├── .gitignore ├── LICENSE ├── README.md ├── assets └── PMN_teaser.png ├── download_data.sh ├── environment-latest.yml ├── environment.yml ├── main.py ├── models ├── __init__.py ├── classifier.py ├── communication.py ├── fgsbir │ ├── model.py │ └── networks.py ├── pmn.py └── transformer.py ├── save_processed_images.py ├── trainer.py └── utils ├── arguments.py ├── coords_utils.py ├── data.py ├── distance_transform.py ├── metrics.py ├── plt_blit.py └── transformations.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Custom 2 | log*/ 3 | data/ 4 | 5 | # Byte-compiled / optimized / DLL files 6 | __pycache__/ 7 | *.py[cod] 8 | *$py.class 9 | 10 | # C extensions 11 | *.so 12 | 13 | # Distribution / packaging 14 | .Python 15 | build/ 16 | develop-eggs/ 17 | dist/ 18 | downloads/ 19 | eggs/ 20 | .eggs/ 21 | lib/ 22 | lib64/ 23 | parts/ 24 | sdist/ 25 | var/ 26 | wheels/ 27 | share/python-wheels/ 28 | *.egg-info/ 29 | .installed.cfg 30 | *.egg 31 | MANIFEST 32 | 33 | # PyInstaller 34 | # Usually these files are written by a python script from a template 35 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 36 | *.manifest 37 | *.spec 38 | 39 | # Installer logs 40 | pip-log.txt 41 | pip-delete-this-directory.txt 42 | 43 | # Unit test / coverage reports 44 | htmlcov/ 45 | .tox/ 46 | .nox/ 47 | .coverage 48 | .coverage.* 49 | .cache 50 | nosetests.xml 51 | coverage.xml 52 | *.cover 53 | *.py,cover 54 | .hypothesis/ 55 | .pytest_cache/ 56 | cover/ 57 | 58 | # Translations 59 | *.mo 60 | *.pot 61 | 62 | # Django stuff: 63 | *.log 64 | local_settings.py 65 | db.sqlite3 66 | db.sqlite3-journal 67 | 68 | # Flask stuff: 69 | instance/ 70 | .webassets-cache 71 | 72 | # Scrapy stuff: 73 | .scrapy 74 | 75 | # Sphinx documentation 76 | docs/_build/ 77 | 78 | # PyBuilder 79 | .pybuilder/ 80 | target/ 81 | 82 | # Jupyter Notebook 83 | .ipynb_checkpoints 84 | 85 | # IPython 86 | profile_default/ 87 | ipython_config.py 88 | 89 | # pyenv 90 | # For a library or package, you might want to ignore these files since the code is 91 | # intended to run in multiple environments; otherwise, check them in: 92 | # .python-version 93 | 94 | # pipenv 95 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 96 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 97 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 98 | # install all needed dependencies. 99 | #Pipfile.lock 100 | 101 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 102 | __pypackages__/ 103 | 104 | # Celery stuff 105 | celerybeat-schedule 106 | celerybeat.pid 107 | 108 | # SageMath parsed files 109 | *.sage.py 110 | 111 | # Environments 112 | .env 113 | .venv 114 | env/ 115 | venv/ 116 | ENV/ 117 | env.bak/ 118 | venv.bak/ 119 | 120 | # Spyder project settings 121 | .spyderproject 122 | .spyproject 123 | 124 | # Rope project settings 125 | .ropeproject 126 | 127 | # mkdocs documentation 128 | /site 129 | 130 | # mypy 131 | .mypy_cache/ 132 | .dmypy.json 133 | dmypy.json 134 | 135 | # Pyre type checker 136 | .pyre/ 137 | 138 | # pytype static type analyzer 139 | .pytype/ 140 | 141 | # Cython debug symbols 142 | cython_debug/ 143 | 144 | # IDEA 145 | .idea/ 146 | -------------------------------------------------------------------------------- /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 | # Abstracting Sketches through Simple Primitives 2 | 3 | [[Blog]](https://www.eml-unitue.de/publication/abstracting-sketches-through-simple-primitives) [[Paper]](https://arxiv.org/abs/2207.13543) 4 | 5 | This is the official PyTorch code for our ECCV 2022 paper on Abstracting Sketches through Simple Primitives. It allows training and testing of our Primitive Matching Network (PMN) on Quickdraw, ShoeV2 and ChairV2 datasets. 6 | 7 | ![PMN Teaser](./assets/PMN_teaser.png) 8 | 9 | 10 | ## Installation 11 | All Python dependencies can be installed into a conda environment with the provided environment.yml file. To setup the code and environment, run the following commands. 12 | 13 | 1. Clone the repository 14 | ```shell 15 | git clone https://github.com/ExplainableML/sketch-primitives.git 16 | cd sketch-primitives 17 | ``` 18 | 2. Create conda environment (you can try to use `environment-latest.yml` if you want to use the latest PyTorch version) 19 | ```shell 20 | conda env create -f environment.yml 21 | ``` 22 | 3. Activate environment 23 | ```shell 24 | conda activate sketch-primitives 25 | ``` 26 | 4. The script `./download_data.sh` downloads Quickdraw, ShoeV2 and ChairV2 into `./data` 27 | 28 | 29 | ## Usage 30 | The following commands are examples for training the PMN model. To see all possible parameters, use `python main.py --help`. By default, logs are saved into `./log` including model checkpoints and tensorboards. Data is expected to be in `./data` unless otherwise specified. 31 | 32 | 33 | ### Classification on Quickdraw 34 | * Train sketch classifier on the original dataset for evaluation purposes. 35 | ``` 36 | python main.py --log-name classifier_quickdraw09 --model-type classifier --dataset quickdraw09 37 | ``` 38 | 39 | * Train the PMN model self-supervised to do sketch abstractions with primitives. 40 | ``` 41 | python main.py --log-name pmn_quickdraw09 --model-type pmn --dataset quickdraw09 42 | ``` 43 | 44 | * Process the dataset with PMN to create a new dataset consisting of primitive reconstructions. 45 | ``` 46 | python save_processed_images.py --log-name pmn_quickdraw09_processed --model-type pmn --dataset quickdraw09 --test ./log/pmn_quickdraw09/model_best_acc.pt --only-save-coords 47 | ``` 48 | 49 | * Evaluate PMN reconstructions using a communication task with limited budget and the previously trained sketch classifier. 50 | ``` 51 | python main.py --log-name pmn_communication_quickdraw09 --model-type communication --dataset quickdraw09 --loss-model-type class --loss-model-ckpt ./log/classifier_quickdraw09/model_best_acc.pt --budget 0.3 --test 1 --preprocessed-root ./log/pmn_quickdraw09_processed/data/coords/ 52 | ``` 53 | 54 | 55 | ### Fine-grainded Sketch-based Image retrieval (FG-SBIR) on ShoeV2 and ChairV2 56 | * Create sketch images with our rendering (to avoid distribution shift). 57 | ``` 58 | python save_processed_images.py --log-name chairv2_processed --model-type none --dataset chairv2 59 | ln -sf $(pwd)/log/chairv2_processed/data/imgs data/chairv2/sketch_picture_files 60 | ``` 61 | 62 | * Train a sketch-based image retrieval model on the original data for use as an evaluation model. 63 | ``` 64 | python main.py --log-name fgsbir_chairv2 --model-type fgsbir --dataset chairv2 --load-fgsbir-photos 65 | ``` 66 | 67 | * Train the PMN model self-supervised to do sketch abstractions with primitives. 68 | ``` 69 | python main.py --log-name pmn_chairv2 --model-type pmn --dataset chairv2 70 | ``` 71 | 72 | * Process the dataset with PMN to create a new dataset consisting of primitive reconstructions including sketch images for a specific budget. 73 | ``` 74 | python save_processed_images.py --log-name pmn_chairv2_processed_budget0.3 --model-type pmn --dataset chairv2 --test ./log/pmn_chairv2/model_best_acc.pt --budget 0.3 --loss-model-type none --add-comm-model --load-fgsbir-photos 75 | ``` 76 | 77 | * Process the dataset with PMN to create a new dataset consisting of primitive reconstructions including sketch images for a specific budget. 78 | ``` 79 | python save_processed_images.py --log-name pmn_chairv2_processed_budget0.3 --model-type pmn --dataset chairv2 --test ./log/pmn_chairv2/model_best_acc.pt --budget 0.3 --loss-model-type none --add-comm-model --load-fgsbir-photos 80 | ln -sf $(pwd)/log/pmn_chairv2_processed_budget0.3/data/imgs data/chairv2/sketch_picture_files 81 | ``` 82 | 83 | * Evaluate PMN reconstructions on the previously trained FG-SBIR model. 84 | ``` 85 | python main.py --log-name pmn_communication_chairv2_budget0.3 --model-type fgsbir --dataset chairv2 --load-fgsbir-photos --test log/fgsbir_chairv2/model_best_acc.pt 86 | ``` 87 | 88 | 89 | ## Citation 90 | If you use this code, please cite 91 | ``` 92 | @inproceedings{alaniz2022primitives, 93 | title = {Abstracting Sketches through Simple Primitives}, 94 | author = {Alaniz, S. and Mancini, M. and Dutta, A. and Marcos, D. and Akata, Z.}, 95 | booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)}, 96 | year = {2022} 97 | } 98 | ``` 99 | -------------------------------------------------------------------------------- /assets/PMN_teaser.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ExplainableML/sketch-primitives/32b025c9bb29bc0354cf2a06075a942ac7712ad6/assets/PMN_teaser.png -------------------------------------------------------------------------------- /download_data.sh: -------------------------------------------------------------------------------- 1 | mkdir -p ./data 2 | cd ./data 3 | 4 | # Download and extract ShoeV2 and ChairV2 5 | wget https://nc.mlcloud.uni-tuebingen.de/index.php/s/ZsydaSGocdxSy2k/download -O datav2.tar.gz 6 | tar xvf datav2.tar.gz 7 | rm datav2.tar.gz 8 | 9 | # Download and extract Quickdraw09 10 | wget https://nc.mlcloud.uni-tuebingen.de/index.php/s/JLeNSFQdGgkEPet/download -O quickdraw09.tar.gz 11 | tar xvf quickdraw09.tar.gz 12 | rm quickdraw09.tar.gz 13 | -------------------------------------------------------------------------------- /environment-latest.yml: -------------------------------------------------------------------------------- 1 | name: sketch-primitives 2 | channels: 3 | - pytorch 4 | - nvidia 5 | - conda-forge 6 | - defaults 7 | dependencies: 8 | - matplotlib 9 | - numpy 10 | - pillow 11 | - pip 12 | - python 13 | - pytorch 14 | - pytorch-cuda 15 | - tar 16 | - tensorboard 17 | - torchvision 18 | - tqdm 19 | - wget 20 | -------------------------------------------------------------------------------- /environment.yml: -------------------------------------------------------------------------------- 1 | name: sketch-primitives 2 | channels: 3 | - pytorch 4 | - conda-forge 5 | - defaults 6 | dependencies: 7 | - cudatoolkit=10.2 8 | - matplotlib=3.5.2 9 | - numpy=1.23.1 10 | - pillow=9.2.0 11 | - pip=22.1.2 12 | - python=3.10.5 13 | - pytorch=1.12.0 14 | - tar=1.34 15 | - tensorboard=2.9.1 16 | - torchvision=0.13.0 17 | - tqdm=4.64.0 18 | - wget=1.20.3 19 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import os 2 | import datetime 3 | 4 | import numpy as np 5 | import torch 6 | 7 | from models import get_model 8 | from utils.data import get_data 9 | from utils.arguments import parser 10 | from trainer import Trainer 11 | 12 | if __name__ == '__main__': 13 | args = parser.parse_args() 14 | 15 | DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu' 16 | 17 | # Add timestamp to logdir 18 | if args.log_name is None: 19 | LOG_DIR = os.path.join(args.log_dir, 20 | datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')) 21 | else: 22 | LOG_DIR = os.path.join(args.log_dir, 23 | args.log_name) 24 | 25 | if args.test is not None: 26 | torch.manual_seed(42) 27 | np.random.seed(42) 28 | 29 | dataloader_coords, n_classes, label_names = get_data( 30 | args.data_dir, args.dataset, batch_size=args.batch_size, 31 | train=True, 32 | preprocessed_root=args.preprocessed_root, 33 | load_fgsbir_photos=args.load_fgsbir_photos, 34 | resample_strokes=not args.no_stroke_resampling, 35 | max_stroke_length=args.max_stroke_length 36 | ) 37 | 38 | test_dataloader_coords, _, _ = get_data( 39 | args.data_dir, args.dataset, batch_size=args.n_log_img, 40 | train=False, 41 | preprocessed_root=args.preprocessed_root, 42 | load_fgsbir_photos=args.load_fgsbir_photos, 43 | resample_strokes=not args.no_stroke_resampling, 44 | max_stroke_length=args.max_stroke_length 45 | ) 46 | 47 | model = get_model(args, n_classes, DEVICE) 48 | 49 | if args.test is not None and not args.model_type == 'communication': 50 | model.load_state_dict(torch.load(args.test)) 51 | 52 | if args.test is not None: 53 | torch.manual_seed(42) 54 | np.random.seed(42) 55 | 56 | trainer = Trainer(args.model_type, model, args.test is None, args.learning_rate, args.epochs, dataloader_coords, test_dataloader_coords, label_names, DEVICE, LOG_DIR, args.log_every, args.n_log_img, args.save_every, args.view_scale, args.weight_decay) 57 | trainer.run() 58 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from .pmn import PrimitiveMatchingNetwork 4 | from .classifier import SketchClassifier 5 | from .communication import Communication 6 | from .fgsbir.model import FGSBIR_Model 7 | 8 | def get_model(args, n_classes, DEVICE): 9 | if args.model_type == 'none': 10 | model = None 11 | elif args.model_type == 'pmn': 12 | model = PrimitiveMatchingNetwork( 13 | temperature=args.temperature, 14 | use_pos_embed=args.use_pos_embed, 15 | use_sinusoid_embed=args.use_sinusoid_embed, 16 | transformations=args.transformations, 17 | use_proportion_scale=args.use_proportion_scale, 18 | transform_bound=args.transform_bound, 19 | Nz=args.Nz, 20 | dt_gamma=args.dt_gamma, 21 | learn_compatibility=args.learn_compatibility) 22 | model = model.to(DEVICE) 23 | 24 | elif args.model_type == 'classifier': 25 | model = SketchClassifier(n_classes=n_classes, use_pos_embed=args.use_pos_embed, 26 | use_sinusoid_embed=args.use_sinusoid_embed).to(DEVICE) 27 | elif args.model_type == 'communication': 28 | if args.loss_model_type == 'class': 29 | classifier = SketchClassifier(n_classes=n_classes, use_pos_embed=args.use_pos_embed, 30 | use_sinusoid_embed=args.use_sinusoid_embed).to(DEVICE) 31 | elif args.loss_model_type == 'fgsbir': 32 | classifier = FGSBIR_Model(backbone=args.fgsbir_backbone, 33 | ce_temp=args.temperature).to(DEVICE) 34 | elif args.loss_model_type == 'none': 35 | classifier = None 36 | else: 37 | raise NotImplementedError 38 | 39 | if args.loss_model_type != 'none': 40 | classifier.load_state_dict(torch.load(args.loss_model_ckpt)) 41 | 42 | if args.loss_model_type == 'class': 43 | classifier = classifier.encoder 44 | 45 | model = Communication(classifier=classifier, 46 | loss_type=args.loss_model_type, 47 | comm_mode=args.comm_mode, 48 | budget=args.budget, 49 | ).to(DEVICE) 50 | elif args.model_type == 'fgsbir': 51 | model = FGSBIR_Model(backbone=args.fgsbir_backbone, 52 | ce_temp=args.temperature, 53 | loss_type=args.loss_model_type).to(DEVICE) 54 | else: 55 | raise NotImplementedError 56 | 57 | return model 58 | 59 | -------------------------------------------------------------------------------- /models/classifier.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torchvision.models import resnet18 5 | 6 | from .transformer import StrokeEncoderTransformer 7 | 8 | class SketchClassifier(nn.Module): 9 | def __init__(self, n_classes, Nz=256, use_pos_embed=False, 10 | use_sinusoid_embed=False, stroke_len=25): 11 | super().__init__() 12 | self.Nz = Nz 13 | self.encoder = StrokeEncoderTransformer( 14 | Nz=Nz, embed_sketch=True, n_layer_strokes=3, use_pos_embed=use_pos_embed, 15 | use_sinusoid_embed=use_sinusoid_embed, out_dim=n_classes, stroke_len=stroke_len 16 | ) 17 | self.criterion = nn.CrossEntropyLoss(reduction='none') 18 | 19 | def forward(self, x_raw, label, n_strokes, scale, translate, stroke_lens, seq_mask, position, **kwargs): 20 | # encode: 21 | cls_logits, _ = self.encoder(x_raw.transpose(0, 1), n_strokes, scale, translate, stroke_lens, seq_mask, position) 22 | loss = self.criterion(cls_logits, label) 23 | 24 | _, pred_main = torch.max(cls_logits, 1) 25 | acc = (label == pred_main).float() 26 | clspred_main_prob = torch.softmax(cls_logits, dim=1).gather(1, pred_main.unsqueeze(1)).squeeze(1) 27 | 28 | output = {'loss': loss, 29 | 'acc_main': acc, 30 | 'cls_pred_main': pred_main, 31 | 'cls_pred_main_prob': clspred_main_prob, 32 | 'pcoords_raw': x_raw, 33 | 'pcoords_n_strokes': n_strokes, 34 | 'pcoords_scale': scale, 35 | 'pcoords_translate': translate, 36 | 'pcoords_seq_mask': seq_mask} 37 | 38 | return output 39 | -------------------------------------------------------------------------------- /models/communication.