├── .gitignore ├── LICENSE ├── README.md ├── checkpoints └── .gitkeep ├── data ├── __init__.py ├── cifar10.py ├── data_loader.py ├── imagenet.py ├── nus_wide.py └── transform.py ├── dhn.py ├── logs └── .gitkeep ├── models ├── __init__.py ├── alexnet.py ├── model_loader.py └── vgg16.py ├── requirements.txt ├── run.py └── utils ├── __init__.py └── evaluate.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Checkpoint 2 | *.pt 3 | 4 | # Script 5 | *.sh 6 | 7 | # Log 8 | *.log 9 | 10 | # Byte-compiled / optimized / DLL files 11 | __pycache__/ 12 | *.py[cod] 13 | *$py.class 14 | 15 | # C extensions 16 | *.so 17 | 18 | # Distribution / packaging 19 | .Python 20 | build/ 21 | develop-eggs/ 22 | dist/ 23 | downloads/ 24 | eggs/ 25 | .eggs/ 26 | lib/ 27 | lib64/ 28 | parts/ 29 | sdist/ 30 | var/ 31 | wheels/ 32 | pip-wheel-metadata/ 33 | share/python-wheels/ 34 | *.egg-info/ 35 | .installed.cfg 36 | *.egg 37 | MANIFEST 38 | 39 | # PyInstaller 40 | # Usually these files are written by a python script from a template 41 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 42 | *.manifest 43 | *.spec 44 | 45 | # Installer logs 46 | pip-log.txt 47 | pip-delete-this-directory.txt 48 | 49 | # Unit test / coverage reports 50 | htmlcov/ 51 | .tox/ 52 | .nox/ 53 | .coverage 54 | .coverage.* 55 | .cache 56 | nosetests.xml 57 | coverage.xml 58 | *.cover 59 | *.py,cover 60 | .hypothesis/ 61 | .pytest_cache/ 62 | 63 | # Translations 64 | *.mo 65 | *.pot 66 | 67 | # Django stuff: 68 | *.log 69 | local_settings.py 70 | db.sqlite3 71 | db.sqlite3-journal 72 | 73 | # Flask stuff: 74 | instance/ 75 | .webassets-cache 76 | 77 | # Scrapy stuff: 78 | .scrapy 79 | 80 | # Sphinx documentation 81 | docs/_build/ 82 | 83 | # PyBuilder 84 | target/ 85 | 86 | # Jupyter Notebook 87 | .ipynb_checkpoints 88 | 89 | # IPython 90 | profile_default/ 91 | ipython_config.py 92 | 93 | # pyenv 94 | .python-version 95 | 96 | # pipenv 97 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 98 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 99 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 100 | # install all needed dependencies. 101 | #Pipfile.lock 102 | 103 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 104 | __pypackages__/ 105 | 106 | # Celery stuff 107 | celerybeat-schedule 108 | celerybeat.pid 109 | 110 | # SageMath parsed files 111 | *.sage.py 112 | 113 | # Environments 114 | .env 115 | .venv 116 | env/ 117 | venv/ 118 | ENV/ 119 | env.bak/ 120 | venv.bak/ 121 | 122 | # Spyder project settings 123 | .spyderproject 124 | .spyproject 125 | 126 | # Rope project settings 127 | .ropeproject 128 | 129 | # mkdocs documentation 130 | /site 131 | 132 | # mypy 133 | .mypy_cache/ 134 | .dmypy.json 135 | dmypy.json 136 | 137 | # Pyre type checker 138 | .pyre/ 139 | -------------------------------------------------------------------------------- /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 | # Deep Hashing Network for Efficient Similarity Retrieval 2 | 3 | ## REQUIREMENTS 4 | `pip install -r requirements.txt` 5 | 6 | 1. pytorch >= 1.0 7 | 2. loguru 8 | 9 | ## DATASETS 10 | 1. [CIFAR-10](https://pan.baidu.com/s/1YJVe-tTfWTSKHMSYnxfjVg) Password: aemd 11 | 2. [NUS-WIDE](https://pan.baidu.com/s/1qVKFQz4_PbQX0CrSWwUwYw) Password: msfv 12 | 3. [Imagenet100](https://pan.baidu.com/s/17koNbdMLIYHgPFEFzjblvQ) Password: xpab 13 | 14 | ## USAGE 15 | ``` 16 | usage: run.py [-h] [--dataset DATASET] [--root ROOT] 17 | [--code-length CODE_LENGTH] [--arch ARCH] 18 | [--batch-size BATCH_SIZE] [--lr LR] [--max-iter MAX_ITER] 19 | [--num-workers NUM_WORKERS] [--topk TOPK] [--gpu GPU] 20 | [--lamda LAMDA] [--seed SEED] 21 | [--evaluate-interval EVALUATE_INTERVAL] 22 | 23 | DHN_PyTorch 24 | 25 | optional arguments: 26 | -h, --help show this help message and exit 27 | --dataset DATASET Dataset name. 28 | --root ROOT Path of dataset 29 | --code-length CODE_LENGTH 30 | Binary hash code length. 31 | --arch ARCH CNN model name.(default: alexnet) 32 | --batch-size BATCH_SIZE 33 | Batch size.(default: 256) 34 | --lr LR Learning rate.(default: 1e-5) 35 | --max-iter MAX_ITER Number of iterations.(default: 500) 36 | --num-workers NUM_WORKERS 37 | Number of loading data threads.(default: 6) 38 | --topk TOPK Calculate map of top k.(default: all) 39 | --gpu GPU Using gpu.(default: False) 40 | --lamda LAMDA Hyper-parameter.(default: 1) 41 | --seed SEED Random seed.(default: 3367) 42 | --evaluate-interval EVALUATE_INTERVAL 43 | Evaluation interval.(default: 10) 44 | ``` 45 | 46 | ## EXPERIMENTS 47 | CNN model: Alexnet. 48 | 49 | cifar10: 1000 query images, 5000 training images, MAP@ALL. 50 | 51 | nus-wide: Top 21 classes, 2100 query images, 10500 training images, MAP@5000. 52 | 53 | imagenet100: Top 100 classes, 5000 query images, 10000 training images, MAP@1000. 