dilation[1])))
|
240 |
+
])
|
241 |
+
self.convs.apply(init_weights)
|
242 |
+
|
243 |
+
def forward(self, x, x_mask=None):
|
244 |
+
for c in self.convs:
|
245 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
246 |
+
if x_mask is not None:
|
247 |
+
xt = xt * x_mask
|
248 |
+
xt = c(xt)
|
249 |
+
x = xt + x
|
250 |
+
if x_mask is not None:
|
251 |
+
x = x * x_mask
|
252 |
+
return x
|
253 |
+
|
254 |
+
def remove_weight_norm(self):
|
255 |
+
for l in self.convs:
|
256 |
+
remove_weight_norm(l)
|
257 |
+
|
258 |
+
|
259 |
+
class Log(nn.Module):
|
260 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
261 |
+
if not reverse:
|
262 |
+
y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
|
263 |
+
logdet = torch.sum(-y, [1, 2])
|
264 |
+
return y, logdet
|
265 |
+
else:
|
266 |
+
x = torch.exp(x) * x_mask
|
267 |
+
return x
|
268 |
+
|
269 |
+
|
270 |
+
class Flip(nn.Module):
|
271 |
+
def forward(self, x, *args, reverse=False, **kwargs):
|
272 |
+
x = torch.flip(x, [1])
|
273 |
+
if not reverse:
|
274 |
+
logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
|
275 |
+
return x, logdet
|
276 |
+
else:
|
277 |
+
return x
|
278 |
+
|
279 |
+
|
280 |
+
class ElementwiseAffine(nn.Module):
|
281 |
+
def __init__(self, channels):
|
282 |
+
super().__init__()
|
283 |
+
self.channels = channels
|
284 |
+
self.m = nn.Parameter(torch.zeros(channels,1))
|
285 |
+
self.logs = nn.Parameter(torch.zeros(channels,1))
|
286 |
+
|
287 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
288 |
+
if not reverse:
|
289 |
+
y = self.m + torch.exp(self.logs) * x
|
290 |
+
y = y * x_mask
|
291 |
+
logdet = torch.sum(self.logs * x_mask, [1,2])
|
292 |
+
return y, logdet
|
293 |
+
else:
|
294 |
+
x = (x - self.m) * torch.exp(-self.logs) * x_mask
|
295 |
+
return x
|
296 |
+
|
297 |
+
|
298 |
+
class ResidualCouplingLayer(nn.Module):
|
299 |
+
def __init__(self,
|
300 |
+
channels,
|
301 |
+
hidden_channels,
|
302 |
+
kernel_size,
|
303 |
+
dilation_rate,
|
304 |
+
n_layers,
|
305 |
+
p_dropout=0,
|
306 |
+
gin_channels=0,
|
307 |
+
mean_only=False):
|
308 |
+
assert channels % 2 == 0, "channels should be divisible by 2"
|
309 |
+
super().__init__()
|
310 |
+
self.channels = channels
|
311 |
+
self.hidden_channels = hidden_channels
|
312 |
+
self.kernel_size = kernel_size
|
313 |
+
self.dilation_rate = dilation_rate
|
314 |
+
self.n_layers = n_layers
|
315 |
+
self.half_channels = channels // 2
|
316 |
+
self.mean_only = mean_only
|
317 |
+
|
318 |
+
self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
|
319 |
+
self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
|
320 |
+
self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
|
321 |
+
self.post.weight.data.zero_()
|
322 |
+
self.post.bias.data.zero_()
|
323 |
+
|
324 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
325 |
+
x0, x1 = torch.split(x, [self.half_channels]*2, 1)
|
326 |
+
h = self.pre(x0) * x_mask
|
327 |
+
h = self.enc(h, x_mask, g=g)
|
328 |
+
stats = self.post(h) * x_mask
|
329 |
+
if not self.mean_only:
|
330 |
+
m, logs = torch.split(stats, [self.half_channels]*2, 1)
|
331 |
+
else:
|
332 |
+
m = stats
|
333 |
+
logs = torch.zeros_like(m)
|
334 |
+
|
335 |
+
if not reverse:
|
336 |
+
x1 = m + x1 * torch.exp(logs) * x_mask
|
337 |
+
x = torch.cat([x0, x1], 1)
|
338 |
+
logdet = torch.sum(logs, [1,2])
|
339 |
+
return x, logdet
|
340 |
+
else:
|
341 |
+
x1 = (x1 - m) * torch.exp(-logs) * x_mask
|
342 |
+
x = torch.cat([x0, x1], 1)
|
343 |
+
return x
|
344 |
+
|
345 |
+
|
346 |
+
class ConvFlow(nn.Module):
|
347 |
+
def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
|
348 |
+
super().__init__()
|
349 |
+
self.in_channels = in_channels
|
350 |
+
self.filter_channels = filter_channels
|
351 |
+
self.kernel_size = kernel_size
|
352 |
+
self.n_layers = n_layers
|
353 |
+
self.num_bins = num_bins
|
354 |
+
self.tail_bound = tail_bound
|
355 |
+
self.half_channels = in_channels // 2
|
356 |
+
|
357 |
+
self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
|
358 |
+
self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
|
359 |
+
self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
|
360 |
+
self.proj.weight.data.zero_()
|
361 |
+
self.proj.bias.data.zero_()
|
362 |
+
|
363 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
364 |
+
x0, x1 = torch.split(x, [self.half_channels]*2, 1)
|
365 |
+
h = self.pre(x0)
|
366 |
+
h = self.convs(h, x_mask, g=g)
|
367 |
+
h = self.proj(h) * x_mask
|
368 |
+
|
369 |
+
b, c, t = x0.shape
|
370 |
+
h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
|
371 |
+
|
372 |
+
unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
|
373 |
+
unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
|
374 |
+
unnormalized_derivatives = h[..., 2 * self.num_bins:]
|
375 |
+
|
376 |
+
x1, logabsdet = piecewise_rational_quadratic_transform(x1,
|
377 |
+
unnormalized_widths,
|
378 |
+
unnormalized_heights,
|
379 |
+
unnormalized_derivatives,
|
380 |
+
inverse=reverse,
|
381 |
+
tails='linear',
|
382 |
+
tail_bound=self.tail_bound
|
383 |
+
)
|
384 |
+
|
385 |
+
x = torch.cat([x0, x1], 1) * x_mask
|
386 |
+
logdet = torch.sum(logabsdet * x_mask, [1,2])
|
387 |
+
if not reverse:
|
388 |
+
return x, logdet
|
389 |
+
else:
|
390 |
+
return x
|
monotonic_align/__init__.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
from .monotonic_align.core import maximum_path_c
|
4 |
+
|
5 |
+
|
6 |
+
def maximum_path(neg_cent, mask):
|
7 |
+
""" Cython optimized version.
|
8 |
+
neg_cent: [b, t_t, t_s]
|
9 |
+
mask: [b, t_t, t_s]
|
10 |
+
"""
|
11 |
+
device = neg_cent.device
|
12 |
+
dtype = neg_cent.dtype
|
13 |
+
neg_cent = neg_cent.data.cpu().numpy().astype(np.float32)
|
14 |
+
path = np.zeros(neg_cent.shape, dtype=np.int32)
|
15 |
+
|
16 |
+
t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(np.int32)
|
17 |
+
t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(np.int32)
|
18 |
+
maximum_path_c(path, neg_cent, t_t_max, t_s_max)
|
19 |
+
return torch.from_numpy(path).to(device=device, dtype=dtype)
|
monotonic_align/core.c
ADDED
The diff for this file is too large to render.
See raw diff
|
|
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 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
sequence = []
|
20 |
+
symbol_to_id = {s: i for i, s in enumerate(symbols)}
|
21 |
+
clean_text = _clean_text(text, cleaner_names)
|
22 |
+
print(clean_text)
|
23 |
+
print(f" length:{len(clean_text)}")
|
24 |
+
for symbol in clean_text:
|
25 |
+
if symbol not in symbol_to_id.keys():
|
26 |
+
continue
|
27 |
+
symbol_id = symbol_to_id[symbol]
|
28 |
+
sequence += [symbol_id]
|
29 |
+
print(f" length:{len(sequence)}")
|
30 |
+
return sequence
|
31 |
+
|
32 |
+
|
33 |
+
def cleaned_text_to_sequence(cleaned_text, symbols):
|
34 |
+
'''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
|
35 |
+
Args:
|
36 |
+
text: string to convert to a sequence
|
37 |
+
Returns:
|
38 |
+
List of integers corresponding to the symbols in the text
|
39 |
+
'''
|
40 |
+
symbol_to_id = {s: i for i, s in enumerate(symbols)}
|
41 |
+
sequence = [symbol_to_id[symbol] for symbol in cleaned_text if symbol in symbol_to_id.keys()]
|
42 |
+
return sequence
|
43 |
+
|
44 |
+
|
45 |
+
def sequence_to_text(sequence):
|
46 |
+
'''Converts a sequence of IDs back to a string'''
|
47 |
+
result = ''
|
48 |
+
for symbol_id in sequence:
|
49 |
+
s = _id_to_symbol[symbol_id]
|
50 |
+
result += s
|
51 |
+
return result
|
52 |
+
|
53 |
+
|
54 |
+
def _clean_text(text, cleaner_names):
|
55 |
+
for name in cleaner_names:
|
56 |
+
cleaner = getattr(cleaners, name)
|
57 |
+
if not cleaner:
|
58 |
+
raise Exception('Unknown cleaner: %s' % name)
|
59 |
+
text = cleaner(text)
|
60 |
+
return text
|
text/cantonese.py
ADDED
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import cn2an
|
3 |
+
import opencc
|
4 |
+
|
5 |
+
|
6 |
+
converter = opencc.OpenCC('jyutjyu')
|
7 |
+
|
8 |
+
# List of (Latin alphabet, ipa) pairs:
|
9 |
+
_latin_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
10 |
+
('A', 'ei˥'),
|
11 |
+
('B', 'biː˥'),
|
12 |
+
('C', 'siː˥'),
|
13 |
+
('D', 'tiː˥'),
|
14 |
+
('E', 'iː˥'),
|
15 |
+
('F', 'e˥fuː˨˩'),
|
16 |
+
('G', 'tsiː˥'),
|
17 |
+
('H', 'ɪk̚˥tsʰyː˨˩'),
|
18 |
+
('I', 'ɐi˥'),
|
19 |
+
('J', 'tsei˥'),
|
20 |
+
('K', 'kʰei˥'),
|
21 |
+
('L', 'e˥llou˨˩'),
|
22 |
+
('M', 'ɛːm˥'),
|
23 |
+
('N', 'ɛːn˥'),
|
24 |
+
('O', 'ou˥'),
|
25 |
+
('P', 'pʰiː˥'),
|
26 |
+
('Q', 'kʰiːu˥'),
|
27 |
+
('R', 'aː˥lou˨˩'),
|
28 |
+
('S', 'ɛː˥siː˨˩'),
|
29 |
+
('T', 'tʰiː˥'),
|
30 |
+
('U', 'juː˥'),
|
31 |
+
('V', 'wiː˥'),
|
32 |
+
('W', 'tʊk̚˥piː˥juː˥'),
|
33 |
+
('X', 'ɪk̚˥siː˨˩'),
|
34 |
+
('Y', 'waːi˥'),
|
35 |
+
('Z', 'iː˨sɛːt̚˥')
|
36 |
+
]]
|
37 |
+
|
38 |
+
|
39 |
+
def number_to_cantonese(text):
|
40 |
+
return re.sub(r'\d+(?:\.?\d+)?', lambda x: cn2an.an2cn(x.group()), text)
|
41 |
+
|
42 |
+
|
43 |
+
def latin_to_ipa(text):
|
44 |
+
for regex, replacement in _latin_to_ipa:
|
45 |
+
text = re.sub(regex, replacement, text)
|
46 |
+
return text
|
47 |
+
|
48 |
+
|
49 |
+
def cantonese_to_ipa(text):
|
50 |
+
text = number_to_cantonese(text.upper())
|
51 |
+
text = converter.convert(text).replace('-','').replace('$',' ')
|
52 |
+
text = re.sub(r'[A-Z]', lambda x: latin_to_ipa(x.group())+' ', text)
|
53 |
+
text = re.sub(r'[、;:]', ',', text)
|
54 |
