├── .gitignore ├── LICENSE ├── README.md ├── ckpt └── .gitignore ├── data ├── .gitignore └── ImageNet1K │ └── .gitignore ├── img ├── fine-tuning_setting.png ├── linear_probing_setting.png ├── mae.png └── pre-training_setting.png ├── lars.py ├── log └── .gitignore ├── main_eval.py ├── main_mae.py ├── model.py ├── requirements.txt ├── util.py └── vit.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | 131 | # Pycharm settings 132 | .idea -------------------------------------------------------------------------------- /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 | # Masked Auto-Encoder (MAE) 2 | 3 | Pytorch implementation of Masked Auto-Encoder: 4 | 5 | * Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick. [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377v1). arXiv 2021. 6 | 7 |
8 | 9 |
10 | 11 | 12 | ## Usage 13 | 14 | 1. Clone to the local. 15 | ``` 16 | > git clone https://github.com/liujiyuan13/MAE-code.git MAE-code 17 | ``` 18 | 2. Install required packages. 19 | ``` 20 | > cd MAE-code 21 | > pip install requirements.txt 22 | ``` 23 | 3. Prepare datasets. 24 | - For *Cifar10*, *Cifar100* and *STL*, skip this step for it will be done automatically; 25 | - For *ImageNet1K*, [download](https://www.image-net.org/download) and unzip the train(val) set into `./data/ImageNet1K/train(val)`. 26 | 4. Set parameters. 27 | - All parameters are kept in `default_args()` function of `main_mae(eval).py` file. 28 | 5. Run the code. 29 | ``` 30 | > python main_mae.py # train MAE encoder 31 | > python main_eval.py # evaluate MAE encoder 32 | ``` 33 | 6. Visualize the ouput. 34 | ``` 35 | > tensorboard --logdir=./log --port 8888 36 | ``` 37 | 38 | 39 | ## Detail 40 | 41 | ### Project structure 42 | 43 | ``` 44 | ... 45 | + ckpt # checkpoint 46 | + data # data folder 47 | + img # store images for README.md 48 | + log # log files 49 | .gitignore 50 | lars.py # LARS optimizer 51 | main_eval.py # main file for evaluation 52 | main_mae.py # main file for MAE training 53 | model.py # model definitions of MAE and EvalNet 54 | README.md 55 | util.py # helper functions 56 | vit.py # definition of vision transformer 57 | ``` 58 | 59 | ### Encoder setting 60 | 61 | In the paper, *ViT-Base*, *ViT-Large* and *ViT-Huge* are used. 62 | You can switch between them by simply changing the parameters in `default_args()`. 63 | Details can be found [here](https://openreview.net/forum?id=YicbFdNTTy) and are listed in following table. 64 | 65 | | Name | Layer Num. | Hidden Size | MLP Size | Head Num. | 66 | |:-----:|:----------:|:-----------:|:-----------:|:---------:| 67 | | Arg | vit_depth | vit_dim | vit_mlp_dim | vit_heads | 68 | | ViT-B | 12 | 768 | 3072 | 12 | 69 | | ViT-L | 24 | 1024 | 4096 | 16 | 70 | | ViT-H | 32 | 1280 | 5120 | 16 | 71 | 72 | ### Evaluation setting 73 | 74 | I implement four network training strategies concerned in the paper, including 75 | - **pre-training** is used to train MAE encoder and done in `main_mae.py`. 76 | - **linear probing** is used to evaluate MAE encoder. During training, MAE encoder is fixed. 77 | + `args.n_partial = 0` 78 | - **partial fine-tuning** is used to evaluate MAE encoder. During training, MAE encoder is partially fixed. 79 | + `args.n_partial = 0.5` --> fine-tuning MLP sub-block with the transformer fixed 80 | + `1<=args.n_partial<=args.vit_depth-1` --> fine-tuning MLP sub-block and last layers of transformer 81 | - **end-to-end fine-tuning** is used to evaluate MAE encoder. During training, MAE encoder is fully trainable. 82 | + `args.n_partial = args.vit_depth` 83 | 84 | Note that the last three strategies are done in `main_eval.py` where parameter `args.n_partial` is located. 85 | 86 | At the same time, I follow the parameter settings in the paper appendix. 87 | Note that **partial fine-tuning** and **end-to-end fine-tuning** use the same setting. 88 | Nevertheless, I replace `RandAug(9, 0.5)` with `RandomResizedCrop` and leave `mixup`, `cutmix` and `drop path` techniques in further implementation. 89 | 90 | 91 | ## Result 92 | 93 | The experiment reproduce will takes a long time and I am unfortunately busy these days. 94 | If you get some results and are willing to contribute, please reach me via email. Thanks! 95 | 96 | By the way, **I have run the code from start to end.** 97 | **It works!** 98 | So don't worry about the implementation errors. 99 | If you find any, please raise issues or email me. 100 | 101 | 102 | ## Licence 103 | 104 | This repository is under [GPL V3](https://github.com/liujiyuan13/MAE-code/blob/main/LICENSE). 105 | 106 | ## About 107 | 108 | Thanks project [*vit-pytorch*](https://github.com/lucidrains/vit-pytorch), [*pytorch-lars*](https://github.com/JosephChenHub/pytorch-lars) and [*DeepLearningExamples*](https://github.com/NVIDIA/DeepLearningExamples) for their codes contribute to this repository a lot! 109 | 110 | Homepage: 111 | 112 | Email: 113 | -------------------------------------------------------------------------------- /ckpt/.gitignore: -------------------------------------------------------------------------------- 1 | * 2 | !.gitignore -------------------------------------------------------------------------------- /data/.gitignore: -------------------------------------------------------------------------------- 1 | * 2 | !.gitignore 3 | !ImageNet1K -------------------------------------------------------------------------------- /data/ImageNet1K/.gitignore: -------------------------------------------------------------------------------- 1 | * 2 | !.gitignore -------------------------------------------------------------------------------- /img/fine-tuning_setting.