54 | 55 | bits | 16 | 32 | 48 | 128 56 | :-: | :-: | :-: | :-: | :-: 57 | cifar10@ALL | 0.7275 | 0.7353 | 0.7302 | 0.7386 58 | nus-wide-tc21@5000 | 0.8194 | 0.8326 | 0.8396 | 0.8443 59 | imagenet100@1000 | 0.2659 | 0.3703 | 0.4122 | 0.4743 60 | 61 | -------------------------------------------------------------------------------- /checkpoints/.gitkeep: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TreezzZ/DHN_PyTorch/79ee001865c0de5e9942cf3083bf3b5fb3036d25/checkpoints/.gitkeep -------------------------------------------------------------------------------- /data/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TreezzZ/DHN_PyTorch/79ee001865c0de5e9942cf3083bf3b5fb3036d25/data/__init__.py -------------------------------------------------------------------------------- /data/cifar10.py: -------------------------------------------------------------------------------- 1 | import os 2 | import torch 3 | import numpy as np 4 | 5 | from torchvision.datasets import ImageFolder 6 | from torch.utils.data.dataloader import DataLoader 7 | from torch.utils.data.dataset import Dataset 8 | from data.transform import train_transform, query_transform, Onehot, encode_onehot 9 | from PIL import Image 10 | 11 | def load_data(root, batch_size, num_workers): 12 | """ 13 | Load cifar-10 dataset. 14 | 15 | Args 16 | root(str): Path of dataset. 17 | batch_size(int): Batch size. 18 | num_workers(int): Number of data loading workers. 19 | 20 | Returns 21 | train_dataloader, query_dataloader, retrieval_dataloader(torch.utils.data.DataLoader): Data loader. 22 | """ 23 | root = os.path.join(root, 'images') 24 | train_dataloader = DataLoader( 25 | ImagenetDataset( 26 | os.path.join(root, 'train'), 27 | transform=train_transform(), 28 | target_transform=Onehot(10), 29 | ), 30 | batch_size=batch_size, 31 | num_workers=num_workers, 32 | shuffle=True, 33 | pin_memory=True, 34 | ) 35 | 36 | query_dataloader = DataLoader( 37 | ImagenetDataset( 38 | os.path.join(root, 'query'), 39 | transform=query_transform(), 40 | target_transform=Onehot(10), 41 | ), 42 | batch_size=batch_size, 43 | num_workers=num_workers, 44 | shuffle=False, 45 | pin_memory=True, 46 | ) 47 | 48 | retrieval_dataloader = DataLoader( 49 | ImagenetDataset( 50 | os.path.join(root, 'database'), 51 | transform=query_transform(), 52 | target_transform=Onehot(10), 53 | ), 54 | batch_size=batch_size, 55 | num_workers=num_workers, 56 | shuffle=False, 57 | pin_memory=True, 58 | ) 59 | 60 | return train_dataloader, query_dataloader, retrieval_dataloader, 61 | 62 | 63 | class ImagenetDataset(Dataset): 64 | classes = None 65 | class_to_idx = None 66 | 67 | def __init__(self, root, transform=None, target_transform=None): 68 | self.root = root 69 | self.transform = transform 70 | self.target_transform = target_transform 71 | self.data = [] 72 | self.targets = [] 73 | 74 | # Assume file alphabet order is the class order 75 | if ImagenetDataset.class_to_idx is None: 76 | ImagenetDataset.classes, ImagenetDataset.class_to_idx = self._find_classes(root) 77 | 78 | for i, cl in enumerate(ImagenetDataset.classes): 79 | cur_class = os.path.join(self.root, cl) 80 | files = os.listdir(cur_class) 81 | files = [os.path.join(cur_class, i) for i in files] 82 | self.data.extend(files) 83 | self.targets.extend([ImagenetDataset.class_to_idx[cl] for i in range(len(files))]) 84 | self.targets = np.asarray(self.targets) 85 | self.onehot_targets = torch.from_numpy(encode_onehot(self.targets, 10)).float() 86 | 87 | def get_onehot_targets(self): 88 | return self.onehot_targets 89 | 90 | def __len__(self): 91 | return len(self.data) 92 | 93 | def __getitem__(self, item): 94 | img, target = self.data[item], self.targets[item] 95 | 96 | img = Image.open(img).convert('RGB') 97 | 98 | if self.transform is not None: 99 | img = self.transform(img) 100 | if self.target_transform is not None: 101 | target = self.target_transform(target) 102 | return img, target, item 103 | 104 | def _find_classes(self, dir): 105 | """ 106 | Finds the class folders in a dataset. 107 | 108 | Args: 109 | dir (string): Root directory path. 110 | 111 | Returns: 112 | tuple: (classes, class_to_idx) where classes are relative to (dir), and class_to_idx is a dictionary. 113 | 114 | Ensures: 115 | No class is a subdirectory of another. 116 | """ 117 | classes = [d.name for d in os.scandir(dir) if d.is_dir()] 118 | classes.sort() 119 | class_to_idx = {classes[i]: i for i in range(len(classes))} 120 | return classes, class_to_idx 121 | 122 | -------------------------------------------------------------------------------- /data/data_loader.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import numpy as np 4 | 5 | from torch.utils.data.dataloader import DataLoader 6 | from torch.utils.data.dataset import Dataset 7 | from PIL import Image, ImageFile 8 | 9 | import data.cifar10 as cifar10 10 | import data.nus_wide as nuswide 11 | import data.imagenet as imagenet 12 | 13 | from data.transform import train_transform, encode_onehot 14 | 15 | ImageFile.LOAD_TRUNCATED_IMAGES = True 16 | 17 | 18 | def load_data(dataset, root, batch_size, num_workers): 19 | """ 20 | Load dataset. 