├── .gitignore ├── LICENSE ├── README.md ├── download_models.sh ├── images ├── boy.png ├── cartoon2.jpg ├── dogsketch.png ├── flowers.png ├── girl2.png ├── girl3.jpg ├── girlmrf.jpg ├── girlsketch.jpg ├── goat.png ├── kid5.png ├── megan.png ├── muse.jpg ├── smallworldI.jpg └── starry_night.jpg ├── lap_style.lua ├── laplacian.py ├── logstat.py ├── multi-lap-test.sh ├── neural-doodle └── doodle.py ├── output ├── boy_girl3_5_0.png ├── boy_girl3_5_100.png ├── combinations │ ├── girlmrf_dogsketch5_1,2_50,100.png │ ├── girlmrf_dogsketch5_1,3_50,400.png │ ├── girlmrf_dogsketch5_1_50.png │ ├── girlmrf_dogsketch5_2,3_100,400.png │ ├── girlmrf_dogsketch5_2,4_100,1600.png │ ├── girlmrf_dogsketch5_2_100.png │ ├── girlmrf_dogsketch_5_0.png │ ├── girlmrf_dogsketch_5_100.png │ └── girlmrf_dogsketch_5_200.png ├── girl2_cartoon2_20_0.jpg ├── girl2_cartoon2_20_100.png ├── girlmrf_girlsketch_5_0.png ├── girlmrf_girlsketch_5_100.png ├── girlmrf_smallworldI_20_0.png ├── girlmrf_smallworldI_20_200.png ├── goat_muse20_1_0.png ├── goat_muse20_2_100.png ├── kid5_smallworldI_20_0.png ├── kid5_smallworldI_20_200.png ├── laplacians │ ├── girl2_car2_0_edge.png │ ├── girl2_car2_edge.png │ └── girl2_edge.png ├── megan_flowers20_0.png ├── megan_flowers20_100.png ├── megan_starry10_0.png └── megan_starry10_100.png ├── single-lap-exp.sh └── tf-neural-style ├── neural_style.py ├── stylize.py └── vgg.py /.gitignore: -------------------------------------------------------------------------------- 1 | commit.bat 2 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU AFFERO GENERAL PUBLIC LICENSE 2 | Version 3, 19 November 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 Affero General Public License is a free, copyleft license for 11 | software and other kinds of works, specifically designed to ensure 12 | cooperation with the community in the case of network server software. 13 | 14 | The licenses for most software and other practical works are designed 15 | to take away your freedom to share and change the works. By contrast, 16 | our General Public Licenses are intended to guarantee your freedom to 17 | share and change all versions of a program--to make sure it remains free 18 | software for all its users. 19 | 20 | When we speak of free software, we are referring to freedom, not 21 | price. Our General Public Licenses are designed to make sure that you 22 | have the freedom to distribute copies of free software (and charge for 23 | them if you wish), that you receive source code or can get it if you 24 | want it, that you can change the software or use pieces of it in new 25 | free programs, and that you know you can do these things. 26 | 27 | Developers that use our General Public Licenses protect your rights 28 | with two steps: (1) assert copyright on the software, and (2) offer 29 | you this License which gives you legal permission to copy, distribute 30 | and/or modify the software. 31 | 32 | A secondary benefit of defending all users' freedom is that 33 | improvements made in alternate versions of the program, if they 34 | receive widespread use, become available for other developers to 35 | incorporate. Many developers of free software are heartened and 36 | encouraged by the resulting cooperation. However, in the case of 37 | software used on network servers, this result may fail to come about. 38 | The GNU General Public License permits making a modified version and 39 | letting the public access it on a server without ever releasing its 40 | source code to the public. 41 | 42 | The GNU Affero General Public License is designed specifically to 43 | ensure that, in such cases, the modified source code becomes available 44 | to the community. It requires the operator of a network server to 45 | provide the source code of the modified version running there to the 46 | users of that server. Therefore, public use of a modified version, on 47 | a publicly accessible server, gives the public access to the source 48 | code of the modified version. 49 | 50 | An older license, called the Affero General Public License and 51 | published by Affero, was designed to accomplish similar goals. This is 52 | a different license, not a version of the Affero GPL, but Affero has 53 | released a new version of the Affero GPL which permits relicensing under 54 | this license. 55 | 56 | The precise terms and conditions for copying, distribution and 57 | modification follow. 58 | 59 | TERMS AND CONDITIONS 60 | 61 | 0. Definitions. 62 | 63 | "This License" refers to version 3 of the GNU Affero General Public License. 64 | 65 | "Copyright" also means copyright-like laws that apply to other kinds of 66 | works, such as semiconductor masks. 67 | 68 | "The Program" refers to any copyrightable work licensed under this 69 | License. Each licensee is addressed as "you". "Licensees" and 70 | "recipients" may be individuals or organizations. 71 | 72 | To "modify" a work means to copy from or adapt all or part of the work 73 | in a fashion requiring copyright permission, other than the making of an 74 | exact copy. The resulting work is called a "modified version" of the 75 | earlier work or a work "based on" the earlier work. 76 | 77 | A "covered work" means either the unmodified Program or a work based 78 | on the Program. 79 | 80 | To "propagate" a work means to do anything with it that, without 81 | permission, would make you directly or secondarily liable for 82 | infringement under applicable copyright law, except executing it on a 83 | computer or modifying a private copy. Propagation includes copying, 84 | distribution (with or without modification), making available to the 85 | public, and in some countries other activities as well. 86 | 87 | To "convey" a work means any kind of propagation that enables other 88 | parties to make or receive copies. Mere interaction with a user through 89 | a computer network, with no transfer of a copy, is not conveying. 90 | 91 | An interactive user interface displays "Appropriate Legal Notices" 92 | to the extent that it includes a convenient and prominently visible 93 | feature that (1) displays an appropriate copyright notice, and (2) 94 | tells the user that there is no warranty for the work (except to the 95 | extent that warranties are provided), that licensees may convey the 96 | work under this License, and how to view a copy of this License. If 97 | the interface presents a list of user commands or options, such as a 98 | menu, a prominent item in the list meets this criterion. 99 | 100 | 1. Source Code. 101 | 102 | The "source code" for a work means the preferred form of the work 103 | for making modifications to it. "Object code" means any non-source 104 | form of a work. 105 | 106 | A "Standard Interface" means an interface that either is an official 107 | standard defined by a recognized standards body, or, in the case of 108 | interfaces specified for a particular programming language, one that 109 | is widely used among developers working in that language. 110 | 111 | The "System Libraries" of an executable work include anything, other 112 | than the work as a whole, that (a) is included in the normal form of 113 | packaging a Major Component, but which is not part of that Major 114 | Component, and (b) serves only to enable use of the work with that 115 | Major Component, or to implement a Standard Interface for which an 116 | implementation is available to the public in source code form. A 117 | "Major Component", in this context, means a major essential component 118 | (kernel, window system, and so on) of the specific operating system 119 | (if any) on which the executable work runs, or a compiler used to 120 | produce the work, or an object code interpreter used to run it. 121 | 122 | The "Corresponding Source" for a work in object code form means all 123 | the source code needed to generate, install, and (for an executable 124 | work) run the object code and to modify the work, including scripts to 125 | control those activities. However, it does not include the work's 126 | System Libraries, or general-purpose tools or generally available free 127 | programs which are used unmodified in performing those activities but 128 | which are not part of the work. For example, Corresponding Source 129 | includes interface definition files associated with source files for 130 | the work, and the source code for shared libraries and dynamically 131 | linked subprograms that the work is specifically designed to require, 132 | such as by intimate data communication or control flow between those 133 | subprograms and other parts of the work. 134 | 135 | The Corresponding Source need not include anything that users 136 | can regenerate automatically from other parts of the Corresponding 137 | Source. 138 | 139 | The Corresponding Source for a work in source code form is that 140 | same work. 141 | 142 | 2. Basic Permissions. 143 | 144 | All rights granted under this License are granted for the term of 145 | copyright on the Program, and are irrevocable provided the stated 146 | conditions are met. This License explicitly affirms your unlimited 147 | permission to run the unmodified Program. The output from running a 148 | covered work is covered by this License only if the output, given its 149 | content, constitutes a covered work. This License acknowledges your 150 | rights of fair use or other equivalent, as provided by copyright law. 151 | 152 | You may make, run and propagate covered works that you do not 153 | convey, without conditions so long as your license otherwise remains 154 | in force. You may convey covered works to others for the sole purpose 155 | of having them make modifications exclusively for you, or provide you 156 | with facilities for running those works, provided that you comply with 157 | the terms of this License in conveying all material for which you do 158 | not control copyright. Those thus making or running the covered works 159 | for you must do so exclusively on your behalf, under your direction 160 | and control, on terms that prohibit them from making any copies of 161 | your copyrighted material outside their relationship with you. 162 | 163 | Conveying under any other circumstances is permitted solely under 164 | the conditions stated below. Sublicensing is not allowed; section 10 165 | makes it unnecessary. 166 | 167 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 168 | 169 | No covered work shall be deemed part of an effective technological 170 | measure under any applicable law fulfilling obligations under article 171 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 172 | similar laws prohibiting or restricting circumvention of such 173 | measures. 174 | 175 | When you convey a covered work, you waive any legal power to forbid 176 | circumvention of technological measures to the extent such circumvention 177 | is effected by exercising rights under this License with respect to 178 | the covered work, and you disclaim any intention to limit operation or 179 | modification of the work as a means of enforcing, against the work's 180 | users, your or third parties' legal rights to forbid circumvention of 181 | technological measures. 182 | 183 | 4. Conveying Verbatim Copies. 184 | 185 | You may convey verbatim copies of the Program's source code as you 186 | receive it, in any medium, provided that you conspicuously and 187 | appropriately publish on each copy an appropriate copyright notice; 188 | keep intact all notices stating that this License and any 189 | non-permissive terms added in accord with section 7 apply to the code; 190 | keep intact all notices of the absence of any warranty; and give all 191 | recipients a copy of this License along with the Program. 192 | 193 | You may charge any price or no price for each copy that you convey, 194 | and you may offer support or warranty protection for a fee. 195 | 196 | 5. Conveying Modified Source Versions. 197 | 198 | You may convey a work based on the Program, or the modifications to 199 | produce it from the Program, in the form of source code under the 200 | terms of section 4, provided that you also meet all of these conditions: 201 | 202 | a) The work must carry prominent notices stating that you modified 203 | it, and giving a relevant date. 204 | 205 | b) The work must carry prominent notices stating that it is 206 | released under this License and any conditions added under section 207 | 7. This requirement modifies the requirement in section 4 to 208 | "keep intact all notices". 209 | 210 | c) You must license the entire work, as a whole, under this 211 | License to anyone who comes into possession of a copy. This 212 | License will therefore apply, along with any applicable section 7 213 | additional terms, to the whole of the work, and all its parts, 214 | regardless of how they are packaged. This License gives no 215 | permission to license the work in any other way, but it does not 216 | invalidate such permission if you have separately received it. 217 | 218 | d) If the work has interactive user interfaces, each must display 219 | Appropriate Legal Notices; however, if the Program has interactive 220 | interfaces that do not display Appropriate Legal Notices, your 221 | work need not make them do so. 222 | 223 | A compilation of a covered work with other separate and independent 224 | works, which are not by their nature extensions of the covered work, 225 | and which are not combined with it such as to form a larger program, 226 | in or on a volume of a storage or distribution medium, is called an 227 | "aggregate" if the compilation and its resulting copyright are not 228 | used to limit the access or legal rights of the compilation's users 229 | beyond what the individual works permit. Inclusion of a covered work 230 | in an aggregate does not cause this License to apply to the other 231 | parts of the aggregate. 232 | 233 | 6. Conveying Non-Source Forms. 234 | 235 | You may convey a covered work in object code form under the terms 236 | of sections 4 and 5, provided that you also convey the 237 | machine-readable Corresponding Source under the terms of this License, 238 | in one of these ways: 239 | 240 | a) Convey the object code in, or embodied in, a physical product 241 | (including a physical distribution medium), accompanied by the 242 | Corresponding Source fixed on a durable physical medium 243 | customarily used for software interchange. 244 | 245 | b) Convey the object code in, or embodied in, a physical product 246 | (including a physical distribution medium), accompanied by a 247 | written offer, valid for at least three years and valid for as 248 | long as you offer spare parts or customer support for that product 249 | model, to give anyone who possesses the object code either (1) a 250 | copy of the Corresponding Source for all the software in the 251 | product that is covered by this License, on a durable physical 252 | medium customarily used for software interchange, for a price no 253 | more than your reasonable cost of physically performing this 254 | conveying of source, or (2) access to copy the 255 | Corresponding Source from a network server at no charge. 256 | 257 | c) Convey individual copies of the object code with a copy of the 258 | written offer to provide the Corresponding Source. This 259 | alternative is allowed only occasionally and noncommercially, and 260 | only if you received the object code with such an offer, in accord 261 | with subsection 6b. 262 | 263 | d) Convey the object code by offering access from a designated 264 | place (gratis or for a charge), and offer equivalent access to the 265 | Corresponding Source in the same way through the same place at no 266 | further charge. You need not require recipients to copy the 267 | Corresponding Source along with the object code. If the place to 268 | copy the object code is a network server, the Corresponding Source 269 | may be on a different server (operated by you or a third party) 270 | that supports equivalent copying facilities, provided you maintain 271 | clear directions next to the object code saying where to find the 272 | Corresponding Source. Regardless of what server hosts the 273 | Corresponding Source, you remain obligated to ensure that it is 274 | available for as long as needed to satisfy these requirements. 275 | 276 | e) Convey the object code using peer-to-peer transmission, provided 277 | you inform other peers where the object code and Corresponding 278 | Source of the work are being offered to the general public at no 279 | charge under subsection 6d. 280 | 281 | A separable portion of the object code, whose source code is excluded 282 | from the Corresponding Source as a System Library, need not be 283 | included in conveying the object code work. 284 | 285 | A "User Product" is either (1) a "consumer product", which means any 286 | tangible personal property which is normally used for personal, family, 287 | or household purposes, or (2) anything designed or sold for incorporation 288 | into a dwelling. In determining whether a product is a consumer product, 289 | doubtful cases shall be resolved in favor of coverage. For a particular 290 | product received by a particular user, "normally used" refers to a 291 | typical or common use of that class of product, regardless of the status 292 | of the particular user or of the way in which the particular user 293 | actually uses, or expects or is expected to use, the product. A product 294 | is a consumer product regardless of whether the product has substantial 295 | commercial, industrial or non-consumer uses, unless such uses represent 296 | the only significant mode of use of the product. 297 | 298 | "Installation Information" for a User Product means any methods, 299 | procedures, authorization keys, or other information required to install 300 | and execute modified versions of a covered work in that User Product from 301 | a modified version of its Corresponding Source. The information must 302 | suffice to ensure that the continued functioning of the modified object 303 | code is in no case prevented or interfered with solely because 304 | modification has been made. 305 | 306 | If you convey an object code work under this section in, or with, or 307 | specifically for use in, a User Product, and the conveying occurs as 308 | part of a transaction in which the right of possession and use of the 309 | User Product is transferred to the recipient in perpetuity or for a 310 | fixed term (regardless of how the transaction is characterized), the 311 | Corresponding Source conveyed under this section must be accompanied 312 | by the Installation Information. But this requirement does not apply 313 | if neither you nor any third party retains the ability to install 314 | modified object code on the User Product (for example, the work has 315 | been installed in ROM). 