├── .gitignore ├── LICENSE ├── Modules ├── .gitignore ├── CBHG.py ├── CBHGNOBN.py ├── ConvNet.py ├── ConvNetNOBN.py ├── FastCBHG.py ├── HighwayNet.py └── __init__.py ├── README.md ├── __init__.py └── model.py /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__/ 2 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU AFFERO GENERAL PUBLIC LICENSE 2 | Version 3, 19 November 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU Affero General Public License is a free, copyleft license for 11 | software and other kinds of works, specifically designed to ensure 12 | cooperation with the community in the case of network server software. 13 | 14 | The licenses for most software and other practical works are designed 15 | to take away your freedom to share and change the works. By contrast, 16 | our General Public Licenses are intended to guarantee your freedom to 17 | share and change all versions of a program--to make sure it remains free 18 | software for all its users. 19 | 20 | When we speak of free software, we are referring to freedom, not 21 | price. Our General Public Licenses are designed to make sure that you 22 | have the freedom to distribute copies of free software (and charge for 23 | them if you wish), that you receive source code or can get it if you 24 | want it, that you can change the software or use pieces of it in new 25 | free programs, and that you know you can do these things. 26 | 27 | Developers that use our General Public Licenses protect your rights 28 | with two steps: (1) assert copyright on the software, and (2) offer 29 | you this License which gives you legal permission to copy, distribute 30 | and/or modify the software. 31 | 32 | A secondary benefit of defending all users' freedom is that 33 | improvements made in alternate versions of the program, if they 34 | receive widespread use, become available for other developers to 35 | incorporate. Many developers of free software are heartened and 36 | encouraged by the resulting cooperation. However, in the case of 37 | software used on network servers, this result may fail to come about. 38 | The GNU General Public License permits making a modified version and 39 | letting the public access it on a server without ever releasing its 40 | source code to the public. 41 | 42 | The GNU Affero General Public License is designed specifically to 43 | ensure that, in such cases, the modified source code becomes available 44 | to the community. It requires the operator of a network server to 45 | provide the source code of the modified version running there to the 46 | users of that server. Therefore, public use of a modified version, on 47 | a publicly accessible server, gives the public access to the source 48 | code of the modified version. 49 | 50 | An older license, called the Affero General Public License and 51 | published by Affero, was designed to accomplish similar goals. This is 52 | a different license, not a version of the Affero GPL, but Affero has 53 | released a new version of the Affero GPL which permits relicensing under 54 | this license. 55 | 56 | The precise terms and conditions for copying, distribution and 57 | modification follow. 58 | 59 | TERMS AND CONDITIONS 60 | 61 | 0. Definitions. 62 | 63 | "This License" refers to version 3 of the GNU Affero General Public License. 64 | 65 | "Copyright" also means copyright-like laws that apply to other kinds of 66 | works, such as semiconductor masks. 67 | 68 | "The Program" refers to any copyrightable work licensed under this 69 | License. Each licensee is addressed as "you". "Licensees" and 70 | "recipients" may be individuals or organizations. 71 | 72 | To "modify" a work means to copy from or adapt all or part of the work 73 | in a fashion requiring copyright permission, other than the making of an 74 | exact copy. The resulting work is called a "modified version" of the 75 | earlier work or a work "based on" the earlier work. 76 | 77 | A "covered work" means either the unmodified Program or a work based 78 | on the Program. 79 | 80 | To "propagate" a work means to do anything with it that, without 81 | permission, would make you directly or secondarily liable for 82 | infringement under applicable copyright law, except executing it on a 83 | computer or modifying a private copy. Propagation includes copying, 84 | distribution (with or without modification), making available to the 85 | public, and in some countries other activities as well. 86 | 87 | To "convey" a work means any kind of propagation that enables other 88 | parties to make or receive copies. Mere interaction with a user through 89 | a computer network, with no transfer of a copy, is not conveying. 90 | 91 | An interactive user interface displays "Appropriate Legal Notices" 92 | to the extent that it includes a convenient and prominently visible 93 | feature that (1) displays an appropriate copyright notice, and (2) 94 | tells the user that there is no warranty for the work (except to the 95 | extent that warranties are provided), that licensees may convey the 96 | work under this License, and how to view a copy of this License. If 97 | the interface presents a list of user commands or options, such as a 98 | menu, a prominent item in the list meets this criterion. 99 | 100 | 1. Source Code. 101 | 102 | The "source code" for a work means the preferred form of the work 103 | for making modifications to it. "Object code" means any non-source 104 | form of a work. 105 | 106 | A "Standard Interface" means an interface that either is an official 107 | standard defined by a recognized standards body, or, in the case of 108 | interfaces specified for a particular programming language, one that 109 | is widely used among developers working in that language. 110 | 111 | The "System Libraries" of an executable work include anything, other 112 | than the work as a whole, that (a) is included in the normal form of 113 | packaging a Major Component, but which is not part of that Major 114 | Component, and (b) serves only to enable use of the work with that 115 | Major Component, or to implement a Standard Interface for which an 116 | implementation is available to the public in source code form. A 117 | "Major Component", in this context, means a major essential component 118 | (kernel, window system, and so on) of the specific operating system 119 | (if any) on which the executable work runs, or a compiler used to 120 | produce the work, or an object code interpreter used to run it. 121 | 122 | The "Corresponding Source" for a work in object code form means all 123 | the source code needed to generate, install, and (for an executable 124 | work) run the object code and to modify the work, including scripts to 125 | control those activities. However, it does not include the work's 126 | System Libraries, or general-purpose tools or generally available free 127 | programs which are used unmodified in performing those activities but 128 | which are not part of the work. For example, Corresponding Source 129 | includes interface definition files associated with source files for 130 | the work, and the source code for shared libraries and dynamically 131 | linked subprograms that the work is specifically designed to require, 132 | such as by intimate data communication or control flow between those 133 | subprograms and other parts of the work. 134 | 135 | The Corresponding Source need not include anything that users 136 | can regenerate automatically from other parts of the Corresponding 137 | Source. 138 | 139 | The Corresponding Source for a work in source code form is that 140 | same work. 141 | 142 | 2. Basic Permissions. 