├── LICENSE ├── README.md ├── captions.txt ├── detect_object.py ├── test_images ├── 1.jpg └── 2.jpg └── train.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 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 | # Image-Captioning-via-YOLOv5-EncoderDecoderwithAttention 2 | 3 | #### Archived : This code was just a fun project! Neither it was propely tuned nor it is properly maintained! No Updates/correction expected! Focusing on other areas, I am not a vision expert. This code runs properly for most of the people, please check if images are getting populated properly for you if there is an error!** 4 | 5 | 6 | Use original Flickr8K dataset. - https://www.kaggle.com/datasets/adityajn105/flickr8k 7 | 8 | PUT 'Images' directory and 'captions.txt' in the same directory as in root of this repo. 9 | 10 | 11 | Attempt for Image Captioning using combination of object detection via YOLOv5 and Encoder Decoder LSTM model on Flickr8K dataset. 12 | 13 | 1. Run to make object crops via YOLOv5 14 | ``` 15 | python detect_object.py 16 | ``` 17 | 2. Run to train - This just takes the Resnet embeddngs of object cropped images detected not any kind of text from YOLO labeller. 18 | ``` 19 | python train.py True 20 | ``` 21 | 3. To evaluate on validation data 22 | ``` 23 | python train.py False 24 | ``` 25 | 4. Sample predictions - 26 | 27 | ![2.jpg](https://github.com/akjayant/Image-Captioning-via-YOLOv5-EncoderDecoderwithAttention/blob/main/test_images/2.jpg) 28 | 29 | ``` 30 | references - [This is a black dog splashing in the water, A black lab with tags frolicks in the water ,A black dog running in the surf,The black dog runs through the water] 31 | 32 | prediction- [[''], ['a'], ['black'], ['dog'], ['is'], ['a'], ['a'], ['water'], ['.'], ['']] 33 | ``` 34 | 35 | ![1.jpg](https://github.com/akjayant/Image-Captioning-via-YOLOv5-EncoderDecoderwithAttention/blob/main/test_images/1.jpg) 36 | ``` 37 | references - [A black dog and a spotted dog are fighting, A black dog and a tri-colored dog playing with each other on the road, 38 | A black dog and a white dog with brown spots are staring at each other in the street,Two dogs of different breeds looking at each other on the road] 39 | 40 | prediction- [[''], ['a'], ['black'], ['and'], ['white'], ['dog'], ['is'], ['running'], ['through'], ['a'], ['.'], ['']] 41 | ``` 42 | 5. Mean BLEU-4 score on validation data is quite low. Suggested Improvements : Use Adam and shuffling of data. Maybe minibatching. Increasing number of datapoints combining other datasets since 8k is quite low smaple size (No of parameters >> no of datapoints, not ideal for neural nets). 43 | 44 | 45 | ## Citation (Flickr8K Dataset) 46 | 47 | Hodosh, Micah, Peter Young, and Julia Hockenmaier. "Framing image description as a ranking task: Data, models and evaluation metrics." Journal of Artificial Intelligence Research 47 (2013): 853-899. 48 | 49 | 50 | -------------------------------------------------------------------------------- /detect_object.