21 | 22 | Args 23 | dataset(str): Dataset name. 24 | root(str): Path of dataset. 25 | num_workers(int): Number of loading data threads. 26 | 27 | Returns 28 | train_dataloader, query_dataloader, retrieval_dataloader(torch.utils.data.DataLoader): Data loader. 29 | """ 30 | if dataset == 'cifar-10': 31 | train_dataloader, query_dataloader, retrieval_dataloader = cifar10.load_data(root, 32 | batch_size, 33 | num_workers, 34 | ) 35 | elif dataset == 'nus-wide-tc21': 36 | train_dataloader, query_dataloader, retrieval_dataloader = nuswide.load_data(root, 37 | batch_size, 38 | num_workers 39 | ) 40 | elif dataset == 'imagenet-tc100': 41 | train_dataloader, query_dataloader, retrieval_dataloader = imagenet.load_data(root, 42 | batch_size, 43 | num_workers, 44 | ) 45 | else: 46 | raise ValueError("Invalid dataset name!") 47 | 48 | return train_dataloader, query_dataloader, retrieval_dataloader 49 | 50 | 51 | def sample_data(dataloader, num_samples, batch_size, root, dataset): 52 | """ 53 | Sample data from dataloder. 54 | 55 | Args 56 | dataloader (torch.utils.data.DataLoader): Dataloader. 57 | num_samples (int): Number of samples. 58 | batch_size (int): Batch size. 59 | root (str): Path of dataset. 60 | sample_index (int): Sample index. 61 | dataset(str): Dataset name. 62 | 63 | Returns 64 | sample_dataloader (torch.utils.data.DataLoader): Sample dataloader. 65 | """ 66 | data = dataloader.dataset.data 67 | targets = dataloader.dataset.targets 68 | 69 | if isinstance(data, list): 70 | data = np.asarray(data) 71 | 72 | sample_index = torch.randperm(len(data))[:num_samples] 73 | data = data[sample_index] 74 | targets = targets[sample_index] 75 | sample = wrap_data(data, targets, batch_size, root, dataset) 76 | 77 | return sample, sample_index 78 | 79 | 80 | def wrap_data(data, targets, batch_size, root, dataset): 81 | """ 82 | Wrap data into dataloader. 83 | 84 | Args 85 | data (np.ndarray): Data. 86 | targets (np.ndarray): Targets. 87 | batch_size (int): Batch size. 88 | root (str): Path of dataset. 89 | dataset(str): Dataset name. 90 | 91 | Returns 92 | dataloader (torch.utils.data.dataloader): Data loader. 93 | """ 94 | class MyDataset(Dataset): 95 | def __init__(self, data, targets, root, dataset): 96 | self.data = data 97 | self.targets = targets 98 | self.root = root 99 | self.transform = train_transform() 100 | self.dataset = dataset 101 | if dataset == 'cifar-10': 102 | self.onehot_targets = encode_onehot(self.targets, 10) 103 | elif dataset == 'imagenet-tc100': 104 | self.onehot_targets = encode_onehot(self.targets, 100) 105 | else: 106 | self.onehot_targets = self.targets 107 | 108 | def __getitem__(self, index): 109 | img = Image.open(os.path.join(self.root, self.data[index])).convert('RGB') 110 | if self.transform is not None: 111 | img = self.transform(img) 112 | 113 | return img, self.targets[index], index 114 | 115 | def __len__(self): 116 | return self.data.shape[0] 117 | 118 | def get_onehot_targets(self): 119 | """ 120 | Return one-hot encoding targets. 121 | """ 122 | return torch.from_numpy(self.onehot_targets).float() 123 | 124 | dataset = MyDataset(data, targets, root, dataset) 125 | dataloader = DataLoader( 126 | dataset, 127 | batch_size=batch_size, 128 | shuffle=True, 129 | pin_memory=True, 130 | ) 131 | 132 | return dataloader 133 | -------------------------------------------------------------------------------- /data/imagenet.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torchvision.transforms as transforms 4 | 5 | import os 6 | 7 | from torch.utils.data import DataLoader 8 | from torch.utils.data.dataset import Dataset 9 | from PIL import Image 10 | from data.transform import encode_onehot, Onehot 11 | 12 | 13 | def load_data(root, batch_size, workers): 14 | """ 15 | Load imagenet dataset 16 | 17 | Args 18 | root (str): Path of imagenet dataset. 19 | batch_size (int): Number of samples in one batch. 20 | workers (int): Number of data loading threads. 21 | 22 | Returns 23 | train_dataloader, query_dataloader, retrieval_dataloader(torch.utils.data.dataloader.DataLoader): Data loader. 