├── LICENSE ├── README.md ├── data ├── Pascal_voc │ └── put voc dataset in here.txt ├── data_set │ ├── Images │ │ ├── 000001.jpg │ │ ├── 000002.jpg │ │ └── 000003.jpg │ ├── Labels │ │ ├── 000001.xml │ │ ├── 000002.xml │ │ └── 000003.xml │ ├── test.txt │ └── train.txt └── output │ └── put weights file in here.txt ├── pascal_voc.py ├── preprocess.py ├── test ├── 01.jpg ├── 02.jpg ├── 03.jpg └── yolo v2 demo.gif ├── test_val.py ├── train_val.py └── yolo ├── __init__.py ├── config.py ├── darknet19.py └── yolo_v2.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 | # YOLO_v2 2 | 3 | This is implementation of [YOLO v2](https://arxiv.org/pdf/1612.08242.pdf) with TensorFlow. 4 | 5 | ## Demo 6 | ![](https://github.com/leeyoshinari/YOLO_v2/blob/master/test/yolo%20v2%20demo.gif) 7 | 8 | ## Installation 9 | 1. Clone YOLO_v2 repository 10 | ```Shell 11 | $ git clone https://github.com/leeyoshinari/YOLO_v2.git 12 | $ cd YOLO_v2 13 | ``` 14 | 15 | 2. Download Pascal VOC2007 dataset, and put the dataset into `data/Pascal_voc`. 16 | 17 | If you download other dataset, you also need to modify file paths. 18 | 19 | 3. Download weights file [yolo_weights](https://drive.google.com/drive/folders/13TWYuNY-XcX9EyoU87dH9XsBKuWcPHHw?usp=sharing) for COCO, and put weight file into `data/output`. 20 | 21 | Or you can also download my training weights file [YOLO_v2](https://drive.google.com/drive/folders/14w9JL74VZivk0iD00I3eQYL67bvNyq0N?usp=sharing) for VOC. 22 | 23 | 4. Modify configuration into `yolo/config.py`. 24 | 25 | 5. Training 26 | ```Shell 27 | $ python train_val.py 28 | ``` 29 | 30 | 6. Test 31 | ```Shell 32 | $ python test_val.py 33 | ``` 34 | 7. For more information to [wiki](https://github.com/leeyoshinari/YOLO_v2/wiki/YOLO_v2). 35 | 36 | ## Darknet-19 37 | Darknet-19 has 19 convolutional layers, it's faster than yolo_v2. If you use darknet-19, you need some modifications. It's easy to modify. 38 | 39 | Please download Darknet-19 weights file for VOC from [darknet-19](https://drive.google.com/open?id=1XWWecDpekQ1t2DjhizF-virWyQCTSUeF). 40 | 41 | ## Training on Your Own Dataset 42 | To train the model on your own dataset, you should need to modify: 43 | 44 | 1. Put all the images into the `Images` folder, put all the labels into the `Labels` folder. Select a part of the image for training, write this part of the image filenames into `train.txt`, the remaining part of the image filenames written in `test.txt`. Then put the `Images`, `Labels`, `train.txt` and `test.txt` into `data/dataset`. Put weight file in `data/output`. 45 | 46 | 2. `config.py:` modify the CLASSES. 47 | 48 | 3. `train.py:` replace`from pascal_voc import Pascal_voc` with `from preprocess import Data_preprocess`, and replace `pre_data = Pascal_voc()` with `pre_data = Data_preprocess()`. 49 | 50 | ## Requirements 51 | 1. Tensorflow 52 | 2. OpenCV 53 | -------------------------------------------------------------------------------- /data/Pascal_voc/put voc dataset in here.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/data/Pascal_voc/put voc dataset in here.txt -------------------------------------------------------------------------------- /data/data_set/Images/000001.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/data/data_set/Images/000001.jpg -------------------------------------------------------------------------------- /data/data_set/Images/000002.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/data/data_set/Images/000002.jpg -------------------------------------------------------------------------------- /data/data_set/Images/000003.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/data/data_set/Images/000003.jpg -------------------------------------------------------------------------------- /data/data_set/Labels/000001.xml: -------------------------------------------------------------------------------- 1 | 2 | VOC2007 3 | 000001.jpg 4 | 5 | The VOC2007 Database 6 | PASCAL VOC2007 7 | flickr 8 | 341012865 9 | 10 | 11 | Fried Camels 12 | Jinky the Fruit Bat 13 | 14 | 15 | 353 16 | 500 17 | 3 18 | 19 | 0 20 | 21 | dog 22 | Left 23 | 1 24 | 0 25 | 26 | 48 27 | 240 28 | 195 29 | 371 30 | 31 | 32 | 33 | person 34 | Left 35 | 1 36 | 0 37 | 38 | 8 39 | 12 40 | 352 41 | 498 42 | 43 | 44 | 45 | -------------------------------------------------------------------------------- /data/data_set/Labels/000002.xml: -------------------------------------------------------------------------------- 1 | 2 | VOC2007 3 | 000002.jpg 4 | 5 | The VOC2007 Database 6 | PASCAL VOC2007 7 | flickr 8 | 329145082 9 | 10 | 11 | hiromori2 12 | Hiroyuki Mori 13 | 14 | 15 | 335 16 | 500 17 | 3 18 | 19 | 0 20 | 21 | train 22 | Unspecified 23 | 0 24 | 0 25 | 26 | 139 27 | 200 28 | 207 29 | 301 30 | 31 | 32 | 33 | -------------------------------------------------------------------------------- /data/data_set/Labels/000003.xml: -------------------------------------------------------------------------------- 1 | 2 | VOC2007 3 | 000003.jpg 4 | 5 | The VOC2007 Database 6 | PASCAL VOC2007 7 | flickr 8 | 138563409 9 | 10 | 11 | RandomEvent101 12 | ? 13 | 14 | 15 | 500 16 | 375 17 | 3 18 | 19 | 0 20 | 21 | sofa 22 | Unspecified 23 | 0 24 | 0 25 | 26 | 123 27 | 155 28 | 215 29 | 195 30 | 31 | 32 | 33 | chair 34 | Left 35 | 0 36 | 0 37 | 38 | 239 39 | 156 40 | 307 41 | 205 42 | 43 | 44 | 45 | -------------------------------------------------------------------------------- /data/data_set/test.txt: -------------------------------------------------------------------------------- 1 | 000002 -------------------------------------------------------------------------------- /data/data_set/train.txt: -------------------------------------------------------------------------------- 1 | 000001 2 | 000003 -------------------------------------------------------------------------------- /data/output/put weights file in here.