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torchvision.models import resnet18 5 | import torchvision.transforms as transforms 6 | import matplotlib.pyplot as plt 7 | 8 | from .transformer import StrokeEncoderTransformer 9 | from utils.plt_blit import BlitManager 10 | from utils.coords_utils import make_image_fast 11 | 12 | def calc_segments_per_sketch(segments, stroke_lens, n_strokes): 13 | if len(segments) == 1: 14 | segments = segments[0] 15 | elif not isinstance(segments, torch.Tensor): 16 | segments = torch.stack(segments).T.flatten() 17 | 18 | grouped_segments = [] 19 | group = [] 20 | cnt = 0 21 | sid = 0 22 | for seg in segments: 23 | group.append(seg) 24 | cnt += seg 25 | if cnt == stroke_lens[sid]: 26 | grouped_segments.append(torch.stack(group)) 27 | cnt = 0 28 | sid += 1 29 | group = [] 30 | 31 | grouped_segments_len = [gs.shape[0] for gs in grouped_segments] 32 | 33 | segments_per_sketch = [] 34 | cnt = 0 35 | for n_s in n_strokes: 36 | segments_per_sketch.append(sum(grouped_segments_len[cnt:cnt+n_s])) 37 | cnt += n_s 38 | segments_per_sketch = torch.tensor(segments_per_sketch, device=n_strokes.device) 39 | 40 | return grouped_segments, grouped_segments_len, segments_per_sketch 41 | 42 | class Communication(nn.Module): 43 | def __init__(self, classifier, Nz=256, comm_mode='human', 44 | loss_type='class', budget=1.0): 45 | super().__init__() 46 | assert comm_mode in ['human', 'random'] 47 | self.Nz = Nz 48 | self.classifier = classifier 49 | if self.classifier is not None: 50 | for p in self.classifier.parameters(): 51 | p.requires_grad = False 52 | 53 | self.criterion = nn.CrossEntropyLoss(reduction='none') 54 | self.comm_mode = comm_mode 55 | self.loss_type = loss_type 56 | self.budget = budget 57 | self.gamma = 0.9 58 | 59 | if self.loss_type == 'fgsbir' and not self.classifier.use_transformer: 60 | fig, ax = plt.subplots(figsize=(2.99, 2.99), constrained_layout=True) 61 | ax.set_axis_off() 62 | fig.add_axes(ax) 63 | self.blit_manager = BlitManager(ax, fig.canvas) 64 | fig.canvas.draw() 65 | 66 | self.img_transform = transforms.Normalize(mean=[0.485, 0.456, 0.406], 67 | std=[0.229, 0.224, 0.225]) 68 | 69 | 70 | def forward(self, x_raw, label, n_strokes, scale, translate, segments=None, stroke_lens=None, seq_mask=None, position=None, **kwargs): 71 | 72 | grouped_segments, grouped_segments_len, segments_per_sketch = calc_segments_per_sketch( 73 | segments, stroke_lens, 74 | kwargs['n_strokes_orig'] if 'n_strokes_orig' in kwargs else n_strokes 75 | ) 76 | 77 | # encode: 78 | ol_size = (n_strokes.shape[0], n_strokes.max().item()) 79 | if self.comm_mode.endswith('human'): 80 | arange_size = ol_size[1] 81 | order_logits = -torch.arange(arange_size, dtype=x_raw.dtype, device=x_raw.device).unsqueeze(0).expand((ol_size[0], arange_size)) 82 | order_logits = order_logits.view(ol_size) 83 | else: # 'random' 84 | order_logits = torch.randn(ol_size, device=x_raw.device) 85 | 86 | padding_mask = torch.zeros(ol_size, device=x_raw.device, dtype=torch.bool) 87 | stroke_id = 0 88 | for i, n_s in enumerate(n_strokes): 89 | pad_len = ol_size[1] - n_s 90 | if pad_len != 0: 91 | padding_mask[i, -pad_len:] = True 92 | stroke_id += n_s 93 | 94 | order_logits = order_logits.masked_fill(padding_mask, float('-inf')) 95 | order_all = torch.full_like(order_logits, float('-inf')) 96 | 97 | final_cls_logits = None 98 | all_cls_logits = [] 99 | cls_input = torch.zeros_like(x_raw) 100 | mask = torch.ones_like(padding_mask) 101 | grad_neg_mask = torch.zeros_like(x_raw[:, 0, 0]) 102 | points_mask = None 103 | losses = [] 104 | 105 | budget_n_strokes = torch.ceil(segments_per_sketch * self.budget).int() 106 | budget_n_strokes = torch.stack([budget_n_strokes, n_strokes]).min(0)[0] 107 | n_iters = max(budget_n_strokes) 108 | 109 | for i in range(n_iters): 110 | done = order_logits.isinf().sum(1) == order_logits.shape[1] 111 | done = done.unsqueeze(1).expand(order_logits.shape) 112 | 113 | order_logits = order_logits.masked_fill(done, 0.) 114 | 115 | next_order = F.one_hot(order_logits.argmax(dim=1), num_classes=order_logits.shape[1]) 116 | next_order = next_order.float() 117 | 118 | next_order_flat = [] 119 | global_stroke_id = 0 120 | for bi, (no, n_s) in enumerate(zip(next_order, n_strokes)): 121 | next_order_flat.append(no[:n_s]) 122 | global_stroke_id += n_s 123 | next_order_flat = torch.cat(next_order_flat) 124 | next_order_flat = next_order_flat.masked_fill(next_order_flat.isnan(), 0.) 125 | 126 | active_sketches = i < budget_n_strokes 127 | active_strokes = torch.cat([a.repeat(n_s) for a, n_s in zip(active_sketches, n_strokes)]) 128 | 129 | grad_neg_mask += active_strokes * next_order_flat 130 | cls_input = x_raw * grad_neg_mask.unsqueeze(-1).unsqueeze(-1) 131 | points_mask_input = None 132 | mask = mask & ~(active_sketches.unsqueeze(1) & next_order.bool()).detach() 133 | if self.loss_type == 'fgsbir': 134 | if not self.classifier.use_transformer: 135 | sketch_images= [] 136 | n_s = list(n_strokes) 137 | # TODO: remove duplicate code 138 | if points_mask_input is None: 139 | for ci, s, t in zip(cls_input.split(n_s), scale.split(n_s), translate.split(n_s)): 140 | sketch_img = make_image_fast(self.blit_manager, (ci.cpu(), s.cpu(), t.cpu(), None), no_axis=True, color='black', linewidth=1, enlarge=2.0, arbitrary_mask=True) 141 | sketch_images.append(self.img_transform(torch.tensor(sketch_img, dtype=torch.float)/255.)) 142 | else: 143 | for ci, s, t, m in zip(cls_input.split(n_s), scale.split(n_s), translate.split(n_s), points_mask_input[:, :-1].split(n_s)): 144 | sketch_img = make_image_fast(self.blit_manager, (ci.cpu(), s.cpu(), t.cpu(), m.cpu()), no_axis=True, color='black', linewidth=1, enlarge=2.0, arbitrary_mask=True) 145 | sketch_images.append(self.img_transform(torch.tensor(sketch_img, dtype=torch.float)/255.)) 146 | kwargs['sketch_img'] = torch.stack(sketch_images).to(cls_input) 147 | outputs = self.classifier(cls_input, n_strokes, scale, translate, seq_mask=seq_mask, position=position, stroke_lens=stroke_lens, mask=torch.cat([mask, torch.zeros_like(mask[:, :1])], dim=1), points_mask=points_mask_input, **kwargs) 148 | cls_loss = outputs['loss'][active_sketches] 149 | sketch_feature = outputs['sketch_feature'] 150 | image_feature = outputs['image_feature'] 151 | elif self.loss_type == 'class': 152 | cls_logits, _ = self.classifier(cls_input.transpose(0, 1), n_strokes, scale, translate, seq_mask=seq_mask, position=position, mask=torch.cat([mask, torch.zeros_like(mask[:, :1])], dim=1), points_mask=points_mask_input, **kwargs) 153 | 154 | all_cls_logits.append(cls_logits) 155 | if final_cls_logits is None: 156 | final_cls_logits = cls_logits.clone().detach() 157 | else: 158 | final_cls_logits[active_sketches] = cls_logits[active_sketches].clone().detach() 159 | 160 | cls_loss = self.criterion(cls_logits[active_sketches], label[active_sketches]) 161 | cls_rank = (cls_logits.argsort(1, descending=True) == label.unsqueeze(1)).nonzero()[:, 1] 162 | elif self.loss_type == 'none': 163 | cls_loss = torch.tensor([0.]) 164 | else: 165 | raise NotImplementedError 166 | 167 | losses.append(cls_loss) 168 | 169 | order_all += 1 170 | order_all = order_all.masked_fill(next_order.bool() & active_sketches.unsqueeze(1), 0.) 171 | 172 | order_logits = order_logits.masked_fill(next_order.bool(), float('-inf')) 173 | 174 | loss = torch.cat(losses) 175 | assert not loss.isnan().any() 176 | 177 | if self.loss_type == 'class': 178 | _, pred_main = torch.max(final_cls_logits, 1) 179 | acc = (label == pred_main).float() 180 | 181 | all_cls_logits = torch.stack(all_cls_logits, dim=1) 182 | _, all_pred_main = torch.max(all_cls_logits, 2) 183 | all_acc = (label.unsqueeze(1) == all_pred_main).float() 184 | clspred_main_prob = torch.softmax(all_cls_logits, dim=2).gather(2, all_pred_main.unsqueeze(2)).squeeze(2) 185 | 186 | x_raw_reordered = [] 187 | scale_reordered = [] 188 | translate_reordered = [] 189 | order_ids_reordered = [] 190 | n_s = n_strokes.cpu().numpy().tolist() 191 | xr = x_raw.split(n_s) 192 | sc = scale.split(n_s) 193 | tr = translate.split(n_s) 194 | pcoords_n_strokes = budget_n_strokes 195 | 196 | for ol, x, s, t, n_s in zip(order_all, xr, sc, tr, pcoords_n_strokes): 197 | assert n_s == (~ol.isinf()).sum() 198 | order = ol.argsort(dim=0, descending=True) 199 | order = order[:n_s] 200 | x_raw_reordered.append(x[order]) 201 | scale_reordered.append(s[order]) 202 | translate_reordered.append(t[order]) 203 | x_raw_reordered = torch.cat(x_raw_reordered) 204 | scale_reordered = torch.cat(scale_reordered) 205 | translate_reordered = torch.cat(translate_reordered) 206 | 207 | if self.loss_type == 'class': 208 | output = {'loss': loss, 209 | 'acc_main': acc, 210 | 'acc_main_all': all_acc, 211 | 'cls_pred_main': all_pred_main, 212 | 'cls_pred_main_prob': clspred_main_prob, 213 | 'pcoords_raw': x_raw_reordered, 214 | 'pcoords_n_strokes': pcoords_n_strokes, 215 | 'pcoords_scale': scale_reordered, 216 | 'pcoords_translate': translate_reordered, 217 | 'pcoords_seq_mask': None, 218 | 'segments_per_sketch': segments_per_sketch, 219 | } 220 | elif self.loss_type == 'fgsbir': 221 | output = {'loss': loss, 222 | 'pcoords_raw': x_raw_reordered, 223 | 'pcoords_n_strokes': pcoords_n_strokes, 224 | 'pcoords_scale': scale_reordered, 225 | 'pcoords_translate': translate_reordered, 226 | 'pcoords_seq_mask': None, 227 | 'segments_per_sketch': segments_per_sketch, 228 | 'sketch_feature': sketch_feature.detach(), 229 | 'image_feature': image_feature.detach(), 230 | } 231 | else: 232 | output = {'loss': loss, 233 | 'pcoords_raw': x_raw_reordered, 234 | 'pcoords_n_strokes': pcoords_n_strokes, 235 | 'pcoords_scale': scale_reordered, 236 | 'pcoords_translate': translate_reordered, 237 | 'pcoords_seq_mask': None, 238 | 'segments_per_sketch': segments_per_sketch, 239 | 'stroke_ids': order_ids_reordered, 240 | } 241 | 242 | return output 243 | 244 | def draw(self, **inputs): 245 | return self.forward(**inputs) 246 | -------------------------------------------------------------------------------- /models/fgsbir/model.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from ..transformer import StrokeEncoderTransformer 3 | from .networks import VGG_Network, InceptionV3_Network, Resnet50_Network 4 | from torch import optim 5 | import torch 6 | import numpy as np 7 | import time 8 | import torch.nn.functional as F 9 | 10 | 11 | class FGSBIR_Model(nn.Module): 12 | def __init__(self, backbone, Nz=256, loss_type='cross_entropy', ce_temp=0.2): 13 | super(FGSBIR_Model, self).__init__() 14 | self.sample_embedding_network = eval(backbone + '_Network(scale=Nz//2)') 15 | 16 | self.loss_type = loss_type 17 | 18 | if self.loss_type == 'cross_entropy': 19 | temperature = torch.tensor([ce_temp]) 20 | self.register_buffer('temperature', temperature) 21 | self.criterion = nn.CrossEntropyLoss(reduction='none') 22 | else: 23 | self.loss = nn.TripletMarginLoss(margin=0.2) 24 | 25 | def forward(self, x_raw, n_strokes, scale, translate, seq_mask, 26 | positive_img, negative_img, sketch_img, positive_name, 27 | mask=None, **kwargs): 28 | 29 | positive_feature = self.sample_embedding_network(positive_img) 30 | sketch_feature = self.sample_embedding_network(sketch_img, finetune=True) 31 | 32 | if self.loss_type == 'cross_entropy': 33 | labels = torch.arange(sketch_feature.shape[0]).to(sketch_feature.device) 34 | 35 | positive_name = np.array(positive_name) 36 | mask = [pn == positive_name for pn in positive_name] 37 | mask = np.stack(mask) 38 | np.fill_diagonal(mask, False) 39 | mask_bool = torch.tensor(mask) 40 | mask = torch.zeros_like(mask_bool, dtype=torch.float) 41 | mask = mask.masked_fill(mask_bool, -np.inf).to(sketch_feature.device) 42 | 43 | similarity_matrix = torch.matmul(sketch_feature, positive_feature.T) / self.temperature 44 | loss = (self.criterion(similarity_matrix+mask, labels) + self.criterion(similarity_matrix.T+mask, labels)) / 2. 45 | else: 46 | negative_feature = self.sample_embedding_network(negative_img) 47 | loss = self.loss(sketch_feature, positive_feature, negative_feature) 48 | 49 | output = {'loss': loss, 50 | 'pcoords_raw': x_raw, 51 | 'pcoords_n_strokes': n_strokes, 52 | 'pcoords_scale': scale, 53 | 'pcoords_translate': translate, 54 | 'pcoords_seq_mask': seq_mask, 55 | 'sketch_feature': sketch_feature.detach(), 56 | 'image_feature': positive_feature.detach()} 57 | 58 | return output 59 | -------------------------------------------------------------------------------- /models/fgsbir/networks.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torchvision.models as backbone_ 3 | import torch.nn.functional as F 4 | import torch 5 | 6 | 7 | 8 | class Resnet50_Network(nn.Module): 9 | def __init__(self): 10 | super(Resnet50_Network, self).__init__() 11 | backbone = backbone_.resnet50(weights=backbone_.ResNet50_Weights.DEFAULT) #resnet50, resnet18, resnet34 12 | 13 | self.features = nn.Sequential() 14 | for name, module in backbone.named_children(): 15 | if name not in ['avgpool', 'fc']: 16 | self.features.add_module(name, module) 17 | self.pool_method = nn.AdaptiveMaxPool2d(1) 18 | 19 | 20 | def forward(self, input, bb_box = None): 21 | x = self.features(input) 22 | x = self.pool_method(x) 23 | x = torch.flatten(x, 1) 24 | return F.normalize(x) 25 | 26 | 27 | class VGG_Network(nn.Module): 28 | def __init__(self): 29 | super(VGG_Network, self).__init__() 30 | self.backbone = backbone_.vgg16(weights=backbone_.VGG16_Weights.DEFAULT).features 31 | self.pool_method = nn.AdaptiveMaxPool2d(1) 32 | 33 | def forward(self, input, bb_box = None): 34 | x = self.backbone(input) 35 | x = self.pool_method(x).view(-1, 512) 36 | return F.normalize(x) 37 | 38 | class InceptionV3_Network(nn.Module): 39 | def __init__(self, scale=128, use_att=False): 40 | super(InceptionV3_Network, self).__init__() 41 | self.use_att = use_att 42 | self.scale = scale 43 | backbone = backbone_.inception_v3(weights=backbone_.Inception_V3_Weights.DEFAULT) 44 | 45 | ## Extract Inception Layers ## 46 | self.Conv2d_1a_3x3 = backbone.Conv2d_1a_3x3 47 | self.Conv2d_2a_3x3 = backbone.Conv2d_2a_3x3 48 | self.Conv2d_2b_3x3 = backbone.Conv2d_2b_3x3 49 | self.Conv2d_3b_1x1 = backbone.Conv2d_3b_1x1 50 | self.Conv2d_4a_3x3 = backbone.Conv2d_4a_3x3 51 | self.Mixed_5b = backbone.Mixed_5b 52 | self.Mixed_5c = backbone.Mixed_5c 53 | self.Mixed_5d = backbone.Mixed_5d 54 | self.Mixed_6a = backbone.Mixed_6a 55 | self.Mixed_6b = backbone.Mixed_6b 56 | self.Mixed_6c = backbone.Mixed_6c 57 | self.Mixed_6d = backbone.Mixed_6d 58 | self.Mixed_6e = backbone.Mixed_6e 59 | 60 | self.Mixed_7a = backbone.Mixed_7a 61 | self.Mixed_7b = backbone.Mixed_7b 62 | self.Mixed_7c = backbone.Mixed_7c 63 | #self.pool_method = nn.AdaptiveMaxPool2d(1) # as default 64 | self.pool_method = nn.AdaptiveAvgPool2d(1) # as default 65 | 66 | if self.use_att: 67 | self.att_conv = nn.Conv2d(2048, 2048, kernel_size=1) 68 | 69 | if self.scale is not None: 70 | self.scale_fn = nn.Linear(2048, self.scale) 71 | 72 | self.finetuned_scale_fn = None 73 | 74 | def forward(self, x, finetune=False): 75 | # N x 3 x 299 x 299 76 | x = self.Conv2d_1a_3x3(x) 77 | # N x 32 x 149 x 149 78 | x = self.Conv2d_2a_3x3(x) 79 | # N x 32 x 147 x 147 80 | x = self.Conv2d_2b_3x3(x) 81 | # N x 64 x 147 x 147 82 | x = F.max_pool2d(x, kernel_size=3, stride=2) 83 | # N x 64 x 73 x 73 84 | x = self.Conv2d_3b_1x1(x) 85 | # N x 80 x 73 x 73 86 | x = self.Conv2d_4a_3x3(x) 87 | # N x 192 x 71 x 71 88 | x = F.max_pool2d(x, kernel_size=3, stride=2) 89 | # N x 192 x 35 x 35 90 | x = self.Mixed_5b(x) 91 | # N x 256 x 35 x 35 92 | x = self.Mixed_5c(x) 93 | # N x 288 x 35 x 35 94 | x = self.Mixed_5d(x) 95 | # N x 288 x 35 x 35 96 | x = self.Mixed_6a(x) 97 | # N x 768 x 17 x 17 98 | x = self.Mixed_6b(x) 99 | # N x 768 x 17 x 17 100 | x = self.Mixed_6c(x) 101 | # N x 768 x 17 x 17 102 | x = self.Mixed_6d(x) 103 | # N x 768 x 17 x 17 104 | x = self.Mixed_6e(x) 105 | # N x 768 x 17 x 17 106 | x = self.Mixed_7a(x) 107 | # N x 1280 x 8 x 8 108 | x = self.Mixed_7b(x) 109 | # N x 2048 x 8 x 8 110 | backbone_tensor = self.Mixed_7c(x) 111 | if self.use_att: 112 | att = self.att_conv(backbone_tensor) 113 | backbone_tensor = backbone_tensor + backbone_tensor * torch.softmax(att, dim=1) 114 | if finetune and self.finetuned_scale_fn is not None: 115 | feature = self.finetuned_scale_fn(backbone_tensor) 116 | else: 117 | feature = self.pool_method(backbone_tensor).view(-1, 2048) 118 | if self.scale is not None: 119 | feature = self.scale_fn(feature) 120 | 121 | return F.normalize(feature) 122 | -------------------------------------------------------------------------------- /models/pmn.