24 | """ 25 | # Data transform 26 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 27 | std=[0.229, 0.224, 0.225]) 28 | train_transform = transforms.Compose([ 29 | transforms.RandomResizedCrop(224), 30 | transforms.RandomHorizontalFlip(), 31 | transforms.ToTensor(), 32 | normalize, 33 | ]) 34 | query_transform = transforms.Compose([ 35 | transforms.Resize(256), 36 | transforms.CenterCrop(224), 37 | transforms.ToTensor(), 38 | normalize, 39 | ]) 40 | 41 | # Construct data loader 42 | train_dir = os.path.join(root, 'train') 43 | query_dir = os.path.join(root, 'query') 44 | database_dir = os.path.join(root, 'database') 45 | 46 | train_dataset = ImagenetDataset( 47 | train_dir, 48 | transform=train_transform, 49 | targets_transform=Onehot(100), 50 | ) 51 | 52 | train_dataloader = DataLoader( 53 | train_dataset, 54 | batch_size=batch_size, 55 | shuffle=True, 56 | num_workers=workers, 57 | pin_memory=True, 58 | ) 59 | 60 | query_dataset = ImagenetDataset( 61 | query_dir, 62 | transform=query_transform, 63 | targets_transform=Onehot(100), 64 | ) 65 | 66 | query_dataloader = DataLoader( 67 | query_dataset, 68 | batch_size=batch_size, 69 | num_workers=workers, 70 | pin_memory=True, 71 | ) 72 | 73 | database_dataset = ImagenetDataset( 74 | database_dir, 75 | transform=query_transform, 76 | targets_transform=Onehot(100), 77 | ) 78 | 79 | database_dataloader = DataLoader( 80 | database_dataset, 81 | batch_size=batch_size, 82 | num_workers=workers, 83 | pin_memory=True, 84 | ) 85 | 86 | return train_dataloader, query_dataloader, database_dataloader 87 | 88 | 89 | class ImagenetDataset(Dataset): 90 | classes = None 91 | class_to_idx = None 92 | 93 | def __init__(self, root, transform=None, targets_transform=None): 94 | self.root = root 95 | self.transform = transform 96 | self.targets_transform = targets_transform 97 | self.imgs = [] 98 | self.targets = [] 99 | 100 | # Assume file alphabet order is the class order 101 | if ImagenetDataset.class_to_idx is None: 102 | ImagenetDataset.classes, ImagenetDataset.class_to_idx = self._find_classes(root) 103 | 104 | for i, cl in enumerate(ImagenetDataset.classes): 105 | cur_class = os.path.join(self.root, cl) 106 | files = os.listdir(cur_class) 107 | files = [os.path.join(cur_class, i) for i in files] 108 | self.imgs.extend(files) 109 | self.targets.extend([ImagenetDataset.class_to_idx[cl] for i in range(len(files))]) 110 | self.targets = np.asarray(self.targets) 111 | self.onehot_targets = torch.from_numpy(encode_onehot(self.targets, 100)).float() 112 | self.data = self.imgs 113 | 114 | def get_onehot_targets(self): 115 | return self.onehot_targets 116 | 117 | def __len__(self): 118 | return len(self.imgs) 119 | 120 | def __getitem__(self, item): 121 | img, target = self.imgs[item], self.targets[item] 122 | 123 | img = Image.open(img).convert('RGB') 124 | 125 | if self.transform is not None: 126 | img = self.transform(img) 127 | if self.targets_transform is not None: 128 | target = self.targets_transform(target) 129 | return img, target, item 130 | 131 | def _find_classes(self, dir): 132 | """ 133 | Finds the class folders in a dataset. 134 | 135 | Args: 136 | dir (string): Root directory path. 137 | 138 | Returns: 139 | tuple: (classes, class_to_idx) where classes are relative to (dir), and class_to_idx is a dictionary. 140 | 141 | Ensures: 142 | No class is a subdirectory of another. 143 | """ 144 | classes = [d.name for d in os.scandir(dir) if d.is_dir()] 145 | classes.sort() 146 | class_to_idx = {classes[i]: i for i in range(len(classes))} 147 | return classes, class_to_idx 148 | 149 | -------------------------------------------------------------------------------- /data/nus_wide.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import numpy as np 4 | 5 | from PIL import Image, ImageFile 6 | from torch.utils.data.dataset import Dataset 7 | from torch.utils.data.dataloader import DataLoader 8 | 9 | from data.transform import train_transform, query_transform 10 | 11 | ImageFile.LOAD_TRUNCATED_IMAGES = True 12 | 13 | 14 | def load_data(root, batch_size, num_workers): 15 | """ 16 | Loading nus-wide dataset. 17 | 18 | Args: 19 | root(str): Path of image files. 20 | batch_size(int): Batch size. 21 | num_workers(int): Number of loading data threads. 22 | 23 | Returns 24 | train_dataloader, query_dataloader, retrieval_dataloader(torch.utils.data.dataloader.DataLoader): Data loader. 25 | """ 26 | query_dataloader = DataLoader( 27 | NusWideDataset( 28 | root, 29 | 'test_img.txt', 30 | 'test_label_onehot.txt', 31 | transform=query_transform(), 32 | ), 33 | batch_size=batch_size, 34 | num_workers=num_workers, 35 | pin_memory=True, 36 | ) 37 | 38 | train_dataloader = DataLoader( 39 | NusWideDataset( 40 | root, 41 | 'train_img.txt', 42 | 'train_label_onehot_tc21.txt', 43 | transform=train_transform(), 44 | ), 45 | shuffle=True, 46 | batch_size=batch_size, 47 | num_workers=num_workers, 48 | pin_memory=True, 49 | ) 50 | 51 | retrieval_dataloader = DataLoader( 52 | NusWideDataset( 53 | root, 54 | 'database_img.txt', 55 | 'database_label_onehot.txt', 56 | transform=query_transform(), 57 | ), 58 | batch_size=batch_size, 59 | num_workers=num_workers, 60 | pin_memory=True, 61 | ) 62 | 63 | return train_dataloader, query_dataloader, retrieval_dataloader 64 | 65 | 66 | class NusWideDataset(Dataset): 67 | """ 68 | Nus-wide dataset, 21 classes. 69 | 70 | Args 71 | root(str): Path of image files. 72 | img_txt(str): Path of txt file containing image file name. 73 | label_txt(str): Path of txt file containing image label. 