+
text = re.sub(r'\s*,\s*', ', ', text)
|
55 |
+
text = re.sub(r'\s*。\s*', '. ', text)
|
56 |
+
text = re.sub(r'\s*?\s*', '? ', text)
|
57 |
+
text = re.sub(r'\s*!\s*', '! ', text)
|
58 |
+
text = re.sub(r'\s*$', '', text)
|
59 |
+
return text
|
text/cleaners.py
ADDED
@@ -0,0 +1,134 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from text.japanese import japanese_to_romaji_with_accent, japanese_to_ipa, japanese_to_ipa2, japanese_to_ipa3
|
3 |
+
from text.korean import latin_to_hangul, number_to_hangul, divide_hangul, korean_to_lazy_ipa, korean_to_ipa
|
4 |
+
from text.mandarin import number_to_chinese, chinese_to_bopomofo, latin_to_bopomofo, chinese_to_romaji, chinese_to_lazy_ipa, chinese_to_ipa, chinese_to_ipa2
|
5 |
+
from text.sanskrit import devanagari_to_ipa
|
6 |
+
from text.english import english_to_lazy_ipa, english_to_ipa2, english_to_lazy_ipa2
|
7 |
+
from text.thai import num_to_thai, latin_to_thai
|
8 |
+
# from text.shanghainese import shanghainese_to_ipa
|
9 |
+
# from text.cantonese import cantonese_to_ipa
|
10 |
+
# from text.ngu_dialect import ngu_dialect_to_ipa
|
11 |
+
|
12 |
+
|
13 |
+
def japanese_cleaners(text):
|
14 |
+
text = japanese_to_romaji_with_accent(text)
|
15 |
+
text = re.sub(r'([A-Za-z])$', r'\1.', text)
|
16 |
+
return text
|
17 |
+
|
18 |
+
|
19 |
+
def japanese_cleaners2(text):
|
20 |
+
return japanese_cleaners(text).replace('ts', 'ʦ').replace('...', '…')
|
21 |
+
|
22 |
+
|
23 |
+
def korean_cleaners(text):
|
24 |
+
'''Pipeline for Korean text'''
|
25 |
+
text = latin_to_hangul(text)
|
26 |
+
text = number_to_hangul(text)
|
27 |
+
text = divide_hangul(text)
|
28 |
+
text = re.sub(r'([\u3131-\u3163])$', r'\1.', text)
|
29 |
+
return text
|
30 |
+
|
31 |
+
|
32 |
+
# def chinese_cleaners(text):
|
33 |
+
# '''Pipeline for Chinese text'''
|
34 |
+
# text = number_to_chinese(text)
|
35 |
+
# text = chinese_to_bopomofo(text)
|
36 |
+
# text = latin_to_bopomofo(text)
|
37 |
+
# text = re.sub(r'([ˉˊˇˋ˙])$', r'\1。', text)
|
38 |
+
# return text
|
39 |
+
|
40 |
+
def chinese_cleaners(text):
|
41 |
+
from pypinyin import Style, pinyin
|
42 |
+
text = text.replace("[ZH]", "")
|
43 |
+
phones = [phone[0] for phone in pinyin(text, style=Style.TONE3)]
|
44 |
+
return ' '.join(phones)
|
45 |
+
|
46 |
+
|
47 |
+
def zh_ja_mixture_cleaners(text):
|
48 |
+
text = re.sub(r'\[ZH\](.*?)\[ZH\]',
|
49 |
+
lambda x: chinese_to_romaji(x.group(1))+' ', text)
|
50 |
+
text = re.sub(r'\[JA\](.*?)\[JA\]', lambda x: japanese_to_romaji_with_accent(
|
51 |
+
x.group(1)).replace('ts', 'ʦ').replace('u', 'ɯ').replace('...', '…')+' ', text)
|
52 |
+
text = re.sub(r'\s+$', '', text)
|
53 |
+
text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
54 |
+
return text
|
55 |
+
|
56 |
+
|
57 |
+
def sanskrit_cleaners(text):
|
58 |
+
text = text.replace('॥', '।').replace('ॐ', 'ओम्')
|
59 |
+
text = re.sub(r'([^।])$', r'\1।', text)
|
60 |
+
return text
|
61 |
+
|
62 |
+
|
63 |
+
def cjks_cleaners(text):
|
64 |
+
text = re.sub(r'\[ZH\](.*?)\[ZH\]',
|
65 |
+
lambda x: chinese_to_lazy_ipa(x.group(1))+' ', text)
|
66 |
+
text = re.sub(r'\[JA\](.*?)\[JA\]',
|
67 |
+
lambda x: japanese_to_ipa(x.group(1))+' ', text)
|
68 |
+
text = re.sub(r'\[KO\](.*?)\[KO\]',
|
69 |
+
lambda x: korean_to_lazy_ipa(x.group(1))+' ', text)
|
70 |
+
text = re.sub(r'\[SA\](.*?)\[SA\]',
|
71 |
+
lambda x: devanagari_to_ipa(x.group(1))+' ', text)
|
72 |
+
text = re.sub(r'\[EN\](.*?)\[EN\]',
|
73 |
+
lambda x: english_to_lazy_ipa(x.group(1))+' ', text)
|
74 |
+
text = re.sub(r'\s+$', '', text)
|
75 |
+
text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
76 |
+
return text
|
77 |
+
|
78 |
+
|
79 |
+
def cjke_cleaners(text):
|
80 |
+
text = re.sub(r'\[ZH\](.*?)\[ZH\]', lambda x: chinese_to_lazy_ipa(x.group(1)).replace(
|
81 |
+
'ʧ', 'tʃ').replace('ʦ', 'ts').replace('ɥan', 'ɥæn')+' ', text)
|
82 |
+
text = re.sub(r'\[JA\](.*?)\[JA\]', lambda x: japanese_to_ipa(x.group(1)).replace('ʧ', 'tʃ').replace(
|
83 |
+
'ʦ', 'ts').replace('ɥan', 'ɥæn').replace('ʥ', 'dz')+' ', text)
|
84 |
+
text = re.sub(r'\[KO\](.*?)\[KO\]',
|
85 |
+
lambda x: korean_to_ipa(x.group(1))+' ', text)
|
86 |
+
text = re.sub(r'\[EN\](.*?)\[EN\]', lambda x: english_to_ipa2(x.group(1)).replace('ɑ', 'a').replace(
|
87 |
+
'ɔ', 'o').replace('ɛ', 'e').replace('ɪ', 'i').replace('ʊ', 'u')+' ', text)
|
88 |
+
text = re.sub(r'\s+$', '', text)
|
89 |
+
text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
90 |
+
return text
|
91 |
+
|
92 |
+
|
93 |
+
def cjke_cleaners2(text):
|
94 |
+
text = re.sub(r'\[ZH\](.*?)\[ZH\]',
|
95 |
+
lambda x: chinese_to_ipa(x.group(1))+' ', text)
|
96 |
+
text = re.sub(r'\[JA\](.*?)\[JA\]',
|
97 |
+
lambda x: japanese_to_ipa2(x.group(1))+' ', text)
|
98 |
+
text = re.sub(r'\[KO\](.*?)\[KO\]',
|
99 |
+
lambda x: korean_to_ipa(x.group(1))+' ', text)
|
100 |
+
text = re.sub(r'\[EN\](.*?)\[EN\]',
|
101 |
+
lambda x: english_to_ipa2(x.group(1))+' ', text)
|
102 |
+
text = re.sub(r'\s+$', '', text)
|
103 |
+
text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
104 |
+
return text
|
105 |
+
|
106 |
+
|
107 |
+
def thai_cleaners(text):
|
108 |
+
text = num_to_thai(text)
|
109 |
+
text = latin_to_thai(text)
|
110 |
+
return text
|
111 |
+
|
112 |
+
|
113 |
+
# def shanghainese_cleaners(text):
|
114 |
+
# text = shanghainese_to_ipa(text)
|
115 |
+
# text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
116 |
+
# return text
|
117 |
+
|
118 |
+
|
119 |
+
# def chinese_dialect_cleaners(text):
|
120 |
+
# text = re.sub(r'\[ZH\](.*?)\[ZH\]',
|
121 |
+
# lambda x: chinese_to_ipa2(x.group(1))+' ', text)
|
122 |
+
# text = re.sub(r'\[JA\](.*?)\[JA\]',
|
123 |
+
# lambda x: japanese_to_ipa3(x.group(1)).replace('Q', 'ʔ')+' ', text)
|
124 |
+
# text = re.sub(r'\[SH\](.*?)\[SH\]', lambda x: shanghainese_to_ipa(x.group(1)).replace('1', '˥˧').replace('5',
|
125 |
+
# '˧˧˦').replace('6', '˩˩˧').replace('7', '˥').replace('8', '˩˨').replace('ᴀ', 'ɐ').replace('ᴇ', 'e')+' ', text)
|
126 |
+
# text = re.sub(r'\[GD\](.*?)\[GD\]',
|
127 |
+
# lambda x: cantonese_to_ipa(x.group(1))+' ', text)
|
128 |
+
# text = re.sub(r'\[EN\](.*?)\[EN\]',
|
129 |
+
# lambda x: english_to_lazy_ipa2(x.group(1))+' ', text)
|
130 |
+
# text = re.sub(r'\[([A-Z]{2})\](.*?)\[\1\]', lambda x: ngu_dialect_to_ipa(x.group(2), x.group(
|
131 |
+
# 1)).replace('ʣ', 'dz').replace('ʥ', 'dʑ').replace('ʦ', 'ts').replace('ʨ', 'tɕ')+' ', text)
|
132 |
+
# text = re.sub(r'\s+$', '', text)
|
133 |
+
# text = re.sub(r'([^\.,!\?\-…~])$', r'\1.', text)
|
134 |
+
# return text
|
text/english.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
|
16 |
+
# Regular expression matching whitespace:
|
17 |
+
|
18 |
+
|
19 |
+
import re
|
20 |
+
import inflect
|
21 |
+
from unidecode import unidecode
|
22 |
+
import eng_to_ipa as ipa
|
23 |
+
_inflect = inflect.engine()
|
24 |
+
_comma_number_re = re.compile(r'([0-9][0-9\,]+[0-9])')
|
25 |
+
_decimal_number_re = re.compile(r'([0-9]+\.[0-9]+)')
|
26 |
+
_pounds_re = re.compile(r'£([0-9\,]*[0-9]+)')
|
27 |
+
_dollars_re = re.compile(r'\$([0-9\.\,]*[0-9]+)')
|
28 |
+
_ordinal_re = re.compile(r'[0-9]+(st|nd|rd|th)')
|
29 |
+
_number_re = re.compile(r'[0-9]+')
|
30 |
+
|
31 |
+
# List of (regular expression, replacement) pairs for abbreviations:
|
32 |
+
_abbreviations = [(re.compile('\\b%s\\.' % x[0], re.IGNORECASE), x[1]) for x in [
|
33 |
+
('mrs', 'misess'),
|
34 |
+
('mr', 'mister'),
|
35 |
+
('dr', 'doctor'),
|
36 |
+
('st', 'saint'),
|
37 |
+
('co', 'company'),
|
38 |
+
('jr', 'junior'),
|
39 |
+
('maj', 'major'),
|
40 |
+
('gen', 'general'),
|
41 |
+
('drs', 'doctors'),
|
42 |
+
('rev', 'reverend'),
|
43 |
+
('lt', 'lieutenant'),
|
44 |
+
('hon', 'honorable'),
|
45 |
+
('sgt', 'sergeant'),
|
46 |
+
('capt', 'captain'),
|
47 |
+
('esq', 'esquire'),
|
48 |
+
('ltd', 'limited'),
|
49 |
+
('col', 'colonel'),
|
50 |
+
('ft', 'fort'),
|
51 |
+
]]
|
52 |
+
|
53 |
+
|
54 |
+
# List of (ipa, lazy ipa) pairs:
|
55 |
+
_lazy_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
56 |
+
('r', 'ɹ'),
|
57 |
+
('æ', 'e'),
|
58 |
+
('ɑ', 'a'),
|
59 |
+
('ɔ', 'o'),
|
60 |
+
('ð', 'z'),
|
61 |
+
('θ', 's'),
|
62 |
+
('ɛ', 'e'),
|
63 |
+
('ɪ', 'i'),
|
64 |
+
('ʊ', 'u'),
|
65 |
+
('ʒ', 'ʥ'),
|
66 |
+
('ʤ', 'ʥ'),
|
67 |
+
('ˈ', '↓'),
|
68 |
+
]]
|
69 |
+
|
70 |
+
# List of (ipa, lazy ipa2) pairs:
|
71 |
+
_lazy_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
|
72 |
+
('r', 'ɹ'),
|
73 |
+
('ð', 'z'),
|
74 |
+
('θ', 's'),
|
75 |
+
('ʒ', 'ʑ'),
|
76 |
+
('ʤ', 'dʑ'),
|
77 |
+
('ˈ', '↓'),
|
78 |
+
]]
|
79 |
+
|
80 |
+
# List of (ipa, ipa2) pairs
|
81 |
+
_ipa_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
|
82 |
+
('r', 'ɹ'),
|
83 |
+
('ʤ', 'dʒ'),
|
84 |
+
('ʧ', 'tʃ')
|
85 |
+
]]
|
86 |
+
|
87 |
+
|
88 |
+
def expand_abbreviations(text):
|
89 |
+
for regex, replacement in _abbreviations:
|
90 |
+
text = re.sub(regex, replacement, text)
|
91 |
+
return text
|
92 |
+
|
93 |
+
|
94 |
+
def collapse_whitespace(text):
|
95 |
+
return re.sub(r'\s+', ' ', text)
|
96 |
+
|
97 |
+
|
98 |
+
def _remove_commas(m):
|
99 |
+
return m.group(1).replace(',', '')
|
100 |
+
|
101 |
+
|
102 |
+
def _expand_decimal_point(m):
|
103 |
+
return m.group(1).replace('.', ' point ')
|
104 |
+
|
105 |
+
|
106 |
+
def _expand_dollars(m):
|
107 |
+
match = m.group(1)
|
108 |
+
parts = match.split('.')