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torch.distributions import Categorical 5 | import numpy as np 6 | 7 | from .transformer import StrokeEncoderTransformer 8 | 9 | from utils.coords_utils import resample_stroke, normalize_strokes 10 | from utils.distance_transform import compute_distance_transform_l2, make_coordinate_grid 11 | from utils.transformations import ToTransformMatrix, get_predefined_primitives, transform_primitives 12 | 13 | 14 | class PrimitiveMatchingNetwork(nn.Module): 15 | def __init__(self, Nz=256, 16 | temperature=0.2, 17 | primitive_names=['square', 'triangle', 'circle', 'line', 'halfcircle', 'u', 'corner', 'dot'], 18 | n_stroke_points=25, 19 | dt_gamma=6., use_pos_embed=False, 20 | use_sinusoid_embed=False, 21 | transform_bound=None, use_proportion_scale=False, 22 | learn_compatibility=False, transformations='rsr'): 23 | super().__init__() 24 | self.Nz = Nz 25 | self.encoder = StrokeEncoderTransformer( 26 | Nz=Nz, embed_sketch=False, n_layer_strokes=6, 27 | use_pos_embed=use_pos_embed, use_sinusoid_embed=use_sinusoid_embed 28 | ) 29 | 30 | #ALL_TRANSFORMATIONS = ['translate', 'rotate', 'scale_full_bounded', 'scale_full', 'uni_scale', 'proportion_scale', 'shear_x', 'shear_y'][::-1] 31 | 32 | transforms = [] 33 | self.parameters_per_transform = [] 34 | for ts in transformations: 35 | if ts == 'r': 36 | transforms.append('rotate') 37 | self.parameters_per_transform.append(1) 38 | elif ts == 's': 39 | if use_proportion_scale: 40 | transforms.append('proportion_scale') 41 | self.parameters_per_transform.append(1) 42 | else: 43 | transforms.append('scale_full') 44 | self.parameters_per_transform.append(2) 45 | elif ts == 'h': 46 | transforms.append('shear_x') 47 | transforms.append('shear_y') 48 | self.parameters_per_transform.append(1) 49 | self.parameters_per_transform.append(1) 50 | 51 | self.transformations = transforms 52 | self.transformation_layers = nn.ModuleList([ToTransformMatrix(t, bound=transform_bound) for t in self.transformations]) 53 | 54 | self.learn_compatibility = learn_compatibility 55 | self.n_stroke_points = n_stroke_points 56 | 57 | self.init_primitives(primitive_names) 58 | 59 | primnet_in = 2 * Nz 60 | if self.learn_compatibility: 61 | primnet_in = primnet_in // 2 62 | self.primnet = nn.Sequential( 63 | nn.Linear(primnet_in, primnet_in//2), 64 | nn.BatchNorm1d(primnet_in//2), 65 | nn.ReLU(inplace=True), 66 | nn.Linear(primnet_in//2, primnet_in//4), 67 | nn.BatchNorm1d(primnet_in//4), 68 | nn.ReLU(inplace=True), 69 | nn.Linear(primnet_in//4, 70 | sum(self.parameters_per_transform)) 71 | ) 72 | 73 | temperature = torch.tensor([temperature]) 74 | self.register_buffer('temperature', temperature) 75 | 76 | self.dt_gamma = dt_gamma 77 | self.criterion = nn.CrossEntropyLoss(reduction='none') 78 | self.params_criterion = nn.MSELoss(reduction='none') 79 | 80 | self.init_distance_transform() 81 | 82 | self.cache_labels_feats = None 83 | self.cache_tgt_labels = None 84 | self.cache_tgt_masks = None 85 | 86 | def init_primitives(self, primitive_names): 87 | self.primitive_names = primitive_names 88 | primitives = get_predefined_primitives( 89 | primitive_names, n_points=self.n_stroke_points 90 | ) 91 | self.register_buffer('primitives', primitives) 92 | self.num_primitives = len(primitive_names) 93 | 94 | prim_is_dot = torch.tensor([torch.unique(pt).nelement() == 1 for pt in self.primitives]) 95 | self.register_buffer('prim_is_dot', prim_is_dot.unsqueeze(0).unsqueeze(-1)) 96 | self.prim_has_dot = self.prim_is_dot.any() 97 | if self.prim_has_dot: 98 | self.prim_dot_id = self.prim_is_dot[0, :, 0].nonzero(as_tuple=False).item() 99 | else: 100 | self.prim_dot_id = None 101 | 102 | primitives_seq_mask = torch.zeros(self.primitives.shape[:2], dtype=torch.bool) 103 | self.register_buffer('primitives_seq_mask', primitives_seq_mask) 104 | primitives_position = torch.linspace(-1., 1., steps=self.primitives.shape[1]).unsqueeze(0).repeat(self.primitives.shape[0], 1) 105 | self.register_buffer('primitives_position', primitives_position) 106 | 107 | def init_distance_transform(self): 108 | self.grid_size = 32 109 | coordinate_grid = make_coordinate_grid(self.grid_size) 110 | grid_sqz = coordinate_grid.repeat(1, self.n_stroke_points-1, 1, 1, 1) 111 | self.register_buffer('grid_sqz', grid_sqz) 112 | 113 | def calc_min_l2_dist(self, recon, target, resample=False, roll=False, invert=False): 114 | l2_dist = torch.linalg.norm(recon - target, dim=-1).mean(-1) 115 | min_l2_dist = l2_dist 116 | 117 | if resample: 118 | recon = torch.stack([resample_stroke(s.cpu(), n=self.n_stroke_points)[..., :2].to(s) for s in recon]) 119 | l2_dist_res = torch.linalg.norm(recon - target, dim=-1).mean(-1) 120 | #min_l2_dist = torch.stack([min_l2_dist, l2_dist_res], dim=-1).min(-1)[0] 121 | min_l2_dist = l2_dist_res 122 | 123 | if invert: 124 | l2_dist_inv = torch.linalg.norm(recon.flip(dims=(1,)) - target, dim=-1).mean(-1) 125 | min_l2_dist = torch.stack([min_l2_dist, l2_dist_inv], dim=-1).min(-1)[0] 126 | 127 | if roll: 128 | best_l2_dist_roll = None 129 | for i in range(1, recon.shape[1]): 130 | recon_roll = recon.roll(i, dims=(1,)) 131 | l2_dist_roll = torch.linalg.norm(recon_roll - target, dim=-1).mean(-1) 132 | if best_l2_dist_roll is None: 133 | best_l2_dist_roll = l2_dist_roll 134 | else: 135 | best_l2_dist_roll = torch.stack([best_l2_dist_roll, l2_dist_roll], dim=-1).min(-1)[0] 136 | if invert: 137 | l2_dist_roll_inv = torch.linalg.norm(recon_roll.flip(dims=(1,)) - target, dim=-1).mean(-1) 138 | best_l2_dist_roll = torch.stack([best_l2_dist_roll, l2_dist_roll_inv], dim=-1).min(-1)[0] 139 | 140 | min_l2_dist = torch.stack([min_l2_dist, best_l2_dist_roll], dim=-1).min(-1)[0] 141 | 142 | out = {'l2_dist_min': min_l2_dist, 'l2_dist_base': l2_dist} 143 | if resample: 144 | out['l2_dist_resample'] = l2_dist_res 145 | 146 | if invert: 147 | out['l2_dist_invert'] = l2_dist_inv 148 | 149 | if roll: 150 | out['l2_dist_roll'] = best_l2_dist_roll 151 | 152 | return out 153 | 154 | def forward(self, x_raw, x_raw_aug=None, 155 | n_strokes=None, scale=None, translate=None, 156 | seq_mask=None, position=None, n_tgt=None, main_mask=None, 157 | **kwargs): 158 | 159 | batch = x_raw 160 | target = x_raw_aug 161 | 162 | n_tgt = max(batch.shape[0], self.primitives.shape[0]) 163 | if self.primitives.shape[0] > batch.shape[0]: 164 | print('WARNING: More primitives than batch strokes. Untested scenario.') 165 | target = self.primitives 166 | tgt_seq_mask = self.primitives_seq_mask 167 | tgt_position = self.primitives_position 168 | 169 | # encode: 170 | z, z_params = self.encoder(batch.transpose(0, 1), seq_mask=seq_mask, position=position)[0].split([self.Nz//2, self.Nz//2], dim=-1) 171 | zt, zt_params = self.encoder(target.transpose(0, 1), seq_mask=tgt_seq_mask, position=tgt_position)[0].split([self.Nz//2, self.Nz//2], dim=-1) 172 | 173 | if self.learn_compatibility: 174 | output = self.info_nce_loss(z, zt, n_tgt, main_mask, ignore_loss=((not self.training) or self.learn_compatibility)) 175 | if self.learn_compatibility: 176 | prim_logits = output['pred_main_logits'] 177 | 178 | output = {} 179 | output['loss'] = 0. 180 | 181 | if not self.learn_compatibility: 182 | z_params = torch.cat([z, z_params], dim=-1) 183 | zt_params = torch.cat([zt, zt_params], dim=-1) 184 | 185 | z_params = z_params.unsqueeze(1) 186 | zt_params = zt_params.unsqueeze(0) 187 | z_params = z_params.expand(z_params.shape[0], zt_params.shape[1], z_params.shape[2]) 188 | zt_params = zt_params.expand(z_params.shape[0], zt_params.shape[1], zt_params.shape[2]) 189 | z_params = z_params.reshape(-1, z_params.shape[-1]) 190 | zt_params = zt_params.reshape(-1, zt_params.shape[-1]) 191 | 192 | primnet_in = torch.cat([z_params, zt_params], dim=-1) 193 | 194 | prims_params = self.primnet(primnet_in) 195 | 196 | prims_params = prims_params.split(self.parameters_per_transform, dim=-1) 197 | prim_params = [] 198 | for p, s, t in zip(prims_params, self.parameters_per_transform, self.transformations): 199 | pt = p.view(-1, self.num_primitives, s) 200 | if t == 'proportion_scale': 201 | pt = torch.cat([pt, -pt], dim=-1) 202 | if self.prim_has_dot: 203 | pt = pt.masked_fill(self.prim_is_dot, 0.) 204 | prim_params.append(pt) 205 | 206 | prims = transform_primitives( 207 | target, self.transformation_layers, prim_params 208 | #target, self.transformation_layers_noop, parameters_aug.split([2, 2], dim=-1) 209 | ) 210 | 211 | prims = prims.squeeze(-3).squeeze(-2)[..., :2] 212 | prims = normalize_strokes(prims)[0] 213 | 214 | if self.training or not self.learn_compatibility: 215 | dt_loss = 10. * compute_distance_transform_l2(batch, prims, self.grid_sqz, batch_mask=seq_mask, gamma=self.dt_gamma) 216 | else: 217 | dt_loss = 0. 218 | if self.learn_compatibility: 219 | if self.training: 220 | if self.prim_has_dot: 221 | loss_mask = dt_loss[:, self.prim_dot_id].isclose(torch.tensor([0.], device=dt_loss.device)).unsqueeze(1) 222 | dt_loss = dt_loss.masked_fill(loss_mask, 0.) 223 | 224 | dt_loss = (dt_loss * torch.softmax(prim_logits, dim=1)).sum(1) 225 | prim_ids = prim_logits.argmax(dim=1, keepdim=True).unsqueeze(-1).unsqueeze(-1) 226 | if self.prim_has_dot: 227 | if self.training: 228 | prim_ids[loss_mask] = self.prim_dot_id 229 | else: 230 | dot_mask = batch.isclose(torch.tensor([0.], device=batch.device)).sum([-2, -1]) == (batch.shape[-2]*batch.shape[-1]) 231 | prim_ids[dot_mask] = self.prim_dot_id 232 | 233 | else: 234 | prim_ids = dt_loss.argmin(dim=1, keepdim=True).unsqueeze(-1).unsqueeze(-1) 235 | if self.prim_has_dot: 236 | loss_mask = dt_loss[:, self.prim_dot_id].isclose(torch.tensor([0.], device=dt_loss.device)).unsqueeze(1) 237 | dt_loss = dt_loss.masked_fill(loss_mask, 0.) 238 | dt_loss = dt_loss.mean(1) 239 | 240 | prims = prims.gather(1, prim_ids.expand(prims.size(0), 1, prims.size(2), prims.size(3))) 241 | prims = prims.squeeze(1) 242 | 243 | output['pcoords_prim_id'] = prim_ids.squeeze(-1) 244 | output['loss'] += dt_loss 245 | 246 | output['pcoords_raw'] = prims 247 | output['pcoords_n_strokes'] = n_strokes 248 | output['pcoords_scale'] = scale 249 | output['pcoords_translate'] = translate 250 | output['pcoords_seq_mask'] = self.primitives_seq_mask[:1].expand(prims.shape[:2]) 251 | output['pcoords_position'] = self.primitives_position[:1].expand(prims.shape[:2]) 252 | 253 | return output 254 | 255 | def info_nce_loss(self, features, targets, n_tgt=None, main_mask=None, 256 | ignore_loss=False): 257 | 258 | if n_tgt is None: 259 | n_tgt = targets.shape[0] 260 | labels = torch.arange(n_tgt).to(features.device) 261 | 262 | features = F.normalize(features, dim=-1) 263 | targets = F.normalize(targets, dim=-1) 264 | 265 | similarity_matrix = torch.matmul(features, targets.T) / self.temperature 266 | main_simmat = similarity_matrix[:n_tgt] 267 | if main_mask is not None: 268 | main_simmat += main_mask 269 | mod_simmat = similarity_matrix.T[:n_tgt] 270 | if not ignore_loss: 271 | loss_main = self.criterion(main_simmat, labels) 272 | loss_mod = self.criterion(mod_simmat, labels) 273 | total_loss = (loss_main + loss_mod) / 2 274 | else: 275 | total_loss = 0. 276 | 277 | labels_mod = labels[:mod_simmat.shape[0]] 278 | 279 | _, pred_main = torch.max(main_simmat, 1) 280 | acc_main = (labels[:pred_main.shape[0]] == pred_main).float() 281 | _, pred_mod = torch.max(mod_simmat, 1) 282 | acc_mod = (labels_mod[:pred_mod.shape[0]] == pred_mod).float() 283 | 284 | output = {'loss': total_loss, 285 | 'acc_main': acc_main, 286 | 'acc_mod': acc_mod, 287 | 'pred_main': pred_main, 288 | 'pred_mod': pred_mod, 289 | 'pred_main_logits': main_simmat} 290 | 291 | return output 292 | 293 | def l2_eval(self, batch): 294 | prims = self.draw(batch) 295 | l2_dist = self.calc_min_l2_dist(prims, batch, resample=True, roll=True, invert=True) 296 | dist_transform = compute_distance_transform_l2(prims, batch, self.grid_sqz, gamma=self.dt_gamma) 297 | return l2_dist['l2_dist_min'], dist_transform 298 | 299 | def draw(self, **inputs): 300 | return self.forward(**inputs) 301 | 302 | def compute_l2(self,prims,batch): 303 | return self.calc_min_l2_dist(prims, batch, resample=True, roll=True, invert=True) 304 | -------------------------------------------------------------------------------- /models/transformer.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import numpy as np 5 | 6 | 7 | class StrokeEncoderTransformer(nn.Module): 8 | def __init__(self, Nz=128, enc_hidden_size=128, dropout=0.1, stroke_len=25, 9 | embed_sketch=True, n_layer_strokes=3, n_layer_sketch=3, 10 | use_pos_embed=True, use_sinusoid_embed=False, fourier_scale=1., 11 | out_dim=None, use_embed_token=True, return_points_embed=False): 12 | super().__init__() 13 | if out_dim is None: 14 | out_dim = Nz 15 | self.enc_hidden_size = enc_hidden_size 16 | self.embed_sketch = embed_sketch 17 | self.stroke_len = stroke_len 18 | self.use_pos_embed = use_pos_embed 19 | self.sinusoid_embed = use_sinusoid_embed 20 | self.use_embed_token = use_embed_token 21 | self.return_points_embed = return_points_embed 22 | # Transformer: 23 | encoder_layer = nn.TransformerEncoderLayer( 24 | enc_hidden_size, 8, dim_feedforward=4*enc_hidden_size, 25 | dropout=dropout, activation='gelu' 26 | ) 27 | self.stroke_transformer = nn.TransformerEncoder( 28 | encoder_layer, num_layers=n_layer_strokes 29 | ) 30 | self.fc_in = nn.Linear(2, enc_hidden_size) 31 | self.fc_out = nn.Linear(enc_hidden_size, out_dim) 32 | if self.use_pos_embed and not self.sinusoid_embed: 33 | self.pos_embed = nn.Parameter(torch.randn(stroke_len+1, 1, enc_hidden_size)) 34 | else: 35 | self.embed_token = nn.Parameter(torch.randn(1, 1, enc_hidden_size)) 36 | if self.use_pos_embed: 37 | self.fc_pos_embed = nn.Linear(enc_hidden_size//(2*3)*2, enc_hidden_size) 38 | 39 | if self.sinusoid_embed: 40 | randn = fourier_scale * torch.randn((1, enc_hidden_size//(2*3))) 41 | #randn = fourier_scale * torch.randn((3, enc_hidden_size//2)) 42 | self.register_buffer('fourier_randn', randn) 43 | 44 | if self.embed_sketch: 45 | if self.sinusoid_embed: 46 | self.fc_transform = nn.Linear((enc_hidden_size//(2*3))*2*3, enc_hidden_size) 47 | #self.fc_transform = nn.Linear(enc_hidden_size, enc_hidden_size) 48 | else: 49 | self.fc_transform = nn.Linear(3, enc_hidden_size) 50 | self.sketch_transformer = nn.TransformerEncoder(encoder_layer, 51 | num_layers=n_layer_sketch) 52 | self.sketch_embed = nn.Parameter(torch.randn(1, enc_hidden_size)) 53 | # active dropout: 54 | self.train() 55 | 56 | def pad_stroke_sequence(self, inputs, stroke_lens): 57 | max_stroke_len = max(stroke_lens) 58 | offset = 0 59 | strokes = [] 60 | mask = torch.ones((len(stroke_lens), max_stroke_len+1), device=inputs.device, dtype=torch.bool) 61 | mask[:, -1] = False 62 | if self.use_pos_embed: 63 | positions = torch.zeros(max_stroke_len, (len(stroke_lens)), device=inputs.device) 64 | else: 65 | positions = None 66 | for i, sl in enumerate(stroke_lens): 67 | strokes.append(torch.cat([inputs[offset:offset+sl], 68 | torch.zeros((max_stroke_len-sl, 2), device=inputs.device)])) 69 | mask[i, :sl] = False 70 | if self.use_pos_embed: 71 | positions[:sl, i] = torch.linspace(-1, 1., steps=sl, device=inputs.device) 72 | offset += sl 73 | return torch.stack(strokes, dim=1), mask, max_stroke_len, positions 74 | 75 | def forward(self, inputs, n_strokes=None, scale=None, translate=None, stroke_lens=None, seq_mask=None, position=None, mask=None, points_mask=None, **kwargs): 76 | if inputs.dim() == 2: 77 | inputs, seq_mask, slen, pos = self.pad_stroke_sequence(inputs.transpose(0, 1), stroke_lens) 78 | if points_mask is None: 79 | points_mask = seq_mask 80 | else: 81 | points_mask = points_mask | seq_mask 82 | elif seq_mask is not None: 83 | slen = inputs.shape[0] 84 | pos = position.transpose(0, 1) 85 | seq_mask = torch.cat([seq_mask, torch.zeros_like(seq_mask[:, :1])], dim=1) 86 | if points_mask is None: 87 | points_mask = seq_mask 88 | else: 89 | points_mask = points_mask | seq_mask 90 | else: 91 | slen = self.stroke_len 92 | embed = self.fc_in(inputs) 93 | if self.use_pos_embed: 94 | if self.sinusoid_embed: 95 | pos_embed = (2.