├── LICENSE ├── README.txt └── src ├── aircraft.py ├── cars.py ├── cub200.py ├── get_conv.py ├── model.py └── 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.txt: -------------------------------------------------------------------------------- 1 | Mean field approximation of Bilinear CNN for Fine-grained recognition 2 | 3 | 4 | DESCRIPTIONS 5 | After getting the deep descriptors of an image, bilinear pooling computes 6 | the sum of the outer product of those deep descriptors. Bilinear pooling 7 | captures all pairwise descriptor interactions, i.e., interactions of 8 | different part, in a translational invariant manner. 9 | 10 | This project aims at accelerating training at the first step. We extract 11 | VGG-16 relu5-3 features from ImageNet pre-trained model in advance and save 12 | them onto disk. At the first step, we train the model directly from the 13 | extracted relu5-3 features. We avoid feed forwarding convolution layers 14 | multiple times. 15 | 16 | 17 | PREREQUIREMENTS 18 | Python3.6 with Numpy supported 19 | PyTorch 20 | 21 | 22 | LAYOUT 23 | ./data/ # Datasets 24 | ./doc/ # Automatically generated documents 25 | ./src/ # Source code 26 | 27 | 28 | USAGE 29 | Step 1. Fine-tune the fc layer only. 30 | # Get relu5-3 features from VGG-16 ImageNet pre-trained model. 31 |    # It gives 75.47% accuracy on CUB. 32 |   $ CUDA_VISIBLE_DEVICES=0 ./src/get_conv.py 33 | $ CUDA_VISIBLE_DEVICES=0,1,2,3 ./src/train.py --base_lr 1e0 \ 34 | --batch_size 64 --epochs 80 --weight_decay 1e-5 \ 35 | | tee "[fc-] base_lr_1e0-weight_decay_1e-5_.log" 36 | 37 | Step 2. Fine-tune all layers. 38 |   # It gives 84.41% accuracy on CUB. 39 | $ CUDA_VISIBLE_DEVICES=0,1,2,3 ./src/train.py --base_lr 1e-2 \ 40 | --batch_size 64 --epochs 80 --weight_decay 1e-5 \ 41 | --pretrained "bcnn_fc_epoch_.pth" \ 42 | | tee "[all-] base_lr_1e-2-weight_decay_1e-5.log" 43 | 44 | 45 | AUTHOR 46 | Hao Zhang: zhangh0214@gmail.com 47 | 48 | 49 | LICENSE 50 | CC BY-SA 3.0 51 | -------------------------------------------------------------------------------- /src/aircraft.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -* 2 | """This module is served as torchvision.datasets to load Aircraft dataset. 3 | 4 | This file is modified from: 5 | https://github.com/vishwakftw/vision. 6 | """ 7 | 8 | 9 | import os 10 | import pickle 11 | 12 | import PIL.Image 13 | import torch 14 | 15 | 16 | __all__ = ['Aircraft', 'AircraftReLU'] 17 | __author__ = 'Hao Zhang' 18 | __copyright__ = '2018 LAMDA' 19 | __date__ = '2018-04-19' 20 | __email__ = 'zhangh0214@gmail.com' 21 | __license__ = 'CC BY-SA 3.0' 22 | __status__ = 'Development' 23 | __updated__ = '2018-04-19' 24 | __version__ = '11.1' 25 | 26 | 27 | class Aircraft(torch.utils.data.Dataset): 28 | """Aircraft dataset. 29 | 30 | Args: 31 | _root, str: Root directory of the dataset. 32 | _train, bool: Load train/test data. 33 | _transform, callable: A function/transform that takes in a PIL.Image 34 | and transforms it. 35 | _target_transform, callable: A function/transform that takes in the 36 | target and transforms it. 37 | _train_data, list of np.ndarray. 38 | _train_labels, list of int. 39 | _test_data, list of np.ndarray. 40 | _test_labels, list of int. 41 | """ 42 | def __init__(self, root, train=True, transform=None, target_transform=None, 43 | download=False): 44 | """Load the dataset. 45 | 46 | Args 47 | root, str: Root directory of the dataset. 48 | train, bool [True]: Load train/test data. 49 | transform, callable [None]: A function/transform that takes in a 50 | PIL.Image and transforms it. 51 | target_transform, callable [None]: A function/transform that takes 52 | in the target and transforms it. 53 | download, bool [False]: If true, downloads the dataset from the 54 | internet and puts it in root directory. If dataset is already 55 | downloaded, it is not downloaded again. 56 | """ 57 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 58 | self._train = train 59 | self._transform = transform 60 | self._target_transform = target_transform 61 | 62 | if self._checkIntegrity(): 63 | print('Files already downloaded and verified.') 64 | elif download: 65 | url = None 66 | self._download(url) 67 | self._extract() 68 | else: 69 | raise RuntimeError( 70 | 'Dataset not found. You can use download=True to download it.') 71 | 72 | # Now load the picked data. 73 | if self._train: 74 | self._train_data, self._train_labels = pickle.load(open( 75 | os.path.join(self._root, 'processed/train.pkl'), 'rb')) 76 | assert (len(self._train_data) == 6667 77 | and len(self._train_labels) == 6667) 78 | else: 79 | self._test_data, self._test_labels = pickle.load(open( 80 | os.path.join(self._root, 'processed/test.pkl'), 'rb')) 81 | assert (len(self._test_data) == 3333 82 | and len(self._test_labels) == 3333) 83 | 84 | def __getitem__(self, index): 85 | """ 86 | Args: 87 | index, int: Index. 88 | 89 | Returns: 90 | image, PIL.Image: Image of the given index. 91 | target, str: target of the given index. 92 | """ 93 | if self._train: 94 | image, target = self._train_data[index], self._train_labels[index] 95 | else: 96 | image, target = self._test_data[index], self._test_labels[index] 97 | # Doing this so that it is consistent with all other datasets. 