|
109 |
+
if len(parts) > 2:
|
110 |
+
return match + ' dollars' # Unexpected format
|
111 |
+
dollars = int(parts[0]) if parts[0] else 0
|
112 |
+
cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
|
113 |
+
if dollars and cents:
|
114 |
+
dollar_unit = 'dollar' if dollars == 1 else 'dollars'
|
115 |
+
cent_unit = 'cent' if cents == 1 else 'cents'
|
116 |
+
return '%s %s, %s %s' % (dollars, dollar_unit, cents, cent_unit)
|
117 |
+
elif dollars:
|
118 |
+
dollar_unit = 'dollar' if dollars == 1 else 'dollars'
|
119 |
+
return '%s %s' % (dollars, dollar_unit)
|
120 |
+
elif cents:
|
121 |
+
cent_unit = 'cent' if cents == 1 else 'cents'
|
122 |
+
return '%s %s' % (cents, cent_unit)
|
123 |
+
else:
|
124 |
+
return 'zero dollars'
|
125 |
+
|
126 |
+
|
127 |
+
def _expand_ordinal(m):
|
128 |
+
return _inflect.number_to_words(m.group(0))
|
129 |
+
|
130 |
+
|
131 |
+
def _expand_number(m):
|
132 |
+
num = int(m.group(0))
|
133 |
+
if num > 1000 and num < 3000:
|
134 |
+
if num == 2000:
|
135 |
+
return 'two thousand'
|
136 |
+
elif num > 2000 and num < 2010:
|
137 |
+
return 'two thousand ' + _inflect.number_to_words(num % 100)
|
138 |
+
elif num % 100 == 0:
|
139 |
+
return _inflect.number_to_words(num // 100) + ' hundred'
|
140 |
+
else:
|
141 |
+
return _inflect.number_to_words(num, andword='', zero='oh', group=2).replace(', ', ' ')
|
142 |
+
else:
|
143 |
+
return _inflect.number_to_words(num, andword='')
|
144 |
+
|
145 |
+
|
146 |
+
def normalize_numbers(text):
|
147 |
+
text = re.sub(_comma_number_re, _remove_commas, text)
|
148 |
+
text = re.sub(_pounds_re, r'\1 pounds', text)
|
149 |
+
text = re.sub(_dollars_re, _expand_dollars, text)
|
150 |
+
text = re.sub(_decimal_number_re, _expand_decimal_point, text)
|
151 |
+
text = re.sub(_ordinal_re, _expand_ordinal, text)
|
152 |
+
text = re.sub(_number_re, _expand_number, text)
|
153 |
+
return text
|
154 |
+
|
155 |
+
|
156 |
+
def mark_dark_l(text):
|
157 |
+
return re.sub(r'l([^aeiouæɑɔəɛɪʊ ]*(?: |$))', lambda x: 'ɫ'+x.group(1), text)
|
158 |
+
|
159 |
+
|
160 |
+
def english_to_ipa(text):
|
161 |
+
text = unidecode(text).lower()
|
162 |
+
text = expand_abbreviations(text)
|
163 |
+
text = normalize_numbers(text)
|
164 |
+
phonemes = ipa.convert(text)
|
165 |
+
phonemes = collapse_whitespace(phonemes)
|
166 |
+
return phonemes
|
167 |
+
|
168 |
+
|
169 |
+
def english_to_lazy_ipa(text):
|
170 |
+
text = english_to_ipa(text)
|
171 |
+
for regex, replacement in _lazy_ipa:
|
172 |
+
text = re.sub(regex, replacement, text)
|
173 |
+
return text
|
174 |
+
|
175 |
+
|
176 |
+
def english_to_ipa2(text):
|
177 |
+
text = english_to_ipa(text)
|
178 |
+
text = mark_dark_l(text)
|
179 |
+
for regex, replacement in _ipa_to_ipa2:
|
180 |
+
text = re.sub(regex, replacement, text)
|
181 |
+
return text.replace('...', '…')
|
182 |
+
|
183 |
+
|
184 |
+
def english_to_lazy_ipa2(text):
|
185 |
+
text = english_to_ipa(text)
|
186 |
+
for regex, replacement in _lazy_ipa2:
|
187 |
+
text = re.sub(regex, replacement, text)
|
188 |
+
return text
|
text/japanese.py
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from unidecode import unidecode
|
3 |
+
import pyopenjtalk
|
4 |
+
|
5 |
+
|
6 |
+
# Regular expression matching Japanese without punctuation marks:
|
7 |
+
_japanese_characters = re.compile(
|
8 |
+
r'[A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
|
9 |
+
|
10 |
+
# Regular expression matching non-Japanese characters or punctuation marks:
|
11 |
+
_japanese_marks = re.compile(
|
12 |
+
r'[^A-Za-z\d\u3005\u3040-\u30ff\u4e00-\u9fff\uff11-\uff19\uff21-\uff3a\uff41-\uff5a\uff66-\uff9d]')
|
13 |
+
|
14 |
+
# List of (symbol, Japanese) pairs for marks:
|
15 |
+
_symbols_to_japanese = [(re.compile('%s' % x[0]), x[1]) for x in [
|
16 |
+
('%', 'パーセント')
|
17 |
+
]]
|
18 |
+
|
19 |
+
# List of (romaji, ipa) pairs for marks:
|
20 |
+
_romaji_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
21 |
+
('ts', 'ʦ'),
|
22 |
+
('u', 'ɯ'),
|
23 |
+
('j', 'ʥ'),
|
24 |
+
('y', 'j'),
|
25 |
+
('ni', 'n^i'),
|
26 |
+
('nj', 'n^'),
|
27 |
+
('hi', 'çi'),
|
28 |
+
('hj', 'ç'),
|
29 |
+
('f', 'ɸ'),
|
30 |
+
('I', 'i*'),
|
31 |
+
('U', 'ɯ*'),
|
32 |
+
('r', 'ɾ')
|
33 |
+
]]
|
34 |
+
|
35 |
+
# List of (romaji, ipa2) pairs for marks:
|
36 |
+
_romaji_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
|
37 |
+
('u', 'ɯ'),
|
38 |
+
('ʧ', 'tʃ'),
|
39 |
+
('j', 'dʑ'),
|
40 |
+
('y', 'j'),
|
41 |
+
('ni', 'n^i'),
|
42 |
+
('nj', 'n^'),
|
43 |
+
('hi', 'çi'),
|
44 |
+
('hj', 'ç'),
|
45 |
+
('f', 'ɸ'),
|
46 |
+
('I', 'i*'),
|
47 |
+
('U', 'ɯ*'),
|
48 |
+
('r', 'ɾ')
|
49 |
+
]]
|
50 |
+
|
51 |
+
# List of (consonant, sokuon) pairs:
|
52 |
+
_real_sokuon = [(re.compile('%s' % x[0]), x[1]) for x in [
|
53 |
+
(r'Q([↑↓]*[kg])', r'k#\1'),
|
54 |
+
(r'Q([↑↓]*[tdjʧ])', r't#\1'),
|
55 |
+
(r'Q([↑↓]*[sʃ])', r's\1'),
|
56 |
+
(r'Q([↑↓]*[pb])', r'p#\1')
|
57 |
+
]]
|
58 |
+
|
59 |
+
# List of (consonant, hatsuon) pairs:
|
60 |
+
_real_hatsuon = [(re.compile('%s' % x[0]), x[1]) for x in [
|
61 |
+
(r'N([↑↓]*[pbm])', r'm\1'),
|
62 |
+
(r'N([↑↓]*[ʧʥj])', r'n^\1'),
|
63 |
+
(r'N([↑↓]*[tdn])', r'n\1'),
|
64 |
+
(r'N([↑↓]*[kg])', r'ŋ\1')
|
65 |
+
]]
|
66 |
+
|
67 |
+
|
68 |
+
def symbols_to_japanese(text):
|
69 |
+
for regex, replacement in _symbols_to_japanese:
|
70 |
+
text = re.sub(regex, replacement, text)
|
71 |
+
return text
|
72 |
+
|
73 |
+
|
74 |
+
def japanese_to_romaji_with_accent(text):
|
75 |
+
'''Reference https://r9y9.github.io/ttslearn/latest/notebooks/ch10_Recipe-Tacotron.html'''
|
76 |
+
text = symbols_to_japanese(text)
|
77 |
+
sentences = re.split(_japanese_marks, text)
|
78 |
+
marks = re.findall(_japanese_marks, text)
|
79 |
+
text = ''
|
80 |
+
for i, sentence in enumerate(sentences):
|
81 |
+
if re.match(_japanese_characters, sentence):
|
82 |
+
if text != '':
|
83 |
+
text += ' '
|
84 |
+
labels = pyopenjtalk.extract_fullcontext(sentence)
|
85 |
+
for n, label in enumerate(labels):
|
86 |
+
phoneme = re.search(r'\-([^\+]*)\+', label).group(1)
|
87 |
+
if phoneme not in ['sil', 'pau']:
|
88 |
+
text += phoneme.replace('ch', 'ʧ').replace('sh',
|
89 |
+
'ʃ').replace('cl', 'Q')
|
90 |
+
else:
|
91 |
+
continue
|
92 |
+
# n_moras = int(re.search(r'/F:(\d+)_', label).group(1))
|
93 |
+
a1 = int(re.search(r"/A:(\-?[0-9]+)\+", label).group(1))
|
94 |
+
a2 = int(re.search(r"\+(\d+)\+", label).group(1))
|
95 |
+
a3 = int(re.search(r"\+(\d+)/", label).group(1))
|
96 |
+
if re.search(r'\-([^\+]*)\+', labels[n + 1]).group(1) in ['sil', 'pau']:
|
97 |
+
a2_next = -1
|
98 |
+
else:
|
99 |
+
a2_next = int(
|
100 |
+
re.search(r"\+(\d+)\+", labels[n + 1]).group(1))
|
101 |
+
# Accent phrase boundary
|
102 |
+
if a3 == 1 and a2_next == 1:
|
103 |
+
text += ' '
|
104 |
+
# Falling
|
105 |
+
elif a1 == 0 and a2_next == a2 + 1:
|
106 |
+
text += '↓'
|
107 |
+
# Rising
|
108 |
+
elif a2 == 1 and a2_next == 2:
|
109 |
+
text += '↑'
|
110 |
+
if i < len(marks):
|
111 |
+
text += unidecode(marks[i]).replace(' ', '')
|
112 |
+
return text
|
113 |
+
|
114 |
+
|
115 |
+
def get_real_sokuon(text):
|
116 |
+
for regex, replacement in _real_sokuon:
|
117 |
+
text = re.sub(regex, replacement, text)
|
118 |
+
return text
|
119 |
+
|
120 |
+
|
121 |
+
def get_real_hatsuon(text):
|
122 |
+
for regex, replacement in _real_hatsuon:
|
123 |
+
text = re.sub(regex, replacement, text)
|
124 |
+
return text
|
125 |
+
|
126 |
+
|
127 |
+
def japanese_to_ipa(text):
|
128 |
+
text = japanese_to_romaji_with_accent(text).replace('...', '…')
|
129 |
+
text = re.sub(
|
130 |
+
r'([aiueo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
|
131 |
+
text = get_real_sokuon(text)
|
132 |
+
text = get_real_hatsuon(text)
|
133 |
+
for regex, replacement in _romaji_to_ipa:
|
134 |
+
text = re.sub(regex, replacement, text)
|
135 |
+
return text
|
136 |
+
|
137 |
+
|
138 |
+
def japanese_to_ipa2(text):
|
139 |
+
text = japanese_to_romaji_with_accent(text).replace('...', '…')
|
140 |
+
text = get_real_sokuon(text)
|
141 |
+
text = get_real_hatsuon(text)
|
142 |
+
for regex, replacement in _romaji_to_ipa2:
|
143 |
+
text = re.sub(regex, replacement, text)
|
144 |
+
return text
|
145 |
+
|
146 |
+
|
147 |
+
def japanese_to_ipa3(text):
|
148 |
+
text = japanese_to_ipa2(text).replace('n^', 'ȵ').replace(
|
149 |
+
'ʃ', 'ɕ').replace('*', '\u0325').replace('#', '\u031a')
|
150 |
+
text = re.sub(
|
151 |
+
r'([aiɯeo])\1+', lambda x: x.group(0)[0]+'ː'*(len(x.group(0))-1), text)
|
152 |
+
text = re.sub(r'((?:^|\s)(?:ts|tɕ|[kpt]))', r'\1ʰ', text)