74 | transform(callable, optional): Transform images. 75 | """ 76 | def __init__(self, root, img_txt, label_txt, transform=None): 77 | self.root = root 78 | self.transform = transform 79 | 80 | img_txt_path = os.path.join(root, img_txt) 81 | label_txt_path = os.path.join(root, label_txt) 82 | 83 | # Read files 84 | with open(img_txt_path, 'r') as f: 85 | self.data = np.array([i.strip() for i in f]) 86 | self.targets = np.loadtxt(label_txt_path, dtype=np.float32) 87 | 88 | def __getitem__(self, index): 89 | img = Image.open(os.path.join(self.root, self.data[index])).convert('RGB') 90 | if self.transform is not None: 91 | img = self.transform(img) 92 | 93 | return img, self.targets[index], index 94 | 95 | def __len__(self): 96 | return len(self.data) 97 | 98 | def get_onehot_targets(self): 99 | return torch.from_numpy(self.targets).float() 100 | 101 | -------------------------------------------------------------------------------- /data/transform.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torchvision.transforms as transforms 3 | import numpy as np 4 | 5 | 6 | def encode_onehot(labels, num_classes=10): 7 | """ 8 | one-hot labels 9 | 10 | Args: 11 | labels (numpy.ndarray): labels. 12 | num_classes (int): Number of classes. 13 | 14 | Returns: 15 | onehot_labels (numpy.ndarray): one-hot labels. 16 | """ 17 | onehot_labels = np.zeros((len(labels), num_classes)) 18 | 19 | for i in range(len(labels)): 20 | onehot_labels[i, labels[i]] = 1 21 | 22 | return onehot_labels 23 | 24 | 25 | class Onehot(object): 26 | def __init__(self, num_classes=10): 27 | self.num_classes = num_classes 28 | 29 | def __call__(self, sample): 30 | target_onehot = torch.zeros(self.num_classes) 31 | target_onehot[sample] = 1 32 | 33 | return target_onehot 34 | 35 | 36 | def train_transform(): 37 | """ 38 | Training images transform. 39 | 40 | Args 41 | None 42 | 43 | Returns 44 | transform(torchvision.transforms): transform 45 | """ 46 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 47 | std=[0.229, 0.224, 0.225]) 48 | return transforms.Compose([ 49 | transforms.RandomResizedCrop(224), 50 | transforms.RandomHorizontalFlip(), 51 | transforms.ToTensor(), 52 | normalize, 53 | ]) 54 | 55 | 56 | def query_transform(): 57 | """ 58 | Query images transform. 59 | 60 | Args 61 | None 62 | 63 | Returns 64 | transform(torchvision.transforms): transform 65 | """ 66 | # Data transform 67 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 68 | std=[0.229, 0.224, 0.225]) 69 | return transforms.Compose([ 70 | transforms.Resize(256), 71 | transforms.CenterCrop(224), 72 | transforms.ToTensor(), 73 | normalize, 74 | ]) 75 | -------------------------------------------------------------------------------- /dhn.py: -------------------------------------------------------------------------------- 1 | import time 2 | import torch 3 | import torch.nn as nn 4 | import torch.optim as optim 5 | 6 | from models.model_loader import load_model 7 | from torch.optim.lr_scheduler import CosineAnnealingLR 8 | from utils.evaluate import mean_average_precision, pr_curve 9 | from loguru import logger 10 | 11 | 12 | def train( 13 | train_dataloader, 14 | query_dataloader, 15 | retrieval_dataloader, 16 | arch, 17 | code_length, 18 | device, 19 | lr, 20 | max_iter, 21 | lamda, 22 | topk, 23 | evaluate_interval, 24 | ): 25 | """ 26 | Training model. 27 | 28 | Args 29 | train_dataloader, query_dataloader, retrieval_dataloader(torch.utils.data.dataloader.DataLoader): Data loader. 30 | arch(str): CNN model name. 31 | code_length(int): Hash code length. 32 | device(torch.device): GPU or CPU. 33 | lr(float): Learning rate. 34 | max_iter(int): Number of iterations. 35 | lamda(float): Hyper-parameters. 36 | topk(int): Compute top k map. 37 | evaluate_interval(int): Interval of evaluation. 38 | 39 | Returns 40 | checkpoint(dict): Checkpoint. 41 | """ 42 | # Load model 43 | model = load_model(arch, code_length).to(device) 44 | 45 | # Create criterion, optimizer, scheduler 46 | criterion = DHNLoss(lamda) 47 | optimizer = optim.RMSprop( 48 | model.parameters(), 49 | lr=lr, 50 | weight_decay=5e-4, 51 | ) 52 | scheduler = CosineAnnealingLR( 53 | optimizer, 54 | max_iter, 55 | lr/100, 56 | ) 57 | 58 | # Initialization 59 | running_loss = 0. 60 | best_map = 0. 61 | training_time = 0. 