├── .gitignore ├── LICENSE ├── README.md ├── download_data.bash ├── download_model.bash ├── img └── talkdown.png ├── requirements.txt └── src ├── bert.py └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | *.egg-info/ 24 | .installed.cfg 25 | *.egg 26 | MANIFEST 27 | 28 | # PyInstaller 29 | # Usually these files are written by a python script from a template 30 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 31 | *.manifest 32 | *.spec 33 | 34 | # Installer logs 35 | pip-log.txt 36 | pip-delete-this-directory.txt 37 | 38 | # Unit test / coverage reports 39 | htmlcov/ 40 | .tox/ 41 | .coverage 42 | .coverage.* 43 | .cache 44 | nosetests.xml 45 | coverage.xml 46 | *.cover 47 | .hypothesis/ 48 | .pytest_cache/ 49 | 50 | # Translations 51 | *.mo 52 | *.pot 53 | 54 | # Django stuff: 55 | *.log 56 | local_settings.py 57 | db.sqlite3 58 | 59 | # Flask stuff: 60 | instance/ 61 | .webassets-cache 62 | 63 | # Scrapy stuff: 64 | .scrapy 65 | 66 | # Sphinx documentation 67 | docs/_build/ 68 | 69 | # PyBuilder 70 | target/ 71 | 72 | # Jupyter Notebook 73 | .ipynb_checkpoints 74 | 75 | # pyenv 76 | .python-version 77 | 78 | # celery beat schedule file 79 | celerybeat-schedule 80 | 81 | # SageMath parsed files 82 | *.sage.py 83 | 84 | # Environments 85 | .env 86 | .venv 87 | env/ 88 | venv/ 89 | ENV/ 90 | env.bak/ 91 | venv.bak/ 92 | 93 | # Spyder project settings 94 | .spyderproject 95 | .spyproject 96 | 97 | # Rope project settings 98 | .ropeproject 99 | 100 | # mkdocs documentation 101 | /site 102 | 103 | # mypy 104 | .mypy_cache/ 105 | 106 | .DS_Store 107 | data 108 | runs 109 | .idea 110 | models 111 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # TalkDown TalkDown: A Corpus for Condescension Detection in Context 2 | 3 | 4 | ## Introduction 5 | 6 | This is the code release for the paper [TalkDown: A Corpus for Condescension Detection in Context](https://www.aclweb.org/anthology/D19-1385) by Zijian Wang and Christopher Potts in proceedings of EMNLP-IJCNLP 2019. 7 | 8 | ## Dependencies 9 | ### Python dependencies 10 | Run `pip install -r requirements.txt`. This codebase requires Python version >= 3.6. 11 | ### Data 12 | Run `bash download_data.bash` to download and uncompress the TalkDown dataset. Or you could use this [link](https://nlp.stanford.edu/~zijwang/talkdown/talkdown.tar.gz). 13 | ### Pretrained model (optional) 14 | Run `bash download_model.bash` to download our best pretrained model to reproduce the result. It is not required if you want to train your model from scratch. 15 | 16 | 17 | ## Sample commands for training and evaluation 18 | 19 | ### Train 20 | You could train a BERT model using the following command. 21 | ``` 22 | python -m src.bert --do_train --use_quoted --use_context --output_dir test 23 | ``` 24 | ### Evaluate 25 | You could evaluate your model using the following command. This command also reproduces our best result in the paper (make sure you have downloaded the pretrained model). 26 | ``` 27 | python -m src.bert --do_eval --use_quoted --use_context --eval_on_test --output_dir pretrained_full 28 | ``` 29 | which should return `Model's F1 is 0.6835111677776263` 30 | ## Citation 31 | 32 | @inproceedings{wang2019talkdown, 33 | author = {Wang, Zijian and Potts, Christopher} 34 | title = {{TalkDown}: A Corpus for Condescension Detection in Context}, 35 | booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing}, 36 | url = {https://www.aclweb.org/anthology/D19-1385}, 37 | year = {2019} 38 | } 39 | 40 | ## Contact 41 | 42 | You may reach out us at zijwang@stanford.edu and cgpotts@stanford.edu. 43 | -------------------------------------------------------------------------------- /download_data.bash: -------------------------------------------------------------------------------- 1 | curl https://nlp.stanford.edu/~zijwang/talkdown/talkdown.tar.gz -o talkdown.tar.gz 2 | mkdir -p data 3 | tar xzf talkdown.tar.gz -C data/ 4 | rm talkdown.tar.gz 5 | echo 'Done!' 6 | -------------------------------------------------------------------------------- /download_model.bash: -------------------------------------------------------------------------------- 1 | curl https://nlp.stanford.edu/~zijwang/talkdown/pretrained_full.tar.gz -o pretrained_full.tar.gz 2 | tar xzf pretrained_full.tar.gz 3 | rm pretrained_full.tar.gz 4 | echo 'Done!' 5 | -------------------------------------------------------------------------------- /img/talkdown.