├── LICENSE ├── README.md ├── ptb_data ├── ptb.test.txt ├── ptb.train.txt └── ptb.valid.txt ├── reader.py ├── softmax.py └── train_lm.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 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 General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | {project} Copyright (C) {year} {fullname} 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # tf-adaptive-softmax-lstm-lm 2 | 3 | This repository shows the experiment result of LSTM language models on PTB (Penn Treebank) and GBW ([Google One Billion Word](https://code.google.com/archive/p/1-billion-word-language-modeling-benchmark/)) using **AdaptiveSoftmax** on TensorFlow. 4 | 5 | ## Adaptive Softmax 6 | 7 | The adaptive softmax is a faster way to train a softmax classifier over a huge number of classes, and can be used for **both training and prediction**. For example, it can be used for training a Language Model with a very huge vocabulary, and the trained languaed model can be used in speech recognition, text generation, and machine translation very efficiently. 8 | 9 | Tha adaptive softmax has been used in the ASR system developed by **Tencent AI Lab**, and achieved about **20x speed up** than full sotfmax in the second pass for rescoing. 10 | 11 | See [Efficient softmax approximation for GPUs](https://arxiv.org/pdf/1609.04309v2.pdf) for detail about the adaptive softmax algorithms. 12 | 13 | ## Implementation 14 | 15 | The implementation of AdaptiveSoftmax on TensorFlow can be found here: [TencentAILab/tensorflow](https://github.com/TencentAILab/tensorflow) 16 | 17 | ## Usage 18 | 19 | Train with AdaptiveSoftmax: 20 | ```shell 21 | python train_lm.py --data_path=ptb_data --gpuid=0 --use_adaptive_softmax=1 22 | ``` 23 | Train with full softmax: 24 | ```shell 25 | python train_lm.py --data_path=ptb_data --gpuid=0 --use_adaptive_softmax=0 26 | ``` 27 | 28 | ## Experiment results 29 | 30 | ### Language Modeling on PTB 31 | With the hyper parameters below, it takes 5min54s to train 20 epochs on PTB corpus, the final perplexity on test set 32 | is *88.51*. With the same parameters and using full softmax, it takes 6min57s to train 20 epochs, and the final perplexity on test set is *89.00*. 33 | 34 | Since the PTB vocabulary size is only 10K, the speed up is not that significant. 35 | 36 | 37 | **hyper parameters:** 38 | ```python 39 | epoch_num = 20 40 | train_batch_size = 128 41 | train_step_size = 20 42 | valid_batch_size = 128 43 | valid_step_size = 20 44 | test_batch_size = 20 45 | test_step_size = 1 46 | word_embedding_dim = 512 47 | lstm_layers = 1 48 | lstm_size = 512 49 | lstm_forget_bias = 0.0 50 | max_grad_norm = 0.25 51 | init_scale = 0.05 52 | learning_rate = 0.2 53 | decay = 0.5 54 | decay_when = 1.0 55 | dropout_prob = 0.5 56 | adagrad_eps = 1e-5 57 | vocab_size = 10001 58 | softmax_type = "AdaptiveSoftmax" 59 | adaptive_softmax_cutoff = [2000, vocab_size] 60 | ``` 61 | **result:** 62 | 63 | | Epoch | Elapse | Train PPL | Valid PPL | Test PPL | 64 | | ----- | ------ | ----------| --------- | -------- | 65 | | 1 | 0min18s | 376.407 | 169.152 | 164.039 | 66 | | 2 | 0min35s | 154.324 | 132.648 | 127.494 | 67 | | 3 | 0min53s | 117.210 | 118.547 | 113.197 | 