62 | 63 | # Training 64 | for it in range(max_iter): 65 | tic = time.time() 66 | for data, targets, index in train_dataloader: 67 | data, targets, index = data.to(device), targets.to(device), index.to(device) 68 | optimizer.zero_grad() 69 | 70 | # Create similarity matrix 71 | S = (targets @ targets.t() > 0).float() 72 | outputs = model(data) 73 | loss = criterion(outputs, S) 74 | 75 | running_loss += loss.item() 76 | loss.backward() 77 | optimizer.step() 78 | scheduler.step() 79 | training_time += time.time() - tic 80 | 81 | # Evaluate 82 | if it % evaluate_interval == evaluate_interval - 1: 83 | # Generate hash code 84 | query_code = generate_code(model, query_dataloader, code_length, device) 85 | retrieval_code = generate_code(model, retrieval_dataloader, code_length, device) 86 | 87 | query_targets = query_dataloader.dataset.get_onehot_targets() 88 | retrieval_targets = retrieval_dataloader.dataset.get_onehot_targets() 89 | 90 | # Compute map 91 | mAP = mean_average_precision( 92 | query_code.to(device), 93 | retrieval_code.to(device), 94 | query_targets.to(device), 95 | retrieval_targets.to(device), 96 | device, 97 | topk, 98 | ) 99 | 100 | # Compute PR curve 101 | P, R = pr_curve( 102 | query_code.to(device), 103 | retrieval_code.to(device), 104 | query_targets.to(device), 105 | retrieval_targets.to(device), 106 | device, 107 | ) 108 | 109 | # Log 110 | logger.info('[iter:{}/{}][loss:{:.2f}][map:{:.4f}][time:{:.2f}]'.format( 111 | it+1, 112 | max_iter, 113 | running_loss / evaluate_interval, 114 | mAP, 115 | training_time, 116 | )) 117 | running_loss = 0. 118 | 119 | # Checkpoint 120 | if best_map < mAP: 121 | best_map = mAP 122 | 123 | checkpoint = { 124 | 'model': model.state_dict(), 125 | 'qB': query_code.cpu(), 126 | 'rB': retrieval_code.cpu(), 127 | 'qL': query_targets.cpu(), 128 | 'rL': retrieval_targets.cpu(), 129 | 'P': P, 130 | 'R': R, 131 | 'map': best_map, 132 | } 133 | 134 | return checkpoint 135 | 136 | 137 | def generate_code(model, dataloader, code_length, device): 138 | """ 139 | Generate hash code 140 | 141 | Args 142 | dataloader(torch.utils.data.dataloader.DataLoader): Data loader. 143 | code_length(int): Hash code length. 144 | device(torch.device): Using gpu or cpu. 145 | 146 | Returns 147 | code(torch.Tensor): Hash code. 148 | """ 149 | model.eval() 150 | with torch.no_grad(): 151 | N = len(dataloader.dataset) 152 | code = torch.zeros([N, code_length]) 153 | for data, _, index in dataloader: 154 | data = data.to(device) 155 | hash_code = model(data) 156 | code[index, :] = hash_code.sign().cpu() 157 | 158 | model.train() 159 | return code 160 | 161 | 162 | class DHNLoss(nn.Module): 163 | """ 164 | DHN loss function. 165 | """ 166 | def __init__(self, lamda): 167 | super(DHNLoss, self).__init__() 168 | self.lamda = lamda 169 | 170 | def forward(self, H, S): 171 | # Inner product 172 | theta = H @ H.t() / 2 173 | 174 | # log(1+e^z) may be overflow when z is large. 175 | # We convert log(1+e^z) to log(1 + e^(-z)) + z. 176 | metric_loss = (torch.log(1 + torch.exp(-(self.lamda * theta).abs())) + theta.clamp(min=0) - self.lamda * S * theta).mean() 177 | quantization_loss = self.logcosh(H.abs() - 1).mean() 178 | 179 | loss = metric_loss + self.lamda * quantization_loss 180 | 181 | return loss 182 | 183 | def logcosh(self, x): 184 | return torch.log(torch.cosh(x)) 185 | 186 | -------------------------------------------------------------------------------- /logs/.gitkeep: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TreezzZ/DHN_PyTorch/79ee001865c0de5e9942cf3083bf3b5fb3036d25/logs/.gitkeep -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TreezzZ/DHN_PyTorch/79ee001865c0de5e9942cf3083bf3b5fb3036d25/models/__init__.py -------------------------------------------------------------------------------- /models/alexnet.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import math 4 | 5 | from torch.hub import load_state_dict_from_url 6 | 7 | 8 | def load_model(code_length): 9 | """ 10 | Load CNN model. 11 | 12 | Args 13 | code_length(int): Hashing code length. 14 | 15 | Returns 16 | model(torch.nn.Module): CNN model. 17 | """ 18 | model = AlexNet(code_length) 19 | state_dict = load_state_dict_from_url('https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth') 20 | model.load_state_dict(state_dict, strict=False) 21 | 22 | return model 23 | 24 | 25 | class AlexNet(nn.Module): 26 | 27 | def __init__(self, code_length): 28 | super(AlexNet, self).__init__() 29 | 30 | self.features = nn.Sequential( 31 | nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2), 32 | nn.ReLU(inplace=True), 33 | nn.MaxPool2d(kernel_size=3, stride=2), 34 | nn.Conv2d(64, 192, kernel_size=5, padding=2), 35 | nn.ReLU(inplace=True), 36 | nn.MaxPool2d(kernel_size=3, stride=2), 37 | nn.Conv2d(192, 384, kernel_size=3, padding=1), 38 | nn.ReLU(inplace=True), 39 | nn.Conv2d(384, 256, kernel_size=3, padding=1), 40 | nn.ReLU(inplace=True), 41 | nn.Conv2d(256, 256, kernel_size=3, padding=1), 42 | nn.ReLU(inplace=True), 43 | nn.MaxPool2d(kernel_size=3, stride=2), 44 | ) 45 | self.avgpool = nn.AdaptiveAvgPool2d((6, 6)) 46 | self.classifier = nn.Sequential( 47 | nn.Dropout(), 48 | nn.Linear(256 * 6 * 6, 4096), 49 | nn.ReLU(inplace=True), 50 | nn.Dropout(), 51 | nn.Linear(4096, 4096), 52 | nn.ReLU(inplace=True), 53 | nn.Linear(4096 ,1000), 54 | ) 55 | self.classifier = self.classifier[:-1] 56 | 57 | self.hash_layer = nn.Linear(4096, code_length) 58 | 59 | def forward(self, x): 60 | x = self.features(x) 61 | x = self.avgpool(x) 62 | x = x.view(x.size(0), 256 * 6 * 6) 63 | x = self.classifier(x) 64 | x = self.hash_layer(x) 65 | 66 | x = torch.tanh(x) 67 | 68 | return x 69 | -------------------------------------------------------------------------------- /models/model_loader.py: -------------------------------------------------------------------------------- 1 | import models.alexnet as alexnet 2 | import models.vgg16 as vgg16 3 | 4 | def load_model(arch, code_length): 5 | """ 6 | Load cnn model. 