98 | image = PIL.Image.fromarray(image) 99 | 100 | if self._transform is not None: 101 | image = self._transform(image) 102 | if self._target_transform is not None: 103 | target = self._target_transform(target) 104 | 105 | return image, target 106 | 107 | def __len__(self): 108 | """Length of the dataset. 109 | 110 | Returns: 111 | length, int: Length of the dataset. 112 | """ 113 | if self._train: 114 | return len(self._train_data) 115 | return len(self._test_data) 116 | 117 | def _checkIntegrity(self): 118 | """Check whether we have already processed the data. 119 | 120 | Returns: 121 | flag, bool: True if we have already processed the data. 122 | """ 123 | return ( 124 | os.path.isfile(os.path.join(self._root, 'processed/train.pkl')) 125 | and os.path.isfile(os.path.join(self._root, 'processed/test.pkl'))) 126 | 127 | def _download(self, url): 128 | raise NotImplementedError 129 | 130 | def _extract(self): 131 | raise NotImplementedError 132 | 133 | 134 | class AircraftReLU(torch.utils.data.Dataset): 135 | """Aircraft relu5-3 dataset. 136 | 137 | Args: 138 | _root, str: Root directory of the dataset. 139 | _train, bool: Load train/test data. 140 | _train_data, list. 141 | _train_labels, list. 142 | _test_data, list. 143 | _test_labels, list. 144 | """ 145 | def __init__(self, root, train=True): 146 | """Load the dataset. 147 | 148 | Args 149 | root, str: Root directory of the dataset. 150 | train, bool [True]: Load train/test data. 151 | """ 152 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 153 | self._train = train 154 | 155 | if self._checkIntegrity(): 156 | print('Aircraft relu5-3 features already prepared.') 157 | else: 158 | raise RuntimeError('Aircraft relu5-3 Dataset not found.' 159 | 'You need to prepare it in advance.') 160 | 161 | # Now load the picked data. 162 | if self._train: 163 | self._train_data, self._train_labels = pickle.load(open( 164 | os.path.join(self._root, 'relu5-3/train.pkl'), 'rb')) 165 | assert (len(self._train_data) == 6667 166 | and len(self._train_labels) == 6667) 167 | else: 168 | self._test_data, self._test_labels = pickle.load(open( 169 | os.path.join(self._root, 'relu5-3/test.pkl'), 'rb')) 170 | assert (len(self._test_data) == 3333 171 | and len(self._test_labels) == 3333) 172 | 173 | def __getitem__(self, index): 174 | """ 175 | Args: 176 | index, int: Index. 177 | 178 | Returns: 179 | feature, torch.Tensor: relu5-3 feature of the given index. 180 | target, int: target of the given index. 181 | """ 182 | if self._train: 183 | return self._train_data[index], self._train_labels[index] 184 | return self._test_data[index], self._test_labels[index] 185 | 186 | def __len__(self): 187 | """Length of the dataset. 188 | 189 | Returns: 190 | length, int: Length of the dataset. 191 | """ 192 | if self._train: 193 | return len(self._train_data) 194 | return len(self._test_data) 195 | 196 | def _checkIntegrity(self): 197 | """Check whether we have already processed the data. 198 | 199 | Returns: 200 | flag, bool: True if we have already processed the data. 201 | """ 202 | return ( 203 | os.path.isfile(os.path.join(self._root, 'relu5-3/train.pkl')) 204 | and os.path.isfile(os.path.join(self._root, 'relu5-3/test.pkl'))) 205 | -------------------------------------------------------------------------------- /src/cars.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -* 2 | """This module is served as torchvision.datasets to load Cars dataset. 3 | 4 | This file is modified from: 5 | https://github.com/vishwakftw/vision. 6 | """ 7 | 8 | 9 | import os 10 | import pickle 11 | 12 | import PIL.Image 13 | import torch 14 | 15 | 16 | __all__ = ['Cars', 'CarsReLU'] 17 | __author__ = 'Hao Zhang' 18 | __copyright__ = '2018 LAMDA' 19 | __date__ = '2018-04-19' 20 | __email__ = 'zhangh0214@gmail.com' 21 | __license__ = 'CC BY-SA 3.0' 22 | __status__ = 'Development' 23 | __updated__ = '2018-04-21' 24 | __version__ = '11.4' 25 | 26 | 27 | class Cars(torch.utils.data.Dataset): 28 | """Cars dataset. 29 | 30 | Args: 31 | _root, str: Root directory of the dataset. 32 | _train, bool: Load train/test data. 33 | _transform, callable: A function/transform that takes in a PIL.Image 34 | and transforms it. 35 | _target_transform, callable: A function/transform that takes in the 36 | target and transforms it. 37 | _train_data, list of np.ndarray. 38 | _train_labels, list of int. 39 | _test_data, list of np.ndarray. 40 | _test_labels, list of int. 41 | """ 42 | def __init__(self, root, train=True, transform=None, target_transform=None, 43 | download=False): 44 | """Load the dataset. 45 | 46 | Args 47 | root, str: Root directory of the dataset. 48 | train, bool [True]: Load train/test data. 49 | transform, callable [None]: A function/transform that takes in a 50 | PIL.Image and transforms it. 51 | target_transform, callable [None]: A function/transform that takes 52 | in the target and transforms it. 53 | download, bool [False]: If true, downloads the dataset from the 54 | internet and puts it in root directory. If dataset is already 55 | downloaded, it is not downloaded again. 