*np.pi*pos.unsqueeze(-1)) @ self.fourier_randn 96 | pos_embed = torch.cat([torch.sin(pos_embed), torch.cos(pos_embed)], dim=-1) 97 | pos_embed = self.fc_pos_embed(pos_embed) 98 | pos_embed = torch.cat([pos_embed, self.embed_token.expand(1, pos_embed.shape[1], pos_embed.shape[2])]) 99 | else: 100 | pos_embed = self.pos_embed 101 | else: 102 | pos_embed = torch.cat([torch.zeros((slen, 1, self.enc_hidden_size), device=embed.device), self.embed_token]) 103 | embed = torch.cat([embed, torch.zeros((1, embed.size(1), embed.size(2)), device=embed.device)]) + pos_embed 104 | embed = self.stroke_transformer(embed, src_key_padding_mask=points_mask) 105 | if self.return_points_embed: 106 | points_embed = embed[:-1] 107 | embed = embed[-1] 108 | 109 | if self.embed_sketch: 110 | transform = torch.cat([scale, translate], dim=-1).squeeze(1) 111 | if self.sinusoid_embed: 112 | transform = (2.*np.pi*transform.unsqueeze(-1)) @ self.fourier_randn 113 | #transform = (2.*np.pi*transform) @ self.fourier_randn 114 | transform = torch.cat([torch.sin(transform), torch.cos(transform)], dim=-1) 115 | transform = transform.flatten(-2) 116 | transform_embed = self.fc_transform(transform) 117 | embed = embed + transform_embed 118 | 119 | embed_batch = torch.split(embed, n_strokes.tolist()) 120 | seq_len = n_strokes.max() 121 | batch = [] 122 | padding_mask = [] 123 | for eb in embed_batch: 124 | pad_len = seq_len-eb.size(0) 125 | padding = torch.zeros((pad_len, eb.size(1)), device=eb.device) 126 | embd = torch.cat([eb, padding]) 127 | if self.use_embed_token: 128 | embd = torch.cat([embd, self.sketch_embed]) 129 | batch.append(embd) 130 | if mask is None: 131 | pad_mask = torch.zeros(seq_len+(1 if self.use_embed_token else 0), 132 | device=eb.device, dtype=torch.bool) 133 | if self.use_embed_token: 134 | pad_mask[-pad_len-1:-1] = True 135 | elif pad_len != 0: 136 | pad_mask[-pad_len:] = True 137 | padding_mask.append(pad_mask) 138 | 139 | batch = torch.stack(batch, dim=1) 140 | if mask is None: 141 | padding_mask = torch.stack(padding_mask, dim=0) 142 | else: 143 | padding_mask = mask 144 | embed = self.sketch_transformer(batch, src_key_padding_mask=padding_mask) 145 | if self.use_embed_token: 146 | embed = embed[-1] 147 | else: 148 | padding_mask = None 149 | 150 | if self.return_points_embed: 151 | return self.fc_out(embed), points_embed, padding_mask 152 | else: 153 | return self.fc_out(embed), padding_mask 154 | -------------------------------------------------------------------------------- /save_processed_images.py: -------------------------------------------------------------------------------- 1 | import datetime 2 | import os 3 | from itertools import islice 4 | 5 | import torch 6 | from tqdm import tqdm 7 | 8 | from models import get_model 9 | from models.communication import calc_segments_per_sketch 10 | from utils.arguments import parser 11 | from utils.data import get_data 12 | 13 | 14 | def split_by_lengths(seq, num): 15 | it = iter(seq) 16 | for x in num: 17 | out = list(islice(it, x)) 18 | if out: 19 | yield out 20 | else: 21 | return # StopIteration 22 | remain = list(it) 23 | if remain: 24 | yield remain 25 | 26 | 27 | if __name__ == "__main__": 28 | args = parser.parse_args() 29 | 30 | DEVICE = "cuda" if torch.cuda.is_available() else "cpu" 31 | 32 | # Add timestamp to logdir 33 | if args.log_name is None: 34 | LOG_DIR = os.path.join( 35 | args.log_dir, datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S") 36 | ) 37 | else: 38 | LOG_DIR = os.path.join(args.log_dir, args.log_name) 39 | 40 | dataloader_coords, n_classes, label_names = get_data( 41 | args.data_dir, 42 | args.dataset, 43 | batch_size=args.batch_size, 44 | train=True, 45 | preprocessed_root=args.preprocessed_root, 46 | load_fgsbir_photos=args.load_fgsbir_photos, 47 | resample_strokes=not args.no_stroke_resampling, 48 | max_stroke_length=args.max_stroke_length, 49 | shuffle=False, 50 | ) 51 | 52 | test_dataloader_coords, _, _ = get_data( 53 | args.data_dir, 54 | args.dataset, 55 | batch_size=args.n_log_img, 56 | train=False, 57 | preprocessed_root=args.preprocessed_root, 58 | load_fgsbir_photos=args.load_fgsbir_photos, 59 | resample_strokes=not args.no_stroke_resampling, 60 | max_stroke_length=args.max_stroke_length, 61 | shuffle=False, 62 | ) 63 | 64 | if args.model_type == "none": 65 | model = None 66 | else: 67 | model = get_model(args, 9, DEVICE) 68 | 69 | if args.test is not None and ( 70 | not args.model_type == "communication" or args.loss_model_type != "none" 71 | ): 72 | model.load_state_dict(torch.load(args.test)) 73 | model.eval() 74 | 75 | model_type = args.model_type 76 | if args.add_comm_model: 77 | args.model_type = "communication" 78 | comm_model = get_model(args, 9, DEVICE) 79 | n_iters = 2 80 | else: 81 | comm_model = None 82 | n_iters = 1 83 | 84 | loaders = [("test", test_dataloader_coords), ("train", dataloader_coords)] 85 | # Preprocess strokes 86 | with torch.no_grad(): 87 | for dset, dset_loader in loaders: 88 | new_coords_list = [] 89 | for batch in tqdm(dset_loader, desc=f"Pre-processing data {dset}"): 90 | for k in batch.keys(): 91 | if k in [ 92 | "x_raw", 93 | "x_raw_aug", 94 | "scale", 95 | "scale_aug", 96 | "translate", 97 | "translate_aug", 98 | "parameters_aug", 99 | ]: 100 | if isinstance(batch[k], list): 101 | batch[k] = batch[k][0] 102 | if k == "n_strokes_aug": 103 | batch[k] = torch.stack(batch[k], dim=1).view(-1) 104 | if isinstance(batch[k], torch.Tensor): 105 | batch[k] = batch[k].to("cuda") 106 | 107 | orig_strokes = batch["n_strokes"] 108 | for citer in range(n_iters): 109 | strokes = batch["n_strokes"] 110 | if model is None: 111 | points = list( 112 | batch["x_raw"].to("cpu").split(list(strokes), dim=0) 113 | ) 114 | scale = list( 115 | batch["scale"].to("cpu").split(list(strokes), dim=0) 116 | ) 117 | translate = list( 118 | batch["translate"].to("cpu").split(list(strokes), dim=0) 119 | ) 120 | stroke_lens = batch["stroke_lens"].split(list(strokes), dim=0) 121 | _, _, segments_per_sketch = calc_segments_per_sketch( 122 | batch["segments"], batch["stroke_lens"], batch["n_strokes"] 123 | ) 124 | segments = batch["segments"].split( 125 | list(segments_per_sketch), dim=0 126 | ) 127 | 128 | new_coords_list += [ 129 | { 130 | "x_raw": x.detach().cpu(), 131 | "label": l.item(), 132 | "n_strokes": n.detach().cpu(), 133 | "scale": s.detach().cpu(), 134 | "translate": t.detach().cpu(), 135 | "segments": se.detach().cpu(), 136 | "stroke_lens": sl.detach().cpu(), 137 | } 138 | for x, l, n, s, t, se, sl in zip( 139 | points, 140 | batch["label"], 141 | strokes, 142 | scale, 143 | translate, 144 | segments, 145 | stroke_lens, 146 | ) 147 | ] 148 | 149 | else: 150 | if comm_model is not None and citer == 1: 151 | new_coords = comm_model(**batch) 152 | else: 153 | new_coords = model.draw(**batch) 154 | 155 | if citer == 1 or model_type in ["pmn", "communication"]: 156 | if comm_model is not None and citer == 0: 157 | for k, v in new_coords.items(): 158 | if k == "pcoords_raw": 159 | batch["x_raw"] = v 160 | elif k.startswith("pcoords"): 161 | batch[k.replace("pcoords_", "")] = v 162 | batch["n_strokes_orig"] = orig_strokes 163 | else: 164 | strokes = new_coords["pcoords_n_strokes"] 165 | if isinstance(new_coords["pcoords_raw"], list): 166 | offset = 0 167 | points = [] 168 | for n_s in strokes: 169 | points.append( 170 | [ 171 | p.detach().cpu() 172 | for p in new_coords["pcoords_raw"] 173 | ][offset : offset + n_s] 174 | ) 175 | offset += n_s 176 | else: 177 | points = list( 178 | new_coords["pcoords_raw"] 179 | .detach() 180 | .to("cpu") 181 | .split(list(strokes), dim=0) 182 | ) 183 | scale = list( 184 | new_coords["pcoords_scale"] 185 | .to("cpu") 186 | .split(list(strokes), dim=0) 187 | ) 188 | translate = list( 189 | new_coords["pcoords_translate"] 190 | .to("cpu") 191 | .split(list(strokes), dim=0) 192 | ) 193 | 194 | stroke_lens = batch["stroke_lens"].split( 195 | list(orig_strokes), dim=0 196 | ) 197 | _, _, segments_per_sketch = calc_segments_per_sketch( 198 | batch["segments"], 199 | batch["stroke_lens"], 200 | batch["n_strokes"], 201 | ) 202 | segments = batch["segments"].split( 203 | list(segments_per_sketch), dim=0 204 | ) 205 | new_coords_list += [ 206 | { 207 | "x_raw": x, 208 | "label": l.item(), 209 | "n_strokes": n.detach().cpu(), 210 | "scale": s.detach().cpu(), 211 | "translate": t.detach().cpu(), 212 | "segments": se.detach().cpu(), 213 | "stroke_lens": sl.detach().cpu(), 214 | } 215 | for x, l, n, s, t, se, sl in zip( 216 | points, 217 | batch["label"], 218 | strokes, 219 | scale, 220 | translate, 221 | segments, 222 | stroke_lens, 223 | ) 224 | ] 225 | 226 | if "pcoords_prim_id" in new_coords: 227 | prim_ids = list( 228 | new_coords["pcoords_prim_id"] 229 | .to("cpu") 230 | .split(list(strokes), dim=0) 231 | ) 232 | for i, pi in enumerate(prim_ids): 233 | new_coords_list[-len(prim_ids) :][i][ 234 | "prim_id" 235 | ] = pi.detach().cpu() 236 | 237 | else: 238 | sketches = list(split_by_lengths((new_coords), strokes)) 239 | split_scale = list( 240 | batch["scale"].split(list(strokes), dim=0) 241 | ) 242 | split_translate = list( 243 | batch["translate"].split(list(strokes), dim=0) 244 | ) 245 | new_strokes = torch.Tensor([]).to("cpu") 246 | scales = torch.Tensor([]).to("cuda") 247 | translates = torch.Tensor([]).to("cuda") 248 | strokes = list(strokes) 249 | for i in range(len(sketches)): 250 | strokes[i] = 0 251 | for j in range(len(sketches[i])): 252 | scales = torch.cat( 253 | [ 254 | scales, 255 | split_scale[i][j].repeat( 256 | len(sketches[i][j]), 1, 1 257 | ), 258 | ], 259 | dim=0, 260 | ) 261 | translates = torch.cat( 262 | [ 263 | translates, 264 | split_translate[i][j].repeat( 265 | len(sketches[i][j]), 1, 1 266 | ), 267 | ], 268 | dim=0, 269 | ) 270 | new_strokes = torch.cat( 271 | [new_strokes, sketches[i][j]], dim=0 272 | ) 273 | strokes[i] += len(sketches[i][j]) 274 | 275 | if comm_model is not None: 276 | batch["x_raw"] = new_strokes.to(batch["x_raw"]) 277 | batch["n_strokes"] = torch.tensor(strokes).to( 278 | batch["n_strokes"] 279 | ) 280 | batch["scale"] = scales.to(batch["scale"]) 281 | batch["translate"] = translates.to(batch["translate"]) 282 | batch["n_strokes_orig"] = orig_strokes 283 | else: 284 | scale = list(scales.split(strokes, dim=0)) 285 | translate = list(translates.split(strokes, dim=0)) 286 | points = list(new_strokes.split(strokes, dim=0)) 287 | strokes = torch.tensor(strokes) 288 | stroke_lens = batch["stroke_lens"].split( 289 | list(batch["n_strokes"]), dim=0 290 | ) 291 | _, _, segments_per_sketch = calc_segments_per_sketch( 292 | batch["segments"], 293 | batch["stroke_lens"], 294 | batch["n_strokes"], 295 | ) 296 | segments = batch["segments"].split( 297 | list(segments_per_sketch), dim=0 298 | ) 299 | 300 | new_coords_list += [ 301 | { 302 | "x_raw": x, 303 | "label": l.item(), 304 | "n_strokes": n, 305 | "scale": s.detach().cpu(), 306 | "translate": t.detach().cpu(), 307 | "segments": se.detach().cpu(), 308 | "stroke_lens": sl.detach().cpu(), 309 | "n_strokes_orig": n_orig.detach().cpu(), 310 | } 311 | for x, l, n, s, t, se, sl, n_orig in zip( 312 | points, 313 | batch["label"], 314 | strokes, 315 | scale, 316 | translate, 317 | segments, 318 | stroke_lens, 319 | orig_strokes, 320 | ) 321 | ] 322 | 323 | dset_loader.dataset.set_preprocessed(new_coords_list) 324 | dset_loader.dataset.save_preprocessed( 325 | os.path.join(LOG_DIR, "data"), 326 | save_images=not args.only_save_coords, 327 | suffix=".npy", 328 | ) 329 | -------------------------------------------------------------------------------- /trainer.py: -------------------------------------------------------------------------------- 1 | import os 2 | from collections import defaultdict 3 | from contextlib import nullcontext 4 | 5 | import numpy as np 6 | import torch 7 | import torch.nn.functional as F 8 | import torchvision.transforms as transforms 9 | from tqdm import tqdm 10 | import matplotlib.pyplot as plt 11 | 12 | from utils.metrics import TbLogger, Accumulator, log_metrics, log_images 13 | from utils.coords_utils import PRIM_COLORS, PRIM_ORDER, make_image_fast 14 | from utils.plt_blit import BlitManager 15 | 16 | 17 | class Trainer: 18 | def __init__(self, model_type, model, train, learning_rate, epochs, 19 | dataloader, test_dataloader, label_names, device, log_dir, 20 | log_every, n_log_img, save_every, view_scale, weight_decay): 21 | self.model_type = model_type 22 | self.model = model 23 | self.train = train 24 | self.epochs = epochs if self.train else 1 25 | self.dataloader = dataloader 26 | self.test_dataloader = test_dataloader 27 | self.label_names = label_names 28 | self.device = device 29 | self.log_dir = log_dir + ('' if self.train else '_test') 30 | self.log_every = log_every 31 | self.n_log_img = n_log_img 32 | self.save_every = save_every 33 | self.view_scale = view_scale 34 | self.best_result = None 35 | 36 | self.tblogger = TbLogger(self.log_dir) 37 | if self.train: 38 | self.optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay) 39 | 40 | self.metrics = defaultdict(Accumulator) 41 | self.update_steps = 0 42 | self.blit_manager = None 43 | 44 | def run(self): 45 | for e in range(self.epochs): 46 | if not self.train: 47 | self.test_dataloader.generator.manual_seed(42) 48 | tqdm_loader = tqdm(self.dataloader if self.train else self.test_dataloader) 49 | 50 | self.update_steps = self.run_epoch(tqdm_loader, self.train, 51 | self.update_steps) 52 | 53 | if self.train: 54 | self.test_dataloader.generator.manual_seed(42) 55 | self.run_epoch(tqdm(self.test_dataloader), False, 0) 56 | 57 | self.tblogger.close() 58 | 59 | def update_metrics(self, outputs, train): 60 | prefix = 'train' if train else 'test' 61 | 62 | if isinstance(outputs['loss'],float): 63 | self.metrics[f'{prefix}/loss'].add(value=outputs['loss']) 64 | else: 65 | self.metrics[f'{prefix}/loss'].add(value=outputs['loss'].mean().item()) 66 | 67 | if 'acc_main' in outputs: 68 | self.metrics[f'{prefix}/acc'].add(value=outputs['acc_main'].mean().item()) 69 | if 'acc_mod' in outputs: 70 | self.metrics[f'{prefix}/acc_mod'].add(value=outputs['acc_mod'].mean().item()) 71 | if 'cls_acc_main' in outputs and 'cls_acc_mod' in outputs and outputs['cls_acc_main'] is not None and outputs['cls_acc_mod'] is not None: 72 | self.metrics[f'{prefix}/cls_acc'].add(value=outputs['cls_acc_main'].mean().item()) 73 | self.metrics[f'{prefix}/cls_acc_mod'].add(value=outputs['cls_acc_mod'].mean().item()) 74 | for k, v in outputs.items(): 75 | if k.startswith('l2_dist'): 76 | self.metrics[f'{prefix}/{k}'].add(value=v.mean().item()) 77 | 78 | def run_epoch(self, loader, train, steps): 79 | if self.model_type == 'communication': 80 | budget_acc = torch.zeros(20) 81 | segment_acc = [] 82 | total_cnt = 0 83 | if self.model_type == 'fgsbir' or (self.model_type == 'communication' and self.model.loss_type == 'fgsbir'): 84 | sketch_names = [] 85 | sketch_features = [] 86 | image_names = [] 87 | images_features = [] 88 | 89 | for