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/zijwang/talkdown/85c44e779273a410f0b8700ac17b0f80272966f1/img/talkdown.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | torch>=1.1.0 2 | pytorch_transformers>=1.1.0 3 | imbalanced-learn==0.4.3 4 | pandas 5 | numpy 6 | gputil 7 | tensorboardX 8 | tqdm 9 | scikit-learn -------------------------------------------------------------------------------- /src/bert.py: -------------------------------------------------------------------------------- 1 | # author: Zijian Wang 2 | # many pieces of code were adapted from `pytorch_transformer` repo 3 | 4 | import argparse 5 | import glob 6 | import json 7 | import sys 8 | 9 | from pytorch_transformers import AdamW, WarmupLinearSchedule 10 | from pytorch_transformers import (WEIGHTS_NAME, BertConfig, 11 | BertForSequenceClassification, BertTokenizer, 12 | RobertaConfig, 13 | RobertaForSequenceClassification, 14 | RobertaTokenizer, 15 | XLMConfig, XLMForSequenceClassification, 16 | XLMTokenizer, XLNetConfig, 17 | XLNetForSequenceClassification, 18 | XLNetTokenizer) 19 | from sklearn.metrics import f1_score 20 | from tensorboardX import SummaryWriter 21 | from torch.utils.data import TensorDataset, DataLoader, SequentialSampler 22 | from tqdm import trange, tqdm 23 | 24 | from .utils import * 25 | 26 | import warnings 27 | 28 | # skip numpy future warning 29 | warnings.simplefilter(action='ignore', category=FutureWarning) 30 | 31 | logging.basicConfig(format='%(asctime)s - %(name)s - %(message)s', 32 | datefmt='%m/%d/%Y %H:%M:%S', 33 | level=logging.INFO) 34 | logger = logging.getLogger(__name__) 35 | 36 | # List all model types here from `pytorch_transformer`. Only bert was tested. 37 | ALL_MODELS = sum( 38 | (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), 39 | ()) 40 | MODEL_CLASSES = { 41 | 'bert': (BertConfig, BertForSequenceClassification, BertTokenizer), 42 | 'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer), 43 | 'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer), 44 | 'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer), 45 | } 46 | 47 | 48 | 49 | def get_dataloader(args, processor, label_list, tokenizer, is_train=False): 50 | if is_train: 51 | logger.info('getting train dataloader') 52 | examples = processor.get_train_examples(args.data_dir, args.train_file, 53 | sampling_strategy=args.sampling_strategy) 54 | else: 55 | logger.info(f'getting {"test" if args.eval_on_test else "dev"} dataloader') 56 | if args.eval_on_test: 57 | examples = processor.get_test_examples(args.data_dir, args.test_file) 58 | else: 59 | examples = processor.get_dev_examples(args.data_dir, args.dev_file) 60 | 61 | features = convert_examples_to_features( 62 | examples, label_list, args.max_seq_length, tokenizer, output_mode='classification', 63 | cls_token_at_end=bool(args.model_type in ['xlnet']), # xlnet has a cls token at the end 64 | cls_token=tokenizer.cls_token, 65 | cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0, 66 | sep_token=tokenizer.sep_token, 67 | sep_token_extra=bool(args.model_type in ['roberta']), 68 | # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805 69 | pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet 70 | pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0], 71 | pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0, 72 | ) 73 | all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) 74 | all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long) 75 | all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long) 76 | all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long) 77 | data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 78 | # Run prediction for full data 79 | if is_train: 80 | sampler = SortedBatchSampler([sum(i) for i in all_input_mask.numpy()], args.train_batch_size) 81 | else: 82 | sampler = SequentialSampler(data) 83 | 84 | dataloader = DataLoader(data, sampler=sampler, 85 | batch_size=max(args.n_gpu, 1) * args.per_gpu_train_batch_size if is_train 86 | else max(args.n_gpu, 1) * args.per_gpu_eval_batch_size, 87 | num_workers=args.n_gpu * 2, pin_memory=True, drop_last=True if is_train else False) 88 | return dataloader 89 | 90 | 91 | def evaluate(eval_dataloader, model, args): 92 | labels = [] 93 | preds = [] 94 | 95 | for batch in tqdm(eval_dataloader): 96 | batch = tuple(t.to(args.device) for t in batch) 97 | inputs = {'input_ids': batch[0], 98 | 'attention_mask': batch[1], 99 | 'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None, 100 | # XLM and RoBERTa don't use segment_ids 101 | 'labels': batch[3]} 102 | 103 | with torch.no_grad(): 104 | outputs = model(**inputs) 105 | _, logits = outputs[:2] 106 | 107 | logits = logits.detach().cpu().numpy() 108 | label_ids = batch[3].to('cpu').numpy() 109 | pred = np.argmax(logits, axis=1) 110 | labels.append(label_ids) 111 | preds.append(pred) 112 | 113 | f1 = f1_score(np.concatenate(labels), np.concatenate(preds), average="macro") 114 | return f1 115 | 116 | def main(): 117 | parser = argparse.ArgumentParser() 118 | 119 | ## Required parameters 120 | parser.add_argument("--data_dir", default='./data/', type=str, required=False, 121 | help="The input data dir. Should contain the jsonl files for the task.") 122 | parser.add_argument("--model_type", default='bert', type=str, required=False, 123 | help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())) 124 | parser.add_argument("--model_name_or_path", default='bert-base-cased', type=str, required=False, 125 | help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join( 126 | ALL_MODELS)) 127 | parser.add_argument("--task_name", default='test', type=str, required=False) 128 | parser.add_argument("--output_dir", default=None, type=str, required=False, 129 | help="The output directory where the model predictions and checkpoints will be written.") 130 | 131 | ## Other parameters 132 | parser.add_argument("--config_name", default="", type=str, 133 | help="Pretrained config name or path if not the same as model_name") 134 | parser.add_argument("--tokenizer_name", default="", type=str, 135 | help="Pretrained tokenizer name or path if not the same as model_name") 136 | parser.add_argument("--cache_dir", default="", type=str, 137 | help="Where do you want to store the pre-trained models downloaded from s3") 138 | parser.add_argument("--max_seq_length", default=512, type=int, 139 | help="The maximum total input sequence length after tokenization. Sequences longer " 140 | "than this will be truncated, sequences shorter will be padded.") 141 | parser.add_argument("--do_train", action='store_true', 142 | help="Whether to run training.") 143 | parser.add_argument("--do_eval", action='store_true', 144 | help="Whether to run eval on the dev set.") 145 | parser.add_argument("--do_lower_case", action='store_true', 146 | help="Set this flag if you are using an uncased model.") 147 | parser.add_argument("--per_gpu_train_batch_size", default=2, type=int, 148 | help="Batch size per GPU/CPU for training.") 149 | parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, 150 | help="Batch size per GPU/CPU for evaluation.") 151 | parser.add_argument('--gradient_accumulation_steps', type=int, default=1, 152 | help="Number of updates steps to accumulate before performing a backward/update pass.") 153 | parser.add_argument("--learning_rate", default=1e-5, type=float, 154 | help="The initial learning rate for Adam.") 155 | parser.add_argument("--weight_decay", default=0.0, type=float, 156 | help="Weight deay if we apply some.") 157 | parser.add_argument("--adam_epsilon", default=1e-8, type=float, 158 | help="Epsilon for Adam optimizer.") 159 | parser.add_argument("--max_grad_norm", default=1.0, type=float, 160 | help="Max gradient norm.") 161 | parser.add_argument("--num_train_epochs", default=3.0, type=float, 162 | help="Total number of training epochs to perform.") 163 | parser.add_argument("--max_steps", default=-1, type=int, 164 | help="If > 0: set total number of training steps to perform. Override num_train_epochs.") 165 | parser.add_argument("--warmup_steps", default=400, type=int, 166 | help="Linear warmup over warmup_steps.") 167 | 168 | parser.add_argument('--logging_steps', type=int, default=50, 169 | help="Log every X updates steps.") 170 | parser.add_argument('--save_steps', type=int, default=1000, 171 | help="Save checkpoint every X updates steps.") 