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torchvision import models 3 | import os 4 | from glob import glob 5 | import torchvision.transforms as transforms 6 | from PIL import Image 7 | from torch.autograd import Variable 8 | #from img2vec_pytorch import Img2Vec 9 | import pandas as pd 10 | 11 | # Model 12 | # Load the pretrained model 13 | resnet_model = models.resnet18(pretrained=True) 14 | 15 | model = torch.hub.load('ultralytics/yolov5', 'yolov5s') 16 | # or yolov5m, yolov5l, yolov5x, custom 17 | 18 | # Images 19 | 20 | img_path = './Images/'# or file, Path, PIL, OpenCV, numpy, list 21 | #output_path = '/media/ashish-j/B/wheat_detection/flick_data/embeddings' 22 | count=0 23 | for file in os.listdir(img_path): 24 | img = os.path.join(img_path,file) 25 | results = model(img) 26 | results.crop(save_dir='./detection_data/'+str(img)+'/') 27 | 28 | 29 | 30 | -------------------------------------------------------------------------------- /test_images/1.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/akjayant/Image-Captioning-via-YOLOv5-EncoderDecoderwithAttention/899223d19dec696d68382d3a68a3c3124d719aa4/test_images/1.jpg -------------------------------------------------------------------------------- /test_images/2.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/akjayant/Image-Captioning-via-YOLOv5-EncoderDecoderwithAttention/899223d19dec696d68382d3a68a3c3124d719aa4/test_images/2.jpg -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | 2 | from __future__ import unicode_literals, print_function, division 3 | import pandas as pd 4 | import torch 5 | from torch import nn, Tensor 6 | import torch.nn.functional as F 7 | from torch.nn import TransformerEncoder, TransformerEncoderLayer 8 | from torch.utils.data import dataset 9 | from sklearn import model_selection 10 | from torchtext.data.utils import get_tokenizer 11 | from torchtext.vocab import build_vocab_from_iterator 12 | from sklearn.feature_extraction.text import CountVectorizer 13 | import os 14 | from glob import glob 15 | from torchvision import models 16 | import torchvision.transforms as transforms 17 | from PIL import Image 18 | from torch.autograd import Variable 19 | import numpy as np 20 | from img2vec_pytorch import Img2Vec 21 | import time 22 | 23 | from io import open 24 | import unicodedata 25 | import string 26 | import re 27 | import random 28 | 29 | import torch 30 | import torch.nn as nn 31 | from torch import optim 32 | import torch.nn.functional as F 33 | from collections import OrderedDict 34 | from tqdm import tqdm 35 | import pickle 36 | import sys 37 | import random 38 | 39 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 40 | resnet_model = models.resnet18(pretrained=True) 41 | layer = resnet_model._modules.get('avgpool') 42 | resnet_model.eval() 43 | 44 | scaler = transforms.Resize((224, 224)) 45 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 46 | std=[0.229, 0.224, 0.225]) 47 | img2vec = Img2Vec(cuda=True) 48 | to_tensor = transforms.ToTensor() 49 | 50 | 51 | 52 | 53 | 54 | 55 | class EncoderRNN(nn.Module): 56 | def __init__(self, input_size, hidden_size): 57 | super(EncoderRNN, self).__init__() 58 | self.hidden_size = hidden_size 59 | 60 | self.embedding = nn.Embedding(input_size, hidden_size) 61 | self.gru = nn.GRU(hidden_size, hidden_size) 62 | 63 | def forward(self, input, hidden): 64 | embedded = input.view(1,1, 512) 65 | #print(hidden.size()) 66 | output = embedded 67 | output, hidden = self.gru(output, hidden) 68 | return output, hidden 69 | 70 | def initHidden(self): 71 | return torch.zeros(1, 1, self.hidden_size, device=device) 72 | 73 | 74 | SOS_token = 0 75 | EOS_token = 1 76 | 77 | 78 | # class Lang: 79 | # def __init__(self, name): 80 | # self.name = name 81 | # self.word2index = {} 