316 | 317 | The requirement to provide Installation Information does not include a 318 | requirement to continue to provide support service, warranty, or updates 319 | for a work that has been modified or installed by the recipient, or for 320 | the User Product in which it has been modified or installed. Access to a 321 | network may be denied when the modification itself materially and 322 | adversely affects the operation of the network or violates the rules and 323 | protocols for communication across the network. 324 | 325 | Corresponding Source conveyed, and Installation Information provided, 326 | in accord with this section must be in a format that is publicly 327 | documented (and with an implementation available to the public in 328 | source code form), and must require no special password or key for 329 | unpacking, reading or copying. 330 | 331 | 7. Additional Terms. 332 | 333 | "Additional permissions" are terms that supplement the terms of this 334 | License by making exceptions from one or more of its conditions. 335 | Additional permissions that are applicable to the entire Program shall 336 | be treated as though they were included in this License, to the extent 337 | that they are valid under applicable law. If additional permissions 338 | apply only to part of the Program, that part may be used separately 339 | under those permissions, but the entire Program remains governed by 340 | this License without regard to the additional permissions. 341 | 342 | When you convey a copy of a covered work, you may at your option 343 | remove any additional permissions from that copy, or from any part of 344 | it. (Additional permissions may be written to require their own 345 | removal in certain cases when you modify the work.) You may place 346 | additional permissions on material, added by you to a covered work, 347 | for which you have or can give appropriate copyright permission. 348 | 349 | Notwithstanding any other provision of this License, for material you 350 | add to a covered work, you may (if authorized by the copyright holders of 351 | that material) supplement the terms of this License with terms: 352 | 353 | a) Disclaiming warranty or limiting liability differently from the 354 | terms of sections 15 and 16 of this License; or 355 | 356 | b) Requiring preservation of specified reasonable legal notices or 357 | author attributions in that material or in the Appropriate Legal 358 | Notices displayed by works containing it; or 359 | 360 | c) Prohibiting misrepresentation of the origin of that material, or 361 | requiring that modified versions of such material be marked in 362 | reasonable ways as different from the original version; or 363 | 364 | d) Limiting the use for publicity purposes of names of licensors or 365 | authors of the material; or 366 | 367 | e) Declining to grant rights under trademark law for use of some 368 | trade names, trademarks, or service marks; or 369 | 370 | f) Requiring indemnification of licensors and authors of that 371 | material by anyone who conveys the material (or modified versions of 372 | it) with contractual assumptions of liability to the recipient, for 373 | any liability that these contractual assumptions directly impose on 374 | those licensors and authors. 375 | 376 | All other non-permissive additional terms are considered "further 377 | restrictions" within the meaning of section 10. If the Program as you 378 | received it, or any part of it, contains a notice stating that it is 379 | governed by this License along with a term that is a further 380 | restriction, you may remove that term. If a license document contains 381 | a further restriction but permits relicensing or conveying under this 382 | License, you may add to a covered work material governed by the terms 383 | of that license document, provided that the further restriction does 384 | not survive such relicensing or conveying. 385 | 386 | If you add terms to a covered work in accord with this section, you 387 | must place, in the relevant source files, a statement of the 388 | additional terms that apply to those files, or a notice indicating 389 | where to find the applicable terms. 390 | 391 | Additional terms, permissive or non-permissive, may be stated in the 392 | form of a separately written license, or stated as exceptions; 393 | the above requirements apply either way. 394 | 395 | 8. Termination. 396 | 397 | You may not propagate or modify a covered work except as expressly 398 | provided under this License. Any attempt otherwise to propagate or 399 | modify it is void, and will automatically terminate your rights under 400 | this License (including any patent licenses granted under the third 401 | paragraph of section 11). 402 | 403 | However, if you cease all violation of this License, then your 404 | license from a particular copyright holder is reinstated (a) 405 | provisionally, unless and until the copyright holder explicitly and 406 | finally terminates your license, and (b) permanently, if the copyright 407 | holder fails to notify you of the violation by some reasonable means 408 | prior to 60 days after the cessation. 409 | 410 | Moreover, your license from a particular copyright holder is 411 | reinstated permanently if the copyright holder notifies you of the 412 | violation by some reasonable means, this is the first time you have 413 | received notice of violation of this License (for any work) from that 414 | copyright holder, and you cure the violation prior to 30 days after 415 | your receipt of the notice. 416 | 417 | Termination of your rights under this section does not terminate the 418 | licenses of parties who have received copies or rights from you under 419 | this License. If your rights have been terminated and not permanently 420 | reinstated, you do not qualify to receive new licenses for the same 421 | material under section 10. 422 | 423 | 9. Acceptance Not Required for Having Copies. 424 | 425 | You are not required to accept this License in order to receive or 426 | run a copy of the Program. Ancillary propagation of a covered work 427 | occurring solely as a consequence of using peer-to-peer transmission 428 | to receive a copy likewise does not require acceptance. However, 429 | nothing other than this License grants you permission to propagate or 430 | modify any covered work. These actions infringe copyright if you do 431 | not accept this License. Therefore, by modifying or propagating a 432 | covered work, you indicate your acceptance of this License to do so. 433 | 434 | 10. Automatic Licensing of Downstream Recipients. 435 | 436 | Each time you convey a covered work, the recipient automatically 437 | receives a license from the original licensors, to run, modify and 438 | propagate that work, subject to this License. You are not responsible 439 | for enforcing compliance by third parties with this License. 440 | 441 | An "entity transaction" is a transaction transferring control of an 442 | organization, or substantially all assets of one, or subdividing an 443 | organization, or merging organizations. If propagation of a covered 444 | work results from an entity transaction, each party to that 445 | transaction who receives a copy of the work also receives whatever 446 | licenses to the work the party's predecessor in interest had or could 447 | give under the previous paragraph, plus a right to possession of the 448 | Corresponding Source of the work from the predecessor in interest, if 449 | the predecessor has it or can get it with reasonable efforts. 450 | 451 | You may not impose any further restrictions on the exercise of the 452 | rights granted or affirmed under this License. For example, you may 453 | not impose a license fee, royalty, or other charge for exercise of 454 | rights granted under this License, and you may not initiate litigation 455 | (including a cross-claim or counterclaim in a lawsuit) alleging that 456 | any patent claim is infringed by making, using, selling, offering for 457 | sale, or importing the Program or any portion of it. 458 | 459 | 11. Patents. 460 | 461 | A "contributor" is a copyright holder who authorizes use under this 462 | License of the Program or a work on which the Program is based. The 463 | work thus licensed is called the contributor's "contributor version". 464 | 465 | A contributor's "essential patent claims" are all patent claims 466 | owned or controlled by the contributor, whether already acquired or 467 | hereafter acquired, that would be infringed by some manner, permitted 468 | by this License, of making, using, or selling its contributor version, 469 | but do not include claims that would be infringed only as a 470 | consequence of further modification of the contributor version. For 471 | purposes of this definition, "control" includes the right to grant 472 | patent sublicenses in a manner consistent with the requirements of 473 | this License. 474 | 475 | Each contributor grants you a non-exclusive, worldwide, royalty-free 476 | patent license under the contributor's essential patent claims, to 477 | make, use, sell, offer for sale, import and otherwise run, modify and 478 | propagate the contents of its contributor version. 479 | 480 | In the following three paragraphs, a "patent license" is any express 481 | agreement or commitment, however denominated, not to enforce a patent 482 | (such as an express permission to practice a patent or covenant not to 483 | sue for patent infringement). To "grant" such a patent license to a 484 | party means to make such an agreement or commitment not to enforce a 485 | patent against the party. 486 | 487 | If you convey a covered work, knowingly relying on a patent license, 488 | and the Corresponding Source of the work is not available for anyone 489 | to copy, free of charge and under the terms of this License, through a 490 | publicly available network server or other readily accessible means, 491 | then you must either (1) cause the Corresponding Source to be so 492 | available, or (2) arrange to deprive yourself of the benefit of the 493 | patent license for this particular work, or (3) arrange, in a manner 494 | consistent with the requirements of this License, to extend the patent 495 | license to downstream recipients. "Knowingly relying" means you have 496 | actual knowledge that, but for the patent license, your conveying the 497 | covered work in a country, or your recipient's use of the covered work 498 | in a country, would infringe one or more identifiable patents in that 499 | country that you have reason to believe are valid. 500 | 501 | If, pursuant to or in connection with a single transaction or 502 | arrangement, you convey, or propagate by procuring conveyance of, a 503 | covered work, and grant a patent license to some of the parties 504 | receiving the covered work authorizing them to use, propagate, modify 505 | or convey a specific copy of the covered work, then the patent license 506 | you grant is automatically extended to all recipients of the covered 507 | work and works based on it. 508 | 509 | A patent license is "discriminatory" if it does not include within 510 | the scope of its coverage, prohibits the exercise of, or is 511 | conditioned on the non-exercise of one or more of the rights that are 512 | specifically granted under this License. You may not convey a covered 513 | work if you are a party to an arrangement with a third party that is 514 | in the business of distributing software, under which you make payment 515 | to the third party based on the extent of your activity of conveying 516 | the work, and under which the third party grants, to any of the 517 | parties who would receive the covered work from you, a discriminatory 518 | patent license (a) in connection with copies of the covered work 519 | conveyed by you (or copies made from those copies), or (b) primarily 520 | for and in connection with specific products or compilations that 521 | contain the covered work, unless you entered into that arrangement, 522 | or that patent license was granted, prior to 28 March 2007. 523 | 524 | Nothing in this License shall be construed as excluding or limiting 525 | any implied license or other defenses to infringement that may 526 | otherwise be available to you under applicable patent law. 527 | 528 | 12. No Surrender of Others' Freedom. 529 | 530 | If conditions are imposed on you (whether by court order, agreement or 531 | otherwise) that contradict the conditions of this License, they do not 532 | excuse you from the conditions of this License. If you cannot convey a 533 | covered work so as to satisfy simultaneously your obligations under this 534 | License and any other pertinent obligations, then as a consequence you may 535 | not convey it at all. For example, if you agree to terms that obligate you 536 | to collect a royalty for further conveying from those to whom you convey 537 | the Program, the only way you could satisfy both those terms and this 538 | License would be to refrain entirely from conveying the Program. 539 | 540 | 13. Remote Network Interaction; Use with the GNU General Public License. 541 | 542 | Notwithstanding any other provision of this License, if you modify the 543 | Program, your modified version must prominently offer all users 544 | interacting with it remotely through a computer network (if your version 545 | supports such interaction) an opportunity to receive the Corresponding 546 | Source of your version by providing access to the Corresponding Source 547 | from a network server at no charge, through some standard or customary 548 | means of facilitating copying of software. This Corresponding Source 549 | shall include the Corresponding Source for any work covered by version 3 550 | of the GNU General Public License that is incorporated pursuant to the 551 | following paragraph. 552 | 553 | Notwithstanding any other provision of this License, you have 554 | permission to link or combine any covered work with a work licensed 555 | under version 3 of the GNU General Public License into a single 556 | combined work, and to convey the resulting work. The terms of this 557 | License will continue to apply to the part which is the covered work, 558 | but the work with which it is combined will remain governed by version 559 | 3 of the GNU General Public License. 560 | 561 | 14. Revised Versions of this License. 562 | 563 | The Free Software Foundation may publish revised and/or new versions of 564 | the GNU Affero General Public License from time to time. Such new versions 565 | will be similar in spirit to the present version, but may differ in detail to 566 | address new problems or concerns. 567 | 568 | Each version is given a distinguishing version number. If the 569 | Program specifies that a certain numbered version of the GNU Affero General 570 | Public License "or any later version" applies to it, you have the 571 | option of following the terms and conditions either of that numbered 572 | version or of any later version published by the Free Software 573 | Foundation. If the Program does not specify a version number of the 574 | GNU Affero General Public License, you may choose any version ever published 575 | by the Free Software Foundation. 576 | 577 | If the Program specifies that a proxy can decide which future 578 | versions of the GNU Affero General Public License can be used, that proxy's 579 | public statement of acceptance of a version permanently authorizes you 580 | to choose that version for the Program. 581 | 582 | Later license versions may give you additional or different 583 | permissions. However, no additional obligations are imposed on any 584 | author or copyright holder as a result of your choosing to follow a 585 | later version. 586 | 587 | 15. Disclaimer of Warranty. 588 | 589 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 590 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 591 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 592 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 593 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 594 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 595 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 596 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 597 | 598 | 16. Limitation of Liability. 599 | 600 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 601 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 602 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 603 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 604 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 605 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 606 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 607 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 608 | SUCH DAMAGES. 609 | 610 | 17. Interpretation of Sections 15 and 16. 611 | 612 | If the disclaimer of warranty and limitation of liability provided 613 | above cannot be given local legal effect according to their terms, 614 | reviewing courts shall apply local law that most closely approximates 615 | an absolute waiver of all civil liability in connection with the 616 | Program, unless a warranty or assumption of liability accompanies a 617 | copy of the Program in return for a fee. 618 | 619 | END OF TERMS AND CONDITIONS 620 | 621 | How to Apply These Terms to Your New Programs 622 | 623 | If you develop a new program, and you want it to be of the greatest 624 | possible use to the public, the best way to achieve this is to make it 625 | free software which everyone can redistribute and change under these terms. 626 | 627 | To do so, attach the following notices to the program. It is safest 628 | to attach them to the start of each source file to most effectively 629 | state the exclusion of warranty; and each file should have at least 630 | the "copyright" line and a pointer to where the full notice is found. 631 | 632 | 633 | Copyright (C) 634 | 635 | This program is free software: you can redistribute it and/or modify 636 | it under the terms of the GNU Affero General Public License as published 637 | by the Free Software Foundation, either version 3 of the License, or 638 | (at your option) any later version. 639 | 640 | This program is distributed in the hope that it will be useful, 641 | but WITHOUT ANY WARRANTY; without even the implied warranty of 642 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 643 | GNU Affero General Public License for more details. 644 | 645 | You should have received a copy of the GNU Affero General Public License 646 | along with this program. If not, see . 647 | 648 | Also add information on how to contact you by electronic and paper mail. 649 | 650 | If your software can interact with users remotely through a computer 651 | network, you should also make sure that it provides a way for users to 652 | get its source. For example, if your program is a web application, its 653 | interface could display a "Source" link that leads users to an archive 654 | of the code. There are many ways you could offer source, and different 655 | solutions will be better for different programs; see section 13 for the 656 | specific requirements. 657 | 658 | You should also get your employer (if you work as a programmer) or school, 659 | if any, to sign a "copyright disclaimer" for the program, if necessary. 660 | For more information on this, and how to apply and follow the GNU AGPL, see 661 | . 662 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Lapstyle: Laplacian-Steered Neural Style Transfer 2 | Code and test images for the paper "[Laplacian-Steered Neural Style Transfer](https://arxiv.org/abs/1707.01253)". 3 | 4 | Lapstyle extends an existing neural style transfer method with one or multiple Laplacian loss layers. The following three neural style transfer implementations have been extended: 5 | 6 | * **lap_style.lua** (*Recommended!*) - https://github.com/jcjohnson/neural-style Gatys-style[1] implemented by Justin Johnson, using the L-BFGS optimization method. 7 | * **tf-neural-style/neural_style.py** - https://github.com/anishathalye/neural-style Gatys-style[1] by Anish Athalye, using Adam. 8 | * **neural-doodle/doodle.py** - https://github.com/alexjc/neural-doodle MRF-CNN[2] implemented by Alex J. Champandard. 9 | 10 | The implementation by Justin Johnson clearly produces the best images (either the original [neural_style.lua](https://github.com/jcjohnson/neural-style/blob/master/neural_style.lua) or the extended [lap_style.lua](https://github.com/askerlee/lapstyle/blob/master/lap_style.lua)). The corresponding content and style losses are also the smallest. Its superiority seems to be ascribed to the L-BFGS optimization, since the algorithm is otherwise identical to Anish Athalye's implementation. 11 | 12 | ## Setup: 13 | The setup procedures are the same as those of each original project. The following procedures for **lap_style.lua** are quoted from https://github.com/jcjohnson/neural-style: 14 |