143 | 144 | All rights granted under this License are granted for the term of 145 | copyright on the Program, and are irrevocable provided the stated 146 | conditions are met. This License explicitly affirms your unlimited 147 | permission to run the unmodified Program. The output from running a 148 | covered work is covered by this License only if the output, given its 149 | content, constitutes a covered work. This License acknowledges your 150 | rights of fair use or other equivalent, as provided by copyright law. 151 | 152 | You may make, run and propagate covered works that you do not 153 | convey, without conditions so long as your license otherwise remains 154 | in force. You may convey covered works to others for the sole purpose 155 | of having them make modifications exclusively for you, or provide you 156 | with facilities for running those works, provided that you comply with 157 | the terms of this License in conveying all material for which you do 158 | not control copyright. Those thus making or running the covered works 159 | for you must do so exclusively on your behalf, under your direction 160 | and control, on terms that prohibit them from making any copies of 161 | your copyrighted material outside their relationship with you. 162 | 163 | Conveying under any other circumstances is permitted solely under 164 | the conditions stated below. Sublicensing is not allowed; section 10 165 | makes it unnecessary. 166 | 167 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 168 | 169 | No covered work shall be deemed part of an effective technological 170 | measure under any applicable law fulfilling obligations under article 171 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 172 | similar laws prohibiting or restricting circumvention of such 173 | measures. 174 | 175 | When you convey a covered work, you waive any legal power to forbid 176 | circumvention of technological measures to the extent such circumvention 177 | is effected by exercising rights under this License with respect to 178 | the covered work, and you disclaim any intention to limit operation or 179 | modification of the work as a means of enforcing, against the work's 180 | users, your or third parties' legal rights to forbid circumvention of 181 | technological measures. 182 | 183 | 4. Conveying Verbatim Copies. 184 | 185 | You may convey verbatim copies of the Program's source code as you 186 | receive it, in any medium, provided that you conspicuously and 187 | appropriately publish on each copy an appropriate copyright notice; 188 | keep intact all notices stating that this License and any 189 | non-permissive terms added in accord with section 7 apply to the code; 190 | keep intact all notices of the absence of any warranty; and give all 191 | recipients a copy of this License along with the Program. 192 | 193 | You may charge any price or no price for each copy that you convey, 194 | and you may offer support or warranty protection for a fee. 195 | 196 | 5. Conveying Modified Source Versions. 197 | 198 | You may convey a work based on the Program, or the modifications to 199 | produce it from the Program, in the form of source code under the 200 | terms of section 4, provided that you also meet all of these conditions: 201 | 202 | a) The work must carry prominent notices stating that you modified 203 | it, and giving a relevant date. 204 | 205 | b) The work must carry prominent notices stating that it is 206 | released under this License and any conditions added under section 207 | 7. This requirement modifies the requirement in section 4 to 208 | "keep intact all notices". 209 | 210 | c) You must license the entire work, as a whole, under this 211 | License to anyone who comes into possession of a copy. This 212 | License will therefore apply, along with any applicable section 7 213 | additional terms, to the whole of the work, and all its parts, 214 | regardless of how they are packaged. This License gives no 215 | permission to license the work in any other way, but it does not 216 | invalidate such permission if you have separately received it. 217 | 218 | d) If the work has interactive user interfaces, each must display 219 | Appropriate Legal Notices; however, if the Program has interactive 220 | interfaces that do not display Appropriate Legal Notices, your 221 | work need not make them do so. 222 | 223 | A compilation of a covered work with other separate and independent 224 | works, which are not by their nature extensions of the covered work, 225 | and which are not combined with it such as to form a larger program, 226 | in or on a volume of a storage or distribution medium, is called an 227 | "aggregate" if the compilation and its resulting copyright are not 228 | used to limit the access or legal rights of the compilation's users 229 | beyond what the individual works permit. Inclusion of a covered work 230 | in an aggregate does not cause this License to apply to the other 231 | parts of the aggregate. 232 | 233 | 6. Conveying Non-Source Forms. 234 | 235 | You may convey a covered work in object code form under the terms 236 | of sections 4 and 5, provided that you also convey the 237 | machine-readable Corresponding Source under the terms of this License, 238 | in one of these ways: 239 | 240 | a) Convey the object code in, or embodied in, a physical product 241 | (including a physical distribution medium), accompanied by the 242 | Corresponding Source fixed on a durable physical medium 243 | customarily used for software interchange. 244 | 245 | b) Convey the object code in, or embodied in, a physical product 246 | (including a physical distribution medium), accompanied by a 247 | written offer, valid for at least three years and valid for as 248 | long as you offer spare parts or customer support for that product 249 | model, to give anyone who possesses the object code either (1) a 250 | copy of the Corresponding Source for all the software in the 251 | product that is covered by this License, on a durable physical 252 | medium customarily used for software interchange, for a price no 253 | more than your reasonable cost of physically performing this 254 | conveying of source, or (2) access to copy the 255 | Corresponding Source from a network server at no charge. 256 | 257 | c) Convey individual copies of the object code with a copy of the 258 | written offer to provide the Corresponding Source. This 259 | alternative is allowed only occasionally and noncommercially, and 260 | only if you received the object code with such an offer, in accord 261 | with subsection 6b. 262 | 263 | d) Convey the object code by offering access from a designated 264 | place (gratis or for a charge), and offer equivalent access to the 265 | Corresponding Source in the same way through the same place at no 266 | further charge. You need not require recipients to copy the 267 | Corresponding Source along with the object code. If the place to 268 | copy the object code is a network server, the Corresponding Source 269 | may be on a different server (operated by you or a third party) 270 | that supports equivalent copying facilities, provided you maintain 271 | clear directions next to the object code saying where to find the 272 | Corresponding Source. Regardless of what server hosts the 273 | Corresponding Source, you remain obligated to ensure that it is 274 | available for as long as needed to satisfy these requirements. 