idx, batch in enumerate(loader): 90 | for k in batch.keys(): 91 | if k in ['x_raw', 'x_raw_aug', 'scale', 'scale_aug', 92 | 'translate', 'translate_aug', 'parameters_aug']: 93 | if isinstance(batch[k], list): 94 | batch[k] = batch[k][0] 95 | if k == 'n_strokes_aug': 96 | batch[k] = torch.stack(batch[k], dim=1).view(-1) 97 | if isinstance(batch[k], torch.Tensor): 98 | batch[k] = batch[k].to(self.device) 99 | 100 | if train: 101 | self.optimizer.zero_grad() 102 | ctx = nullcontext() 103 | else: 104 | self.model.eval() 105 | ctx = torch.no_grad() 106 | 107 | with ctx: 108 | outputs = self.run_model(batch, train) 109 | 110 | if train: 111 | outputs['loss'].mean().backward() 112 | self.optimizer.step() 113 | else: 114 | self.model.train() 115 | 116 | self.update_metrics(outputs, train=train) 117 | 118 | steps += 1 119 | 120 | if steps % 100 == 0 and train: 121 | log_metrics(self.metrics, self.tblogger, steps) 122 | self.metrics = defaultdict(Accumulator) 123 | 124 | if steps % self.log_every == 0: 125 | postfix = '' #if train else f'/{steps}' 126 | all_log_images = self.log_images(batch, outputs, train, 127 | coords_type='x_raw' if 'pcoords_raw' in outputs else 'x') 128 | 129 | for phase, log_imgs, n_img in all_log_images: 130 | img_tensor = F.interpolate(log_imgs, size=(self.view_scale, self.view_scale)) 131 | log_images(f'image_drawing_{phase}{postfix}', img_tensor, n_img, self.tblogger, steps if train else self.update_steps) 132 | 133 | if steps % self.save_every == 0: 134 | torch.save(self.model.state_dict(), os.path.join(self.log_dir, 'model_latest.pt')) 135 | 136 | if self.model_type == 'communication' and self.model.loss_type == 'class': 137 | if outputs['segments_per_sketch'] is None: 138 | seq_lens = batch['n_strokes'].detach().cpu() 139 | max_seq_lens = outputs['pcoords_n_strokes'].detach().cpu() 140 | else: 141 | seq_lens = outputs['segments_per_sketch'].detach().cpu() 142 | max_seq_lens = outputs['pcoords_n_strokes'].detach().cpu() 143 | 144 | acc_all = outputs['acc_main_all'].detach().cpu() 145 | budget = torch.arange(seq_lens.max()).unsqueeze(0).repeat(seq_lens.shape[0], 1) 146 | budget = (budget+1) / seq_lens.unsqueeze(1) 147 | for b in range(20): 148 | curr_budget = ((b+1)/20) 149 | if self.model.budget is not None and curr_budget > self.model.budget: 150 | budget_acc[b] = budget_acc[b-1] 151 | continue 152 | budget_mask = budget >= curr_budget 153 | budget_arg = budget_mask.float().argmax(dim=1) 154 | budget_arg = torch.stack([budget_arg, max_seq_lens-1]).min(0)[0] 155 | budget_acc[b] += acc_all.gather(1, budget_arg.unsqueeze(1)).sum() 156 | 157 | if not train: 158 | for aa, bud, seq in zip (acc_all, budget, seq_lens): 159 | if self.model.budget is None: 160 | sl = seq 161 | else: 162 | sl = int(np.ceil(seq * self.model.budget)) 163 | segment_acc.append(list(zip(bud[:sl].numpy(), aa[:sl].numpy()))) 164 | 165 | total_cnt += acc_all.shape[0] 166 | 167 | if self.model_type == 'fgsbir' or (self.model_type == 'communication' and self.model.loss_type == 'fgsbir'): 168 | sketch_features.extend(outputs['sketch_feature'].cpu()) 169 | sketch_names.extend(batch['positive_name']) 170 | 171 | for name, img_feat in zip(batch['positive_name'], outputs['image_feature'].cpu()): 172 | if name not in image_names: 173 | image_names.append(name) 174 | images_features.append(img_feat) 175 | 176 | if self.model_type == 'fgsbir' or (self.model_type == 'communication' and self.model.loss_type == 'fgsbir'): 177 | rank = [] 178 | images_features = torch.stack(images_features) 179 | 180 | for name, sketch_feature in zip(sketch_names, sketch_features): 181 | position_query = image_names.index(name) 182 | 183 | if self.model.loss_type == 'cross_entropy': 184 | distance = -torch.matmul(sketch_feature.unsqueeze(0), images_features.T) 185 | target_distance = -torch.matmul(sketch_feature.unsqueeze(0), images_features[position_query].unsqueeze(1)) 186 | else: 187 | distance = F.pairwise_distance(sketch_feature.unsqueeze(0), images_features) 188 | target_distance = F.pairwise_distance(sketch_feature.unsqueeze(0), 189 | images_features[position_query].unsqueeze(0)) 190 | 191 | rank.append(distance.le(target_distance).sum().item()) 192 | 193 | rank = np.array(rank) 194 | top1 = (rank <= 1).sum() / rank.shape[0] 195 | top5 = (rank <= 5).sum() / rank.shape[0] 196 | top10 = (rank <= 10).sum() / rank.shape[0] 197 | prefix = 'train' if train else 'test' 198 | self.metrics[f'{prefix}/acc'].add(value=top1) 199 | self.metrics[f'{prefix}/acc5'].add(value=top5) 200 | self.metrics[f'{prefix}/acc10'].add(value=top10) 201 | 202 | if self.model_type == 'communication' and self.model.loss_type == 'class': 203 | prefix = 'train' if train else 'test' 204 | for b, acc in enumerate(budget_acc): 205 | self.metrics[f'{prefix}/budget_acc_{(b+1)/20}'].add(value=acc.item()/total_cnt) 206 | 207 | 208 | aggregated = log_metrics(self.metrics, self.tblogger, steps if train else self.update_steps) 209 | self.metrics = defaultdict(Accumulator) 210 | if not train: 211 | is_best = False 212 | if 'test/acc' in aggregated: 213 | current_acc = aggregated['test/acc'] 214 | if self.best_result is None or self.best_result < current_acc: 215 | self.best_result = current_acc 216 | is_best = True 217 | 218 | else: 219 | current_loss = aggregated['test/loss'] 220 | if self.best_result is None or self.best_result > current_loss: 221 | self.best_result = current_loss 222 | is_best = True 223 | 224 | if is_best: 225 | torch.save(self.model.state_dict(), os.path.join(self.log_dir, 'model_best_acc.pt')) 226 | if 'test/acc' in aggregated: 227 | with open(os.path.join(self.log_dir, 'test_acc.txt'), 'w') as f: 228 | f.write(f'Acc: {self.best_result}\n') 229 | if 'test/acc5' in aggregated: 230 | f.write(f'Acc 5: {aggregated["test/acc5"]}\n') 231 | if 'test/acc10' in aggregated: 232 | f.write(f'Acc 10: {aggregated["test/acc10"]}\n') 233 | if self.model_type == 'communication' and self.model.loss_type == 'class': 234 | f.write('Budget Acc: ' + ' '.join([str(ba.item()/total_cnt) for ba in budget_acc])) 235 | 236 | if self.model_type == 'communication' and self.model.loss_type == 'class': 237 | segments = sorted(set([s for segs in segment_acc for s, _ in segs])) 238 | seg_acc = np.zeros(len(segments)) 239 | for segs in segment_acc: 240 | last_idx = 0 241 | last_acc = 0. 242 | for bud, acc in segs: 243 | curr_idx = segments.index(bud) 244 | seg_acc[last_idx:curr_idx] += last_acc 245 | last_idx = curr_idx 246 | last_acc = acc 247 | seg_acc[last_idx:] += last_acc 248 | #segment_acc = [[k, v] for k, v in sorted(segment_acc.items())] 249 | #segment_acc = np.array(segment_acc) 250 | seg_acc = seg_acc / total_cnt 251 | with open(os.path.join(self.log_dir, 'budget_acc.txt'), 'w') as f: 252 | for bud, acc in zip(segments, seg_acc): 253 | f.write(f'{bud} {acc}\n') 254 | 255 | return steps 256 | 257 | def init_blit_manager(self): 258 | fig, ax = plt.subplots() 259 | self.blit_manager = BlitManager(ax, fig.canvas) 260 | fig.canvas.draw() 261 | 262 | def log_images(self, inputs, outputs, train, coords_type='x'): 263 | if self.blit_manager is None: 264 | self.init_blit_manager() 265 | 266 | if coords_type == 'x': 267 | x_input = inputs['x'].detach().cpu().numpy() 268 | if 'x_aug' in inputs: 269 | x_aug_input = inputs['x_aug'].detach().cpu().numpy() 270 | pcoords = outputs['pcoords'].detach().cpu().numpy() 271 | elif coords_type == 'x_raw': 272 | def preprocess_raw(inputs, base, prefix='', postfix=''): 273 | if isinstance(inputs[prefix+base+postfix], list): 274 | x_raw = [x.detach().cpu().numpy() for x in inputs[prefix+base+postfix]] 275 | else: 276 | x_raw = inputs[prefix+base+postfix].detach().cpu().numpy() 277 | scale = inputs[prefix+'scale'+postfix].detach().cpu().numpy() 278 | translate = inputs[prefix+'translate'+postfix].detach().cpu().numpy() 279 | seq_mask = inputs[prefix+'seq_mask'+postfix] 280 | if seq_mask is not None: 281 | seq_mask= seq_mask.detach().cpu().numpy() 282 | if prefix+'prim_id'+postfix in inputs and self.model_type == 'pmn': 283 | prim_id = inputs[prefix+'prim_id'+postfix].detach().cpu().numpy() 284 | color_per_id = [PRIM_COLORS[PRIM_ORDER.index(prim_name)] for prim_name in self.model.primitive_names[:-1] + ['circle']] 285 | else: 286 | prim_id = None 287 | offset = 0 288 | x_input = [] 289 | colors = [] 290 | for nstr in inputs[prefix+'n_strokes'+postfix]: 291 | x_input.append((x_raw[offset:offset+nstr], 292 | scale[offset:offset+nstr], 293 | translate[offset:offset+nstr], 294 | seq_mask[offset:offset+nstr] if seq_mask is not None else seq_mask)) 295 | if prim_id is None: 296 | colors.append(None) 297 | else: 298 | colors.append([color_per_id[pi.item()] for pi in prim_id[offset:offset+nstr]]) 299 | offset += nstr 300 | return x_input, colors 301 | x_input, _ = preprocess_raw(inputs, 'x_raw') 302 | if 'x_aug' in inputs or 'x_raw_aug' in inputs: 303 | x_aug_input, _ = preprocess_raw(inputs, 'x_raw', postfix='_aug') 304 | pcoords, pcoords_colors = preprocess_raw(outputs, 'raw', prefix='pcoords_') 305 | else: 306 | raise NotImplementedError 307 | 308 | gt = {aug: [] for aug in range(outputs['n_augment'])} 309 | pred = [] 310 | in_img = [] 311 | if 'x_img' in inputs: 312 | orig_img = [] 313 | for idx in range(min(self.n_log_img, len(x_input))): 314 | if 'x_img' in inputs: 315 | orig_img.append(inputs['x_img'][idx].to('cpu')) 316 | in_img.append(torch.tensor(make_image_fast(self.blit_manager, x_input[idx], self.label_names[inputs['label'][idx]], 1.))) 317 | for aug in range(outputs['n_augment']): 318 | if 'cls_pred_mod' in outputs and outputs['cls_pred_mod'] is not None: 319 | cls_name = self.label_names[outputs['cls_pred_mod'][outputs['n_augment']*idx+aug]] 320 | cls_prob = outputs['cls_pred_mod_prob'][outputs['n_augment']*idx+aug] 321 | else: 322 | cls_name = None 323 | cls_prob = None 324 | gt[aug].append(torch.tensor(make_image_fast(self.blit_manager, x_aug_input[outputs['n_augment']*idx+aug], cls_name, cls_prob))) 325 | 326 | if self.model_type in ['pmn', 'classifier', 'fgsbir']: 327 | if 'cls_pred_main' in outputs and outputs['cls_pred_main'] is not None: 328 | cls_name = self.label_names[outputs['cls_pred_main'][idx]] 329 | cls_prob = outputs['cls_pred_main_prob'][idx] 330 | else: 331 | cls_name = None 332 | cls_prob = None 333 | pred.append(torch.tensor(make_image_fast(self.blit_manager, pcoords[idx], cls_name, cls_prob, color=pcoords_colors[idx]))) 334 | elif self.model_type == 'communication': 335 | max_seq_len = (outputs['pcoords_n_strokes'][:min(self.n_log_img, len(x_input))]).max().item() 336 | pred_idx = [] 337 | for seq_idx in range(max_seq_len): 338 | if 'cls_pred_main' in outputs and outputs['cls_pred_main'] is not None: 339 | cls_name = self.label_names[outputs['cls_pred_main'][idx][seq_idx]] 340 | cls_prob = outputs['cls_pred_main_prob'][idx][seq_idx] 341 | else: 342 | cls_name = None 343 | cls_prob = None 344 | pred_idx.append(torch.tensor(make_image_fast(self.blit_manager, pcoords[idx], cls_name, cls_prob, max_seq_len=seq_idx+1))) 345 | pred.append(pred_idx) 346 | else: 347 | raise NotImplementedError 348 | 349 | if self.model_type == 'communication': 350 | pred = list(map(list, zip(*pred))) 351 | pred = [item for sublist in pred for item in sublist] 352 | 353 | log_imgs_train = [torch.stack(in_img,dim=0), torch.stack(pred,dim=0)] + [torch.stack(gt[aug],dim=0) for aug in range(outputs['n_augment'])] 354 | 355 | if self.model_type == 'fgsbir': 356 | inv_transform = transforms.Compose([ 357 | transforms.Normalize(mean = [ 0., 0., 0. ], 358 | std = [ 1/0.229, 1/0.224, 1/0.225 ]), 359 | transforms.Normalize(mean = [ -0.485, -0.456, -0.406 ], 360 | std = [ 1., 1., 1. ]), 361 | transforms.Resize(480), 362 | transforms.Pad((80, 0)), 363 | ]) 364 | log_imgs_train.append((inv_transform(inputs['positive_img']) * 255).byte().cpu()) 365 | log_imgs_train.append((inv_transform(inputs['negative_img']) * 255).byte().cpu()) 366 | 367 | log_imgs_train = torch.cat(log_imgs_train,dim=0) 368 | 369 | imgs = [('train' if train else 'test', log_imgs_train, min(self.n_log_img, len(x_input)))] 370 | return imgs 371 | 372 | 373 | def run_model(self, inputs, train): 374 | outputs = self.model(**inputs) 375 | #outputs['n_augment'] = 1 if self.model_type == 'pmn' and not self.model.finetune_prims and (not self.model.train_prims or self.model.learned_prims) else 0 376 | outputs['n_augment'] = 0 377 | 378 | 379 | return outputs 380 | -------------------------------------------------------------------------------- /utils/arguments.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | parser = argparse.ArgumentParser() 4 | 5 | parser.add_argument('--data-dir', default='./data', help='Root directory for the data') 6 | parser.add_argument('--preprocessed-root', default=None, help='Path to preprocessed data') 7 | parser.add_argument('--dataset', default='quickdraw09', help='Dataset to train/test') 8 | parser.add_argument('--model-type', default='pmn', help='Model type') 9 | parser.add_argument('--learning-rate', type=float, default=0.0001, help='Learning rate') 10 | parser.add_argument('--batch-size', type=int, default=32, help='Training batch size') 11 | parser.add_argument('--epochs', type=int, default=500, help='Number of epochs') 12 | parser.add_argument('--weight-decay', type=float, default=0., help='Weight decay for optimizer') 13 | parser.add_argument('--Nz', type=int, default=256, help='Latent dimension') 14 | parser.add_argument('--dt-gamma', type=float, default=6., help='Gamma for distance transform') 15 | parser.add_argument('--temperature', type=float, default=0.2, help='Temperature for contrastive loss') 16 | parser.add_argument('--transformations', type=str, default='rsr') 17 | parser.add_argument('--use-proportion-scale', action='store_true', default=False) 18 | parser.add_argument('--transform-bound', type=float, default=None) 19 | parser.add_argument('--learn-compatibility', action='store_true', default=False) 20 | parser.add_argument('--n-log-img', type=int, default=50, help='Number of images to display in the logger') 21 | parser.add_argument('--log-dir', default='./log', help='Root directory for the logs') 22 | parser.add_argument('--log-every', type=int, default=500, help='How often to log') 23 | parser.add_argument('--save-every', type=int, default=5000, help='How often to save model') 24 | parser.add_argument('--log-name', default='test', help='Name of the logs') 25 | parser.add_argument('--loss-model-ckpt', type=str, default=None, help='Checkpoint used as the loss model') 26 | parser.add_argument('--loss-model-type', default='cross_entropy', help='Loss to use for the loss model') 27 | parser.add_argument('--view-scale', type=int, default=256, help='Size of log images') 28 | parser.add_argument('--test', default=None, type=str, help='Checkpoint used for test set evaluation') 29 | 30 | # Position embedding 31 | parser.add_argument('--use-pos-embed', action='store_true', default=False, help='Use positional embeddings') 32 | parser.add_argument('--use-sinusoid-embed', action='store_true', default=False, help='Use sinusoid positional embeddings') 33 | 34 | #FGSBIR 35 | parser.add_argument('--fgsbir-backbone', default='InceptionV3') 36 | parser.add_argument('--load-fgsbir-photos', action='store_true', default=False, help='Load photo images for retrieval task') 37 | 38 | #Communication 39 | parser.add_argument('--comm-mode', default='human') 40 | parser.add_argument('--budget', type=float, default=1.0) 41 | 42 | # Data optimization 43 | parser.add_argument('--no-stroke-resampling', action='store_true', default=False) 44 | parser.add_argument('--max-stroke-length', type=int, default=25) 45 | 46 | # Save processed images 47 | parser.add_argument('--add-comm-model', action='store_true', default=False) 48 | parser.add_argument('--only-save-coords', action='store_true', default=False) 49 | -------------------------------------------------------------------------------- /utils/coords_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from matplotlib.figure import Figure 4 | from matplotlib.axes import Axes 5 | 6 | PRIM_COLORS = ['#E69F00', '#009E73', '#CC79A7', '#56B4E9', '#D55E00', '#0072B2', '#785EF0'] 7 | PRIM_ORDER = ['halfcircle', 'circle', 'line', 'corner', 'triangle', 'square', 'u'] 8 | 9 | def get_stroke_stats(strokes, scale_margin=0.01): 10 | mins = strokes.min(dim=-2, keepdim=True)[0] 11 | maxs = strokes.max(dim=-2, keepdim=True)[0] 12 | sizes = (maxs - mins) 13 | scale = sizes.max(dim=-1, keepdim=True)[0] 14 | scale += scale * scale_margin 15 | # replace 0. scales with margin 16 | scale = scale.masked_fill(scale.isclose(torch.tensor([0.], device=scale.device)), scale_margin) 17 | translation = (mins + maxs) / 2. 