15 | Dependencies: 16 | * [torch7](https://github.com/torch/torch7) 17 | * [loadcaffe](https://github.com/szagoruyko/loadcaffe) 18 | 19 | Optional dependencies: 20 | * For CUDA backend: 21 | * CUDA 6.5+ 22 | * [cunn](https://github.com/torch/cunn) 23 | * For cuDNN backend: 24 | * [cudnn.torch](https://github.com/soumith/cudnn.torch) 25 | * For OpenCL backend: 26 | * [cltorch](https://github.com/hughperkins/cltorch) 27 | * [clnn](https://github.com/hughperkins/clnn) 28 | 29 | After installing dependencies, you'll need to run the following script to download the VGG model: 30 | ``` 31 | sh models/download_models.sh 32 | ``` 33 | This will download the original [VGG-19 model](https://gist.github.com/ksimonyan/3785162f95cd2d5fee77#file-readme-md). 34 |
35 | 36 | ### Sample usage: 37 | ``` 38 | th lap_style.lua -style_image images/flowers.png -content_image images/megan.png -output_image output/megan_flowers20_100.png -content_weight 20 -lap_layers 2 -lap_weights 100 39 | ``` 40 | 41 | ### Sample images: 42 |

43 | 44 |
45 | 46 | 47 |

48 |

49 | 50 |
51 | 52 | 53 |

54 | 55 | The four images in each group are: 1) content image, 2) style image, 3) image synthesized with the original Gatys-style, and 4) image synthesized with Lapstyle. 56 | 57 | Note: although photo-realistic style transfer[3] (https://github.com/luanfujun/deep-photo-styletransfer) performs amazingly well on their test images, it doesn't work on the images we tested. Seems that in order to make it work well, the content image and the style image has to have highly similar layout and semantic contents. 58 | 59 | ### Citation 60 | You are welcome to cite the paper (https://arxiv.org/abs/1707.01253) with this bibtex: 61 | 62 | ``` 63 | @InProceedings{lapstyle, 64 | author = {Shaohua Li and Xinxing Xu and Liqiang Nie and Tat-Seng Chua}, 65 | title = {Laplacian-Steered Neural Style Transfer}, 66 | booktitle = {Proceedings of the ACM Multimedia Conference (MM), to appear.}, 67 | year = {2017}, 68 | } 69 | ``` 70 | 71 | ### References 72 | [1] Leon A Gatys, Alexander S Ecker,and Matthias Bethge. 2016. Image style transfer using convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2414–2423. 73 | 74 | [2] Chuan Li and Michael Wand. 2016. Combining markov random fields and convolutional neural networks for image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2479–2486. 75 | 76 | [3] Fujun Luan, Sylvain Paris, Eli Shechtman, and Kavita Bala. 2017. Deep Photo Style Transfer. arXiv preprint arXiv:1703.07511 (2017). 77 | -------------------------------------------------------------------------------- /download_models.sh: -------------------------------------------------------------------------------- 1 | cd models 2 | wget -c https://gist.githubusercontent.com/ksimonyan/3785162f95cd2d5fee77/raw/bb2b4fe0a9bb0669211cf3d0bc949dfdda173e9e/VGG_ILSVRC_19_layers_deploy.prototxt 3 | wget -c --no-check-certificate https://bethgelab.org/media/uploads/deeptextures/vgg_normalised.caffemodel 4 | wget -c http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel 5 | cd .. 6 | -------------------------------------------------------------------------------- /images/boy.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/boy.png -------------------------------------------------------------------------------- /images/cartoon2.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/cartoon2.jpg -------------------------------------------------------------------------------- /images/dogsketch.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/dogsketch.png -------------------------------------------------------------------------------- /images/flowers.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/flowers.png -------------------------------------------------------------------------------- /images/girl2.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/girl2.png -------------------------------------------------------------------------------- /images/girl3.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/girl3.jpg -------------------------------------------------------------------------------- /images/girlmrf.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/girlmrf.jpg -------------------------------------------------------------------------------- /images/girlsketch.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/girlsketch.jpg -------------------------------------------------------------------------------- /images/goat.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/goat.png -------------------------------------------------------------------------------- /images/kid5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/kid5.png -------------------------------------------------------------------------------- /images/megan.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/megan.png -------------------------------------------------------------------------------- /images/muse.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/muse.jpg -------------------------------------------------------------------------------- /images/smallworldI.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/smallworldI.jpg -------------------------------------------------------------------------------- /images/starry_night.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/images/starry_night.jpg -------------------------------------------------------------------------------- /lap_style.lua: -------------------------------------------------------------------------------- 1 | require 'torch' 2 | require 'nn' 3 | require 'image' 4 | require 'optim' 5 | require 'math' 6 | 7 | require 'loadcaffe' 8 | 9 | 10 | local cmd = torch.CmdLine() 11 | 12 | -- Basic options 13 | cmd:option('-style_image', 'examples/inputs/seated-nude.jpg', 14 | 'Style target image') 15 | cmd:option('-style_blend_weights', 'nil') 16 | cmd:option('-content_image', 'examples/inputs/tubingen.jpg', 17 | 'Content target image') 18 | -- cmd:option('-image_size', 512, 'Maximum height / width of generated image') 19 | cmd:option('-gpu', '0', 'Zero-indexed ID of the GPU to use; for CPU mode set -gpu = -1') 20 | cmd:option('-multigpu_strategy', '', 'Index of layers to split the network across GPUs') 21 | 22 | -- Optimization options 23 | cmd:option('-content_weight', 5e0) 24 | cmd:option('-style_weight', 1e2) 25 | 26 | cmd:option('-lap_weights', '100') 27 | cmd:option('-lap_layers', '2') 28 | cmd:option('-lap_nobp', false) 29 | 30 | cmd:option('-tv_weight', 1e-3) 31 | cmd:option('-num_iterations', 1000) 32 | cmd:option('-normalize_gradients', false) 33 | cmd:option('-init', 'random', 'random|image') 34 | cmd:option('-init_image', '') 35 | cmd:option('-optimizer', 'lbfgs', 'lbfgs|adam') 36 | cmd:option('-learning_rate', 1e1) 37 | cmd:option('-lbfgs_num_correction', 0) 38 | 39 | -- Output options 40 | cmd:option('-print_iter', 50) 41 | cmd:option('-save_iter', 100) 42 | cmd:option('-output_image', 'out.png') 43 | 44 | -- Other options 45 | cmd:option('-style_scale', 1.0) 46 | cmd:option('-original_colors', 0) 47 | cmd:option('-pooling', 'max', 'max|avg') 48 | cmd:option('-proto_file', 'models/VGG_ILSVRC_19_layers_deploy.prototxt') 49 | cmd:option('-model_file', 'models/VGG_ILSVRC_19_layers.caffemodel') 50 | cmd:option('-backend', 'nn', 'nn|cudnn|clnn') 51 | cmd:option('-cudnn_autotune', false) 52 | cmd:option('-seed', -1) 53 | 54 | cmd:option('-content_layers', 'relu4_2', 'layers for content') 55 | cmd:option('-style_layers', 'relu1_1,relu2_1,relu3_1,relu4_1,relu5_1', 'layers for style') 56 | 57 | 58 | local function main(params) 59 | local dtype, multigpu = setup_gpu(params) 60 | 61 | local loadcaffe_backend = params.backend 62 | if params.backend == 'clnn' then loadcaffe_backend = 'nn' end 63 | local cnn = loadcaffe.load(params.proto_file, params.model_file, loadcaffe_backend):type(dtype) 64 | 65 | local content_image = image.load(params.content_image, 3) 66 | -- content_image = image.scale(content_image, params.image_size, 'bilinear') 67 | local content_image_caffe = preprocess(content_image):float() 68 | 69 | -- local style_size = math.ceil(params.style_scale * params.image_size) 70 | local style_image_list = params.style_image:split(',') 71 | local style_images_caffe = {} 72 | for _, img_path in ipairs(style_image_list) do 73 | local img = image.load(img_path, 3) 74 | -- img = image.scale(img, style_size, 'bilinear') 75 | local img_caffe = preprocess(img):float() 76 | table.insert(style_images_caffe, img_caffe) 77 | end 78 | 79 | local init_image = nil 80 | if params.init_image ~= '' then 81 | init_image = image.load(params.init_image, 3) 82 | local H, W = content_image:size(2), content_image:size(3) 83 | init_image = image.scale(init_image, W, H, 'bilinear') 84 | init_image = preprocess(init_image):float() 85 | end 86 | 87 | -- Handle style blending weights for multiple style inputs 88 | local style_blend_weights = nil 89 | if params.style_blend_weights == 'nil' then 90 | -- Style blending not specified, so use equal weighting 91 | style_blend_weights = {} 92 | for i = 1, #style_image_list do 93 | table.insert(style_blend_weights, 1.0) 94 | end 95 | else 96 | style_blend_weights = params.style_blend_weights:split(',') 97 | assert(#style_blend_weights == #style_image_list, 98 | '-style_blend_weights and -style_images must have the same number of elements') 99 | end 100 | -- Normalize the style blending weights so they sum to 1 101 | local style_blend_sum = 0 102 | for i = 1, #style_blend_weights do 103 | style_blend_weights[i] = tonumber(style_blend_weights[i]) 104 | style_blend_sum = style_blend_sum + style_blend_weights[i] 105 | end 106 | for i = 1, #style_blend_weights do 107 | style_blend_weights[i] = style_blend_weights[i] / style_blend_sum 108 | end 109 | 110 | local content_layers = params.content_layers:split(",") 111 | local style_layers = params.style_layers:split(",") 112 | 113 | local lap_layers = params.lap_layers:split(",") 114 | for i = 1, #lap_layers do 115 | lap_layers[i] = tonumber(lap_layers[i]) 116 | end 117 | local lap_weights = params.lap_weights:split(",") 118 | for i = 1, #lap_weights do 119 | lap_weights[i] = tonumber(lap_weights[i]) 120 | end 121 | assert( #lap_layers == #lap_weights ) 122 | 123 | -- Set up the network, inserting style and content loss modules 124 | local content_losses, style_losses, lap_losses = {}, {}, {} 125 | local next_content_idx, next_style_idx = 1, 1 126 | local net1 = nn.Sequential() 127 | if params.tv_weight > 0 then 128 | local tv_mod = nn.TVLoss(params.tv_weight):type(dtype) 129 | net1:add(tv_mod) 130 | end 131 | for i = 1, #cnn do 132 | if next_content_idx <= #content_layers or next_style_idx <= #style_layers then 133 | local layer = cnn:get(i) 134 | local name = layer.name 135 | local layer_type = torch.type(layer) 136 | local is_pooling = (layer_type == 'cudnn.SpatialMaxPooling' or layer_type == 'nn.SpatialMaxPooling') 137 | if is_pooling and params.pooling == 'avg' then 138 | assert(layer.padW == 0 and layer.padH == 0) 139 | local kW, kH = layer.kW, layer.kH 140 | local dW, dH = layer.dW, layer.dH 141 | local avg_pool_layer = nn.SpatialAveragePooling(kW, kH, dW, dH):type(dtype) 142 | local msg = 'Replacing max pooling at layer %d with average pooling' 143 | print(string.format(msg, i)) 144 | net1:add(avg_pool_layer) 145 | else 146 | net1:add(layer) 147 | end 148 | if name == content_layers[next_content_idx] then 149 | print("Setting up content layer", i, ":", layer.name) 150 | local norm = params.normalize_gradients 151 | local loss_module = nn.ContentLoss(params.content_weight, norm):type(dtype) 152 | net1:add(loss_module) 153 | table.insert(content_losses, loss_module) 154 | next_content_idx = next_content_idx + 1 155 | end 156 | if name == style_layers[next_style_idx] then 157 | print("Setting up style layer ", i, ":", layer.name) 158 | local norm = params.normalize_gradients 159 | local loss_module = nn.StyleLoss(params.style_weight, norm):type(dtype) 160 | net1:add(loss_module) 161 | table.insert(style_losses, loss_module) 162 | next_style_idx = next_style_idx + 1 163 | end 164 | end 165 | end 166 | 167 | local net = nn.ConcatTable() 168 | net:add(net1) 169 | 170 | --[[ 171 | local function lapofGauss(size, sigma) 172 | local center = 0.5 * size + 0.5 173 | local logauss = torch.Tensor(size,size) 174 | for i=1,size do 175 | for j=1,size do 176 | xsq = math.pow((i-center)/sigma,2)/2 177 | ysq = math.pow((j-center)/sigma,2)/2 178 | derivCoef = 1 - (xsq + ysq) 179 | logauss[i][j] = derivCoef * math.exp(-(xsq + ysq)) 180 | end 181 | end 182 | logauss = logauss - logauss:sum() / (size*size) 183 | logauss = logauss / logauss[{math.floor(center),math.floor(center)}] 184 | return logauss 185 | end 186 | 187 | local ks = 5 188 | logauss = lapofGauss(ks,0.5) 189 | lapW = torch.Tensor(1,3,ks,ks) 190 | for i = 1, 3 do 191 | lapW[{1,i,{},{}}] = logauss 192 | end 193 | 194 | --]] 195 | 196 | -- conv weight shape: out_channels, in_channels, kernel_size[0], kernel_size[1] 197 | local ks = 3 198 | laplacian = torch.Tensor(3,3):zero() 199 | laplacian[{1,2}] = -1.0 200 | laplacian[{2,1}] = -1.0 201 | laplacian[{2,2}] = 4.0 202 | laplacian[{2,3}] = -1.0 203 | laplacian[{3,2}] = -1.0 204 | lapW = torch.Tensor(1,3,3,3) 205 | for i = 1, 3 do 206 | lapW[{1,i,{},{}}] = laplacian 207 | end 208 | 209 | stride = math.floor(ks/2) 210 | for i = 1, #lap_layers do 211 | local netlap = nn.Sequential() 212 | local ps = math.pow(2, lap_layers[i]) 213 | local avg_pool_layer = nn.SpatialAveragePooling(ps, ps, ps, ps):type(dtype) 214 | netlap:add(avg_pool_layer) 215 | local conv_layer = nn.SpatialConvolution(3,1,ks,ks,stride,stride):type(dtype) 216 | conv_layer.weight = lapW 217 | netlap:add(conv_layer) 218 | local norm = params.normalize_gradients 219 | local loss_module = nn.ContentLoss(lap_weights[i], norm):type(dtype) 220 | netlap:add(loss_module) 221 | table.insert(content_losses, loss_module) 222 | table.insert(lap_losses, loss_module) 223 | net:add(netlap) 224 | end 225 | 226 | -- print(string.format('content losses: %d', #content_losses)) 227 | 228 | if multigpu then 229 | net = setup_multi_gpu(net, params) 230 | end 231 | net:type(dtype) 232 | 233 | -- Capture content targets 234 | for i = 1, #content_losses do 235 | content_losses[i].mode = 'capture' 236 | end 237 | print 'Capturing content targets' 238 | print(net) 239 | content_image_caffe = content_image_caffe:type(dtype) 240 | net:forward(content_image_caffe:type(dtype)) 241 | 242 | -- Capture style targets 243 | for i = 1, #content_losses do 244 | content_losses[i].mode = 'none' 245 | end 246 | for i = 1, #style_images_caffe do 247 | print(string.format('Capturing style target %d', i)) 248 | for j = 1, #style_losses do 249 | style_losses[j].mode = 'capture' 250 | style_losses[j].blend_weight = style_blend_weights[i] 251 | end 252 | net:forward(style_images_caffe[i]:type(dtype)) 253 | end 254 | 255 | -- Set all loss modules to loss mode 256 | for i = 1, #content_losses do 257 | content_losses[i].mode = 'loss' 258 | end 259 | 260 | if params.lap_nobp then 261 | for i = 1, #lap_losses do 262 | lap_losses[i].strength_bp = 0 263 | end 264 | end 265 | 266 | for i = 1, #style_losses do 267 | style_losses[i].mode = 'loss' 268 | end 269 | 270 | -- We don't need the base CNN anymore, so clean it up to save memory. 