56 | """ 57 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 58 | self._train = train 59 | self._transform = transform 60 | self._target_transform = target_transform 61 | 62 | if self._checkIntegrity(): 63 | print('Files already downloaded and verified.') 64 | elif download: 65 | url = None 66 | self._download(url) 67 | self._extract() 68 | else: 69 | raise RuntimeError( 70 | 'Dataset not found. You can use download=True to download it.') 71 | 72 | # Now load the picked data. 73 | if self._train: 74 | self._train_data, self._train_labels = pickle.load(open( 75 | os.path.join(self._root, 'processed/train.pkl'), 'rb')) 76 | assert (len(self._train_data) == 8144 77 | and len(self._train_labels) == 8144) 78 | else: 79 | self._test_data, self._test_labels = pickle.load(open( 80 | os.path.join(self._root, 'processed/test.pkl'), 'rb')) 81 | assert (len(self._test_data) == 8041 82 | and len(self._test_labels) == 8041) 83 | 84 | def __getitem__(self, index): 85 | """ 86 | Args: 87 | index, int: Index. 88 | 89 | Returns: 90 | image, PIL.Image: Image of the given index. 91 | target, str: target of the given index. 92 | """ 93 | if self._train: 94 | image, target = self._train_data[index], self._train_labels[index] 95 | else: 96 | image, target = self._test_data[index], self._test_labels[index] 97 | # Doing this so that it is consistent with all other datasets. 98 | image = PIL.Image.fromarray(image) 99 | 100 | if self._transform is not None: 101 | image = self._transform(image) 102 | if self._target_transform is not None: 103 | target = self._target_transform(target) 104 | 105 | return image, target 106 | 107 | def __len__(self): 108 | """Length of the dataset. 109 | 110 | Returns: 111 | length, int: Length of the dataset. 112 | """ 113 | if self._train: 114 | return len(self._train_data) 115 | return len(self._test_data) 116 | 117 | def _checkIntegrity(self): 118 | """Check whether we have already processed the data. 119 | 120 | Returns: 121 | flag, bool: True if we have already processed the data. 122 | """ 123 | return ( 124 | os.path.isfile(os.path.join(self._root, 'processed/train.pkl')) 125 | and os.path.isfile(os.path.join(self._root, 'processed/test.pkl'))) 126 | 127 | def _download(self, url): 128 | raise NotImplementedError 129 | 130 | def _extract(self): 131 | raise NotImplementedError 132 | 133 | 134 | class CarsReLU(torch.utils.data.Dataset): 135 | """Cars relu5-3 dataset. 136 | 137 | Args: 138 | _root, str: Root directory of the dataset. 139 | _train, bool: Load train/test data. 140 | _train_data, list. 141 | _train_labels, list. 142 | _test_data, list. 143 | _test_labels, list. 144 | """ 145 | def __init__(self, root, train=True): 146 | """Load the dataset. 147 | 148 | Args 149 | root, str: Root directory of the dataset. 150 | train, bool [True]: Load train/test data. 151 | """ 152 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 153 | self._train = train 154 | 155 | if self._checkIntegrity(): 156 | print('Cars relu5-3 features already prepared.') 157 | else: 158 | raise RuntimeError('Cars relu5-3 Dataset not found.' 159 | 'You need to prepare it in advance.') 160 | 161 | # Now load the picked data. 162 | if self._train: 163 | self._train_data, self._train_labels = pickle.load(open( 164 | os.path.join(self._root, 'relu5-3/train.pkl'), 'rb')) 165 | assert (len(self._train_data) == 8144 166 | and len(self._train_labels) == 8144) 167 | else: 168 | self._test_data, self._test_labels = pickle.load(open( 169 | os.path.join(self._root, 'relu5-3/test.pkl'), 'rb')) 170 | assert (len(self._test_data) == 8041 171 | and len(self._test_labels) == 8041) 172 | 173 | def __getitem__(self, index): 174 | """ 175 | Args: 176 | index, int: Index. 177 | 178 | Returns: 179 | feature, torch.Tensor: relu5-3 feature of the given index. 180 | target, int: target of the given index. 181 | """ 182 | if self._train: 183 | return self._train_data[index], self._train_labels[index] 184 | return self._test_data[index], self._test_labels[index] 185 | 186 | def __len__(self): 187 | """Length of the dataset. 188 | 189 | Returns: 190 | length, int: Length of the dataset. 191 | """ 192 | if self._train: 193 | return len(self._train_data) 194 | return len(self._test_data) 195 | 196 | def _checkIntegrity(self): 197 | """Check whether we have already processed the data. 198 | 199 | Returns: 200 | flag, bool: True if we have already processed the data. 201 | """ 202 | return ( 203 | os.path.isfile(os.path.join(self._root, 'relu5-3/train.pkl')) 204 | and os.path.isfile(os.path.join(self._root, 'relu5-3/test.pkl'))) 205 | -------------------------------------------------------------------------------- /src/cub200.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -* 2 | """This module is served as torchvision.datasets to load CUB200-2011. 3 | 4 | CUB200-2011 dataset has 11,788 images of 200 bird species. The project page 5 | is as follows. 6 | http://www.vision.caltech.edu/visipedia/CUB-200-2011.html 7 | - Images are contained in the directory data/cub200/raw/images/, 8 | with 200 subdirectories. 