68 | | 4 | 1min11s | 98.662 | 111.791 | 106.373 | 69 | | 5 | 1min28s | 87.366 | 107.808 | 102.588 | 70 | | 6 | 1min46s | 79.448 | 105.028 | 100.024 | 71 | | 7 | 2min04s | 73.749 | 103.705 | 98.220 | 72 | | 8 | 2min21s | 69.392 | 102.939 | 96.931 | 73 | | 9 | 2min39s | 62.737 | 100.174 | 94.043 | 74 | | 10 | 2min57s | 59.423 | 99.412 | 93.153 | 75 | | 11 | 3min15s | 56.634 | 97.600 | 91.271 | 76 | | 12 | 3min32s | 55.036 | 97.388 | 91.061 | 77 | | 13 | 3min50s | 54.002 | 96.127 | 89.796 | 78 | | 14 | 4min08s | 53.232 | 96.170 | 89.805 | 79 | | 15 | 4min25s | 52.844 | 95.461 | 89.130 | 80 | | 16 | 4min43s | 52.488 | 95.085 | 88.788 | 81 | | 17 | 5min01s | 52.314 | 94.905 | 88.615 | 82 | | 18 | 5min18s | 52.172 | 94.835 | 88.553 | 83 | | 19 | 5min36s | 52.038 | 94.806 | 88.526 | 84 | | 20 | 5min54s | 51.998 | 94.788 | 88.510 | 85 | 86 | 87 | ### Language Modeling on Google 1Billion Word corpus 88 | 89 | **hyper parameters:** 90 | ```python 91 | word_embedding_dim = 256 92 | train_batch_size = 256 93 | train_step_size = 20 94 | valid_batch_size = 256 95 | valid_step_size = 20 96 | test_batch_size = 128 97 | test_step_size = 1 98 | lstm_layers = 1 99 | lstm_size = 2048 100 | lstm_forget_bias = 1.0 101 | max_grad_norm = 0.25 102 | init_scale = 0.05 103 | learning_rate = 0.1 104 | decay = 0.5 105 | decay_when = 1.0 106 | dropout_prob = 0.01 107 | adagrad_eps = 1e-5 108 | vocab_size = 793471 109 | softmax_type = "AdaptiveSoftmax" 110 | adaptive_softmax_cutoff = [4000,40000,200000, vocab_size] 111 | ``` 112 | **result:** 113 | 114 | On GBW corpus, we achived a perplexcity of 43.24 after 5 epochs, taking about two days to train on 2 GPUs with synchronous gradient updates. 115 | 116 | | Epoch | Elapse | Train PPL | Valid PPL | Test PPL | 117 | | ----- | ------ | --------- | --------- | -------- | 118 | | 1 | 9h56min| 51.428 | 52.727 | 49.553 | 119 | | 2 |19h53min| 45.141 | 48.683 | 45.639 | 120 | | 3 |29h51min| 42.605 | 47.379 | 44.332 | 121 | | 4 |39h48min| 41.119 | 46.822 | 43.743 | 122 | | 5 |49h45min| 38.757 | 46.402 | 43.241 | 123 | | 6 |59h42min| 37.664 | 46.334 | 43.119 | 124 | | 7 |69h40min| 37.139 | 46.337 | 43.101 | 125 | | 8 |79h37min| 36.884 | 46.342 | 43.097 | 126 | 127 | ##Reference 128 | 129 | [1] Grave E, Joulin A, Cissé M, et al. Efficient softmax approximation for GPUs[J]. arXiv preprint arXiv:1609.04309, 2016. 130 | 131 | [2] https://github.com/facebookresearch/adaptive-softmax 132 | -------------------------------------------------------------------------------- /reader.py: -------------------------------------------------------------------------------- 1 | import os 2 | import time 3 | import numpy as np 4 | 5 | def INFO_LOG(info): 6 | print "[%s]%s" % (time.strftime("%Y-%m-%d %X", time.localtime()), info) 7 | 8 | class Vocab(object): 9 | def __init__(self): 10 | self.BOS = "" 11 | self.EOS = "" 12 | self.UNK = "" 13 | 14 | def buildFromFiles(self, files): 15 | if type(files) is not list: 16 | raise ValueError("buildFromFiles input type error") 17 | 18 | INFO_LOG("build vocabulary from files ...") 