172 | parser.add_argument("--eval_all_checkpoints", action='store_true', 173 | help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number") 174 | parser.add_argument("--no_cuda", action='store_true', 175 | help="Avoid using CUDA when available") 176 | parser.add_argument('--overwrite_output_dir', action='store_true', 177 | help="Overwrite the content of the output directory") 178 | parser.add_argument('--overwrite_cache', action='store_true', 179 | help="Overwrite the cached training and evaluation sets") 180 | parser.add_argument('--seed', type=int, default=42, 181 | help="Random seed for initialization") 182 | 183 | parser.add_argument('--fp16', action='store_true', 184 | help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") 185 | parser.add_argument('--fp16_opt_level', type=str, default='O1', 186 | help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." 187 | "See details at https://nvidia.github.io/apex/amp.html") 188 | 189 | parser.add_argument("--eval_on_test", action='store_true', help="Whether to evaluate on test (final) or dev set") 190 | 191 | parser.add_argument("--use_quoted", 192 | action='store_true', 193 | help="Whether to use quoted part as dataset") 194 | 195 | parser.add_argument("--use_context", 196 | action='store_true', 197 | help="Whether to use context part as dataset") 198 | 199 | parser.add_argument("--tag", 200 | default='exp_0', 201 | type=str, 202 | help="The name of this experiment") 203 | 204 | parser.add_argument("--sampling_strategy", type=float, default=-1, 205 | help='The oversampling ratio for training dataset. 1 means oversampling to balance and -1 means ' 206 | 'no oversampling') 207 | 208 | parser.add_argument("--train_file", type=str, default="imbalanced_train.jsonl") 209 | parser.add_argument("--dev_file", type=str, default="imbalanced_dev.jsonl") 210 | parser.add_argument("--test_file", type=str, default="imbalanced_test.jsonl") 211 | 212 | args = parser.parse_args() 213 | 214 | if args.do_train and args.output_dir is None: 215 | setattr(args, 'output_dir', os.path.join('models', args.task_name, args.tag)) 216 | 217 | if args.do_train: 218 | logger.info(f"Model will be saved to {args.output_dir}") 219 | else: 220 | logger.info(f"Model will be loaded to {args.output_dir}") 221 | 222 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 223 | args.device = device 224 | args.n_gpu = torch.cuda.device_count() 225 | 226 | logger.warning(f"Device: {device}, n_gpu: {args.n_gpu}, 16-bits training: {args.fp16}") 227 | 228 | set_seed(args) 229 | 230 | if not os.path.exists(args.output_dir) or os.listdir(args.output_dir): 231 | os.makedirs(args.output_dir, exist_ok=True) 232 | 233 | processor = CondProcessor(use_quoted=args.use_quoted, use_context=args.use_context) 234 | label_list = processor.get_labels() 235 | num_labels = len(label_list) 236 | args.model_type = args.model_type.lower() 237 | 238 | tb_writer = SummaryWriter(logdir=f'runs/{args.task_name}_{args.tag}') 239 | 240 | args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) 241 | args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps 242 | args.output_mode = 'classification' 243 | 244 | config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type] 245 | config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, 246 | num_labels=num_labels, finetuning_task=args.task_name) 247 | tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, 248 | do_lower_case=args.do_lower_case) 249 | model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), 250 | config=config) 251 | model.to(device) 252 | 253 | eval_f1s = [] 254 | 255 | if args.do_train: 256 | 257 | # get data 258 | train_dataloader = get_dataloader(args, processor, label_list, tokenizer, is_train=True) 259 | 260 | # get training setting 261 | 262 | t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs 263 | 264 | param_optimizer = list(model.named_parameters()) 265 | no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] 266 | optimizer_grouped_parameters = [ 267 | {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01}, 268 | {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} 269 | ] 270 | 271 | optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) 272 | scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total) 273 | 274 | if args.fp16: 275 | try: 276 | from apex import amp 277 | except ImportError: 278 | raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") 279 | model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) 280 | 281 | if args.n_gpu > 1: 282 | model = torch.nn.DataParallel(model) 283 | 284 | logger.info("***** Running training *****") 285 | logger.info(" Num examples = %d", len(train_dataloader.dataset)) 286 | logger.info(" Num Epochs = %d", args.num_train_epochs) 287 | logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) 288 | logger.info(" Total train batch size = %d", 289 | args.train_batch_size * args.gradient_accumulation_steps) 290 | logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) 291 | logger.info(" Total optimization steps = %d", t_total) 292 | 293 | global_step = 0 294 | tr_loss, logging_loss = 0.0, 0.0 295 | 296 | model.zero_grad() 297 | 298 | # reset seed before training 299 | set_seed(args) 300 | 301 | # train 302 | for _ in trange(int(args.num_train_epochs), desc="Epoch"): 303 | model.train() 304 | 305 | for step, batch in tqdm(enumerate(train_dataloader), total=len(train_dataloader), mininterval=5): 306 | 307 | min_len = torch.min(torch.tensor(args.max_seq_length), torch.max(torch.sum(batch[1], dim=1)) + 2) 308 | batch = [ 309 | t.to(device, non_blocking=True) if len(t.shape) == 1 else t[:, :min_len].contiguous().to(device, 310 | non_blocking=True) 311 | for t in batch] 312 | 313 | inputs = {'input_ids': batch[0], 314 | 'attention_mask': batch[1], 315 | 'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None, 316 | # XLM don't use segment_ids 317 | 'labels': batch[3]} 318 | outputs = model(**inputs) 319 | 320 | loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc) 321 | 322 | if args.n_gpu > 1: 323 | loss = loss.mean() # mean() to average on multi-gpu parallel training 324 | if args.gradient_accumulation_steps > 1: 325 | loss = loss / args.gradient_accumulation_steps 326 | 327 | if args.fp16: 328 | with amp.scale_loss(loss, optimizer) as scaled_loss: 329 | scaled_loss.backward() 330 | torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) 331 | else: 332 | loss.backward() 333 | torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) 334 | 335 | tr_loss += loss.item() 336 | 337 | if (step + 1) % args.gradient_accumulation_steps == 0: 338 | optimizer.step() 339 | scheduler.step() # Update learning rate schedule 340 | 341 | model.zero_grad() 342 | global_step += 1 343 | 344 | if args.logging_steps > 0 and global_step % args.logging_steps == 0: 345 | # Log metrics 346 | tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) 347 | tb_writer.add_scalar('loss', (tr_loss - logging_loss) / args.logging_steps, global_step) 348 | logging_loss = tr_loss 349 | 350 | if args.save_steps > 0 and global_step % args.save_steps == 0: 351 | # Save model checkpoint 352 | output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) 353 | if not os.path.exists(output_dir): 354 | os.makedirs(output_dir) 355 | model_to_save = model.module if hasattr(model, 356 | 'module') else model # Take care of parallel training 357 | model_to_save.save_pretrained(output_dir) 358 | torch.save(args, os.path.join(output_dir, 'training_args.bin')) 359 | logger.info("Saving model checkpoint to %s", output_dir) 360 | 361 | tb_writer.close() 362 | 363 | output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) 364 | if not os.path.exists(output_dir): 365 | os.makedirs(output_dir) 366 | model_to_save = model.module if hasattr(model, 367 | 'module') else model # Take care of distributed/parallel training 368 | model_to_save.save_pretrained(output_dir) 369 | torch.save(args, os.path.join(output_dir, 'training_args.bin')) 370 | logger.info("Saving model checkpoint to %s", output_dir) 371 | 372 | if args.do_eval: 373 | eval_dataloader = get_dataloader(args, processor, label_list, tokenizer, is_train=False) 374 | 375 | if args.eval_all_checkpoints: 376 | best_f1 = -1 377 | checkpoints = list(os.path.dirname(c) for c in 378 | sorted(glob.glob(args.output_dir + '/checkpoint*/' + WEIGHTS_NAME, recursive=True))) 379 | logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging 380 | logger.info("Evaluate the following checkpoints: %s", checkpoints) 381 | 382 | for checkpoint in checkpoints: 383 | model = model_class.from_pretrained(checkpoint) 384 | model.to(device) 385 | model.eval() 386 | if args.n_gpu > 1: 387 | model = torch.nn.DataParallel(model) 388 | 389 | f1 = evaluate(eval_dataloader, model, args) 390 | 391 | logger.info(f"{checkpoint}'s F1 is {f1}") 392 | eval_f1s.append(f1) 393 | if f1 > best_f1: 394 | 395 | best_f1 = f1 396 | logger.info("Best F1 %s" % best_f1) 397 | result = {'eval_f1': f1, 