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/data/output/put weights file in here.txt -------------------------------------------------------------------------------- /pascal_voc.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #------------------------------------------------------------------------------------ 5 | import os 6 | import cv2 7 | import numpy as np 8 | import yolo.config as cfg 9 | import xml.etree.ElementTree as ET 10 | 11 | class Pascal_voc(object): 12 | def __init__(self): 13 | self.pascal_voc = os.path.join(cfg.DATA_DIR, 'Pascal_voc') 14 | self.image_size = cfg.IMAGE_SIZE 15 | self.batch_size = cfg.BATCH_SIZE 16 | self.cell_size = cfg.CELL_SIZE 17 | self.classes = cfg.CLASSES 18 | self.num_classes = len(self.classes) 19 | self.box_per_cell = cfg.BOX_PRE_CELL 20 | self.class_to_ind = dict(zip(self.classes, range(self.num_classes))) 21 | 22 | self.count = 0 23 | self.epoch = 1 24 | self.count_t = 0 25 | 26 | def load_labels(self, model): 27 | if model == 'train': 28 | self.devkil_path = os.path.join(self.pascal_voc, 'VOCdevkit') 29 | self.data_path = os.path.join(self.devkil_path, 'VOC2007') 30 | txtname = os.path.join(self.data_path, 'ImageSets', 'Main', 'trainval.txt') 31 | if model == 'test': 32 | self.devkil_path = os.path.join(self.pascal_voc, 'VOCdevkit-test') 33 | self.data_path = os.path.join(self.devkil_path, 'VOC2007') 34 | txtname = os.path.join(self.data_path, 'ImageSets', 'Main', 'test.txt') 35 | 36 | with open(txtname, 'r') as f: 37 | image_ind = [x.strip() for x in f.readlines()] 38 | 39 | labels = [] 40 | for ind in image_ind: 41 | label, num = self.load_data(ind) 42 | if num == 0: 43 | continue 44 | imagename = os.path.join(self.data_path, 'JPEGImages', ind + '.jpg') 45 | labels.append({'imagename': imagename, 'labels': label}) 46 | np.random.shuffle(labels) 47 | return labels 48 | 49 | 50 | def load_data(self, index): 51 | label = np.zeros([self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 52 | filename = os.path.join(self.data_path, 'Annotations', index + '.xml') 53 | tree = ET.parse(filename) 54 | image_size = tree.find('size') 55 | image_width = float(image_size.find('width').text) 56 | image_height = float(image_size.find('height').text) 57 | h_ratio = 1.0 * self.image_size / image_height 58 | w_ratio = 1.0 * self.image_size / image_width 59 | 60 | objects = tree.findall('object') 61 | for obj in objects: 62 | box = obj.find('bndbox') 63 | x1 = max(min((float(box.find('xmin').text)) * w_ratio, self.image_size), 0) 64 | y1 = max(min((float(box.find('ymin').text)) * h_ratio, self.image_size), 0) 65 | x2 = max(min((float(box.find('xmax').text)) * w_ratio, self.image_size), 0) 66 | y2 = max(min((float(box.find('ymax').text)) * h_ratio, self.image_size), 0) 67 | class_ind = self.class_to_ind[obj.find('name').text.lower().strip()] 68 | boxes = [0.5 * (x1 + x2) / self.image_size, 0.5 * (y1 + y2) / self.image_size, np.sqrt((x2 - x1) / self.image_size), np.sqrt((y2 - y1) / self.image_size)] 69 | cx = 1.0 * boxes[0] * self.cell_size 70 | cy = 1.0 * boxes[1] * self.cell_size 71 | xind = int(np.floor(cx)) 72 | yind = int(np.floor(cy)) 73 | 74 | label[yind, xind, :, 0] = 1 75 | label[yind, xind, :, 1:5] = boxes 76 | label[yind, xind, :, 5 + class_ind] = 1 77 | 78 | return label, len(objects) 79 | 80 | 81 | def next_batches(self, label): 82 | images = np.zeros([self.batch_size, self.image_size, self.image_size, 3]) 83 | labels = np.zeros([self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 84 | num = 0 85 | while num < self.batch_size: 86 | imagename = label[self.count]['imagename'] 87 | images[num, :, :, :] = self.image_read(imagename) 88 | labels[num, :, :, :, :] = label[self.count]['labels'] 89 | num += 1 90 | self.count += 1 91 | if self.count >= len(label): 92 | np.random.shuffle(label) 93 | self.count = 0 94 | self.epoch += 1 95 | return images, labels 96 | 97 | 98 | def next_batches_test(self, label): 99 | images = np.zeros([self.batch_size, self.image_size, self.image_size, 3]) 100 | labels = np.zeros([self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 101 | num = 0 102 | while num < self.batch_size: 103 | imagename = label[self.count_t]['imagename'] 104 | images[num, :, :, :] = self.image_read(imagename) 105 | labels[num, :, :, :, :] = label[self.count_t]['labels'] 106 | num += 1 107 | self.count_t += 1 108 | if self.count_t >= len(label): 109 | self.count_t = 0 110 | return images, labels 111 | 112 | 113 | def image_read(self, imagename): 114 | image = cv2.imread(imagename) 115 | image = cv2.resize(image, (self.image_size, self.image_size)) 116 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32) 117 | image = image / 255.0 * 2.0 - 1.0 118 | return image 119 | -------------------------------------------------------------------------------- /preprocess.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #------------------------------------------------------------------------------------ 5 | import os 6 | import cv2 7 | import numpy as np 8 | import yolo.config as cfg 9 | import xml.etree.ElementTree as ET 10 | 11 | class Data_preprocess(object): 12 | def __init__(self): 13 | self.data_path = os.path.join(cfg.DATA_DIR, cfg.DATA_SET) 14 | self.image_size = cfg.IMAGE_SIZE 15 | self.batch_size = cfg.BATCH_SIZE 16 | self.cell_size = cfg.CELL_SIZE 17 | self.classes = cfg.CLASSES 18 | self.num_classes = len(self.classes) 19 | self.box_per_cell = cfg.BOX_PRE_CELL 20 | self.class_to_ind = dict(zip(self.classes, range(self.num_classes))) 21 | 22 | self.count = 0 23 | self.epoch = 1 24 | self.count_t = 0 25 | 26 | def load_labels(self, model): 27 | if model == 'train': 28 | txtname = os.path.join(self.data_path, 'train.txt') 29 | if model == 'test': 30 | txtname = os.path.join(self.data_path, 