19 | self.word_cnt = {self.BOS: 0, self.EOS: 0} 20 | for _file in files: 21 | line_num = 0 22 | for line in open(_file): 23 | line_num += 1 24 | for w in line.strip().replace('', self.UNK).split(): 25 | if self.word_cnt.has_key(w): 26 | self.word_cnt[w] += 1 27 | else: 28 | self.word_cnt[w] = 1 29 | self.word_cnt[self.BOS] += line_num 30 | self.word_cnt[self.EOS] += line_num 31 | count_pairs = sorted(self.word_cnt.items(), key = lambda x: (-x[1], x[0])) 32 | self.words, _ = list(zip(*count_pairs)) 33 | self.word2id = dict(zip(self.words, range(len(self.words)))) 34 | self.UNK_ID = self.word2id[self.UNK] 35 | INFO_LOG("vocab size: {}".format(self.size())) 36 | 37 | def encode(self, sentence): 38 | return [self.word2id[w] if self.word2id.has_key(w) else self.UNK_ID for w in sentence] 39 | 40 | def decode(self, ids): 41 | return [self.words[_id] for _id in ids] 42 | 43 | def size(self): 44 | return len(self.words) 45 | 46 | class Reader(object): 47 | def __init__(self, data_path): 48 | self.train_file = os.path.join(data_path, 'ptb.train.txt') 49 | self.valid_file = os.path.join(data_path, 'ptb.valid.txt') 50 | self.test_file = os.path.join(data_path, 'ptb.test.txt') 51 | 52 | self.vocab = Vocab() 53 | self.vocab.buildFromFiles([self.train_file]) 54 | 55 | def getVocabSize(self): 56 | return self.vocab.size() 57 | 58 | def yieldSpliceBatch(self, tag, batch_size, step_size): 59 | eos_index = self.vocab.word2id[self.vocab.EOS] 60 | unk_index = self.vocab.word2id[self.vocab.UNK] 61 | if tag == 'Train': 62 | _file = self.train_file 63 | elif tag == 'Valid': 64 | _file = self.valid_file 65 | else: 66 | _file = self.test_file 67 | 68 | INFO_LOG("File: %s" % _file) 69 | data = [] 70 | line_num = 0 71 | for line in open(_file): 72 | tokens = line.strip().split() 73 | data += self.vocab.encode(tokens) + [eos_index] 74 | line_num += 1 75 | total_token = len(data) 76 | token_num = (total_token - line_num) 77 | 78 | data_len = len(data) 79 | batch_len = data_len // batch_size 80 | batch_num = (batch_len - 1) // step_size 81 | if batch_num == 0: 82 | raise ValueError("batch_num == 0, decrease batch_size or step_size") 83 | 84 | INFO_LOG(" {} sentence, {}/{} tokens with/out {}".format(line_num, total_token, token_num, self.vocab.EOS)) 85 | 86 | used_token = batch_num * batch_size * step_size 87 | INFO_LOG(" {} batches, {}*{}*{} = {}({:.2%}) tokens will be used".format(batch_num, 88 | batch_num, batch_size, step_size, used_token, float(used_token) / total_token)) 89 | 90 | word_data = np.zeros([batch_size, batch_len], dtype=np.int32) 91 | for j in range(batch_size): 92 | index = j * batch_len 93 | word_data[j] = data[index : index + batch_len] 94 | for batch_id in range(batch_num): 95 | index = step_size * batch_id 96 | x = word_data[:, index : index + step_size] 97 | y = word_data[:, index + 1 : index + step_size + 1] 98 | n = batch_size * step_size 99 | yield(batch_id, batch_num, x, y, n) 100 | 101 | -------------------------------------------------------------------------------- /softmax.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python -u 2 | #encoding=utf8 3 | #Author: yangsaiyong@gmail.com 4 | #Update: 2018.10.17 5 | 6 | import tensorflow as tf 7 | import numpy as np 8 | 9 | class FullSoftmax(object): 10 | def __init__(self, input_dim, vocab_size, initializer=None, name=None): 11 | with tf.variable_scope(name or type(self).