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/liujiyuan13/MAE-code/7dd4344c3c1885f85262342603d2dece99dc755f/img/fine-tuning_setting.png -------------------------------------------------------------------------------- /img/linear_probing_setting.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/liujiyuan13/MAE-code/7dd4344c3c1885f85262342603d2dece99dc755f/img/linear_probing_setting.png -------------------------------------------------------------------------------- /img/mae.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/liujiyuan13/MAE-code/7dd4344c3c1885f85262342603d2dece99dc755f/img/mae.png -------------------------------------------------------------------------------- /img/pre-training_setting.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/liujiyuan13/MAE-code/7dd4344c3c1885f85262342603d2dece99dc755f/img/pre-training_setting.png -------------------------------------------------------------------------------- /lars.py: -------------------------------------------------------------------------------- 1 | """ 2 | This is from https://github.com/JosephChenHub/pytorch-lars. 3 | """ 4 | 5 | import torch 6 | from torch.optim.optimizer import Optimizer 7 | 8 | 9 | class LARS(Optimizer): 10 | r"""Implements layer-wise adaptive rate scaling for SGD. 11 | Args: 12 | params (iterable): iterable of parameters to optimize or dicts defining 13 | parameter groups 14 | lr (float): base learning rate (\gamma_0) 15 | momentum (float, optional): momentum factor (default: 0) ("m") 16 | weight_decay (float, optional): weight decay (L2 penalty) (default: 0) 17 | ("\beta") 18 | 19 | Based on Algorithm 1 of the following paper by You, Gitman, and Ginsburg. 20 | Large batch training of convolutional networks with layer-wise adaptive rate scaling. ICLR'18: 21 | https://openreview.net/pdf?id=rJ4uaX2aW 22 | 23 | The LARS algorithm can be written as 24 | .. math:: 25 | \begin{aligned} 26 | v_{t+1} & = \mu * v_{t} + (1.0 - \mu) * (g_{t} + \beta * w_{t}), \\ 27 | w_{t+1} & = w_{t} - lr * ||w_{t}|| / ||v_{t+1}|| * v_{t+1}, 28 | \end{aligned} 29 | where :math:`w`, :math:`g`, :math:`v` and :math:`\mu` denote the 30 | parameters, gradient, velocity, and momentum respectively. 31 | 32 | Example: 33 | >>> optimizer = LARS(model.parameters(), lr=0.1) 34 | >>> optimizer.zero_grad() 35 | >>> loss_fn(model(input), target).backward() 36 | >>> optimizer.step() 37 | """ 38 | def __init__(self, params, lr, momentum=.9, 39 | weight_decay=.0005, dampening = 0): 40 | if lr < 0.0: 41 | raise ValueError("Invalid learning rate: {}".format(lr)) 42 | if momentum < 0.0: 43 | raise ValueError("Invalid momentum value: {}".format(momentum)) 44 | if weight_decay < 0.0: 45 | raise ValueError("Invalid weight_decay value: {}" 46 | .format(weight_decay)) 47 | #if eta < 0.0: 48 | # raise ValueError("Invalid eta value:{}".format(eta)) 49 | 50 | defaults = dict(lr=lr, momentum = momentum, 51 | weight_decay = weight_decay, 52 | dampening = dampening) 53 | 54 | super(LARS, self).__init__(params, defaults) 55 | 56 | @torch.no_grad() 57 | def step(self, closure=None): 58 | """Performs a single optimization step. 59 | Arguments: 60 | closure (callable, optional): A closure that reevaluates the model 61 | and returns the loss. 62 | """ 63 | loss = None 64 | if closure is not None: 65 | loss = closure() 66 | 67 | 68 | for group in self.param_groups: 69 | weight_decay = group['weight_decay'] 70 | momentum = group['momentum'] 71 | lr = group['lr'] 72 | dampening = group['dampening'] 73 | 74 | for p in group['params']: 75 | if p.grad is None: 76 | continue 77 | 78 | param_state = self.state[p] 79 | # gradient 80 | d_p = p.grad.data 81 | weight_norm = torch.norm(p.data) 82 | 83 | # update the velocity 84 | if 'momentum_buffer' not in param_state: 85 | buf = param_state['momentum_buffer'] = torch.zeros_like(p.data) 86 | else: 87 | buf = param_state['momentum_buffer'] 88 | # l2 regularization 89 | if weight_decay != 0: 90 | d_p.add_(p, alpha=weight_decay) 91 | 92 | buf.mul_(momentum).add_(d_p, alpha = 1.0 - dampening) 93 | v_norm = torch.norm(buf) 94 | 95 | local_lr = lr * weight_norm / (1e-6 + v_norm) 96 | 97 | # Update the weight 98 | p.add_(buf, alpha = -local_lr) 99 | 100 | 101 | return loss 102 | 103 | 104 | 105 | -------------------------------------------------------------------------------- /log/.gitignore: -------------------------------------------------------------------------------- 1 | * 2 | !.gitignore -------------------------------------------------------------------------------- /main_eval.py: -------------------------------------------------------------------------------- 1 | ''' 2 | This is written by Jiyuan Liu, Dec. 21, 2021. 3 | Homepage: https://liujiyuan13.github.io. 4 | Email: liujiyuan13@163.com. 5 | All rights reserved. 