├── .gitignore ├── LICENSE ├── README.md ├── attention.py ├── base_model.py ├── classifier.py ├── data ├── train_ids.pkl └── val_ids.pkl ├── dataset.py ├── fc.py ├── language_model.py ├── main.py ├── tools ├── compute_softscore.py ├── create_dictionary.py ├── detection_features_converter.py ├── download.sh └── process.sh ├── train.py └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | data 2 | *.pyc 3 | *.ipynb 4 | logs/ 5 | new_logs/ 6 | task.sh 7 | *.npy 8 | *.pth 9 | new_logs* 10 | *.txt 11 | models/ 12 | saved_models/ 13 | -------------------------------------------------------------------------------- /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 | 635 | Copyright (C) 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 | Copyright (C) 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 | ## Bottom-Up and Top-Down Attention for Visual Question Answering 2 | 3 | An efficient PyTorch implementation of the winning entry of the [2017 VQA Challenge](http://www.visualqa.org/challenge.html). 4 | 5 | The implementation follows the VQA system described in "Bottom-Up and 6 | Top-Down Attention for Image Captioning and Visual Question Answering" 7 | (https://arxiv.org/abs/1707.07998) and "Tips and Tricks for Visual 8 | Question Answering: Learnings from the 2017 Challenge" 9 | (https://arxiv.org/abs/1708.02711). 10 | 11 | ## Results 12 | 13 | | Model | Validation Accuracy | Training Time 14 | | --- | --- | -- | 15 | | Reported Model | 63.15 | 12 - 18 hours (Tesla K40) | 16 | | Implemented Model | **63.58** | 40 - 50 minutes (Titan Xp) | 17 | 18 | The accuracy was calculated using the [VQA evaluation metric](http://www.visualqa.org/evaluation.html). 19 | 20 | ## About 21 | 22 | This is part of a project done at CMU for the course 11-777 23 | Advanced Multimodal Machine Learning and a joint work between Hengyuan Hu, 24 | Alex Xiao, and Henry Huang. 25 | 26 | As part of our project, we implemented bottom up attention as a strong VQA baseline. We were planning to integrate object 27 | detection with VQA and were very glad to see that Peter Anderson and 28 | Damien Teney et al. had already done that beautifully. 29 | We hope this clean and 30 | efficient implementation can serve as a useful baseline for future VQA 31 | explorations. 32 | 33 | ## Implementation Details 34 | 35 | Our implementation follows the overall structure of the papers but with 36 | the following simplifications: 37 | 38 | 1. We don't use extra data from [Visual Genome](http://visualgenome.org/). 39 | 2. We use only a fixed number of objects per image (K=36). 40 | 3. We use a simple, single stream classifier without pre-training. 41 | 4. We use the simple ReLU activation instead of gated tanh. 42 | 43 | The first two points greatly reduce the training time. Our 44 | implementation takes around 200 seconds per epoch on a single Titan Xp while 45 | the one described in the paper takes 1 hour per epoch. 46 | 47 | The third point is simply because we feel the two stream classifier 48 | and pre-training in the original paper is over-complicated and not 49 | necessary. 50 | 51 | For the non-linear activation unit, we tried gated tanh but couldn't 52 | make it work. We also tried gated linear unit (GLU) and it works better than 53 | ReLU. Eventually we choose ReLU due to its simplicity and since the gain 54 | from using GLU is too small to justify the fact that GLU doubles the 55 | number of parameters. 56 | 57 | With these simplifications we would expect the performance to drop. For 58 | reference, the best result on validation set reported in the paper is 59 | 63.15. The reported result without extra data from visual genome is 60 | 62.48, the result using only 36 objects per image is 62.82, the result 61 | using two steam classifier but not pre-trained is 62.28 and the result 62 | using ReLU is 61.63. These numbers are cited from the Table 1 of the 63 | paper: "Tips and Tricks for Visual Question Answering: Learnings from 64 | the 2017 Challenge". With all the above simplification aggregated, our 65 | first implementation got around 59-60 on validation set. 66 | 67 | To shrink the gap, we added some simple but powerful 68 | modifications. Including: 69 | 70 | 1. Add dropout to alleviate overfitting 71 | 2. Double the number of neurons 72 | 3. Add weight normalization (BN seems not work well here) 73 | 4. Switch to Adamax optimizer 74 | 5. Gradient clipping 75 | 76 | These small modifications bring the number back to ~62.80. We further 77 | change the concatenation based attention module in the original paper 78 | to a projection based module. This new attention module is inspired by 79 | the paper "Modeling Relationships in Referential Expressions with 80 | Compositional Modular Networks" 81 | (https://arxiv.org/pdf/1611.09978.pdf), but with some modifications 82 | (implemented in attention.NewAttention). With 83 | the help of this new attention, we boost the performance to ~63.58, 84 | surpassing the reported best result with no extra data and less 85 | computation cost. 86 | 87 | ## Usage 88 | 89 | #### Prerequisites 90 | 91 | Make sure you are on a machine with a NVIDIA GPU and Python 2 with about 70 GB disk space. 92 | 93 | 1. Install [PyTorch v0.3](http://pytorch.org/) with CUDA and Python 2.7. 94 | 2. Install [h5py](http://docs.h5py.org/en/latest/build.html). 95 | 96 | #### Data Setup 97 | 98 | All data should be downloaded to a 'data/' directory in the root 99 | directory of this repository. 