18 | aspect = sizes / scale 19 | return scale, translation, aspect 20 | 21 | def normalize_strokes(strokes): 22 | scale, translation, _ = get_stroke_stats(strokes) 23 | strokes = (strokes - translation) / scale * 2. 24 | return strokes, scale, translation 25 | 26 | def preprocess_transformer(seq, resample_strokes=False, resample_size=25, 27 | points_per_primitive=3): 28 | assert seq[0, 2] == 1., seq[0] 29 | 30 | strokes = split_seq_into_strokes(seq, max_seq_len=None if resample_strokes else resample_size) 31 | segments_len = [int(np.ceil(len(s)/points_per_primitive)) for s in strokes] 32 | if resample_strokes: 33 | strokes = [resample_stroke(s, resample_size) for s in strokes] 34 | strokes = torch.stack(strokes) 35 | strokes = (strokes[:, :, :2] - 127.5) / 127.5 36 | strokes, scale, translation = normalize_strokes(strokes) 37 | stroke_lens = tuple([torch.tensor(resample_size).unsqueeze(0) for _ in strokes]) 38 | mask = torch.zeros(strokes.shape[:2], dtype=torch.bool) 39 | position = torch.linspace(-1., 1., steps=resample_size).unsqueeze(0).repeat(len(strokes), 1) 40 | else: 41 | strokes = [(s[:, :2] - 127.5) / 127.5 for s in strokes] 42 | strokes, scale, translation = zip(*[normalize_strokes(s) for s in strokes]) 43 | scale, translation = torch.stack(scale), torch.stack(translation) 44 | stroke_lens = tuple([torch.tensor(len(s)).unsqueeze(0) for s in strokes]) 45 | strokes_tensor = torch.zeros((len(strokes), resample_size, 2)) 46 | mask = torch.ones(strokes_tensor.shape[:2], dtype=torch.bool) 47 | position = torch.zeros(strokes_tensor.shape[:2]) 48 | for i, s in enumerate(strokes): 49 | slen = len(s) 50 | strokes_tensor[i, :slen] = s 51 | mask[i, :slen] = False 52 | position[i, :slen] = torch.linspace(-1., 1., steps=slen) 53 | strokes = strokes_tensor 54 | 55 | segments = [] 56 | for i, sl in enumerate(segments_len): 57 | if resample_strokes: 58 | sl = min(resample_size, sl) 59 | seg = [] 60 | cnt = stroke_lens[i].item() 61 | for n_remain in range(sl, 0, -1): 62 | seg.append(int(np.round(cnt / n_remain))) 63 | cnt -= seg[-1] 64 | segments.append(seg) 65 | 66 | return strokes, len(strokes), scale, translation, segments, stroke_lens, mask, position 67 | 68 | def preprocess_sequence(seq, Nmax=99): 69 | assert seq[0, 2] == 1., seq[0] 70 | nonzero = torch.nonzero(seq.sum(-1) == 0.) 71 | if nonzero.nelement() == 0: 72 | len_seq = Nmax 73 | else: 74 | len_seq = nonzero[0].item() - 1 75 | if seq.shape[0] < Nmax+1: 76 | seq = torch.cat([seq, torch.zeros(Nmax+1-seq.shape[0], 4)]) 77 | new_seq = (seq[:,:2]- 127.5) / 127.5 78 | new_seq = new_seq[1:] - new_seq[:-1] 79 | new_seq = torch.cat([new_seq, seq[1:, 2:], torch.zeros((Nmax, 1))], dim=1) 80 | new_seq[len_seq-1:,4] = 1 81 | new_seq[len_seq-1:,2:4] = 0 82 | new_seq[len_seq:,0:2] = 0 83 | return new_seq, len_seq 84 | 85 | 86 | def split_seq_into_strokes(orig_seq, max_seq_len=None): 87 | split_idx = [0]+list(torch.where(orig_seq[:,-1]>0)[0]+1) 88 | split_idx = [(split_idx[i+1]-split_idx[i]).item() for i in range(len(split_idx)-1)] 89 | pad_len = orig_seq.shape[0] - sum(split_idx) 90 | if pad_len > 0: 91 | split_idx.append(pad_len) 92 | 93 | current_strokes = list(torch.split(orig_seq, split_idx)) 94 | if max_seq_len is not None: 95 | updated_strokes = [] 96 | for stroke in current_strokes: 97 | if stroke.shape[0] > max_seq_len and not (stroke[0] == 0.).all().item(): 98 | n_splits = int(np.ceil(stroke.shape[0] / max_seq_len)) 99 | if stroke.shape[0] > (max_seq_len*n_splits-n_splits+1): 100 | n_splits += 1 101 | 102 | new_strokes = [] 103 | offset = 0 104 | for n_remain in range(n_splits, 0, -1): 105 | stroke_len = int(np.round((stroke.shape[0]+n_remain-1-offset) / n_remain)) 106 | new_strokes.append(stroke[offset:offset+stroke_len]) 107 | offset = offset + stroke_len - 1 108 | 109 | updated_strokes.extend(new_strokes) 110 | else: 111 | updated_strokes.append(stroke) 112 | current_strokes = updated_strokes 113 | 114 | is_padded = (current_strokes[-1][0] == 0.).all().item() 115 | if is_padded: 116 | current_strokes = current_strokes[:-1] 117 | 118 | pnt_strokes = [] 119 | for i, stroke in enumerate(current_strokes): 120 | if stroke.shape[0] == 1: 121 | pnt_strokes.append(i) 122 | for i in reversed(pnt_strokes): 123 | del current_strokes[i] 124 | 125 | return current_strokes 126 | 127 | 128 | def resample_strokes(current_strokes, resample_ratios, max_points): 129 | avail_points = max_points - sum([c.shape[0] for c in current_strokes]) 130 | new_strokes = [] 131 | for curr_stroke in current_strokes: 132 | n_pnts = curr_stroke.shape[0] 133 | ratio = np.random.rand() * (resample_ratios[1] - resample_ratios[0]) + resample_ratios[0] 134 | while (new_n_pnts := max(2, int(np.around(n_pnts * ratio)))) - n_pnts > avail_points: 135 | ratio = np.random.rand() * (resample_ratios[1] - resample_ratios[0]) + resample_ratios[0] 136 | new_strokes.append(resample_stroke(curr_stroke.clone(), new_n_pnts)) 137 | avail_points -= new_n_pnts - n_pnts 138 | return new_strokes 139 | 140 | 141 | def make_image_fast(blit_manager, sequence, class_name=None, class_prob=None, no_axis=False, color=None, max_seq_len=None, linewidth=3, enlarge=1.0, arbitrary_mask=False): 142 | if isinstance(sequence, tuple): 143 | # raw stroke-based coordinates 144 | seq, scale, translate, seq_mask = sequence 145 | if isinstance(seq, list): 146 | strokes = [(se / 2 * sc + tr) * enlarge for se, sc, tr in zip(seq, scale, translate)] 147 | else: 148 | strokes = (seq / 2 * scale + translate) * enlarge 149 | check_end = False 150 | else: 151 | # relative coordinates 152 | sequence[:, 0] = np.cumsum(sequence[:, 0]) 153 | sequence[:, 1] = np.cumsum(sequence[:, 1]) 154 | """plot drawing with separated strokes""" 155 | strokes = np.split(sequence, np.where(sequence[:, 3] + sequence[:, 4] > 0)[0] + 1) 156 | strokes[0] = np.concatenate([strokes[0][0][np.newaxis], strokes[0]], axis=0) 157 | strokes[0][0, :2] = 0. 158 | check_end = True 159 | seq_mask = None 160 | 161 | if max_seq_len is not None: 162 | if max_seq_len > len(strokes): 163 | strokes = [] 164 | else: 165 | strokes = strokes[:max_seq_len] 166 | 167 | if isinstance(blit_manager, Axes): 168 | ax = blit_manager 169 | artists = [] 170 | else: 171 | ax = blit_manager.ax 172 | artists = blit_manager.get_artists() 173 | if len(artists) == 0: 174 | ax.set_xlim(-2, 2) 175 | ax.set_ylim(-2, 2) 176 | ax.set_aspect('equal') 177 | if no_axis: 178 | ax.set_axis_off() 179 | txt = ax.text(-2.2, 2.2, '', size=20, animated=True) 180 | blit_manager.add_artist(txt) 181 | else: 182 | txt = artists[0] 183 | 184 | if not no_axis and class_name is not None and class_prob is not None: 185 | txt.set_text(f'{class_name} ({class_prob:.2%})') 186 | else: 187 | txt.set_text('') 188 | 189 | if arbitrary_mask and seq_mask is not None: 190 | new_strokes = [] 191 | for s, m in zip(strokes, seq_mask): 192 | used = np.argwhere(~m).flatten() 193 | stroke_sections = [] 194 | start = None 195 | for u in used: 196 | if start is None: 197 | start = u 198 | elif u != last_u + 1: 199 | stroke_sections.append((start, last_u)) 200 | start = u 201 | last_u = u 202 | if start is not None: 203 | stroke_sections.append((start, last_u)) 204 | 205 | for sec_start, sec_end in stroke_sections: 206 | new_strokes.append(s[sec_start:sec_end+1]) 207 | 208 | strokes = new_strokes 209 | seq_mask = None 210 | 211 | n_strokes = 0 212 | for i, s in enumerate(strokes): 213 | if s.size == 0: 214 | break 215 | 216 | x = s[:, 0] 217 | y = -s[:, 1] 218 | if seq_mask is not None: 219 | seq_len = np.argwhere(seq_mask[i]).flatten() 220 | if len(seq_len) > 0: 221 | x = x[:seq_len[0]] 222 | y = y[:seq_len[0]] 223 | 224 | if (x == x[0]).all() and (y == y[0]).all(): 225 | marker = 'o' 226 | else: 227 | marker = '' 228 | 229 | if color is not None: 230 | if isinstance(color, list): 231 | c = color[i] 232 | else: 233 | c = color 234 | 235 | if i+1 < len(artists): 236 | artists[i+1].set_data(x, y) 237 | artists[i+1].set_marker(marker) 238 | if color is not None: 239 | artists[i+1].set_color(c) 240 | else: 241 | if color is not None: 242 | (line,) = ax.plot(x, y, marker=marker, linewidth=linewidth, animated=True, color=c) 243 | else: 244 | (line,) = ax.plot(x, y, marker=marker, linewidth=linewidth, animated=True) 245 | blit_manager.add_artist(line) 246 | 247 | n_strokes += 1 248 | if check_end and s[-1, 4] == 1: 249 | break 250 | 251 | # don't show unused artists 252 | for i in range(n_strokes+1, len(artists)): 253 | artists[i].set_data([], []) 254 | 255 | if not isinstance(blit_manager, Axes): 256 | blit_manager.update() 257 | # grab the pixel buffer and dump it into a numpy array 258 | img = np.array(blit_manager.canvas.renderer.buffer_rgba())[:, :, :3] 259 | img = np.moveaxis(img, -1, 0) 260 | 261 | return img 262 | 263 | 264 | def resample_data(edge_list, n=20, random=True, endpoint=False): 265 | total_edge_length = sum([e[0] for e in edge_list]) 266 | if random: 267 | locations = sorted(np.random.rand(n) * total_edge_length) 268 | else: 269 | if endpoint: 270 | locations = np.linspace(0., total_edge_length, num=n) 271 | else: 272 | locations = np.linspace(0., total_edge_length, num=n+1)[:-1] 273 | 274 | points = [] 275 | loc_i = 0 276 | curr_length = 0 277 | for edge in edge_list: 278 | start_pnt = edge[1] 279 | end_pnt = edge[2] 280 | edge_length = edge[0] 281 | while loc_i < n and (rel_length := locations[loc_i] - curr_length) <= edge_length: 282 | if edge_length == 0.: 283 | pnt = start_pnt 284 | else: 285 | pnt = start_pnt + rel_length/edge_length * (end_pnt - start_pnt) 286 | points.append(pnt) 287 | loc_i += 1 288 | 289 | curr_length += edge[0] 290 | 291 | points = np.array(points) 292 | assert (~np.isnan(points)).all() 293 | 294 | return points 295 | 296 | def resample_stroke(stroke, n): 297 | pnts = stroke[:, :2].numpy() 298 | edge_list = [[np.linalg.norm(pnts[i] - pnts[i+1]), pnts[i], pnts[i+1]] for i in range(pnts.shape[0]-1)] 299 | re_pnts = resample_data(edge_list, n, random=False, endpoint=True) 300 | new_stroke = torch.from_numpy(re_pnts) 301 | states = torch.zeros(new_stroke.shape, dtype=new_stroke.dtype) 302 | states[:, 0] = 1. 303 | states[-1, 0] = 0. 304 | states[-1, 1] = 1. 305 | new_stroke = torch.cat([new_stroke, states], dim=1) 306 | return new_stroke 307 | -------------------------------------------------------------------------------- /utils/data.py: -------------------------------------------------------------------------------- 1 | import os 2 | import collections 3 | from itertools import chain, combinations 4 | import pickle 5 | import re 6 | 7 | import matplotlib.pyplot as plt 8 | import numpy as np 9 | import torch 10 | import torchvision 11 | import torch.utils.data as data 12 | import torchvision.transforms as transforms 13 | import torchvision.transforms.functional as F 14 | from tqdm import tqdm 15 | from PIL import Image 16 | import copy 17 | 18 | from .coords_utils import preprocess_sequence, preprocess_transformer, make_image_fast 19 | from .transformations import transform_primitives, ToTransformMatrix 20 | from .plt_blit import BlitManager 21 | 22 | from torch.utils.data._utils.collate import default_collate 23 | 24 | 25 | def sketch_collate(batch): 26 | r"""Puts each data field into a tensor with outer dimension batch size""" 27 | 28 | elem = batch[0] 29 | if isinstance(elem, collections.abc.Mapping): 30 | return {key: sketch_collate([d[key] for d in batch]) for key in elem} 31 | elif isinstance(elem, collections.abc.Sequence) and not isinstance(elem, str): 32 | # check to make sure that the elements in batch have consistent size 33 | it = iter(batch) 34 | elem = next(it) 35 | elem_size = len(elem) 36 | elem_dims = elem[0].dim() if isinstance(elem[0], torch.Tensor) else 0 37 | if not all(len(elem) == elem_size for elem in it) or elem_dims > 1 or isinstance(elem, tuple): 38 | batch = sum(batch, tuple()) 39 | 40 | if elem_dims > 0: 41 | return torch.cat(batch) 42 | else: 43 | transposed = zip(*batch) 44 | return [sketch_collate(samples) for samples in transposed] 45 | else: 46 | return default_collate(batch) 47 | 48 | 49 | 50 | 51 | class PreprocessedSketchDataset(data.Dataset): 52 | def __init__(self, coordinate_path_root, sketch_list, idx_class=0, 53 | original_strokes_list=None, load_fgsbir_photos=False, 54 | original_path_root=None): 55 | self.coordinate_path_root = coordinate_path_root 56 | self.load_fgsbir_photos = load_fgsbir_photos 57 | 58 | with open(sketch_list) as sketch_url_file: 59 | sketch_url_list = sketch_url_file.readlines() 60 | 61 | self.coordinate_urls, sketch_url_list = self.get_coords_list(coordinate_path_root, sketch_url_list) 62 | 63 | self.labels = [int(sketch_url.strip().split(' ')[1]) 64 | for sketch_url in sketch_url_list] 65 | label_names = [sketch_url.strip().split(' ')[0].split('/')[idx_class + 1] 66 | for sketch_url in sketch_url_list] 67 | 68 | if self.load_fgsbir_photos: 69 | assert original_path_root is not None 70 | alt_coordinate_urls, _ = self.get_coords_list(os.path.join(original_path_root, 'picture_files'), sketch_url_list) 71 | self.picture_urls = [re.sub('_[^_]*\.pt', '.png', cu.replace('/train/', '/').replace('/test/', '/')) for cu in alt_coordinate_urls] 72 | self.unique_picture_urls = set(self.picture_urls) 73 | 74 | self.img_transform = transforms.Compose([ 75 | transforms.Resize(299), transforms.ToTensor(), 76 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 77 | std=[0.229, 0.224, 0.225]) 78 | ]) 79 | 80 | self.original_strokes = None 81 | 82 | self.label_names = np.unique(label_names) 83 | 84 | self.n_classes = len(self.label_names) 85 | self.keys = ['x_raw', 'scale', 'translate', 'seq_mask', 'position'] 86 | 87 | 88 | 89 | def __len__(self): 90 | return len(self.coordinate_urls) 91 | 92 | def get_coords_list(self, coordinate_path_root, sketch_url_list): 93 | return [os.path.join(coordinate_path_root, (sketch_url.strip( 94 | ).split(' ')[0]).replace('.npy', '.pt')) for sketch_url in sketch_url_list], sketch_url_list 95 | 96 | 97 | def __getitem__(self, item): 98 | coordinate = torch.load(self.coordinate_urls[item]) 99 | assert coordinate['label'] == self.labels[item] 100 | 101 | if isinstance(coordinate['x_raw'], list): 102 | split_scale = list(coordinate['scale'].split(1, dim=0)) 103 | split_translate = list(coordinate['translate'].split(1, dim=0)) 104 | coordinate['n_strokes'] = 0 105 | for i in range(len(coordinate['x_raw'])): 106 | split_scale[i] = split_scale[i].repeat(coordinate['x_raw'][i].shape[0],1,1) 107 | split_translate[i] = split_translate[i].repeat(coordinate['x_raw'][i].shape[0],1,1) 108 | coordinate['n_strokes']+=coordinate['x_raw'][i].shape[0] 109 | coordinate['x_raw'] = torch.cat(coordinate['x_raw'],dim=0).float() 110 | coordinate['scale'] = torch.cat(split_scale,dim=0).float() 111 | coordinate['translate'] = torch.cat(split_translate,dim=0).float() 112 | 113 | for k in coordinate.keys(): 114 | if isinstance(coordinate[k], np.ndarray): 115 | coordinate[k] = torch.from_numpy(coordinate[k]).float() 116 | elif k == 'scale' and isinstance(coordinate[k], float): 117 | coordinate[k] = torch.tensor([[[coordinate[k]]]]) 118 | elif k == 'stroke_lens' and isinstance(coordinate[k], int): 119 | coordinate[k] = torch.tensor([coordinate[k]]) 120 | 121 | coordinate['seq_mask'] = torch.zeros(coordinate['x_raw'].shape[:2], dtype=torch.bool) 122 | coordinate['position'] = torch.linspace(-1., 1., steps=coordinate['x_raw'].shape[1]).unsqueeze(0).repeat(coordinate['x_raw'].shape[0], 1) 123 | data_dict = {k: coordinate[k].split(1, dim=0) for k in self.keys} 124 | data_dict['label'] = coordinate['label'] 125 | data_dict['n_strokes'] = coordinate['n_strokes'] 126 | if isinstance(coordinate['segments'], int): 127 | data_dict['segments'] = tuple([torch.tensor(coordinate['segments']).unsqueeze(0)]) 128 | else: 129 | data_dict['segments'] = tuple([c.unsqueeze(0) for c in coordinate['segments']]) 130 | data_dict['stroke_lens'] = tuple([s.unsqueeze(0) for s in coordinate['stroke_lens']]) 131 | 132 | if 'prim_id' in coordinate: 133 | if isinstance(coordinate['prim_id'], int): 134 | coordinate['prim_id'] = torch.tensor([[[coordinate['prim_id']]]]) 135 | data_dict['prim_id'] = coordinate['prim_id'].split(1, dim=0) 136 | 137 | if 'n_strokes_orig' in coordinate: 138 | data_dict['n_strokes_orig'] = coordinate['n_strokes_orig'] 139 | 140 | if self.load_fgsbir_photos: 141 | # TODO: duplicate code 142 | positive_path = self.picture_urls[item] 143 | uniq_urls = set(self.unique_picture_urls) 144 | uniq_urls.remove(positive_path) 145 | negative_path = np.random.choice(list(uniq_urls)) 146 | positive_img = Image.open(positive_path).convert('RGB') 147 | negative_img = Image.open(negative_path).convert('RGB') 148 | positive_img = self.img_transform(positive_img) 149 | negative_img = self.img_transform(negative_img) 150 | data_dict['positive_img'] = positive_img 151 | data_dict['negative_img'] = negative_img 152 | data_dict['positive_name'] = positive_path.split('/')[-1].split('.')