|
153 |
+
return text
|
text/korean.py
ADDED
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from jamo import h2j, j2hcj
|
3 |
+
import ko_pron
|
4 |
+
|
5 |
+
|
6 |
+
# This is a list of Korean classifiers preceded by pure Korean numerals.
|
7 |
+
_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'
|
8 |
+
|
9 |
+
# List of (hangul, hangul divided) pairs:
|
10 |
+
_hangul_divided = [(re.compile('%s' % x[0]), x[1]) for x in [
|
11 |
+
('ㄳ', 'ㄱㅅ'),
|
12 |
+
('ㄵ', 'ㄴㅈ'),
|
13 |
+
('ㄶ', 'ㄴㅎ'),
|
14 |
+
('ㄺ', 'ㄹㄱ'),
|
15 |
+
('ㄻ', 'ㄹㅁ'),
|
16 |
+
('ㄼ', 'ㄹㅂ'),
|
17 |
+
('ㄽ', 'ㄹㅅ'),
|
18 |
+
('ㄾ', 'ㄹㅌ'),
|
19 |
+
('ㄿ', 'ㄹㅍ'),
|
20 |
+
('ㅀ', 'ㄹㅎ'),
|
21 |
+
('ㅄ', 'ㅂㅅ'),
|
22 |
+
('ㅘ', 'ㅗㅏ'),
|
23 |
+
('ㅙ', 'ㅗㅐ'),
|
24 |
+
('ㅚ', 'ㅗㅣ'),
|
25 |
+
('ㅝ', 'ㅜㅓ'),
|
26 |
+
('ㅞ', 'ㅜㅔ'),
|
27 |
+
('ㅟ', 'ㅜㅣ'),
|
28 |
+
('ㅢ', 'ㅡㅣ'),
|
29 |
+
('ㅑ', 'ㅣㅏ'),
|
30 |
+
('ㅒ', 'ㅣㅐ'),
|
31 |
+
('ㅕ', 'ㅣㅓ'),
|
32 |
+
('ㅖ', 'ㅣㅔ'),
|
33 |
+
('ㅛ', 'ㅣㅗ'),
|
34 |
+
('ㅠ', 'ㅣㅜ')
|
35 |
+
]]
|
36 |
+
|
37 |
+
# List of (Latin alphabet, hangul) pairs:
|
38 |
+
_latin_to_hangul = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
|
39 |
+
('a', '에이'),
|
40 |
+
('b', '비'),
|
41 |
+
('c', '시'),
|
42 |
+
('d', '디'),
|
43 |
+
('e', '이'),
|
44 |
+
('f', '에프'),
|
45 |
+
('g', '지'),
|
46 |
+
('h', '에이치'),
|
47 |
+
('i', '아이'),
|
48 |
+
('j', '제이'),
|
49 |
+
('k', '케이'),
|
50 |
+
('l', '엘'),
|
51 |
+
('m', '엠'),
|
52 |
+
('n', '엔'),
|
53 |
+
('o', '오'),
|
54 |
+
('p', '피'),
|
55 |
+
('q', '큐'),
|
56 |
+
('r', '아르'),
|
57 |
+
('s', '에스'),
|
58 |
+
('t', '티'),
|
59 |
+
('u', '유'),
|
60 |
+
('v', '브이'),
|
61 |
+
('w', '더블유'),
|
62 |
+
('x', '엑스'),
|
63 |
+
('y', '와이'),
|
64 |
+
('z', '제트')
|
65 |
+
]]
|
66 |
+
|
67 |
+
# List of (ipa, lazy ipa) pairs:
|
68 |
+
_ipa_to_lazy_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
|
69 |
+
('t͡ɕ','ʧ'),
|
70 |
+
('d͡ʑ','ʥ'),
|
71 |
+
('ɲ','n^'),
|
72 |
+
('ɕ','ʃ'),
|
73 |
+
('ʷ','w'),
|
74 |
+
('ɭ','l`'),
|
75 |
+
('ʎ','ɾ'),
|
76 |
+
('ɣ','ŋ'),
|
77 |
+
('ɰ','ɯ'),
|
78 |
+
('ʝ','j'),
|
79 |
+
('ʌ','ə'),
|
80 |
+
('ɡ','g'),
|
81 |
+
('\u031a','#'),
|
82 |
+
('\u0348','='),
|
83 |
+
('\u031e',''),
|
84 |
+
('\u0320',''),
|
85 |
+
('\u0339','')
|
86 |
+
]]
|
87 |
+
|
88 |
+
|
89 |
+
def latin_to_hangul(text):
|
90 |
+
for regex, replacement in _latin_to_hangul:
|
91 |
+
text = re.sub(regex, replacement, text)
|
92 |
+
return text
|
93 |
+
|
94 |
+
|
95 |
+
def divide_hangul(text):
|
96 |
+
text = j2hcj(h2j(text))
|
97 |
+
for regex, replacement in _hangul_divided:
|
98 |
+
text = re.sub(regex, replacement, text)
|
99 |
+
return text
|
100 |
+
|
101 |
+
|
102 |
+
def hangul_number(num, sino=True):
|
103 |
+
'''Reference https://github.com/Kyubyong/g2pK'''
|
104 |
+
num = re.sub(',', '', num)
|
105 |
+
|
106 |
+
if num == '0':
|
107 |
+
return '영'
|
108 |
+
if not sino and num == '20':
|
109 |
+
return '스무'
|
110 |
+
|
111 |
+
digits = '123456789'
|
112 |
+
names = '일이삼사오육칠팔구'
|
113 |
+
digit2name = {d: n for d, n in zip(digits, names)}
|
114 |
+
|
115 |
+
modifiers = '한 두 세 네 다섯 여섯 일곱 여덟 아홉'
|
116 |
+
decimals = '열 스물 서른 마흔 쉰 예순 일흔 여든 아흔'
|
117 |
+
digit2mod = {d: mod for d, mod in zip(digits, modifiers.split())}
|
118 |
+
digit2dec = {d: dec for d, dec in zip(digits, decimals.split())}
|
119 |
+
|
120 |
+
spelledout = []
|
121 |
+
for i, digit in enumerate(num):
|
122 |
+
i = len(num) - i - 1
|
123 |
+
if sino:
|
124 |
+
if i == 0:
|
125 |
+
name = digit2name.get(digit, '')
|
126 |
+
elif i == 1:
|
127 |
+
name = digit2name.get(digit, '') + '십'
|
128 |
+
name = name.replace('일십', '십')
|
129 |
+
else:
|
130 |
+
if i == 0:
|
131 |
+
name = digit2mod.get(digit, '')
|
132 |
+
elif i == 1:
|
133 |
+
name = digit2dec.get(digit, '')
|
134 |
+
if digit == '0':
|
135 |
+
if i % 4 == 0:
|
136 |
+
last_three = spelledout[-min(3, len(spelledout)):]
|
137 |
+
if ''.join(last_three) == '':
|
138 |
+
spelledout.append('')
|
139 |
+
continue
|
140 |
+
else:
|
141 |
+
spelledout.append('')
|
142 |
+
continue
|
143 |
+
if i == 2:
|
144 |
+
name = digit2name.get(digit, '') + '백'
|
145 |
+
name = name.replace('일백', '백')
|
146 |
+
elif i == 3:
|
147 |
+
name = digit2name.get(digit, '') + '천'
|
148 |
+
name = name.replace('일천', '천')
|
149 |
+
elif i == 4:
|
150 |
+
name = digit2name.get(digit, '') + '만'
|
151 |
+
name = name.replace('일만', '만')
|
152 |
+
elif i == 5:
|
153 |
+
name = digit2name.get(digit, '') + '십'
|
154 |
+
name = name.replace('일십', '십')
|
155 |
+
elif i == 6:
|
156 |
+
name = digit2name.get(digit, '') + '백'
|
157 |
+
name = name.replace('일백', '백')
|
158 |
+
elif i == 7:
|
159 |
+
name = digit2name.get(digit, '') + '천'
|
160 |
+
name = name.replace('일천', '천')
|
161 |
+
elif i == 8:
|
162 |
+
name = digit2name.get(digit, '') + '억'
|
163 |
+
elif i == 9:
|
164 |
+
name = digit2name.get(digit, '') + '십'
|
165 |
+
elif i == 10:
|
166 |
+
name = digit2name.get(digit, '') + '백'
|
167 |
+
elif i == 11:
|
168 |
+
name = digit2name.get(digit, '') + '천'
|
169 |
+
elif i == 12:
|
170 |
+
name = digit2name.get(digit, '') + '조'
|
171 |
+
elif i == 13:
|
172 |
+
name = digit2name.get(digit, '') + '십'
|
173 |
+
elif i == 14:
|
174 |
+
name = digit2name.get(digit, '') + '백'
|
175 |
+
elif i == 15:
|
176 |
+
name = digit2name.get(digit, '') + '천'
|
177 |
+
spelledout.append(name)
|
178 |
+
return ''.join(elem for elem in spelledout)
|
179 |
+
|
180 |
+
|
181 |
+
def number_to_hangul(text):
|
182 |
+
'''Reference https://github.com/Kyubyong/g2pK'''
|
183 |
+
tokens = set(re.findall(r'(\d[\d,]*)([\uac00-\ud71f]+)', text))
|
184 |
+
for token in tokens:
|
185 |
+
num, classifier = token
|
186 |
+
if classifier[:2] in _korean_classifiers or classifier[0] in _korean_classifiers:
|
187 |
+
spelledout = hangul_number(num, sino=False)
|
188 |
+
else:
|
189 |
+
spelledout = hangul_number(num, sino=True)
|
190 |
+
text = text.replace(f'{num}{classifier}', f'{spelledout}{classifier}')
|
191 |
+
# digit by digit for remaining digits
|
192 |
+
digits = '0123456789'
|
193 |
+
names = '영일이삼사오육칠팔구'
|
194 |
+
for d, n in zip(digits, names):
|
195 |
+
text = text.replace(d, n)
|
196 |
+
return text
|
197 |
+
|
198 |
+
|
199 |
+
def korean_to_lazy_ipa(text):
|
200 |
+
text = latin_to_hangul(text)
|
201 |
+
text = number_to_hangul(text)
|
202 |
+
text=re.sub('[\uac00-\ud7af]+',lambda x:ko_pron.romanise(x.group(0),'ipa').split('] ~ [')[0],text)
|
203 |
+
for regex, replacement in _ipa_to_lazy_ipa:
|
204 |
+
text = re.sub(regex, replacement, text)
|
205 |
+
return text
|
206 |
+
|
207 |
+
|
208 |
+
def korean_to_ipa(text):
|
209 |
+
text = korean_to_lazy_ipa(text)
|
210 |
+
return text.replace('ʧ','tʃ').replace('ʥ','dʑ')
|
text/mandarin.py
ADDED
@@ -0,0 +1,326 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import re
|
4 |
+
from pypinyin import lazy_pinyin, BOPOMOFO
|
5 |
+
import jieba
|
6 |
+
import cn2an
|
7 |
+
import logging
|
8 |
+
|
9 |
+
|
10 |
+
# List of (Latin alphabet, bopomofo) pairs:
|
11 |
+
_latin_to_bopomofo = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
|
12 |
+
('a', 'ㄟˉ'),
|
13 |
+
('b', 'ㄅㄧˋ'),
|
14 |
+
('c', 'ㄙㄧˉ'),
|
15 |
+
('d', 'ㄉㄧˋ'),
|
16 |
+
('e', 'ㄧˋ'),
|
17 |
+
('f', 'ㄝˊㄈㄨˋ'),
|
18 |
+
('g', 'ㄐㄧˋ'),
|
19 |
+
('h', 'ㄝˇㄑㄩˋ'),
|
20 |
+
('i', 'ㄞˋ'),
|
21 |
+
('j', 'ㄐㄟˋ'),
|
22 |
+
('k', 'ㄎㄟˋ'),
|
23 |
+
('l', 'ㄝˊㄛˋ'),
|
24 |
+
('m', 'ㄝˊㄇㄨˋ'),
|
25 |
+
('n', 'ㄣˉ'),
|
26 |
+
('o', 'ㄡˉ'),
|
27 |
+
('p', 'ㄆㄧˉ'),
|
28 |
+
('q', 'ㄎㄧㄡˉ'),
|
29 |
+
('r', 'ㄚˋ'),
|
30 |
+
('s', 'ㄝˊㄙˋ'),
|
31 |
+
('t', 'ㄊㄧˋ'),
|
32 |
+
('u', 'ㄧㄡˉ'),
|
33 |
+
('v', 'ㄨㄧˉ'),
|
34 |
+
('w', 'ㄉㄚˋㄅㄨˋㄌㄧㄡˋ'),
|
35 |
+
('x', 'ㄝˉㄎㄨˋㄙˋ'),
|
36 |
+
('y', 'ㄨㄞˋ'),
|
37 |
+
('z', 'ㄗㄟˋ')
|
38 |
+
]]
|
39 |
+
|
40 |
+
# List of (bopomofo, romaji) pairs:
|
41 |
+
_bopomofo_to_romaji = [(re.compile('%s' % x[0]), x[1]) for x in [
|
42 |
+
('ㄅㄛ', 'p⁼wo'),
|
43 |
+
('ㄆㄛ', 'pʰwo'),
|
44 |
+
('ㄇㄛ', 'mwo'),
|
45 |
+
('ㄈㄛ', 'fwo'),
|
46 |
+
('ㄅ', 'p⁼'),
|
47 |
+
('ㄆ', 'pʰ'),
|
48 |
+
('ㄇ', 'm'),
|
49 |
+
('ㄈ', 'f'),
|
50 |
+
('ㄉ', 't⁼'),
|
51 |
+
('ㄊ', 'tʰ'),
|
52 |
+
('ㄋ', 'n'),
|
53 |
+
('ㄌ', 'l'),
|
54 |
+
('ㄍ', 'k⁼'),
|
55 |
+
('ㄎ', 'kʰ'),
|
56 |
+
('ㄏ', 'h'),
|
57 |
+
('ㄐ', 'ʧ⁼'),
|
58 |
+
('ㄑ', 'ʧʰ'),
|
59 |
+
('ㄒ', 'ʃ'),
|
60 |
+
('ㄓ', 'ʦ`⁼'),
|
61 |
+
('ㄔ', 'ʦ`ʰ'),
|
62 |
+
('ㄕ', 's`'),
|
63 |
+
('ㄖ', 'ɹ`'),
|
64 |
+
('ㄗ', 'ʦ⁼'),
|
65 |
+
('ㄘ', 'ʦʰ'),
|
66 |
+