100 | 101 | The easiest way to download the data is to run the provided script 102 | `tools/download.sh` from the repository root. The features are 103 | provided by and downloaded from the original authors' 104 | [repo](https://github.com/peteanderson80/bottom-up-attention). If the 105 | script does not work, it should be easy to examine the script and 106 | modify the steps outlined in it according to your needs. Then run 107 | `tools/process.sh` from the repository root to process the data to the 108 | correct format. 109 | 110 | #### Training 111 | 112 | Simply run `python main.py` to start training. The training and 113 | validation scores will be printed every epoch, and the best model will 114 | be saved under the directory "saved_models". The default flags should 115 | give you the result provided in the table above. 116 | -------------------------------------------------------------------------------- /attention.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.nn.utils.weight_norm import weight_norm 4 | from fc import FCNet 5 | 6 | 7 | class Attention(nn.Module): 8 | def __init__(self, v_dim, q_dim, num_hid): 9 | super(Attention, self).__init__() 10 | self.nonlinear = FCNet([v_dim + q_dim, num_hid]) 11 | self.linear = weight_norm(nn.Linear(num_hid, 1), dim=None) 12 | 13 | def forward(self, v, q): 14 | """ 15 | v: [batch, k, vdim] 16 | q: [batch, qdim] 17 | """ 18 | logits = self.logits(v, q) 19 | w = nn.functional.softmax(logits, 1) 20 | return w 21 | 22 | def logits(self, v, q): 23 | num_objs = v.size(1) 24 | q = q.unsqueeze(1).repeat(1, num_objs, 1) 25 | vq = torch.cat((v, q), 2) 26 | joint_repr = self.nonlinear(vq) 27 | logits = self.linear(joint_repr) 28 | return logits 29 | 30 | 31 | class NewAttention(nn.Module): 32 | def __init__(self, v_dim, q_dim, num_hid, dropout=0.2): 33 | super(NewAttention, self).__init__() 34 | 35 | self.v_proj = FCNet([v_dim, num_hid]) 36 | self.q_proj = FCNet([q_dim, num_hid]) 37 | self.dropout = nn.Dropout(dropout) 38 | self.linear = weight_norm(nn.Linear(q_dim, 1), dim=None) 39 | 40 | def forward(self, v, q): 41 | """ 42 | v: [batch, k, vdim] 43 | q: [batch, qdim] 44 | """ 45 | logits = self.logits(v, q) 46 | w = nn.functional.softmax(logits, 1) 47 | return w 48 | 49 | def logits(self, v, q): 50 | batch, k, _ = v.size() 51 | v_proj = self.v_proj(v) # [batch, k, qdim] 52 | q_proj = self.q_proj(q).unsqueeze(1).repeat(1, k, 1) 53 | joint_repr = v_proj * q_proj 54 | joint_repr = self.dropout(joint_repr) 55 | logits = self.linear(joint_repr) 56 | return logits 57 | -------------------------------------------------------------------------------- /base_model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from attention import Attention, NewAttention 4 | from language_model import WordEmbedding, QuestionEmbedding 5 | from classifier import SimpleClassifier 6 | from fc import FCNet 7 | 8 | 9 | class BaseModel(nn.Module): 10 | def __init__(self, w_emb, q_emb, v_att, q_net, v_net, classifier): 11 | super(BaseModel, self).__init__() 12 | self.w_emb = w_emb 13 | self.q_emb = q_emb 14 | self.v_att = v_att 15 | self.q_net = q_net 16 | self.v_net = v_net 17 | self.classifier = classifier 18 | 19 | def forward(self, v, b, q, labels): 20 | """Forward 21 | 22 | v: [batch, num_objs, obj_dim] 23 | b: [batch, num_objs, b_dim] 24 | q: [batch_size, seq_length] 25 | 26 | return: logits, not probs 27 | """ 28 | w_emb = self.w_emb(q) 29 | q_emb = self.q_emb(w_emb) # [batch, q_dim] 30 | 31 | att = self.v_att(v, q_emb) 32 | v_emb = (att * v).sum(1) # [batch, v_dim] 33 | 34 | q_repr = self.q_net(q_emb) 35 | v_repr = self.v_net(v_emb) 36 | joint_repr = q_repr * v_repr 37 | logits = self.classifier(joint_repr) 38 | return logits 39 | 40 | 41 | def build_baseline0(dataset, num_hid): 42 | w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0) 43 | q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0) 44 | v_att = Attention(dataset.v_dim, q_emb.num_hid, num_hid) 45 | q_net = FCNet([num_hid, num_hid]) 46 | v_net = FCNet([dataset.v_dim, num_hid]) 47 | classifier = SimpleClassifier( 48 | num_hid, 2 * num_hid, dataset.num_ans_candidates, 0.5) 49 | return BaseModel(w_emb, q_emb, v_att, q_net, v_net, classifier) 50 | 51 | 52 | def build_baseline0_newatt(dataset, num_hid): 53 | w_emb = WordEmbedding(dataset.dictionary.ntoken, 300, 0.0) 54 | q_emb = QuestionEmbedding(300, num_hid, 1, False, 0.0) 55 | v_att = NewAttention(dataset.v_dim, q_emb.num_hid, num_hid) 56 | q_net = FCNet([q_emb.num_hid, num_hid]) 57 | v_net = FCNet([dataset.v_dim, num_hid]) 58 | classifier = SimpleClassifier( 59 | num_hid, num_hid * 2, dataset.num_ans_candidates, 0.5) 60 | return BaseModel(w_emb, q_emb, v_att, q_net, v_net, classifier) 61 | -------------------------------------------------------------------------------- /classifier.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from torch.nn.utils.weight_norm import weight_norm 3 | 4 | 5 | class SimpleClassifier(nn.Module): 6 | def __init__(self, in_dim, hid_dim, out_dim, dropout): 7 | super(SimpleClassifier, self).__init__() 8 | layers = [ 9 | weight_norm(nn.Linear(in_dim, hid_dim), dim=None), 10 | nn.ReLU(), 11 | nn.Dropout(dropout, inplace=True), 12 | weight_norm(nn.Linear(hid_dim, out_dim), dim=None) 13 | ] 14 | self.main = nn.Sequential(*layers) 15 | 16 | def forward(self, x): 17 | logits = self.main(x) 18 | return logits 19 | -------------------------------------------------------------------------------- /dataset.