6 | ''' 7 | 8 | import time 9 | import math 10 | import argparse 11 | import torch 12 | import tensorboard_logger 13 | 14 | from vit import ViT 15 | from lars import LARS 16 | from model import EvalNet, LabelSmoothing 17 | from util import * 18 | 19 | # for re-produce 20 | set_seed(0) 21 | 22 | 23 | def build_model(args): 24 | ''' 25 | build EvalNet model and restore weights 26 | :param args: model args 27 | :return: model 28 | ''' 29 | # build encoder 30 | v = ViT(image_size=args.image_size, 31 | patch_size=args.patch_size, 32 | num_classes=args.n_class, 33 | dim=args.vit_dim, 34 | depth=args.vit_depth, 35 | heads=args.vit_heads, 36 | mlp_dim=args.vit_mlp_dim).to(args.device) 37 | 38 | # build linear probing 39 | enet = EvalNet(encoder=v, 40 | n_class=args.n_class, 41 | masking_ratio=0, 42 | device=args.device).to(args.device) 43 | 44 | # restore weights 45 | state_dict_encoder = enet.encoder.state_dict() 46 | state_dict_loaded = torch.load(args.ckpt)['model'] 47 | for k in state_dict_encoder.keys(): 48 | state_dict_encoder[k] = state_dict_loaded['encoder.' + k] 49 | enet.encoder.load_state_dict(state_dict_encoder) 50 | 51 | return enet 52 | 53 | 54 | def train(args): 55 | ''' 56 | train the model 57 | :param args: parameters 58 | :return: 59 | ''' 60 | # load data 61 | data_loader, args.n_class = load_data(args.data_dir, 62 | args.data_name, 63 | image_size=args.image_size, 64 | batch_size=args.batch_size, 65 | n_worker=args.n_worker, 66 | is_train=True) 67 | test_loader, args.n_class = load_data(args.data_dir, 68 | args.data_name, 69 | image_size=args.image_size, 70 | batch_size=args.batch_size, 71 | n_worker=args.n_worker, 72 | is_train=False) 73 | 74 | # build model 75 | model = build_model(args) 76 | 77 | # build optimizer 78 | if args.n_partial == 0: 79 | # optimizer = torch.optim.SGD(model.parameters(), 80 | # lr=args.base_lr, 81 | # weight_decay=args.weight_decay, 82 | # momentum=args.momentum) 83 | optimizer = LARS(model.parameters(), 84 | lr=args.base_lr, 85 | weight_decay=args.weight_decay, 86 | momentum=args.momentum) 87 | else: 88 | optimizer = torch.optim.AdamW(model.parameters(), 89 | lr=args.base_lr, 90 | weight_decay=args.weight_decay, 91 | betas=args.momentum) 92 | 93 | # learning rate scheduler: warmup + consine 94 | def lr_lambda(epoch): 95 | if epoch < args.epochs_warmup: 96 | p = epoch / args.epochs_warmup 97 | lr = args.warmup_from + p * (args.warmup_to - args.warmup_from) 98 | else: 99 | eta_min = args.lr * (args.lr_decay_rate ** 3) 100 | lr = eta_min + (args.lr - eta_min) * (1 + math.cos(math.pi * epoch / args.epochs)) / 2 101 | return lr 102 | 103 | scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lr_lambda) 104 | 105 | # tensorboard 106 | tb_logger = tensorboard_logger.Logger(logdir=args.tb_folder, flush_secs=2) 107 | 108 | for epoch in range(1, args.epochs + 1): 109 | # set training mode 110 | model.encoder.eval() 111 | model.fc.train() 112 | if args.n_partial == 0.5 or (type(args.n_partial) is int and 1 <= args.n_partial <= args.vit_depth): 113 | model.encoder.mlp_head.train() 114 | for i in range(1, int(args.n_partial)+1): 115 | model.encoder.transformer.layers[args.vit_depth-i].train() 116 | elif args.n_partial == 0: 117 | pass 118 | else: 119 | raise ValueError('please check requirements of \'args.n_partial\'.') 120 | 121 | # records 122 | ts = time.time() 123 | losses = AverageMeter() 124 | 125 | # train by epoch 126 | for idx, (images, targets) in enumerate(data_loader): 127 | # put images into device 128 | images, targets = images.to(args.device), targets.to(args.device) 129 | # forward 130 | output = model(images) 131 | # compute loss 132 | if args.label_smoothing: 133 | criterion = LabelSmoothing(smoothing=args.smoothing) # use label smoothing technique 134 | else: 135 | criterion = torch.nn.CrossEntropyLoss() # common and simplest one 136 | loss = criterion(output, targets) 137 | # back propagation 138 | optimizer.zero_grad() 139 | loss.backward() 140 | optimizer.step() 141 | scheduler.step() 142 | # record 143 | losses.update(loss.item(), args.batch_size) 144 | 145 | # log 146 | tb_logger.log_value('loss_eval_partial_{}'.format(args.n_partial), losses.avg, epoch) 147 | 148 | # eval 149 | if epoch % args.eval_freq == 0: 150 | acc = test(args, model=model, data_loader=test_loader) 151 | tb_logger.log_value('acc_eval_partial_{}'.format(args.n_partial), acc, epoch) 152 | 153 | # print 154 | if epoch % args.print_freq == 0: 155 | print('- epoch {:3d}, time, {:.2f}s, loss {:.4f}'.format(epoch, time.time() - ts, losses.avg)) 156 | 157 | # save the last checkpoint 158 | save_file = os.path.join(args.ckpt_folder, 'enet_partial_{}.ckpt'.format(args.n_partial)) 159 | save_ckpt(model, optimizer, args, epoch, save_file=save_file) 160 | 161 | 162 | def test(args, model=None, ckpt_path=None, data_loader=None): 163 | ''' 164 | train the model 165 | :param args: args 166 | :param model: the test model 167 | :param ckpt_path: checkpoint path, if model is given, this is deactivated 168 | :param data_loader: data loader 169 | :return: accuracy 170 | ''' 171 | 172 | # load data 173 | if data_loader is None: 174 | data_loader, args.n_class = load_data(args.data_dir, 175 | args.data_name, 176 | image_size=args.image_size, 177 | batch_size=args.batch_size, 178 | n_worker=args.n_worker, 179 | is_train=False) 180 | 181 | # restore mae model 182 | assert model is not None or ckpt_path is not None 183 | if model is None: 184 | model = build_model(args) 185 | model = load_ckpt(model, ckpt_path) 186 | model.eval() 187 | 188 | # test 189 | accs = AverageMeter() 190 | with torch.no_grad(): 191 | for idx, (images, targets) in enumerate(data_loader): 192 | # put images into device 193 | images = images.to(args.device) 194 | # forward 195 | output = model(images) 196 | # eval 197 | _, y_pred = torch.max(output, dim=1) 198 | acc = accuracy(targets.detach().cpu().numpy(), y_pred.detach().cpu().numpy()) 199 | # record 200 | accs.update(acc, args.batch_size) 201 | 202 | return accs.avg 203 | 204 | 205 | def default_args(data_name, trail=0, ckpt_file='last.ckpt'): 206 | ''' 207 | for default parameters. tune them upon your options 208 | :param data_name: dataset name, such as 'imagenet' 209 | :param trail: an int indicator to specify different runnings 210 | :param ckpt_file: path of the trained MAE model 211 | :return: 212 | ''' 213 | # params 214 | args = argparse.ArgumentParser().parse_args() 215 | 216 | # device 217 | args.