82 | # self.word2count = {} 83 | # self.index2word = {0: "SOS", 1: "EOS"} 84 | # self.n_words = 2 # Count SOS and EOS 85 | 86 | # def addSentence(self, sentence): 87 | # for word in sentence.split(' '): 88 | # self.addWord(word) 89 | 90 | # def addWord(self, word): 91 | # if word not in self.word2index: 92 | # self.word2index[word] = self.n_words 93 | # self.word2count[word] = 1 94 | # self.index2word[self.n_words] = word 95 | # self.n_words += 1 96 | # else: 97 | # self.word2count[word] += 1 98 | # # Turn a Unicode string to plain ASCII, thanks to 99 | # # https://stackoverflow.com/a/518232/2809427 100 | # def unicodeToAscii(s): 101 | # return ''.join( 102 | # c for c in unicodedata.normalize('NFD', s) 103 | # if unicodedata.category(c) != 'Mn' 104 | # ) 105 | 106 | # # Lowercase, trim, and remove non-letter characters 107 | 108 | 109 | # def normalizeString(s): 110 | # s = unicodeToAscii(s.lower().strip()) 111 | # s = re.sub(r"([.!?])", r" \1", s) 112 | # s = re.sub(r"[^a-zA-Z.!?]+", r" ", s) 113 | # return s 114 | 115 | 116 | # def readLangs(lang1, lang2, reverse=False): 117 | # print("Reading lines...") 118 | 119 | # # Read the file and split into lines 120 | # lines = open('data/%s-%s.txt' % (lang1, lang2), encoding='utf-8').\ 121 | # read().strip().split('\n') 122 | 123 | # # Split every line into pairs and normalize 124 | # pairs = [[normalizeString(s) for s in l.split('\t')] for l in lines] 125 | 126 | # # Reverse pairs, make Lang instances 127 | # if reverse: 128 | # pairs = [list(reversed(p)) for p in pairs] 129 | # input_lang = Lang(lang2) 130 | # output_lang = Lang(lang1) 131 | # else: 132 | # input_lang = Lang(lang1) 133 | # output_lang = Lang(lang2) 134 | 135 | # return input_lang, output_lang, pairs 136 | 137 | MAX_LENGTH = 25 138 | 139 | # eng_prefixes = ( 140 | # "i am ", "i m ", 141 | # "he is", "he s ", 142 | # "she is", "she s ", 143 | # "you are", "you re ", 144 | # "we are", "we re ", 145 | # "they are", "they re " 146 | # ) 147 | 148 | 149 | # def filterPair(p): 150 | # return len(p[0].split(' ')) < MAX_LENGTH and \ 151 | # len(p[1].split(' ')) < MAX_LENGTH and \ 152 | # p[1].startswith(eng_prefixes) 153 | 154 | 155 | # def filterPairs(pairs): 156 | # return [pair for pair in pairs if filterPair(pair)] 157 | 158 | # def prepareData(lang1, lang2, reverse=False): 159 | # input_lang, output_lang, pairs = readLangs(lang1, lang2, reverse) 160 | # print("Read %s sentence pairs" % len(pairs)) 161 | # pairs = filterPairs(pairs) 162 | # print("Trimmed to %s sentence pairs" % len(pairs)) 163 | # print("Counting words...") 164 | # for pair in pairs: 165 | # input_lang.addSentence(pair[0]) 166 | # output_lang.addSentence(pair[1]) 167 | # print("Counted words:") 168 | # print(input_lang.name, input_lang.n_words) 169 | # print(output_lang.name, output_lang.n_words) 170 | # return input_lang, output_lang, pairs 171 | 172 | 173 | # input_lang, output_lang, pairs = prepareData('eng', 'fra', True) 174 | # print(random.choice(pairs)) 175 | 176 | # def indexesFromSentence(lang, sentence): 177 | # return [lang.word2index[word] for word in sentence.split(' ')] 178 | 179 | 180 | # def tensorFromSentence(lang, sentence): 181 | # indexes = indexesFromSentence(lang, sentence) 182 | # indexes.append(EOS_token) 183 | # return torch.tensor(indexes, dtype=torch.long, device=device).view(-1, 1) 184 | 185 | 186 | # def tensorsFromPair(pair): 187 | # input_tensor = tensorFromSentence(input_lang, pair[0]) 188 | # target_tensor = tensorFromSentence(output_lang, pair[1]) 189 | # return (input_tensor, target_tensor) 190 | class AttnDecoderRNN(nn.Module): 191 | def __init__(self, hidden_size, output_size, dropout_p=0.1, max_length=MAX_LENGTH): 192 | super(AttnDecoderRNN, self).