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import os 3 | import json 4 | import cPickle 5 | import numpy as np 6 | import utils 7 | import h5py 8 | import torch 9 | from torch.utils.data import Dataset 10 | 11 | 12 | class Dictionary(object): 13 | def __init__(self, word2idx=None, idx2word=None): 14 | if word2idx is None: 15 | word2idx = {} 16 | if idx2word is None: 17 | idx2word = [] 18 | self.word2idx = word2idx 19 | self.idx2word = idx2word 20 | 21 | @property 22 | def ntoken(self): 23 | return len(self.word2idx) 24 | 25 | @property 26 | def padding_idx(self): 27 | return len(self.word2idx) 28 | 29 | def tokenize(self, sentence, add_word): 30 | sentence = sentence.lower() 31 | sentence = sentence.replace(',', '').replace('?', '').replace('\'s', ' \'s') 32 | words = sentence.split() 33 | tokens = [] 34 | if add_word: 35 | for w in words: 36 | tokens.append(self.add_word(w)) 37 | else: 38 | for w in words: 39 | tokens.append(self.word2idx[w]) 40 | return tokens 41 | 42 | def dump_to_file(self, path): 43 | cPickle.dump([self.word2idx, self.idx2word], open(path, 'wb')) 44 | print('dictionary dumped to %s' % path) 45 | 46 | @classmethod 47 | def load_from_file(cls, path): 48 | print('loading dictionary from %s' % path) 49 | word2idx, idx2word = cPickle.load(open(path, 'rb')) 50 | d = cls(word2idx, idx2word) 51 | return d 52 | 53 | def add_word(self, word): 54 | if word not in self.word2idx: 55 | self.idx2word.append(word) 56 | self.word2idx[word] = len(self.idx2word) - 1 57 | return self.word2idx[word] 58 | 59 | def __len__(self): 60 | return len(self.idx2word) 61 | 62 | 63 | def _create_entry(img, question, answer): 64 | answer.pop('image_id') 65 | answer.pop('question_id') 66 | entry = { 67 | 'question_id' : question['question_id'], 68 | 'image_id' : question['image_id'], 69 | 'image' : img, 70 | 'question' : question['question'], 71 | 'answer' : answer} 72 | return entry 73 | 74 | 75 | def _load_dataset(dataroot, name, img_id2val): 76 | """Load entries 77 | 78 | img_id2val: dict {img_id -> val} val can be used to retrieve image or features 79 | dataroot: root path of dataset 80 | name: 'train', 'val' 81 | """ 82 | question_path = os.path.join( 83 | dataroot, 'v2_OpenEnded_mscoco_%s2014_questions.json' % name) 84 | questions = sorted(json.load(open(question_path))['questions'], 85 | key=lambda x: x['question_id']) 86 | answer_path = os.path.join(dataroot, 'cache', '%s_target.pkl' % name) 87 | answers = cPickle.load(open(answer_path, 'rb')) 88 | answers = sorted(answers, key=lambda x: x['question_id']) 89 | 90 | utils.assert_eq(len(questions), len(answers)) 91 | entries = [] 92 | for question, answer in zip(questions, answers): 93 | utils.assert_eq(question['question_id'], answer['question_id']) 94 | utils.assert_eq(question['image_id'], answer['image_id']) 95 | img_id = question['image_id'] 96 | entries.append(_create_entry(img_id2val[img_id], question, answer)) 97 | 98 | return entries 99 | 100 | 101 | class VQAFeatureDataset(Dataset): 102 | def __init__(self, name, dictionary, dataroot='data'): 103 | super(VQAFeatureDataset, self).__init__() 104 | assert name in ['train', 'val'] 105 | 106 | ans2label_path = os.path.join(dataroot, 'cache', 'trainval_ans2label.pkl') 107 | label2ans_path = os.path.join(dataroot, 'cache', 'trainval_label2ans.pkl') 108 | self.ans2label = cPickle.load(open(ans2label_path, 'rb')) 109 | self.label2ans = cPickle.load(open(label2ans_path, 'rb')) 110 | self.num_ans_candidates = len(self.ans2label) 111 | 112 | self.dictionary = dictionary 113 | 114 | self.img_id2idx = cPickle.load( 115 | open(os.path.join(dataroot, '%s36_imgid2idx.pkl' % name))) 116 | print('loading features from h5 file') 117 | h5_path = os.path.join(dataroot, '%s36.hdf5' % name) 118 | with h5py.File(h5_path, 'r') as hf: 119 | self.features = np.array(hf.get('image_features')) 120 | self.spatials = np.array(hf.get('spatial_features')) 121 | 122 | self.entries = _load_dataset(dataroot, name, self.img_id2idx) 123 | 124 | self.tokenize() 125 | self.tensorize() 126 | self.v_dim = self.features.size(2) 127 | self.s_dim = self.spatials.size(2) 128 | 129 | def tokenize(self, max_length=14): 130 | """Tokenizes the questions. 131 | 132 | This will add q_token in each entry of the dataset. 133 | -1 represent nil, and should be treated as padding_idx in embedding 134 | """ 135 | for entry in self.entries: 136 | tokens = self.dictionary.tokenize(entry['question'], False) 137 | tokens = tokens[:max_length] 138 | if len(tokens) < max_length: 139 | # Note here we pad in front of the sentence 140 | padding = [self.dictionary.padding_idx] * (max_length - len(tokens)) 141 | tokens = padding + tokens 142 | utils.assert_eq(len(tokens), max_length) 143 | entry['q_token'] = tokens 144 | 145 | def tensorize(self): 146 | self.features = torch.from_numpy(self.features) 147 | self.spatials = torch.from_numpy(self.spatials) 148 | 149 | for entry in self.entries: 150 | question = torch.from_numpy(np.array(entry['q_token'])) 151 | entry['q_token'] = question 152 | 153 | answer = entry['answer'] 154 | labels = np.array(answer['labels']) 155 | scores = np.array(answer['scores'], dtype=np.float32) 156 | if len(labels): 157 | labels = torch.from_numpy(labels) 158 | scores = torch.from_numpy(scores) 159 | entry['answer']['labels'] = labels 160 | entry['answer']['scores'] = scores 161 | else: 162 | entry['answer']['labels'] = None 163 | entry['answer']['scores'] = None 164 | 165 | def __getitem__(self, index): 166 | entry = self.entries[index] 167 | features = self.features[entry['image']] 168 | spatials = self.spatials[entry['image']] 169 | 170 | question = entry['q_token'] 171 | answer = entry['answer'] 172 | labels = answer['labels'] 173 | scores = answer['scores'] 174 | target = torch.zeros(self.num_ans_candidates) 175 | if labels is not None: 176 | target.scatter_(0, labels, scores) 177 | 178 | return features, spatials, question, target 179 | 180 | def __len__(self): 181 | return len(self.entries) 182 | -------------------------------------------------------------------------------- /fc.