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') 218 | 219 | # data 220 | args.data_dir = 'data' 221 | args.data_name = data_name 222 | args.image_size = 256 223 | args.n_worker = 8 224 | 225 | # model 226 | args.patch_size = 32 227 | args.vit_dim = 768 228 | args.vit_depth = 12 229 | args.vit_heads = 12 230 | args.vit_mlp_dim = 3072 231 | args.masking_ratio = 0 # the paper recommended to use uncorrupted images 232 | 233 | # linear probing or partial fine-tuning or fine-tuning 234 | # - 0: linear probing, the encoder is fixed 235 | # - 0.5: fine-tuning MLP sub-block with the transformer fixed 236 | # - 1~(args.vit_depth-1): partial fine-tuning, including MLP sub-block and last layers of transformer 237 | # - args.vit_depth: fine-tuning, including MLP sub-block and all layers of transformer 238 | args.n_partial = 0 239 | 240 | # train 241 | if args.n_partial == 0: 242 | args.batch_size = 16384 243 | args.epochs = 90 244 | args.base_lr = 1e-1 245 | args.lr = args.base_lr * args.batch_size / 256 246 | args.weight_decay = 0 247 | args.momentum = 0.9 248 | args.epochs_warmup = 10 249 | else: 250 | args.batch_size = 1024 251 | args.epochs = 100 252 | args.base_lr = 1e-3 253 | args.lr = args.base_lr * args.batch_size / 256 254 | args.weight_decay = 5e-2 255 | args.momentum = (0.9, 0.999) 256 | args.epochs_warmup = 5 257 | args.warmup_from = 1e-4 258 | args.lr_decay_rate = 1e-2 259 | eta_min = args.lr * (args.lr_decay_rate ** 3) 260 | args.warmup_to = eta_min + (args.lr - eta_min) * (1 + math.cos(math.pi * args.epochs_warmup / args.epochs)) / 2 261 | 262 | # extra 263 | args.label_smoothing = True 264 | args.smoothing = 0.1 265 | 266 | # print and save 267 | args.print_freq = 5 268 | args.eval_freq = 5 269 | 270 | # tensorboard 271 | args.tb_folder = os.path.join('log', '{}_{}'.format(args.data_name, trail)) 272 | if not os.path.isdir(args.tb_folder): 273 | os.makedirs(args.tb_folder) 274 | 275 | # ckpt 276 | args.ckpt_folder = os.path.join('ckpt', '{}_{}'.format(args.data_name, trail)) 277 | args.ckpt = os.path.join(args.ckpt_folder, ckpt_file) 278 | 279 | return args 280 | 281 | 282 | if __name__ == '__main__': 283 | 284 | data_name = 'imagenet' 285 | train(default_args(data_name)) 286 | -------------------------------------------------------------------------------- /main_mae.py: -------------------------------------------------------------------------------- 1 | ''' 2 | This is written by Jiyuan Liu, Dec. 21, 2021. 3 | Homepage: https://liujiyuan13.github.io. 4 | Email: liujiyuan13@163.com. 5 | All rights reserved. 6 | ''' 7 | 8 | import time 9 | import math 10 | import argparse 11 | import torch 12 | import tensorboard_logger 13 | 14 | from vit import ViT 15 | from model import MAE 16 | from util import * 17 | 18 | # for re-produce 19 | set_seed(0) 20 | 21 | 22 | def build_model(args): 23 | ''' 24 | build MAE model. 25 | :param args: model args 26 | :return: model 27 | ''' 28 | # build model 29 | v = ViT(image_size=args.image_size, 30 | patch_size=args.patch_size, 31 | num_classes=args.n_class, 32 | dim=args.vit_dim, 33 | depth=args.vit_depth, 34 | heads=args.vit_heads, 35 | mlp_dim=args.vit_mlp_dim) 36 | 37 | mae = MAE(encoder=v, 38 | masking_ratio=args.masking_ratio, 39 | decoder_dim=args.decoder_dim, 40 | decoder_depth=args.decoder_depth, 41 | device=args.device).to(args.device) 42 | 43 | return mae 44 | 45 | 46 | def train(args): 47 | ''' 48 | train the model 49 | :param args: parameters 50 | :return: 51 | ''' 52 | # load data 53 | data_loader, args.n_class = load_data(args.data_dir, 54 | args.data_name, 55 | image_size=args.image_size, 56 | batch_size=args.batch_size, 57 | n_worker=args.n_worker, 58 | is_train=True) 59 | 60 | # build mae model 61 | model = build_model(args) 62 | model.train() 63 | 64 | # build optimizer 65 | optimizer = torch.optim.AdamW(model.parameters(), 66 | lr=args.base_lr, 67 | weight_decay=args.weight_decay, 68 | betas=args.momentum) 69 | 70 | # learning rate scheduler: warmup + consine 71 | def lr_lambda(epoch): 72 | if epoch < args.epochs_warmup: 73 | p = epoch / args.epochs_warmup 74 | lr = args.warmup_from + p * (args.warmup_to - args.warmup_from) 75 | else: 76 | eta_min = args.lr * (args.lr_decay_rate ** 3) 77 | lr = eta_min + (args.lr - eta_min) * (1 + math.cos(math.pi * epoch / args.epochs)) / 2 78 | return lr 79 | 80 | scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lr_lambda) 81 | 82 | # tensorboard 83 | tb_logger = tensorboard_logger.Logger(logdir=args.tb_folder, flush_secs=2) 84 | 85 | for epoch in range(1, args.epochs + 1): 86 | # records 87 | ts = time.time() 88 | losses = AverageMeter() 89 | 90 | # train by epoch 91 | for idx, (images, targets) in enumerate(data_loader): 92 | # put images into device 93 | images = images.to(args.device) 94 | # forward 95 | loss = model(images) 96 | # back propagation 97 | optimizer.zero_grad() 98 | loss.backward() 99 | optimizer.step() 100 | scheduler.step() 101 | # record 102 | losses.update(loss.item(), args.batch_size) 103 | 104 | # log 105 | tb_logger.log_value('loss', losses.avg, epoch) 106 | 107 | # print 108 | if epoch % args.print_freq == 0: 109 | print('- epoch {:3d}, time, {:.2f}s, loss {:.4f}'.format(epoch, time.time() - ts, losses.avg)) 110 | 111 | # save checkpoint 112 | if epoch % args.save_freq == 0: 113 | save_file = os.path.join(args.ckpt_folder, 'epoch_{:d}.ckpt'.format(epoch)) 114 | save_ckpt(model, optimizer, args, epoch, save_file=save_file) 115 | 116 | # save the last checkpoint 117 | save_file = os.path.join(args.ckpt_folder, 'last.ckpt') 118 | save_ckpt(model, optimizer, args, epoch, save_file=save_file) 119 | 120 | 121 | def default_args(data_name, trail=0): 122 | ''' 123 | for default parameters. tune them upon your options 124 | :param data_name: dataset name, such as 'imagenet' 125 | :param trail: an int indicator to specify different runnings 126 | :return: 127 | ''' 128 | # params 129 | args = argparse.ArgumentParser().parse_args() 130 | 131 | # device 132 | args.