__init__() 193 | self.hidden_size = hidden_size 194 | self.output_size = output_size 195 | self.dropout_p = dropout_p 196 | self.max_length = max_length 197 | 198 | self.embedding = nn.Embedding(self.output_size, self.hidden_size) 199 | self.attn = nn.Linear(self.hidden_size * 2, self.max_length) 200 | self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size) 201 | self.dropout = nn.Dropout(self.dropout_p) 202 | self.gru = nn.GRU(self.hidden_size, self.hidden_size) 203 | self.out = nn.Linear(self.hidden_size, self.output_size) 204 | 205 | def forward(self, input, hidden, encoder_outputs): 206 | embedded = self.embedding(input).view(1, 1, -1) 207 | embedded = self.dropout(embedded) 208 | 209 | attn_weights = F.softmax( 210 | self.attn(torch.cat((embedded[0], hidden[0]), 1)), dim=1) 211 | attn_applied = torch.bmm(attn_weights.unsqueeze(0), 212 | encoder_outputs.unsqueeze(0)) 213 | 214 | output = torch.cat((embedded[0], attn_applied[0]), 1) 215 | output = self.attn_combine(output).unsqueeze(0) 216 | 217 | output = F.relu(output) 218 | output, hidden = self.gru(output, hidden) 219 | 220 | output = F.log_softmax(self.out(output[0]), dim=1) 221 | return output, hidden, attn_weights 222 | 223 | def initHidden(self): 224 | return torch.zeros(1, 1, self.hidden_size, device=device) 225 | 226 | 227 | teacher_forcing_ratio = 0.5 228 | 229 | 230 | def train(input_tensor, target_tensor, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length=MAX_LENGTH): 231 | encoder_hidden = encoder.initHidden() 232 | 233 | encoder_optimizer.zero_grad() 234 | decoder_optimizer.zero_grad() 235 | 236 | input_length = input_tensor.size(0) 237 | target_length = target_tensor.size(0) 238 | 239 | encoder_outputs = torch.zeros(max_length, encoder.hidden_size, device=device) 240 | 241 | loss = 0 242 | 243 | for ei in range(input_length): 244 | encoder_output, encoder_hidden = encoder( 245 | input_tensor[ei], encoder_hidden) 246 | #print(encoder_output[0,0]) 247 | encoder_outputs[ei] = encoder_output[0, 0] 248 | 249 | decoder_input = torch.tensor([[SOS_token]], device=device) 250 | 251 | decoder_hidden = encoder_hidden 252 | 253 | use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False 254 | 255 | if use_teacher_forcing: 256 | # Teacher forcing: Feed the target as the next input 257 | for di in range(target_length): 258 | decoder_output, decoder_hidden, decoder_attention = decoder( 259 | decoder_input, decoder_hidden, encoder_outputs) 260 | loss += criterion(decoder_output, target_tensor[di]) 261 | decoder_input = target_tensor[di] # Teacher forcing 262 | 263 | else: 264 | # Without teacher forcing: use its own predictions as the next input 265 | for di in range(target_length): 266 | decoder_output, decoder_hidden, decoder_attention = decoder( 267 | decoder_input, decoder_hidden, encoder_outputs) 268 | topv, topi = decoder_output.topk(1) 269 | decoder_input = topi.squeeze().detach() # detach from history as input 270 | 271 | loss += criterion(decoder_output, target_tensor[di]) 272 | #print(decoder_input.item()) 273 | if