271 | cnn = nil 272 | for i=1, #net.modules do 273 | local module = net.modules[i] 274 | if torch.type(module) == 'nn.SpatialConvolutionMM' then 275 | -- remove these, not used, but uses gpu memory 276 | module.gradWeight = nil 277 | module.gradBias = nil 278 | end 279 | end 280 | collectgarbage() 281 | 282 | -- Initialize the image 283 | if params.seed >= 0 then 284 | torch.manualSeed(params.seed) 285 | end 286 | local img = nil 287 | if params.init == 'random' then 288 | img = torch.randn(content_image:size()):float():mul(0.001) 289 | elseif params.init == 'image' then 290 | if init_image then 291 | img = init_image:clone() 292 | else 293 | img = content_image_caffe:clone() 294 | end 295 | else 296 | error('Invalid init type') 297 | end 298 | img = img:type(dtype) 299 | 300 | -- Run it through the network once to get the proper size for the gradient 301 | -- All the gradients will come from the extra loss modules, so we just pass 302 | -- zeros into the top of the net on the backward pass. 303 | local y = net:forward(img) 304 | local dy = img.new(#y):zero() 305 | 306 | -- Declaring this here lets us access it in maybe_print 307 | local optim_state = nil 308 | if params.optimizer == 'lbfgs' then 309 | optim_state = { 310 | maxIter = params.num_iterations, 311 | verbose=true, 312 | tolX=-1, 313 | tolFun=-1, 314 | } 315 | if params.lbfgs_num_correction > 0 then 316 | optim_state.nCorrection = params.lbfgs_num_correction 317 | end 318 | elseif params.optimizer == 'adam' then 319 | optim_state = { 320 | learningRate = params.learning_rate, 321 | } 322 | else 323 | error(string.format('Unrecognized optimizer "%s"', params.optimizer)) 324 | end 325 | 326 | local function maybe_print(t, loss) 327 | local verbose = params.print_iter > 0 and 328 | (t % params.print_iter == 0 or t == params.num_iterations - 1) 329 | if verbose then 330 | print(string.format('Iteration %d / %d', t, params.num_iterations)) 331 | for i, loss_module in ipairs(content_losses) do 332 | print(string.format(' Content %d loss: %f', i, loss_module.loss)) 333 | end 334 | for i, loss_module in ipairs(style_losses) do 335 | print(string.format(' Style %d loss: %f', i, loss_module.loss)) 336 | end 337 | print(string.format(' Total loss: %f', loss)) 338 | end 339 | end 340 | 341 | local function maybe_save(t) 342 | local should_save = params.save_iter > 0 and t % params.save_iter == 0 343 | should_save = should_save or t == params.num_iterations 344 | if should_save then 345 | local disp = deprocess(img:double()) 346 | disp = image.minmax{tensor=disp, min=0, max=1} 347 | local filename = build_filename(params.output_image, t) 348 | if t == params.num_iterations then 349 | filename = params.output_image 350 | end 351 | 352 | -- Maybe perform postprocessing for color-independent style transfer 353 | if params.original_colors == 1 then 354 | disp = original_colors(content_image, disp) 355 | end 356 | 357 | image.save(filename, disp) 358 | end 359 | end 360 | 361 | -- Function to evaluate loss and gradient. We run the net forward and 362 | -- backward to get the gradient, and sum up losses from the loss modules. 363 | -- optim.lbfgs internally handles iteration and calls this function many 364 | -- times, so we manually count the number of iterations to handle printing 365 | -- and saving intermediate results. 366 | local num_calls = 0 367 | local function feval(x) 368 | net:forward(x) 369 | local grad = net:updateGradInput(x, dy) 370 | local loss = 0 371 | for _, mod in ipairs(content_losses) do 372 | loss = loss + mod.loss 373 | end 374 | for _, mod in ipairs(style_losses) do 375 | loss = loss + mod.loss 376 | end 377 | maybe_print(num_calls, loss) 378 | num_calls = num_calls + 1 379 | maybe_save(num_calls) 380 | 381 | collectgarbage() 382 | -- optim.lbfgs expects a vector for gradients 383 | return loss, grad:view(grad:nElement()) 384 | end 385 | 386 | -- Run optimization. 387 | if params.optimizer == 'lbfgs' then 388 | print('Running optimization with L-BFGS') 389 | local x, losses = optim.lbfgs(feval, img, optim_state) 390 | elseif params.optimizer == 'adam' then 391 | print('Running optimization with ADAM') 392 | for t = 1, params.num_iterations do 393 | local x, losses = optim.adam(feval, img, optim_state) 394 | end 395 | end 396 | end 397 | 398 | 399 | function setup_gpu(params) 400 | local multigpu = false 401 | if params.gpu:find(',') then 402 | multigpu = true 403 | params.gpu = params.gpu:split(',') 404 | for i = 1, #params.gpu do 405 | params.gpu[i] = tonumber(params.gpu[i]) + 1 406 | end 407 | else 408 | params.gpu = tonumber(params.gpu) + 1 409 | end 410 | local dtype = 'torch.FloatTensor' 411 | if multigpu or params.gpu > 0 then 412 | if params.backend ~= 'clnn' then 413 | require 'cutorch' 414 | require 'cunn' 415 | if multigpu then 416 | cutorch.setDevice(params.gpu[1]) 417 | else 418 | cutorch.setDevice(params.gpu) 419 | end 420 | dtype = 'torch.CudaTensor' 421 | else 422 | require 'clnn' 423 | require 'cltorch' 424 | if multigpu then 425 | cltorch.setDevice(params.gpu[1]) 426 | else 427 | cltorch.setDevice(params.gpu) 428 | end 429 | dtype = torch.Tensor():cl():type() 430 | end 431 | else 432 | params.backend = 'nn' 433 | end 434 | 435 | if params.backend == 'cudnn' then 436 | require 'cudnn' 437 | if params.cudnn_autotune then 438 | cudnn.benchmark = true 439 | end 440 | cudnn.SpatialConvolution.accGradParameters = nn.SpatialConvolutionMM.accGradParameters -- ie: nop 441 | end 442 | return dtype, multigpu 443 | end 444 | 445 | 446 | function setup_multi_gpu(net, params) 447 | local DEFAULT_STRATEGIES = { 448 | [2] = {3}, 449 | } 450 | local gpu_splits = nil 451 | if params.multigpu_strategy == '' then 452 | -- Use a default strategy 453 | gpu_splits = DEFAULT_STRATEGIES[#params.gpu] 454 | -- Offset the default strategy by one if we are using TV 455 | if params.tv_weight > 0 then 456 | for i = 1, #gpu_splits do gpu_splits[i] = gpu_splits[i] + 1 end 457 | end 458 | else 459 | -- Use the user-specified multigpu strategy 460 | gpu_splits = params.multigpu_strategy:split(',') 461 | for i = 1, #gpu_splits do 462 | gpu_splits[i] = tonumber(gpu_splits[i]) 463 | end 464 | end 465 | assert(gpu_splits ~= nil, 'Must specify -multigpu_strategy') 466 | local gpus = params.gpu 467 | 468 | local cur_chunk = nn.Sequential() 469 | local chunks = {} 470 | for i = 1, #net do 471 | cur_chunk:add(net:get(i)) 472 | if i == gpu_splits[1] then 473 | table.remove(gpu_splits, 1) 474 | table.insert(chunks, cur_chunk) 475 | cur_chunk = nn.Sequential() 476 | end 477 | end 478 | table.insert(chunks, cur_chunk) 479 | assert(#chunks == #gpus) 480 | 481 | local new_net = nn.Sequential() 482 | for i = 1, #chunks do 483 | local out_device = nil 484 | if i == #chunks then 485 | out_device = gpus[1] 486 | end 487 | new_net:add(nn.GPU(chunks[i], gpus[i], out_device)) 488 | end 489 | 490 | return new_net 491 | end 492 | 493 | 494 | function build_filename(output_image, iteration) 495 | local ext = paths.extname(output_image) 496 | local basename = paths.basename(output_image, ext) 497 | local directory = paths.dirname(output_image) 498 | return string.format('%s/%s_%d.%s',directory, basename, iteration, ext) 499 | end 500 | 501 | 502 | -- Preprocess an image before passing it to a Caffe model. 503 | -- We need to rescale from [0, 1] to [0, 255], convert from RGB to BGR, 504 | -- and subtract the mean pixel. 505 | function preprocess(img) 506 | local mean_pixel = torch.DoubleTensor({103.939, 116.779, 123.68}) 507 | local perm = torch.LongTensor{3, 2, 1} 508 | img = img:index(1, perm):mul(256.0) 509 | mean_pixel = mean_pixel:view(3, 1, 1):expandAs(img) 510 | img:add(-1, mean_pixel) 511 | return img 512 | end 513 | 514 | 515 | -- Undo the above preprocessing. 516 | function deprocess(img) 517 | local mean_pixel = torch.DoubleTensor({103.939, 116.779, 123.68}) 518 | mean_pixel = mean_pixel:view(3, 1, 1):expandAs(img) 519 | img = img + mean_pixel 520 | local perm = torch.LongTensor{3, 2, 1} 521 | img = img:index(1, perm):div(256.0) 522 | return img 523 | end 524 | 525 | 526 | -- Combine the Y channel of the generated image and the UV channels of the 527 | -- content image to perform color-independent style transfer. 528 | function original_colors(content, generated) 529 | local generated_y = image.rgb2yuv(generated)[{{1, 1}}] 530 | local content_uv = image.rgb2yuv(content)[{{2, 3}}] 531 | return image.yuv2rgb(torch.cat(generated_y, content_uv, 1)) 532 | end 533 | 534 | 535 | -- Define an nn Module to compute content loss in-place 536 | local ContentLoss, parent = torch.class('nn.ContentLoss', 'nn.Module') 537 | 538 | function ContentLoss:__init(strength, normalize) 539 | parent.__init(self) 540 | self.strength = strength 541 | -- strength when doing BP. used to disable lap_losses when -lap_nobp is specified 542 | self.strength_bp = strength 543 | self.target = torch.Tensor() 544 | self.normalize = normalize or false 545 | self.loss = 0 546 | self.crit = nn.MSECriterion() 547 | self.mode = 'none' 548 | end 549 | 550 | function ContentLoss:updateOutput(input) 551 | if self.mode == 'loss' then 552 | self.loss = self.crit:forward(input, self.target) * self.strength 553 | elseif self.mode == 'capture' then 554 | self.target:resizeAs(input):copy(input) 555 | end 556 | self.output = input 557 | return self.output 558 | end 559 | 560 | function ContentLoss:updateGradInput(input, gradOutput) 561 | if self.mode == 'loss' then 562 | if input:nElement() == self.target:nElement() then 563 | self.gradInput = self.crit:backward(input, self.target) 564 | end 565 | if self.normalize then 566 | self.gradInput:div(torch.norm(self.gradInput, 1) + 1e-8) 567 | end 568 | self.gradInput:mul(self.strength_bp) 569 | self.gradInput:add(gradOutput) 570 | else 571 | self.gradInput:resizeAs(gradOutput):copy(gradOutput) 572 | end 573 | return self.gradInput 574 | end 575 | 576 | 577 | local Gram, parent = torch.class('nn.GramMatrix', 'nn.Module') 578 | 579 | function Gram:__init() 580 | parent.__init(self) 581 | end 582 | 583 | function Gram:updateOutput(input) 584 | assert(input:dim() == 3) 585 | local C, H, W = input:size(1), input:size(2), input:size(3) 586 | local x_flat = input:view(C, H * W) 587 | self.output:resize(C, C) 588 | self.output:mm(x_flat, x_flat:t()) 589 | return self.output 590 | end 591 | 592 | function Gram:updateGradInput(input, gradOutput) 593 | assert(input:dim() == 3 and input:size(1)) 594 | local C, H, W = input:size(1), input:size(2), input:size(3) 595 | local x_flat = input:view(C, H * W) 596 | self.gradInput:resize(C, H * W):mm(gradOutput, x_flat) 597 | self.gradInput:addmm(gradOutput:t(), x_flat) 598 | self.gradInput = self.gradInput:view(C, H, W) 599 | return self.gradInput 600 | end 601 | 602 | 603 | -- Define an nn Module to compute style loss in-place 604 | local StyleLoss, parent = torch.class('nn.StyleLoss', 'nn.Module') 605 | 606 | function StyleLoss:__init(strength, normalize) 607 | parent.__init(self) 608 | self.normalize = normalize or false 609 | self.strength = strength 610 | self.target = torch.Tensor() 611 | self.mode = 'none' 612 | self.loss = 0 613 | 614 | self.gram = nn.GramMatrix() 615 | self.blend_weight = nil 616 | self.G = nil 617 | self.crit = nn.MSECriterion() 618 | end 619 | 620 | function StyleLoss:updateOutput(input) 621 | self.G = self.gram:forward(input) 622 | self.G:div(input:nElement()) 623 | if self.mode == 'capture' then 624 | if self.blend_weight == nil then 625 | self.target:resizeAs(self.G):copy(self.G) 626 | elseif self.target:nElement() == 0 then 627 | self.target:resizeAs(self.G):copy(self.G):mul(self.blend_weight) 628 | else 629 | self.target:add(self.blend_weight, self.G) 630 | end 631 | elseif self.mode == 'loss' then 632 | self.loss = self.strength * self.crit:forward(self.G, self.target) 633 | end 634 | self.output = input 635 | return self.output 636 | end 637 | 638 | function StyleLoss:updateGradInput(input, gradOutput) 639 | if self.mode == 'loss' then 640 | local dG = self.crit:backward(self.G, self.target) 641 | dG:div(input:nElement()) 642 | self.gradInput = self.gram:backward(input, dG) 643 | if self.normalize then 644 | self.gradInput:div(torch.norm(self.gradInput, 1) + 1e-8) 645 | end 646 | self.gradInput:mul(self.strength) 647 | self.gradInput:add(gradOutput) 648 | else 649 | self.gradInput = gradOutput 650 | end 651 | return self.gradInput 652 | end 653 | 654 | 655 | local TVLoss, parent = torch.class('nn.TVLoss', 'nn.Module') 656 | 657 | function TVLoss:__init(strength) 658 | parent.__init(self) 659 | self.strength = strength 660 | self.x_diff = torch.Tensor() 661 | self.y_diff = torch.Tensor() 662 | end 663 | 664 | function TVLoss:updateOutput(input) 665 | self.output = input 666 | return self.output 667 | end 668 | 669 | -- TV loss backward pass inspired by kaishengtai/neuralart 670 | function TVLoss:updateGradInput(input, gradOutput) 671 | self.gradInput:resizeAs(input):zero() 672 | local C, H, W = input:size(1), input:size(2), input:size(3) 673 | self.x_diff:resize(3, H - 1, W - 1) 674 | self.y_diff:resize(3, H - 1, W - 1) 675 | self.x_diff:copy(input[{{}, {1, -2}, {1, -2}}]) 676 | self.x_diff:add(-1, input[{{}, {1, -2}, {2, -1}}]) 677 | self.y_diff:copy(input[{{}, {1, -2}, {1, -2}}]) 678 | self.y_diff:add(-1, input[{{}, {2, -1}, {1, -2}}]) 679 | self.gradInput[{{}, {1, -2}, {1, -2}}]:add(self.x_diff):add(self.y_diff) 680 | self.gradInput[{{}, {1, -2}, {2, -1}}]:add(-1, self.x_diff) 681 | self.gradInput[{{}, {2, -1}, {1, -2}}]:add(-1, self.y_diff) 682 | self.gradInput:mul(self.strength) 683 | self.gradInput:add(gradOutput) 684 | return self.gradInput 685 | end 686 | 687 | 688 | local params = cmd:parse(arg) 689 | main(params) 690 | -------------------------------------------------------------------------------- /laplacian.py: -------------------------------------------------------------------------------- 1 | # Generate image Laplacian of a given image, save it as a .png file for visualization 2 | 3 | import numpy as np 4 | import scipy.optimize, scipy.ndimage, scipy.misc 5 | import lasagne 6 | from lasagne.layers import Conv2DLayer, InputLayer 7 | import sys 8 | import pdb 9 | 10 | def prepare_image(image): 11 | """Given an image loaded from disk, turn it into a representation compatible with the model. 12 | The format is (b,c,y,x) with batch=1 for a single image, channels=3 for RGB, and y,x matching 13 | the resolution. 14 | """ 15 | image = np.swapaxes(np.swapaxes(image, 1, 2), 0, 1)[::-1, :, :] 16 | image = image.astype(np.float32) 17 | return image[np.newaxis] 18 | 19 | def finalize_image(image, scale): 20 | """Based on the output of the neural network, convert it into an image format that can be saved 21 | to disk -- shuffling dimensions as appropriate. 22 | """ 23 | image = np.swapaxes(np.swapaxes(image[::-1], 0, 1), 1, 2) 24 | # ignore the sign, scale up the pixel value to make it more striking 25 | image = np.abs(image) * scale 26 | image = np.clip(image, 0, 255).astype('uint8') 27 | return image 28 | 29 | infilename = sys.argv[1] 30 | outfilename = sys.argv[2] 31 | if len(sys.argv) > 3: 32 | scale = int( sys.argv[3] ) 33 | else: 34 | scale = 3 35 | 36 | net = {} 37 | # once a channel among 'R','G','B', 38 | # if all channels are fed to laplacian conv at the same time, 39 | # the output will be sum of the outputs from all input channels 40 | net['img'] = InputLayer((None, 1, None, None)) 41 | net['laplacian'] = Conv2DLayer(net['img'], 1, 3, pad=1) 42 | laplacian = np.array( [ [0,-1,0], [-1,4,-1], [0,-1,0] ], dtype=np.float32 ) 43 | W = np.zeros( (1, 1, 3, 3), dtype=np.float32 ) 44 | W[0,0] = laplacian 45 | net['laplacian'].W.set_value(W) 46 | 47 | orig_img = scipy.ndimage.imread(infilename, mode='RGB') 48 | img = prepare_image(orig_img) 49 | 50 | output = [] 51 | for i in xrange(3): 52 | img_chan = img[:, [i], :, :] 53 | tensor_input = { net['img']: img_chan } 54 | out_chan = lasagne.layers.get_output( net['laplacian'], tensor_input ) 55 | output.append( out_chan.eval()[0,0] ) 56 | 57 | output = np.array(output) 58 | 59 | out_img = finalize_image(output, scale) 60 | scipy.misc.toimage(out_img, cmin=0, cmax=255).save(outfilename) 61 | -------------------------------------------------------------------------------- /logstat.py: -------------------------------------------------------------------------------- 1 | # Get statistics from lap_style.lua logs, used to generate Table 1 in the lapstyle paper 2 | import re 3 | import os 4 | import sys 5 | import numpy as np 6 | 7 | contentimageAll = ( 'megan.png', 'kid5.png', 'goat.png', 'girlmrf.jpg', 'boy.png' ) 8 | styleimageAll = ( 'flowers.png', 'smallworldI.jpg', 'muse.png', 'girlsketch.png', 'girl3.jpg' ) 9 | #sigs = ("20_2_100.log", "20_2_100_nobp.log") 10 | #contentLayer2contentLossType = [ -1, 1, 2, -1 ] 11 | 12 | sigs = ("relu2_2,relu4_2_20_100_nobp.log",) 13 | contentLayer2contentLossType = [ -1, 0, 1, 2 ] 14 | 15 | logpath =sys.argv[1] 16 | readlogcount = 0 17 | 18 | max_layer_num = 9 19 | init_losses = [] 20 | end_losses = [] 21 | end_losses_nobp = [] 22 | 23 | for i, contentFile in enumerate(contentimageAll): 24 | styleFile = styleimageAll[i] 25 | contentName = os.path.splitext(contentFile)[0] 26 | styleName = os.path.splitext(styleFile)[0] 27 | for sig in sigs: 28 | logfilename = "%s/%s_%s%s" %(logpath, contentName, styleName, sig) 29 | if not os.path.isfile(logfilename): 30 | print "'%s' doesn't exist!" %(logfilename) 31 | else: 32 | if sig[-8:] == 'nobp.log': 33 | nobp = True 34 | else: 35 | nobp = False 36 | 37 | LOG = open(logfilename) 38 | readlogcount += 1 39 | iter_type = 'none' 40 | style_loss_sum = 0 41 | content_loss_sum = 0 42 | lap_loss_sum = 0 43 | 44 | for line in LOG: 45 | line = line.strip() 46 | # Iteration 0 / 1000 47 | result = re.match(r"Iteration (\d+) / (\d+)", line) 48 | if result: 49 | if iter_type == 'init': 50 | loss_block = [1.0, lap_loss_sum/total_loss, total_loss/lap_loss_sum, content_loss_sum/lap_loss_sum, style_loss_sum/lap_loss_sum] 51 | init_losses.append(loss_block) 52 | lap_loss_sum0 = lap_loss_sum 53 | style_loss_sum = 0 54 | content_loss_sum = 0 55 | lap_loss_sum = 0 56 | curr_iter = int(result.group(1)) 57 | total_iter = int(result.group(2)) 58 | if curr_iter == 0: 59 | iter_type = 'init' 60 | elif curr_iter == total_iter - 1: 61 | iter_type = 'end' 62 | else: 63 | iter_type = 'mid' 64 | 65 | continue 66 | if iter_type == 'init' or iter_type == 'end': 67 | # Content 1 loss: 4858483.750000 68 | result = re.match(r"(Content|Style) (\d+) loss: ([0-9.]+)", line) 69 | if result: 70 | loss = float(result.group(3)) 71 | loss_type = result.group(1) 72 | layer_no = int(result.group(2)) 73 | if loss_type == 'Content': 74 | contentLossType = contentLayer2contentLossType[layer_no] 75 | if contentLossType == 1: 76 | content_loss_sum += loss 77 | if contentLossType == 2: 78 | lap_loss_sum += loss 79 | else: 80 | style_loss_sum += loss 81 | else: 82 | result = re.match(r"Total loss: ([0-9.]+)", line) 83 | if result: 84 | total_loss = float(result.group(1)) 85 | 86 | if iter_type == 'end': 87 | loss_block = [lap_loss_sum/lap_loss_sum0, lap_loss_sum/total_loss, total_loss/lap_loss_sum0, 88 | content_loss_sum/lap_loss_sum0, style_loss_sum/lap_loss_sum0] 89 | if nobp: 90 | end_losses_nobp.append(loss_block) 91 | else: 92 | end_losses.append(loss_block) 93 | 94 | print "%d log files read" %readlogcount 95 | init_losses = np.array(init_losses) 96 | #arr_labels = ('init_losses', 'end_losses', 'end_losses_nobp') 97 | # always init_losses_nobp == init_losses. So no need to print it 98 | #for i,arr in enumerate([init_losses, end_losses, end_losses_nobp]): 99 | arr_labels = ('init_losses', 'end_losses_nobp') 100 | for i,arr in enumerate([init_losses, end_losses_nobp]): 101 | arr = np.array(arr) 102 | arr_label = arr_labels[i] 103 | loss_labels = ('lap', 'lap-frac', 'total', 'content', 'style') 104 | print "%s:" %arr_label 105 | for j,loss_label in enumerate(loss_labels): 106 | losses = arr[:,j] 107 | mean = np.mean(losses) 108 | std = np.std(losses) 109 | print "%s: %.5f (%.5f)" %(loss_label, mean, std) 110 | print losses 111 | -------------------------------------------------------------------------------- /multi-lap-test.