9 | - Format of images.txt: 10 | - Format of train_test_split.txt: 11 | - Format of classes.txt: 12 | - Format of iamge_class_labels.txt: 13 | 14 | This file is modified from: 15 | https://github.com/vishwakftw/vision. 16 | """ 17 | 18 | 19 | import os 20 | import pickle 21 | 22 | import numpy as np 23 | import PIL.Image 24 | import torch 25 | 26 | 27 | __all__ = ['CUB200', 'CUB200ReLU'] 28 | __author__ = 'Hao Zhang' 29 | __copyright__ = '2018 LAMDA' 30 | __date__ = '2018-01-09' 31 | __email__ = 'zhangh0214@gmail.com' 32 | __license__ = 'CC BY-SA 3.0' 33 | __status__ = 'Development' 34 | __updated__ = '2018-03-04' 35 | __version__ = '6.0' 36 | 37 | 38 | class CUB200(torch.utils.data.Dataset): 39 | """CUB200 dataset. 40 | 41 | Args: 42 | _root, str: Root directory of the dataset. 43 | _train, bool: Load train/test data. 44 | _transform, callable: A function/transform that takes in a PIL.Image 45 | and transforms it. 46 | _target_transform, callable: A function/transform that takes in the 47 | target and transforms it. 48 | _train_data, list of np.ndarray. 49 | _train_labels, list of int. 50 | _test_data, list of np.ndarray. 51 | _test_labels, list of int. 52 | """ 53 | def __init__(self, root, train=True, transform=None, target_transform=None, 54 | download=False): 55 | """Load the dataset. 56 | 57 | Args 58 | root, str: Root directory of the dataset. 59 | train, bool [True]: Load train/test data. 60 | transform, callable [None]: A function/transform that takes in a 61 | PIL.Image and transforms it. 62 | target_transform, callable [None]: A function/transform that takes 63 | in the target and transforms it. 64 | download, bool [False]: If true, downloads the dataset from the 65 | internet and puts it in root directory. If dataset is already 66 | downloaded, it is not downloaded again. 67 | """ 68 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 69 | self._train = train 70 | self._transform = transform 71 | self._target_transform = target_transform 72 | 73 | if self._checkIntegrity(): 74 | print('Files already downloaded and verified.') 75 | elif download: 76 | url = ('http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/' 77 | 'CUB_200_2011.tgz') 78 | self._download(url) 79 | self._extract() 80 | else: 81 | raise RuntimeError( 82 | 'Dataset not found. You can use download=True to download it.') 83 | 84 | # Now load the picked data. 85 | if self._train: 86 | self._train_data, self._train_labels = pickle.load(open( 87 | os.path.join(self._root, 'processed/train.pkl'), 'rb')) 88 | assert (len(self._train_data) == 5994 89 | and len(self._train_labels) == 5994) 90 | else: 91 | self._test_data, self._test_labels = pickle.load(open( 92 | os.path.join(self._root, 'processed/test.pkl'), 'rb')) 93 | assert (len(self._test_data) == 5794 94 | and len(self._test_labels) == 5794) 95 | 96 | def __getitem__(self, index): 97 | """ 98 | Args: 99 | index, int: Index. 100 | 101 | Returns: 102 | image, PIL.Image: Image of the given index. 103 | target, str: target of the given index. 104 | """ 105 | if self._train: 106 | image, target = self._train_data[index], self._train_labels[index] 107 | else: 108 | image, target = self._test_data[index], self._test_labels[index] 109 | # Doing this so that it is consistent with all other datasets. 110 | image = PIL.Image.fromarray(image) 111 | 112 | if self._transform is not None: 113 | image = self._transform(image) 114 | if self._target_transform is not None: 115 | target = self._target_transform(target) 116 | 117 | return image, target 118 | 119 | def __len__(self): 120 | """Length of the dataset. 121 | 122 | Returns: 123 | length, int: Length of the dataset. 124 | """ 125 | if self._train: 126 | return len(self._train_data) 127 | return len(self._test_data) 128 | 129 | def _checkIntegrity(self): 130 | """Check whether we have already processed the data. 131 | 132 | Returns: 133 | flag, bool: True if we have already processed the data. 134 | """ 135 | return ( 136 | os.path.isfile(os.path.join(self._root, 'processed/train.pkl')) 137 | and os.path.isfile(os.path.join(self._root, 'processed/test.pkl'))) 138 | 139 | def _download(self, url): 140 | """Download and uncompress the tar.gz file from a given URL. 141 | 142 | Args: 143 | url, str: URL to be downloaded. 144 | """ 145 | import six.moves 146 | import tarfile 147 | 148 | raw_path = os.path.join(self._root, 'raw') 149 | processed_path = os.path.join(self._root, 'processed') 150 | if not os.path.isdir(raw_path): 151 | os.mkdir(raw_path, mode=0o775) 152 | if not os.path.isdir(processed_path): 153 | os.mkdir(processed_path, mode=0x775) 154 | 155 | # Downloads file. 156 | fpath = os.path.join(self._root, 'raw/CUB_200_2011.tgz') 157 | try: 158 | print('Downloading ' + url + ' to ' + fpath) 159 | six.moves.urllib.request.urlretrieve(url, fpath) 160 | except six.moves.urllib.error.URLError: 161 | if url[:5] == 'https:': 162 | self._url = self._url.replace('https:', 'http:') 163 | print('Failed download. Trying https -> http instead.') 