__name__, initializer=initializer): 12 | self.softmax_w = tf.get_variable("softmax_w", [input_dim, vocab_size]) 13 | self.softmax_b = tf.get_variable("softmax_b", [vocab_size], \ 14 | initializer=tf.constant_initializer(0.0, dtype=tf.float32)) 15 | 16 | def loss(self, inputs, labels, name='loss'): 17 | logits = tf.matmul(inputs, self.softmax_w) + self.softmax_b 18 | loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels, name=name) 19 | return loss, [loss] 20 | 21 | def softmax(self, inputs, name='softmax'): 22 | logits = tf.matmul(inputs, self.softmax_w) + self.softmax_b 23 | return tf.nn.softmax(logits, name=name) 24 | 25 | def log_softmax(self, inputs, name='log_softmax'): 26 | logits = tf.matmul(inputs, self.softmax_w) + self.softmax_b 27 | return tf.nn.log_softmax(logits, name=name) 28 | 29 | class AdaptiveSoftmax(object): 30 | def __init__(self, input_dim, cutoff, project_factor=4, project_dims=None, initializer=None, name=None): 31 | self.cluster_num = len(cutoff) - 1 32 | if project_dims: 33 | assert(len(project_dims) == self.cluster_num) 34 | else: 35 | project_dims = [] 36 | tail_project_factor = project_factor 37 | for i in range(self.cluster_num): 38 | dim = max(1, input_dim / tail_project_factor) 39 | project_dims.append(dim) 40 | tail_project_factor *= project_factor 41 | 42 | self.cutoff = cutoff 43 | with tf.variable_scope(name or type(self).__name__, initializer=initializer): 44 | head_dim = cutoff[0] + self.cluster_num 45 | self.head_w = tf.get_variable("adaptive_softmax_head_w", [input_dim, head_dim]) 46 | 47 | self.tail_w = [] 48 | for i in range(self.cluster_num): 49 | project_dim = project_dims[i] 50 | tail_dim = cutoff[i + 1] - cutoff[i] 51 | self.tail_w.append([ 52 | tf.get_variable("adaptive_softmax_tail{}_proj_w".format(i+1), [input_dim, project_dim]), 53 | tf.get_variable("adaptive_softmax_tail{}_w".format(i+1), [project_dim, tail_dim]) 54 | ]) 55 | 56 | def loss(self, inputs, labels, name='loss'): 57 | # Get tail masks and update head labels 58 | training_losses = [] 59 | head_labels = labels 60 | ones = tf.ones([tf.size(labels)], dtype=tf.int32) 61 | for i in range(self.cluster_num): 62 | mask = tf.logical_and(tf.greater_equal(labels, self.cutoff[i]), tf.less(labels, self.cutoff[i + 1])) 63 | 64 | # Update head labels 65 | head_labels = tf.where(mask, ones * (self.cutoff[0] + i), head_labels) 66 | 67 | # Compute tail loss 68 | tail_inputs = tf.boolean_mask(inputs, mask) 69 | tail_logits = tf.matmul(tf.matmul(tail_inputs, self.tail_w[i][0]), self.tail_w[i][1]) 70 | tail_labels = tf.boolean_mask(labels - self.cutoff[i], mask) 71 | tail_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=tail_logits, labels=tail_labels) 72 | training_losses.append(tail_loss) 73 | aligned_tail_loss = tf.SparseTensor(tf.squeeze(tf.where(mask)), tail_loss, [tf.size(labels, out_type=tf.int64)]) 74 | loss = tf.sparse_tensor_to_dense(aligned_tail_loss) if i == 0 else \ 75 | loss + tf.sparse_tensor_to_dense(aligned_tail_loss) 76 | 77 | # Compute head loss 78 | head_logits = tf.matmul(inputs, self.head_w) # (sample_num, head_size) 79 | head_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=head_logits, labels=head_labels) #(sample_num) 80 | training_losses.append(head_loss) 