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') 133 | 134 | # data 135 | args.data_dir = 'data' 136 | args.data_name = data_name 137 | args.image_size = 256 138 | args.n_worker = 8 139 | 140 | # model 141 | # - use ViT-Base whose parameters are referred from "Dosovitskiy et al. An Image is Worth 16x16 Words: Transformers 142 | # - for Image Recognition at Scale. ICLR 2021. https://openreview.net/forum?id=YicbFdNTTy". 143 | args.patch_size = 32 144 | args.vit_dim = 768 145 | args.vit_depth = 12 146 | args.vit_heads = 12 147 | args.vit_mlp_dim = 3072 148 | args.masking_ratio = 0.75 # the paper recommended 75% masked patches 149 | args.decoder_dim = 512 # paper showed good results with 512 150 | args.decoder_depth = 8 # paper showed good results with 8 151 | 152 | # train 153 | args.batch_size = 4096 154 | args.epochs = 800 155 | args.base_lr = 1.5e-4 156 | args.lr = args.base_lr * args.batch_size / 256 157 | args.weight_decay = 5e-2 158 | args.momentum = (0.9, 0.95) 159 | args.epochs_warmup = 40 160 | args.warmup_from = 1e-4 161 | args.lr_decay_rate = 1e-2 162 | eta_min = args.lr * (args.lr_decay_rate ** 3) 163 | args.warmup_to = eta_min + (args.lr - eta_min) * (1 + math.cos(math.pi * args.epochs_warmup / args.epochs)) / 2 164 | 165 | # print and save 166 | args.print_freq = 5 167 | args.save_freq = 100 168 | 169 | # tensorboard 170 | args.tb_folder = os.path.join('log', '{}_{}'.format(args.data_name, trail)) 171 | if not os.path.isdir(args.tb_folder): 172 | os.makedirs(args.tb_folder) 173 | 174 | # ckpt 175 | args.ckpt_folder = os.path.join('ckpt', '{}_{}'.format(args.data_name, trail)) 176 | if not os.path.isdir(args.ckpt_folder): 177 | os.makedirs(args.ckpt_folder) 178 | 179 | return args 180 | 181 | 182 | if __name__ == '__main__': 183 | 184 | data_name = 'imagenet' 185 | train(default_args(data_name)) 186 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | ''' 2 | This is written by Jiyuan Liu, Dec. 21, 2021. 3 | Homepage: https://liujiyuan13.github.io. 4 | Email: liujiyuan13@163.com. 5 | All rights reserved. 6 | ''' 7 | 8 | import torch 9 | from torch import nn 10 | import torch.nn.functional as F 11 | from einops import repeat 12 | 13 | from vit import Transformer 14 | 15 | 16 | class MAE(nn.Module): 17 | ''' 18 | the implementation from https://github.com/lucidrains/vit-pytorch. 19 | ''' 20 | def __init__(self, 21 | *, 22 | encoder, 23 | decoder_dim, 24 | masking_ratio=0.75, 25 | decoder_depth=1, 26 | decoder_heads=8, 27 | decoder_dim_head=64, 28 | device='cpu'): 29 | super().__init__() 30 | # common 31 | self.device = device 32 | assert 0 <= masking_ratio < 1, 'masking ratio must be kept between 0 and 1' 33 | self.masking_ratio = masking_ratio 34 | 35 | # extract some hyperparameters and functions from encoder (vision transformer to be trained) 36 | self.encoder = encoder 37 | num_patches, encoder_dim = encoder.pos_embedding.shape[-2:] 38 | self.to_patch, self.patch_to_emb = encoder.to_patch_embedding[:2] 39 | pixel_values_per_patch = self.patch_to_emb.weight.shape[-1] 40 | 41 | # decoder parameters 42 | self.enc_to_dec = nn.Linear(encoder_dim, decoder_dim) if encoder_dim != decoder_dim else nn.Identity() 43 | self.mask_token = nn.Parameter(torch.randn(decoder_dim)) 44 | self.decoder = Transformer(dim=decoder_dim, 45 | depth=decoder_depth, 46 | heads=decoder_heads, 47 | dim_head=decoder_dim_head, 48 | mlp_dim=decoder_dim * 4) 49 | self.decoder_pos_emb = nn.Embedding(num_patches, decoder_dim) 50 | self.to_pixels = nn.Linear(decoder_dim, pixel_values_per_patch) 51 | 52 | def forward(self, img): 53 | # get patches 54 | patches = self.to_patch(img) 55 | batch, num_patches, *_ = patches.shape 56 | 57 | # patch to encoder tokens and add positions 58 | tokens = self.patch_to_emb(patches) 59 | tokens = tokens + self.encoder.pos_embedding[:, 1:(num_patches + 1)] 60 | 61 | # calculate of patches needed to be masked, and get random indices, dividing it up for mask vs unmasked 62 | num_masked = int(self.masking_ratio * num_patches) 63 | rand_indices = torch.rand(batch, num_patches, device=self.device).argsort(dim=-1) 64 | masked_indices, unmasked_indices = rand_indices[:, :num_masked], rand_indices[:, num_masked:] 65 | 66 | # get the unmasked tokens to be encoded 67 | batch_range = torch.arange(batch, device=self.device)[:, None] 68 | tokens = tokens[batch_range, unmasked_indices] 69 | 70 | # get the patches to be masked for the final reconstruction loss 71 | masked_patches = patches[batch_range, masked_indices] 72 | 73 | # attend with vision transformer 74 | encoded_tokens = self.encoder.transformer(tokens) 75 | 76 | # project encoder to decoder dimensions, if they are not equal - the paper says you can get away with a smaller dimension for decoder 77 | decoder_tokens = self.enc_to_dec(encoded_tokens) 78 | 79 | # repeat mask tokens for number of masked, and add the positions using the masked indices derived above 80 | mask_tokens = repeat(self.mask_token, 'd -> b n d', b=batch, n=num_masked) 81 | mask_tokens = mask_tokens + self.decoder_pos_emb(masked_indices) 82 | 83 | # concat the masked tokens to the decoder tokens and attend with decoder 84 | decoder_tokens = torch.cat((mask_tokens, decoder_tokens), dim=1) 85 | decoded_tokens = self.decoder(decoder_tokens) 86 | 87 | # splice out the mask tokens and project to pixel values 88 | mask_tokens = decoded_tokens[:, :num_masked] 89 | pred_pixel_values = self.to_pixels(mask_tokens) 90 | 91 | # calculate reconstruction loss 92 | recon_loss = F.mse_loss(pred_pixel_values, masked_patches) 93 | return recon_loss 94 | 95 | 96 | class EvalNet(nn.Module): 97 | ''' 98 | the encoder of masked auto-encoder + linear layer. 