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import torch.nn as nn 3 | from torch.nn.utils.weight_norm import weight_norm 4 | 5 | 6 | class FCNet(nn.Module): 7 | """Simple class for non-linear fully connect network 8 | """ 9 | def __init__(self, dims): 10 | super(FCNet, self).__init__() 11 | 12 | layers = [] 13 | for i in range(len(dims)-2): 14 | in_dim = dims[i] 15 | out_dim = dims[i+1] 16 | layers.append(weight_norm(nn.Linear(in_dim, out_dim), dim=None)) 17 | layers.append(nn.ReLU()) 18 | layers.append(weight_norm(nn.Linear(dims[-2], dims[-1]), dim=None)) 19 | layers.append(nn.ReLU()) 20 | 21 | self.main = nn.Sequential(*layers) 22 | 23 | def forward(self, x): 24 | return self.main(x) 25 | 26 | 27 | if __name__ == '__main__': 28 | fc1 = FCNet([10, 20, 10]) 29 | print(fc1) 30 | 31 | print('============') 32 | fc2 = FCNet([10, 20]) 33 | print(fc2) 34 | -------------------------------------------------------------------------------- /language_model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.autograd import Variable 4 | import numpy as np 5 | 6 | 7 | class WordEmbedding(nn.Module): 8 | """Word Embedding 9 | 10 | The ntoken-th dim is used for padding_idx, which agrees *implicitly* 11 | with the definition in Dictionary. 12 | """ 13 | def __init__(self, ntoken, emb_dim, dropout): 14 | super(WordEmbedding, self).__init__() 15 | self.emb = nn.Embedding(ntoken+1, emb_dim, padding_idx=ntoken) 16 | self.dropout = nn.Dropout(dropout) 17 | self.ntoken = ntoken 18 | self.emb_dim = emb_dim 19 | 20 | def init_embedding(self, np_file): 21 | weight_init = torch.from_numpy(np.load(np_file)) 22 | assert weight_init.shape == (self.ntoken, self.emb_dim) 23 | self.emb.weight.data[:self.ntoken] = weight_init 24 | 25 | def forward(self, x): 26 | emb = self.emb(x) 27 | emb = self.dropout(emb) 28 | return emb 29 | 30 | 31 | class QuestionEmbedding(nn.Module): 32 | def __init__(self, in_dim, num_hid, nlayers, bidirect, dropout, rnn_type='GRU'): 33 | """Module for question embedding 34 | """ 35 | super(QuestionEmbedding, self).__init__() 36 | assert rnn_type == 'LSTM' or rnn_type == 'GRU' 37 | rnn_cls = nn.LSTM if rnn_type == 'LSTM' else nn.GRU 38 | 39 | self.rnn = rnn_cls( 40 | in_dim, num_hid, nlayers, 41 | bidirectional=bidirect, 42 | dropout=dropout, 43 | batch_first=True) 44 | 45 | self.in_dim = in_dim 46 | self.num_hid = num_hid 47 | self.nlayers = nlayers 48 | self.rnn_type = rnn_type 49 | self.ndirections = 1 + int(bidirect) 50 | 51 | def init_hidden(self, batch): 52 | # just to get the type of tensor 53 | weight = next(self.parameters()).data 54 | hid_shape = (self.nlayers * self.ndirections, batch, self.num_hid) 55 | if self.rnn_type == 'LSTM': 56 | return (Variable(weight.new(*hid_shape).zero_()), 57 | Variable(weight.new(*hid_shape).zero_())) 58 | else: 59 | return Variable(weight.new(*hid_shape).zero_()) 60 | 61 | def forward(self, x): 62 | # x: [batch, sequence, in_dim] 63 | batch = x.size(0) 64 | hidden = self.init_hidden(batch) 65 | self.rnn.flatten_parameters() 66 | output, hidden = self.rnn(x, hidden) 67 | 68 | if self.ndirections == 1: 69 | return output[:, -1] 70 | 71 | forward_ = output[:, -1, :self.num_hid] 72 | backward = output[:, 0, self.num_hid:] 73 | return torch.cat((forward_, backward), dim=1) 74 | 75 | def forward_all(self, x): 76 | # x: [batch, sequence, in_dim] 77 | batch = x.size(0) 78 | hidden = self.init_hidden(batch) 79 | self.rnn.flatten_parameters() 80 | output, hidden = self.rnn(x, hidden) 81 | return output 82 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import torch 3 | import torch.nn as nn 4 | from torch.utils.data import DataLoader 5 | import numpy as np 6 | 7 | from dataset import Dictionary, VQAFeatureDataset 8 | import base_model 9 | from train import train 10 | import utils 11 | 12 | 13 | def parse_args(): 14 | parser = argparse.ArgumentParser() 15 | parser.add_argument('--epochs', type=int, default=30) 16 | parser.add_argument('--num_hid', type=int, default=1024) 17 | parser.add_argument('--model', type=str, default='baseline0_newatt') 18 | parser.add_argument('--output', type=str, default='saved_models/exp0') 19 | parser.add_argument('--batch_size', type=int, default=512) 20 | parser.add_argument('--seed', type=int, default=1111, help='random seed') 21 | args = parser.parse_args() 22 | return args 23 | 24 | 25 | if __name__ == '__main__': 26 | args = parse_args() 27 | 28 | torch.manual_seed(args.seed) 29 | torch.cuda.manual_seed(args.seed) 30 | torch.backends.cudnn.benchmark = True 31 | 32 | dictionary = Dictionary.load_from_file('data/dictionary.pkl') 33 | train_dset = VQAFeatureDataset('train', dictionary) 34 | eval_dset = VQAFeatureDataset('val', dictionary) 35 | batch_size = args.batch_size 36 | 37 | constructor = 'build_%s' % args.model 38 | model = getattr(base_model, constructor)(train_dset, args.num_hid).cuda() 39 | model.w_emb.init_embedding('data/glove6b_init_300d.npy') 40 | 41 | model = nn.DataParallel(model).cuda() 42 | 43 | train_loader = DataLoader(train_dset, batch_size, shuffle=True, num_workers=1) 44 | eval_loader = DataLoader(eval_dset, batch_size, shuffle=True, num_workers=1) 45 | train(model, train_loader, eval_loader, args.epochs, args.output) 46 | -------------------------------------------------------------------------------- /tools/compute_softscore.