('ㄙ', 's'),
|
67 |
+
('ㄚ', 'a'),
|
68 |
+
('ㄛ', 'o'),
|
69 |
+
('ㄜ', 'ə'),
|
70 |
+
('ㄝ', 'e'),
|
71 |
+
('ㄞ', 'ai'),
|
72 |
+
('ㄟ', 'ei'),
|
73 |
+
('ㄠ', 'au'),
|
74 |
+
('ㄡ', 'ou'),
|
75 |
+
('ㄧㄢ', 'yeNN'),
|
76 |
+
('ㄢ', 'aNN'),
|
77 |
+
('ㄧㄣ', 'iNN'),
|
78 |
+
('ㄣ', 'əNN'),
|
79 |
+
('ㄤ', 'aNg'),
|
80 |
+
('ㄧㄥ', 'iNg'),
|
81 |
+
('ㄨㄥ', 'uNg'),
|
82 |
+
('ㄩㄥ', 'yuNg'),
|
83 |
+
('ㄥ', 'əNg'),
|
84 |
+
('ㄦ', 'əɻ'),
|
85 |
+
('ㄧ', 'i'),
|
86 |
+
('ㄨ', 'u'),
|
87 |
+
('ㄩ', 'ɥ'),
|
88 |
+
('ˉ', '→'),
|
89 |
+
('ˊ', '↑'),
|
90 |
+
('ˇ', '↓↑'),
|
91 |
+
('ˋ', '↓'),
|
92 |
+
('˙', ''),
|
93 |
+
(',', ','),
|
94 |
+
('。', '.'),
|
95 |
+
('!', '!'),
|
96 |
+
('?', '?'),
|
97 |
+
('—', '-')
|
98 |
+
]]
|
99 |
+
|
100 |
+
# List of (romaji, ipa) pairs:
|
101 |
+
_romaji_to_ipa = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
|
102 |
+
('ʃy', 'ʃ'),
|
103 |
+
('ʧʰy', 'ʧʰ'),
|
104 |
+
('ʧ⁼y', 'ʧ⁼'),
|
105 |
+
('NN', 'n'),
|
106 |
+
('Ng', 'ŋ'),
|
107 |
+
('y', 'j'),
|
108 |
+
('h', 'x')
|
109 |
+
]]
|
110 |
+
|
111 |
+
# List of (bopomofo, ipa) pairs:
|
112 |
+
_bopomofo_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
113 |
+
('ㄅㄛ', 'p⁼wo'),
|
114 |
+
('ㄆㄛ', 'pʰwo'),
|
115 |
+
('ㄇㄛ', 'mwo'),
|
116 |
+
('ㄈㄛ', 'fwo'),
|
117 |
+
('ㄅ', 'p⁼'),
|
118 |
+
('ㄆ', 'pʰ'),
|
119 |
+
('ㄇ', 'm'),
|
120 |
+
('ㄈ', 'f'),
|
121 |
+
('ㄉ', 't⁼'),
|
122 |
+
('ㄊ', 'tʰ'),
|
123 |
+
('ㄋ', 'n'),
|
124 |
+
('ㄌ', 'l'),
|
125 |
+
('ㄍ', 'k⁼'),
|
126 |
+
('ㄎ', 'kʰ'),
|
127 |
+
('ㄏ', 'x'),
|
128 |
+
('ㄐ', 'tʃ⁼'),
|
129 |
+
('ㄑ', 'tʃʰ'),
|
130 |
+
('ㄒ', 'ʃ'),
|
131 |
+
('ㄓ', 'ts`⁼'),
|
132 |
+
('ㄔ', 'ts`ʰ'),
|
133 |
+
('ㄕ', 's`'),
|
134 |
+
('ㄖ', 'ɹ`'),
|
135 |
+
('ㄗ', 'ts⁼'),
|
136 |
+
('ㄘ', 'tsʰ'),
|
137 |
+
('ㄙ', 's'),
|
138 |
+
('ㄚ', 'a'),
|
139 |
+
('ㄛ', 'o'),
|
140 |
+
('ㄜ', 'ə'),
|
141 |
+
('ㄝ', 'ɛ'),
|
142 |
+
('ㄞ', 'aɪ'),
|
143 |
+
('ㄟ', 'eɪ'),
|
144 |
+
('ㄠ', 'ɑʊ'),
|
145 |
+
('ㄡ', 'oʊ'),
|
146 |
+
('ㄧㄢ', 'jɛn'),
|
147 |
+
('ㄩㄢ', 'ɥæn'),
|
148 |
+
('ㄢ', 'an'),
|
149 |
+
('ㄧㄣ', 'in'),
|
150 |
+
('ㄩㄣ', 'ɥn'),
|
151 |
+
('ㄣ', 'ən'),
|
152 |
+
('ㄤ', 'ɑŋ'),
|
153 |
+
('ㄧㄥ', 'iŋ'),
|
154 |
+
('ㄨㄥ', 'ʊŋ'),
|
155 |
+
('ㄩㄥ', 'jʊŋ'),
|
156 |
+
('ㄥ', 'əŋ'),
|
157 |
+
('ㄦ', 'əɻ'),
|
158 |
+
('ㄧ', 'i'),
|
159 |
+
('ㄨ', 'u'),
|
160 |
+
('ㄩ', 'ɥ'),
|
161 |
+
('ˉ', '→'),
|
162 |
+
('ˊ', '↑'),
|
163 |
+
('ˇ', '↓↑'),
|
164 |
+
('ˋ', '↓'),
|
165 |
+
('˙', ''),
|
166 |
+
(',', ','),
|
167 |
+
('。', '.'),
|
168 |
+
('!', '!'),
|
169 |
+
('?', '?'),
|
170 |
+
('—', '-')
|
171 |
+
]]
|
172 |
+
|
173 |
+
# List of (bopomofo, ipa2) pairs:
|
174 |
+
_bopomofo_to_ipa2 = [(re.compile('%s' % x[0]), x[1]) for x in [
|
175 |
+
('ㄅㄛ', 'pwo'),
|
176 |
+
('ㄆㄛ', 'pʰwo'),
|
177 |
+
('ㄇㄛ', 'mwo'),
|
178 |
+
('ㄈㄛ', 'fwo'),
|
179 |
+
('ㄅ', 'p'),
|
180 |
+
('ㄆ', 'pʰ'),
|
181 |
+
('ㄇ', 'm'),
|
182 |
+
('ㄈ', 'f'),
|
183 |
+
('ㄉ', 't'),
|
184 |
+
('ㄊ', 'tʰ'),
|
185 |
+
('ㄋ', 'n'),
|
186 |
+
('ㄌ', 'l'),
|
187 |
+
('ㄍ', 'k'),
|
188 |
+
('ㄎ', 'kʰ'),
|
189 |
+
('ㄏ', 'h'),
|
190 |
+
('ㄐ', 'tɕ'),
|
191 |
+
('ㄑ', 'tɕʰ'),
|
192 |
+
('ㄒ', 'ɕ'),
|
193 |
+
('ㄓ', 'tʂ'),
|
194 |
+
('ㄔ', 'tʂʰ'),
|
195 |
+
('ㄕ', 'ʂ'),
|
196 |
+
('ㄖ', 'ɻ'),
|
197 |
+
('ㄗ', 'ts'),
|
198 |
+
('ㄘ', 'tsʰ'),
|
199 |
+
('ㄙ', 's'),
|
200 |
+
('ㄚ', 'a'),
|
201 |
+
('ㄛ', 'o'),
|
202 |
+
('ㄜ', 'ɤ'),
|
203 |
+
('ㄝ', 'ɛ'),
|
204 |
+
('ㄞ', 'aɪ'),
|
205 |
+
('ㄟ', 'eɪ'),
|
206 |
+
('ㄠ', 'ɑʊ'),
|
207 |
+
('ㄡ', 'oʊ'),
|
208 |
+
('ㄧㄢ', 'jɛn'),
|
209 |
+
('ㄩㄢ', 'yæn'),
|
210 |
+
('ㄢ', 'an'),
|
211 |
+
('ㄧㄣ', 'in'),
|
212 |
+
('ㄩㄣ', 'yn'),
|
213 |
+
('ㄣ', 'ən'),
|
214 |
+
('ㄤ', 'ɑŋ'),
|
215 |
+
('ㄧㄥ', 'iŋ'),
|
216 |
+
('ㄨㄥ', 'ʊŋ'),
|
217 |
+
('ㄩㄥ', 'jʊŋ'),
|
218 |
+
('ㄥ', 'ɤŋ'),
|
219 |
+
('ㄦ', 'əɻ'),
|
220 |
+
('ㄧ', 'i'),
|
221 |
+
('ㄨ', 'u'),
|
222 |
+
('ㄩ', 'y'),
|
223 |
+
('ˉ', '˥'),
|
224 |
+
('ˊ', '˧˥'),
|
225 |
+
('ˇ', '˨˩˦'),
|
226 |
+
('ˋ', '˥˩'),
|
227 |
+
('˙', ''),
|
228 |
+
(',', ','),
|
229 |
+
('。', '.'),
|
230 |
+
('!', '!'),
|
231 |
+
('?', '?'),
|
232 |
+
('—', '-')
|
233 |
+
]]
|
234 |
+
|
235 |
+
|
236 |
+
def number_to_chinese(text):
|
237 |
+
numbers = re.findall(r'\d+(?:\.?\d+)?', text)
|
238 |
+
for number in numbers:
|
239 |
+
text = text.replace(number, cn2an.an2cn(number), 1)
|
240 |
+
return text
|
241 |
+
|
242 |
+
|
243 |
+
def chinese_to_bopomofo(text):
|
244 |
+
text = text.replace('、', ',').replace(';', ',').replace(':', ',')
|
245 |
+
words = jieba.lcut(text, cut_all=False)
|
246 |
+
text = ''
|
247 |
+
for word in words:
|
248 |
+
bopomofos = lazy_pinyin(word, BOPOMOFO)
|
249 |
+
if not re.search('[\u4e00-\u9fff]', word):
|
250 |
+
text += word
|
251 |
+
continue
|
252 |
+
for i in range(len(bopomofos)):
|
253 |
+
bopomofos[i] = re.sub(r'([\u3105-\u3129])$', r'\1ˉ', bopomofos[i])
|
254 |
+
if text != '':
|
255 |
+
text += ' '
|
256 |
+
text += ''.join(bopomofos)
|
257 |
+
return text
|
258 |
+
|
259 |
+
|
260 |
+
def latin_to_bopomofo(text):
|
261 |
+
for regex, replacement in _latin_to_bopomofo:
|
262 |
+
text = re.sub(regex, replacement, text)
|
263 |
+
return text
|
264 |
+
|
265 |
+
|
266 |
+
def bopomofo_to_romaji(text):
|
267 |
+
for regex, replacement in _bopomofo_to_romaji:
|
268 |
+
text = re.sub(regex, replacement, text)
|
269 |
+
return text
|
270 |
+
|
271 |
+
|
272 |
+
def bopomofo_to_ipa(text):
|
273 |
+
for regex, replacement in _bopomofo_to_ipa:
|
274 |
+
text = re.sub(regex, replacement, text)
|
275 |
+
return text
|
276 |
+
|
277 |
+
|
278 |
+
def bopomofo_to_ipa2(text):
|
279 |
+
for regex, replacement in _bopomofo_to_ipa2:
|
280 |
+
text = re.sub(regex, replacement, text)
|
281 |
+
return text
|
282 |
+
|
283 |
+
|
284 |
+
def chinese_to_romaji(text):
|
285 |
+
text = number_to_chinese(text)
|
286 |
+
text = chinese_to_bopomofo(text)
|
287 |
+
text = latin_to_bopomofo(text)
|
288 |
+
text = bopomofo_to_romaji(text)
|
289 |
+
text = re.sub('i([aoe])', r'y\1', text)
|
290 |
+
text = re.sub('u([aoəe])', r'w\1', text)
|
291 |
+
text = re.sub('([ʦsɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
|
292 |
+
r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
|
293 |
+
text = re.sub('([ʦs][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
|
294 |
+
return text
|
295 |
+
|
296 |
+
|
297 |
+
def chinese_to_lazy_ipa(text):
|
298 |
+
text = chinese_to_romaji(text)
|
299 |
+
for regex, replacement in _romaji_to_ipa:
|
300 |
+
text = re.sub(regex, replacement, text)
|
301 |
+
return text
|
302 |
+
|
303 |
+
|
304 |
+
def chinese_to_ipa(text):
|
305 |
+
text = number_to_chinese(text)
|
306 |
+
text = chinese_to_bopomofo(text)
|
307 |
+
text = latin_to_bopomofo(text)
|
308 |
+
text = bopomofo_to_ipa(text)
|
309 |
+
text = re.sub('i([aoe])', r'j\1', text)
|
310 |
+
text = re.sub('u([aoəe])', r'w\1', text)
|
311 |
+
text = re.sub('([sɹ]`[⁼ʰ]?)([→↓↑ ]+|$)',
|
312 |
+
r'\1ɹ`\2', text).replace('ɻ', 'ɹ`')
|
313 |
+
text = re.sub('([s][⁼ʰ]?)([→↓↑ ]+|$)', r'\1ɹ\2', text)
|
314 |
+
return text
|
315 |
+
|
316 |
+
|
317 |
+
def chinese_to_ipa2(text):
|
318 |
+
text = number_to_chinese(text)
|
319 |
+
text = chinese_to_bopomofo(text)
|
320 |
+
text = latin_to_bopomofo(text)
|
321 |
+
text = bopomofo_to_ipa2(text)
|
322 |
+
text = re.sub(r'i([aoe])', r'j\1', text)
|
323 |
+
text = re.sub(r'u([aoəe])', r'w\1', text)
|
324 |
+
text = re.sub(r'([ʂɹ]ʰ?)([˩˨˧˦˥ ]+|$)', r'\1ʅ\2', text)
|
325 |
+
text = re.sub(r'(sʰ?)([˩˨˧˦˥ ]+|$)', r'\1ɿ\2', text)
|
326 |
+
return text
|
text/ngu_dialect.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import opencc
|
3 |
+
|
4 |
+
|
5 |
+
dialects = {'SZ': 'suzhou', 'WX': 'wuxi', 'CZ': 'changzhou', 'HZ': 'hangzhou',
|
6 |
+
'SX': 'shaoxing', 'NB': 'ningbo', 'JJ': 'jingjiang', 'YX': 'yixing',
|
7 |
+
'JD': 'jiading', 'ZR': 'zhenru', 'PH': 'pinghu', 'TX': 'tongxiang',
|
8 |
+
'JS': 'jiashan', 'HN': 'xiashi', 'LP': 'linping', 'XS': 'xiaoshan',
|
9 |
+
'FY': 'fuyang', 'RA': 'ruao', 'CX': 'cixi', 'SM': 'sanmen',
|
10 |
+
'TT': 'tiantai', 'WZ': 'wenzhou', 'SC': 'suichang', 'YB': 'youbu'}
|
11 |
+
|
12 |
+
converters = {}
|
13 |
+
|
14 |
+
for dialect in dialects.values():
|
15 |
+
try:
|
16 |
+
converters[dialect] = opencc.OpenCC(dialect)
|
17 |
+
except:
|
18 |
+
pass
|
19 |
+
|
20 |
+
|
21 |
+
def ngu_dialect_to_ipa(text, dialect):
|
22 |
+
dialect = dialects[dialect]
|
23 |
+
text = converters[dialect].convert(text).replace('-','').replace('$',' ')
|
24 |
+