81 | loss = tf.add(loss, head_loss, name=name) 82 | 83 | return loss, training_losses 84 | 85 | def softmax(self, inputs, name='softmax'): 86 | head_logits = tf.matmul(inputs, self.head_w) 87 | head_softmax = tf.nn.softmax(head_logits) 88 | softmax_list = [head_softmax[:, :self.cutoff[0]]] 89 | for i in range(self.cluster_num): 90 | tail_logits = tf.matmul(tf.matmul(inputs, self.tail_w[i][0]), self.tail_w[i][1]) 91 | tail_softmax = tf.nn.softmax(tail_logits) 92 | index = self.cutoff[0] + i 93 | softmax_list.append(tail_softmax * head_softmax[:, index:index+1]) 94 | return tf.concat(softmax_list, axis=1, name=name) 95 | 96 | def log_softmax(self, inputs, name='log_softmax'): 97 | head_logits = tf.matmul(inputs, self.head_w) 98 | head_logsoftmax = tf.nn.log_softmax(head_logits) 99 | logsoftmax_list = [head_logsoftmax[:, :self.cutoff[0]]] 100 | for i in range(self.cluster_num): 101 | tail_logits = tf.matmul(tf.matmul(inputs, self.tail_w[i][0]), self.tail_w[i][1]) 102 | tail_logsoftmax = tf.nn.log_softmax(tail_logits) 103 | index = self.cutoff[0] + i 104 | logsoftmax_list.append(tail_logsoftmax + head_logsoftmax[:, index:index+1]) 105 | return tf.concat(logsoftmax_list, axis=1, name=name) 106 | 107 | -------------------------------------------------------------------------------- /train_lm.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python -u 2 | #encoding=utf8 3 | #Author: yangsaiyong@gmail.com 4 | #Update: 2018.10.17 5 | 6 | import os 7 | import time 8 | 9 | import numpy as np 10 | import tensorflow as tf 11 | 12 | from tensorflow.contrib.rnn import BasicLSTMCell 13 | from tensorflow.contrib.rnn import MultiRNNCell 14 | from tensorflow.contrib.rnn import DropoutWrapper 15 | 16 | from reader import * 17 | import softmax 18 | 19 | flags = tf.flags 20 | flags.DEFINE_string("data_path", "ptb_data", "Where the training/test data is stored.") 21 | flags.DEFINE_bool("use_adaptive_softmax", True, "Train using adaptive softmax") 22 | flags.DEFINE_integer("gpuid", 0, "GPU ID") 23 | 24 | FLAGS = flags.FLAGS 25 | 26 | def parse_device(gpuid): 27 | os.environ["CUDA_VISIBLE_DEVICES"] = "%s" % gpuid 28 | print "Use GPU:" 29 | print "device {} => /gpu:{}".format(gpuid, 0) 30 | return "/gpu:0" 31 | 32 | class LSTMLM(object): 33 | def __init__(self, config, mode, device, reuse=None): 34 | self.config = config 35 | self.mode = mode 36 | if mode == "Train": 37 | self.is_training = True 38 | self.batch_size = self.config.train_batch_size 39 | self.step_size = self.config.train_step_size 40 | elif mode == "Valid": 41 | self.is_training = False 42 | self.batch_size = self.config.valid_batch_size 43 | self.step_size = self.config.valid_step_size 44 | else: 45 | self.is_training = False 46 | self.batch_size = self.config.test_batch_size 47 | self.step_size = self.config.test_step_size 48 | 49 | vocab_size = config.vocab_size 50 | embed_dim = config.word_embedding_dim 51 | lstm_size = config.lstm_size 52 | lstm_layers = config.lstm_layers 53 | lstm_forget_bias = config.lstm_forget_bias 54 | batch_size = self.batch_size 55 | step_size = self.step_size 56 | 57 | with tf.device(device), tf.name_scope(mode), tf.variable_scope("LSTMLM", reuse=reuse): 58 | # INPUTS and TARGETS 59 | self.inputs = tf.placeholder(tf.int32, [batch_size, step_size]) 60 | self.targets = tf.placeholder(tf.int32, [batch_size, step_size]) 61 | 62 | # Inititial state 63 | self.initial_state = tf.placeholder(tf.float32, 64 | [batch_size, lstm_size * 2 * lstm_layers]) 65 | 66 | # WORD EMBEDDING 67 | stdv = np.sqrt(1. / vocab_size) 68 | self.word_embedding = tf.get_variable("word_embedding", [ 69 | vocab_size, embed_dim], initializer=tf.random_uniform_initializer(-stdv, stdv)) 70 | inputs = tf.nn.embedding_lookup(self.word_embedding, self.inputs) 71 | 72 | # INPUT DROPOUT 73 | if self.is_training and self.config.dropout_prob > 0: 74 | inputs = tf.nn.dropout(inputs, keep_prob=1 - config.dropout_prob) 75 | 76 | # LSTM 77 | lstm_cell = BasicLSTMCell(lstm_size, forget_bias=lstm_forget_bias, state_is_tuple=False) 78 | if self.is_training and config.dropout_prob > 0: 79 | lstm_cell = DropoutWrapper(lstm_cell, output_keep_prob=1. - config.dropout_prob) 80 | cell = MultiRNNCell([lstm_cell] * lstm_layers, state_is_tuple=False) 81 | 82 | #inputs = tf.unstack(inputs, axis=1) 83 | output, self.final_state = tf.nn.dynamic_rnn(cell, inputs, initial_state=self.initial_state) 84 | 85 | #output = tf.reshape(tf.concat(1, outputs), [-1, lstm_size]) 86 | output = tf.reshape(output, [-1, lstm_size]) 87 | 88 | # Softmax & loss 89 | if config.softmax_type == 'AdaptiveSoftmax': 90 | cutoff = config.adaptive_softmax_cutoff 91 | softmax_layer = softmax.AdaptiveSoftmax(lstm_size, cutoff) 92 | else: 93 | softmax_layer = softmax.FullSoftmax(lstm_size, vocab_size) 94 | self.loss, training_losses = softmax_layer.loss(output, tf.reshape(self.targets, [-1]), 'loss') 95 | 96 | self.cost = tf.reduce_sum(self.loss) 97 | 98 | if self.is_training: 99 | self.lr = tf.Variable(0.0, trainable=False) 100 | optimizer = tf.train.AdagradOptimizer(self.lr, config.adagrad_eps) 101 | losses = [tf.reduce_sum(loss) / batch_size for loss in training_losses] 102 | trainable_vars = tf.trainable_variables() 103 | grads = tf.gradients(losses, trainable_vars) 104 | grads = [tf.clip_by_norm(grad, config.max_grad_norm) \ 105 | if grad is not None else grad for grad in grads] 106 | self.eval_op = optimizer.apply_gradients(zip(grads, trainable_vars)) 107 | else: 108 | self.eval_op = tf.no_op() 109 | 110 | def update_lr(self, session, learning_rate): 111 | if self.is_training: 112 | session.run(tf.assign(self.lr, learning_rate)) 113 | 114 | def get_initial_state(self): 115 | return np.zeros([self.batch_size, self.config.lstm_size * 2 * self.config.lstm_layers], dtype=np.float32) 116 | 117 | 118 | class Config(object): 119 | epoch_num = 20 120 | train_batch_size = 128 121 | train_step_size = 20 122 | valid_batch_size = 128 123 | valid_step_size = 20 124 | test_batch_size = 20 125 | test_step_size = 1 126 | word_embedding_dim = 512 127 | lstm_layers = 1 128 | lstm_size = 512 129 | lstm_forget_bias = 0.0 130 | max_grad_norm = 0.25 131 | init_scale = 0.05 132 | learning_rate = 0.2 133 | decay = 0.5 134 | decay_when = 1.0 135 | dropout_prob = 0.5 136 | adagrad_eps = 1e-5 137 | vocab_size = 10001 138 | softmax_type = "AdaptiveSoftmax" 139 | adaptive_softmax_cutoff = [2000, vocab_size] 140 | 141 | class LearningRateUpdater(object): 142 | def __init__(self, init_lr, decay_rate, decay_when): 143 | self._init_lr = init_lr 144 | self._decay_rate = decay_rate 145 | self._decay_when = decay_when 146 | self._current_lr = init_lr 147 | self._last_ppl = -1 148 | 149 | def get_lr(self): 150 | return self._current_lr 151 | 152 | def update(self, cur_ppl): 153 | if self._last_ppl > 0 and self._last_ppl - cur_ppl < self._decay_when: 154 | current_lr = self._current_lr * self._decay_rate 155 | INFO_LOG("learning rate: {} ==> {}".format(self._current_lr, current_lr)) 156 | self._current_lr = current_lr 157 | self._last_ppl = cur_ppl 158 | 159 | def run(session, model, reader, verbose=True): 160 | state = model.get_initial_state() 161 | total_cost = 0 162 | total_word_cnt = 0 163 | start_time = time.time() 164 | 165 | for batch in reader.yieldSpliceBatch(model.mode, model.batch_size, model.step_size): 166 | batch_id, batch_num, x, y, word_cnt = batch 167 | feed = {model.inputs: x, model.targets:y, model.initial_state: state} 168 | cost, state, _ = session.run([model.cost, model.final_state, model.eval_op], feed) 169 | total_cost += cost 170 | total_word_cnt += word_cnt 171 | if verbose and (batch_id % max(10, batch_num//10)) == 0: 172 | ppl = np.exp(total_cost / total_word_cnt) 173 | wps = total_word_cnt / (time.time() - start_time) 174 | print " [%5d/%d]ppl: %.3f speed: %.0f wps costs %.3f words %d" % ( 175 | batch_id, batch_num, ppl, wps, total_cost, total_word_cnt) 176 | return total_cost, total_word_cnt, np.exp(total_cost / total_word_cnt) 177 | 178 | def main(_): 179 | reader = Reader(FLAGS.data_path) 180 | config = Config() 181 | 182 | if FLAGS.use_adaptive_softmax: 183 | config.softmax_type = 'AdaptiveSoftmax' 184 | else: 185 | config.softmax_type = 'FullSoftmax' 186 | 187 | device = parse_device(FLAGS.gpuid) 188 | 189 | lr_updater = LearningRateUpdater(config.learning_rate, config.decay, config.decay_when) 190 | 191 | graph = tf.Graph() 192 | with graph.as_default(): 193 | trainm = LSTMLM(config, device=device, mode="Train", reuse=False) 194 | validm = LSTMLM(config, device=device, mode="Valid", reuse=True) 195 | testm = LSTMLM(config, device=device, mode="Test", reuse=True) 196 | 197 | session_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False) 198 | session_config.gpu_options.allow_growth = True 199 | with tf.Session(graph=graph, config=session_config) as session: 200 | session.run(tf.global_variables_initializer()) 201 | for epoch in range(config.epoch_num): 202 | trainm.update_lr(session, lr_updater.get_lr()) 203 | INFO_LOG("Epoch {}, learning rate: {}".format(epoch + 1, lr_updater.get_lr())) 204 | cost, word_cnt, ppl = run(session, trainm, reader) 205 | INFO_LOG("Epoch %d Train perplexity %.3f words %d" % (epoch + 1, ppl, word_cnt)) 206 | 207 | cost, word_cnt, ppl = run(session, validm, reader) 208 | INFO_LOG("Epoch %d Valid perplexity %.3f words %d" % (epoch + 1, ppl, word_cnt)) 209 | 210 | lr_updater.update(ppl) 211 | cost, word_cnt, ppl = run(session, testm, reader) 212 | INFO_LOG("Epoch %d Test perplexity %.3f words %d" % (epoch + 1, ppl, word_cnt)) 213 | 214 | if __name__ == '__main__': 215 | tf.app.run() 216 | 217 | --------------------------------------------------------------------------------