[0] 153 | 154 | if self.original_strokes is not None: 155 | data_dict['original_n_strokes'] = self.original_strokes[item] 156 | data_dict['name'] = item 157 | return data_dict 158 | 159 | def save_img(self, path, suffix='.pt', process=True): 160 | fig, ax = plt.subplots(figsize=(2.24, 2.24)) 161 | ax.set_axis_off() 162 | fig.add_axes(ax) 163 | self.blit_manager = BlitManager(ax, fig.canvas) 164 | fig.canvas.draw() 165 | for item in tqdm(range(len(self.coordinate_urls)), desc='Saving images'): 166 | file_path = self.coordinate_urls[item] 167 | relative_path = os.path.relpath(file_path, self.coordinate_path_root) 168 | file_path = os.path.join(path, relative_path).replace(suffix, '.png') 169 | os.makedirs(os.path.dirname(file_path), exist_ok=True) 170 | coordinate = torch.load(self.coordinate_urls[item]) 171 | sequence_raw = coordinate['x_raw'] 172 | scale = coordinate['scale'] 173 | translate = coordinate['translate'] 174 | 175 | if isinstance(sequence_raw, list): 176 | sequence_img = torch.cat( 177 | [((sequence_raw[i] / sequence_raw[i].abs().max()) / 2 * scale[i] + translate[i]) for i in 178 | range(len(sequence_raw))], dim=0).numpy() 179 | else: 180 | sequence_img = (sequence_raw / 2 * scale + translate).numpy() 181 | x_img = make_image_fast(self.blit_manager, (sequence_img, 3.0, 0.0, None), no_axis=True, color='black') 182 | im = Image.fromarray(np.moveaxis(x_img, 0, -1).astype(np.uint8)) 183 | im.save(file_path) 184 | 185 | def add_strokes_to_list(self, file_name): 186 | full_text = "" 187 | for item in tqdm(range(len(self.coordinate_urls)), desc='Updating file'): 188 | data = self.__getitem__(item) 189 | n_strokes = data['n_strokes'] 190 | label = data['label'] 191 | file_path = self.coordinate_urls[item] 192 | relative_path = os.path.relpath(file_path, self.coordinate_path_root) 193 | text = relative_path + ' ' + str(label) + ' ' + str(n_strokes.item()) 194 | full_text+=text+'\n' 195 | 196 | with open(file_name, 'w') as file: 197 | file.write(full_text) 198 | 199 | def save_preprocessed(self, root, suffix='.pt'): 200 | self.save_img(os.path.join(root, 'imgs'), suffix) 201 | #self.save_coords(os.path.join(root, 'coords'), suffix) 202 | 203 | 204 | class SketchDataset(data.Dataset): 205 | def __init__(self, coordinate_path_root, sketch_list, 206 | idx_class=0, preprocessed=None, 207 | Nmax=99, load_fgsbir_photos=False, 208 | points_per_primitives=3, resample_strokes=False, 209 | max_stroke_length=25): 210 | self.coordinate_path_root = coordinate_path_root 211 | self.preprocessed = preprocessed 212 | self.Nmax = Nmax 213 | self.load_fgsbir_photos = load_fgsbir_photos 214 | self.points_per_primitive = points_per_primitives 215 | self.resample_strokes = resample_strokes 216 | self.max_stroke_length = max_stroke_length 217 | 218 | with open(sketch_list) as sketch_url_file: 219 | sketch_url_list = sketch_url_file.readlines() 220 | 221 | self.coordinate_urls, sketch_url_list = self.get_coords_list(coordinate_path_root, sketch_url_list) 222 | 223 | self.labels = [int(sketch_url.strip().split(' ')[-1]) 224 | for sketch_url in sketch_url_list] 225 | label_names = [sketch_url.strip().split(' ')[0].split('/')[idx_class + 1] 226 | for sketch_url in sketch_url_list] 227 | self.label_names = np.unique(label_names) 228 | 229 | if self.load_fgsbir_photos: 230 | self.sketch_picture_urls = [re.sub('.npy', '.png', cu.replace('coordinate_files', 'sketch_picture_files')) for cu in self.coordinate_urls] 231 | self.picture_urls = [re.sub('_[^_]*\.npy', '.png', cu.replace('coordinate_files', 'picture_files').replace('/train/', '/').replace('/test/', '/')) for cu in self.coordinate_urls] 232 | self.unique_picture_urls = set(self.picture_urls) 233 | 234 | self.img_transform = transforms.Compose([ 235 | transforms.Resize(299), transforms.ToTensor(), 236 | transforms.Normalize(mean=[0.485, 0.456, 0.406], 237 | std=[0.229, 0.224, 0.225]), 238 | #transforms.RandomHorizontalFlip(), 239 | ]) 240 | 241 | self.n_classes = len(self.label_names) 242 | 243 | def __len__(self): 244 | return len(self.coordinate_urls) 245 | 246 | def set_preprocessed(self, preprocessed): 247 | print('Setting preprocessed') 248 | self.preprocessed = preprocessed 249 | 250 | def get_coords_list(self, coordinate_path_root, sketch_url_list): 251 | return [os.path.join(coordinate_path_root, (sketch_url.strip( 252 | ).split(' ')[0]).replace('png', 'npy')) for sketch_url in sketch_url_list], sketch_url_list 253 | 254 | def read_coordinates(self, coordinate_urls, idx, normalize=False): 255 | coordinate = np.load(coordinate_urls[idx], encoding='latin1', allow_pickle=True) 256 | 257 | if coordinate.dtype == 'object': 258 | coordinate = coordinate[0] 259 | 260 | coordinate = coordinate.astype('float32') 261 | coordinate = torch.from_numpy(coordinate) 262 | if normalize: 263 | coordinate[:, :2] = ((coordinate[:, :2] - coordinate[:, :2].min()) / ( 264 | coordinate[:, :2] - coordinate[:, :2].min()).max()) * 255 265 | 266 | return coordinate 267 | 268 | def save_img(self, path, suffix='.svg', process=True): 269 | fig, ax = plt.subplots(figsize=(2.99, 2.99), constrained_layout=True) 270 | ax.set_axis_off() 271 | fig.add_axes(ax) 272 | self.blit_manager = BlitManager(ax, fig.canvas) 273 | fig.canvas.draw() 274 | for item in tqdm(range(len(self.coordinate_urls)), desc='Saving images'): 275 | file_path = self.coordinate_urls[item] 276 | relative_path = os.path.relpath(file_path, self.coordinate_path_root) 277 | file_path = os.path.join(path, relative_path).replace(suffix, '.png') 278 | os.makedirs(os.path.dirname(file_path), exist_ok=True) 279 | coordinate = self.read_coordinates(self.coordinate_urls, item) 280 | sequence_raw = self.preprocessed[item]['x_raw'] 281 | scale = self.preprocessed[item]['scale'] 282 | translate = self.preprocessed[item]['translate'] 283 | 284 | x_img = make_image_fast(self.blit_manager, (sequence_raw, scale, translate, None), no_axis=True, color='black', linewidth=1, enlarge=2.0) 285 | im = Image.fromarray(np.moveaxis(x_img, 0, -1).astype(np.uint8)) 286 | im.save(file_path) 287 | 288 | def save_coords(self, path, suffix='.svg'): 289 | for item in tqdm(range(len(self.coordinate_urls)), desc='Saving coords'): 290 | file_path = self.coordinate_urls[item] 291 | relative_path = os.path.relpath(file_path, self.coordinate_path_root) 292 | file_path = os.path.join(path, relative_path).replace(suffix, '.pt') 293 | os.makedirs(os.path.dirname(file_path), exist_ok=True) 294 | coordinate = self.preprocessed[item] 295 | for k in coordinate.keys(): 296 | if isinstance(coordinate[k], torch.Tensor): 297 | if coordinate[k].numel() == 1: 298 | coordinate[k] = coordinate[k].item() 299 | else: 300 | coordinate[k] = coordinate[k].detach().cpu().numpy() 301 | coordinate['path'] = relative_path 302 | torch.save(coordinate, file_path) 303 | 304 | def add_strokes_to_list(self, file_name): 305 | full_text = "" 306 | for item in tqdm(range(len(self.coordinate_urls)), desc='Updating file'): 307 | coordinate = self.read_coordinates(self.coordinate_urls, item) 308 | file_path = self.coordinate_urls[item] 309 | relative_path = os.path.relpath(file_path, self.coordinate_path_root) 310 | _, n_strokes, _, _ = preprocess_transformer(coordinate) 311 | text = relative_path + ' ' + str(self.labels[item]) + ' ' + str(n_strokes) 312 | full_text+=text+'\n' 313 | with open(file_name, 'w') as file: 314 | file.write(full_text) 315 | 316 | 317 | def save_preprocessed(self, root, suffix='.svg', save_images=True): 318 | if save_images: 319 | self.save_img(os.path.join(root, 'imgs'), suffix) 320 | self.save_coords(os.path.join(root, 'coords'), suffix) 321 | 322 | def __getitem__(self, item): 323 | label = self.labels[item] 324 | 325 | coordinate = self.read_coordinates(self.coordinate_urls, item) 326 | 327 | sequence, length = preprocess_sequence(coordinate, Nmax=self.Nmax) 328 | sequence_raw, n_strokes, scale, translate, segments, stroke_lens, mask, position = preprocess_transformer( 329 | coordinate, resample_strokes=self.resample_strokes, 330 | points_per_primitive=self.points_per_primitive, 331 | resample_size=self.max_stroke_length 332 | ) 333 | if self.preprocessed is not None: 334 | sequence_raw = self.preprocessed[item] 335 | # TODO: check if all calculated values match with preprocessed 336 | assert False 337 | 338 | sequence_raw = sequence_raw.split(1, dim=0) 339 | scale = scale.split(1, dim=0) 340 | translate = translate.split(1, dim=0) 341 | position = position.split(1, dim=0) 342 | if mask is not None: 343 | mask = mask.split(1, dim=0) 344 | 345 | data_dict = { 346 | 'x': sequence, 347 | 'length': length, 348 | 'label': label, 349 | 'x_raw': sequence_raw, 350 | 'n_strokes': n_strokes, 351 | 'scale': scale, 352 | 'translate': translate, 353 | 'segments': tuple([torch.tensor(s).unsqueeze(0) for seg in segments for s in seg]), 354 | 'stroke_lens': stroke_lens, 355 | 'seq_mask': mask, 356 | 'position': position, 357 | 'data_id': item 358 | } 359 | 360 | if self.load_fgsbir_photos: 361 | positive_path = self.picture_urls[item] 362 | sketch_img_path = self.sketch_picture_urls[item] 363 | uniq_urls = set(self.unique_picture_urls) 364 | uniq_urls.remove(positive_path) 365 | negative_path = np.random.choice(list(uniq_urls)) 366 | positive_img = Image.open(positive_path).convert('RGB') 367 | negative_img = Image.open(negative_path).convert('RGB') 368 | if os.path.exists(sketch_img_path): 369 | sketch_img = Image.open(sketch_img_path).convert('RGB') 370 | else: 371 | sketch_img = None 372 | positive_img = self.img_transform(positive_img) 373 | negative_img = self.img_transform(negative_img) 374 | if sketch_img is not None: 375 | sketch_img = self.img_transform(sketch_img) 376 | data_dict['sketch_img'] = sketch_img 377 | else: 378 | data_dict['sketch_img'] = -1 379 | data_dict['positive_img'] = positive_img 380 | data_dict['negative_img'] = negative_img 381 | data_dict['positive_name'] = positive_path.split('/')[-1].split('.')[0] 382 | 383 | return data_dict 384 | 385 | 386 | def get_data(data_root='./data', name='MNIST', batch_size=128, train=True, 387 | shuffle=True, 388 | preprocessed_root=None, 389 | original_strokes_list=None, load_fgsbir_photos=False, 390 | resample_strokes=False, max_stroke_length=25): 391 | 392 | if name == 'quickdraw': 393 | data_dir = os.path.join(data_root, 'quickdraw') 394 | 395 | data_dir_coords = os.path.join(data_dir, 'coordinate_files') 396 | data_dir = os.path.join(data_dir, 'picture_files') 397 | if train: 398 | sketch_list = os.path.join(data_dir_coords, 'train.txt') 399 | if preprocessed_root is not None: 400 | data_coords = PreprocessedSketchDataset(coordinate_path_root=preprocessed_root, 401 | sketch_list=sketch_list, 402 | original_strokes_list=original_strokes_list) 403 | else: 404 | data_coords = SketchDataset( 405 | data_dir_coords, sketch_list=sketch_list, 406 | resample_strokes=resample_strokes, 407 | max_stroke_length=max_stroke_length 408 | ) 409 | else: 410 | sketch_list = os.path.join(data_dir_coords, 'test.txt') 411 | if preprocessed_root is not None: 412 | data_coords = PreprocessedSketchDataset(coordinate_path_root=preprocessed_root, 413 | sketch_list=sketch_list, 414 | original_strokes_list=original_strokes_list) 415 | else: 416 | data_coords = SketchDataset( 417 | data_dir_coords, sketch_list=sketch_list, 418 | resample_strokes=resample_strokes, 419 | max_stroke_length=max_stroke_length 420 | ) 421 | 422 | elif name == 'quickdraw09': 423 | data_dir_coords = os.path.join(data_root, 'quickdraw09') 424 | 425 | if train: 426 | sketch_list = os.path.join(data_dir_coords, 'train.txt') 427 | 428 | if preprocessed_root is not None: 429 | data_coords = PreprocessedSketchDataset( 430 | coordinate_path_root=preprocessed_root, 431 | sketch_list=sketch_list, 432 | original_strokes_list=original_strokes_list) 433 | else: 434 | data_coords = SketchDataset( 435 | data_dir_coords, sketch_list=sketch_list, 436 | Nmax=160, 437 | resample_strokes=resample_strokes, 438 | max_stroke_length=max_stroke_length 439 | ) 440 | else: 441 | sketch_list = os.path.join(data_dir_coords, 'test.txt') 442 | if preprocessed_root is not None: 443 | data_coords = PreprocessedSketchDataset( 444 | coordinate_path_root=preprocessed_root, 445 | sketch_list=sketch_list, 446 | original_strokes_list=original_strokes_list) 447 | else: 448 | data_coords = SketchDataset( 449 | data_dir_coords, sketch_list=sketch_list, 450 | Nmax=160, 451 | resample_strokes=resample_strokes, 452 | max_stroke_length=max_stroke_length 453 | ) 454 | 455 | 456 | elif name in ['chairv2', 'shoev2']: 457 | data_dir = os.path.join(data_root, name) 458 | data_dir_coords = os.path.join(data_dir, 'coordinate_files') 459 | 460 | if train: 461 | sketch_list = os.path.join(data_dir_coords, 'train.txt') 462 | else: 463 | sketch_list = os.path.join(data_dir_coords, 'test.txt') 464 | 465 | if preprocessed_root is not None: 466 | data_coords = PreprocessedSketchDataset( 467 | coordinate_path_root=preprocessed_root, sketch_list=sketch_list, 468 | original_strokes_list=original_strokes_list, 469 | load_fgsbir_photos=load_fgsbir_photos, original_path_root=data_dir) 470 | else: 471 | data_coords = SketchDataset( 472 | data_dir_coords, sketch_list=sketch_list, 473 | Nmax=520, load_fgsbir_photos=load_fgsbir_photos, 474 | resample_strokes=resample_strokes, 475 | max_stroke_length=max_stroke_length 476 | ) 477 | 478 | else: 479 | raise NotImplementedError 480 | 481 | if not train: 482 | generator = torch.Generator() 483 | generator.manual_seed(42) 484 | else: 485 | generator = None 486 | 487 | # Initialize dataloaders 488 | loader_coords = torch.utils.data.DataLoader(data_coords, batch_size, shuffle=shuffle, num_workers=8, 489 | generator=generator, collate_fn=sketch_collate) 490 | 491 | return loader_coords, data_coords.n_classes, data_coords.label_names 492 | -------------------------------------------------------------------------------- /utils/distance_transform.