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash -x 2 | # test lap_style.lua with the combination of multiple laplacian layers, on a single image pair 3 | 4 | contentimg=$1 5 | styleimg=$2 6 | contentweight=$3 7 | contentFilename=$(basename "$contentimg") 8 | contentname="${contentFilename%.*}" 9 | styleFilename=$(basename "$styleimg") 10 | stylename="${styleFilename%.*}" 11 | laplayersAll=( 1 1 2 1,2 1,3 2,3 2,4) 12 | lapweightsAll=(0 50 100 50,100 50,400 100,400 100,1600) 13 | configLen=${#laplayersAll[@]} 14 | for (( i=0; i<$configLen; i++ )); do 15 | laplayers=${laplayersAll[$i]} 16 | lapweights=${lapweightsAll[$i]} 17 | outsig="${contentname}_${stylename}${contentweight}_${laplayers}_${lapweights}" 18 | th lap_style.lua -style_image images/$styleimg -content_image images/$contentimg -output_image output/${outsig}.png -content_weight $contentweight -lap_layers $laplayers -lap_weights $lapweights 2>&1| tee output/${outsig}.log 19 | done 20 | -------------------------------------------------------------------------------- /neural-doodle/doodle.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # 3 | # Neural Doodle! 4 | # Copyright (c) 2017, Shaohua Li 5 | # Copyright (c) 2016, Alex J. Champandard. 6 | # Please refer to https://github.com/alexjc/neural-doodle for a user manual. 7 | 8 | ''' Dimensionalities and other numbers are based on the command: 9 | python3 doodle-orig.py --style samples/Gogh.jpg --content samples/Seth.png \ 10 | --output SethAsGogh.png --device=gpu0 --phases=4 --iterations=40 11 | ''' 12 | 13 | import os 14 | import sys 15 | import bz2 16 | import math 17 | import time 18 | import pickle 19 | import argparse 20 | import itertools 21 | import collections 22 | import pdb 23 | 24 | 25 | # Configure all options first so we can custom load other libraries (Theano) based on device specified by user. 26 | parser = argparse.ArgumentParser(description='Generate a new image by applying style onto a content image.', 27 | formatter_class=argparse.ArgumentDefaultsHelpFormatter) 28 | add_arg = parser.add_argument 29 | 30 | add_arg('--content', default=None, type=str, help='Content image path as optimization target.') 31 | add_arg('--content-weight', default=30.0, type=float, help='Weight of content relative to style.') 32 | add_arg('--lap-weight', default=100.0, type=float, help='Weight of laplace content relative to style.') 33 | # consider add '1_laplace' to content_layers, to force similar laplacian of the new image and the content image 34 | add_arg('--content-layers', default='4_2,1_laplace', type=str, help='The layer with which to match content.') 35 | add_arg('--style', default=None, type=str, help='Style image path to extract patches.') 36 | add_arg('--style-weight', default=15.0, type=float, help='Weight of style relative to content.') 37 | add_arg('--style-layers', default='3_1,4_1', type=str, help='The layers to match style patches.') 38 | add_arg('--semantic-ext', default='_sem.png', type=str, help='File extension for the semantic maps.') 39 | add_arg('--semantic-weight', default=10.0, type=float, help='Global weight of semantics vs. features.') 40 | add_arg('--output', default='output.png', type=str, help='Output image path to save once done.') 41 | add_arg('--output-size', default=None, type=str, help='Size of the output image, e.g. 512x512.') 42 | add_arg('--phases', default=3, type=int, help='Number of image scales to process in phases.') 43 | add_arg('--slices', default=2, type=int, help='Split patches up into this number of batches.') 44 | add_arg('--cache', default=0, type=int, help='Whether to compute matches only once.') 45 | add_arg('--smoothness', default=1E+0, type=float, help='Weight of image smoothing scheme.') 46 | add_arg('--variety', default=0.0, type=float, help='Bias toward selecting diverse patches, e.g. 0.5.') 47 | add_arg('--seed', default='noise', type=str, help='Seed image path, "noise" or "content".') 48 | add_arg('--seed-range', default='16:240', type=str, help='Random colors chosen in range, e.g. 0:255.') 49 | add_arg('--iterations', default=100, type=int, help='Number of iterations to run each resolution.') 50 | add_arg('--device', default='gpu', type=str, help='Index of the GPU number to use, for theano.') 51 | add_arg('--print-every', default=10, type=int, help='How often to log statistics to stdout.') 52 | add_arg('--save-every', default=10, type=int, help='How frequently to save PNG into `frames`.') 53 | args = parser.parse_args() 54 | 55 | 56 | #---------------------------------------------------------------------------------------------------------------------- 57 | 58 | # Color coded output helps visualize the information a little better, plus looks cool! 59 | class ansi: 60 | BOLD = '\033[1;97m' 61 | WHITE = '\033[0;97m' 62 | YELLOW = '\033[0;33m' 63 | YELLOW_B = '\033[0;33m' 64 | RED = '\033[0;31m' 65 | RED_B = '\033[1;31m' 66 | BLUE = '\033[0;94m' 67 | BLUE_B = '\033[1;94m' 68 | CYAN = '\033[0;36m' 69 | CYAN_B = '\033[1;36m' 70 | ENDC = '\033[0m' 71 | 72 | def error(message, *lines): 73 | string = "\n{}ERROR: " + message + "{}\n" + "\n".join(lines) + "{}\n" 74 | print(string.format(ansi.RED_B, ansi.RED, ansi.ENDC)) 75 | sys.exit(-1) 76 | 77 | print('{}Neural Doodle for semantic style transfer.{}'.format(ansi.CYAN_B, ansi.ENDC)) 78 | 79 | # Load the underlying deep learning libraries based on the device specified. If you specify THEANO_FLAGS manually, 80 | # the code assumes you know what you are doing and they are not overriden! 81 | os.environ.setdefault('THEANO_FLAGS', 'floatX=float32,device={},force_device=True,'\ 82 | 'print_active_device=False'.format(args.device)) 83 | 84 | # Scientific & Imaging Libraries 85 | import numpy as np 86 | import scipy.optimize, scipy.ndimage, scipy.misc 87 | import PIL 88 | 89 | # Numeric Computing (GPU) 90 | import theano 91 | import theano.tensor as T 92 | import theano.tensor.nnet.neighbours 93 | 94 | # Support ansi colors in Windows too. 95 | if sys.platform == 'win32': 96 | import colorama 97 | 98 | # Deep Learning Framework 99 | import lasagne 100 | from lasagne.layers import Conv2DLayer as ConvLayer, Pool2DLayer as PoolLayer 101 | from lasagne.layers import InputLayer, ConcatLayer 102 | 103 | print('{} - Using device `{}` for processing the images.{}'.format(ansi.CYAN, theano.config.device, ansi.ENDC)) 104 | 105 | 106 | #---------------------------------------------------------------------------------------------------------------------- 107 | # Convolutional Neural Network 108 | #---------------------------------------------------------------------------------------------------------------------- 109 | class Model(object): 110 | """Store all the data related to the neural network (aka. "model"). This is currently based on VGG19. 111 | """ 112 | 113 | def __init__(self): 114 | self.pixel_mean = np.array([103.939, 116.779, 123.680], dtype=np.float32).reshape((3,1,1)) 115 | 116 | self.setup_model() 117 | self.load_data() 118 | 119 | def setup_model(self, input=None): 120 | """Use lasagne to create a network of convolution layers, first using VGG19 as the framework 121 | and then adding augmentations for Semantic Style Transfer. 122 | """ 123 | net, self.channels = {}, {} 124 | 125 | # Primary network for the main image. These are convolution only, and stop at layer 4_2 (rest unused). 126 | net['img'] = input or InputLayer((None, 3, None, None)) 127 | net['conv1_1'] = ConvLayer(net['img'], 64, 3, pad=1) 128 | net['conv1_2'] = ConvLayer(net['conv1_1'], 64, 3, pad=1) 129 | net['pool1'] = PoolLayer(net['conv1_2'], 2, mode='average_exc_pad') 130 | net['conv2_1'] = ConvLayer(net['pool1'], 128, 3, pad=1) 131 | net['conv2_2'] = ConvLayer(net['conv2_1'], 128, 3, pad=1) 132 | net['pool2'] = PoolLayer(net['conv2_2'], 2, mode='average_exc_pad') 133 | net['conv3_1'] = ConvLayer(net['pool2'], 256, 3, pad=1) 134 | net['conv3_2'] = ConvLayer(net['conv3_1'], 256, 3, pad=1) 135 | net['conv3_3'] = ConvLayer(net['conv3_2'], 256, 3, pad=1) 136 | net['conv3_4'] = ConvLayer(net['conv3_3'], 256, 3, pad=1) 137 | net['pool3'] = PoolLayer(net['conv3_4'], 2, mode='average_exc_pad') 138 | net['conv4_1'] = ConvLayer(net['pool3'], 512, 3, pad=1) 139 | # 512 filters, 3x3 140 | net['conv4_2'] = ConvLayer(net['conv4_1'], 512, 3, pad=1) 141 | net['conv4_3'] = ConvLayer(net['conv4_2'], 512, 3, pad=1) 142 | net['conv4_4'] = ConvLayer(net['conv4_3'], 512, 3, pad=1) 143 | net['pool4'] = PoolLayer(net['conv4_4'], 2, mode='average_exc_pad') 144 | net['conv5_1'] = ConvLayer(net['pool4'], 512, 3, pad=1) 145 | net['conv5_2'] = ConvLayer(net['conv5_1'], 512, 3, pad=1) 146 | net['conv5_3'] = ConvLayer(net['conv5_2'], 512, 3, pad=1) 147 | net['conv5_4'] = ConvLayer(net['conv5_3'], 512, 3, pad=1) 148 | net['main'] = net['conv5_4'] 149 | 150 | # Auxiliary network for the semantic layers, and the nearest neighbors calculations. 151 | # batch size: 1. num_filters: 1 (temporary. will be updated later) 152 | net['map'] = InputLayer((1, 1, None, None)) 153 | for j, i in itertools.product(range(5), range(4)): 154 | if j < 2 and i > 1: continue 155 | suffix = '%i_%i' % (j+1, i+1) 156 | 157 | if i == 0: 158 | # PoolLayer: Pool2DLayer. 2**j: pooling region. mode: mean pooling (majority voting) 159 | # all pooling layers net['map1'], ..., net['map5'] are fed by the same net['map'] 160 | # net['map*'] are at decreasing granularity, in sync with conv* 161 | # pooling layer keeps the channel num of the input layer net['map']. In the example, 3 162 | net['map%i'%(j+1)] = PoolLayer(net['map'], 2**j, mode='average_exc_pad') 163 | # channels[layer]: number of filters 164 | self.channels[suffix] = net['conv'+suffix].num_filters 165 | 166 | # sem* is used to calculate style loss 167 | # most sem* layers are not used, except 2 layers 168 | # semantic_weight default: 10 169 | if args.semantic_weight > 0.0: 170 | # sem1_1 <= conv1_1 + map1 171 | # sem1_2 <= conv1_2 + map1 172 | # sem2_1 <= conv2_1 + map2 173 | # sem2_2 <= conv2_2 + map2 174 | # sem3_1 <= conv3_1 + map3 175 | # sem3_2 <= conv3_2 + map3 176 | # sem3_3 <= conv3_3 + map3 177 | # sem3_4 <= conv3_4 + map3 178 | # sem4_1 <= conv4_1 + map4 179 | # sem4_2 <= conv4_2 + map4 180 | # sem4_3 <= conv4_3 + map4 181 | # sem4_4 <= conv4_4 + map4 182 | # sem5_1 <= conv5_1 + map5 183 | # sem5_2 <= conv5_2 + map5 184 | # sem5_3 <= conv5_3 + map5 185 | # sem5_4 <= conv5_4 + map5 186 | net['sem'+suffix] = ConcatLayer([net['conv'+suffix], net['map%i'%(j+1)]]) 187 | else: 188 | net['sem'+suffix] = net['conv'+suffix] 189 | 190 | net['dup'+suffix] = InputLayer(net['sem'+suffix].output_shape) 191 | # num_filters=1 is only a placeholder, will be updated to the number of patches in each slice 192 | # nn filter size: 3*3 193 | net['nn'+suffix] = ConvLayer(net['dup'+suffix], 1, 3, b=None, pad=0, flip_filters=False) 194 | shape = net['nn'+suffix].W.get_value().shape 195 | net['nn'+suffix].W = theano.shared( net['nn'+suffix].W.get_value(), broadcastable=[False]* len(shape) ) 196 | 197 | # laplacian layer 198 | # do pooling first to reduce effective image size and reduce required memory & running time 199 | net['pool1_1'] = PoolLayer(net['img'], 2, mode='average_exc_pad') 200 | net['conv1_laplace'] = ConvLayer(net['pool1_1'], 1, 3, pad=1) 201 | laplacian = np.array( [ [0,-1,0], [-1,4,-1], [0,-1,0] ], dtype=theano.config.floatX ) 202 | W = np.zeros((1, 3, 3, 3), dtype=theano.config.floatX) 203 | for t in range(3): 204 | W[0,t] = laplacian 205 | net['conv1_laplace'].W.set_value(W) 206 | 207 | net['sem1_laplace'] = net['conv1_laplace'] 208 | net['dup1_laplace'] = InputLayer(net['sem1_laplace'].output_shape) 209 | net['nn1_laplace'] = ConvLayer(net['dup1_laplace'], 1, 3, b=None, pad=0, flip_filters=False) 210 | shape = net['nn1_laplace'].W.get_value().shape 211 | net['nn1_laplace'].W = theano.shared( net['nn1_laplace'].W.get_value(), broadcastable=[False]* len(shape) ) 212 | self.channels['1_laplace'] = net['conv1_laplace'].num_filters 213 | self.network = net 214 | 215 | def load_data(self): 216 | """Open the serialized parameters from a pre-trained network, and load them into the model created. 217 | """ 218 | vgg19_file = os.path.join(os.path.dirname(__file__), 'vgg19_conv.pkl.bz2') 219 | if not os.path.exists(vgg19_file): 220 | error("Model file with pre-trained convolution layers not found. Download here...", 221 | "https://github.com/alexjc/neural-doodle/releases/download/v0.0/vgg19_conv.pkl.bz2") 222 | 223 | data = pickle.load(bz2.open(vgg19_file, 'rb')) 224 | # layers before and including conv5_4. 16 conv layers, 32 parameter arrays 225 | params = lasagne.layers.get_all_param_values(self.network['main']) 226 | lasagne.layers.set_all_param_values(self.network['main'], data[:len(params)]) 227 | 228 | # layers = [ sem3_1, sem4_1, conv4_2 ] & sem1_laplace 229 | def setup(self, layers): 230 | """Setup the inputs and outputs, knowing the layers that are required by the optimization algorithm. 231 | """ 232 | self.tensor_img = T.tensor4() 233 | self.tensor_map = T.tensor4() 234 | # self.network['img']: image input; self.network['map']: semantic input 235 | tensor_inputs = {self.network['img']: self.tensor_img, self.network['map']: self.tensor_map} 236 | # outputs from 3 layers sem3_1, sem4_1, conv4_2, given tensor_img and tensor_map 237 | outputs = lasagne.layers.get_output([self.network[l] for l in layers], tensor_inputs) 238 | self.tensor_outputs = {k: v for k, v in zip(layers, outputs)} 239 | 240 | # type: sem or conv 241 | # type+l: sem3_1, sem4_1 or conv3_1, conv4_1 242 | def get_outputs(self, type, layers): 243 | """Fetch the output tensors for the network layers. 244 | """ 245 | return [self.tensor_outputs[type+l] for l in layers] 246 | 247 | def prepare_image(self, image): 248 | """Given an image loaded from disk, turn it into a representation compatible with the model. 249 | The format is (b,c,y,x) with batch=1 for a single image, channels=3 for RGB, and y,x matching 250 | the resolution. 251 | """ 252 | image = np.swapaxes(np.swapaxes(image, 1, 2), 0, 1)[::-1, :, :] 253 | image = image.astype(np.float32) - self.pixel_mean 254 | return image[np.newaxis] 255 | 256 | def finalize_image(self, image, resolution): 257 | """Based on the output of the neural network, convert it into an image format that can be saved 258 | to disk -- shuffling dimensions as appropriate. 259 | """ 260 | image = np.swapaxes(np.swapaxes(image[::-1], 0, 1), 1, 2) 261 | image = np.clip(image, 0, 255).astype('uint8') 262 | return scipy.misc.imresize(image, resolution, interp='bicubic') 263 | 264 | 265 | #---------------------------------------------------------------------------------------------------------------------- 266 | # Semantic Style Transfer 267 | #---------------------------------------------------------------------------------------------------------------------- 268 | class NeuralGenerator(object): 269 | """This is the main part of the application that generates an image using optimization and LBFGS. 270 | The images will be processed at increasing resolutions in the run() method. 271 | """ 272 | 273 | def __init__(self): 274 | """Constructor sets up global variables, loads and validates files, then builds the model. 275 | """ 276 | self.start_time = time.time() 277 | self.style_cache = {} 278 | # style_layers: 3_1, 4_1, 1_laplace 279 | self.style_layers = args.style_layers.split(',') 280 | self.content_layers = args.content_layers.split(',') 281 | self.used_layers = self.style_layers + self.content_layers 282 | 283 | # Prepare file output and load files specified as input. 284 | if args.save_every is not None: 285 | os.makedirs('frames', exist_ok=True) 286 | if args.output is not None and os.path.isfile(args.output): 287 | os.remove(args.output) 288 | 289 | print(ansi.CYAN, end='') 290 | target = args.content or args.output 291 | self.content_img_original, self.content_map_original = self.load_images('content', target) 292 | self.style_img_original, self.style_map_original = self.load_images('style', args.style) 293 | 294 | if self.content_map_original is None and self.content_img_original is None: 295 | print(" - No content files found; result depends on seed only.") 296 | print(ansi.ENDC, end='') 297 | 298 | # Display some useful errors if the user's input can't be understood. 299 | if self.style_img_original is None: 300 | error("Couldn't find style image as expected.", 301 | " - Try making sure `{}` exists and is a valid image.".format(args.style)) 302 | 303 | if self.content_map_original is not None and self.style_map_original is None: 304 | basename, _ = os.path.splitext(args.style) 305 | error("Expecting a semantic map for the input style image too.", 306 | " - Try creating the file `{}_sem.png` with your annotations.".format(basename)) 307 | 308 | if self.style_map_original is not None and self.content_map_original is None: 309 | basename, _ = os.path.splitext(target) 310 | error("Expecting a semantic map for the input content image too.", 311 | " - Try creating the file `{}_sem.png` with your annotations.".format(basename)) 312 | 313 | if self.content_map_original is None: 314 | if self.content_img_original is None and args.output_size: 315 | shape = tuple([int(i) for i in args.output_size.split('x')]) 316 | else: 317 | shape = self.style_img_original.shape[:2] 318 | 319 | self.content_map_original = np.zeros(shape+(3,)) 320 | args.semantic_weight = 0.0 321 | 322 | if self.style_map_original is None: 323 | self.style_map_original = np.zeros(self.style_img_original.shape[:2]+(3,)) 324 | args.semantic_weight = 0.0 325 | 326 | if self.content_img_original is None: 327 | self.content_img_original = np.zeros(self.content_map_original.shape[:2]+(3,)) 328 | args.content_weight = 0.0 329 | 330 | if self.content_map_original.shape[2] != self.style_map_original.shape[2]: 331 | error("Mismatch in number of channels for style and content semantic map.", 332 | " - Make sure both images are RGB, RGBA, or L.") 333 | 334 | # Finalize the parameters based on what we loaded, then create the model. 335 | args.semantic_weight = math.sqrt(9.0 / args.semantic_weight) if args.semantic_weight else 0.0 336 | self.model = Model() 337 | 338 | 339 | #------------------------------------------------------------------------------------------------------------------ 340 | # Helper Functions 341 | #------------------------------------------------------------------------------------------------------------------ 342 | 343 | def load_images(self, name, filename): 344 | """If the image and map files exist, load them. Otherwise they'll be set to default values later. 