164 | print('Downloading ' + url + ' to ' + fpath) 165 | six.moves.urllib.request.urlretrieve(url, fpath) 166 | 167 | # Extract file. 168 | cwd = os.getcwd() 169 | tar = tarfile.open(fpath, 'r:gz') 170 | os.chdir(os.path.join(self._root, 'raw')) 171 | tar.extractall() 172 | tar.close() 173 | os.chdir(cwd) 174 | 175 | def _extract(self): 176 | """Prepare the data for train/test split and save onto disk.""" 177 | image_path = os.path.join(self._root, 'raw/CUB_200_2011/images/') 178 | # Format of images.txt: 179 | id2name = np.genfromtxt(os.path.join( 180 | self._root, 'raw/CUB_200_2011/images.txt'), dtype=str) 181 | # Format of train_test_split.txt: 182 | id2train = np.genfromtxt(os.path.join( 183 | self._root, 'raw/CUB_200_2011/train_test_split.txt'), dtype=int) 184 | 185 | train_data = [] 186 | train_labels = [] 187 | test_data = [] 188 | test_labels = [] 189 | for id_ in range(id2name.shape[0]): 190 | image = PIL.Image.open(os.path.join(image_path, id2name[id_, 1])) 191 | label = int(id2name[id_, 1][:3]) - 1 # Label starts with 0 192 | 193 | # Convert gray scale image to RGB image. 194 | if image.getbands()[0] == 'L': 195 | image = image.convert('RGB') 196 | image_np = np.array(image) 197 | image.close() 198 | 199 | if id2train[id_, 1] == 1: 200 | train_data.append(image_np) 201 | train_labels.append(label) 202 | else: 203 | test_data.append(image_np) 204 | test_labels.append(label) 205 | 206 | pickle.dump((train_data, train_labels), 207 | open(os.path.join(self._root, 'processed/train.pkl'), 'wb')) 208 | pickle.dump((test_data, test_labels), 209 | open(os.path.join(self._root, 'processed/test.pkl'), 'wb')) 210 | 211 | 212 | class CUB200ReLU(torch.utils.data.Dataset): 213 | """CUB200 relu5-3 dataset. 214 | 215 | Args: 216 | _root, str: Root directory of the dataset. 217 | _train, bool: Load train/test data. 218 | _train_data, list. 219 | _train_labels, list. 220 | _test_data, list. 221 | _test_labels, list. 222 | """ 223 | def __init__(self, root, train=True): 224 | """Load the dataset. 225 | 226 | Args 227 | root, str: Root directory of the dataset. 228 | train, bool [True]: Load train/test data. 229 | """ 230 | self._root = os.path.expanduser(root) # Replace ~ by the complete dir 231 | self._train = train 232 | 233 | if self._checkIntegrity(): 234 | print('CUB200 relu5-3 features already prepared.') 235 | else: 236 | raise RuntimeError('CUB200 relu5-3 Dataset not found.' 237 | 'You need to prepare it in advance.') 238 | 239 | # Now load the picked data. 240 | if self._train: 241 | self._train_data, self._train_labels = torch.load( 242 | os.path.join(self._root, 'relu5-3', 'train.pth')) 243 | assert (len(self._train_data) == 5994 244 | and len(self._train_labels) == 5994) 245 | else: 246 | self._test_data, self._test_labels = torch.load( 247 | os.path.join(self._root, 'relu5-3', 'test.pth')) 248 | assert (len(self._test_data) == 5794 249 | and len(self._test_labels) == 5794) 250 | 251 | def __getitem__(self, index): 252 | """ 253 | Args: 254 | index, int: Index. 255 | 256 | Returns: 257 | feature, torch.Tensor: relu5-3 feature of the given index. 258 | target, int: target of the given index. 259 | """ 260 | if self._train: 261 | return self._train_data[index], self._train_labels[index] 262 | return self._test_data[index], self._test_labels[index] 263 | 264 | def __len__(self): 265 | """Length of the dataset. 266 | 267 | Returns: 268 | length, int: Length of the dataset. 269 | """ 270 | if self._train: 271 | return len(self._train_data) 272 | return len(self._test_data) 273 | 274 | def _checkIntegrity(self): 275 | """Check whether we have already processed the data. 276 | 277 | Returns: 278 | flag, bool: True if we have already processed the data. 279 | """ 280 | return ( 281 | os.path.isfile(os.path.join(self._root, 'relu5-3', 'train.pth')) 282 | and os.path.isfile(os.path.join(self._root, 'relu5-3', 'test.pth'))) 283 | -------------------------------------------------------------------------------- /src/get_conv.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | """Get relu5-3 features for CUB/Aircraft/Cars dataset. 4 | 5 | Used for the fc process to speed up training. 6 | """ 7 | 8 | 9 | import os 10 | 11 | import torch 12 | import torchvision 13 | 14 | import cub200 15 | 16 | torch.set_default_dtype(torch.float32) 17 | torch.set_default_tensor_type(torch.FloatTensor) 18 | torch.manual_seed(0) 19 | torch.cuda.manual_seed_all(0) 20 | torch.backends.cudnn.benchmark = True 21 | 22 | 23 | __all__ = ['VGGManager'] 24 | __author__ = 'Hao Zhang' 25 | __copyright__ = '2018 LAMDA' 26 | __date__ = '2018-03-04' 27 | __email__ = 'zhangh0214@gmail.com' 28 | __license__ = 'CC BY-SA 3.0' 29 | __status__ = 'Development' 30 | __updated__ = '2018-05-19' 31 | __version__ = '13.1' 32 | 33 | 34 | class VGGManager(object): 35 | """Manager class to extract VGG-16 relu5-3 features. 36 | 37 | Attributes: 38 | _paths, dict: Useful paths. 39 | _net, torch.nn.Module: VGG-16 truncated at relu5-3. 40 | _train_loader, torch.utils.data.DataLoader: Training data. 41 | _test_loader, torch.utils.data.DataLoader: Testing data. 42 | """ 43 | def __init__(self, paths): 44 | """Prepare the network and data. 45 | 46 | Args: 47 | paths, dict: Useful paths. 48 | """ 49 | print('Prepare the network and data.') 50 | 51 | # Configurations. 52 | self._paths = paths 53 | 54 | # Network. 