'test.txt') 31 | 32 | with open(txtname, 'r') as f: 33 | image_ind = [x.strip() for x in f.readlines()] 34 | 35 | labels = [] 36 | for ind in image_ind: 37 | label, num = self.load_data(ind) 38 | if num == 0: 39 | continue 40 | imagename = os.path.join(self.data_path, 'Images', ind + '.jpg') 41 | labels.append({'imagename': imagename, 'labels': label}) 42 | np.random.shuffle(labels) 43 | return labels 44 | 45 | 46 | def load_data(self, index): 47 | label = np.zeros([self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 48 | filename = os.path.join(self.data_path, 'Annotations', index + '.xml') 49 | tree = ET.parse(filename) 50 | image_size = tree.find('size') 51 | image_width = float(image_size.find('width').text) 52 | image_height = float(image_size.find('height').text) 53 | h_ratio = 1.0 * self.image_size / image_height 54 | w_ratio = 1.0 * self.image_size / image_width 55 | 56 | objects = tree.findall('object') 57 | for obj in objects: 58 | box = obj.find('bndbox') 59 | x1 = max(min((float(box.find('xmin').text)) * w_ratio, self.image_size), 0) 60 | y1 = max(min((float(box.find('ymin').text)) * h_ratio, self.image_size), 0) 61 | x2 = max(min((float(box.find('xmax').text)) * w_ratio, self.image_size), 0) 62 | y2 = max(min((float(box.find('ymax').text)) * h_ratio, self.image_size), 0) 63 | class_ind = self.class_to_ind[obj.find('name').text.lower().strip()] 64 | boxes = [0.5 * (x1 + x2) / self.image_size, 0.5 * (y1 + y2) / self.image_size, np.sqrt((x2 - x1) / self.image_size), np.sqrt((y2 - y1) / self.image_size)] 65 | cx = 1.0 * boxes[0] * self.cell_size 66 | cy = 1.0 * boxes[1] * self.cell_size 67 | xind = int(np.floor(cx)) 68 | yind = int(np.floor(cy)) 69 | 70 | label[yind, xind, :, 0] = 1 71 | label[yind, xind, :, 1:5] = boxes 72 | label[yind, xind, :, 5 + class_ind] = 1 73 | 74 | return label, len(objects) 75 | 76 | 77 | def next_batches(self, label): 78 | images = np.zeros([self.batch_size, self.image_size, self.image_size, 3]) 79 | labels = np.zeros([self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 80 | num = 0 81 | while num < self.batch_size: 82 | imagename = label[self.count]['imagename'] 83 | images[num, :, :, :] = self.image_read(imagename) 84 | labels[num, :, :, :, :] = label[self.count]['labels'] 85 | num += 1 86 | self.count += 1 87 | if self.count >= len(label): 88 | np.random.shuffle(label) 89 | self.count = 0 90 | self.epoch += 1 91 | return images, labels 92 | 93 | 94 | def next_batches_test(self, label): 95 | images = np.zeros([self.batch_size, self.image_size, self.image_size, 3]) 96 | labels = np.zeros([self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 97 | num = 0 98 | while num < self.batch_size: 99 | imagename = label[self.count_t]['imagename'] 100 | images[num, :, :, :] = self.image_read(imagename) 101 | labels[num, :, :, :, :] = label[self.count_t]['labels'] 102 | num += 1 103 | self.count_t += 1 104 | if self.count_t >= len(label): 105 | self.count_t = 0 106 | return images, labels 107 | 108 | 109 | def image_read(self, imagename): 110 | image = cv2.imread(imagename) 111 | image = cv2.resize(image, (self.image_size, self.image_size)) 112 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32) 113 | image = image / 255.0 * 2.0 - 1.0 114 | return image 115 | -------------------------------------------------------------------------------- /test/01.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/test/01.jpg -------------------------------------------------------------------------------- /test/02.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/test/02.jpg -------------------------------------------------------------------------------- /test/03.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/test/03.jpg -------------------------------------------------------------------------------- /test/yolo v2 demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/test/yolo v2 demo.gif -------------------------------------------------------------------------------- /test_val.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #----------------------------------------------------------------------------------- 5 | import tensorflow as tf 6 | import numpy as np 7 | import argparse 8 | import colorsys 9 | import cv2 10 | import os 11 | 12 | import yolo.config as cfg 13 | from yolo.yolo_v2 import yolo_v2 14 | # from yolo.darknet19 import Darknet19 15 | 16 | class Detector(object): 17 | def __init__(self, yolo, weights_file): 18 | self.yolo = yolo 19 | self.classes = cfg.CLASSES 20 | self.num_classes = len(self.classes) 21 | self.image_size = cfg.IMAGE_SIZE 22 | self.cell_size = cfg.CELL_SIZE 23 | self.batch_size = cfg.BATCH_SIZE 24 | self.box_per_cell = cfg.BOX_PRE_CELL 25 | self.threshold = cfg.THRESHOLD 26 | self.anchor = cfg.ANCHOR 27 | 28 | self.sess = tf.Session() 29 | self.sess.run(tf.global_variables_initializer()) 30 | 31 | print('Restore weights from: ' + weights_file) 32 | self.saver = tf.train.Saver() 33 | self.saver.restore(self.sess, weights_file) 34 | 35 | def detect(self, image): 36 | image_h, image_w, _ = image.shape 37 | image = cv2.resize(image, (self.image_size, self.image_size)) 38 | image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32) 39 | image = image / 255.0 * 2.0 - 1.0 40 | image = np.reshape(image, [1, self.image_size, self.image_size, 3]) 41 | 42 | output = self.sess.run(self.yolo.logits, feed_dict = {self.yolo.images: image}) 43 | 44 | results = self.calc_output(output) 45 | 46 | for i in range(len(results)): 47 | results[i][1] *= (1.0 * image_w / self.image_size) 48 | results[i][2] *= (1.0 * image_h / self.image_size) 49 | results[i][3] *= (1.0 * image_w / self.image_size) 50 | results[i][4] *= (1.0 * image_h / self.image_size) 51 | 52 | return results 53 | 54 | 55 | def calc_output(self, output): 56 | output = np.reshape(output, [self.cell_size, self.cell_size, self.box_per_cell, 5 + self.num_classes]) 57 | boxes = np.reshape(output[:, :, :, :4], [self.cell_size, self.cell_size, self.box_per_cell, 4]) #boxes coordinate 58 | boxes = self.get_boxes(boxes) * self.image_size 59 | 60 | confidence = np.reshape(output[:, :, :, 4], [self.cell_size, self.cell_size, self.box_per_cell]) #the confidence of the each anchor boxes 61 | confidence = 1.0 / (1.0 + np.exp(-1.0 * confidence)) 62 | confidence = np.tile(np.expand_dims(confidence, 3), (1, 1, 1, self.num_classes)) 63 | 64 | classes = np.reshape(output[:, :, :, 5:], [self.cell_size, self.cell_size, self.box_per_cell, self.num_classes]) #classes 65 | classes = np.exp(classes) / np.tile(np.expand_dims(np.sum(np.exp(classes), axis=3), axis=3), (1, 1, 1, self.num_classes)) 66 | 67 | probs = classes * confidence 68 | 69 | filter_probs = np.array(probs >= self.threshold, dtype = 'bool') 70 | filter_index = np.nonzero(filter_probs) 71 | box_filter = boxes[filter_index[0], filter_index[1], filter_index[2]] 72 | probs_filter = probs[filter_probs] 73 | classes_num = np.argmax(filter_probs, axis = 3)[filter_index[0], filter_index[1], filter_index[2]] 74 | 75 | sort_num = np.array(np.argsort(probs_filter))[::-1] 76 | box_filter = box_filter[sort_num] 77 | probs_filter = probs_filter[sort_num] 78 | classes_num = classes_num[sort_num] 79 | 80 | for i in range(len(probs_filter)): 81 | if probs_filter[i] == 0: 82 | continue 83 | for j in range(i+1, len(probs_filter)): 84 | if self.calc_iou(box_filter[i], box_filter[j]) > 0.5: 85 | probs_filter[j] = 0.0 86 | 87 | filter_probs = np.array(probs_filter > 0, dtype = 'bool') 88 | probs_filter = probs_filter[filter_probs] 89 | box_filter = box_filter[filter_probs] 90 | classes_num = classes_num[filter_probs] 91 | 92 | results = [] 93 | for i in range(len(probs_filter)): 94 | results.append([self.classes[classes_num[i]], box_filter[i][0], box_filter[i][1], 95 | box_filter[i][2], box_filter[i][3], probs_filter[i]]) 96 | 97 | return results 98 | 99 | def get_boxes(self, boxes): 100 | offset = np.transpose(np.reshape(np.array([np.arange(self.cell_size)] * self.cell_size * self.box_per_cell), 101 | [self.box_per_cell, self.cell_size, self.cell_size]), (1, 2, 0)) 102 | boxes1 = np.stack([(1.0 / (1.0 + np.exp(-1.0 * boxes[:, :, :, 0])) + offset) / self.cell_size, 103 | (1.0 / (1.0 + np.exp(-1.0 * boxes[:, :, :, 1])) + np.transpose(offset, (1, 0, 2))) / self.cell_size, 104 | np.exp(boxes[:, :, :, 2]) * np.reshape(self.anchor[:5], [1, 1, 5]) / self.cell_size, 105 | np.exp(boxes[:, :, :, 3]) * np.reshape(self.anchor[5:], [1, 1, 5]) / self.cell_size]) 106 | 107 | return np.transpose(boxes1, (1, 2, 3, 0)) 108 | 109 | 110 | def calc_iou(self, box1, box2): 111 | width = min(box1[0] + 0.5 * box1[2], box2[0] + 0.5 * box2[2]) - max(box1[0] - 0.5 * box1[2], box2[0] - 0.5 * box2[2]) 112 | height = min(box1[1] + 0.5 * box1[3], box2[1] + 0.5 * box2[3]) - max(box1[1] - 0.5 * box1[3], box2[1] - 0.5 * box2[3]) 113 | 114 | if width <= 0 or height <= 0: 115 | intersection = 0 116 | else: 117 | intersection = width * height 118 | 119 | return intersection / (box1[2] * box1[3] + box2[2] * box2[3] - intersection) 120 | 121 | def random_colors(self, N, bright=True): 122 | brightness = 1.0 if bright else 0.7 123 | hsv = [(i / N, 1, brightness) for i in range(N)] 124 | colors = list(map(lambda c: colorsys.hsv_to_rgb(*c), hsv)) 125 | np.random.shuffle(colors) 126 | return colors 127 | 128 | 129 | def draw(self, image, result): 130 | image_h, image_w, _ = image.shape 131 | colors = self.random_colors(len(result)) 132 | for i in range(len(result)): 133 | xmin = max(int(result[i][1] - 0.5 * result[i][3]), 0) 134 | ymin = max(int(result[i][2] - 0.5 * result[i][4]), 0) 135 | xmax = min(int(result[i][1] + 0.5 * result[i][3]), image_w) 136 | ymax = min(int(result[i][2] + 0.5 * result[i][4]), image_h) 137 | color = tuple([rgb * 255 for rgb in colors[i]]) 138 | cv2.rectangle(image, (xmin, ymin), (xmax, ymax), color, 1) 139 | cv2.putText(image, result[i][0] + ':%.2f' % result[i][5], (xmin + 1, ymin + 8), cv2.FONT_HERSHEY_COMPLEX_SMALL, 0.5, color, 1) 140 | print(result[i][0], ':%.2f%%' % (result[i][5] * 100 )) 141 | 142 | 143 | def image_detect(self, imagename): 144 | image = cv2.imread(imagename) 145 | result = self.detect(image) 146 | self.draw(image, result) 147 | cv2.imshow('Image', image) 148 | cv2.waitKey(0) 149 | 150 | 151 | def video_detect(self, cap): 152 | while(1): 153 | ret, image = cap.read() 154 | if not ret: 155 | print('Cannot capture images from device') 156 | break 157 | 158 | result = self.detect(image) 159 | self.draw(image, result) 160 | cv2.imshow('Image', image) 161 | 162 | if cv2.waitKey(10) & 0xFF == ord('q'): 163 | break 164 | cap.release() 165 | cv2.destroyAllWindows() 166 | 167 | 168 | def main(): 169 | parser = argparse.ArgumentParser() 170 | parser.add_argument('--weights', default = 'yolo_v2.ckpt', type = str) # darknet-19.ckpt 171 | parser.add_argument('--weight_dir', default = 'output', type = str) 172 | parser.add_argument('--data_dir', default = 'data', type = str) 173 | parser.add_argument('--gpu', default = '', type = str) # which gpu to be selected 174 | args = parser.parse_args() 175 | 176 | os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu # configure gpu 177 | weights_file = os.path.join(args.data_dir, args.weight_dir, args.weights) 178 | yolo = yolo_v2(False) # 'False' mean 'test' 179 | # yolo = Darknet19(False) 180 | 181 | detector = Detector(yolo, weights_file) 182 | 183 | #detect the video 184 | #cap = cv2.VideoCapture('asd.mp4') 185 | #cap = cv2.VideoCapture(0) 186 | #detector.video_detect(cap) 187 | 188 | #detect the image 189 | imagename = './test/01.jpg' 190 | detector.image_detect(imagename) 191 | 192 | if __name__ == '__main__': 193 | main() 194 | -------------------------------------------------------------------------------- /train_val.