decoder_input.item() == EOS_token: 274 | break 275 | 276 | loss.backward() 277 | 278 | encoder_optimizer.step() 279 | decoder_optimizer.step() 280 | 281 | return loss.item() / target_length 282 | 283 | 284 | import matplotlib.pyplot as plt 285 | plt.switch_backend('agg') 286 | import matplotlib.ticker as ticker 287 | import numpy as np 288 | 289 | import time 290 | import math 291 | 292 | 293 | def asMinutes(s): 294 | m = math.floor(s / 60) 295 | s -= m * 60 296 | return '%dm %ds' % (m, s) 297 | 298 | 299 | def timeSince(since, percent): 300 | now = time.time() 301 | s = now - since 302 | es = s / (percent) 303 | rs = es - s 304 | return '%s (- %s)' % (asMinutes(s), asMinutes(rs)) 305 | 306 | def showPlot(points): 307 | plt.figure() 308 | fig, ax = plt.subplots() 309 | # this locator puts ticks at regular intervals 310 | loc = ticker.MultipleLocator(base=0.2) 311 | ax.yaxis.set_major_locator(loc) 312 | plt.plot(points) 313 | 314 | def trainIters(encoder, decoder, n_iters, print_every=1000, plot_every=100, learning_rate=0.01): 315 | start = time.time() 316 | plot_losses = [] 317 | print_loss_total = 0 # Reset every print_every 318 | plot_loss_total = 0 # Reset every plot_every 319 | 320 | encoder_optimizer = optim.SGD(encoder.parameters(), lr=learning_rate) 321 | decoder_optimizer = optim.SGD(decoder.parameters(), lr=learning_rate) 322 | # training_pairs = [tensorsFromPair(random.choice(pairs)) 323 | # for i in range(n_iters)] 324 | print("Data---------") 325 | # print(training_pairs[0]) 326 | criterion = nn.NLLLoss() 327 | 328 | for iter in tqdm(range(1, n_iters + 1)): 329 | #training_pair = training_pairs[iter - 1] 330 | input_paths = joined_df['image_seq'][iter-1] 331 | sos_img = np.zeros([1,512]) 332 | eos_img = np.ones([1,512]) 333 | input_np=[] 334 | #print(input_paths) 335 | for i in input_paths.split(','): 336 | #print(i) 337 | img = Image.open(i) 338 | v = img2vec.get_vec(img,tensor=False).reshape(1,512) 339 | input_np.append(v) 340 | 341 | input_tensor = torch.tensor(input_np).float().to(device) 342 | #print(input_tensor) 343 | target_tensor = torch.tensor(joined_df['caption_seq'][iter-1]).to(device) 344 | #print(target_tensor) 345 | loss = train(input_tensor, target_tensor, encoder, 346 | decoder, encoder_optimizer, decoder_optimizer, criterion) 347 | print_loss_total += loss 348 | plot_loss_total += loss 349 | 350 | if iter % print_every == 0: 351 | torch.save(encoder1.state_dict(),"encoder_model.pkl") 352 | torch.save(attn_decoder1.state_dict(),"decoder_model.pkl") 353 | print_loss_avg = print_loss_total / print_every 354 | print_loss_total = 0 355 | print('%s (%d %d%%) %.4f' % (timeSince(start, iter / n_iters), 356 | iter, iter / n_iters * 100, print_loss_avg)) 357 | 358 | if iter % plot_every == 0: 359 | plot_loss_avg = plot_loss_total / plot_every 360 | plot_losses.append(plot_loss_avg) 361 | plot_loss_total = 0 362 | 363 | showPlot(plot_losses) 364 | 365 | 366 | 367 | 368 | 369 | 370 | 371 | 372 | crop_path = './detection_data/Images/' 373 | 374 | #---------feature extraction------------------ 375 | train_flag=sys.argv[1] 376 | if train_flag=='True': 377 | 378 | image_dict = OrderedDict() 379 | files_list = [y for x in os.walk(crop_path) for y in glob(os.path.join(x[0], '*.jpg'))] 380 | files_names = [os.path.basename(i) for i in files_list] 381 | files_names_set = set(files_names) 382 | print(files_list) 383 | 