55 | self._net = torchvision.models.vgg16(pretrained=True).features 56 | self._net = torch.nn.Sequential(*list(self._net.children())[:-2]) 57 | self._net = self._net.cuda() 58 | 59 | # Data. 60 | # NOTE: Resize such that the short edge is 448, and then ceter crop 448. 61 | train_transforms = torchvision.transforms.Compose([ 62 | torchvision.transforms.Resize(size=(448, 448)), 63 | # torchvision.transforms.CenterCrop(size=448), 64 | torchvision.transforms.ToTensor(), 65 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 66 | std=(0.229, 0.224, 0.225)) 67 | ]) 68 | test_transforms = torchvision.transforms.Compose([ 69 | torchvision.transforms.Resize(size=(448, 448)), 70 | # torchvision.transforms.CenterCrop(size=448), 71 | torchvision.transforms.ToTensor(), 72 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 73 | std=(0.229, 0.224, 0.225)) 74 | ]) 75 | train_data = cub200.CUB200( 76 | root=self._paths['cub200'], train=True, transform=train_transforms, 77 | download=True) 78 | test_data = cub200.CUB200( 79 | root=self._paths['cub200'], train=False, transform=test_transforms, 80 | download=True) 81 | self._train_loader = torch.utils.data.DataLoader( 82 | train_data, batch_size=1, shuffle=False, num_workers=0, 83 | pin_memory=False) 84 | self._test_loader = torch.utils.data.DataLoader( 85 | test_data, batch_size=1, shuffle=False, num_workers=0, 86 | pin_memory=False) 87 | 88 | def getFeature(self, phase, size): 89 | """Get relu5-3 features and save it onto disk. 90 | 91 | Args: 92 | phase, str: Train or test. 93 | size, int: Dataset size. 94 | """ 95 | print('Get relu5-3 feaures for %s data.' % phase) 96 | if phase not in ['train', 'test']: 97 | raise RuntimeError('phase should be train/test.') 98 | with torch.no_grad(): 99 | all_data = [] # list 100 | all_label = [] # list 101 | data_loader = (self._train_loader if phase == 'train' 102 | else self._test_loader) 103 | for instance, label in data_loader: 104 | # Data. 105 | instance = instance.cuda() 106 | assert instance.size() == (1, 3, 448, 448) 107 | assert label.size() == (1,) 108 | 109 | # Forward pass 110 | feature = self._net(instance) 111 | assert feature.size() == (1, 512, 28, 28) 112 | 113 | all_data.append(torch.squeeze(feature, dim=0).cpu()) 114 | all_label.append(label.item()) 115 | assert len(all_data) == size and len(all_label) == size 116 | torch.save((all_data, all_label), os.path.join( 117 | self._paths['cub200'], 'relu5-3', '%s.pth' % phase)) 118 | 119 | def main(): 120 | """The main function.""" 121 | project_root = os.popen('pwd').read().strip() 122 | paths = { 123 | 'cub200': os.path.join(project_root, 'data', 'cub200'), 124 | 'aircraft': os.path.join(project_root, 'data', 'aircraft'), 125 | 'cars': os.path.join(project_root, 'data', 'cars'), 126 | } 127 | for d in paths: 128 | assert os.path.isdir(paths[d]) 129 | 130 | manager = VGGManager(paths) 131 | manager.getFeature('train', 5994) 132 | manager.getFeature('test', 5794) 133 | 134 | 135 | if __name__ == '__main__': 136 | main() 137 | -------------------------------------------------------------------------------- /src/model.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """Mean field B-CNN model.""" 3 | 4 | 5 | import torch 6 | import torchvision 7 | 8 | torch.set_default_dtype(torch.float32) 9 | torch.set_default_tensor_type(torch.FloatTensor) 10 | torch.manual_seed(0) 11 | torch.cuda.manual_seed_all(0) 12 | torch.backends.cudnn.benckmark = True 13 | 14 | 15 | __all__ = ['BCNN'] 16 | __author__ = 'Hao Zhang' 17 | __copyright__ = '2018 LAMDA' 18 | __date__ = '2018-01-09' 19 | __email__ = 'zhangh0214@gmail.com' 20 | __license__ = 'CC BY-SA 3.0' 21 | __status__ = 'Development' 22 | __updated__ = '2018-05-21' 23 | __version__ = '13.7' 24 | 25 | 26 | class BCNN(torch.nn.Module): 27 | """Mean field B-CNN model. 28 | 29 | The B-CNN model is illustrated as follows. 30 | conv1^2 (64) -> pool1 -> conv2^2 (128) -> pool2 -> conv3^3 (256) -> pool3 31 | -> conv4^3 (512) -> pool4 -> conv5^3 (512) -> mean field bilinear pooling 32 | -> fc. 33 | 34 | The network accepts a 3*448*448 input, and the relu5-3 activation has shape 35 | 512*28*28 since we down-sample 4 times. 36 | 37 | Attributes: 38 | _is_all, bool: In the all/fc phase. 39 | features, torch.nn.Module: Convolution and pooling layers. 40 | bn, torch.nn.Module. 41 | gap_pool, torch.nn.Module. 42 | mf_relu, torch.nn.Module. 43 | mf_pool, torch.nn.Module. 44 | fc, torch.nn.Module. 45 | """ 46 | def __init__(self, num_classes, is_all): 47 | """Declare all needed layers. 48 | 49 | Args: 50 | num_classes, int. 51 | is_all, bool: In the all/fc phase. 52 | """ 53 | torch.nn.Module.__init__(self) 54 | self._is_all = is_all 55 | 56 | if self._is_all: 57 | # Convolution and pooling layers of VGG-16. 58 | self.features = torchvision.models.vgg16(pretrained=True).features 59 | self.features = torch.nn.Sequential(*list(self.features.children()) 60 | [:-2]) # Remove pool5. 61 | 62 | # Mean filed pooling layer. 63 | self.relu5_3 = torch.nn.ReLU(inplace=False) 64 | 65 | # Classification layer. 66 | self.fc = torch.nn.Linear( 67 | in_features=512 * 512, out_features=num_classes, bias=True) 68 | 69 | if not self._is_all: 70 | self.apply(BCNN._initParameter) 71 | 72 | def _initParameter(module): 73 | """Initialize the weight and bias for each module. 