275 | 276 | e) Convey the object code using peer-to-peer transmission, provided 277 | you inform other peers where the object code and Corresponding 278 | Source of the work are being offered to the general public at no 279 | charge under subsection 6d. 280 | 281 | A separable portion of the object code, whose source code is excluded 282 | from the Corresponding Source as a System Library, need not be 283 | included in conveying the object code work. 284 | 285 | A "User Product" is either (1) a "consumer product", which means any 286 | tangible personal property which is normally used for personal, family, 287 | or household purposes, or (2) anything designed or sold for incorporation 288 | into a dwelling. In determining whether a product is a consumer product, 289 | doubtful cases shall be resolved in favor of coverage. For a particular 290 | product received by a particular user, "normally used" refers to a 291 | typical or common use of that class of product, regardless of the status 292 | of the particular user or of the way in which the particular user 293 | actually uses, or expects or is expected to use, the product. A product 294 | is a consumer product regardless of whether the product has substantial 295 | commercial, industrial or non-consumer uses, unless such uses represent 296 | the only significant mode of use of the product. 297 | 298 | "Installation Information" for a User Product means any methods, 299 | procedures, authorization keys, or other information required to install 300 | and execute modified versions of a covered work in that User Product from 301 | a modified version of its Corresponding Source. The information must 302 | suffice to ensure that the continued functioning of the modified object 303 | code is in no case prevented or interfered with solely because 304 | modification has been made. 305 | 306 | If you convey an object code work under this section in, or with, or 307 | specifically for use in, a User Product, and the conveying occurs as 308 | part of a transaction in which the right of possession and use of the 309 | User Product is transferred to the recipient in perpetuity or for a 310 | fixed term (regardless of how the transaction is characterized), the 311 | Corresponding Source conveyed under this section must be accompanied 312 | by the Installation Information. But this requirement does not apply 313 | if neither you nor any third party retains the ability to install 314 | modified object code on the User Product (for example, the work has 315 | been installed in ROM). 316 | 317 | The requirement to provide Installation Information does not include a 318 | requirement to continue to provide support service, warranty, or updates 319 | for a work that has been modified or installed by the recipient, or for 320 | the User Product in which it has been modified or installed. Access to a 321 | network may be denied when the modification itself materially and 322 | adversely affects the operation of the network or violates the rules and 323 | protocols for communication across the network. 324 | 325 | Corresponding Source conveyed, and Installation Information provided, 326 | in accord with this section must be in a format that is publicly 327 | documented (and with an implementation available to the public in 328 | source code form), and must require no special password or key for 329 | unpacking, reading or copying. 330 | 331 | 7. Additional Terms. 332 | 333 | "Additional permissions" are terms that supplement the terms of this 334 | License by making exceptions from one or more of its conditions. 335 | Additional permissions that are applicable to the entire Program shall 336 | be treated as though they were included in this License, to the extent 337 | that they are valid under applicable law. If additional permissions 338 | apply only to part of the Program, that part may be used separately 339 | under those permissions, but the entire Program remains governed by 340 | this License without regard to the additional permissions. 341 | 342 | When you convey a copy of a covered work, you may at your option 343 | remove any additional permissions from that copy, or from any part of 344 | it. (Additional permissions may be written to require their own 345 | removal in certain cases when you modify the work.) You may place 346 | additional permissions on material, added by you to a covered work, 347 | for which you have or can give appropriate copyright permission. 348 | 349 | Notwithstanding any other provision of this License, for material you 350 | add to a covered work, you may (if authorized by the copyright holders of 351 | that material) supplement the terms of this License with terms: 352 | 353 | a) Disclaiming warranty or limiting liability differently from the 354 | terms of sections 15 and 16 of this License; or 355 | 356 | b) Requiring preservation of specified reasonable legal notices or 357 | author attributions in that material or in the Appropriate Legal 358 | Notices displayed by works containing it; or 359 | 360 | c) Prohibiting misrepresentation of the origin of that material, or 361 | requiring that modified versions of such material be marked in 362 | reasonable ways as different from the original version; or 363 | 364 | d) Limiting the use for publicity purposes of names of licensors or 365 | authors of the material; or 366 | 367 | e) Declining to grant rights under trademark law for use of some 368 | trade names, trademarks, or service marks; or 369 | 370 | f) Requiring indemnification of licensors and authors of that 371 | material by anyone who conveys the material (or modified versions of 372 | it) with contractual assumptions of liability to the recipient, for 373 | any liability that these contractual assumptions directly impose on 374 | those licensors and authors. 375 | 376 | All other non-permissive additional terms are considered "further 377 | restrictions" within the meaning of section 10. If the Program as you 378 | received it, or any part of it, contains a notice stating that it is 379 | governed by this License along with a term that is a further 380 | restriction, you may remove that term. If a license document contains 381 | a further restriction but permits relicensing or conveying under this 382 | License, you may add to a covered work material governed by the terms 383 | of that license document, provided that the further restriction does 384 | not survive such relicensing or conveying. 385 | 386 | If you add terms to a covered work in accord with this section, you 387 | must place, in the relevant source files, a statement of the 388 | additional terms that apply to those files, or a notice indicating 389 | where to find the applicable terms. 390 | 391 | Additional terms, permissive or non-permissive, may be stated in the 392 | form of a separately written license, or stated as exceptions; 393 | the above requirements apply either way. 394 | 395 | 8. Termination. 396 | 397 | You may not propagate or modify a covered work except as expressly 398 | provided under this License. Any attempt otherwise to propagate or 399 | modify it is void, and will automatically terminate your rights under 400 | this License (including any patent licenses granted under the third 401 | paragraph of section 11). 402 | 403 | However, if you cease all violation of this License, then your 404 | license from a particular copyright holder is reinstated (a) 405 | provisionally, unless and until the copyright holder explicitly and 406 | finally terminates your license, and (b) permanently, if the copyright 407 | holder fails to notify you of the violation by some reasonable means 408 | prior to 60 days after the cessation. 409 | 410 | Moreover, your license from a particular copyright holder is 411 | reinstated permanently if the copyright holder notifies you of the 412 | violation by some reasonable means, this is the first time you have 413 | received notice of violation of this License (for any work) from that 414 | copyright holder, and you cure the violation prior to 30 days after 415 | your receipt of the notice. 