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #----------------------------------------------------------------------------------- 5 | import tensorflow as tf 6 | import numpy as np 7 | import argparse 8 | import datetime 9 | import time 10 | import os 11 | import yolo.config as cfg 12 | 13 | from pascal_voc import Pascal_voc 14 | from six.moves import xrange 15 | from yolo.yolo_v2 import yolo_v2 16 | # from yolo.darknet19 import Darknet19 17 | 18 | class Train(object): 19 | def __init__(self, yolo, data): 20 | self.yolo = yolo 21 | self.data = data 22 | self.num_class = len(cfg.CLASSES) 23 | self.max_step = cfg.MAX_ITER 24 | self.saver_iter = cfg.SAVER_ITER 25 | self.summary_iter = cfg.SUMMARY_ITER 26 | self.initial_learn_rate = cfg.LEARN_RATE 27 | self.output_dir = os.path.join(cfg.DATA_DIR, 'output') 28 | weight_file = os.path.join(self.output_dir, cfg.WEIGHTS_FILE) 29 | 30 | self.variable_to_restore = tf.global_variables() 31 | self.saver = tf.train.Saver(self.variable_to_restore) 32 | self.summary_op = tf.summary.merge_all() 33 | self.writer = tf.summary.FileWriter(self.output_dir) 34 | 35 | self.global_step = tf.get_variable('global_step', [], initializer=tf.constant_initializer(0), trainable=False) 36 | self.learn_rate = tf.train.exponential_decay(self.initial_learn_rate, self.global_step, 20000, 0.1, name='learn_rate') 37 | # self.global_step = tf.Variable(0, trainable = False) 38 | # self.learn_rate = tf.train.piecewise_constant(self.global_step, [100, 190, 10000, 15500], [1e-3, 5e-3, 1e-2, 1e-3, 1e-4]) 39 | self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learn_rate).minimize(self.yolo.total_loss, global_step=self.global_step) 40 | 41 | self.average_op = tf.train.ExponentialMovingAverage(0.999).apply(tf.trainable_variables()) 42 | with tf.control_dependencies([self.optimizer]): 43 | self.train_op = tf.group(self.average_op) 44 | 45 | config = tf.ConfigProto(gpu_options=tf.GPUOptions()) 46 | self.sess = tf.Session(config=config) 47 | self.sess.run(tf.global_variables_initializer()) 48 | 49 | print('Restore weights from:', weight_file) 50 | self.saver.restore(self.sess, weight_file) 51 | self.writer.add_graph(self.sess.graph) 52 | 53 | def train(self): 54 | labels_train = self.data.load_labels('train') 55 | labels_test = self.data.load_labels('test') 56 | 57 | num = 5 58 | initial_time = time.time() 59 | 60 | for step in xrange(0, self.max_step + 1): 61 | images, labels = self.data.next_batches(labels_train) 62 | feed_dict = {self.yolo.images: images, self.yolo.labels: labels} 63 | 64 | if step % self.summary_iter == 0: 65 | if step % 50 == 0: 66 | summary_, loss, _ = self.sess.run([self.summary_op, self.yolo.total_loss, self.train_op], feed_dict = feed_dict) 67 | sum_loss = 0 68 | 69 | for i in range(num): 70 | images_t, labels_t = self.data.next_batches_test(labels_test) 71 | feed_dict_t = {self.yolo.images: images_t, self.yolo.labels: labels_t} 72 | loss_t = self.sess.run(self.yolo.total_loss, feed_dict=feed_dict_t) 73 | sum_loss += loss_t 74 | 75 | log_str = ('{} Epoch: {}, Step: {}, train_Loss: {:.4f}, test_Loss: {:.4f}, Remain: {}').format( 76 | datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'), self.data.epoch, int(step), loss, sum_loss/num, self.remain(step, initial_time)) 77 | print(log_str) 78 | 79 | if loss < 1e4: 80 | pass 81 | else: 82 | print('loss > 1e04') 83 | break 84 | 85 | else: 86 | summary_, _ = self.sess.run([self.summary_op, self.train_op], feed_dict = feed_dict) 87 | 88 | self.writer.add_summary(summary_, step) 89 | 90 | else: 91 | self.sess.run(self.train_op, feed_dict = feed_dict) 92 | 93 | if step % self.saver_iter == 0: 94 | self.saver.save(self.sess, self.output_dir + '/yolo_v2.ckpt', global_step = step) 95 | 96 | def remain(self, i, start): 97 | if i == 0: 98 | remain_time = 0 99 | else: 100 | remain_time = (time.time() - start) * (self.max_step - i) / i 101 | return str(datetime.timedelta(seconds = int(remain_time))) 102 | 103 | 104 | def main(): 105 | parser = argparse.ArgumentParser() 106 | parser.add_argument('--weights', default = 'yolo_v2.ckpt', type = str) # darknet-19.ckpt 107 | parser.add_argument('--gpu', default = '', type = str) # which gpu to be selected 108 | args = parser.parse_args() 109 | 110 | if args.gpu is not None: 111 | cfg.GPU = args.gpu 112 | 113 | if args.weights is not None: 114 | cfg.WEIGHTS_FILE = args.weights 115 | 116 | os.environ['CUDA_VISIBLE_DEVICES'] = cfg.GPU 117 | yolo = yolo_v2() 118 | # yolo = Darknet19() 119 | pre_data = Pascal_voc() 120 | 121 | train = Train(yolo, pre_data) 122 | 123 | print('start training ...') 124 | train.train() 125 | print('successful training.') 126 | 127 | 128 | if __name__ == '__main__': 129 | main() 130 | -------------------------------------------------------------------------------- /yolo/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/leeyoshinari/YOLO_v2/aa9bf78ecf461a1a3924e466024f8417d23197aa/yolo/__init__.py -------------------------------------------------------------------------------- /yolo/config.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | 5 | DATA_DIR = 'data' 6 | DATA_SET = 'data_set' 7 | WEIGHTS_FILE = 'yolo_weights.ckpt' 8 | 9 | CLASSES = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 10 | 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 11 | 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 12 | 'train', 'tvmonitor'] 13 | 14 | #ANCHOR = [0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828] 15 | ANCHOR = [0.57273, 1.87446, 3.33843, 7.88282, 9.77052, 0.677385, 2.06253, 5.47434, 3.52778, 9.16828] 16 | 17 | GPU = '' 18 | 19 | IMAGE_SIZE = 416 #The size of the input images 20 | 21 | LEARN_RATE = 0.0001 #The learn_rate of training 22 | MAX_ITER = 20000 #The max step 23 | SUMMARY_ITER = 5 #Every 'summary_iter' step output a summary 24 | SAVER_ITER = 50 #Every 'saver_iter' step save a weights 25 | 26 | BOX_PRE_CELL = 5 #The number of BoundingBoxs predicted by each grid 27 | CELL_SIZE = 13 #The size of the last layer #(batch_size, 13, 13, ?) 