345 | """ 346 | basename, _ = os.path.splitext(filename) 347 | mapname = basename + args.semantic_ext 348 | img = scipy.ndimage.imread(filename, mode='RGB') if os.path.exists(filename) else None 349 | map = scipy.ndimage.imread(mapname) if os.path.exists(mapname) and args.semantic_weight > 0.0 else None 350 | 351 | if img is not None: print(' - Loading `{}` for {} data.'.format(filename, name)) 352 | if map is not None: print(' - Adding `{}` as semantic map.'.format(mapname)) 353 | 354 | if img is not None and map is not None and img.shape[:2] != map.shape[:2]: 355 | error("The {} image and its semantic map have different resolutions. Either:".format(name), 356 | " - Resize {} to {}, or\n - Resize {} to {}."\ 357 | .format(filename, map.shape[1::-1], mapname, img.shape[1::-1])) 358 | return img, map 359 | 360 | def compile(self, arguments, function): 361 | """Build a Theano function that will run the specified expression on the GPU. 362 | """ 363 | return theano.function(list(arguments), function, on_unused_input='ignore') 364 | 365 | def compute_norms(self, backend, layer, array): 366 | # self.model.channels[layer]: num of filters of conv_layer 367 | # e.g. layer=3_1, then it's the num of filters of conv3_1 368 | # if layer==sem3_1(4_1): first half, i.e., conv3_1(4_1) output norms 369 | # if layer==conv3_1(4_1): all the output norms 370 | ni = backend.sqrt(backend.sum(array[:,:self.model.channels[layer]] ** 2.0, axis=(1,), keepdims=True)) 371 | # if layer==sem3_1(4_1): second half, i.e., map3_1(4_1) output norms 372 | # if layer==conv3_1(4_1): empty array 373 | ns = backend.sqrt(backend.sum(array[:,self.model.channels[layer]:] ** 2.0, axis=(1,), keepdims=True)) 374 | return [ni] + [ns] 375 | 376 | def normalize_components(self, layer, array, norms): 377 | if args.style_weight > 0.0: 378 | array[:,:self.model.channels[layer]] /= (norms[0] * 3.0) 379 | if args.semantic_weight > 0.0: 380 | array[:,self.model.channels[layer]:] /= (norms[1] * args.semantic_weight) 381 | 382 | 383 | #------------------------------------------------------------------------------------------------------------------ 384 | # Initialization & Setup 385 | #------------------------------------------------------------------------------------------------------------------ 386 | 387 | def rescale_image(self, img, scale): 388 | """Re-implementing skimage.transform.scale without the extra dependency. Saves a lot of space and hassle! 389 | """ 390 | output = scipy.misc.toimage(img, cmin=0.0, cmax=255) 391 | output.thumbnail((int(output.size[0]*scale), int(output.size[1]*scale)), PIL.Image.ANTIALIAS) 392 | return np.asarray(output) 393 | 394 | def prepare_content(self, scale=1.0): 395 | """Called each phase of the optimization, rescale the original content image and its map to use as inputs. 396 | """ 397 | content_img = self.rescale_image(self.content_img_original, scale) 398 | self.content_img = self.model.prepare_image(content_img) 399 | 400 | content_map = self.rescale_image(self.content_map_original, scale) 401 | self.content_map = content_map.transpose((2, 0, 1))[np.newaxis].astype(np.float32) 402 | 403 | def prepare_style(self, scale=1.0): 404 | """Called each phase of the optimization, process the style image according to the scale, then run it 405 | through the model to extract intermediate outputs (e.g. sem4_1) and turn them into patches. 406 | """ 407 | style_img = self.rescale_image(self.style_img_original, scale) 408 | self.style_img = self.model.prepare_image(style_img) 409 | 410 | style_map = self.rescale_image(self.style_map_original, scale) 411 | # style_map.shape: (63, 52, 3) 412 | self.style_map = style_map.transpose((2, 0, 1))[np.newaxis].astype(np.float32) 413 | # self.style_map.shape: (1, 3, 63, 52). channel num: 3 414 | 415 | # Compile a function to run on the GPU to extract patches for all layers at once. 416 | # layer_outputs: [ ('3_1', sem3_1_output), ('4_1', sem4_1_output) ] 417 | # sem3_1 = conv3_1 + map3, sem4_1 = conv4_1 + map4 418 | layer_outputs = zip(self.style_layers, self.model.get_outputs('sem', self.style_layers)) 419 | # ext_patches: symbolic results from extracting patches from output of 420 | # 'sem3_1' & 'sem4_1'. Input to the network is required 421 | ext_patches = self.do_extract_patches(layer_outputs) 422 | # extractor: 6 symbolic operators that take two inputs 423 | # Assign inputs to tensor_img & tensor_map. Get output from tensor_outputs 424 | # Then extract patches from tensor_outputs 425 | extractor = self.compile([self.model.tensor_img, self.model.tensor_map], ext_patches) 426 | # feed two inputs into each of the 6 symbolic operators 427 | # Assign tensor_img = self.style_img, tensor_map = self.style_map 428 | 429 | # self.style_map.shape: (1, 3, 63, 52). channel num: 3 430 | result = extractor(self.style_img, self.style_map) 431 | # result: conv output of sem*, given style_img | style_map 432 | # a list of 6 arrays in shapes: 433 | # result[0::3] result[1::3] result[2::3] 434 | # sem3_1 (143, 259, 3, 3), (143, 1, 3, 3), (143, 1, 3, 3), 435 | # sem4_1 (20, 515, 3, 3), (20, 1, 3, 3), (20, 1, 3, 3) 436 | # sem3_1_outneighbs conv3_1_neibnorms map3_1_neibnorms 437 | # sem4_1_outneighbs conv4_1_neibnorms map4_1_neibnorms 438 | # Store all the style patches layer by layer, split to match slice size and cast to 16-bit for size. 439 | self.style_data = {} 440 | for layer, *data in zip(self.style_layers, result[0::3], result[1::3], result[2::3]): 441 | # data[0]: neighbors of sem* output 442 | # data[1]: norms of sem* output 443 | # data[2]: norms of semantic pooling layer output 444 | patches = data[0] 445 | l = self.model.network['nn'+layer] 446 | # args.slices: 'Split patches up into this number of batches.' Default: 2 447 | # nn3_1(4_1).num_filters: initialized to the patch num in each slice. 448 | # the original num_filters = '1' has been changed 449 | # nn3_1(4_1).W will be set to each patch later 450 | l.num_filters = patches.shape[0] // args.slices 451 | # self.style_data['3_1']: [0]: sem3_1_outneighbs, [1]: conv3_1_neibnorms, 452 | # [2]: map3_1_neibnorms, [3]: zeros(660) - storing matching history used in evaluate_slices() 453 | self.style_data[layer] = [d[:l.num_filters*args.slices].astype(np.float16) for d in data]\ 454 | + [np.zeros((patches.shape[0],), dtype=np.float16)] 455 | print(' - Style layer {}: {} patches in {:,}kb.'.format(layer, patches.shape, patches.size//1000)) 456 | 457 | # style_data['3_1'][0]: (142, 259, 3, 3) 458 | # style_data['4_1'][0]: (20, 515, 3, 3) 459 | # num_filters - nn3_1: 71, nn4_1: 10 460 | 461 | def prepare_optimization(self): 462 | """Optimization requires a function to compute the error (aka. loss) which is done in multiple components. 463 | Here we compile a function to run on the GPU that returns all components separately. 464 | """ 465 | 466 | # Feed-forward calculation only, returns the result of the convolution post-activation 467 | self.compute_features = self.compile([self.model.tensor_img, self.model.tensor_map], 468 | self.model.get_outputs('sem', self.style_layers)) 469 | 470 | # Patch matching calculation that uses only pre-calculated features and a slice of the patches. 471 | 472 | # create empty Theano shared variable for matcher_history of 3_1, 4_1 and 1_laplace 473 | self.matcher_tensors = {l: lasagne.utils.shared_empty(dim=4) for l in self.style_layers} 474 | # matcher_history will be set to style_data[-1] later 475 | self.matcher_history = {l: T.vector() for l in self.style_layers} 476 | # dup3_1 = matcher_tensors['3_1'], dup4_1 = matcher_tensors['4_1'] 477 | # matcher_tensors will be set to normalized current_features in evaluate(): 478 | # sem3_1, sem4_1 conv output given content_img 479 | # so dup* = normalized conv output of sem* on content_img 480 | # dup* is the input of sem* 481 | self.matcher_inputs = {self.model.network['dup'+l]: self.matcher_tensors[l] for l in self.style_layers} 482 | # nn_layers = ['nn3_1', 'nn4_1'] 483 | # nn*.W will be set to normalized style patches in evaluate_slices() <- evaluate() <- fmin_l_bfgs_b() 484 | nn_layers = [self.model.network['nn'+l] for l in self.style_layers] 485 | # 3_1 => nn3_1_output, 4_1 => nn4_1_output, given dup* = conv output of sem* on content_img 486 | self.matcher_outputs = dict(zip(self.style_layers, lasagne.layers.get_output(nn_layers, self.matcher_inputs))) 487 | 488 | # conv of nn* computes the cross correlation between conv outputs of sem*, given content_img and style_img, respectively 489 | # do_match_patches() will find the patch indices corresponding to max conv output 490 | self.compute_matches = {l: self.compile([self.matcher_history[l]], self.do_match_patches(l))\ 491 | for l in self.style_layers} 492 | 493 | self.tensor_matches = [T.tensor4() for l in self.style_layers] 494 | # Build a list of Theano expressions that, once summed up, compute the total error. 495 | self.losses = self.content_loss() + self.total_variation_loss() + self.style_loss() 496 | # losses = [('smooth', 'img', Elemwise{mul,no_inplace}.0), ('style', '3_1', Elemwise{mul,no_inplace}.0), 497 | # ('style', '4_1', Elemwise{mul,no_inplace}.0)] 498 | # 'smooth': total variation loss 499 | # Let Theano automatically compute the gradient of the error, used by LBFGS to update image pixels. 500 | grad = T.grad(sum([l[-1] for l in self.losses]), self.model.tensor_img) 501 | # Create a single function that returns the gradient and the individual errors components. 502 | self.compute_grad_and_losses = theano.function( 503 | [self.model.tensor_img, self.model.tensor_map] + self.tensor_matches, 504 | [grad] + [l[-1] for l in self.losses], on_unused_input='ignore') 505 | 506 | 507 | #------------------------------------------------------------------------------------------------------------------ 508 | # Theano Computation 509 | #------------------------------------------------------------------------------------------------------------------ 510 | 511 | # layers: [ ('3_1', sem(conv)3_1_output), ('4_1', sem(conv)4_1_output) ] 512 | # extract 3*3 patches from the conv output 513 | # size = the filter size in nn* layers. So the patches will be assigned to nn*.W 514 | def do_extract_patches(self, layers, size=3, stride=1): 515 | """This function builds a Theano expression that will get compiled an run on the GPU. It extracts 3x3 patches 516 | from the intermediate outputs in the model. 517 | """ 518 | results = [] 519 | # l: '3_1', f: sem3_1_output 520 | # sem3_1_output.shape: [1,259,15,13] 521 | # 259: conv3_1: 256 + map3_1: 3 522 | # patches.shape: [10300,9]: 10300 = 259 * (15-2) * (13-2) 523 | # patches shape after reshape: [143,259,3,3] 524 | for l, f in layers: 525 | # Use a Theano helper function to extract "neighbors" of specific size, seems a bit slower than doing 526 | # it manually but much simpler! 527 | # images2neibs() gets small patches in (size*size) from sem(conv)3(4)_1_output, 528 | # i.e. small patches in the neural encoding tensor 529 | # first dimension of "patches" is "filter", so in sem*, the first dimensionality is num_filters of conv* + map* 530 | patches = theano.tensor.nnet.neighbours.images2neibs(f, (size, size), (stride, stride), mode='valid') 531 | # Make sure the patches are in the shape required to insert them into the model as another layer. 532 | patches = patches.reshape((-1, patches.shape[0] // f.shape[1], size, size)).dimshuffle((1, 0, 2, 3)) 533 | # Calculate the magnitude that we'll use for normalization at runtime, then store... 534 | # each call of compute_norms() returns two arrays 535 | # if l==sem*, conv* neighbor norms and map* neighbor norms 536 | # if l==conv*, conv* neighbor norms and an empty array 537 | results.extend([patches] + self.compute_norms(T, l, patches)) 538 | # self.compute_norms(T, l, patches) returns the (symbolic) magnitude for normalization 539 | return results 540 | 541 | def do_match_patches(self, layer): 542 | # Use node in the model to compute the result of the normalized cross-correlation, using results from the 543 | # nearest-neighbor layers called 'nn3_1' and 'nn4_1'. 544 | # dist: distances (nn_layer outputs) 545 | # dist from nn3_1: [1, 71, 14, 9] 546 | # dist from nn4_1: [1, 10, 6, 3] 547 | # matcher takes normalized content features as matcher_inputs, and normalized sem* features as weights 548 | # matcher_outputs are normalized cross-correlations between normalized content features and sem* features 549 | dist = self.matcher_outputs[layer] 550 | # dist.shape is now (71,126) or (10, 18) 551 | dist = dist.reshape((dist.shape[1], -1)) 552 | # Compute the score of each patch, taking into account statistics from previous iteration. This equalizes 553 | # the chances of the patches being selected when the user requests more variety. 554 | # matcher_history[layer] <= dist.max(axis=1), the best dist for each filter in the last run 555 | offset = self.matcher_history[layer].reshape((-1, 1)) 556 | scores = (dist - offset * args.variety) 557 | # Pick the best style patches for each patch in the current image, the result is an array of indices. 558 | # Also return the maximum value along both axis, used to compare slices and add patch variety. 559 | # axis 0: filter num, axis 1: patch num 560 | # argmax(axis=0): 126(18), argmax for each patch 561 | # argmax(axis=1): 71(10), argmax for each filter 562 | return [scores.argmax(axis=0), scores.max(axis=0), dist.max(axis=1)] 563 | 564 | 565 | #------------------------------------------------------------------------------------------------------------------ 566 | # Error/Loss Functions 567 | #------------------------------------------------------------------------------------------------------------------ 568 | 569 | def content_loss(self): 570 | """Return a list of Theano expressions for the error function, measuring how different the current image is 571 | from the reference content that was loaded. 572 | """ 573 | 574 | content_loss = [] 575 | if args.content_weight == 0.0: 576 | return content_loss 577 | 578 | # First extract all the features we need from the model, these results after convolution. 579 | # content_layers: conv4_2 (only one layer) 580 | extractor = theano.function([self.model.tensor_img], self.model.get_outputs('conv', self.content_layers)) 581 | # conv output of the original content image. 512d 582 | result = extractor(self.content_img) 583 | 584 | # Build a list of loss components that compute the mean squared error by comparing current result to desired. 585 | for l, ref in zip(self.content_layers, result): 586 | # input: tensor_img = current_img, tensor_map = content_map 587 | # layer: output from conv4_2 588 | layer = self.model.tensor_outputs['conv'+l] 589 | loss = T.mean((layer - ref) ** 2.0) 590 | 591 | if 'lap' not in l: 592 | content_loss.append(('content', l, args.content_weight * loss)) 593 | else: 594 | content_loss.append(('content', l, args.lap_weight * loss)) 595 | 596 | print(' - Content layer conv{}: {} features in {:,}kb.'.format(l, ref.shape[1], ref.size//1000)) 597 | return content_loss 598 | 599 | def style_loss(self): 600 | """Returns a list of loss components as Theano expressions. Finds the best style patch for each patch in the 601 | current image using normalized cross-correlation, then computes the mean squared error for all patches. 602 | """ 603 | style_loss = [] 604 | if args.style_weight == 0.0: 605 | return style_loss 606 | 607 | # Extract the patches from the current image, as well as their magnitude. 608 | result = self.do_extract_patches(zip(self.style_layers, self.model.get_outputs('conv', self.style_layers))) 609 | 610 | # Multiple style layers are optimized separately, usually conv3_1 and conv4_1. Semantic data not used here. 611 | # Semantic data is only used for selecting nearest neighbor style patches 612 | # tensor_matches = current_best, i.e. indices of best matching patches in each layer 613 | for l, matches, patches in zip(self.style_layers, self.tensor_matches, result[0::3]): 614 | # Compute the mean squared error between the current patch and the best matching style patch. 615 | # Ignore the last channels (from semantic map) so errors returned are indicative of image only. 616 | # matches = tensor_matches[i] = current_best[i] 617 | loss = T.mean((patches - matches[:,:self.model.channels[l]]) ** 2.0) 618 | if 'laplace' not in l: 619 | style_loss.append(('style', l, args.style_weight * loss)) 620 | else: 621 | style_loss.append(('style', l, args.style_lapweight * loss)) 622 | 623 | return style_loss 624 | 625 | def total_variation_loss(self): 626 | """Return a loss component as Theano expression for the smoothness prior on the result image. 627 | """ 628 | # here tensor_img is always current_img 629 | x = self.model.tensor_img 630 | loss = (((x[:,:,:-1,:-1] - x[:,:,1:,:-1])**2 + (x[:,:,:-1,:-1] - x[:,:,:-1,1:])**2)**1.25).mean() 631 | return [('smooth', 'img', args.smoothness * loss)] 632 | 633 | #------------------------------------------------------------------------------------------------------------------ 634 | # Optimization Loop 635 | #------------------------------------------------------------------------------------------------------------------ 636 | 637 | def iterate_batches(self, *arrays, batch_size): 638 | """Break down the data in arrays batch by batch and return them as a generator. 639 | """ 640 | total_size = arrays[0].shape[0] 641 | indices = np.arange(total_size) 642 | for index in range(0, total_size, batch_size): 643 | excerpt = indices[index:index + batch_size] 644 | yield excerpt, [a[excerpt] for a in arrays] 645 | 646 | # argument f is not directly used in this function 647 | # f is set to matcher_tensors['3_1'('4_1')]. 648 | # dup3_1(4_1) = matcher_tensors['3_1'('4_1')], input of nn3_1(4_1) 649 | # input to nn3_1: (1, 259, 16, 11) 650 | # input to nn4_1: (1, 515, 8, 5) 651 | # output from nn3_1: [1, 71, 14, 9] 652 | # output from nn4_1: [1, 10, 6, 3] 653 | # best_idx from nn3_1: 126 (length) 654 | # best_idx from nn4_1: 18 (length) 655 | def evaluate_slices(self, f, l): 656 | # if cache is on, only use previously saved matches 657 | if args.cache and l in self.style_cache: 658 | return self.style_cache[l] 659 | 660 | # style_data: style patches & norms, i.e. 'sem3(4)_1' output given style_img & style_map 661 | # these patches will be set to the weights of 'nn3(4)_1' 662 | layer, data = self.model.network['nn'+l], self.style_data[l] 663 | # history: 143 or 20. Effective: 142 or 20 664 | history = data[-1] 665 | 666 | best_idx, best_val = None, 0.0 667 | # bp: sem3(4)_1_outneighbs, bi: conv3(4)_1_neibnorms, 668 | # bs: map3(4)_1_neibnorms, bh: (660) - history slice (not used: history is directly accessed) 669 | for idx, (bp, bi, bs, bh) in self.iterate_batches(*data, batch_size=layer.num_filters): 670 | weights = bp.astype(np.float32) 671 | # after normalization the conv result will be the cross correlation between sem* patches and the img patches 672 | self.normalize_components(l, weights, (bi, bs)) 673 | # weights.shape: (71, 259, 3, 3) 674 | # nn3_1(4_1) num_filters (output channels): 71. input channels: 259 675 | # print nn3_1.input_shape: (1,257,None,None). 