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import os 3 | import sys 4 | import json 5 | import numpy as np 6 | import re 7 | import cPickle 8 | 9 | sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) 10 | from dataset import Dictionary 11 | import utils 12 | 13 | 14 | contractions = { 15 | "aint": "ain't", "arent": "aren't", "cant": "can't", "couldve": 16 | "could've", "couldnt": "couldn't", "couldn'tve": "couldn't've", 17 | "couldnt've": "couldn't've", "didnt": "didn't", "doesnt": 18 | "doesn't", "dont": "don't", "hadnt": "hadn't", "hadnt've": 19 | "hadn't've", "hadn'tve": "hadn't've", "hasnt": "hasn't", "havent": 20 | "haven't", "hed": "he'd", "hed've": "he'd've", "he'dve": 21 | "he'd've", "hes": "he's", "howd": "how'd", "howll": "how'll", 22 | "hows": "how's", "Id've": "I'd've", "I'dve": "I'd've", "Im": 23 | "I'm", "Ive": "I've", "isnt": "isn't", "itd": "it'd", "itd've": 24 | "it'd've", "it'dve": "it'd've", "itll": "it'll", "let's": "let's", 25 | "maam": "ma'am", "mightnt": "mightn't", "mightnt've": 26 | "mightn't've", "mightn'tve": "mightn't've", "mightve": "might've", 27 | "mustnt": "mustn't", "mustve": "must've", "neednt": "needn't", 28 | "notve": "not've", "oclock": "o'clock", "oughtnt": "oughtn't", 29 | "ow's'at": "'ow's'at", "'ows'at": "'ow's'at", "'ow'sat": 30 | "'ow's'at", "shant": "shan't", "shed've": "she'd've", "she'dve": 31 | "she'd've", "she's": "she's", "shouldve": "should've", "shouldnt": 32 | "shouldn't", "shouldnt've": "shouldn't've", "shouldn'tve": 33 | "shouldn't've", "somebody'd": "somebodyd", "somebodyd've": 34 | "somebody'd've", "somebody'dve": "somebody'd've", "somebodyll": 35 | "somebody'll", "somebodys": "somebody's", "someoned": "someone'd", 36 | "someoned've": "someone'd've", "someone'dve": "someone'd've", 37 | "someonell": "someone'll", "someones": "someone's", "somethingd": 38 | "something'd", "somethingd've": "something'd've", "something'dve": 39 | "something'd've", "somethingll": "something'll", "thats": 40 | "that's", "thered": "there'd", "thered've": "there'd've", 41 | "there'dve": "there'd've", "therere": "there're", "theres": 42 | "there's", "theyd": "they'd", "theyd've": "they'd've", "they'dve": 43 | "they'd've", "theyll": "they'll", "theyre": "they're", "theyve": 44 | "they've", "twas": "'twas", "wasnt": "wasn't", "wed've": 45 | "we'd've", "we'dve": "we'd've", "weve": "we've", "werent": 46 | "weren't", "whatll": "what'll", "whatre": "what're", "whats": 47 | "what's", "whatve": "what've", "whens": "when's", "whered": 48 | "where'd", "wheres": "where's", "whereve": "where've", "whod": 49 | "who'd", "whod've": "who'd've", "who'dve": "who'd've", "wholl": 50 | "who'll", "whos": "who's", "whove": "who've", "whyll": "why'll", 51 | "whyre": "why're", "whys": "why's", "wont": "won't", "wouldve": 52 | "would've", "wouldnt": "wouldn't", "wouldnt've": "wouldn't've", 53 | "wouldn'tve": "wouldn't've", "yall": "y'all", "yall'll": 54 | "y'all'll", "y'allll": "y'all'll", "yall'd've": "y'all'd've", 55 | "y'alld've": "y'all'd've", "y'all'dve": "y'all'd've", "youd": 56 | "you'd", "youd've": "you'd've", "you'dve": "you'd've", "youll": 57 | "you'll", "youre": "you're", "youve": "you've" 58 | } 59 | 60 | manual_map = { 'none': '0', 61 | 'zero': '0', 62 | 'one': '1', 63 | 'two': '2', 64 | 'three': '3', 65 | 'four': '4', 66 | 'five': '5', 67 | 'six': '6', 68 | 'seven': '7', 69 | 'eight': '8', 70 | 'nine': '9', 71 | 'ten': '10'} 72 | articles = ['a', 'an', 'the'] 73 | period_strip = re.compile("(?!<=\d)(\.)(?!\d)") 74 | comma_strip = re.compile("(\d)(\,)(\d)") 75 | punct = [';', r"/", '[', ']', '"', '{', '}', 76 | '(', ')', '=', '+', '\\', '_', '-', 77 | '>', '<', '@', '`', ',', '?', '!'] 78 | 79 | 80 | def get_score(occurences): 81 | if occurences == 0: 82 | return 0 83 | elif occurences == 1: 84 | return 0.3 85 | elif occurences == 2: 86 | return 0.6 87 | elif occurences == 3: 88 | return 0.9 89 | else: 90 | return 1 91 | 92 | 93 | def process_punctuation(inText): 94 | outText = inText 95 | for p in punct: 96 | if (p + ' ' in inText or ' ' + p in inText) \ 97 | or (re.search(comma_strip, inText) != None): 98 | outText = outText.replace(p, '') 99 | else: 100 | outText = outText.replace(p, ' ') 101 | outText = period_strip.sub("", outText, re.UNICODE) 102 | return outText 103 | 104 | 105 | def process_digit_article(inText): 106 | outText = [] 107 | tempText = inText.lower().split() 108 | for word in tempText: 109 | word = manual_map.setdefault(word, word) 110 | if word not in articles: 111 | outText.append(word) 112 | else: 113 | pass 114 | for wordId, word in enumerate(outText): 115 | if word in contractions: 116 | outText[wordId] = contractions[word] 117 | outText = ' '.join(outText) 118 | return outText 119 | 120 | 121 | def multiple_replace(text, wordDict): 122 | for key in wordDict: 123 | text = text.replace(key, wordDict[key]) 124 | return text 125 | 126 | 127 | def preprocess_answer(answer): 128 | answer = process_digit_article(process_punctuation(answer)) 129 | answer = answer.replace(',', '') 130 | return answer 131 | 132 | 133 | def filter_answers(answers_dset, min_occurence): 134 | """This will change the answer to preprocessed version 135 | """ 136 | occurence = {} 137 | 138 | for ans_entry in answers_dset: 139 | answers = ans_entry['answers'] 140 | gtruth = ans_entry['multiple_choice_answer'] 141 | gtruth = preprocess_answer(gtruth) 142 | if gtruth not in occurence: 143 | occurence[gtruth] = set() 144 | occurence[gtruth].add(ans_entry['question_id']) 145 | for answer in occurence.keys(): 146 | if len(occurence[answer]) < min_occurence: 147 | occurence.pop(answer) 148 | 149 | print('Num of answers that appear >= %d times: %d' % ( 150 | min_occurence, len(occurence))) 151 | return occurence 152 | 153 | 154 | def create_ans2label(occurence, name, cache_root='data/cache'): 155 | """Note that this will also create label2ans.pkl at the same time 156 | 157 | occurence: dict {answer -> whatever} 158 | name: prefix of the output file 159 | cache_root: str 160 | """ 161 | ans2label = {} 162 | label2ans = [] 163 | label = 0 164 | for answer in occurence: 165 | label2ans.append(answer) 166 | ans2label[answer] = label 167 | label += 1 168 | 169 | utils.create_dir(cache_root) 170 | 171 | cache_file = os.path.join(cache_root, name+'_ans2label.pkl') 172 | cPickle.dump(ans2label, open(cache_file, 'wb')) 173 | cache_file = os.path.join(cache_root, name+'_label2ans.pkl') 174 | cPickle.dump(label2ans, open(cache_file, 'wb')) 175 | return ans2label 176 | 177 | 178 | def compute_target(answers_dset, ans2label, name, cache_root='data/cache'): 179 | """Augment answers_dset with soft score as label 180 | 181 | ***answers_dset should be preprocessed*** 182 | 183 | Write result into a cache file 184 | """ 185 | target = [] 186 | for ans_entry in answers_dset: 187 | answers = ans_entry['answers'] 188 | answer_count = {} 189 | for answer in answers: 190 | answer_ = answer['answer'] 191 | answer_count[answer_] = answer_count.get(answer_, 0) + 1 192 | 193 | labels = [] 194 | scores = [] 195 | for answer in answer_count: 196 | if answer not in ans2label: 197 | continue 198 | labels.append(ans2label[answer]) 199 | score = get_score(answer_count[answer]) 200 | scores.append(score) 201 | 202 | target.append({ 203 | 'question_id': ans_entry['question_id'], 204 | 'image_id': ans_entry['image_id'], 205 | 'labels': labels, 206 | 'scores': scores 207 | }) 208 | 209 | utils.create_dir(cache_root) 210 | cache_file = os.path.join(cache_root, name+'_target.pkl') 211 | cPickle.dump(target, open(cache_file, 'wb')) 212 | return target 213 | 214 | 215 | def get_answer(qid, answers): 216 | for ans in answers: 217 | if ans['question_id'] == qid: 218 | return ans 219 | 220 | 221 | def get_question(qid, questions): 222 | for question in questions: 223 | if question['question_id'] == qid: 224 | return question 225 | 226 | 227 | if __name__ == '__main__': 228 | train_answer_file = 'data/v2_mscoco_train2014_annotations.json' 229 | train_answers = json.load(open(train_answer_file))['annotations'] 230 | 231 | val_answer_file = 'data/v2_mscoco_val2014_annotations.json' 232 | val_answers = json.load(open(val_answer_file))['annotations'] 233 | 234 | train_question_file = 'data/v2_OpenEnded_mscoco_train2014_questions.json' 235 | train_questions = json.load(open(train_question_file))['questions'] 236 | 237 | val_question_file = 'data/v2_OpenEnded_mscoco_val2014_questions.json' 238 | val_questions = json.load(open(val_question_file))['questions'] 239 | 240 | answers = train_answers + val_answers 241 | occurence = filter_answers(answers, 9) 242 | ans2label = create_ans2label(occurence, 'trainval') 243 | compute_target(train_answers, ans2label, 'train') 244 | compute_target(val_answers, ans2label, 'val') 245 | -------------------------------------------------------------------------------- /tools/create_dictionary.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import os 3 | import sys 4 | import json 5 | import numpy as np 6 | sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) 7 | from dataset import Dictionary 8 | 9 | 10 | def create_dictionary(dataroot): 11 | dictionary = Dictionary() 12 | questions = [] 13 | files = [ 14 | 'v2_OpenEnded_mscoco_train2014_questions.json', 15 | 'v2_OpenEnded_mscoco_val2014_questions.json', 16 | 'v2_OpenEnded_mscoco_test2015_questions.json', 17 | 'v2_OpenEnded_mscoco_test-dev2015_questions.json' 18 | ] 19 | for path in files: 20 | question_path = os.path.join(dataroot, path) 21 | qs = json.load(open(question_path))['questions'] 22 | for q in qs: 23 | dictionary.tokenize(q['question'], True) 24 | return dictionary 25 | 26 | 27 | def create_glove_embedding_init(idx2word, glove_file): 28 | word2emb = {} 29 | with open(glove_file, 'r') as f: 30 | entries = f.readlines() 31 | emb_dim = len(entries[0].split(' ')) - 1 32 | print('embedding dim is %d' % emb_dim) 33 | weights = np.zeros((len(idx2word), emb_dim), dtype=np.float32) 34 | 35 | for entry in entries: 36 | vals = entry.split(' ') 37 | word = vals[0] 38 | vals = map(float, vals[1:]) 39 | word2emb[word] = np.array(vals) 40 | for idx, word in enumerate(idx2word): 41 | if word not in word2emb: 42 | continue 43 | weights[idx] = word2emb[word] 44 | return weights, word2emb 45 | 46 | 47 | if __name__ == '__main__': 48 | d = create_dictionary('data') 49 | d.dump_to_file('data/dictionary.pkl') 50 | 51 | d = Dictionary.load_from_file('data/dictionary.pkl') 52 | emb_dim = 300 53 | glove_file = 'data/glove/glove.6B.%dd.txt' % emb_dim 54 | weights, word2emb = create_glove_embedding_init(d.idx2word, glove_file) 55 | np.save('data/glove6b_init_%dd.npy' % emb_dim, weights) 56 | -------------------------------------------------------------------------------- /tools/detection_features_converter.py: -------------------------------------------------------------------------------- 1 | """ 2 | Reads in a tsv file with pre-trained bottom up attention features and 3 | stores it in HDF5 format. Also store {image_id: feature_idx} 4 | as a pickle file. 5 | 6 | Hierarchy of HDF5 file: 7 | 8 | { 'image_features': num_images x num_boxes x 2048 array of features 9 | 'image_bb': num_images x num_boxes x 4 array of bounding boxes } 10 | """ 11 | from __future__ import print_function 12 | 13 | import os 14 | import sys 15 | sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) 16 | 17 | import base64 18 | import csv 19 | import h5py 20 | import cPickle 21 | import numpy as np 22 | import utils 23 | 24 | 25 | csv.field_size_limit(sys.maxsize) 26 | 27 | FIELDNAMES = ['image_id', 'image_w', 'image_h', 'num_boxes', 'boxes', 'features'] 28 | infile = 'data/trainval_36/trainval_resnet101_faster_rcnn_genome_36.tsv' 29 | train_data_file = 'data/train36.hdf5' 30 | val_data_file = 'data/val36.hdf5' 31 | train_indices_file = 'data/train36_imgid2idx.pkl' 32 | val_indices_file = 'data/val36_imgid2idx.pkl' 33 | train_ids_file = 'data/train_ids.pkl' 34 | val_ids_file = 'data/val_ids.pkl' 35 | 36 | feature_length = 2048 37 | num_fixed_boxes = 36 38 | 39 | 40 | if __name__ == '__main__': 41 | h_train = h5py.File(train_data_file, "w") 42 | h_val = h5py.File(val_data_file, "w") 43 | 44 | if os.path.exists(train_ids_file) and os.path.exists(val_ids_file): 45 | train_imgids = cPickle.load(open(train_ids_file)) 46 | val_imgids = cPickle.load(open(val_ids_file)) 47 | else: 48 | train_imgids = utils.load_imageid('data/train2014') 49 | val_imgids = utils.load_imageid('data/val2014') 50 | cPickle.dump(train_imgids, open(train_ids_file, 'wb')) 51 | cPickle.dump(val_imgids, open(val_ids_file, 'wb')) 52 | 53 | train_indices = {} 54 | val_indices = {} 55 | 56 | train_img_features = h_train.create_dataset( 57 | 'image_features', (len(train_imgids), num_fixed_boxes, feature_length), 'f') 58 | train_img_bb = h_train.create_dataset( 59 | 'image_bb', (len(train_imgids), num_fixed_boxes, 4), 'f') 60 | train_spatial_img_features = h_train.create_dataset( 61 | 'spatial_features', (len(train_imgids), num_fixed_boxes, 6), 'f') 62 | 63 | val_img_bb = h_val.create_dataset( 64 | 'image_bb', (len(val_imgids), num_fixed_boxes, 4), 'f') 65 | val_img_features = h_val.create_dataset( 66 | 'image_features', (len(val_imgids), num_fixed_boxes, feature_length), 'f') 67 | val_spatial_img_features = h_val.create_dataset( 68 | 'spatial_features', (len(val_imgids), num_fixed_boxes, 6), 'f') 69 | 70 | train_counter = 0 71 | val_counter = 0 72 | 73 | print("reading tsv...") 74 | with open(infile, "r+b") as tsv_in_file: 75 | reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames=FIELDNAMES) 76 | for item in reader: 77 | item['num_boxes'] = int(item['num_boxes']) 78 | image_id = int(item['image_id']) 79 | image_w = float(item['image_w']) 80 | image_h = float(item['image_h']) 81 | bboxes = np.frombuffer( 82 | base64.decodestring(item['boxes']), 83 | dtype=np.float32).reshape((item['num_boxes'], -1)) 84 | 85 | box_width = bboxes[:, 2] - bboxes[:, 0] 86 | box_height = bboxes[:, 3] - bboxes[:, 1] 87 | scaled_width = box_width / image_w 88 | scaled_height = box_height / image_h 89 | scaled_x = bboxes[:, 0] / image_w 90 | scaled_y = bboxes[:, 1] / image_h 91 | 92 | box_width = box_width[..., np.newaxis] 93 | box_height = box_height[..., np.newaxis] 94 | scaled_width = scaled_width[..., np.newaxis] 95 | scaled_height = scaled_height[..., np.newaxis] 96 | scaled_x = scaled_x[..., np.newaxis] 97 | scaled_y = scaled_y[..., np.newaxis] 98 | 99 | spatial_features = np.concatenate( 100 | (scaled_x, 101 | scaled_y, 102 | scaled_x + scaled_width, 103 | scaled_y + scaled_height, 104 | scaled_width, 105 | scaled_height), 106 | axis=1) 107 | 108 | if image_id in train_imgids: 109 | train_imgids.remove(image_id) 110 | train_indices[image_id] = train_counter 111 | train_img_bb[train_counter, :, :] = bboxes 112 | train_img_features[train_counter, :, :] = np.frombuffer( 113 | base64.decodestring(item['features']), 114 | dtype=np.float32).reshape((item['num_boxes'], -1)) 115 | train_spatial_img_features[train_counter, :, :] = spatial_features 116 | train_counter += 1 117 | elif image_id in val_imgids: 118 | val_imgids.remove(image_id) 119 | val_indices[image_id] = val_counter 120 | val_img_bb[val_counter, :, :] = bboxes 121 | val_img_features[val_counter, :, :] = np.frombuffer( 122 | base64.decodestring(item['features']), 123 | dtype=np.float32).reshape((item['num_boxes'], -1)) 124 | val_spatial_img_features[val_counter, :, :] = spatial_features 125 | val_counter += 1 126 | else: 127 | assert False, 'Unknown image id: %d' % image_id 128 | 129 | if len(train_imgids) != 0: 130 | print('Warning: train_image_ids is not empty') 131 | 132 | if len(val_imgids) != 0: 133 | print('Warning: val_image_ids is not empty') 134 | 135 | cPickle.dump(train_indices, open(train_indices_file, 'wb')) 136 | cPickle.dump(val_indices, open(val_indices_file, 'wb')) 137 | h_train.close() 138 | h_val.close() 139 | print("done!") 140 | -------------------------------------------------------------------------------- /tools/download.sh: -------------------------------------------------------------------------------- 1 | ## Script for downloading data 2 | 3 | # GloVe Vectors 4 | wget -P data http://nlp.stanford.edu/data/glove.6B.zip 5 | unzip data/glove.6B.zip -d data/glove 6 | rm data/glove.6B.zip 7 | 8 | # Questions 9 | wget -P data http://visualqa.org/data/mscoco/vqa/v2_Questions_Train_mscoco.zip 10 | unzip data/v2_Questions_Train_mscoco.zip -d data 11 | rm data/v2_Questions_Train_mscoco.zip 12 | 13 | wget -P data http://visualqa.org/data/mscoco/vqa/v2_Questions_Val_mscoco.zip 14 | unzip data/v2_Questions_Val_mscoco.zip -d data 15 | rm data/v2_Questions_Val_mscoco.zip 16 | 17 | wget -P data http://visualqa.org/data/mscoco/vqa/v2_Questions_Test_mscoco.zip 18 | unzip data/v2_Questions_Test_mscoco.zip -d data 19 | rm data/v2_Questions_Test_mscoco.zip 20 | 21 | # Annotations 22 | wget -P data http://visualqa.org/data/mscoco/vqa/v2_Annotations_Train_mscoco.zip 23 | unzip data/v2_Annotations_Train_mscoco.zip -d data 24 | rm data/v2_Annotations_Train_mscoco.zip 25 | 26 | wget -P data http://visualqa.org/data/mscoco/vqa/v2_Annotations_Val_mscoco.zip 27 | unzip data/v2_Annotations_Val_mscoco.zip -d data 28 | rm data/v2_Annotations_Val_mscoco.zip 29 | 30 | # Image Features 31 | wget -P data https://imagecaption.blob.core.windows.net/imagecaption/trainval_36.zip 32 | unzip data/trainval_36.zip -d data 33 | rm data/trainval_36.zip 34 | -------------------------------------------------------------------------------- /tools/process.sh: -------------------------------------------------------------------------------- 1 | # Process data 2 | 3 | python tools/create_dictionary.py 4 | python tools/compute_softscore.py 5 | python tools/detection_features_converter.py 6 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import os 2 | import time 3 | import torch 4 | import torch.nn as nn 5 | import utils 6 | from torch.autograd import Variable 7 | 8 | 9 | def instance_bce_with_logits(logits, labels): 10 | assert logits.dim() == 2 11 | 12 | loss = nn.functional.binary_cross_entropy_with_logits(logits, labels) 13 | loss *= labels.size(1) 14 | return loss 15 | 16 | 17 | def compute_score_with_logits(logits, labels): 18 | logits = torch.max(logits, 1)[1].data # argmax 19 | one_hots = torch.zeros(*labels.size()).cuda() 20 | one_hots.scatter_(1, logits.view(-1, 1), 1) 21 | scores = (one_hots * labels) 22 | return scores 23 | 24 | 25 | def train(model, train_loader, eval_loader, num_epochs, output): 26 | utils.create_dir(output) 27 | optim = torch.optim.Adamax(model.parameters()) 28 | logger = utils.Logger(os.path.join(output, 'log.txt')) 29 | best_eval_score = 0 30 | 31 | for epoch in range(num_epochs): 32 | total_loss = 0 33 | train_score = 0 34 | t = time.time() 35 | 36 | for i, (v, b, q, a) in enumerate(train_loader): 37 | v = Variable(v).cuda() 38 | b = Variable(b).cuda() 39 | q = Variable(q).cuda() 40 | a = Variable(a).cuda() 41 | 42 | pred = model(v, b, q, a) 43 | loss = instance_bce_with_logits(pred, a) 44 | loss.backward() 45 | nn.utils.clip_grad_norm(model.parameters(), 0.25) 46 | optim.step() 47 | optim.zero_grad() 48 | 49 | batch_score = compute_score_with_logits(pred, a.data).sum() 50 | total_loss += loss.data[0] * v.size(0) 51 | train_score += batch_score 52 | 53 | total_loss /= len(train_loader.dataset) 54 | train_score = 100 * train_score / len(train_loader.dataset) 55 | model.train(False) 56 | eval_score, bound = evaluate(model, eval_loader) 57 | model.train(True) 58 | 59 | logger.write('epoch %d, time: %.2f' % (epoch, time.time()-t)) 60 | logger.write('\ttrain_loss: %.2f, score: %.2f' % (total_loss, train_score)) 61 | logger.write('\teval score: %.2f (%.2f)' % (100 * eval_score, 100 * bound)) 62 | 63 | if eval_score > best_eval_score: 64 | model_path = os.path.join(output, 'model.pth') 65 | torch.save(model.state_dict(), model_path) 66 | best_eval_score = eval_score 67 | 68 | 69 | def evaluate(model, dataloader): 70 | score = 0 71 | upper_bound = 0 72 | num_data = 0 73 | for v, b, q, a in iter(dataloader): 74 | v = Variable(v, volatile=True).cuda() 75 | b = Variable(b, volatile=True).cuda() 76 | q = Variable(q, volatile=True).cuda() 77 | pred = model(v, b, q, None) 78 | batch_score = compute_score_with_logits(pred, a.cuda()).sum() 79 | score += batch_score 80 | upper_bound += (a.max(1)[0]).sum() 81 | num_data += pred.size(0) 82 | 83 | score = score / len(dataloader.dataset) 84 | upper_bound = upper_bound / len(dataloader.dataset) 85 | return score, upper_bound 86 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import errno 4 | import os 5 | import numpy as np 6 | from PIL import Image 7 | import torch 8 | import torch.nn as nn 9 | 10 | 11 | EPS = 1e-7 12 | 13 | 14 | def assert_eq(real, expected): 15 | assert real == expected, '%s (true) vs %s (expected)' % (real, expected) 16 | 17 | 18 | def assert_array_eq(real, expected): 19 | assert (np.abs(real-expected) < EPS).all(), \ 20 | '%s (true) vs %s (expected)' % (real, expected) 21 | 22 | 23 | def load_folder(folder, suffix): 24 | imgs = [] 25 | for f in sorted(os.listdir(folder)): 26 | if f.endswith(suffix): 27 | imgs.append(os.path.join(folder, f)) 28 | return imgs 29 | 30 | 31 | def load_imageid(folder): 32 | images = load_folder(folder, 'jpg') 33 | img_ids = set() 34 | for img in images: 35 | img_id = int(img.split('/')[-1].split('.')[0].split('_')[-1]) 36 | img_ids.add(img_id) 37 | return img_ids 38 | 39 | 40 | def pil_loader(path): 41 | with open(path, 'rb') as f: 42 | with Image.open(f) as img: 43 | return img.convert('RGB') 44 | 45 | 46 | def weights_init(m): 47 | """custom weights initialization.""" 48 | cname = m.__class__ 49 | if cname == nn.Linear or cname == nn.Conv2d or cname == nn.ConvTranspose2d: 50 | m.weight.data.normal_(0.0, 0.02) 51 | elif cname == nn.BatchNorm2d: 52 | m.weight.data.normal_(1.0, 0.02) 53 | m.bias.data.fill_(0) 54 | else: 55 | print('%s is not initialized.' % cname) 56 | 57 | 58 | def init_net(net, net_file): 59 | if net_file: 60 | net.load_state_dict(torch.load(net_file)) 61 | else: 62 | net.apply(weights_init) 63 | 64 | 65 | def create_dir(path): 66 | if not os.path.exists(path): 67 | try: 68 | os.makedirs(path) 69 | except OSError as exc: 70 | if exc.errno != errno.EEXIST: 71 | raise 72 | 73 | 74 | class Logger(object): 75 | def __init__(self, output_name): 76 | dirname = os.path.dirname(output_name) 77 | if not os.path.exists(dirname): 78 | os.mkdir(dirname) 79 | 80 | self.log_file = open(output_name, 'w') 81 | self.infos = {} 82 | 83 | def append(self, key, val): 84 | vals = self.infos.setdefault(key, []) 85 | vals.append(val) 86 | 87 | def log(self, extra_msg=''): 88 | msgs = [extra_msg] 89 | for key, vals in self.infos.iteritems(): 90 | msgs.append('%s %.6f' % (key, np.mean(vals))) 91 | msg = '\n'.join(msgs) 92 | self.log_file.write(msg + '\n') 93 | self.log_file.flush() 94 | self.infos = {} 95 | return msg 96 | 97 | def write(self, msg): 98 | self.log_file.write(msg + '\n') 99 | self.log_file.flush() 100 | print(msg) 101 | --------------------------------------------------------------------------------