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | 4 | def compute_distance_transform_l2(batch, prims, grid_sqz, batch_mask=None, 5 | prims_mask=None, gamma=3.): 6 | gt_dist = distance_transform(batch, grid_sqz, mask=batch_mask, gamma=gamma) 7 | if prims.dim() == 4: 8 | dims = prims.shape 9 | prims = prims.view(-1, prims.size(2), prims.size(3)) 10 | else: 11 | dims = None 12 | prim_dist = distance_transform(prims, grid_sqz, mask=prims_mask, gamma=gamma) 13 | if dims is not None: 14 | prim_dist = prim_dist.view(dims[0], dims[1], prim_dist.size(-2), prim_dist.size(-1)) 15 | gt_dist = gt_dist.unsqueeze(1).expand(prim_dist.size()) 16 | return F.mse_loss(gt_dist, prim_dist, reduction='none').mean([-2, -1]) 17 | 18 | def make_coordinate_grid(dim, scale=1.5): 19 | """ 20 | Create a meshgrid [-1,1] x [-1,1] of given spatial_size. 21 | """ 22 | h, w = (dim, dim) 23 | x = torch.arange(w).float() 24 | y = torch.arange(h).float() 25 | 26 | yy = y.view(-1, 1).repeat(1, w) 27 | xx = x.view(1, -1).repeat(h, 1) 28 | 29 | meshed = (torch.cat([xx.unsqueeze_(2), yy.unsqueeze_(2)], 2).unsqueeze(0).unsqueeze(0) / (h - 1) * 2) - 1 30 | meshed *= scale 31 | 32 | return meshed 33 | 34 | def distance_transform(kps, grid_sqz, gamma=3., glb=False, mask=None): 35 | grid_size = grid_sqz.shape[-2] 36 | grid_sqz = grid_sqz.repeat(kps.size(0),1, 1, 1, 1) 37 | # kps = torch.index_select(kps_, 2, torch.LongTensor([1, 0]).to(kps_.device)) 38 | pi_set = kps[:, :-1].unsqueeze(-2).unsqueeze(-2) 39 | pj_set = kps[:, 1:].unsqueeze(-2).unsqueeze(-2) 40 | 41 | # Compute r 42 | v_set = (pi_set - pj_set).repeat(1, 1, grid_size, grid_size, 1) 43 | v_norm = (v_set.pow(2)).sum(-1).unsqueeze(-1) 44 | u_set = (grid_sqz - pj_set) 45 | 46 | uv = torch.bmm(u_set.view(-1, 1, 2), (v_set).view(-1, 2, 1)).view(kps.shape[0], -1, grid_size, grid_size, 1) 47 | rs = torch.clamp(uv / v_norm, 0, 1)#.detach() 48 | rs.masked_scatter_(rs.isnan(), uv) 49 | #rs = rs.detach() 50 | 51 | betas = ((u_set - rs * v_set).pow(2)).sum(-1) 52 | betas = torch.exp(-gamma * betas) 53 | 54 | if mask is not None: 55 | betas = betas * (~mask[:, 1:]).float().unsqueeze(-1).unsqueeze(-1) 56 | 57 | betas = betas.max(1)[0] 58 | if glb: 59 | betas = betas.max(0)[0] 60 | 61 | return betas 62 | -------------------------------------------------------------------------------- /utils/metrics.py: -------------------------------------------------------------------------------- 1 | from collections import defaultdict 2 | from typing import Dict 3 | import os 4 | import numpy as np 5 | 6 | import torch 7 | import torchvision 8 | 9 | from torch.utils.tensorboard import SummaryWriter 10 | 11 | 12 | class TbLogger(object): 13 | def __init__(self, log_dir: str): 14 | self.log_dir = os.path.expanduser(log_dir) 15 | self.tb_writer = SummaryWriter(log_dir=log_dir, flush_secs=10) 16 | 17 | def log_metrics(self, metrics: Dict, global_step): 18 | metrics = metrics.copy() 19 | # Scalars and images 20 | for key, value in metrics.items(): 21 | if 'image' in key: 22 | # 3D shape of [3, H, W] with channels R G B 23 | self.tb_writer.add_image(**value, global_step=global_step) 24 | elif isinstance(value, np.ndarray) and len(value) > 1: 25 | for n, v in enumerate(value): 26 | self.tb_writer.add_scalar(f'{key}_{n}', v, global_step) 27 | else: 28 | self.tb_writer.add_scalar(key, value, global_step) 29 | 30 | def close(self): 31 | self.tb_writer.close() 32 | 33 | 34 | class Accumulator(object): 35 | def __init__(self, name='accumulator'): 36 | self.name = name 37 | self.value = 0.0 38 | self.count = 0.0 39 | 40 | def add(self, value, count=1): 41 | if isinstance(value, torch.Tensor): 42 | assert value.dim() == 0 43 | value = value.item() 44 | self.value += value * count 45 | self.count += count 46 | 47 | def reset(self): 48 | self.value = 0.0 49 | self.count = 0.0 50 | 51 | @property 52 | def avg(self): 53 | return self.value / self.count 54 | 55 | 56 | def log_metrics(metrics, summary_logger, global_step): 57 | aggregated = dict() 58 | for name, v in metrics.items(): 59 | if isinstance(v, Accumulator): 60 | v = v.avg 61 | aggregated[name] = v 62 | summary_logger.log_metrics(metrics=aggregated, global_step=global_step) 63 | return aggregated 64 | 65 | 66 | def log_images(tag, img_tensor, n_row, summary_logger, global_step): 67 | metrics = defaultdict(Accumulator) 68 | img_tensor = torchvision.utils.make_grid(img_tensor, nrow=n_row) 69 | if img_tensor.dtype != torch.uint8: 70 | # Add 0.5 after unnormalizing to [0, 255] to round to nearest integer 71 | img_tensor = img_tensor.mul(255).add_(0.5).clamp_(0, 255) 72 | img_tensor = img_tensor.to(dtype=torch.uint8) 73 | metrics[tag] = {'tag': tag, 'img_tensor': img_tensor} 74 | log_metrics(metrics, summary_logger, global_step) 75 | -------------------------------------------------------------------------------- /utils/plt_blit.py: -------------------------------------------------------------------------------- 1 | class BlitManager: 2 | def __init__(self, ax, canvas, animated_artists=()): 3 | """ 4 | Parameters 5 | ---------- 6 | canvas : FigureCanvasAgg 7 | The canvas to work with, this only works for sub-classes of the Agg 8 | canvas which have the `~FigureCanvasAgg.copy_from_bbox` and 9 | `~FigureCanvasAgg.restore_region` methods. 10 | 11 | animated_artists : Iterable[Artist] 12 | List of the artists to manage 13 | """ 14 | self.ax = ax 15 | self.canvas = canvas 16 | self._bg = None 17 | self._artists = [] 18 | 19 | for a in animated_artists: 20 | self.add_artist(a) 21 | # grab the background on every draw 22 | self.cid = canvas.mpl_connect("draw_event", self.on_draw) 23 | 24 | def on_draw(self, event): 25 | """Callback to register with 'draw_event'.""" 26 | cv = self.canvas 27 | if event is not None: 28 | if event.canvas != cv: 29 | raise RuntimeError 30 | self._bg = cv.copy_from_bbox(cv.figure.bbox) 31 | self._draw_animated() 32 | 33 | def get_artists(self): 34 | return self._artists 35 | 36 | def add_artist(self, art): 37 | """ 38 | Add an artist to be managed. 39 | 40 | Parameters 41 | ---------- 42 | art : Artist 43 | 44 | The artist to be added. Will be set to 'animated' (just 45 | to be safe). *art* must be in the figure associated with 46 | the canvas this class is managing. 47 | 48 | """ 49 | if art.figure != self.canvas.figure: 50 | raise RuntimeError 51 | art.set_animated(True) 52 | self._artists.append(art) 53 | 54 | def _draw_animated(self): 55 | """Draw all of the animated artists.""" 56 | fig = self.canvas.figure 57 | for a in self._artists: 58 | fig.draw_artist(a) 59 | 60 | def update(self): 61 | """Update the screen with animated artists.""" 62 | cv = self.canvas 63 | fig = cv.figure 64 | # paranoia in case we missed the draw event, 65 | if self._bg is None: 66 | self.on_draw(None) 67 | else: 68 | # restore the background 69 | cv.restore_region(self._bg) 70 | # draw all of the animated artists 71 | self._draw_animated() 72 | # update the GUI state 73 | cv.blit(fig.bbox) 74 | # let the GUI event loop process anything it has to do 75 | cv.flush_events() 76 | -------------------------------------------------------------------------------- /utils/transformations.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | 6 | from utils.coords_utils import normalize_strokes 7 | 8 | ALLOWED_TRANSFORMATIONS = ['identity', 'translate_x', 'translate_y', 'translate', 'scale', 'scale_full_bounded', 'scale_full', 'uni_scale', 'proportion_scale', 'shear_x', 'shear_y', 'rotate'] 9 | 10 | 11 | class CosSin(nn.Module): 12 | def __init__(self, dim): 13 | super(CosSin, self).__init__() 14 | self.dim = dim 15 | 16 | def forward(self, input): 17 | return torch.cat([torch.cos(input), torch.sin(input)], dim=self.dim) 18 | 19 | 20 | class ToTransformMatrix(nn.Module): 21 | def __init__(self, transformation, no_activation=False, bound=None): 22 | super(ToTransformMatrix, self).__init__() 23 | 24 | assert transformation in ALLOWED_TRANSFORMATIONS 25 | 26 | 27 | if transformation == 'identity': 28 | mul = [ 29 | [[0, 0, 0], 30 | [0, 0, 0], 31 | [0, 0, 0]] 32 | ] 33 | bias = [[1, 0, 0], 34 | [0, 1, 0], 35 | [0, 0, 1]] 36 | 37 | def identity(x): 38 | return x 39 | activation = identity 40 | 41 | elif transformation == 'translate': 42 | mul = [ 43 | [[0, 0, 1], 44 | [0, 0, 0], 45 | [0, 0, 0]], 46 | 47 | [[0, 0, 0], 48 | [0, 0, 1], 49 | [0, 0, 0]] 50 | ] 51 | bias = [[1, 0, 0], 52 | [0, 1, 0], 53 | [0, 0, 1]] 54 | #activation = nn.Tanh() 55 | def translate_act(x): 56 | return torch.tanh(x)# * 0.9 57 | activation = translate_act 58 | 59 | elif transformation == 'translate_x': 60 | mul = [ 61 | [[0, 0, 1], 62 | [0, 0, 0], 63 | [0, 0, 0]] 64 | ] 65 | bias = [[1, 0, 0], 66 | [0, 1, 0], 67 | [0, 0, 1]] 68 | #activation = nn.Tanh() 69 | def translate_act(x): 70 | return torch.tanh(x)# * 0.9 71 | activation = translate_act 72 | 73 | elif transformation == 'translate_y': 74 | mul = [ 75 | [[0, 0, 0], 76 | [0, 0, 1], 77 | [0, 0, 0]] 78 | ] 79 | bias = [[1, 0, 0], 80 | [0, 1, 0], 81 | [0, 0, 1]] 82 | #activation = nn.Tanh() 83 | def translate_act(x): 84 | return torch.tanh(x)# * 0.9 85 | activation = translate_act 86 | 87 | elif transformation == 'scale_full': 88 | mul = [ 89 | [[1, 0, 0], 90 | [0, 0, 0], 91 | [0, 0, 0]], 92 | 93 | [[0, 0, 0], 94 | [0, 1, 0], 95 | [0, 0, 0]] 96 | ] 97 | 98 | bias = [[0, 0, 0], 99 | [0, 0, 0], 100 | [0, 0, 1]] 101 | 102 | activation = torch.exp 103 | 104 | elif transformation == 'scale_full_bounded': 105 | mul = [ 106 | [[1, 0, 0], 107 | [0, 0, 0], 108 | [0, 0, 0]], 109 | 110 | [[0, 0, 0], 111 | [0, 1, 0], 112 | [0, 0, 0]] 113 | ] 114 | 115 | bias = [[0, 0, 0], 116 | [0, 0, 0], 117 | [0, 0, 1]] 118 | 119 | def scale_act(x): 120 | return torch.exp(x) + (bound if bound is not None else 0.) 121 | activation = scale_act 122 | 123 | elif transformation == 'scale': 124 | mul = [ 125 | [[1, 0, 0], 126 | [0, 0, 0], 127 | [0, 0, 0]], 128 | 129 | [[0, 0, 0], 130 | [0, 1, 0], 131 | [0, 0, 0]] 132 | ] 133 | 134 | bias = [[0, 0, 0], 135 | [0, 0, 0], 136 | [0, 0, 1]] 137 | 138 | def scale_act(x): 139 | return 2. * torch.sigmoid(x) + 0.1 140 | activation = scale_act 141 | 142 | elif transformation == 'uni_scale': 143 | mul = [ 144 | [[1, 0, 0], 145 | [0, 1, 0], 146 | [0, 0, 0]] 147 | ] 148 | 149 | bias = [[0, 0, 0], 150 | [0, 0, 0], 151 | [0, 0, 1]] 152 | 153 | def scale_act(x): 154 | return 2. * torch.sigmoid(x) + 0.02 155 | activation = scale_act 156 | 157 | elif transformation == 'proportion_scale': 158 | mul = [ 159 | [[-1, 0, 0], 160 | [0, 0, 0], 161 | [0, 0, 0]], 162 | 163 | [[0, 0, 0], 164 | [0, -1, 0], 165 | [0, 0, 0]] 166 | ] 167 | 168 | bias = [[1. + (bound if bound is not None else 0.02), 0, 0], 169 | [0, 1. + (bound if bound is not None else 0.02) , 0], 170 | [0, 0, 1]] 171 | 172 | def scale_act(x): 173 | return F.relu(torch.tanh(x)) 174 | activation = scale_act 175 | 176 | elif transformation == 'shear_x': 177 | mul = [ 178 | [[0, 1, 0], 179 | [0, 0, 0], 180 | [0, 0, 0]] 181 | ] 182 | bias = [[1, 0, 0], 183 | [0, 1, 0], 184 | [0, 0, 1]] 185 | activation = nn.Tanh() 186 | 187 | elif transformation == 'shear_y': 188 | mul = [ 189 | [[0, 0, 0], 190 | [1, 0, 0], 191 | [0, 0, 0]] 192 | ] 193 | bias = [[1, 0, 0], 194 | [0, 1, 0], 195 | [0, 0, 1]] 196 | activation = nn.Tanh() 197 | 198 | elif transformation == 'rotate': 199 | mul = [ 200 | [[1, 0, 0], 201 | [0, 1, 0], 202 | [0, 0, 0]], 203 | 204 | [[0, -1, 0], 205 | [1, 0, 0], 206 | [0, 0, 0]] 207 | ] 208 | bias = [[0, 0, 0], 209 | [0, 0, 0], 210 | [0, 0, 1]] 211 | activation = CosSin(dim=2) 212 | 213 | self.mul = nn.Parameter(torch.tensor(mul).float().unsqueeze(0).unsqueeze(1), 214 | requires_grad=False) 215 | self.bias = nn.Parameter(torch.tensor(bias).float().unsqueeze(0).unsqueeze(1), 216 | requires_grad=False) 217 | if no_activation: 218 | self.activation = None 219 | else: 220 | self.activation = activation 221 | 222 | def forward(self, input): 223 | if not (self.activation is None): 224 | input = self.activation(input) 225 | return (input.unsqueeze(3).unsqueeze(4) * self.mul).sum(2) + self.bias 226 | 227 | 228 | def add_edge_points(points, n_points): 229 | points = np.array(points) 230 | new_points = [] 231 | pnts_per_egdge = (n_points-1) // (len(points)-1) 232 | extra_points = (n_points-1) % (len(points)-1) 233 | for i in range(len(points)-1): 234 | if extra_points > 0: 235 | extra_points -= 1 236 | ep = 1 237 | else: 238 | ep = 0 239 | edge_points = np.linspace(points[i], points[i+1], pnts_per_egdge+ep, endpoint=False) 240 | new_points.append(edge_points) 241 | new_points.append(points[-1][np.newaxis]) 242 | new_points = np.concatenate(new_points) 243 | return new_points 244 | 245 | 246 | def create_circle(n_points, scale=0.5, percent=1.): 247 | x = np.linspace(0., 2*np.pi*percent, n_points) 248 | points = np.stack([np.sin(x), np.cos(x)], axis=1) 249 | points = points * scale 250 | return points 251 | 252 | 253 | def get_predefined_primitives(primitive_names, n_points=25, normalize=True, 254 | scale=1.): 255 | points = [] 256 | 257 | for pname in primitive_names: 258 | # Square 259 | if pname == 'square': 260 | pnts = [[-0.5, -0.5], [-0.5, 0.5], [0.5, 0.5], [0.5, -0.5], [-0.5, -0.5]] 261 | pnts = add_edge_points(pnts, n_points) 262 | points.append(torch.from_numpy(pnts).float()) 263 | 264 | # Triangle 265 | elif pname == 'triangle': 266 | pnts = [[-0.5, 0.44], [0.5, 0.44], [0., -0.44], [-0.5, 0.44]] 267 | pnts = add_edge_points(pnts, n_points) 268 | points.append(torch.from_numpy(pnts).float()) 269 | 270 | # Circle 271 | elif pname == 'circle': 272 | pnts = create_circle(n_points) 273 | points.append(torch.from_numpy(pnts).float()) 274 | 275 | # Half-circle 276 | elif pname == 'halfcircle': 277 | pnts = create_circle(n_points, percent=0.5) 278 | points.append(torch.from_numpy(pnts).float()) 279 | 280 | # Line 281 | elif pname == 'line': 282 | pnts = [[-0.5, 0.], [0.5, 0.]] 283 | pnts = add_edge_points(pnts, n_points) 284 | points.append(torch.from_numpy(pnts).float()) 285 | 286 | # Corner 287 | elif pname == 'corner': 288 | pnts = [[-0.5, -0.5], [-0.5, 0.5], [0.5, 0.5]] 289 | pnts = add_edge_points(pnts, n_points) 290 | points.append(torch.from_numpy(pnts).float()) 291 | 292 | # u 293 | elif pname == 'u': 294 | pnts = [[-0.5, -0.5], [-0.5, 0.5], [0.5, 0.5], [0.5, -0.5]] 295 | pnts = add_edge_points(pnts, n_points) 296 | points.append(torch.from_numpy(pnts).float()) 297 | 298 | # dot (single point 299 | elif pname == 'dot': 300 | pnts = [[0., 0.]] 301 | pnts = np.array(pnts).repeat(n_points, axis=0) 302 | points.append(torch.from_numpy(pnts).float()) 303 | 304 | primitives = torch.stack(points) 305 | if normalize: 306 | primitives, _, _ = normalize_strokes(primitives) 307 | 308 | if scale != 1.: 309 | primitives = primitives * scale 310 | 311 | return primitives 312 | 313 | def compute_full_transformation(transform_layers, transformation_parameters): 314 | mat = None 315 | for t_layer, params in zip(transform_layers, transformation_parameters): 316 | t_mat = t_layer(params) 317 | 318 | if mat is None: 319 | mat = t_mat 320 | else: 321 | mat = t_mat @ mat 322 | 323 | return mat 324 | 325 | 326 | def transform_primitives(primitives, transform_layers, transforms_params, prim_ids=None): 327 | full_transforms = compute_full_transformation(transform_layers, transforms_params) 328 | 329 | if prim_ids is None: 330 | if primitives.size(0) != transforms_params[0].size(0): 331 | full_prim_htan = primitives.unsqueeze(0).expand([transforms_params[0].size(0)]+list(primitives.size())) 332 | else: 333 | full_prim_htan = primitives.unsqueeze(2) 334 | else: 335 | full_prim_htan = primitives.unsqueeze(0).repeat(prim_ids.size(0), 1, 1, 1).gather(1, prim_ids.repeat(1, 1, primitives.size(1), primitives.size(2))) 336 | 337 | full_prim_htan = torch.cat( 338 | [full_prim_htan, torch.ones(full_prim_htan.size()[:-1]).unsqueeze(-1).to(full_prim_htan.device)], dim=-1) 339 | 340 | full_prim_htan = full_prim_htan.unsqueeze(-2) @ full_transforms.unsqueeze(-3).transpose(-2, -1) 341 | 342 | return full_prim_htan 343 | --------------------------------------------------------------------------------