99 | ''' 100 | def __init__(self, encoder, n_class, masking_ratio=0, device='cpu'): 101 | super(EvalNet, self).__init__() 102 | # common 103 | self.device = device 104 | assert 0 <= masking_ratio < 1, 'masking ratio must be kept between 0 and 1' 105 | self.masking_ratio = masking_ratio 106 | 107 | # extract some hyperparameters and functions from encoder (vision transformer to be trained) 108 | self.encoder = encoder 109 | num_patches, encoder_dim = encoder.pos_embedding.shape[-2:] 110 | self.to_patch, self.patch_to_emb = encoder.to_patch_embedding[:2] 111 | 112 | # linear layer 113 | self.fc = nn.Linear((num_patches - 1) * encoder_dim, n_class) 114 | 115 | def forward(self, img): 116 | # get patches 117 | patches = self.to_patch(img) 118 | batch, num_patches, *_ = patches.shape 119 | 120 | # patch to encoder tokens and add positions 121 | tokens = self.patch_to_emb(patches) 122 | tokens = tokens + self.encoder.pos_embedding[:, 1:(num_patches + 1)] 123 | 124 | # calculate of patches needed to be masked, and get random indices, dividing it up for mask vs unmasked 125 | num_masked = int(self.masking_ratio * num_patches) 126 | rand_indices = torch.rand(batch, num_patches, device=self.device).argsort(dim=-1) 127 | masked_indices, unmasked_indices = rand_indices[:, :num_masked], rand_indices[:, num_masked:] 128 | 129 | # get the unmasked tokens to be encoded 130 | batch_range = torch.arange(batch, device=self.device)[:, None] 131 | tokens = tokens[batch_range, unmasked_indices] 132 | 133 | # attend with vision transformer 134 | encoded_tokens = self.encoder.transformer(tokens) 135 | 136 | # feed to linear probing 137 | latent_fea = encoded_tokens.flatten(start_dim=1) 138 | output = self.fc(latent_fea) 139 | 140 | return output 141 | 142 | 143 | class LabelSmoothing(nn.Module): 144 | """ 145 | NLL loss with label smoothing from https://github.com/NVIDIA/DeepLearningExamples. 146 | """ 147 | def __init__(self, smoothing=0.0): 148 | """ 149 | Constructor for the LabelSmoothing module. 150 | :param smoothing: label smoothing factor 151 | """ 152 | super(LabelSmoothing, self).__init__() 153 | self.confidence = 1.0 - smoothing 154 | self.smoothing = smoothing 155 | 156 | def forward(self, x, target): 157 | logprobs = torch.nn.functional.log_softmax(x, dim=-1) 158 | nll_loss = -logprobs.gather(dim=-1, index=target.unsqueeze(1)) 159 | nll_loss = nll_loss.squeeze(1) 160 | smooth_loss = -logprobs.mean(dim=-1) 161 | loss = self.confidence * nll_loss + self.smoothing * smooth_loss 162 | return loss.mean() -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | absl-py @ file:///tmp/build/80754af9/absl-py_1607439979954/work 2 | aiohttp @ file:///C:/ci/aiohttp_1607109732912/work 3 | albumentations==0.5.2 4 | async-timeout==3.0.1 5 | attrs @ file:///tmp/build/80754af9/attrs_1604765588209/work 6 | blinker==1.4 7 | brotlipy==0.7.0 8 | cachetools @ file:///tmp/build/80754af9/cachetools_1611600262290/work 9 | certifi==2021.5.30 10 | cffi @ file:///C:/ci/cffi_1606255207413/work 11 | chardet @ file:///C:/ci/chardet_1605303259695/work 12 | click @ file:///home/linux1/recipes/ci/click_1610990599742/work 13 | cmake==3.21.2 14 | colorama @ file:///tmp/build/80754af9/colorama_1607707115595/work 15 | cryptography==2.9.2 16 | cycler==0.10.0 17 | daal4py==2021.2.2 18 | decorator @ file:///tmp/build/80754af9/decorator_1617916966915/work 19 | docopt==0.6.2 20 | einops==0.3.2 21 | faiss==1.7.0 22 | google-auth @ file:///tmp/build/80754af9/google-auth_1607969906642/work 23 | google-auth-oauthlib @ file:///tmp/build/80754af9/google-auth-oauthlib_1603929124518/work 24 | grpcio @ file:///C:/ci/grpcio_1597406403308/work 25 | h5py==2.10.0 26 | hdf5storage==0.1.18 27 | idna @ file:///home/linux1/recipes/ci/idna_1610986105248/work 28 | imageio==2.9.0 29 | imgaug==0.4.0 30 | importlib-metadata @ file:///tmp/build/80754af9/importlib-metadata_1602276842396/work 31 | joblib @ file:///home/conda/feedstock_root/build_artifacts/joblib_1607956439537/work 32 | kiwisolver @ file:///C:/ci/kiwisolver_1612282618948/work 33 | Markdown @ file:///C:/ci/markdown_1605111187600/work 34 | mat73==0.50 35 | matplotlib @ file:///C:/ci/matplotlib-base_1603356257853/work 36 | mkl-fft==1.2.0 37 | mkl-random==1.1.1 38 | mkl-service==2.3.0 39 | multidict @ file:///C:/ci/multidict_1600456486794/work 40 | munkres==1.1.4 41 | networkx==2.6.3 42 | numpy @ file:///C:/ci/numpy_and_numpy_base_1603468620949/work 43 | oauthlib==3.1.0 44 | olefile==0.46 45 | opencv-python==4.5.3.56 46 | opencv-python-headless==4.5.3.56 47 | packaging==21.0 48 | pandas @ file:///C:/ci/pandas_1613686255372/work 49 | Pillow @ file:///C:/ci/pillow_1609786872067/work 50 | pipreqs==0.4.10 51 | protobuf==3.14.0 52 | pyasn1==0.4.8 53 | pyasn1-modules==0.2.8 54 | pycparser @ file:///tmp/build/80754af9/pycparser_1594388511720/work 55 | PyJWT @ file:///C:/ci/pyjwt_1610911411733/work 56 | pyOpenSSL @ file:///tmp/build/80754af9/pyopenssl_1608057966937/work 57 | pyparsing @ file:///home/linux1/recipes/ci/pyparsing_1610983426697/work 58 | pyreadline==2.1 59 | PySocks @ file:///C:/ci/pysocks_1594394709107/work 60 | python-dateutil @ file:///home/ktietz/src/ci/python-dateutil_1611928101742/work 61 | pytorch-metric-learning==0.9.95 62 | pytorch-ranger==0.1.1 63 | pytz @ file:///tmp/build/80754af9/pytz_1612215392582/work 64 | PyWavelets==1.1.1 65 | PyYAML==5.4.1 66 | requests @ file:///tmp/build/80754af9/requests_1608241421344/work 67 | requests-oauthlib==1.3.0 68 | rsa @ file:///tmp/build/80754af9/rsa_1610483308194/work 69 | SciencePlots==1.0.6 70 | scikit-image==0.18.3 71 | scikit-learn @ file:///C:/ci/scikit-learn_1622739441385/work 72 | scikit-learn-intelex==2021.20210504.131156 73 | scipy @ file:///C:/ci/scipy_1597686737426/work 74 | seaborn @ file:///tmp/build/80754af9/seaborn_1608578541026/work 75 | Shapely==1.7.1 76 | six @ file:///C:/ci/six_1605205426665/work 77 | tensorboard @ file:///home/builder/ktietz/conda/conda-bld/tensorboard_1604313476433/work/tmp_pip_dir 78 | tensorboard-logger==0.1.0 79 | tensorboard-plugin-wit==1.6.0 80 | threadpoolctl @ file:///tmp/tmp79xdzxkt/threadpoolctl-2.1.0-py3-none-any.whl 81 | tifffile==2021.8.30 82 | timm==0.3.4 83 | torch==1.7.1 84 | torch-ema @ git+https://github.com/fadel/pytorch_ema@3985995e523aa25dd3cff7e7984130eef90a4282 85 | torch-lr-finder==0.2.1 86 | torch-optimizer==0.0.1a17 87 | torchaudio==0.7.2 88 | torchvision==0.8.2 89 | tornado @ file:///C:/ci/tornado_1606935947090/work 90 | tqdm @ file:///tmp/build/80754af9/tqdm_1611857934208/work 91 | typing-extensions @ file:///tmp/build/80754af9/typing_extensions_1598376058250/work 92 | urllib3 @ file:///tmp/build/80754af9/urllib3_1611694770489/work 93 | validators @ file:///tmp/build/80754af9/validators_1612286467315/work 94 | Werkzeug @ file:///home/ktietz/src/ci/werkzeug_1611932622770/work 95 | win-inet-pton @ file:///C:/ci/win_inet_pton_1605306165655/work 96 | wincertstore==0.2 97 | yarg==0.1.9 98 | yarl @ file:///C:/ci/yarl_1598045271760/work 99 | zipp @ file:///tmp/build/80754af9/zipp_1604001098328/work 100 | -------------------------------------------------------------------------------- /util.py: -------------------------------------------------------------------------------- 1 | ''' 2 | This is written by Jiyuan Liu, Dec. 21, 2021. 3 | Homepage: https://liujiyuan13.github.io. 4 | Email: liujiyuan13@163.com. 5 | All rights reserved. 6 | ''' 7 | 8 | import os 9 | import numpy as np 10 | import random 11 | import torch 12 | import torch.backends.cudnn as cudnn 13 | from torchvision import transforms, datasets 14 | from torch.utils.data import DataLoader 15 | 16 | 17 | def set_seed(seed=0): 18 | """ 19 | set seed for torch. 20 | @param seed: int, default 0 21 | """ 22 | random.seed(seed) 23 | os.environ['PYTHONHASHSEED'] = str(seed) 24 | np.random.seed(seed) 25 | torch.manual_seed(seed) 26 | torch.cuda.manual_seed(seed) 27 | torch.cuda.manual_seed_all(seed) # if you are using multi-GPU. 28 | torch.backends.cudnn.benchmark = False 29 | torch.backends.cudnn.deterministic = True 30 | torch.backends.cudnn.enabled = False 31 | 32 | 33 | def load_data(data_dir, data_name, is_train, image_size, batch_size, n_worker): 34 | """ 35 | load data. 36 | @param data_dir: data dir, data folder 37 | @param data_name: data name 38 | @param is_train: train data or test data 39 | @param image_size: image size 40 | @param batch_size: batch size 41 | @param n_worker: number of workers to load data 42 | @return: data_loader: loader for train data; 43 | n_class: number of data classes 44 | """ 45 | 46 | # load data 47 | if data_name is 'cifar10': 48 | MEAN, STD = (0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010) 49 | transform = transforms.Compose([ 50 | transforms.RandomResizedCrop(image_size), 51 | transforms.ToTensor(), 52 | transforms.Normalize(mean=MEAN, std=STD) 53 | ]) 54 | data = datasets.CIFAR10(data_dir, transform=transform, train=is_train, download=True) 55 | elif data_name is 'cifar100': 56 | MEAN, STD = (0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761) 57 | transform = transforms.Compose([ 58 | transforms.RandomResizedCrop(image_size), 59 | transforms.ToTensor(), 60 | transforms.Normalize(mean=MEAN, std=STD) 61 | ]) 62 | data = datasets.CIFAR100(data_dir, transform=transform, train=is_train, download=True) 63 | elif data_name is 'stl10': 64 | transform = transforms.Compose([ 65 | transforms.RandomResizedCrop(image_size), 66 | transforms.ToTensor() 67 | ]) 68 | data = datasets.STL10(data_dir, transform=transform, split='train' if is_train else 'test', download=True) 69 | elif data_name is 'imagenet': 70 | MEAN, STD = (0.485, 0.456, 0.406), (0.229, 0.224, 0.225) # constants in timm.data.constants 71 | transform = transforms.Compose([ 72 | transforms.RandomResizedCrop(image_size), 73 | transforms.ToTensor(), 74 | transforms.Normalize(mean=MEAN, std=STD) 75 | ]) 76 | data = datasets.ImageFolder(os.path.join(data_dir, 'ImageNet1K', 'train' if is_train else 'val'), transform=transform) 77 | else: 78 | raise Exception(data_name + ': not supported yet.') 79 | 80 | # obtain class number from test data 81 | n_class = len(set(data.targets)) 82 | 83 | # create data loader 84 | data_loader = DataLoader(data, 85 | batch_size=batch_size, 86 | shuffle=True, 87 | num_workers=n_worker, 88 | pin_memory=True, 89 | drop_last=True) 90 | 91 | return data_loader, n_class 92 | 93 | 94 | def save_ckpt(model, optimizer, args, epoch, save_file): 95 | ''' 96 | save checkpoint 97 | :param model: target model 98 | :param optimizer: used optimizer 99 | :param args: training parameters 100 | :param epoch: save at which epoch 101 | :param save_file: file path 102 | :return: 103 | ''' 104 | ckpt = { 105 | 'args': args, 106 | 'model': model.state_dict(), 107 | 'optimizer': optimizer.state_dict(), 108 | 'epoch': epoch, 109 | } 110 | torch.save(ckpt, save_file) 111 | del ckpt 112 | 113 | 114 | def load_ckpt(model, load_file): 115 | ''' 116 | load ckpt to model 117 | :param model: target model 118 | :param load_file: file path 119 | :return: the loaded model 120 | ''' 121 | ckpt = torch.load(load_file) 122 | model.load_state_dict(ckpt['model']) 123 | del ckpt 124 | return model 125 | 126 | 127 | def accuracy(y_true, y_pred): 128 | """ 129 | compute classification accuracy. 