28 | BATCH_SIZE = 32 #The batch size of each training 29 | ALPHA = 0.1 30 | 31 | THRESHOLD = 0.3 #The threshold of the probability of the classes 32 | -------------------------------------------------------------------------------- /yolo/darknet19.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #----------------------------------------------------------------------------------- 5 | 6 | import tensorflow as tf 7 | import numpy as np 8 | import yolo.config as cfg 9 | 10 | class Darknet19(object): 11 | def __init__(self, isTraining = True): 12 | self.classes = cfg.CLASSES 13 | self.num_class = len(self.classes) 14 | 15 | self.box_per_cell = cfg.BOX_PRE_CELL 16 | self.cell_size = cfg.CELL_SIZE 17 | self.batch_size = cfg.BATCH_SIZE 18 | self.image_size = cfg.IMAGE_SIZE 19 | self.anchor = cfg.ANCHOR 20 | self.alpha = cfg.ALPHA 21 | 22 | self.class_scale = 1.0 23 | self.object_scale = 5.0 24 | self.noobject_scale = 1.0 25 | self.coordinate_scale = 1.0 26 | 27 | self.offset = np.transpose(np.reshape(np.array([np.arange(self.cell_size)] * self.cell_size * self.box_per_cell), 28 | [self.box_per_cell, self.cell_size, self.cell_size]), (1, 2, 0)) 29 | self.offset = tf.reshape(tf.constant(self.offset, dtype=tf.float32), [1, self.cell_size, self.cell_size, self.box_per_cell]) 30 | self.offset = tf.tile(self.offset, (self.batch_size, 1, 1, 1)) 31 | 32 | self.images = tf.placeholder(tf.float32, [None, self.image_size, self.image_size, 3], name='images') 33 | self.logits = self.build_networks(self.images) 34 | 35 | if isTraining: 36 | self.labels = tf.placeholder(tf.float32, [None, self.cell_size, self.cell_size, self.box_per_cell, self.num_class + 5], name = 'labels') 37 | self.total_loss = self.loss_layer(self.logits, self.labels) 38 | tf.summary.scalar('total_loss', self.total_loss) 39 | 40 | def build_networks(self, inputs): 41 | net = self.conv_layer(inputs, [3, 3, 3, 32], name = '0_conv') 42 | net = self.pooling_layer(net, name = '1_pool') 43 | 44 | net = self.conv_layer(net, [3, 3, 32, 64], name = '2_conv') 45 | net = self.pooling_layer(net, name = '3_pool') 46 | 47 | net = self.conv_layer(net, [3, 3, 64, 128], name = '4_conv') 48 | net = self.conv_layer(net, [1, 1, 128, 64], name = '5_conv') 49 | net = self.conv_layer(net, [3, 3, 64, 128], name = '6_conv') 50 | net = self.pooling_layer(net, name = '7_pool') 51 | 52 | net = self.conv_layer(net, [3, 3, 128, 256], name = '8_conv') 53 | net = self.conv_layer(net, [1, 1, 256, 128], name = '9_conv') 54 | net = self.conv_layer(net, [3, 3, 128, 256], name = '10_conv') 55 | net = self.pooling_layer(net, name = '11_pool') 56 | 57 | net = self.conv_layer(net, [3, 3, 256, 512], name = '12_conv') 58 | net = self.conv_layer(net, [1, 1, 512, 256], name = '13_conv') 59 | net = self.conv_layer(net, [3, 3, 256, 512], name = '14_conv') 60 | net = self.conv_layer(net, [1, 1, 512, 256], name = '15_conv') 61 | net16 = self.conv_layer(net, [3, 3, 256, 512], name = '16_conv') 62 | net = self.pooling_layer(net16, name = '17_pool') 63 | 64 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '18_conv') 65 | net = self.conv_layer(net, [1, 1, 1024, 512], name = '19_conv') 66 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '20_conv') 67 | net = self.conv_layer(net, [1, 1, 1024, 512], name = '21_conv') 68 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '22_conv') 69 | 70 | net = self.conv_layer(net, [1, 1, 1024, self.box_per_cell * (self.num_class + 5)], batch_norm=False, name = '23_conv') 71 | 72 | return net 73 | 74 | 75 | def conv_layer(self, inputs, shape, batch_norm = True, name = '0_conv'): 76 | weight = tf.Variable(tf.truncated_normal(shape, stddev=0.1), name='weight') 77 | biases = tf.Variable(tf.constant(0.1, shape=[shape[3]]), name='biases') 78 | 79 | conv = tf.nn.conv2d(inputs, weight, strides=[1, 1, 1, 1], padding='SAME', name=name) 80 | 81 | if batch_norm: 82 | depth = shape[3] 83 | scale = tf.Variable(tf.ones([depth, ], dtype='float32'), name='scale') 84 | shift = tf.Variable(tf.zeros([depth, ], dtype='float32'), name='shift') 85 | mean = tf.Variable(tf.ones([depth, ], dtype='float32'), name='rolling_mean') 86 | variance = tf.Variable(tf.ones([depth, ], dtype='float32'), name='rolling_variance') 87 | 88 | conv_bn = tf.nn.batch_normalization(conv, mean, variance, shift, scale, 1e-05) 89 | conv = tf.add(conv_bn, biases) 90 | conv = tf.maximum(self.alpha * conv, conv) 91 | else: 92 | conv = tf.add(conv, biases) 93 | 94 | return conv 95 | 96 | 97 | def pooling_layer(self, inputs, name = '1_pool'): 98 | pool = tf.nn.max_pool(inputs, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = 'SAME', name = name) 99 | return pool 100 | 101 | 102 | def loss_layer(self, predict, label): 103 | predict = tf.reshape(predict, [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class + 5]) 104 | box_coordinate = tf.reshape(predict[:, :, :, :, :4], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 4]) 105 | box_confidence = tf.reshape(predict[:, :, :, :, 4], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 1]) 106 | box_classes = tf.reshape(predict[:, :, :, :, 5:], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class]) 107 | 108 | boxes1 = tf.stack([(1.0 / (1.0 + tf.exp(-1.0 * box_coordinate[:, :, :, :, 0])) + self.offset) / self.cell_size, 109 | (1.0 / (1.0 + tf.exp(-1.0 * box_coordinate[:, :, :, :, 1])) + tf.transpose(self.offset, (0, 2, 1, 3))) / self.cell_size, 110 | tf.sqrt(tf.exp(box_coordinate[:, :, :, :, 2]) * np.reshape(self.anchor[:5], [1, 