416 | 417 | Termination of your rights under this section does not terminate the 418 | licenses of parties who have received copies or rights from you under 419 | this License. If your rights have been terminated and not permanently 420 | reinstated, you do not qualify to receive new licenses for the same 421 | material under section 10. 422 | 423 | 9. Acceptance Not Required for Having Copies. 424 | 425 | You are not required to accept this License in order to receive or 426 | run a copy of the Program. Ancillary propagation of a covered work 427 | occurring solely as a consequence of using peer-to-peer transmission 428 | to receive a copy likewise does not require acceptance. However, 429 | nothing other than this License grants you permission to propagate or 430 | modify any covered work. These actions infringe copyright if you do 431 | not accept this License. Therefore, by modifying or propagating a 432 | covered work, you indicate your acceptance of this License to do so. 433 | 434 | 10. Automatic Licensing of Downstream Recipients. 435 | 436 | Each time you convey a covered work, the recipient automatically 437 | receives a license from the original licensors, to run, modify and 438 | propagate that work, subject to this License. You are not responsible 439 | for enforcing compliance by third parties with this License. 440 | 441 | An "entity transaction" is a transaction transferring control of an 442 | organization, or substantially all assets of one, or subdividing an 443 | organization, or merging organizations. If propagation of a covered 444 | work results from an entity transaction, each party to that 445 | transaction who receives a copy of the work also receives whatever 446 | licenses to the work the party's predecessor in interest had or could 447 | give under the previous paragraph, plus a right to possession of the 448 | Corresponding Source of the work from the predecessor in interest, if 449 | the predecessor has it or can get it with reasonable efforts. 450 | 451 | You may not impose any further restrictions on the exercise of the 452 | rights granted or affirmed under this License. For example, you may 453 | not impose a license fee, royalty, or other charge for exercise of 454 | rights granted under this License, and you may not initiate litigation 455 | (including a cross-claim or counterclaim in a lawsuit) alleging that 456 | any patent claim is infringed by making, using, selling, offering for 457 | sale, or importing the Program or any portion of it. 458 | 459 | 11. Patents. 460 | 461 | A "contributor" is a copyright holder who authorizes use under this 462 | License of the Program or a work on which the Program is based. The 463 | work thus licensed is called the contributor's "contributor version". 464 | 465 | A contributor's "essential patent claims" are all patent claims 466 | owned or controlled by the contributor, whether already acquired or 467 | hereafter acquired, that would be infringed by some manner, permitted 468 | by this License, of making, using, or selling its contributor version, 469 | but do not include claims that would be infringed only as a 470 | consequence of further modification of the contributor version. For 471 | purposes of this definition, "control" includes the right to grant 472 | patent sublicenses in a manner consistent with the requirements of 473 | this License. 474 | 475 | Each contributor grants you a non-exclusive, worldwide, royalty-free 476 | patent license under the contributor's essential patent claims, to 477 | make, use, sell, offer for sale, import and otherwise run, modify and 478 | propagate the contents of its contributor version. 479 | 480 | In the following three paragraphs, a "patent license" is any express 481 | agreement or commitment, however denominated, not to enforce a patent 482 | (such as an express permission to practice a patent or covenant not to 483 | sue for patent infringement). To "grant" such a patent license to a 484 | party means to make such an agreement or commitment not to enforce a 485 | patent against the party. 486 | 487 | If you convey a covered work, knowingly relying on a patent license, 488 | and the Corresponding Source of the work is not available for anyone 489 | to copy, free of charge and under the terms of this License, through a 490 | publicly available network server or other readily accessible means, 491 | then you must either (1) cause the Corresponding Source to be so 492 | available, or (2) arrange to deprive yourself of the benefit of the 493 | patent license for this particular work, or (3) arrange, in a manner 494 | consistent with the requirements of this License, to extend the patent 495 | license to downstream recipients. "Knowingly relying" means you have 496 | actual knowledge that, but for the patent license, your conveying the 497 | covered work in a country, or your recipient's use of the covered work 498 | in a country, would infringe one or more identifiable patents in that 499 | country that you have reason to believe are valid. 500 | 501 | If, pursuant to or in connection with a single transaction or 502 | arrangement, you convey, or propagate by procuring conveyance of, a 503 | covered work, and grant a patent license to some of the parties 504 | receiving the covered work authorizing them to use, propagate, modify 505 | or convey a specific copy of the covered work, then the patent license 506 | you grant is automatically extended to all recipients of the covered 507 | work and works based on it. 508 | 509 | A patent license is "discriminatory" if it does not include within 510 | the scope of its coverage, prohibits the exercise of, or is 511 | conditioned on the non-exercise of one or more of the rights that are 512 | specifically granted under this License. You may not convey a covered 513 | work if you are a party to an arrangement with a third party that is 514 | in the business of distributing software, under which you make payment 515 | to the third party based on the extent of your activity of conveying 516 | the work, and under which the third party grants, to any of the 517 | parties who would receive the covered work from you, a discriminatory 518 | patent license (a) in connection with copies of the covered work 519 | conveyed by you (or copies made from those copies), or (b) primarily 520 | for and in connection with specific products or compilations that 521 | contain the covered work, unless you entered into that arrangement, 522 | or that patent license was granted, prior to 28 March 2007. 523 | 524 | Nothing in this License shall be construed as excluding or limiting 525 | any implied license or other defenses to infringement that may 526 | otherwise be available to you under applicable patent law. 527 | 528 | 12. No Surrender of Others' Freedom. 529 | 530 | If conditions are imposed on you (whether by court order, agreement or 531 | otherwise) that contradict the conditions of this License, they do not 532 | excuse you from the conditions of this License. If you cannot convey a 533 | covered work so as to satisfy simultaneously your obligations under this 534 | License and any other pertinent obligations, then as a consequence you may 535 | not convey it at all. For example, if you agree to terms that obligate you 536 | to collect a royalty for further conveying from those to whom you convey 537 | the Program, the only way you could satisfy both those terms and this 538 | License would be to refrain entirely from conveying the Program. 539 | 540 | 13. Remote Network Interaction; Use with the GNU General Public License. 