676 | # 257 is based on the initialized map shape: 1,1,..,.. 677 | # however the actual map shape is 1,3,..,... so nn3_1.input_shape should be 1,259,..,.. 678 | layer.W.set_value(weights) 679 | 680 | # history[idx] works as offsets to the correlations 681 | # encourage matches that are different from the previous iteration, to boost diversity 682 | # max indices, max scores (dist - offset), max dists 683 | # (126)/(18) (126)/(18) (71)/(10) 684 | cur_idx, cur_val, cur_match = self.compute_matches[l](history[idx]) 685 | 686 | if best_idx is None: 687 | best_idx, best_val = cur_idx, cur_val 688 | else: 689 | i = np.where(cur_val > best_val) 690 | # update those indices where cur_val are larger. i is a boolean array 691 | best_idx[i] = idx[cur_idx[i]] 692 | best_val[i] = cur_val[i] 693 | 694 | # idx: indices for the current batch. 0..71 & 71..142 695 | history[idx] = cur_match 696 | 697 | if args.cache: 698 | self.style_cache[l] = best_idx 699 | return best_idx 700 | 701 | def varshape(self): 702 | return [ self.matcher_outputs[l].shape for l in self.style_layers ] 703 | 704 | def evaluate(self, Xn): 705 | """Callback for the L-BFGS optimization that computes the loss and gradients on the GPU. 706 | """ 707 | # Xn: content_img or noisy input (depending on the program argument) 708 | # Adjust the representation to be compatible with the model before computing results. 709 | # demean the original content image as the initialized transfer image 710 | current_img = Xn.reshape(self.content_img.shape).astype(np.float32) - self.model.pixel_mean 711 | # tensor_img = content_img, tensor_map = content_map 712 | # current_features: sem3_1, sem4_1 conv output given content_img 713 | current_features = self.compute_features(current_img, self.content_map) 714 | 715 | # Iterate through each of the style layers one by one, computing best matches. 716 | current_best = [] 717 | for l, f in zip(self.style_layers, current_features): 718 | # content patches are normalized here 719 | self.normalize_components(l, f, self.compute_norms(np, l, f)) 720 | self.matcher_tensors[l].set_value(f) 721 | # input to nn3_1: (1, 259, 16, 11) 722 | # input to nn4_1: (1, 515, 8, 5) 723 | 724 | # Compute best matching patches in this style layer, going through all slices. 725 | warmup = bool(args.variety > 0.0 and self.iteration == 0) 726 | for _ in range(2 if warmup else 1): 727 | best_idx = self.evaluate_slices(f, l) 728 | 729 | patches = self.style_data[l][0] 730 | current_best.append(patches[best_idx].astype(np.float32)) 731 | # current_best: (126, 259, 3, 3), (18, 515, 3, 3) 732 | # output from nn3_1: [1, 71, 14, 9] 733 | # output from nn4_1: [1, 10, 6, 3] 734 | #varshape = self.compile( [self.model.tensor_img, self.model.tensor_map], self.varshape() ) 735 | #shapes = varshape(current_img, self.content_map) 736 | 737 | # tensor_img = current_img, tensor_map = content_map, tensor_matches = current_best 738 | grads, *losses = self.compute_grad_and_losses(current_img, self.content_map, *current_best) 739 | if np.isnan(grads).any(): 740 | raise OverflowError("Optimization diverged; try using a different device or parameters.") 741 | 742 | # Use magnitude of gradients as an estimate for overall quality. 743 | self.error = self.error * 0.9 + 0.1 * min(np.abs(grads).max(), 255.0) 744 | loss = sum(losses) 745 | 746 | # Dump the image to disk if requested by the user. 747 | if args.save_every and self.frame % args.save_every == 0: 748 | frame = Xn.reshape(self.content_img.shape[1:]) 749 | resolution = self.content_img_original.shape 750 | image = scipy.misc.toimage(self.model.finalize_image(frame, resolution), cmin=0, cmax=255) 751 | image.save('frames/%04d.png'%self.frame) 752 | 753 | # Print more information to the console every few iterations. 754 | if args.print_every and self.frame % args.print_every == 0: 755 | print('{:>3} {}loss{} {:8.2e} '.format(self.frame, ansi.BOLD, ansi.ENDC, loss / 1000.0), end='') 756 | category = '' 757 | for v, l in zip(losses, self.losses): 758 | if l[0] == 'smooth': 759 | continue 760 | if l[0] != category: 761 | print(' {}{}{}'.format(ansi.BOLD, l[0], ansi.ENDC), end='') 762 | category = l[0] 763 | print(' {}{}{} {:8.2e} '.format(ansi.BOLD, l[1], ansi.ENDC, v / 1000.0), end='') 764 | 765 | current_time = time.time() 766 | quality = 100.0 - 100.0 * np.sqrt(self.error / 255.0) 767 | print(' {}quality{} {: >4.1f}% '.format(ansi.BOLD, ansi.ENDC, quality), end='') 768 | print(' {}time{} {:3.1f}s '.format(ansi.BOLD, ansi.ENDC, current_time - self.iter_time), flush=True) 769 | self.iter_time = current_time 770 | 771 | # Update counters and timers. 772 | self.frame += 1 773 | self.iteration += 1 774 | 775 | # Return the data in the right format for L-BFGS. 776 | return loss, np.array(grads).flatten().astype(np.float64) 777 | 778 | def run(self): 779 | """The main entry point for the application, runs through multiple phases at increasing resolutions. 780 | """ 781 | self.frame, Xn = 0, None 782 | for i in range(args.phases): 783 | self.error = 255.0 784 | scale = 1.0 / 2.0 ** (args.phases - 1 - i) 785 | 786 | shape = self.content_img_original.shape 787 | print('\n{}Phase #{}: resolution {}x{} scale {}{}'\ 788 | .format(ansi.BLUE_B, i, int(shape[1]*scale), int(shape[0]*scale), scale, ansi.BLUE)) 789 | 790 | # Precompute all necessary data for the various layers, put patches in place into augmented network. 791 | self.model.setup(layers=['sem'+l for l in self.style_layers] + ['conv'+l for l in self.content_layers]) 792 | self.prepare_content(scale) 793 | self.prepare_style(scale) 794 | 795 | # Now setup the model with the new data, ready for the optimization loop. 796 | self.model.setup(layers=['sem'+l for l in self.style_layers] + ['conv'+l for l in self.used_layers]) 797 | self.prepare_optimization() 798 | print('{}'.format(ansi.ENDC)) 799 | 800 | # Setup the seed for the optimization as specified by the user. 801 | shape = self.content_img.shape[2:] 802 | if args.seed == 'content': 803 | Xn = self.content_img[0] + self.model.pixel_mean 804 | if args.seed == 'noise': 805 | bounds = [int(i) for i in args.seed_range.split(':')] 806 | Xn = np.random.uniform(bounds[0], bounds[1], shape + (3,)).astype(np.float32) 807 | if args.seed == 'previous': 808 | Xn = scipy.misc.imresize(Xn[0], shape, interp='bicubic') 809 | Xn = Xn.transpose((2, 0, 1))[np.newaxis] 810 | if os.path.exists(args.seed): 811 | seed_image = scipy.ndimage.imread(args.seed, mode='RGB') 812 | seed_image = scipy.misc.imresize(seed_image, shape, interp='bicubic') 813 | self.seed_image = self.model.prepare_image(seed_image) 814 | Xn = self.seed_image[0] + self.model.pixel_mean 815 | if Xn is None: 816 | error("Seed for optimization was not found. You can either...", 817 | " - Set the `--seed` to `content` or `noise`.", " - Specify `--seed` as a valid filename.") 818 | 819 | # Optimization algorithm needs min and max bounds to prevent divergence. 820 | data_bounds = np.zeros((np.product(Xn.shape), 2), dtype=np.float64) 821 | data_bounds[:] = (0.0, 255.0) 822 | 823 | self.iter_time, self.iteration, interrupt = time.time(), 0, False 824 | # Xn: the input content image 825 | # output new Xn: the best image that minimizes the losses 826 | try: 827 | Xn, Vn, info = scipy.optimize.fmin_l_bfgs_b( 828 | self.evaluate, 829 | Xn.astype(np.float64).flatten(), 830 | bounds=data_bounds, 831 | factr=0.0, pgtol=0.0, # Disable automatic termination, set low threshold. 832 | m=5, # Maximum correlations kept in memory by algorithm. 833 | maxfun=args.iterations-1, # Limit number of calls to evaluate(). 834 | iprint=-1) # Handle our own logging of information. 835 | except OverflowError: 836 | error("The optimization diverged and NaNs were encountered.", 837 | " - Try using a different `--device` or change the parameters.", 838 | " - Make sure libraries are updated to work around platform bugs.") 839 | except KeyboardInterrupt: 840 | interrupt = True 841 | 842 | args.seed = 'previous' 843 | resolution = self.content_img.shape 844 | Xn = Xn.reshape(resolution) 845 | 846 | output = self.model.finalize_image(Xn[0], self.content_img_original.shape) 847 | scipy.misc.toimage(output, cmin=0, cmax=255).save(args.output) 848 | if interrupt: break 849 | 850 | status = "finished in" if not interrupt else "interrupted at" 851 | print('\n{}Optimization {} {:3.1f}s, average pixel error {:3.1f}!{}\n'\ 852 | .format(ansi.CYAN, status, time.time() - self.start_time, self.error, ansi.ENDC)) 853 | 854 | 855 | if __name__ == "__main__": 856 | generator = NeuralGenerator() 857 | generator.run() 858 | -------------------------------------------------------------------------------- /output/boy_girl3_5_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/boy_girl3_5_0.png -------------------------------------------------------------------------------- /output/boy_girl3_5_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/boy_girl3_5_100.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_1,2_50,100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_1,2_50,100.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_1,3_50,400.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_1,3_50,400.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_1_50.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_1_50.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_2,3_100,400.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_2,3_100,400.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_2,4_100,1600.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_2,4_100,1600.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch5_2_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch5_2_100.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch_5_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch_5_0.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch_5_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch_5_100.png -------------------------------------------------------------------------------- /output/combinations/girlmrf_dogsketch_5_200.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/combinations/girlmrf_dogsketch_5_200.png -------------------------------------------------------------------------------- /output/girl2_cartoon2_20_0.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girl2_cartoon2_20_0.jpg -------------------------------------------------------------------------------- /output/girl2_cartoon2_20_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girl2_cartoon2_20_100.png -------------------------------------------------------------------------------- /output/girlmrf_girlsketch_5_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girlmrf_girlsketch_5_0.png -------------------------------------------------------------------------------- /output/girlmrf_girlsketch_5_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girlmrf_girlsketch_5_100.png -------------------------------------------------------------------------------- /output/girlmrf_smallworldI_20_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girlmrf_smallworldI_20_0.png -------------------------------------------------------------------------------- /output/girlmrf_smallworldI_20_200.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/girlmrf_smallworldI_20_200.png -------------------------------------------------------------------------------- /output/goat_muse20_1_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/goat_muse20_1_0.png -------------------------------------------------------------------------------- /output/goat_muse20_2_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/goat_muse20_2_100.png -------------------------------------------------------------------------------- /output/kid5_smallworldI_20_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/kid5_smallworldI_20_0.png -------------------------------------------------------------------------------- /output/kid5_smallworldI_20_200.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/kid5_smallworldI_20_200.png -------------------------------------------------------------------------------- /output/laplacians/girl2_car2_0_edge.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/laplacians/girl2_car2_0_edge.png -------------------------------------------------------------------------------- /output/laplacians/girl2_car2_edge.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/laplacians/girl2_car2_edge.png -------------------------------------------------------------------------------- /output/laplacians/girl2_edge.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/laplacians/girl2_edge.png -------------------------------------------------------------------------------- /output/megan_flowers20_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/megan_flowers20_0.png -------------------------------------------------------------------------------- /output/megan_flowers20_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/megan_flowers20_100.png -------------------------------------------------------------------------------- /output/megan_starry10_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/megan_starry10_0.png -------------------------------------------------------------------------------- /output/megan_starry10_100.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/askerlee/lapstyle/60489ea862f2d53316a1dd83b6af7fff8324df38/output/megan_starry10_100.png -------------------------------------------------------------------------------- /single-lap-exp.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash -x 2 | # test lap_style.lua with default settings on five test image pairs. 3 | # You can specify a different laplacian layer(s) using laplayersAll and lapweightsAll. 4 | 5 | contentimageAll=( megan.png kid5.png goat.png girlmrf.jpg boy.png) 6 | styleimageAll=( flowers.png smallworldI.jpg muse.jpg girlsketch.jpg girl3.jpg) 7 | inputLen=${#contentimageAll[@]} 8 | contentweight=20 9 | #laplayersAll=( 2 2,4) 10 | #lapweightsAll=(100 100,1600) 11 | laplayersAll=( 2) 12 | lapweightsAll=(100) 13 | configLen=${#laplayersAll[@]} 14 | for (( i=0; i<$inputLen; i++ )); do 15 | contentimage=${contentimageAll[$i]} 16 | contentname="${contentimage%.*}" 17 | styleimage=${styleimageAll[$i]} 18 | stylename="${styleimage%.*}" 19 | for (( j=0; j<$configLen; j++ )); do 20 | laplayers=${laplayersAll[$j]} 21 | lapweights=${lapweightsAll[$j]} 22 | outsig="${contentname}_${stylename}${contentweight}_${laplayers}_${lapweights}" 23 | th lap_style.lua -style_image images/$styleimage -content_image images/$contentimage -output_image output/${outsig}.png -content_weight $contentweight -lap_layers $laplayers -lap_weights $lapweights 2>&1| tee output/${outsig}.log 24 | outsig="${contentname}_${stylename}${contentweight}_${laplayers}_${lapweights}_nobp" 25 | th lap_style.lua -style_image images/$styleimage -content_image images/$contentimage -output_image output/${outsig}.png -content_weight $contentweight -lap_layers $laplayers -lap_weights $lapweights -lap_nobp 2>&1| tee output/${outsig}.log 26 | done 27 | done 28 | -------------------------------------------------------------------------------- /tf-neural-style/neural_style.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2015-2017 Anish Athalye. Released under GPLv3. 2 | # Revised by Shaohua Li in Apr 2017. 3 | # Please refer to https://github.com/anishathalye/neural-style for a user manual 4 | 5 | import os 6 | 7 | import numpy as np 8 | import scipy.misc 9 | 10 | from stylize import stylize 11 | 12 | import math 13 | from argparse import ArgumentParser 14 | 15 | from PIL import Image 16 | 17 | # default arguments 18 | CONTENT_WEIGHT = 5.0 19 | LAP_LAYERS = '2' 20 | LAP_WEIGHT = 5e1 21 | CONTENT_WEIGHT_BLEND = 1.0 22 | STYLE_WEIGHT = 5e2 23 | TV_WEIGHT = 1e2 24 | STYLE_LAYER_WEIGHT_EXP = 1.0 25 | LEARNING_RATE = 1e1 26 | BETA1 = 0.9 27 | BETA2 = 0.999 28 | EPSILON = 1e-08 29 | STYLE_SCALE = 1.0 30 | ITERATIONS = 1000 31 | VGG_PATH = 'imagenet-vgg-verydeep-19.mat' 32 | POOLING = 'max' 33 | CHECKPOINT_ITERATIONS = 100 34 | PRINT_ITERATIONS = 50 35 | 36 | def build_parser(): 37 | parser = ArgumentParser() 38 | parser.add_argument('--content', 39 | dest='content', help='content image', 40 | metavar='CONTENT', required=True) 41 | parser.add_argument('--styles', 42 | dest='styles', 43 | nargs='+', help='one or more style images', 44 | metavar='STYLE', required=True) 45 | parser.add_argument('--output', 46 | dest='output', help='output path', 47 | metavar='OUTPUT', required=True) 48 | parser.add_argument('--iterations', type=int, 49 | dest='iterations', help='iterations (default %(default)s)', 50 | metavar='ITERATIONS', default=ITERATIONS) 51 | parser.add_argument('--print-iterations', type=int, 52 | dest='print_iterations', help='statistics printing frequency', 53 | metavar='PRINT_ITERATIONS', default=PRINT_ITERATIONS) 54 | parser.add_argument('--checkpoint-output', 55 | dest='checkpoint_output', help='checkpoint output format, e.g. output%%s.jpg', 56 | metavar='OUTPUT') 57 | parser.add_argument('--checkpoint-iterations', type=int, 58 | dest='checkpoint_iterations', help='checkpoint frequency', 59 | metavar='CHECKPOINT_ITERATIONS', default=CHECKPOINT_ITERATIONS) 60 | parser.add_argument('--width', type=int, 61 | dest='width', help='output width', 62 | metavar='WIDTH') 63 | parser.add_argument('--style-scales', type=float, 64 | dest='style_scales', 65 | nargs='+', help='one or more style scales', 66 | metavar='STYLE_SCALE') 67 | parser.add_argument('--network', 68 | dest='network', help='path to network parameters (default %(default)s)', 69 | metavar='VGG_PATH', default=VGG_PATH) 70 | parser.add_argument('--content-weight-blend', type=float, 71 | dest='content_weight_blend', help='content weight blend, conv4_2 * blend + conv5_2 * (1-blend) (default %(default)s)', 72 | metavar='CONTENT_WEIGHT_BLEND', default=CONTENT_WEIGHT_BLEND) 73 | parser.add_argument('--lap-layers', type=str, 74 | dest='lap_layers', help='Laplacian layers (default %(default)s)', 75 | metavar='LAP_LAYERS', default=LAP_LAYERS) 76 | parser.add_argument('--lap-weight', type=float, 77 | dest='lap_weight', help='laplacian weight (default %(default)s)', 78 | metavar='LAP_WEIGHT', default=LAP_WEIGHT) 79 | parser.add_argument('--content-weight', type=float, 80 | dest='content_weight', help='content weight (default %(default)s)', 81 | metavar='CONTENT_WEIGHT', default=CONTENT_WEIGHT) 82 | parser.add_argument('--style-weight', type=float, 83 | dest='style_weight', help='style weight (default %(default)s)', 84 | metavar='STYLE_WEIGHT', default=STYLE_WEIGHT) 85 | parser.add_argument('--style-layer-weight-exp', type=float, 86 | dest='style_layer_weight_exp', help='style layer weight exponentional increase - weight(layer) = weight_exp*weight(layer) (default %(default)s)', 87 | metavar='STYLE_LAYER_WEIGHT_EXP', default=STYLE_LAYER_WEIGHT_EXP) 88 | parser.add_argument('--style-blend-weights', type=float, 89 | dest='style_blend_weights', help='style blending weights', 90 | nargs='+', metavar='STYLE_BLEND_WEIGHT') 91 | parser.add_argument('--tv-weight', type=float, 92 | dest='tv_weight', help='total variation regularization weight (default %(default)s)', 93 | metavar='TV_WEIGHT', default=TV_WEIGHT) 94 | parser.add_argument('--learning-rate', type=float, 95 | dest='learning_rate', help='learning rate (default %(default)s)', 96 | metavar='LEARNING_RATE', default=LEARNING_RATE) 97 | parser.add_argument('--beta1', type=float, 98 | dest='beta1', help='Adam: beta1 parameter (default %(default)s)', 99 | metavar='BETA1', default=BETA1) 100 | parser.add_argument('--beta2', type=float, 101 | dest='beta2', help='Adam: beta2 parameter (default %(default)s)', 102 | metavar='BETA2', default=BETA2) 103 | parser.add_argument('--eps', type=float, 104 | dest='epsilon', help='Adam: epsilon parameter (default %(default)s)', 105 | metavar='EPSILON', default=EPSILON) 106 | parser.add_argument('--initial', 107 | dest='initial', help='initial image', 108 | metavar='INITIAL') 109 | parser.add_argument('--initial-noiseblend', type=float, 110 | dest='initial_noiseblend', help='ratio of blending initial image with normalized noise (if no initial image specified, content image is used) (default %(default)s)', 111 | metavar='INITIAL_NOISEBLEND') 112 | parser.add_argument('--preserve-colors', action='store_true', 113 | dest='preserve_colors', help='style-only transfer (preserving colors) - if color transfer is not needed') 114 | parser.add_argument('--pooling', 115 | dest='pooling', help='pooling layer configuration: max or avg (default %(default)s)', 116 | metavar='POOLING', default=POOLING) 117 | return parser 118 | 119 | 120 | def main(): 121 | parser = build_parser() 122 | options = parser.parse_args() 123 | 124 | if not os.path.isfile(options.network): 125 | parser.error("Network %s does not exist. (Did you forget to download it?)" % options.network) 126 | 127 | content_image = imread(options.content) 128 | style_images = [imread(style) for style in options.styles] 129 | 130 | width = options.width 131 | if width is not None: 132 | new_shape = (int(math.floor(float(content_image.shape[0]) / 133 | content_image.shape[1] * width)), width) 134 | content_image = scipy.misc.imresize(content_image, new_shape) 135 | target_shape = content_image.shape 136 | for i in range(len(style_images)): 137 | style_scale = STYLE_SCALE 138 | if options.style_scales is not None: 139 | style_scale = options.style_scales[i] 140 | style_images[i] = scipy.misc.imresize(style_images[i], style_scale * 141 | target_shape[1] / style_images[i].shape[1]) 142 | 143 | style_blend_weights = options.style_blend_weights 144 | if style_blend_weights is None: 145 | # default is equal weights 146 | style_blend_weights = [1.0/len(style_images) for _ in style_images] 147 | else: 148 | total_blend_weight = sum(style_blend_weights) 149 | style_blend_weights = [weight/total_blend_weight 150 | for weight in style_blend_weights] 151 | 152 | options.lap_layers = options.lap_layers.split(',') 153 | 154 | initial = options.initial 155 | if initial is not None: 156 | initial = scipy.misc.imresize(imread(initial), content_image.shape[:2]) 157 | # Initial guess is specified, but not noiseblend - no noise should be blended 158 | if options.initial_noiseblend is None: 159 | options.initial_noiseblend = 0.0 160 | else: 161 | # Neither inital, nor noiseblend is provided, falling back to random generated initial guess 162 | if options.initial_noiseblend is None: 163 | options.initial_noiseblend = 1.0 164 | if options.initial_noiseblend < 1.0: 165 | initial = content_image 166 | 167 | if options.checkpoint_output and "%s" not in options.checkpoint_output: 168 | parser.error("To save intermediate images, the checkpoint output " 169 | "parameter must contain `%s` (e.g. `foo%s.jpg`)") 170 | 171 | for iteration, image in stylize( 172 | network=options.network, 173 | initial=initial, 174 | initial_noiseblend=options.initial_noiseblend, 175 | content=content_image, 176 | styles=style_images, 177 | preserve_colors=options.preserve_colors, 178 | iterations=options.iterations, 179 | content_weight=options.content_weight, 180 | content_weight_blend=options.content_weight_blend, 181 | lap_layers=options.lap_layers, 182 | lap_weight=options.lap_weight, 183 | style_weight=options.style_weight, 184 | style_layer_weight_exp=options.style_layer_weight_exp, 185 | style_blend_weights=style_blend_weights, 186 | tv_weight=options.tv_weight, 187 | learning_rate=options.learning_rate, 188 | beta1=options.beta1, 189 | beta2=options.beta2, 190 | epsilon=options.epsilon, 191 | pooling=options.pooling, 192 | print_iterations=options.print_iterations, 193 | checkpoint_iterations=options.checkpoint_iterations 194 | ): 195 | output_file = None 196 | combined_rgb = image 197 | if iteration is not None: 198 | if options.checkpoint_output: 199 | output_file = options.checkpoint_output % iteration 200 | else: 201 | output_file = options.output 202 | if output_file: 203 | imsave(output_file, combined_rgb) 204 | 205 | 206 | def imread(path): 207 | img = scipy.misc.imread(path).astype(np.float) 208 | if len(img.shape) == 2: 209 | # grayscale 210 | img = np.dstack((img,img,img)) 211 | elif img.shape[2] == 4: 212 | # PNG with alpha channel 213 | img = img[:,:,:3] 214 | return img 215 | 216 | 217 | def imsave(path, img): 218 | img = np.clip(img, 0, 255).astype(np.uint8) 219 | Image.fromarray(img).save(path, quality=95) 220 | 221 | if __name__ == '__main__': 222 | main() 223 | -------------------------------------------------------------------------------- /tf-neural-style/stylize.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2015-2017 Anish Athalye. Released under GPLv3. 2 | 3 | import vgg 4 | 5 | import tensorflow as tf 6 | import numpy as np 7 | 8 | from sys import stderr 9 | 10 | from PIL import Image 11 | 12 | CONTENT_LAYERS = ['relu4_2', 'relu5_2'] 13 | STYLE_LAYERS = ('relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1') 14 | 15 | try: 16 | reduce 17 | except NameError: 18 | from functools import reduce 19 | 20 | 21 | def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations, 22 | content_weight, content_weight_blend, lap_weight, lap_layers, 23 | style_weight, style_layer_weight_exp, style_blend_weights, tv_weight, 24 | learning_rate, beta1, beta2, epsilon, pooling, 25 | print_iterations=None, checkpoint_iterations=None): 26 | """ 27 | Stylize images. 28 | 29 | This function yields tuples (iteration, image); `iteration` is None 30 | if this is the final image (the last iteration). Other tuples are yielded 31 | every `checkpoint_iterations` iterations. 32 | 33 | :rtype: iterator[tuple[int|None,image]] 34 | """ 35 | shape = (1,) + content.shape 36 | style_shapes = [(1,) + style.shape for style in styles] 37 | content_features = {} 38 | style_features = [{} for _ in styles] 39 | 40 | vgg_weights, vgg_mean_pixel = vgg.load_net(network) 41 | 42 | layer_weight = 1.0 43 | style_layers_weights = {} 44 | for style_layer in STYLE_LAYERS: 45 | style_layers_weights[style_layer] = layer_weight 46 | layer_weight *= style_layer_weight_exp 47 | 48 | # normalize style layer weights 49 | layer_weights_sum = 0 50 | for style_layer in STYLE_LAYERS: 51 | layer_weights_sum += style_layers_weights[style_layer] 52 | for style_layer in STYLE_LAYERS: 53 | style_layers_weights[style_layer] /= layer_weights_sum 54 | 55 | global CONTENT_LAYERS 56 | for lap_layer in lap_layers: 57 | CONTENT_LAYERS.append( 'lap'+lap_layer ) 58 | 59 | # compute content features in feedforward mode 60 | g = tf.Graph() 61 | with g.as_default(), g.device('/cpu:0'), tf.Session() as sess: 62 | image = tf.placeholder('float', shape=shape) 63 | net = vgg.net_preloaded(vgg_weights, image, pooling) 64 | content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)]) 65 | for layer in CONTENT_LAYERS: 66 | content_features[layer] = net[layer].eval(feed_dict={image: content_pre}) 67 | 68 | if len(lap_layers) > 0: 69 | stderr.write('Lap layers: %s\n' %(', '.join(lap_layers))) 70 | 71 | # compute style features in feedforward mode 72 | for i in range(len(styles)): 73 | g = tf.Graph() 74 | with g.as_default(), g.device('/cpu:0'), tf.Session() as sess: 75 | image = tf.placeholder('float', shape=style_shapes[i]) 76 | net = vgg.net_preloaded(vgg_weights, image, pooling) 77 | style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)]) 78 | for layer in STYLE_LAYERS: 79 | features = net[layer].eval(feed_dict={image: style_pre}) 80 | features = np.reshape(features, (-1, features.shape[3])) 81 | gram = np.matmul(features.T, features) / features.size 82 | style_features[i][layer] = gram 83 | 84 | initial_content_noise_coeff = 1.0 - initial_noiseblend 85 | #content_weight_step = (content_weight_max - content_weight_min) * 1.0 / iterations 86 | #stderr.write('content weight step: %.2f\n' % content_weight_step) 87 | 88 | # make stylized image using backpropogation 89 | with tf.Graph().as_default(): 90 | if initial is None: 91 | noise = np.random.normal(size=shape, scale=np.std(content) * 0.1) 92 | initial = tf.random_normal(shape) * 0.256 93 | else: 94 | initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)]) 95 | initial = initial.astype('float32') 96 | noise = np.random.normal(size=shape, scale=np.std(content) * 0.1) 97 | initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff) 98 | image = tf.Variable(initial) 99 | net = vgg.net_preloaded(vgg_weights, image, pooling) 100 | #content_weight = tf.constant(content_weight_max, dtype=tf.float32) 101 | 102 | # content loss 103 | content_layers_weights = {} 104 | content_layers_weights['relu4_2'] = content_weight_blend 105 | content_layers_weights['relu5_2'] = 1.0 - content_weight_blend 106 | 107 | content_loss = 0 108 | lap_loss = 0 109 | content_losses = [] 110 | lap_losses = [] 111 | 112 | for content_layer in CONTENT_LAYERS: 113 | if 'lap' in content_layer: 114 | loss = lap_weight * (2 * tf.nn.l2_loss( net[content_layer] - 115 | content_features[content_layer] ) / 116 | content_features[content_layer].size ) 117 | lap_losses.append(loss) 118 | else: 119 | weight = content_weight 120 | content_layer_weight = content_layers_weights[content_layer] 121 | loss = content_layer_weight * weight * (2 * tf.nn.l2_loss( 122 | net[content_layer] - content_features[content_layer]) / 123 | content_features[content_layer].size) 124 | content_losses.append(loss) 125 | 126 | content_loss += reduce(tf.add, content_losses) 127 | lap_loss += reduce(tf.add, lap_losses) 128 | 129 | # style loss 130 | style_loss = 0 131 | for i in range(len(styles)): 132 | style_losses = [] 133 | for style_layer in STYLE_LAYERS: 134 | layer = net[style_layer] 135 | _, height, width, number = map(lambda i: i.value, layer.get_shape()) 136 | size = height * width * number 137 | feats = tf.reshape(layer, (-1, number)) 138 | gram = tf.matmul(tf.transpose(feats), feats) / size 139 | style_gram = style_features[i][style_layer] 140 | style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size) 141 | style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses) 142 | 143 | # total variation denoising 144 | tv_y_size = _tensor_size(image[:,1:,:,:]) 145 | tv_x_size = _tensor_size(image[:,:,1:,:]) 146 | tv_loss = tv_weight * 2 * ( 147 | (tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) / 148 | tv_y_size) + 149 | (tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) / 150 | tv_x_size)) 151 | # overall loss 152 | loss = content_loss + style_loss + tv_loss + lap_loss 153 | 154 | # optimizer setup 155 | train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss) 156 | 157 | def print_progress(): 158 | stderr.write(' content loss: %g\n' % content_loss.eval()) 159 | #stderr.write(' content weight: %g\n' % content_weight.eval()) 160 | stderr.write(' laplacian loss: %g\n' % lap_loss.eval()) 161 | stderr.write(' style loss: %g\n' % style_loss.eval()) 162 | stderr.write(' tv loss: %g\n' % tv_loss.eval()) 163 | stderr.write(' total loss: %g\n' % loss.eval()) 164 | 165 | # optimization 166 | best_loss = float('inf') 167 | best = None 168 | with tf.Session() as sess: 169 | sess.run(tf.global_variables_initializer()) 170 | stderr.write('Optimization started...\n') 171 | if (print_iterations and print_iterations != 0): 172 | print_progress() 173 | for i in range(iterations): 174 | stderr.write('Iteration %4d/%4d\n' % (i + 1, iterations)) 175 | train_step.run() 176 | #content_weight = tf.constant( tf.add( content_weight, -content_weight_step ).eval() ) 177 | 178 | last_step = (i == iterations - 1) 179 | if last_step or (print_iterations and i % print_iterations == 0): 180 | print_progress() 181 | 182 | if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step: 183 | this_loss = loss.eval() 184 | if this_loss < best_loss: 185 | best_loss = this_loss 186 | best = image.eval() 187 | 188 | img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel) 189 | 190 | if preserve_colors and preserve_colors == True: 191 | original_image = np.clip(content, 0, 255) 192 | styled_image = np.clip(img_out, 0, 255) 193 | 194 | # Luminosity transfer steps: 195 | # 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114) 196 | # 2. Convert stylized grayscale into YUV (YCbCr) 197 | # 3. Convert original image into YUV (YCbCr) 198 | # 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V) 199 | # 5. Convert recombined image from YUV back to RGB 200 | 201 | # 1 202 | styled_grayscale = rgb2gray(styled_image) 203 | styled_grayscale_rgb = gray2rgb(styled_grayscale) 204 | 205 | # 2 206 | styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr')) 207 | 208 | # 3 209 | original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr')) 210 | 211 | # 4 212 | w, h, _ = original_image.shape 213 | combined_yuv = np.empty((w, h, 3), dtype=np.uint8) 214 | combined_yuv[..., 0] = styled_grayscale_yuv[..., 0] 215 | combined_yuv[..., 1] = original_yuv[..., 1] 216 | combined_yuv[..., 2] = original_yuv[..., 2] 217 | 218 | # 5 219 | img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB')) 220 | 221 | 222 | yield ( 223 | (None if last_step else i), 224 | img_out 225 | ) 226 | 227 | 228 | def _tensor_size(tensor): 229 | from operator import mul 230 | return reduce(mul, (d.value for d in tensor.get_shape()), 1) 231 | 232 | def rgb2gray(rgb): 233 | return np.dot(rgb[...,:3], [0.299, 0.587, 0.114]) 234 | 235 | def gray2rgb(gray): 236 | w, h = gray.shape 237 | rgb = np.empty((w, h, 3), dtype=np.float32) 238 | rgb[:, :, 2] = rgb[:, :, 1] = rgb[:, :, 0] = gray 239 | return rgb 240 | 241 | def l1_loss(tensor): 242 | return tf.reduce_sum(tf.abs(tensor)) 243 | -------------------------------------------------------------------------------- /tf-neural-style/vgg.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) 2015-2017 Anish Athalye. Released under GPLv3. 2 | 3 | import tensorflow as tf 4 | import numpy as np 5 | import scipy.io 6 | 7 | VGG19_LAYERS = ( 8 | 'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 9 | 10 | 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2', 11 | 12 | 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 13 | 'relu3_3', 'conv3_4', 'relu3_4', 'pool3', 14 | 15 | 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 16 | 'relu4_3', 'conv4_4', 'relu4_4', 'pool4', 17 | 18 | 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 19 | 'relu5_3', 'conv5_4', 'relu5_4' 20 | ) 21 | LAPLACIAN_LAYERS = ( 'pool_lap1', 'lap1', 'pool_lap2', 'lap2', 'pool_lap3', 'lap3' ) 22 | 23 | def load_net(data_path): 24 | data = scipy.io.loadmat(data_path) 25 | mean = data['normalization'][0][0][0] 26 | mean_pixel = np.mean(mean, axis=(0, 1)) 27 | weights = data['layers'][0] 28 | return weights, mean_pixel 29 | 30 | def net_preloaded(weights, input_image, pooling): 31 | net = {} 32 | current = input_image 33 | for i, name in enumerate(VGG19_LAYERS): 34 | kind = name[:4] 35 | if kind == 'conv': 36 | kernels, bias = weights[i][0][0][0][0] 37 | # matconvnet: weights are [width, height, in_channels, out_channels] 38 | # tensorflow: weights are [height, width, in_channels, out_channels] 39 | kernels = np.transpose(kernels, (1, 0, 2, 3)) 40 | bias = bias.reshape(-1) 41 | current = _conv_layer(current, kernels, bias) 42 | elif kind == 'relu': 43 | current = tf.nn.relu(current) 44 | elif kind == 'pool': 45 | current = _pool_layer(current, pooling) 46 | net[name] = current 47 | 48 | laplacian = np.array( [ [0,-1,0], [-1,4,-1], [0,-1,0] ], dtype=np.float32 ) 49 | lapW = np.zeros( (3, 3, 3, 1), dtype=np.float32 ) 50 | for t in range(3): 51 | lapW[:,:,t,0] = laplacian 52 | 53 | for i in range(1,4): 54 | net['pool_lap%d'%i] = _pool_layer( input_image, 'avg', 2**i ) 55 | #net['lap%d'%i] = _conv_layer(net['pool_lap%d'%i], lapW, [0.0]) 56 | net['lap%d'%i] = _lap_layer(net['pool_lap%d'%i]) 57 | assert len(net) == len(VGG19_LAYERS) + len(LAPLACIAN_LAYERS) 58 | return net 59 | 60 | def _conv_layer(input, weights, bias): 61 | conv = tf.nn.conv2d(input, tf.constant(weights), strides=(1, 1, 1, 1), 62 | padding='SAME') 63 | return tf.nn.bias_add(conv, bias) 64 | 65 | 66 | def _pool_layer(input, pooling, poolsize=2): 67 | ksize = (1, poolsize, poolsize, 1) 68 | if pooling == 'avg': 69 | return tf.nn.avg_pool(input, ksize, strides=(1, poolsize, poolsize, 1), 70 | padding='SAME') 71 | else: 72 | return tf.nn.max_pool(input, ksize, strides=(1, poolsize, poolsize, 1), 73 | padding='SAME') 74 | 75 | def _lap_layer(input): 76 | laplacian = np.array( [ [0,-1,0], [-1,4,-1], [0,-1,0] ], dtype=np.float32 ) 77 | lapW = np.zeros( (3, 3, 1, 1), dtype=np.float32 ) 78 | lapW[:,:,0,0] = laplacian 79 | color_outs = [] 80 | for i in range(3): 81 | color = input[:,:,:,i] 82 | color4d = tf.expand_dims(color, -1) 83 | color_out = _conv_layer(color4d, lapW, [0.0]) 84 | color_outs.append(color_out) 85 | output = tf.concat(color_outs, axis=-1) 86 | output = tf.abs(output) 87 | sum_output = tf.reduce_sum(output, reduction_indices=[3], keep_dims=False) 88 | #cut_cond = tf.less( max_output, tf.ones(tf.shape(max_output)) * thres ) 89 | #cut_output = tf.where( cut_cond, tf.zeros(tf.shape(max_output)), max_output ) 90 | return sum_output 91 | 92 | 93 | def preprocess(image, mean_pixel): 94 | return image - mean_pixel 95 | 96 | 97 | def unprocess(image, mean_pixel): 98 | return image + mean_pixel 99 | --------------------------------------------------------------------------------