74 | 75 | Args: 76 | module, torch.nn.Module. 77 | """ 78 | if isinstance(module, torch.nn.BatchNorm2d): 79 | torch.nn.init.constant_(module.weight, val=1.0) 80 | torch.nn.init.constant_(module.bias, val=0.0) 81 | elif isinstance(module, torch.nn.Conv2d): 82 | torch.nn.init.kaiming_normal_(module.weight, a=0, mode='fan_out', 83 | nonlinearity='relu') 84 | if module.bias is not None: 85 | torch.nn.init.constant_(module.bias, val=0.0) 86 | elif isinstance(module, torch.nn.Linear): 87 | if module.bias is not None: 88 | torch.nn.init.constant_(module.bias, val=0.0) 89 | 90 | def forward(self, X): 91 | """Forward pass of the network. 92 | 93 | Args: 94 | X, torch.Tensor (N*3*448*448). 95 | 96 | Returns: 97 | score, torch.Tensor (N*200). 98 | """ 99 | # Input. 100 | N = X.size()[0] 101 | if self._is_all: 102 | assert X.size() == (N, 3, 448, 448) 103 | X = self.features(X) 104 | assert X.size() == (N, 512, 28, 28) 105 | 106 | # The main branch. 107 | X = self.relu5_3(X) 108 | assert X.size() == (N, 512, 28, 28) 109 | 110 | # Classical bilinear pooling. 111 | X = torch.reshape(X, (N, 512, 28 * 28)) 112 | X = torch.bmm(X, torch.transpose(X, 1, 2)) / (28 * 28) 113 | assert X.size() == (N, 512, 512) 114 | X = torch.reshape(X, (N, 512 * 512)) 115 | 116 | # Normalization. 117 | # X = torch.sign(X) * torch.sqrt(torch.abs(X) + 1e-5) 118 | X = torch.sqrt(X + 1e-5) 119 | X = torch.nn.functional.normalize(X) 120 | 121 | # Classification. 122 | X = self.fc(X) 123 | return X 124 | -------------------------------------------------------------------------------- /src/train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | """Fine-tune all layers for bilinear CNN. 4 | 5 | This is the second step. 6 | """ 7 | 8 | 9 | import os 10 | import time 11 | 12 | import torch 13 | import torchvision 14 | 15 | import cub200 16 | import model 17 | 18 | torch.set_default_dtype(torch.float32) 19 | torch.set_default_tensor_type(torch.FloatTensor) 20 | torch.manual_seed(0) 21 | torch.cuda.manual_seed_all(0) 22 | torch.backends.cudnn.benchmark = True 23 | 24 | 25 | __all__ = ['BCNNManager'] 26 | __author__ = 'Hao Zhang' 27 | __copyright__ = '2018 LAMDA' 28 | __date__ = '2018-01-11' 29 | __email__ = 'zhangh0214@gmail.com' 30 | __license__ = 'CC BY-SA 3.0' 31 | __status__ = 'Development' 32 | __updated__ = '2018-05-19' 33 | __version__ = '13.1' 34 | 35 | 36 | class BCNNManager(object): 37 | """Manager class to train bilinear CNN. 38 | 39 | Attributes: 40 | _is_all, bool: In the all/fc phase. 41 | _options, dict: Hyperparameters. 42 | _paths, dict: Useful paths. 43 | _net, torch.nn.Module: Bilinear CNN. 44 | _criterion, torch.nn.Module: Cross-entropy loss. 45 | _optimizer, torch.optim.Optimizer: SGD with momentum. 46 | _scheduler, tirch.optim.lr_scheduler: Reduce learning rate when plateau. 47 | _train_loader, torch.utils.data.DataLoader. 48 | _test_loader, torch.utils.data.DataLoader. 49 | """ 50 | def __init__(self, options, paths): 51 | """Prepare the network, criterion, optimizer, and data. 52 | 53 | Args: 54 | options, dict: Hyperparameters. 55 | paths, dict: Useful paths. 56 | """ 57 | print('Prepare the network and data.') 58 | 59 | # Configurations. 60 | self._options = options 61 | self._paths = paths 62 | 63 | # Network. 64 | if self._paths['pretrained'] is not None: 65 | self._net = torch.nn.DataParallel( 66 | model.BCNN(num_classes=200, is_all=True)).cuda() 67 | self._net.load_state_dict(torch.load(self._paths['pretrained']), 68 | strict=False) 69 | else: 70 | self._net = torch.nn.DataParallel( 71 | model.BCNN(num_classes=200, is_all=False)).cuda() 72 | print(self._net) 73 | self._criterion = torch.nn.CrossEntropyLoss().cuda() 74 | 75 | # Optimizer. 76 | self._optimizer = torch.optim.SGD( 77 | self._net.parameters(), lr=self._options['base_lr'], 78 | momentum=0.9, weight_decay=self._options['weight_decay']) 79 | self._scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( 80 | self._optimizer, mode='max', factor=0.1, patience=8, verbose=True, 81 | threshold=1e-4) 82 | 83 | # Data. 84 | if self._paths['pretrained'] is not None: 85 | train_transforms = torchvision.transforms.Compose([ 86 | torchvision.transforms.RandomResizedCrop(size=448, 87 | scale=(0.8, 1.0)), 88 | # torchvision.transforms.Resize(size=448), 89 | # torchvision.transforms.RandomCrop(size=448), 90 | torchvision.transforms.RandomHorizontalFlip(), 91 | torchvision.transforms.ToTensor(), 92 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 93 | std=(0.229, 0.224, 0.225)), 94 | 95 | ]) 96 | test_transforms = torchvision.transforms.Compose([ 97 | # torchvision.transforms.Resize(size=448), 98 | # torchvision.transforms.CenterCrop(size=448), 99 | torchvision.transforms.Resize(size=(448, 448)), 100 | torchvision.transforms.ToTensor(), 101 | torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), 102 | std=(0.229, 0.224, 0.225)), 103 | ]) 104 | train_data = cub200.CUB200( 105 | root=self._paths['cub200'], train=True, 106 | transform=train_transforms, download=True) 107 | test_data = cub200.CUB200( 108 | root=self._paths['cub200'], train=False, 109 | transform=test_transforms, download=True) 110 | else: 111 | train_data = cub200.CUB200ReLU( 112 | root=self._paths['cub200'], train=True) 113 | test_data = cub200.CUB200ReLU( 114 | root=self._paths['cub200'], train=False) 115 | self._train_loader = torch.utils.data.DataLoader( 116 | train_data, batch_size=self._options['batch_size'], shuffle=True, 117 | num_workers=4, pin_memory=False) 118 | self._test_loader = torch.utils.data.DataLoader( 119 | test_data, 120 | batch_size=(64 if self._paths['pretrained'] is not None else 4096), 121 | shuffle=False, num_workers=4, pin_memory=False) 122 | 123 | def train(self): 124 | """Train the network.""" 