text = re.sub(r'[、;:]', ',', text)
|
25 |
+
text = re.sub(r'\s*,\s*', ', ', text)
|
26 |
+
text = re.sub(r'\s*。\s*', '. ', text)
|
27 |
+
text = re.sub(r'\s*?\s*', '? ', text)
|
28 |
+
text = re.sub(r'\s*!\s*', '! ', text)
|
29 |
+
text = re.sub(r'\s*$', '', text)
|
30 |
+
return text
|
text/sanskrit.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from indic_transliteration import sanscript
|
3 |
+
|
4 |
+
|
5 |
+
# List of (iast, ipa) pairs:
|
6 |
+
_iast_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
7 |
+
('a', 'ə'),
|
8 |
+
('ā', 'aː'),
|
9 |
+
('ī', 'iː'),
|
10 |
+
('ū', 'uː'),
|
11 |
+
('ṛ', 'ɹ`'),
|
12 |
+
('ṝ', 'ɹ`ː'),
|
13 |
+
('ḷ', 'l`'),
|
14 |
+
('ḹ', 'l`ː'),
|
15 |
+
('e', 'eː'),
|
16 |
+
('o', 'oː'),
|
17 |
+
('k', 'k⁼'),
|
18 |
+
('k⁼h', 'kʰ'),
|
19 |
+
('g', 'g⁼'),
|
20 |
+
('g⁼h', 'gʰ'),
|
21 |
+
('ṅ', 'ŋ'),
|
22 |
+
('c', 'ʧ⁼'),
|
23 |
+
('ʧ⁼h', 'ʧʰ'),
|
24 |
+
('j', 'ʥ⁼'),
|
25 |
+
('ʥ⁼h', 'ʥʰ'),
|
26 |
+
('ñ', 'n^'),
|
27 |
+
('ṭ', 't`⁼'),
|
28 |
+
('t`⁼h', 't`ʰ'),
|
29 |
+
('ḍ', 'd`⁼'),
|
30 |
+
('d`⁼h', 'd`ʰ'),
|
31 |
+
('ṇ', 'n`'),
|
32 |
+
('t', 't⁼'),
|
33 |
+
('t⁼h', 'tʰ'),
|
34 |
+
('d', 'd⁼'),
|
35 |
+
('d⁼h', 'dʰ'),
|
36 |
+
('p', 'p⁼'),
|
37 |
+
('p⁼h', 'pʰ'),
|
38 |
+
('b', 'b⁼'),
|
39 |
+
('b⁼h', 'bʰ'),
|
40 |
+
('y', 'j'),
|
41 |
+
('ś', 'ʃ'),
|
42 |
+
('ṣ', 's`'),
|
43 |
+
('r', 'ɾ'),
|
44 |
+
('l̤', 'l`'),
|
45 |
+
('h', 'ɦ'),
|
46 |
+
("'", ''),
|
47 |
+
('~', '^'),
|
48 |
+
('ṃ', '^')
|
49 |
+
]]
|
50 |
+
|
51 |
+
|
52 |
+
def devanagari_to_ipa(text):
|
53 |
+
text = text.replace('ॐ', 'ओम्')
|
54 |
+
text = re.sub(r'\s*।\s*$', '.', text)
|
55 |
+
text = re.sub(r'\s*।\s*', ', ', text)
|
56 |
+
text = re.sub(r'\s*॥', '.', text)
|
57 |
+
text = sanscript.transliterate(text, sanscript.DEVANAGARI, sanscript.IAST)
|
58 |
+
for regex, replacement in _iast_to_ipa:
|
59 |
+
text = re.sub(regex, replacement, text)
|
60 |
+
text = re.sub('(.)[`ː]*ḥ', lambda x: x.group(0)
|
61 |
+
[:-1]+'h'+x.group(1)+'*', text)
|
62 |
+
return text
|
text/shanghainese.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import cn2an
|
3 |
+
import opencc
|
4 |
+
|
5 |
+
|
6 |
+
converter = opencc.OpenCC('zaonhe')
|
7 |
+
|
8 |
+
# List of (Latin alphabet, ipa) pairs:
|
9 |
+
_latin_to_ipa = [(re.compile('%s' % x[0]), x[1]) for x in [
|
10 |
+
('A', 'ᴇ'),
|
11 |
+
('B', 'bi'),
|
12 |
+
('C', 'si'),
|
13 |
+
('D', 'di'),
|
14 |
+
('E', 'i'),
|
15 |
+
('F', 'ᴇf'),
|
16 |
+
('G', 'dʑi'),
|
17 |
+
('H', 'ᴇtɕʰ'),
|
18 |
+
('I', 'ᴀi'),
|
19 |
+
('J', 'dʑᴇ'),
|
20 |
+
('K', 'kʰᴇ'),
|
21 |
+
('L', 'ᴇl'),
|
22 |
+
('M', 'ᴇm'),
|
23 |
+
('N', 'ᴇn'),
|
24 |
+
('O', 'o'),
|
25 |
+
('P', 'pʰi'),
|
26 |
+
('Q', 'kʰiu'),
|
27 |
+
('R', 'ᴀl'),
|
28 |
+
('S', 'ᴇs'),
|
29 |
+
('T', 'tʰi'),
|
30 |
+
('U', 'ɦiu'),
|
31 |
+
('V', 'vi'),
|
32 |
+
('W', 'dᴀbɤliu'),
|
33 |
+
('X', 'ᴇks'),
|
34 |
+
('Y', 'uᴀi'),
|
35 |
+
('Z', 'zᴇ')
|
36 |
+
]]
|
37 |
+
|
38 |
+
|
39 |
+
def _number_to_shanghainese(num):
|
40 |
+
num = cn2an.an2cn(num).replace('一十','十').replace('二十', '廿').replace('二', '两')
|
41 |
+
return re.sub(r'((?:^|[^三四五六七八九])十|廿)两', r'\1二', num)
|
42 |
+
|
43 |
+
|
44 |
+
def number_to_shanghainese(text):
|
45 |
+
return re.sub(r'\d+(?:\.?\d+)?', lambda x: _number_to_shanghainese(x.group()), text)
|
46 |
+
|
47 |
+
|
48 |
+
def latin_to_ipa(text):
|
49 |
+
for regex, replacement in _latin_to_ipa:
|
50 |
+
text = re.sub(regex, replacement, text)
|
51 |
+
return text
|
52 |
+
|
53 |
+
|
54 |
+
def shanghainese_to_ipa(text):
|
55 |
+
text = number_to_shanghainese(text.upper())
|
56 |
+
text = converter.convert(text).replace('-','').replace('$',' ')
|
57 |
+
text = re.sub(r'[A-Z]', lambda x: latin_to_ipa(x.group())+' ', text)
|
58 |
+
text = re.sub(r'[、;:]', ',', text)
|
59 |
+
text = re.sub(r'\s*,\s*', ', ', text)
|
60 |
+
text = re.sub(r'\s*。\s*', '. ', text)
|
61 |
+
text = re.sub(r'\s*?\s*', '? ', text)
|
62 |
+
text = re.sub(r'\s*!\s*', '! ', text)
|
63 |
+
text = re.sub(r'\s*$', '', text)
|
64 |
+
return text
|
text/symbols.py
ADDED
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
|
18 |
+
'''# korean_cleaners
|
19 |
+
_pad = '_'
|
20 |
+
_punctuation = ',.!?…~'
|
21 |
+
_letters = 'ㄱㄴㄷㄹㅁㅂㅅㅇㅈㅊㅋㅌㅍㅎㄲㄸㅃㅆㅉㅏㅓㅗㅜㅡㅣㅐㅔ '
|
22 |
+
'''
|
23 |
+
|
24 |
+
'''# chinese_cleaners
|
25 |
+
_pad = '_'
|
26 |
+
_punctuation = ',。!?—…'
|
27 |
+
_letters = 'ㄅㄆㄇㄈㄉㄊㄋㄌㄍㄎㄏㄐㄑㄒㄓㄔㄕㄖㄗㄘㄙㄚㄛㄜㄝㄞㄟㄠㄡㄢㄣㄤㄥㄦㄧㄨㄩˉˊˇˋ˙ '
|
28 |
+
'''
|
29 |
+
|
30 |
+
# # zh_ja_mixture_cleaners
|
31 |
+
# _pad = '_'
|
32 |
+
# _punctuation = ',.!?-~…'
|
33 |
+
# _letters = 'AEINOQUabdefghijklmnoprstuvwyzʃʧʦɯɹəɥ⁼ʰ`→↓↑ '
|
34 |
+
|
35 |
+
|
36 |
+
'''# sanskrit_cleaners
|
37 |
+
_pad = '_'
|
38 |
+
_punctuation = '।'
|
39 |
+
_letters = 'ँंःअआइईउऊऋएऐओऔकखगघङचछजझञटठडढणतथदधनपफबभमयरलळवशषसहऽािीुूृॄेैोौ्ॠॢ '
|
40 |
+
'''
|
41 |
+
|
42 |
+
'''# cjks_cleaners
|
43 |
+
_pad = '_'
|
44 |
+
_punctuation = ',.!?-~…'
|
45 |
+
_letters = 'NQabdefghijklmnopstuvwxyzʃʧʥʦɯɹəɥçɸɾβŋɦː⁼ʰ`^#*=→↓↑ '
|
46 |
+
'''
|
47 |
+
|
48 |
+
'''# thai_cleaners
|
49 |
+
_pad = '_'
|
50 |
+
_punctuation = '.!? '
|
51 |
+
_letters = 'กขฃคฆงจฉชซฌญฎฏฐฑฒณดตถทธนบปผฝพฟภมยรฤลวศษสหฬอฮฯะัาำิีึืุูเแโใไๅๆ็่้๊๋์'
|
52 |
+
'''
|
53 |
+
|
54 |
+
# # cjke_cleaners2
|
55 |
+
_pad = '_'
|
56 |
+
_punctuation = ',.!?-~…'
|
57 |
+
_letters = 'NQabdefghijklmnopstuvwxyzɑæʃʑçɯɪɔɛɹðəɫɥɸʊɾʒθβŋɦ⁼ʰ`^#*=ˈˌ→↓↑ '
|
58 |
+
|
59 |
+
|
60 |
+
'''# shanghainese_cleaners
|
61 |
+
_pad = '_'
|
62 |
+
_punctuation = ',.!?…'
|
63 |
+
_letters = 'abdfghiklmnopstuvyzøŋȵɑɔɕəɤɦɪɿʑʔʰ̩̃ᴀᴇ15678 '
|
64 |
+
'''
|
65 |
+
|
66 |
+
'''# chinese_dialect_cleaners
|
67 |
+
_pad = '_'
|
68 |
+
_punctuation = ',.!?~…─'
|
69 |
+
_letters = '#Nabdefghijklmnoprstuvwxyzæçøŋœȵɐɑɒɓɔɕɗɘəɚɛɜɣɤɦɪɭɯɵɷɸɻɾɿʂʅʊʋʌʏʑʔʦʮʰʷˀː˥˦˧˨˩̥̩̃̚ᴀᴇ↑↓∅ⱼ '
|
70 |
+
'''
|
71 |
+
|
72 |
+
# Export all symbols:
|
73 |
+
symbols = [_pad] + list(_punctuation) + list(_letters)
|
74 |
+
|
75 |
+
# Special symbol ids
|
76 |
+
SPACE_ID = symbols.index(" ")
|
text/thai.py
ADDED
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from num_thai.thainumbers import NumThai
|
3 |
+
|
4 |
+
|
5 |
+
num = NumThai()
|
6 |
+
|
7 |
+
# List of (Latin alphabet, Thai) pairs:
|
8 |
+
_latin_to_thai = [(re.compile('%s' % x[0], re.IGNORECASE), x[1]) for x in [
|
9 |
+
('a', 'เอ'),
|
10 |
+
('b','บี'),
|
11 |
+
('c','ซี'),
|
12 |
+
('d','ดี'),
|
13 |
+
('e','อี'),
|
14 |
+
('f','เอฟ'),
|
15 |
+
('g','จี'),
|
16 |
+
('h','เอช'),
|
17 |
+
('i','ไอ'),
|
18 |
+
('j','เจ'),
|
19 |
+
('k','เค'),
|
20 |
+
('l','แอล'),
|
21 |
+
('m','เอ็ม'),
|
22 |
+
('n','เอ็น'),
|
23 |
+
('o','โอ'),
|
24 |
+
('p','พี'),
|
25 |
+
('q','คิว'),
|
26 |
+
('r','แอร์'),
|
27 |
+
('s','เอส'),
|
28 |
+
('t','ที'),
|
29 |
+
('u','ยู'),
|
30 |
+
('v','วี'),
|
31 |
+
('w','ดับเบิลยู'),
|
32 |
+
('x','เอ็กซ์'),
|
33 |
+
('y','วาย'),
|
34 |
+
('z','ซี')
|
35 |
+
]]
|
36 |
+
|
37 |
+
|
38 |
+
def num_to_thai(text):
|
39 |
+
return re.sub(r'(?:\d+(?:,?\d+)?)+(?:\.\d+(?:,?\d+)?)?', lambda x: ''.join(num.NumberToTextThai(float(x.group(0).replace(',', '')))), text)
|
40 |
+
|
41 |
+
def latin_to_thai(text):
|
42 |
+
for regex, replacement in _latin_to_thai:
|
43 |
+
text = re.sub(regex, replacement, text)
|
44 |
+
return text
|
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 +1,486 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import glob
|
3 |
+
import sys
|
4 |
+
import argparse
|
5 |
+
import logging
|
6 |
+
import json
|
7 |
+
import subprocess
|
8 |
+
import numpy as np
|
9 |
+
from scipy.io.wavfile import read
|
10 |
+
import torch
|
11 |
+
import regex as re
|
12 |
+
|
13 |
+
import loralib as lora
|
14 |
+
|
15 |
+
MATPLOTLIB_FLAG = False
|
16 |
+
|
17 |
+
logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
|
18 |
+
logger = logging
|
19 |
+
|
20 |
+
|
21 |
+
|
22 |
+
zh_pattern = re.compile(r'[\u4e00-\u9fa5]')
|
23 |
+
en_pattern = re.compile(r'[a-zA-Z]')
|
24 |
+
jp_pattern = re.compile(r'[\u3040-\u30ff\u31f0-\u31ff]')
|
25 |
+
kr_pattern = re.compile(r'[\uac00-\ud7af\u1100-\u11ff\u3130-\u318f\ua960-\ua97f]')
|
26 |
+
num_pattern=re.compile(r'[0-9]')
|
27 |
+
comma=r"(?<=[.。!!??;;,,、::'\"‘“”’()()《》「」~——])" #向前匹配但固定长度
|
28 |
+
tags={'ZH':'[ZH]','EN':'[EN]','JP':'[JA]','KR':'[KR]'}
|
29 |
+
|
30 |
+
def tag_cjke(text):
|
31 |
+
'''为中英日韩加tag,中日正则分不开,故先分句分离中日再识别,以应对大部分情况'''
|
32 |
+
sentences = re.split(r"([.。!!??;;,,、::'\"‘“”’()()【】《》「」~——]+ *(?![0-9]))", text) #分句,排除小数点