384 | c=0 385 | original_file_path = './Images/' 386 | for i in range(len(files_list)): 387 | base_name = files_list[i].split("/")[-4] 388 | b_idx = base_name 389 | orig_file = os.path.join(original_file_path,base_name) 390 | image_dict[base_name] = orig_file 391 | if base_name==files_names[i][:len(base_name)]: 392 | if base_name not in image_dict.keys(): 393 | image_dict[base_name] = str(files_list[i]) 394 | #image_dict[base_name] += ","+str(orig_file) 395 | else: 396 | image_dict[b_idx] += ","+str(files_list[i]) 397 | #print(image_dict[b_idx]) 398 | 399 | 400 | image_dict_keys = [k for k in image_dict.keys()] 401 | 402 | 403 | image_df = pd.DataFrame.from_dict(image_dict,orient='index',columns=['image_seq']) 404 | image_df['image'] = image_df.index 405 | 406 | 407 | embedding_input_size = len(files_list) 408 | def first_caption(x): 409 | return x[0] 410 | df = pd.read_csv("captions.txt",names=['image','caption'],sep=',',skiprows=1) 411 | #df = df.groupby("image")['caption'].apply(list).reset_index(name='caption') 412 | #df['caption']=df['caption'].apply(first_caption) 413 | print(df.shape) 414 | print(df.head()) 415 | print(image_df.head()) 416 | joined_df = pd.merge(df,image_df,on='image') 417 | #joined_df.to_csv("data.csv") 418 | 419 | # print("training pairs") 420 | # n_iters = 50 421 | # training_pairs = [tensorsFromPair(random.choice(pairs)) 422 | # for i in range(n_iters)] 423 | # print(training_pairs[0]) 424 | 425 | words = [] 426 | for idx, row in joined_df.iterrows(): 427 | words.extend(joined_df['caption'][idx].lower().split(" ")) 428 | #paragraph context 429 | 430 | words = set(words) 431 | words = list(words) 432 | n_words = len(words)+2 433 | 434 | words_idx = {} 435 | words_idx['']=0 436 | words_idx['']=1 437 | for i in range(len(words)): 438 | words_idx[words[i]]= i+2 439 | 440 | #print(words_idx) 441 | 442 | a_file = open("word_dict.pkl", "wb") 443 | pickle.dump(words_idx, a_file) 444 | 445 | 446 | def tokenizer(x): 447 | embed = [[0]] 448 | 449 | for i in x.lower().split(" "): 450 | 451 | embed.append([words_idx[i]]) 452 | 453 | embed.append([1]) 454 | return embed 455 | joined_df['caption_seq']=joined_df['caption'].apply(tokenizer) 456 | def randomizer(x): 457 | l = x.split(',') 458 | 459 | print('bs',l) 460 | random.shuffle(l) 461 | print('as',l) 462 | return l 463 | #joined_df['image_seq']=joined_df['image_seq'].apply(randomizer) 464 | 465 | 466 | print(joined_df.head()) 467 | print(joined_df.shape) 468 | print(n_words,embedding_input_size) 469 | joined_df.to_csv("last_train_data.csv",sep='\t') 470 | 471 | 472 | hidden_size = 512 473 | encoder1 = EncoderRNN(embedding_input_size, hidden_size).to(device) 474 | attn_decoder1 = AttnDecoderRNN(hidden_size, n_words, dropout_p=0.1).to(device) 475 | trainIters(encoder1, attn_decoder1, 39900, print_every=1000) 476 | torch.save(encoder1.state_dict(),"encoder_model.pkl") 477 | torch.save(attn_decoder1.state_dict(),"decoder_model.pkl") 478 | 479 | else: 480 | import ast 481 | from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction 482 | smoothie = SmoothingFunction().method4 483 | a_file = open("word_dict.pkl", "rb") 484 | words_idx = pickle.load(a_file) 485 | 486 | max_length=25 487 | hidden_size = 512 488 | n_words=8873 489 | embedding_input_size=42446 490 | encoder1 = EncoderRNN(embedding_input_size, hidden_size).to(device) 491 | attn_decoder1 = AttnDecoderRNN(hidden_size, n_words, dropout_p=0.1).to(device) 492 | def eval(jdf,encoder,decoder,i): 493 | input_paths = jdf['image_seq'][i] 494 | sos_img = np.zeros([1,512]) 495 | eos_img = np.ones([1,512]) 496 | 497 | input_np =[] 498 | #input_paths = ast.literal_eval(input_paths) 499 | for i in input_paths.split(","): 500 | 501 | img = Image.open(i) 502 | v = img2vec.get_vec(img,tensor=False).reshape(1,512) 503 | input_np.append(v) 504 | 505 | 506 | input_tensor = torch.tensor(input_np).float().to(device) 507 | input_length = input_tensor.size()[0] 508 | encoder_hidden = encoder.initHidden() 509 | 510 | encoder_outputs = torch.zeros(max_length, encoder.hidden_size, device=device) 511 | 512 | for ei in range(input_length): 513 | encoder_output, encoder_hidden = encoder(input_tensor[ei], 514 | encoder_hidden) 515 | encoder_outputs[ei] += encoder_output[0, 0] 516 | 517 | decoder_input = torch.tensor([[SOS_token]], device=device) # SOS 518 | 519 | decoder_hidden = encoder_hidden 520 | 521 | decoded_words = [] 522 | decoder_attentions = torch.zeros(max_length, max_length) 523 | inv = {} 524 | for key, val in words_idx.items(): 525 | inv[val] = inv.get(val, []) + [key] 526 | for di in range(max_length): 527 | decoder_output, decoder_hidden, decoder_attention = decoder( 528 | decoder_input, decoder_hidden, encoder_outputs) 529 | #decoder_attentions[di] = decoder_attention.data 530 | topv, topi = decoder_output.data.topk(1) 531 | if topi.item() == EOS_token: 532 | decoded_words.append(['']) 533 | break 534 | else: 535 | 536 | decoded_words.append(inv[topi.item()]) 537 | 538 | decoder_input = topi.squeeze().detach() 539 | 540 | return decoded_words 541 | 542 | 543 | ld = pd.read_csv("last_train_data.csv",sep='\t') 544 | ec = torch.load("encoder_model.pkl") 545 | encoder1.load_state_dict(ec) 546 | dc = torch.load("decoder_model.pkl") 547 | attn_decoder1.load_state_dict(dc) 548 | #print(joined_df.head(50)) 549 | preds = {} 550 | refs = {} 551 | print("Prediction-------------------------------------------------") 552 | for i in range(87,100): 553 | print("correct",ld['caption'][i]) 554 | print("prediction-",eval(ld,encoder1,attn_decoder1,i)) 555 | print("-------") 556 | 557 | print("---------------VALIDATION BLEU SCORE----------------------------------- ") 558 | reference_df = pd.read_csv("captions.txt",names=['image','caption'],sep=',',skiprows=1) 559 | reference_df = reference_df[39900:] 560 | reference_df = reference_df.groupby("image")['caption'].apply(list).reset_index(name='caption') 561 | #print(reference_df.head()) 562 | for i in range(reference_df.shape[0]): 563 | refs[reference_df['image'][i]]=(reference_df['caption'][i],i) 564 | 565 | #print(refs) 566 | 567 | mean_bleu_score = 0 568 | n=0 569 | for i in refs.keys(): 570 | true,k = refs[i] 571 | trues=[] 572 | for j in true: 573 | trues.append(j.split(" ")) 574 | 575 | hypothesis = eval(ld,encoder1,attn_decoder1,k) 576 | try: 577 | hypothesis.remove(['']) 578 | hypothesis.remove(['']) 579 | except: 580 | pass 581 | preds = [] 582 | for l in hypothesis: 583 | preds.append(l[0]) 584 | 585 | mean_bleu_score+=sentence_bleu(trues,preds,smoothing_function=smoothie) 586 | n+=1 587 | 588 | print("Mean Bleu Score=",mean_bleu_score/n) 589 | 590 | 591 | 592 | 593 | 594 | 595 | 596 | 597 | 598 | 599 | 600 | 601 | 602 | 603 | --------------------------------------------------------------------------------