1, 1, 5]) / self.cell_size), 111 | tf.sqrt(tf.exp(box_coordinate[:, :, :, :, 3]) * np.reshape(self.anchor[5:], [1, 1, 1, 5]) / self.cell_size)]) 112 | box_coor_trans = tf.transpose(boxes1, (1, 2, 3, 4, 0)) 113 | box_confidence = 1.0 / (1.0 + tf.exp(-1.0 * box_confidence)) 114 | box_classes = tf.nn.softmax(box_classes) 115 | 116 | response = tf.reshape(label[:, :, :, :, 0], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell]) 117 | boxes = tf.reshape(label[:, :, :, :, 1:5], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 4]) 118 | classes = tf.reshape(label[:, :, :, :, 5:], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class]) 119 | 120 | iou = self.calc_iou(box_coor_trans, boxes) 121 | best_box = tf.to_float(tf.equal(iou, tf.reduce_max(iou, axis=-1, keep_dims=True))) 122 | confs = tf.expand_dims(best_box * response, axis = 4) 123 | 124 | conid = self.noobject_scale * (1.0 - confs) + self.object_scale * confs 125 | cooid = self.coordinate_scale * confs 126 | proid = self.class_scale * confs 127 | 128 | coo_loss = cooid * tf.square(box_coor_trans - boxes) 129 | con_loss = conid * tf.square(box_confidence - confs) 130 | pro_loss = proid * tf.square(box_classes - classes) 131 | 132 | loss = tf.concat([coo_loss, con_loss, pro_loss], axis = 4) 133 | loss = tf.reduce_mean(tf.reduce_sum(loss, axis = [1, 2, 3, 4]), name = 'loss') 134 | 135 | return loss 136 | 137 | 138 | def calc_iou(self, boxes1, boxes2): 139 | boxx = tf.square(boxes1[:, :, :, :, 2:4]) 140 | boxes1_square = boxx[:, :, :, :, 0] * boxx[:, :, :, :, 1] 141 | box = tf.stack([boxes1[:, :, :, :, 0] - boxx[:, :, :, :, 0] * 0.5, 142 | boxes1[:, :, :, :, 1] - boxx[:, :, :, :, 1] * 0.5, 143 | boxes1[:, :, :, :, 0] + boxx[:, :, :, :, 0] * 0.5, 144 | boxes1[:, :, :, :, 1] + boxx[:, :, :, :, 1] * 0.5]) 145 | boxes1 = tf.transpose(box, (1, 2, 3, 4, 0)) 146 | 147 | boxx = tf.square(boxes2[:, :, :, :, 2:4]) 148 | boxes2_square = boxx[:, :, :, :, 0] * boxx[:, :, :, :, 1] 149 | box = tf.stack([boxes2[:, :, :, :, 0] - boxx[:, :, :, :, 0] * 0.5, 150 | boxes2[:, :, :, :, 1] - boxx[:, :, :, :, 1] * 0.5, 151 | boxes2[:, :, :, :, 0] + boxx[:, :, :, :, 0] * 0.5, 152 | boxes2[:, :, :, :, 1] + boxx[:, :, :, :, 1] * 0.5]) 153 | boxes2 = tf.transpose(box, (1, 2, 3, 4, 0)) 154 | 155 | left_up = tf.maximum(boxes1[:, :, :, :, :2], boxes2[:, :, :, :, :2]) 156 | right_down = tf.minimum(boxes1[:, :, :, :, 2:], boxes2[:, :, :, :, 2:]) 157 | 158 | intersection = tf.maximum(right_down - left_up, 0.0) 159 | inter_square = intersection[:, :, :, :, 0] * intersection[:, :, :, :, 1] 160 | union_square = boxes1_square + boxes2_square - inter_square 161 | 162 | return tf.clip_by_value(1.0 * inter_square / union_square, 0.0, 1.0) 163 | -------------------------------------------------------------------------------- /yolo/yolo_v2.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding:utf-8 -*- 3 | # Author:leeyoshinari 4 | #----------------------------------------------------------------------------------- 5 | 6 | import tensorflow as tf 7 | import numpy as np 8 | import yolo.config as cfg 9 | 10 | class yolo_v2(object): 11 | def __init__(self, isTraining = True): 12 | self.classes = cfg.CLASSES 13 | self.num_class = len(self.classes) 14 | 15 | self.box_per_cell = cfg.BOX_PRE_CELL 16 | self.cell_size = cfg.CELL_SIZE 17 | self.batch_size = cfg.BATCH_SIZE 18 | self.image_size = cfg.IMAGE_SIZE 19 | self.anchor = cfg.ANCHOR 20 | self.alpha = cfg.ALPHA 21 | 22 | self.class_scale = 1.0 23 | self.object_scale = 5.0 24 | self.noobject_scale = 1.0 25 | self.coordinate_scale = 1.0 26 | 27 | self.offset = np.transpose(np.reshape(np.array([np.arange(self.cell_size)] * self.cell_size * self.box_per_cell), 28 | [self.box_per_cell, self.cell_size, self.cell_size]), (1, 2, 0)) 29 | self.offset = tf.reshape(tf.constant(self.offset, dtype=tf.float32), [1, self.cell_size, self.cell_size, self.box_per_cell]) 30 | self.offset = tf.tile(self.offset, (self.batch_size, 1, 1, 1)) 31 | 32 | self.images = tf.placeholder(tf.float32, [None, self.image_size, self.image_size, 3], name='images') 33 | self.logits = self.build_networks(self.images) 34 | 35 | if isTraining: 36 | self.labels = tf.placeholder(tf.float32, [None, self.cell_size, self.cell_size, self.box_per_cell, self.num_class + 5], name = 'labels') 37 | self.total_loss = self.loss_layer(self.logits, self.labels) 38 | tf.summary.scalar('total_loss', self.total_loss) 39 | 40 | def build_networks(self, inputs): 41 | net = self.conv_layer(inputs, [3, 3, 3, 32], name = '0_conv') 42 | net = self.pooling_layer(net, name = '1_pool') 43 | 44 | net = self.conv_layer(net, [3, 3, 32, 64], name = '2_conv') 45 | net = self.pooling_layer(net, name = '3_pool') 46 | 47 | net = self.conv_layer(net, [3, 3, 64, 128], name = '4_conv') 48 | net = self.conv_layer(net, [1, 1, 128, 64], name = '5_conv') 49 | net = self.conv_layer(net, [3, 3, 64, 128], name = '6_conv') 50 | net = self.pooling_layer(net, name = '7_pool') 51 | 52 | net = self.conv_layer(net, [3, 3, 128, 256], name = '8_conv') 53 | net = self.conv_layer(net, [1, 1, 256, 128], name = '9_conv') 54 | net = self.conv_layer(net, [3, 3, 128, 256], name = '10_conv') 55 | net = self.pooling_layer(net, name = '11_pool') 56 | 57 | net = self.conv_layer(net, [3, 3, 256, 512], name = '12_conv') 58 | net = self.conv_layer(net, [1, 1, 512, 256], name = '13_conv') 59 | net = self.conv_layer(net, [3, 3, 256, 512], name = '14_conv') 60 | net = self.conv_layer(net, [1, 1, 512, 256], name = '15_conv') 61 | net16 = self.conv_layer(net, [3, 3, 256, 512], name = '16_conv') 62 | net = self.pooling_layer(net16, name = '17_pool') 63 | 64 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '18_conv') 65 | net = self.conv_layer(net, [1, 1, 1024, 512], name = '19_conv') 66 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '20_conv') 