541 | 542 | Notwithstanding any other provision of this License, if you modify the 543 | Program, your modified version must prominently offer all users 544 | interacting with it remotely through a computer network (if your version 545 | supports such interaction) an opportunity to receive the Corresponding 546 | Source of your version by providing access to the Corresponding Source 547 | from a network server at no charge, through some standard or customary 548 | means of facilitating copying of software. This Corresponding Source 549 | shall include the Corresponding Source for any work covered by version 3 550 | of the GNU General Public License that is incorporated pursuant to the 551 | following paragraph. 552 | 553 | Notwithstanding any other provision of this License, you have 554 | permission to link or combine any covered work with a work licensed 555 | under version 3 of the GNU General Public License into a single 556 | combined work, and to convey the resulting work. The terms of this 557 | License will continue to apply to the part which is the covered work, 558 | but the work with which it is combined will remain governed by version 559 | 3 of the GNU General Public License. 560 | 561 | 14. Revised Versions of this License. 562 | 563 | The Free Software Foundation may publish revised and/or new versions of 564 | the GNU Affero General Public License from time to time. Such new versions 565 | will be similar in spirit to the present version, but may differ in detail to 566 | address new problems or concerns. 567 | 568 | Each version is given a distinguishing version number. If the 569 | Program specifies that a certain numbered version of the GNU Affero General 570 | Public License "or any later version" applies to it, you have the 571 | option of following the terms and conditions either of that numbered 572 | version or of any later version published by the Free Software 573 | Foundation. If the Program does not specify a version number of the 574 | GNU Affero General Public License, you may choose any version ever published 575 | by the Free Software Foundation. 576 | 577 | If the Program specifies that a proxy can decide which future 578 | versions of the GNU Affero General Public License can be used, that proxy's 579 | public statement of acceptance of a version permanently authorizes you 580 | to choose that version for the Program. 581 | 582 | Later license versions may give you additional or different 583 | permissions. However, no additional obligations are imposed on any 584 | author or copyright holder as a result of your choosing to follow a 585 | later version. 586 | 587 | 15. Disclaimer of Warranty. 588 | 589 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 590 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 591 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 592 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 593 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 594 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 595 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 596 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 597 | 598 | 16. Limitation of Liability. 599 | 600 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 601 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 602 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 603 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 604 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 605 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 606 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 607 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 608 | SUCH DAMAGES. 609 | 610 | 17. Interpretation of Sections 15 and 16. 611 | 612 | If the disclaimer of warranty and limitation of liability provided 613 | above cannot be given local legal effect according to their terms, 614 | reviewing courts shall apply local law that most closely approximates 615 | an absolute waiver of all civil liability in connection with the 616 | Program, unless a warranty or assumption of liability accompanies a 617 | copy of the Program in return for a fee. 618 | 619 | END OF TERMS AND CONDITIONS 620 | 621 | How to Apply These Terms to Your New Programs 622 | 623 | If you develop a new program, and you want it to be of the greatest 624 | possible use to the public, the best way to achieve this is to make it 625 | free software which everyone can redistribute and change under these terms. 626 | 627 | To do so, attach the following notices to the program. It is safest 628 | to attach them to the start of each source file to most effectively 629 | state the exclusion of warranty; and each file should have at least 630 | the "copyright" line and a pointer to where the full notice is found. 631 | 632 | 633 | Copyright (C) 634 | 635 | This program is free software: you can redistribute it and/or modify 636 | it under the terms of the GNU Affero General Public License as published 637 | by the Free Software Foundation, either version 3 of the License, or 638 | (at your option) any later version. 639 | 640 | This program is distributed in the hope that it will be useful, 641 | but WITHOUT ANY WARRANTY; without even the implied warranty of 642 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 643 | GNU Affero General Public License for more details. 644 | 645 | You should have received a copy of the GNU Affero General Public License 646 | along with this program. If not, see . 647 | 648 | Also add information on how to contact you by electronic and paper mail. 649 | 650 | If your software can interact with users remotely through a computer 651 | network, you should also make sure that it provides a way for users to 652 | get its source. For example, if your program is a web application, its 653 | interface could display a "Source" link that leads users to an archive 654 | of the code. There are many ways you could offer source, and different 655 | solutions will be better for different programs; see section 13 for the 656 | specific requirements. 657 | 658 | You should also get your employer (if you work as a programmer) or school, 659 | if any, to sign a "copyright disclaimer" for the program, if necessary. 660 | For more information on this, and how to apply and follow the GNU AGPL, see 661 | . 662 | -------------------------------------------------------------------------------- /Modules/.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__/ 2 | -------------------------------------------------------------------------------- /Modules/CBHG.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from tensorflow.python.ops import array_ops 3 | from TFCommon.RNNCell import GRUCell 4 | from Tacotron.Modules import ConvNet, HighwayNet 5 | 6 | bidirectional_dynamic_rnn = tf.nn.bidirectional_dynamic_rnn 7 | 8 | Conv1dBankWithMaxPool = ConvNet.Conv1dBankWithMaxPool 9 | Conv1dProjection = ConvNet.Conv1dProjection 10 | FCHighwayNet = HighwayNet.FCHighwayNet 11 | 12 | class CBHG(object): 13 | """CBHG Net 14 | """ 15 | 16 | def __init__(self, bank_K, proj_unit, highway_layers=4): 17 | """ 18 | Args: 19 | bank_K: int 20 | proj_unit: a pair of int 21 | """ 22 | self.__bank_K = bank_K 23 | self.__proj_unit = proj_unit 24 | self.__highway_layers = highway_layers 25 | 26 | @property 27 | def bank_K(self): 28 | return self.__bank_K 29 | 30 | @property 31 | def proj_unit(self): 32 | return self.__proj_unit 33 | 34 | @property 35 | def highway_layers(self): 36 | return self.