130 | # Arguments 131 | y: true labels, numpy.array with shape `(n_samples,)` 132 | y_pred: predicted labels, numpy.array with shape `(n_samples,)` 133 | # Return 134 | accuracy, in [0,1] 135 | """ 136 | assert y_pred.size == y_true.size 137 | y_true, y_pred = y_true.astype(np.int64), y_pred.astype(np.int64) 138 | return sum(np.equal(y_true, y_pred)) / y_true.size 139 | 140 | 141 | class AverageMeter(object): 142 | ''' 143 | compute and store the average and current value 144 | ''' 145 | def __init__(self): 146 | self.reset() 147 | 148 | def reset(self): 149 | self.val = 0 150 | self.avg = 0 151 | self.sum = 0 152 | self.count = 0 153 | 154 | def update(self, val, n=1): 155 | self.val = val 156 | self.sum += val * n 157 | self.count += n 158 | self.avg = self.sum / self.count -------------------------------------------------------------------------------- /vit.py: -------------------------------------------------------------------------------- 1 | ''' 2 | This is from https://github.com/lucidrains/vit-pytorch. 3 | ''' 4 | 5 | import torch 6 | from torch import nn 7 | 8 | from einops import rearrange, repeat 9 | from einops.layers.torch import Rearrange 10 | 11 | # helpers 12 | 13 | def pair(t): 14 | return t if isinstance(t, tuple) else (t, t) 15 | 16 | 17 | # classes 18 | 19 | 20 | class PreNorm(nn.Module): 21 | def __init__(self, dim, fn): 22 | super().__init__() 23 | self.norm = nn.LayerNorm(dim) 24 | self.fn = fn 25 | def forward(self, x, **kwargs): 26 | return self.fn(self.norm(x), **kwargs) 27 | 28 | 29 | class FeedForward(nn.Module): 30 | def __init__(self, dim, hidden_dim, dropout = 0.): 31 | super().__init__() 32 | self.net = nn.Sequential( 33 | nn.Linear(dim, hidden_dim), 34 | nn.GELU(), 35 | nn.Dropout(dropout), 36 | nn.Linear(hidden_dim, dim), 37 | nn.Dropout(dropout) 38 | ) 39 | 40 | def forward(self, x): 41 | return self.net(x) 42 | 43 | 44 | class Attention(nn.Module): 45 | def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.): 46 | super().__init__() 47 | inner_dim = dim_head * heads 48 | project_out = not (heads == 1 and dim_head == dim) 49 | 50 | self.heads = heads 51 | self.scale = dim_head ** -0.5 52 | 53 | self.attend = nn.Softmax(dim = -1) 54 | self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False) 55 | 56 | self.to_out = nn.Sequential( 57 | nn.Linear(inner_dim, dim), 58 | nn.Dropout(dropout) 59 | ) if project_out else nn.Identity() 60 | 61 | def forward(self, x): 62 | qkv = self.to_qkv(x).chunk(3, dim = -1) 63 | q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv) 64 | 65 | dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale 66 | 67 | attn = self.attend(dots) 68 | 69 | out = torch.matmul(attn, v) 70 | out = rearrange(out, 'b h n d -> b n (h d)') 71 | return self.to_out(out) 72 | 73 | 74 | class Transformer(nn.Module): 75 | def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.): 76 | super().__init__() 77 | self.layers = nn.ModuleList([]) 78 | for _ in range(depth): 79 | self.layers.append(nn.ModuleList([ 80 | PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)), 81 | PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)) 82 | ])) 83 | 84 | def forward(self, x): 85 | for attn, ff in self.layers: 86 | x = attn(x) + x 87 | x = ff(x) + x 88 | return x 89 | 90 | 91 | class ViT(nn.Module): 92 | def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.): 93 | super().__init__() 94 | image_height, image_width = pair(image_size) 95 | patch_height, patch_width = pair(patch_size) 96 | 97 | assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.' 98 | 99 | num_patches = (image_height // patch_height) * (image_width // patch_width) 100 | patch_dim = channels * patch_height * patch_width 101 | assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)' 102 | 103 | self.to_patch_embedding = nn.Sequential( 104 | Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width), 105 | nn.Linear(patch_dim, dim), 106 | ) 107 | 108 | self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim)) 109 | self.cls_token = nn.Parameter(torch.randn(1, 1, dim)) 110 | self.dropout = nn.Dropout(emb_dropout) 111 | 112 | self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout) 113 | 114 | self.pool = pool 115 | self.to_latent = nn.Identity() 116 | 117 | self.mlp_head = nn.Sequential( 118 | nn.LayerNorm(dim), 119 | nn.Linear(dim, num_classes) 120 | ) 121 | 122 | def forward(self, img): 123 | x = self.to_patch_embedding(img) 124 | b, n, _ = x.shape 125 | 126 | cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b) 127 | x = torch.cat((cls_tokens, x), dim=1) 128 | x += self.pos_embedding[:, :(n + 1)] 129 | x = self.dropout(x) 130 | 131 | x = self.transformer(x) 132 | 133 | x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0] 134 | 135 | x = self.to_latent(x) 136 | return self.mlp_head(x) 137 | --------------------------------------------------------------------------------