|
33 |
+
sentences.append("")
|
34 |
+
sentences = ["".join(i) for i in zip(sentences[0::2],sentences[1::2])]
|
35 |
+
# print(sentences)
|
36 |
+
prev_lang=None
|
37 |
+
tagged_text = ""
|
38 |
+
for s in sentences:
|
39 |
+
#全为符号跳过
|
40 |
+
nu = re.sub(r'[\s\p{P}]+', '', s, flags=re.U).strip()
|
41 |
+
if len(nu)==0:
|
42 |
+
continue
|
43 |
+
s = re.sub(r'[()()《》「」【】‘“”’]+', '', s)
|
44 |
+
jp=re.findall(jp_pattern, s)
|
45 |
+
#本句含日语字符判断为日语
|
46 |
+
if len(jp)>0:
|
47 |
+
prev_lang,tagged_jke=tag_jke(s,prev_lang)
|
48 |
+
tagged_text +=tagged_jke
|
49 |
+
else:
|
50 |
+
prev_lang,tagged_cke=tag_cke(s,prev_lang)
|
51 |
+
tagged_text +=tagged_cke
|
52 |
+
return tagged_text
|
53 |
+
|
54 |
+
def tag_jke(text,prev_sentence=None):
|
55 |
+
'''为英日韩加tag'''
|
56 |
+
# 初始化标记变量
|
57 |
+
tagged_text = ""
|
58 |
+
prev_lang = None
|
59 |
+
tagged=0
|
60 |
+
# 遍历文本
|
61 |
+
for char in text:
|
62 |
+
# 判断当前字符属于哪种语言
|
63 |
+
if jp_pattern.match(char):
|
64 |
+
lang = "JP"
|
65 |
+
elif zh_pattern.match(char):
|
66 |
+
lang = "JP"
|
67 |
+
elif kr_pattern.match(char):
|
68 |
+
lang = "KR"
|
69 |
+
elif en_pattern.match(char):
|
70 |
+
lang = "EN"
|
71 |
+
# elif num_pattern.match(char):
|
72 |
+
# lang = prev_sentence
|
73 |
+
else:
|
74 |
+
lang = None
|
75 |
+
tagged_text += char
|
76 |
+
continue
|
77 |
+
# 如果当前语言与上一个语言不同,就添加标记
|
78 |
+
if lang != prev_lang:
|
79 |
+
tagged=1
|
80 |
+
if prev_lang==None: # 开头
|
81 |
+
tagged_text =tags[lang]+tagged_text
|
82 |
+
else:
|
83 |
+
tagged_text =tagged_text+tags[prev_lang]+tags[lang]
|
84 |
+
|
85 |
+
# 重置标记变量
|
86 |
+
prev_lang = lang
|
87 |
+
|
88 |
+
# 添加当前字符到标记文本中
|
89 |
+
tagged_text += char
|
90 |
+
|
91 |
+
# 在最后一个语言的结尾添加对应的标记
|
92 |
+
if prev_lang:
|
93 |
+
tagged_text += tags[prev_lang]
|
94 |
+
if not tagged:
|
95 |
+
prev_lang=prev_sentence
|
96 |
+
tagged_text =tags[prev_lang]+tagged_text+tags[prev_lang]
|
97 |
+
|
98 |
+
return prev_lang,tagged_text
|
99 |
+
|
100 |
+
def tag_cke(text,prev_sentence=None):
|
101 |
+
'''为中英韩加tag'''
|
102 |
+
# 初始化标记变量
|
103 |
+
tagged_text = ""
|
104 |
+
prev_lang = None
|
105 |
+
# 是否全略过未标签
|
106 |
+
tagged=0
|
107 |
+
|
108 |
+
# 遍历文本
|
109 |
+
for char in text:
|
110 |
+
# 判断当前字符属于哪种语言
|
111 |
+
if zh_pattern.match(char):
|
112 |
+
lang = "ZH"
|
113 |
+
elif kr_pattern.match(char):
|
114 |
+
lang = "KR"
|
115 |
+
elif en_pattern.match(char):
|
116 |
+
lang = "EN"
|
117 |
+
# elif num_pattern.match(char):
|
118 |
+
# lang = prev_sentence
|
119 |
+
else:
|
120 |
+
# 略过
|
121 |
+
lang = None
|
122 |
+
tagged_text += char
|
123 |
+
continue
|
124 |
+
|
125 |
+
# 如果当前语言与上一个语言不同,添加标记
|
126 |
+
if lang != prev_lang:
|
127 |
+
tagged=1
|
128 |
+
if prev_lang==None: # 开头
|
129 |
+
tagged_text =tags[lang]+tagged_text
|
130 |
+
else:
|
131 |
+
tagged_text =tagged_text+tags[prev_lang]+tags[lang]
|
132 |
+
|
133 |
+
# 重置标记变量
|
134 |
+
prev_lang = lang
|
135 |
+
|
136 |
+
# 添加当前字符到标记文本中
|
137 |
+
tagged_text += char
|
138 |
+
|
139 |
+
# 在最后一个语言的结尾添加对应的标记
|
140 |
+
if prev_lang:
|
141 |
+
tagged_text += tags[prev_lang]
|
142 |
+
# 未标签则继承上一句标签
|
143 |
+
if tagged==0:
|
144 |
+
prev_lang=prev_sentence
|
145 |
+
tagged_text =tags[prev_lang]+tagged_text+tags[prev_lang]
|
146 |
+
return prev_lang,tagged_text
|
147 |
+
|
148 |
+
|
149 |
+
def load_lora_checkpoint(checkpoint_path, model, optimizer=None, generator_path = "./pretrained_models/G_latest.pth"):
|
150 |
+
assert os.path.isfile(checkpoint_path)
|
151 |
+
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
|
152 |
+
iteration = checkpoint_dict['iteration']
|
153 |
+
learning_rate = checkpoint_dict['learning_rate']
|
154 |
+
if optimizer is not None:
|
155 |
+
optimizer.load_state_dict(checkpoint_dict['optimizer'])
|
156 |
+
# saved_state_dict = checkpoint_dict['model']
|
157 |
+
generator_state_dict = torch.load(generator_path)['model']
|
158 |
+
lora_state_dict = checkpoint_dict['model']
|
159 |
+
new_state_dict = {}
|
160 |
+
for k, v in lora_state_dict.items():
|
161 |
+
try:
|
162 |
+
if k == 'emb_g.weight':
|
163 |
+
if drop_speaker_emb:
|
164 |
+
new_state_dict[k] = v
|
165 |
+
continue
|
166 |
+
v[:lora_state_dict[k].shape[0], :] = lora_state_dict[k]
|
167 |
+
new_state_dict[k] = v
|
168 |
+
else:
|
169 |
+
new_state_dict[k] = lora_state_dict[k]
|
170 |
+
except:
|
171 |
+
logger.info("%s is not in the checkpoint" % k)
|
172 |
+
new_state_dict[k] = v
|
173 |
+
if hasattr(model, 'module'):
|
174 |
+
model.module.load_state_dict(generator_state_dict, strict=False)
|
175 |
+
model.module.load_state_dict(new_state_dict, strict=False)
|
176 |
+
# lora.mark_only_lora_as_trainable(model.module)
|
177 |
+
else:
|
178 |
+
model.load_state_dict(generator_state_dict, strict=False)
|
179 |
+
model.load_state_dict(new_state_dict, strict=False)
|
180 |
+
# lora.mark_only_lora_as_trainable(model)
|
181 |
+
logger.info("Loaded checkpoint '{}' (iteration {})" .format(
|
182 |
+
checkpoint_path, iteration))
|
183 |
+
|
184 |
+
return model, optimizer, learning_rate, iteration
|
185 |
+
|
186 |
+
def save_lora_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
|
187 |
+
logger.info("Saving model and optimizer state at iteration {} to {}".format(
|
188 |
+
iteration, checkpoint_path))
|
189 |
+
if hasattr(model, 'module'):
|
190 |
+
state_dict = lora.lora_state_dict(model.module)
|
191 |
+
else:
|
192 |
+
state_dict = lora.lora_state_dict(model)
|
193 |
+
torch.save({'model': state_dict,
|
194 |
+
'iteration': iteration,
|
195 |
+
'optimizer': optimizer.state_dict() if optimizer is not None else None,
|
196 |
+
'learning_rate': learning_rate}, checkpoint_path)
|
197 |
+
|
198 |
+
def load_lora_checkpoint_fix(checkpoint_path, model, optimizer=None, drop_speaker_emb=False):
|
199 |
+
assert os.path.isfile(checkpoint_path)
|
200 |
+
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
|
201 |
+
iteration = checkpoint_dict['iteration']
|
202 |
+
learning_rate = checkpoint_dict['learning_rate']
|
203 |
+
if optimizer is not None:
|
204 |
+
optimizer.load_state_dict(checkpoint_dict['optimizer'])
|
205 |
+
saved_state_dict = checkpoint_dict['model']
|
206 |
+
if hasattr(model, 'module'):
|
207 |
+
state_dict = model.module.state_dict()
|
208 |
+
else:
|
209 |
+
state_dict = model.state_dict()
|
210 |
+
new_state_dict = {}
|
211 |
+
for k, v in state_dict.items():
|
212 |
+
try:
|
213 |
+
if k == 'emb_g.weight':
|
214 |
+
if drop_speaker_emb:
|
215 |
+
new_state_dict[k] = v
|
216 |
+
continue
|
217 |
+
v[:saved_state_dict[k].shape[0], :] = saved_state_dict[k]
|
218 |
+
new_state_dict[k] = v
|
219 |
+
else:
|
220 |
+
new_state_dict[k] = saved_state_dict[k]
|
221 |
+
except:
|
222 |
+
logger.info("%s is not in the checkpoint" % k)
|
223 |
+
new_state_dict[k] = v
|
224 |
+
if hasattr(model, 'module'):
|
225 |
+
model.module.load_state_dict(new_state_dict, strict=False)
|
226 |
+
lora.mark_only_lora_as_trainable(model.module)
|
227 |
+
else:
|
228 |
+
model.load_state_dict(new_state_dict, strict=False)
|
229 |
+
lora.mark_only_lora_as_trainable(model)
|
230 |
+
logger.info("Loaded checkpoint '{}' (iteration {})".format(
|
231 |
+
checkpoint_path, iteration))
|
232 |
+
return model, optimizer, learning_rate, iteration
|
233 |
+
|
234 |
+
def load_checkpoint(checkpoint_path, model, optimizer=None, drop_speaker_emb=False):
|
235 |
+
assert os.path.isfile(checkpoint_path)
|
236 |
+
checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
|
237 |
+
iteration = checkpoint_dict['iteration']
|
238 |
+
learning_rate = checkpoint_dict['learning_rate']
|
239 |
+
if optimizer is not None:
|
240 |
+
optimizer.load_state_dict(checkpoint_dict['optimizer'])
|
241 |
+
saved_state_dict = checkpoint_dict['model']
|
242 |
+
if hasattr(model, 'module'):
|
243 |
+
state_dict = model.module.state_dict()
|
244 |
+
else:
|
245 |
+
state_dict = model.state_dict()
|
246 |
+
new_state_dict = {}
|
247 |
+
for k, v in state_dict.items():
|
248 |
+
try:
|
249 |
+
if k == 'emb_g.weight':
|
250 |
+
if drop_speaker_emb:
|
251 |
+
new_state_dict[k] = v
|
252 |
+
continue
|
253 |
+
v[:saved_state_dict[k].shape[0], :] = saved_state_dict[k]
|
254 |
+
new_state_dict[k] = v
|
255 |
+
else:
|
256 |
+
new_state_dict[k] = saved_state_dict[k]
|
257 |
+
except:
|
258 |
+
logger.info("%s is not in the checkpoint" % k)
|
259 |