398 | 'ckpt': checkpoint} 399 | 400 | output_eval_file = os.path.join(args.output_dir, "eval_results.txt") 401 | with open(output_eval_file, "w") as writer: 402 | logger.info("***** Eval results *****") 403 | for key in sorted(result.keys()): 404 | logger.info(" %s = %s", key, str(result[key])) 405 | writer.write("%s = %s\n" % (key, str(result[key]))) 406 | 407 | model_to_save = model.module if hasattr(model, 'module') else model # 408 | output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME) 409 | torch.save(model_to_save.state_dict(), output_model_file) 410 | output_param_file = os.path.join(args.output_dir, "param") 411 | with open(output_param_file, 'w') as f: 412 | json.dump(args.__dict__, f, indent=2, sort_keys=True) 413 | 414 | else: 415 | model = model_class.from_pretrained(args.output_dir) 416 | model.to(device) 417 | model.eval() 418 | if args.n_gpu > 1: 419 | model = torch.nn.DataParallel(model) 420 | f1 = evaluate(eval_dataloader, model, args) 421 | logger.info(f"Model's F1 is {f1}") 422 | 423 | if __name__ == "__main__": 424 | main() 425 | -------------------------------------------------------------------------------- /src/utils.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import random 4 | from collections import Counter 5 | 6 | import GPUtil 7 | import numpy as np 8 | import pandas as pd 9 | import torch 10 | from imblearn.over_sampling import RandomOverSampler 11 | from torch.utils.data import Sampler 12 | 13 | logging.basicConfig(format='%(asctime)s - %(name)s - %(message)s', 14 | datefmt='%m/%d/%Y %H:%M:%S', 15 | level=logging.INFO) 16 | logger = logging.getLogger(__name__) 17 | 18 | 19 | class InputExample(object): 20 | """A single training/test example for simple sequence classification.""" 21 | 22 | def __init__(self, guid, text_a, text_b=None, label=None, extra=None): 23 | """Constructs a InputExample. 24 | Args: 25 | guid: Unique id for the example. 26 | text_a: string. The untokenized text of the first sequence. For single 27 | sequence tasks, only this sequence must be specified. 28 | text_b: (Optional) string. The untokenized text of the second sequence. 29 | Only must be specified for sequence pair tasks. 30 | label: (Optional) string. The label of the example. This should be 31 | specified for train and dev examples, but not for test examples. 32 | """ 33 | self.guid = guid 34 | self.text_a = text_a 35 | self.text_b = text_b 36 | self.label = label 37 | self.extra = extra 38 | 39 | 40 | class InputFeatures(object): 41 | """A single set of features of data.""" 42 | 43 | def __init__(self, input_ids, input_mask, segment_ids, label_id): 44 | self.input_ids = input_ids 45 | self.input_mask = input_mask 46 | self.segment_ids = segment_ids 47 | self.label_id = label_id 48 | 49 | 50 | def _truncate_seq_pair(tokens_a, tokens_b, max_length): 51 | """Truncates a sequence pair in place to the maximum length.""" 52 | 53 | # This is a simple heuristic which will always truncate the longer sequence 54 | # one token at a time. This makes more sense than truncating an equal percent 55 | # of tokens from each, since if one sequence is very short then each token 56 | # that's truncated likely contains more information than a longer sequence. 57 | if len(tokens_a) > max_length: 58 | tokens_a = tokens_a[:max_length] 59 | tokens_b = [] 60 | return tokens_a, tokens_b 61 | total_length = len(tokens_a) + len(tokens_b) 62 | if total_length > max_length: 63 | tokens_b = tokens_b[total_length - max_length:] 64 | while True: 65 | cnt = 0 66 | if len(tokens_b) and tokens_b[0] not in ["?", "!", ".", ",", ";", ""]: 67 | tokens_b.pop(0) 68 | cnt += 1 69 | else: 70 | break 71 | if cnt > 10: 72 | print(tokens_b) 73 | return tokens_a, tokens_b 74 | 75 | 76 | class CondProcessor: 77 | """Processor for the condescension dataset""" 78 | 79 | def __init__(self, use_quoted=True, use_context=False): 80 | assert use_quoted or use_context 81 | self.use_quoted = use_quoted 82 | self.use_context = use_context 83 | 84 | def get_train_examples(self, data_dir, filename, sampling_strategy=-1): 85 | """See base class.""" 86 | logger.info("Get train data") 87 | return self._create_examples( 88 | self._read_jsonl(os.path.join(data_dir, filename)), "train", sampling_strategy=sampling_strategy) 89 | 90 | def get_dev_examples(self, data_dir, filename): 91 | """See base class.""" 