__highway_layers 37 | 38 | def __call__(self, inputs, sequence_length=None, is_training=True, time_major=None): 39 | assert time_major is not None, "[*] You must specify whether is time_major or not!" 40 | if time_major: 41 | inputs = tf.transpose(inputs, perm=(1,0,2)) # Use batch major data. 42 | assert inputs.get_shape()[-1] == self.proj_unit[1], "[!] input's shape is not the same as ConvProj's output!" 43 | 44 | ### for correctness. 45 | if sequence_length is not None: 46 | mask = tf.expand_dims(array_ops.sequence_mask(sequence_length, tf.shape(inputs)[1], tf.float32), -1) 47 | inputs = inputs * mask 48 | 49 | ConvBankWithPool = Conv1dBankWithMaxPool(self.bank_K) 50 | ConvProj = Conv1dProjection(self.proj_unit) 51 | Highway = FCHighwayNet(self.highway_layers) 52 | rnn_cell_fw = GRUCell(self.proj_unit[1]) 53 | rnn_cell_bw = GRUCell(self.proj_unit[1]) 54 | 55 | ### calculate 56 | # conv net 57 | output_0 = ConvBankWithPool(inputs, is_training) 58 | 59 | ### for correctness. 60 | if sequence_length is not None: 61 | output_0 = output_0 * mask 62 | 63 | output_1 = ConvProj(output_0, is_training) 64 | # residual connect 65 | res_output = tf.identity(inputs) + output_1 66 | 67 | # highway net 68 | highway_output = Highway(res_output) 69 | 70 | # biGRU 71 | # batch major 72 | final_output, *_ = bidirectional_dynamic_rnn(rnn_cell_fw, rnn_cell_bw, highway_output, sequence_length=sequence_length, time_major=False, dtype=tf.float32) 73 | final_output = tf.concat(final_output, axis=-1) 74 | if time_major: 75 | final_output = tf.transpose(final_output, perm=(1,0,2)) 76 | 77 | return final_output 78 | 79 | -------------------------------------------------------------------------------- /Modules/CBHGNOBN.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from tensorflow.python.ops import array_ops 3 | from TFCommon.RNNCell import GRUCell 4 | from Tacotron.Modules import ConvNetNOBN, HighwayNet 5 | 6 | bidirectional_dynamic_rnn = tf.nn.bidirectional_dynamic_rnn 7 | 8 | Conv1dBankWithMaxPool = ConvNetNOBN.Conv1dBankWithMaxPool 9 | Conv1dProjection = ConvNetNOBN.Conv1dProjection 10 | FCHighwayNet = HighwayNet.FCHighwayNet 11 | 12 | class CBHG(object): 13 | """CBHG Net 14 | """ 15 | 16 | def __init__(self, bank_K, proj_unit, highway_layers=4): 17 | """ 18 | Args: 19 | bank_K: int 20 | proj_unit: a pair of int 21 | """ 22 | self.__bank_K = bank_K 23 | self.__proj_unit = proj_unit 24 | self.__highway_layers = highway_layers 25 | 26 | @property 27 | def bank_K(self): 28 | return self.__bank_K 29 | 30 | @property 31 | def proj_unit(self): 32 | return self.__proj_unit 33 | 34 | @property 35 | def highway_layers(self): 36 | return self.__highway_layers 37 | 38 | def __call__(self, inputs, sequence_length=None, is_training=True, time_major=None): 39 | assert time_major is not None, "[*] You must specify whether is time_major or not!" 40 | if time_major: 41 | inputs = tf.transpose(inputs, perm=(1,0,2)) # Use batch major data. 42 | assert inputs.get_shape()[-1] == self.proj_unit[1], "[!] input's shape is not the same as ConvProj's output!" 43 | 44 | ### for correctness. 45 | if sequence_length is not None: 46 | mask = tf.expand_dims(array_ops.sequence_mask(sequence_length, tf.shape(inputs)[1], tf.float32), -1) 47 | inputs = inputs * mask 48 | 49 | ConvBankWithPool = Conv1dBankWithMaxPool(self.bank_K) 50 | ConvProj = Conv1dProjection(self.proj_unit) 51 | Highway = FCHighwayNet(self.highway_layers) 52 | rnn_cell_fw = GRUCell(self.proj_unit[1]) 53 | rnn_cell_bw = GRUCell(self.proj_unit[1]) 54 | 55 | ### calculate 56 | # conv net 57 | output_0 = ConvBankWithPool(inputs, is_training) 58 | 59 | ### for correctness. 60 | if sequence_length is not None: 61 | output_0 = output_0 * mask 62 | 63 | output_1 = ConvProj(output_0, is_training) 64 | # residual connect 65 | res_output = tf.identity(inputs) + output_1 66 | 67 | # highway net 68 | highway_output = Highway(res_output) 69 | 70 | # biGRU 71 | # batch major 72 | final_output, *_ = bidirectional_dynamic_rnn(rnn_cell_fw, rnn_cell_bw, highway_output, sequence_length=sequence_length, time_major=False, dtype=tf.float32) 73 | final_output = tf.concat(final_output, axis=-1) 74 | if time_major: 75 | final_output = tf.transpose(final_output, perm=(1,0,2)) 76 | 77 | return final_output 78 | 79 | -------------------------------------------------------------------------------- /Modules/ConvNet.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from six.moves import xrange 3 | 4 | 5 | def __conv1d__(inputs, width, stride, in_channels, out_channels): 6 | return tf.layers.conv1d(inputs, out_channels, width, stride, 'SAME') 7 | 8 | def __conv1d_alone_time__(inputs, width, in_channels, out_channels): 9 | return __conv1d__(inputs, width, 1, in_channels, out_channels) 10 | 11 | class Conv1dBankWithMaxPool(object): 12 | """Conv1d Bank. 13 | The output is max_pooled along time. 14 | """ 15 | 16 | def __init__(self, K, activation=tf.nn.relu): 17 | self.__K = K 18 | self.__activation = activation 19 | 20 | @property 21 | def K(self): 22 | return self.__K 23 | 24 | @property 25 | def activation(self): 26 | return self.__activation 27 | 28 | def __call__(self, inputs, is_training=True, scope=None): 29 | """ 30 | Args: 31 | inputs: with shape -> (batch_size, time_step/width, units/channels) 32 | """ 33 | with tf.variable_scope(scope or type(self).__name__): 34 | in_channels = inputs.shape[-1].value 35 | conv_lst = [] 36 | for idk in xrange(1, self.K + 1): 37 | with tf.variable_scope('inner_conv_%d' % idk): 38 | conv_k = self.activation(__conv1d_alone_time__(inputs, idk, in_channels, in_channels)) 39 | norm_k = tf.contrib.layers.batch_norm(conv_k, is_training=is_training, updates_collections=None) 40 | conv_lst.append(norm_k) 41 | 42 | stacked_conv = tf.stack(conv_lst, axis=-1) # shape -> (batch_size, time_step/width, units/channels, K/height) 43 | re_shape = [tf.shape(stacked_conv)[0], tf.shape(stacked_conv)[1], 1, in_channels * self.K] 44 | stacked_conv = tf.reshape(stacked_conv, shape=re_shape) # shape -> (batch_size, time_step/width, 1, units*K/channels) 45 | 46 | ### max pool along time 47 | ksize = [1, 2, 1, 1] 48 | strid = [1, 1, 1, 1] 49 | pooled_conv = tf.squeeze(tf.nn.max_pool(stacked_conv, ksize, strid, 'SAME'), axis=2) # shape -> (batch_size, time_step/width, units*K/channels) 50 | 51 | return pooled_conv 52 | 53 | class Conv1dProjection(object): 54 | """Conv1d Projection 55 | """ 56 | 57 | def __init__(self, proj_unit, width=3, activation=tf.nn.relu): 58 | self.__proj_unit = proj_unit 59 | self.__width = width 60 | self.__activation = activation 61 | 62 | @property 63 | def proj_unit(self): 64 | return self.__proj_unit 65 | 66 | @property 67 | def width(self): 68 | return self.__width 69 | 70 | @property 71 | def activation(self): 72 | return self.__activation 73 | 74 | def __call__(self, inputs, is_training=True, scope=None): 75 | """ 76 | Args: 77 | inputs: with shape -> (batch_size, time_step/width, units/channels) 78 | """ 79 | with tf.variable_scope(scope or type(self).__name__): 80 | filter_width = self.width 81 | proj_0 = self.proj_unit[0] 82 | proj_1 = self.proj_unit[1] 83 | in_channels = inputs.get_shape()[-1].value 84 | with tf.variable_scope('inner_conv_with_acti'): 85 | conv_a = self.activation(__conv1d_alone_time__(inputs, filter_width, in_channels, proj_0)) 86 | norm_a = tf.contrib.layers.batch_norm(conv_a, is_training=is_training, updates_collections=None) 87 | with tf.variable_scope('inner_conv_linear'): 88 | conv_l = __conv1d_alone_time__(norm_a, filter_width, proj_0, proj_1) 89 | norm_l = tf.contrib.layers.batch_norm(conv_l, is_training=is_training, updates_collections=None) 90 | 91 | return norm_l 92 | 93 | -------------------------------------------------------------------------------- /Modules/ConvNetNOBN.