7 | 8 | Args 9 | arch(str): CNN model name. 10 | code_length(int): Hash code length. 11 | 12 | Returns 13 | model(torch.nn.Module): CNN model. 14 | """ 15 | if arch == 'alexnet': 16 | model = alexnet.load_model(code_length) 17 | elif arch == 'vgg16': 18 | model = vgg16.load_model(code_length) 19 | else: 20 | raise ValueError('Invalid model name!') 21 | 22 | return model 23 | 24 | -------------------------------------------------------------------------------- /models/vgg16.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | from torch.hub import load_state_dict_from_url 5 | 6 | 7 | def load_model(code_length): 8 | """ 9 | Load vgg16 model. 10 | 11 | Args 12 | code_length (int): Hash code length. 13 | 14 | Returns 15 | model (torch.nn.Module): VGG16 model. 16 | """ 17 | model = VGG(make_layers(cfgs['D'], batch_norm=False), code_length) 18 | model.load_state_dict( 19 | load_state_dict_from_url('https://download.pytorch.org/models/vgg16-397923af.pth'), 20 | strict=False, 21 | ) 22 | 23 | return model 24 | 25 | 26 | class VGG(nn.Module): 27 | 28 | def __init__(self, features, code_length): 29 | super(VGG, self).__init__() 30 | 31 | self.features = features 32 | self.avgpool = nn.AdaptiveAvgPool2d((7, 7)) 33 | self.classifier = nn.Sequential( 34 | nn.Linear(512 * 7 * 7, 4096), 35 | nn.ReLU(True), 36 | nn.Dropout(), 37 | nn.Linear(4096, 4096), 38 | nn.ReLU(True), 39 | nn.Dropout(), 40 | nn.Linear(4096, 1000), 41 | ) 42 | self.classifier = self.classifier[:-1] 43 | 44 | self.hash_layer = nn.Linear(4096, code_length), 45 | 46 | def forward(self, x): 47 | x = self.features(x) 48 | x = self.avgpool(x) 49 | x = torch.flatten(x, 1) 50 | x = self.classifier(x) 51 | x = self.hash_layer(x) 52 | 53 | x = torch.tanh(x) 54 | 55 | return x 56 | 57 | 58 | def make_layers(cfg, batch_norm=False): 59 | layers = [] 60 | in_channels = 3 61 | for v in cfg: 62 | if v == 'M': 63 | layers += [nn.MaxPool2d(kernel_size=2, stride=2)] 64 | else: 65 | conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) 66 | if batch_norm: 67 | layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)] 68 | else: 69 | layers += [conv2d, nn.ReLU(inplace=True)] 70 | in_channels = v 71 | return nn.Sequential(*layers) 72 | 73 | 74 | cfgs = { 75 | 'A': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 76 | 'B': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'], 77 | 'D': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 78 | 'E': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], 79 | } 80 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | torch 2 | torchvision 3 | loguru 4 | -------------------------------------------------------------------------------- /run.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import numpy as np 4 | import random 5 | import argparse 6 | import dhn 7 | 8 | from loguru import logger 9 | from data.data_loader import load_data 10 | 11 | 12 | def run(): 13 | # Load config 14 | args = load_config() 15 | logger.add('logs/{}_model_{}_code_{}_lamda_{}.log'.format( 16 | args.dataset, 17 | args.arch, 18 | args.code_length, 19 | args.lamda, 20 | ), 21 | rotation='500 MB', 22 | level='INFO', 23 | ) 24 | logger.info(args) 25 | 26 | # Set seed 27 | torch.backends.cudnn.benchmark = True 28 | random.seed(args.seed) 29 | torch.manual_seed(args.seed) 30 | torch.cuda.manual_seed(args.seed) 31 | np.random.seed(args.seed) 32 | 33 | # Load dataset 34 | train_dataloader, query_dataloader, retrieval_dataloader = load_data( 35 | args.dataset, 36 | args.root, 37 | args.batch_size, 38 | args.num_workers, 39 | ) 40 | 41 | # Training 42 | checkpoint = dhn.train( 43 | train_dataloader, 44 | query_dataloader, 45 | retrieval_dataloader, 46 | args.arch, 47 | args.code_length, 48 | args.device, 49 | args.lr, 50 | args.max_iter, 51 | args.lamda, 52 | args.topk, 53 | args.evaluate_interval, 54 | ) 55 | logger.info('[code_length:{}][map:{:.4f}]'.format(args.code_length, checkpoint['map'])) 56 | 57 | # Save checkpoint 58 | torch.save( 59 | checkpoint, 60 | os.path.join('checkpoints', '{}_model_{}_code_{}_lamda_{}_map_{:.4f}.pt'.format( 61 | args.dataset, 62 | args.arch, 63 | args.code_length, 64 | args.lamda, 65 | checkpoint['map']), 66 | ) 67 | ) 68 | 69 | 70 | def load_config(): 71 | """ 72 | Load configuration. 73 | 74 | Args 75 | None 76 | 77 | Returns 78 | args(argparse.ArgumentParser): Configuration. 