+
new_state_dict[k] = v
|
260 |
+
if hasattr(model, 'module'):
|
261 |
+
model.module.load_state_dict(new_state_dict)
|
262 |
+
else:
|
263 |
+
model.load_state_dict(new_state_dict)
|
264 |
+
logger.info("Loaded checkpoint '{}' (iteration {})".format(
|
265 |
+
checkpoint_path, iteration))
|
266 |
+
return model, optimizer, learning_rate, iteration
|
267 |
+
|
268 |
+
|
269 |
+
def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
|
270 |
+
logger.info("Saving model and optimizer state at iteration {} to {}".format(
|
271 |
+
iteration, checkpoint_path))
|
272 |
+
if hasattr(model, 'module'):
|
273 |
+
state_dict = model.module.state_dict()
|
274 |
+
else:
|
275 |
+
state_dict = model.state_dict()
|
276 |
+
torch.save({'model': state_dict,
|
277 |
+
'iteration': iteration,
|
278 |
+
'optimizer': optimizer.state_dict() if optimizer is not None else None,
|
279 |
+
'learning_rate': learning_rate}, checkpoint_path)
|
280 |
+
|
281 |
+
|
282 |
+
def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sampling_rate=22050):
|
283 |
+
for k, v in scalars.items():
|
284 |
+
writer.add_scalar(k, v, global_step)
|
285 |
+
for k, v in histograms.items():
|
286 |
+
writer.add_histogram(k, v, global_step)
|
287 |
+
for k, v in images.items():
|
288 |
+
writer.add_image(k, v, global_step, dataformats='HWC')
|
289 |
+
for k, v in audios.items():
|
290 |
+
writer.add_audio(k, v, global_step, audio_sampling_rate)
|
291 |
+
|
292 |
+
|
293 |
+
def latest_checkpoint_path(dir_path, regex="G_*.pth"):
|
294 |
+
f_list = glob.glob(os.path.join(dir_path, regex))
|
295 |
+
f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
|
296 |
+
x = f_list[-1]
|
297 |
+
print(x)
|
298 |
+
return x
|
299 |
+
|
300 |
+
|
301 |
+
def plot_spectrogram_to_numpy(spectrogram):
|
302 |
+
global MATPLOTLIB_FLAG
|
303 |
+
if not MATPLOTLIB_FLAG:
|
304 |
+
import matplotlib
|
305 |
+
matplotlib.use("Agg")
|
306 |
+
MATPLOTLIB_FLAG = True
|
307 |
+
mpl_logger = logging.getLogger('matplotlib')
|
308 |
+
mpl_logger.setLevel(logging.WARNING)
|
309 |
+
import matplotlib.pylab as plt
|
310 |
+
import numpy as np
|
311 |
+
|
312 |
+
fig, ax = plt.subplots(figsize=(10, 2))
|
313 |
+
im = ax.imshow(spectrogram, aspect="auto", origin="lower",
|
314 |
+
interpolation='none')
|
315 |
+
plt.colorbar(im, ax=ax)
|
316 |
+
plt.xlabel("Frames")
|
317 |
+
plt.ylabel("Channels")
|
318 |
+
plt.tight_layout()
|
319 |
+
|
320 |
+
fig.canvas.draw()
|
321 |
+
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
|
322 |
+
data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
|
323 |
+
plt.close()
|
324 |
+
return data
|
325 |
+
|
326 |
+
|
327 |
+
def plot_alignment_to_numpy(alignment, info=None):
|
328 |
+
global MATPLOTLIB_FLAG
|
329 |
+
if not MATPLOTLIB_FLAG:
|
330 |
+
import matplotlib
|
331 |
+
matplotlib.use("Agg")
|
332 |
+
MATPLOTLIB_FLAG = True
|
333 |
+
mpl_logger = logging.getLogger('matplotlib')
|
334 |
+
mpl_logger.setLevel(logging.WARNING)
|
335 |
+
import matplotlib.pylab as plt
|
336 |
+
import numpy as np
|
337 |
+
|
338 |
+
fig, ax = plt.subplots(figsize=(6, 4))
|
339 |
+
im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
|
340 |
+
interpolation='none')
|
341 |
+
fig.colorbar(im, ax=ax)
|
342 |
+
xlabel = 'Decoder timestep'
|
343 |
+
if info is not None:
|
344 |
+
xlabel += '\n\n' + info
|
345 |
+
plt.xlabel(xlabel)
|
346 |
+
plt.ylabel('Encoder timestep')
|
347 |
+
plt.tight_layout()
|
348 |
+
|
349 |
+
fig.canvas.draw()
|
350 |
+
data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
|
351 |
+
data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
|
352 |
+
plt.close()
|
353 |
+
return data
|
354 |
+
|
355 |
+
|
356 |
+
def load_wav_to_torch(full_path):
|
357 |
+
sampling_rate, data = read(full_path)
|
358 |
+
return torch.FloatTensor(data.astype(np.float32)), sampling_rate
|
359 |
+
|
360 |
+
|
361 |
+
def load_filepaths_and_text(filename, split="|"):
|
362 |
+
with open(filename, encoding='utf-8') as f:
|
363 |
+
filepaths_and_text = [line.strip().split(split) for line in f]
|
364 |
+
return filepaths_and_text
|
365 |
+
|
366 |
+
|
367 |
+
def get_hparams(init=True):
|
368 |
+
parser = argparse.ArgumentParser()
|
369 |
+
parser.add_argument('-c', '--config', type=str, default="./configs/modified_finetune_speaker.json",
|
370 |
+
help='JSON file for configuration')
|
371 |
+
parser.add_argument('-m', '--model', type=str, default="pretrained_models",
|
372 |
+
help='Model name')
|
373 |
+
parser.add_argument('-n', '--max_epochs', type=int, default=50,
|
374 |
+
help='finetune epochs')
|
375 |
+
parser.add_argument('--drop_speaker_embed', type=bool, default=False, help='whether to drop existing characters')
|
376 |
+
|
377 |
+
args = parser.parse_args()
|
378 |
+
model_dir = os.path.join("./", args.model)
|
379 |
+
|
380 |
+
if not os.path.exists(model_dir):
|
381 |
+
os.makedirs(model_dir)
|
382 |
+
|
383 |
+
config_path = args.config
|
384 |
+
config_save_path = os.path.join(model_dir, "config.json")
|
385 |
+
if init:
|
386 |
+
with open(config_path, "r") as f:
|
387 |
+
data = f.read()
|
388 |
+
with open(config_save_path, "w") as f:
|
389 |
+
f.write(data)
|
390 |
+
else:
|
391 |
+
with open(config_save_path, "r") as f:
|
392 |
+
data = f.read()
|
393 |
+
config = json.loads(data)
|
394 |
+
|
395 |
+
hparams = HParams(**config)
|
396 |
+
hparams.model_dir = model_dir
|
397 |
+
hparams.max_epochs = args.max_epochs
|
398 |
+
hparams.drop_speaker_embed = args.drop_speaker_embed
|
399 |
+
return hparams
|
400 |
+
|
401 |
+
|
402 |
+
def get_hparams_from_dir(model_dir):
|
403 |
+
config_save_path = os.path.join(model_dir, "config.json")
|
404 |
+
with open(config_save_path, "r") as f:
|
405 |
+
data = f.read()
|
406 |
+
config = json.loads(data)
|
407 |
+
|
408 |
+
hparams = HParams(**config)
|
409 |
+
hparams.model_dir = model_dir
|
410 |
+
return hparams
|
411 |
+
|
412 |
+
|
413 |
+
def get_hparams_from_file(config_path):
|
414 |
+
with open(config_path, "r", encoding="utf-8") as f:
|
415 |
+
data = f.read()
|
416 |
+
config = json.loads(data)
|
417 |
+
|
418 |
+
hparams = HParams(**config)
|
419 |
+
return hparams
|
420 |
+
|
421 |
+
|
422 |
+
def check_git_hash(model_dir):
|
423 |
+
source_dir = os.path.dirname(os.path.realpath(__file__))
|
424 |
+
if not os.path.exists(os.path.join(source_dir, ".git")):
|
425 |
+
logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
|
426 |
+
source_dir
|
427 |
+
))
|
428 |
+
return
|
429 |
+
|
430 |
+
cur_hash = subprocess.getoutput("git rev-parse HEAD")
|
431 |
+
|
432 |
+
path = os.path.join(model_dir, "githash")
|
433 |
+
if os.path.exists(path):
|
434 |
+
saved_hash = open(path).read()
|
435 |
+
if saved_hash != cur_hash:
|
436 |
+
logger.warn("git hash values are different. {}(saved) != {}(current)".format(
|
437 |
+
saved_hash[:8], cur_hash[:8]))
|
438 |
+
else:
|
439 |
+
open(path, "w").write(cur_hash)
|
440 |
+
|
441 |
+
|
442 |
+
def get_logger(model_dir, filename="train.log"):
|
443 |
+
global logger
|
444 |
+
logger = logging.getLogger(os.path.basename(model_dir))
|
445 |
+
logger.setLevel(logging.DEBUG)
|
446 |
+
|
447 |
+
formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
|
448 |
+
if not os.path.exists(model_dir):
|
449 |
+
os.makedirs(model_dir)
|
450 |
+
h = logging.FileHandler(os.path.join(model_dir, filename))
|
451 |
+
h.setLevel(logging.DEBUG)
|
452 |
+
h.setFormatter(formatter)
|
453 |
+
logger.addHandler(h)
|
454 |
+
return logger
|
455 |
+
|
456 |
+
|
457 |
+
class HParams():
|
458 |
+
def __init__(self, **kwargs):
|
459 |
+
for k, v in kwargs.items():
|
460 |
+
if type(v) == dict:
|
461 |
+
v = HParams(**v)
|
462 |
+
self[k] = v
|
463 |
+
|
464 |
+
def keys(self):
|
465 |
+
return self.__dict__.keys()
|
466 |
+
|
467 |
+
def items(self):
|
468 |
+
return self.__dict__.items()
|
469 |
+
|
470 |
+
def values(self):
|
471 |
+
return self.__dict__.values()
|
472 |
+
|
473 |
+
def __len__(self):
|
474 |
+
return len(self.__dict__)
|
475 |
+
|
476 |
+
def __getitem__(self, key):
|
477 |
+
return getattr(self, key)
|
478 |
+
|
479 |
+
def __setitem__(self, key, value):
|
480 |
+
return setattr(self, key, value)
|
481 |
+
|
482 |
+
def __contains__(self, key):
|
483 |
+
return key in self.__dict__
|
484 |
+
|
485 |
+
def __repr__(self):
|
486 |
+
return self.__dict__.__repr__()
|