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from six.moves import xrange 3 | 4 | 5 | def __conv1d__(inputs, width, stride, in_channels, out_channels): 6 | filter_1d = tf.get_variable(name='filter', shape=(width, in_channels, out_channels)) 7 | return tf.nn.conv1d(inputs, filter_1d, stride, 'SAME') 8 | 9 | def __conv1d_alone_time__(inputs, width, in_channels, out_channels): 10 | return __conv1d__(inputs, width, 1, in_channels, out_channels) 11 | 12 | class Conv1dBankWithMaxPool(object): 13 | """Conv1d Bank. 14 | The output is max_pooled along time. 15 | """ 16 | 17 | def __init__(self, K, activation=tf.nn.relu): 18 | self.__K = K 19 | self.__activation = activation 20 | 21 | @property 22 | def K(self): 23 | return self.__K 24 | 25 | @property 26 | def activation(self): 27 | return self.__activation 28 | 29 | def __call__(self, inputs, is_training=True, scope=None): 30 | """ 31 | Args: 32 | inputs: with shape -> (batch_size, time_step/width, units/channels) 33 | """ 34 | with tf.variable_scope(scope or type(self).__name__): 35 | in_channels = inputs.shape[-1].value 36 | conv_lst = [] 37 | for idk in xrange(1, self.K + 1): 38 | with tf.variable_scope('inner_conv_%d' % idk): 39 | conv_k = self.activation(__conv1d_alone_time__(inputs, idk, in_channels, in_channels)) 40 | conv_lst.append(conv_k) 41 | 42 | stacked_conv = tf.stack(conv_lst, axis=-1) # shape -> (batch_size, time_step/width, units/channels, K/height) 43 | #re_shape = tf.shape(stacked_conv)[:2] + [1, in_channels * self.K] 44 | re_shape = [tf.shape(stacked_conv)[0], tf.shape(stacked_conv)[1], 1, in_channels * self.K] 45 | stacked_conv = tf.reshape(stacked_conv, shape=re_shape) # shape -> (batch_size, time_step/width, 1, units*K/channels) 46 | 47 | ### max pool along time 48 | ksize = [1, 2, 1, 1] 49 | strid = [1, 1, 1, 1] 50 | pooled_conv = tf.squeeze(tf.nn.max_pool(stacked_conv, ksize, strid, 'SAME'), axis=2) # shape -> (batch_size, time_step/width, units*K/channels) 51 | 52 | return pooled_conv 53 | 54 | class Conv1dProjection(object): 55 | """Conv1d Projection 56 | """ 57 | 58 | def __init__(self, proj_unit, width=3, activation=tf.nn.relu): 59 | self.__proj_unit = proj_unit 60 | self.__width = width 61 | self.__activation = activation 62 | 63 | @property 64 | def proj_unit(self): 65 | return self.__proj_unit 66 | 67 | @property 68 | def width(self): 69 | return self.__width 70 | 71 | @property 72 | def activation(self): 73 | return self.__activation 74 | 75 | def __call__(self, inputs, is_training=True, scope=None): 76 | """ 77 | Args: 78 | inputs: with shape -> (batch_size, time_step/width, units/channels) 79 | """ 80 | with tf.variable_scope(scope or type(self).__name__): 81 | filter_width = self.width 82 | proj_0 = self.proj_unit[0] 83 | proj_1 = self.proj_unit[1] 84 | in_channels = inputs.get_shape()[-1].value 85 | with tf.variable_scope('inner_conv_with_acti'): 86 | conv_a = self.activation(__conv1d_alone_time__(inputs, filter_width, in_channels, proj_0)) 87 | with tf.variable_scope('inner_conv_linear'): 88 | conv_l = __conv1d_alone_time__(conv_a, filter_width, proj_0, proj_1) 89 | 90 | return conv_l 91 | 92 | -------------------------------------------------------------------------------- /Modules/FastCBHG.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from tensorflow.python.ops import array_ops 3 | from tensorflow.contrib.rnn import FusedRNNCellAdaptor, TimeReversedFusedRNN 4 | from TFCommon.RNNCell import FastGRUCell as GRUCell 5 | from Tacotron.Modules import ConvNet, HighwayNet 6 | 7 | bidirectional_dynamic_rnn = tf.nn.bidirectional_dynamic_rnn 8 | 9 | Conv1dBankWithMaxPool = ConvNet.Conv1dBankWithMaxPool 10 | Conv1dProjection = ConvNet.Conv1dProjection 11 | FCHighwayNet = HighwayNet.FCHighwayNet 12 | 13 | class CBHG(object): 14 | """CBHG Net 15 | """ 16 | 17 | def __init__(self, bank_K, proj_unit, highway_layers=4): 18 | """ 19 | Args: 20 | bank_K: int 21 | proj_unit: a pair of int 22 | """ 23 | self.__bank_K = bank_K 24 | self.__proj_unit = proj_unit 25 | self.__highway_layers = highway_layers 26 | 27 | @property 28 | def bank_K(self): 29 | return self.__bank_K 30 | 31 | @property 32 | def proj_unit(self): 33 | return self.__proj_unit 34 | 35 | @property 36 | def highway_layers(self): 37 | return self.__highway_layers 38 | 39 | def __call__(self, inputs, sequence_length=None, is_training=True, time_major=None): 40 | assert time_major is not None, "[*] You must specify whether is time_major or not!" 41 | if time_major: 42 | inputs = tf.transpose(inputs, perm=(1, 0, 2)) # Use batch major data. 43 | assert inputs.get_shape()[-1] == self.proj_unit[1], "[!] input's shape is not the same as ConvProj's output!" 44 | 45 | ### for correctness. 46 | if sequence_length is not None: 47 | mask = tf.expand_dims(array_ops.sequence_mask(sequence_length, tf.shape(inputs)[1], tf.float32), -1) 48 | inputs = inputs * mask 49 | 50 | ConvBankWithPool = Conv1dBankWithMaxPool(self.bank_K) 51 | ConvProj = Conv1dProjection(self.proj_unit) 52 | Highway = FCHighwayNet(self.highway_layers) 53 | cell = GRUCell(self.proj_unit[1]) 54 | fw_cell = FusedRNNCellAdaptor(cell) 55 | bw_cell = TimeReversedFusedRNN(fw_cell) 56 | 57 | ### calculate 58 | # conv net 59 | output_0 = ConvBankWithPool(inputs, is_training) 60 | 61 | ### for correctness. 62 | if sequence_length is not None: 63 | output_0 = output_0 * mask 64 | 65 | output_1 = ConvProj(output_0, is_training) 66 | # residual connect 67 | res_output = tf.identity(inputs) + output_1 68 | 69 | # highway net 70 | highway_output = Highway(res_output) 71 | 72 | # biGRU 73 | # time major 74 | bGRUinp = tf.transpose(highway_output, perm=(1, 0, 2)) 75 | fw_out, _ = fw_cell(bGRUinp, sequence_length=sequence_length, scope="fw", dtype=tf.float32) 76 | bw_out, _ = bw_cell(bGRUinp, sequence_length=sequence_length, scope="bw", dtype=tf.float32) 77 | final_output = tf.concat([fw_out, bw_out], axis=-1) 78 | 79 | if not time_major: 80 | final_output = tf.transpose(final_output, perm=(1,0,2)) 81 | 82 | return final_output 83 | 84 | -------------------------------------------------------------------------------- /Modules/HighwayNet.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from six.moves import xrange 3 | 4 | class FCHighwayNet(object): 5 | """Implements Highway Networks. 6 | 7 | Rupesh Kumar Srivastava, Klaus Greff, Ju ̈rgen Schmidhuber. 8 | "Highway Networks." 9 | https://arxiv.org/abs/1505.00387 10 | """ 11 | 12 | def __init__(self, layer_num, activation=tf.nn.relu): 13 | self.__layer_num = layer_num 14 | self.__activation = activation 15 | 16 | @property 17 | def layer_num(self): 18 | return self.__layer_num 19 | 20 | @property 21 | def activation(self): 22 | return self.__activation 23 | 24 | def __flow_layer(self, inputs, units): 25 | H = tf.layers.dense(name='H', inputs=inputs, units=units, activation=self.activation) 26 | T = tf.layers.dense(name='T', inputs=inputs, units=units, activation=tf.sigmoid) 27 | y = H * T + inputs * (1 - T) 28 | return y 29 | 30 | def __call__(self, inputs, scope=None): 31 | with tf.variable_scope(scope or type(self).__name__): 32 | x_l = inputs 33 | units = inputs.shape[-1].value 34 | for idx in xrange(1, self.layer_num + 1): 35 | with tf.variable_scope('inner_layer_%d' % idx): 36 | y_l = self.__flow_layer(x_l, units) 37 | x_l = y_l 38 | 39 | return y_l 40 | 41 | -------------------------------------------------------------------------------- /Modules/__init__.py: -------------------------------------------------------------------------------- 1 | __all__ = ["CBHG"] 2 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Tacotron 2 | TACOTRON: TOWARDS END-TO-END SPEECH SYNTHESIS 3 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | __all__ = ["Modules", "model"] 2 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from TFCommon.Model import Model 3 | from Tacotron.Modules.CBHG import CBHG 4 | from TFCommon.Attention import BahdanauAttentionModule as AttModule 5 | from TFCommon.RNNCell import GRUCell, ResidualWrapper 6 | from TFCommon.Layers import EmbeddingLayer 7 | 8 | class Tacotron(Model): 9 | """Tacotron 10 | """ 11 | def __init__(self, r=2, lambda_l1=0., is_training=True): 12 | super(Tacotron, self).