67 | net = self.conv_layer(net, [1, 1, 1024, 512], name = '21_conv') 68 | net = self.conv_layer(net, [3, 3, 512, 1024], name = '22_conv') 69 | 70 | net = self.conv_layer(net, [3, 3, 1024, 1024], name = '23_conv') 71 | net24 = self.conv_layer(net, [3, 3, 1024, 1024], name = '24_conv') 72 | 73 | net = self.conv_layer(net16, [1, 1, 512, 64], name = '26_conv') 74 | net = self.reorg(net) 75 | 76 | net = tf.concat([net, net24], 3) 77 | 78 | net = self.conv_layer(net, [3, 3, int(net.get_shape()[3]), 1024], name = '29_conv') 79 | net = self.conv_layer(net, [1, 1, 1024, self.box_per_cell * (self.num_class + 5)], batch_norm=False, name = '30_conv') 80 | 81 | return net 82 | 83 | 84 | def conv_layer(self, inputs, shape, batch_norm = True, name = '0_conv'): 85 | weight = tf.Variable(tf.truncated_normal(shape, stddev=0.1), name='weight') 86 | biases = tf.Variable(tf.constant(0.1, shape=[shape[3]]), name='biases') 87 | 88 | conv = tf.nn.conv2d(inputs, weight, strides=[1, 1, 1, 1], padding='SAME', name=name) 89 | 90 | if batch_norm: 91 | depth = shape[3] 92 | scale = tf.Variable(tf.ones([depth, ], dtype='float32'), name='scale') 93 | shift = tf.Variable(tf.zeros([depth, ], dtype='float32'), name='shift') 94 | mean = tf.Variable(tf.ones([depth, ], dtype='float32'), name='rolling_mean') 95 | variance = tf.Variable(tf.ones([depth, ], dtype='float32'), name='rolling_variance') 96 | 97 | conv_bn = tf.nn.batch_normalization(conv, mean, variance, shift, scale, 1e-05) 98 | conv = tf.add(conv_bn, biases) 99 | conv = tf.maximum(self.alpha * conv, conv) 100 | else: 101 | conv = tf.add(conv, biases) 102 | 103 | return conv 104 | 105 | 106 | def pooling_layer(self, inputs, name = '1_pool'): 107 | pool = tf.nn.max_pool(inputs, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = 'SAME', name = name) 108 | return pool 109 | 110 | 111 | def reorg(self, inputs): 112 | outputs_1 = inputs[:, ::2, ::2, :] 113 | outputs_2 = inputs[:, ::2, 1::2, :] 114 | outputs_3 = inputs[:, 1::2, ::2, :] 115 | outputs_4 = inputs[:, 1::2, 1::2, :] 116 | output = tf.concat([outputs_1, outputs_2, outputs_3, outputs_4], axis = 3) 117 | return output 118 | 119 | 120 | def loss_layer(self, predict, label): 121 | predict = tf.reshape(predict, [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class + 5]) 122 | box_coordinate = tf.reshape(predict[:, :, :, :, :4], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 4]) 123 | box_confidence = tf.reshape(predict[:, :, :, :, 4], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 1]) 124 | box_classes = tf.reshape(predict[:, :, :, :, 5:], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class]) 125 | 126 | boxes1 = tf.stack([(1.0 / (1.0 + tf.exp(-1.0 * box_coordinate[:, :, :, :, 0])) + self.offset) / self.cell_size, 127 | (1.0 / (1.0 + tf.exp(-1.0 * box_coordinate[:, :, :, :, 1])) + tf.transpose(self.offset, (0, 2, 1, 3))) / self.cell_size, 128 | tf.sqrt(tf.exp(box_coordinate[:, :, :, :, 2]) * np.reshape(self.anchor[:5], [1, 1, 1, 5]) / self.cell_size), 129 | tf.sqrt(tf.exp(box_coordinate[:, :, :, :, 3]) * np.reshape(self.anchor[5:], [1, 1, 1, 5]) / self.cell_size)]) 130 | box_coor_trans = tf.transpose(boxes1, (1, 2, 3, 4, 0)) 131 | box_confidence = 1.0 / (1.0 + tf.exp(-1.0 * box_confidence)) 132 | box_classes = tf.nn.softmax(box_classes) 133 | 134 | response = tf.reshape(label[:, :, :, :, 0], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell]) 135 | boxes = tf.reshape(label[:, :, :, :, 1:5], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, 4]) 136 | classes = tf.reshape(label[:, :, :, :, 5:], [self.batch_size, self.cell_size, self.cell_size, self.box_per_cell, self.num_class]) 137 | 138 | iou = self.calc_iou(box_coor_trans, boxes) 139 | best_box = tf.to_float(tf.equal(iou, tf.reduce_max(iou, axis=-1, keep_dims=True))) 140 | confs = tf.expand_dims(best_box * response, axis = 4) 141 | 142 | conid = self.noobject_scale * (1.0 - confs) + self.object_scale * confs 143 | cooid = self.coordinate_scale * confs 144 | proid = self.class_scale * confs 145 | 146 | coo_loss = cooid * tf.square(box_coor_trans - boxes) 147 | con_loss = conid * tf.square(box_confidence - confs) 148 | pro_loss = proid * tf.square(box_classes - classes) 149 | 150 | loss = tf.concat([coo_loss, con_loss, pro_loss], axis = 4) 151 | loss = tf.reduce_mean(tf.reduce_sum(loss, axis = [1, 2, 3, 4]), name = 'loss') 152 | 153 | return loss 154 | 155 | 156 | def calc_iou(self, boxes1, boxes2): 157 | boxx = tf.square(boxes1[:, :, :, :, 2:4]) 158 | boxes1_square = boxx[:, :, :, :, 0] * boxx[:, :, :, :, 1] 159 | box = tf.stack([boxes1[:, :, :, :, 0] - boxx[:, :, :, :, 0] * 0.5, 160 | boxes1[:, :, :, :, 1] - boxx[:, :, :, :, 1] * 0.5, 161 | boxes1[:, :, :, :, 0] + boxx[:, :, :, :, 0] * 0.5, 162 | boxes1[:, :, :, :, 1] + boxx[:, :, :, :, 1] * 0.5]) 163 | boxes1 = tf.transpose(box, (1, 2, 3, 4, 0)) 164 | 165 | boxx = tf.square(boxes2[:, :, :, :, 2:4]) 166 | boxes2_square = boxx[:, :, :, :, 0] * boxx[:, :, :, :, 1] 167 | box = tf.stack([boxes2[:, :, :, :, 0] - boxx[:, :, :, :, 0] * 0.5, 168 | boxes2[:, :, :, :, 1] - boxx[:, :, :, :, 1] * 0.5, 169 | boxes2[:, :, :, :, 0] + boxx[:, :, :, :, 0] * 0.5, 170 | boxes2[:, :, :, :, 1] + boxx[:, :, :, :, 1] * 0.5]) 171 | boxes2 = tf.transpose(box, (1, 2, 3, 4, 0)) 172 | 173 | left_up = tf.maximum(boxes1[:, :, :, :, :2], boxes2[:, :, :, :, :2]) 174 | right_down = tf.minimum(boxes1[:, :, :, :, 2:], boxes2[:, :, :, :, 2:]) 175 | 176 | intersection = tf.maximum(right_down - left_up, 0.0) 177 | inter_square = intersection[:, :, :, :, 0] * intersection[:, :, :, :, 1] 178 | union_square = boxes1_square + boxes2_square - inter_square 179 | 180 | return tf.clip_by_value(1.0 * inter_square / union_square, 0.0, 1.0) 181 | --------------------------------------------------------------------------------