79 | """ 80 | parser = argparse.ArgumentParser(description='DHN_PyTorch') 81 | parser.add_argument('--dataset', 82 | help='Dataset name.') 83 | parser.add_argument('--root', 84 | help='Path of dataset') 85 | parser.add_argument('--code-length', type=int, 86 | help='Binary hash code length.') 87 | parser.add_argument('--arch', default='alexnet', type=str, 88 | help='CNN model name.(default: alexnet)') 89 | parser.add_argument('--batch-size', default=256, type=int, 90 | help='Batch size.(default: 256)') 91 | parser.add_argument('--lr', default=1e-5, type=float, 92 | help='Learning rate.(default: 1e-5)') 93 | parser.add_argument('--max-iter', default=500, type=int, 94 | help='Number of iterations.(default: 500)') 95 | parser.add_argument('--num-workers', default=6, type=int, 96 | help='Number of loading data threads.(default: 6)') 97 | parser.add_argument('--topk', default=-1, type=int, 98 | help='Calculate map of top k.(default: all)') 99 | parser.add_argument('--gpu', default=None, type=int, 100 | help='Using gpu.(default: False)') 101 | parser.add_argument('--lamda', default=1, type=float, 102 | help='Hyper-parameter.(default: 1)') 103 | parser.add_argument('--seed', default=3367, type=int, 104 | help='Random seed.(default: 3367)') 105 | parser.add_argument('--evaluate-interval', default=10, type=int, 106 | help='Evaluation interval.(default: 10)') 107 | 108 | args = parser.parse_args() 109 | 110 | # GPU 111 | if args.gpu is None: 112 | args.device = torch.device("cpu") 113 | else: 114 | args.device = torch.device("cuda:%d" % args.gpu) 115 | 116 | return args 117 | 118 | 119 | if __name__ == '__main__': 120 | run() 121 | 122 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TreezzZ/DHN_PyTorch/79ee001865c0de5e9942cf3083bf3b5fb3036d25/utils/__init__.py -------------------------------------------------------------------------------- /utils/evaluate.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def mean_average_precision(query_code, 5 | database_code, 6 | query_labels, 7 | database_labels, 8 | device, 9 | topk=None, 10 | ): 11 | """ 12 | Calculate mean average precision(map). 13 | 14 | Args: 15 | query_code (torch.Tensor): Query data hash code. 16 | database_code (torch.Tensor): Database data hash code. 17 | query_labels (torch.Tensor): Query data targets, one-hot 18 | database_labels (torch.Tensor): Database data targets, one-host 19 | device (torch.device): Using CPU or GPU. 20 | topk (int): Calculate top k data map. 21 | 22 | Returns: 23 | meanAP (float): Mean Average Precision. 24 | """ 25 | num_query = query_labels.shape[0] 26 | mean_AP = 0.0 27 | 28 | for i in range(num_query): 29 | # Retrieve images from database 30 | retrieval = (query_labels[i, :] @ database_labels.t() > 0).float() 31 | 32 | # Calculate hamming distance 33 | hamming_dist = 0.5 * (database_code.shape[1] - query_code[i, :] @ database_code.t()) 34 | 35 | # Arrange position according to hamming distance 36 | retrieval = retrieval[torch.argsort(hamming_dist)][:topk] 37 | 38 | # Retrieval count 39 | retrieval_cnt = retrieval.sum().int().item() 40 | 41 | # Can not retrieve images 42 | if retrieval_cnt == 0: 43 | continue 44 | 45 | # Generate score for every position 46 | score = torch.linspace(1, retrieval_cnt, retrieval_cnt).to(device) 47 | 48 | # Acquire index 49 | index = (torch.nonzero(retrieval == 1).squeeze() + 1.0).float() 50 | 51 | mean_AP += (score / index).mean() 52 | 53 | mean_AP = mean_AP / num_query 54 | return mean_AP 55 | 56 | 57 | def pr_curve(query_code, retrieval_code, query_targets, retrieval_targets, device): 58 | """ 59 | P-R curve. 60 | 61 | Args 62 | query_code(torch.Tensor): Query hash code. 63 | retrieval_code(torch.Tensor): Retrieval hash code. 64 | query_targets(torch.Tensor): Query targets. 65 | retrieval_targets(torch.Tensor): Retrieval targets. 66 | device (torch.device): Using CPU or GPU. 67 | 68 | Returns 69 | P(torch.Tensor): Precision. 70 | R(torch.Tensor): Recall. 71 | """ 72 | num_query = query_code.shape[0] 73 | num_bit = query_code.shape[1] 74 | P = torch.zeros(num_query, num_bit + 1).to(device) 75 | R = torch.zeros(num_query, num_bit + 1).to(device) 76 | for i in range(num_query): 77 | gnd = (query_targets[i].unsqueeze(0).mm(retrieval_targets.t()) > 0).float().squeeze() 78 | tsum = torch.sum(gnd) 79 | if tsum == 0: 80 | continue 81 | hamm = 0.5 * (retrieval_code.shape[1] - query_code[i, :] @ retrieval_code.t()) 82 | tmp = (hamm <= torch.arange(0, num_bit + 1).reshape(-1, 1).float().to(device)).float() 83 | total = tmp.sum(dim=-1) 84 | total = total + (total == 0).float() * 0.1 85 | t = gnd * tmp 86 | count = t.sum(dim=-1) 87 | p = count / total 88 | r = count / tsum 89 | P[i] = p 90 | R[i] = r 91 | mask = (P > 0).float().sum(dim=0) 92 | mask = mask + (mask == 0).float() * 0.1 93 | P = P.sum(dim=0) / mask 94 | R = R.sum(dim=0) / mask 95 | 96 | return P, R 97 | 98 | --------------------------------------------------------------------------------