__init__() 13 | self.__r = r 14 | self.lambda_l1 = lambda_l1 15 | self.__is_training = is_training 16 | 17 | @property 18 | def r(self): 19 | return self.__r 20 | 21 | @property 22 | def is_training(self): 23 | return self.__is_training 24 | 25 | def build_forward(self, inputs, outputs, embed_class, time_major=None): 26 | assert time_major is not None, "[*] You must specify whether is time_major or not!" 27 | if time_major: 28 | input_time_steps = tf.shape(inputs)[0] 29 | else: 30 | input_time_steps = tf.shape(inputs)[1] 31 | 32 | global_step = tf.Variable(0, name='global_step', trainable=False) 33 | self.global_step = global_step 34 | 35 | ### Encoder ### 36 | ### Embedding 37 | with tf.variable_scope('encoder'): 38 | embeded = EmbeddingLayer(embed_class, 256)(inputs, scope='chr-emb') 39 | with tf.variable_scope('pre-net'): 40 | pre_0 = tf.layers.dropout(tf.layers.dense(embeded, 256, tf.nn.relu)) 41 | pre_1 = tf.layers.dropout(tf.layers.dense(pre_0, 128, tf.nn.relu)) 42 | with tf.variable_scope('CBHG'): 43 | cbhg_net = CBHG(16, (128, 128)) 44 | cbhg_out = cbhg_net(pre_1, self.is_training, time_major) 45 | 46 | with tf.variable_scope('decoder'): 47 | with tf.variable_scope('att-memory'): 48 | att_module = AttModule(256, cbhg_out, time_major=time_major) 49 | att_rnn = GRUCell(256) 50 | dec_rnn_0 = GRUCell(256) 51 | dec_rnn_1 = ResidualWrapper(GRUCell(256)) 52 | ### prepare loop 53 | with tf.variable_scope('loop'): 54 | if not time_major: 55 | outputs = tf.transpose(outputs, perm=(1,0,2)) 56 | max_time_steps = tf.shape(outputs)[0] 57 | reduced_time_steps = tf.div(max_time_steps - 1, self.r) + 1 58 | batch_size = tf.shape(outputs)[1] 59 | output_dim = outputs.shape[-1].value 60 | pad_indic = tf.zeros(shape=(self.r, batch_size, output_dim), dtype=tf.float32) 61 | indic = tf.concat([pad_indic, outputs], axis=0) 62 | pred_indic = tf.zeros(shape=(batch_size, output_dim), dtype=tf.float32) 63 | att_rnn_state = att_rnn.init_state(batch_size, tf.float32) 64 | dec_rnn_0_state = dec_rnn_0.init_state(batch_size, tf.float32) 65 | dec_rnn_1_state = dec_rnn_1.init_state(batch_size, tf.float32) 66 | state_tup = tuple([att_rnn_state, dec_rnn_0_state, dec_rnn_1_state]) 67 | ### prepare tensor array 68 | output_ta = tf.TensorArray(size=reduced_time_steps, dtype=tf.float32) 69 | alpha_ta = tf.TensorArray(size=reduced_time_steps, dtype=tf.float32) 70 | 71 | time = tf.constant(0, dtype=tf.int32) 72 | cond = lambda time, *_: tf.less(time, reduced_time_steps) 73 | def body(time, pred_indic, output_ta, alpha_ta, state_tup): 74 | ### get indication 75 | if self.is_training: 76 | this_indic = indic[self.r * time] # 只用最后一帧 77 | else: 78 | this_indic = pred_indic 79 | ### pre-net 80 | with tf.variable_scope('pre-net'): 81 | pre_0 = tf.layers.dropout(tf.layers.dense(this_indic, 256, tf.nn.relu)) 82 | pre_1 = tf.layers.dropout(tf.layers.dense(pre_0, 128, tf.nn.relu)) 83 | with tf.variable_scope('att-rnn'): 84 | att_rnn_out, att_rnn_state = att_rnn(pre_1, state_tup[0]) 85 | with tf.variable_scope('att-query'): 86 | query = att_rnn_out 87 | context, alpha = att_module(query) 88 | alpha_ta = alpha_ta.write(time, alpha) 89 | with tf.variable_scope('decoder-rnn'): 90 | with tf.variable_scope('cell-0'): 91 | dec_rnn_0_inp = tf.concat([context, att_rnn_out], axis=-1) 92 | dec_rnn_0_out, dec_rnn_0_state = dec_rnn_0(dec_rnn_0_inp, state_tup[1]) 93 | with tf.variable_scope('cell-1'): 94 | dec_rnn_1_out, dec_rnn_1_state = dec_rnn_1(dec_rnn_0_out, state_tup[2]) 95 | with tf.variable_scope('dense-out'): 96 | dense_out = tf.layers.dense(dec_rnn_1_out, output_dim) 97 | out_mgc_lf0 = tf.reshape(\ 98 | tf.layers.dense(dec_rnn_1_out, (output_dim - 1) * self.r), 99 | shape=(batch_size, self.r, output_dim - 1)) 100 | out_vuv = tf.reshape(\ 101 | tf.layers.dense(dec_rnn_1_out, self.r, tf.sigmoid), 102 | shape=(batch_size, self.r, 1)) 103 | dense_out = tf.concat([out_mgc_lf0, out_vuv], axis=-1) 104 | output_ta = output_ta.write(time, dense_out) 105 | if not self.is_training: 106 | mgc_lf0_indic = out_mgc_lf0[:, -1] 107 | vuv_indic = tf.round(out_vuv[:, -1]) 108 | pred_indic = tf.concat([mgc_lf0_indic, vuv_indic], axis=-1) 109 | 110 | state_tup = tuple([att_rnn_state, dec_rnn_0_state, dec_rnn_1_state]) 111 | return tf.add(time, 1), pred_indic, output_ta, alpha_ta, state_tup 112 | 113 | ### run loop 114 | _, _, final_output_ta, final_alpha_ta, *_ = tf.while_loop(cond, body, [time, pred_indic, output_ta, alpha_ta, state_tup]) 115 | 116 | final_output = tf.reshape(final_output_ta.stack(), shape=(reduced_time_steps, batch_size, self.r, output_dim)) 117 | final_output = tf.reshape(tf.transpose(final_output, perm=(0, 2, 1, 3)), shape=(reduced_time_steps * self.r, batch_size, output_dim)) 118 | final_output = final_output[:max_time_steps] # time major 119 | final_alpha = tf.reshape(final_alpha_ta.stack(), shape=(reduced_time_steps, batch_size, input_time_steps)) 120 | final_alpha = tf.transpose(final_alpha, perm=(1, 0, 2)) # batch major 121 | 122 | 123 | self.pred_out = final_output 124 | self.alpha_img = tf.expand_dims(final_alpha, -1) 125 | 126 | self.loss_mgc_lf0 = tf.losses.mean_squared_error(outputs[:, :, :-1], final_output[:, :, :-1]) 127 | self.loss_vuv = tf.losses.sigmoid_cross_entropy(outputs[:, :, -1], final_output[:, :, -1]) 128 | l1_reg = tf.contrib.layers.l1_regularizer(self.lambda_l1) 129 | l1_loss_vars = [item for item in tf.trainable_variables() if "decoder-rnn" in item.name or "dense-out" in item.name] 130 | self.l1_loss = tf.contrib.layers.apply_regularization(l1_reg, l1_loss_vars) 131 | self.loss = self.loss_mgc_lf0 + self.loss_vuv + self.l1_loss 132 | 133 | def build_backprop(self): 134 | with tf.variable_scope("backprop"): 135 | self.lr_start = tf.constant(0.001) 136 | self.learning_rate = tf.Variable(self.lr_start, name='learning_rate', trainable=False) 137 | self.opt = tf.train.AdamOptimizer(self.learning_rate) 138 | self.upd = self.opt.minimize(self.loss, global_step=self.global_step) 139 | return self.upd, self.lr_schedule_op 140 | 141 | def lr_schedule_op(self): 142 | lr_stage_0 = self.lr_start 143 | lr_stage_1 = tf.constant(0.0005) 144 | lr_stage_2 = tf.constant(0.0003) 145 | lr_state_3 = tf.constant(0.0001) 146 | gate_0 = tf.constant(int(5e5), dtype=tf.int32) 147 | gate_1 = tf.constant(int(1e6), dtype=tf.int32) 148 | gate_2 = tf.constant(int(2e6), dtype=tf.int32) 149 | def f1(): return lr_stage_0 150 | def f2(): return lr_stage_1 151 | def f3(): return lr_stage_2 152 | def f4(): return lr_stage_3 153 | new_lr = case([(tf.less(self.global_step, gate_0), f1), (tf.less(self.global_step, gate_1), f2),\ 154 | (tf.less(self.global_step, gate_2), f3)], 155 | default=f4, exclusive=False) 156 | return self.learning_rate.assign(new_lr) 157 | 158 | def summary(self, suffix, num_img=2): 159 | sums = [] 160 | sums.append(tf.summary.scalar('%s/loss_mgc_lf0' % suffix, self.loss_mgc_lf0)) 161 | sums.append(tf.summary.scalar('%s/loss_vuv' % suffix, self.loss_vuv)) 162 | sums.append(tf.summary.scalar('%s/loss_l1' % suffix, self.l1_loss)) 163 | sums.append(tf.summary.scalar('%s/loss' % suffix, self.loss)) 164 | sums.append(tf.summary.image('%s/alpha' % suffix, self.alpha_img[:num_img])) 165 | merged = tf.summary.merge(sums) 166 | return merged 167 | 168 | 169 | --------------------------------------------------------------------------------