125 | print('Training.') 126 | self._net.train() 127 | best_acc = 0.0 128 | best_epoch = None 129 | print('Epoch\tTrain loss\tTrain acc\tTest acc\tTime') 130 | for t in range(self._options['epochs']): 131 | epoch_loss = [] 132 | num_correct = 0 133 | num_total = 0 134 | tic = time.time() 135 | for instances, labels in self._train_loader: 136 | # Data. 137 | instances = instances.cuda() 138 | labels = labels.cuda() 139 | 140 | # Forward pass. 141 | score = self._net(instances) 142 | loss = self._criterion(score, labels) 143 | 144 | with torch.no_grad(): 145 | epoch_loss.append(loss.item()) 146 | # Prediction. 147 | prediction = torch.argmax(score, dim=1) 148 | num_total += labels.size(0) 149 | num_correct += torch.sum(prediction == labels).item() 150 | 151 | # Backward pass. 152 | self._optimizer.zero_grad() 153 | loss.backward() 154 | self._optimizer.step() 155 | del instances, labels, score, loss, prediction 156 | train_acc = 100 * num_correct / num_total 157 | test_acc = self._accuracy(self._test_loader) 158 | if test_acc > best_acc: 159 | best_acc = test_acc 160 | best_epoch = t + 1 161 | print('*', end='') 162 | save_path = os.path.join( 163 | self._paths['model'], 164 | 'bcnn_%s_epoch_%d.pth' % ( 165 | 'all' if self._paths['pretrained'] is not None 166 | else 'fc', t + 1)) 167 | torch.save(self._net.state_dict(), save_path) 168 | toc = time.time() 169 | print('%d\t%4.3f\t\t%4.2f%%\t\t%4.2f%%\t\t%4.2f min' % 170 | (t + 1, sum(epoch_loss) / len(epoch_loss), train_acc, 171 | test_acc, (toc - tic) / 60)) 172 | self._scheduler.step(test_acc) 173 | print('Best at epoch %d, test accuaray %4.2f' % (best_epoch, best_acc)) 174 | 175 | def _accuracy(self, data_loader): 176 | """Compute the train/test accuracy. 177 | 178 | Args: 179 | data_loader: Train/Test DataLoader. 180 | 181 | Returns: 182 | Train/Test accuracy in percentage. 183 | """ 184 | with torch.no_grad(): 185 | self._net.eval() 186 | num_correct = 0 187 | num_total = 0 188 | for instances, labels in data_loader: 189 | # Data. 190 | instances = instances.cuda() 191 | labels = labels.cuda() 192 | 193 | # Forward pass. 194 | score = self._net(instances) 195 | 196 | # Predictions. 197 | prediction = torch.argmax(score, dim=1) 198 | num_total += labels.size(0) 199 | num_correct += torch.sum(prediction == labels).item() 200 | self._net.train() # Set the model to training phase 201 | return 100 * num_correct / num_total 202 | 203 | 204 | def main(): 205 | """The main function.""" 206 | import argparse 207 | parser = argparse.ArgumentParser( 208 | description='Train mean field bilinear CNN on CUB200.') 209 | parser.add_argument('--base_lr', dest='base_lr', type=float, required=True, 210 | help='Base learning rate for training.') 211 | parser.add_argument('--batch_size', dest='batch_size', type=int, 212 | required=True, help='Batch size.') 213 | parser.add_argument('--epochs', dest='epochs', type=int, required=True, 214 | help='Epochs for training.') 215 | parser.add_argument('--weight_decay', dest='weight_decay', type=float, 216 | required=True, help='Weight decay.') 217 | parser.add_argument('--pretrained', dest='pretrained', type=str, 218 | required=False, help='Pre-trained model.') 219 | args = parser.parse_args() 220 | if args.base_lr <= 0: 221 | raise AttributeError('--base_lr parameter must >0.') 222 | if args.batch_size <= 0: 223 | raise AttributeError('--batch_size parameter must >0.') 224 | if args.epochs < 0: 225 | raise AttributeError('--epochs parameter must >=0.') 226 | if args.weight_decay <= 0: 227 | raise AttributeError('--weight_decay parameter must >0.') 228 | 229 | project_root = os.popen('pwd').read().strip() 230 | options = { 231 | 'base_lr': args.base_lr, 232 | 'batch_size': args.batch_size, 233 | 'epochs': args.epochs, 234 | 'weight_decay': args.weight_decay, 235 | } 236 | paths = { 237 | 'cub200': os.path.join(project_root, 'data', 'cub200'), 238 | 'aircraft': os.path.join(project_root, 'data', 'aircraft'), 239 | 'model': os.path.join(project_root, 'model'), 240 | 'pretrained': (os.path.join(project_root, 'model', args.pretrained) 241 | if args.pretrained else None), 242 | } 243 | for d in paths: 244 | if d == 'pretrained': 245 | assert paths[d] is None or os.path.isfile(paths[d]) 246 | else: 247 | assert os.path.isdir(paths[d]) 248 | 249 | manager = BCNNManager(options, paths) 250 | manager.train() 251 | 252 | 253 | if __name__ == '__main__': 254 | main() 255 | --------------------------------------------------------------------------------