92 | logger.info("Get dev data") 93 | return self._create_examples( 94 | self._read_jsonl(os.path.join(data_dir, filename)), "dev") 95 | 96 | def get_test_examples(self, data_dir, filename): 97 | """See base class.""" 98 | logger.info("Get test data") 99 | return self._create_examples( 100 | self._read_jsonl(os.path.join(data_dir, filename)), "test") 101 | 102 | @classmethod 103 | def get_labels(cls): 104 | """See base class.""" 105 | return ["0", "1"] 106 | 107 | @classmethod 108 | def _read_jsonl(cls, input_file): 109 | """Reads a tab separated value file.""" 110 | logger.debug("trying to load pickle file %s" % input_file) 111 | df = pd.read_json(input_file, orient='records', lines=True) 112 | return df 113 | 114 | def _create_examples(self, df, set_type, sampling_strategy=-1): 115 | """Creates examples for the training and dev sets.""" 116 | examples = [] 117 | if set_type == "train": 118 | cnt = Counter(df.label) 119 | if cnt[True] != cnt[False]: 120 | logger.info(f'training dataset: {cnt}') 121 | if sampling_strategy == -1: 122 | logger.info('no oversampling') 123 | pass 124 | else: 125 | logger.info(f"setting sampling strategy to {sampling_strategy}") 126 | ros = RandomOverSampler(random_state=42, sampling_strategy=sampling_strategy) 127 | ids, _ = ros.fit_resample(np.arange(len(df)).reshape(-1, 1), df.label) 128 | df = df.iloc[ids.reshape(-1)] 129 | logger.info(f'Now training dataset: {Counter(df.label)}') 130 | for idx, row in enumerate(df.itertuples()): 131 | if self.use_quoted: 132 | text_a = row.quotedpost 133 | text_b = row.post[:row.start_offset] if self.use_context else None 134 | else: 135 | text_a = row.post[:row.start_offset] 136 | text_b = None 137 | 138 | label = 1 if row.label is True else 0 139 | guid = "%s-%s" % (set_type, idx) 140 | 141 | examples.append( 142 | InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) 143 | 144 | return examples 145 | 146 | 147 | def convert_examples_to_features(examples, label_list, max_seq_length, 148 | tokenizer, output_mode='classification', 149 | cls_token_at_end=False, 150 | cls_token='[CLS]', 151 | cls_token_segment_id=1, 152 | sep_token='[SEP]', 153 | sep_token_extra=False, 154 | pad_on_left=False, 155 | pad_token=0, 156 | pad_token_segment_id=0, 157 | sequence_a_segment_id=0, 158 | sequence_b_segment_id=1, 159 | mask_padding_with_zero=True): 160 | """ Loads a data file into a list of `InputBatch`s 161 | `cls_token_at_end` define the location of the CLS token: 162 | - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP] 163 | - True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS] 164 | `cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet) 165 | """ 166 | 167 | label_map = {label: i for i, label in enumerate(label_list)} 168 | 169 | features = [] 170 | for (ex_index, example) in enumerate(examples): 171 | if ex_index % 10000 == 0: 172 | logger.info("Writing example %d of %d" % (ex_index, len(examples))) 173 | 174 | tokens_a = tokenizer.tokenize(example.text_a) 175 | 176 | tokens_b = None 177 | if example.text_b: 178 | tokens_b = tokenizer.tokenize(example.text_b) 179 | # Modifies `tokens_a` and `tokens_b` in place so that the total 180 | # length is less than the specified length. 181 | # Account for [CLS], [SEP], [SEP] with "- 3". " -4" for RoBERTa. 182 | special_tokens_count = 4 if sep_token_extra else 3 183 | tokens_a, tokens_b = _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - special_tokens_count) 184 | else: 185 | # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa. 186 | special_tokens_count = 3 if sep_token_extra else 2 187 | if len(tokens_a) > max_seq_length - special_tokens_count: 188 | tokens_a = tokens_a[:(max_seq_length - special_tokens_count)] 189 | 190 | # The convention in BERT is: 191 | # (a) For sequence pairs: 192 | # tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP] 193 | # type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1 194 | # (b) For single sequences: 195 | # tokens: [CLS] the dog is hairy . [SEP] 196 | # type_ids: 0 0 0 0 0 0 0 197 | # 198 | # Where "type_ids" are used to indicate whether this is the first 199 | # sequence or the second sequence. The embedding vectors for `type=0` and 200 | # `type=1` were learned during pre-training and are added to the wordpiece 201 | # embedding vector (and position vector). This is not *strictly* necessary 202 | # since the [SEP] token unambiguously separates the sequences, but it makes 203 | # it easier for the model to learn the concept of sequences. 204 | # 205 | # For classification tasks, the first vector (corresponding to [CLS]) is 206 | # used as as the "sentence vector". Note that this only makes sense because 207 | # the entire model is fine-tuned. 208 | tokens = tokens_a + [sep_token] 209 | if sep_token_extra: 210 | # roberta uses an extra separator b/w pairs of sentences 211 | tokens += [sep_token] 212 | segment_ids = [sequence_a_segment_id] * len(tokens) 213 | 214 | if tokens_b: 215 | tokens += tokens_b + [sep_token] 216 | segment_ids += [sequence_b_segment_id] * (len(tokens_b) + 1) 217 | 218 | if cls_token_at_end: 219 | tokens = tokens + [cls_token] 220 | segment_ids = segment_ids + [cls_token_segment_id] 221 | else: 222 | tokens = [cls_token] + tokens 223 | segment_ids = [cls_token_segment_id] + segment_ids 224 | 225 | input_ids = tokenizer.convert_tokens_to_ids(tokens) 226 | 227 | # The mask has 1 for real tokens and 0 for padding tokens. Only real 228 | # tokens are attended to. 229 | input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids) 230 | 231 | # Zero-pad up to the sequence length. 232 | padding_length = max_seq_length - len(input_ids) 233 | if pad_on_left: 234 | input_ids = ([pad_token] * padding_length) + input_ids 235 | input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask 236 | segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids 237 | else: 238 | input_ids = input_ids + ([pad_token] * padding_length) 239 | input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length) 240 | segment_ids = segment_ids + ([pad_token_segment_id] * padding_length) 241 | 242 | # print(len(input_ids), max_seq_length) 243 | assert len(input_ids) == max_seq_length 244 | assert len(input_mask) == max_seq_length 245 | assert len(segment_ids) == max_seq_length 246 | 247 | if output_mode == "classification": 248 | label_id = label_map[str(example.label)] 249 | elif output_mode == "regression": 250 | label_id = float(example.label) 251 | else: 252 | raise KeyError(output_mode) 253 | 254 | features.append( 255 | InputFeatures(input_ids=input_ids, 256 | input_mask=input_mask, 257 | segment_ids=segment_ids, 258 | label_id=label_id)) 259 | return features 260 | 261 | 262 | def set_seed(args): 263 | logger.info(f'setting seed to {args.seed}') 264 | random.seed(args.seed) 265 | np.random.seed(args.seed) 266 | torch.manual_seed(args.seed) 267 | if args.n_gpu > 0: 268 | torch.cuda.manual_seed_all(args.seed) 269 | 270 | 271 | class SortedBatchSampler(Sampler): 272 | 273 | def __init__(self, lengths, batch_size, shuffle=True): 274 | self.lengths = lengths 275 | self.batch_size = batch_size 276 | self.shuffle = shuffle 277 | 278 | def __iter__(self): 279 | lengths = np.array( 280 | [(-l, np.random.random()) for l in self.lengths], 281 | dtype=[('l1', np.int_), ('rand', np.float_)] 282 | ) 283 | indices = np.argsort(lengths, order=('l1', 'rand')) 284 | batches = [indices[i:i + self.batch_size] 285 | for i in range(0, len(indices), self.batch_size)] 286 | if self.shuffle: 287 | np.random.shuffle(batches) 288 | return iter([i for batch in batches for i in batch]) 289 | 290 | def __len__(self): 291 | return len(self.lengths) 292 | 293 | 294 | def gpu_util(): 295 | """ 296 | As of Aug. 24th, 2019, the official GPUtil package (if installed using `pip install gputil`) does not come with 297 | the correct functionality to show GPU util with customized `attrList`. You may want to download from github and 298 | install from the source code. 299 | """ 300 | 301 | GPUtil.showUtilization(attrList=[[{'attr': 'id', 'name': 'ID'}, 302 | {'attr': 'name', 'name': 'Name', 'transform': lambda x: x.replace("GeForce", "")}, 303 | {'attr': 'load', 'name': 'GPU util.', 'suffix': '%', 304 | 'transform': lambda x: x * 100, 'precision': 0}, 305 | {'attr': 'memoryUtil', 'name': 'Mem. util.', 'suffix': '%', 306 | 'transform': lambda x: x * 100, 'precision': 0}], 307 | [{'attr': 'memoryTotal', 'name': 'Mem. total', 'suffix': 'MB', 'precision': 0}, 308 | {'attr': 'memoryUsed', 'name': 'Mem. used', 'suffix': 'MB', 'precision': 0}, 309 | {'attr': 'memoryFree', 'name': 'Mem. free', 'suffix': 'MB', 'precision': 0}]] 310 | ) 311 | --------------------------------------------------------------------------------