├── .gitignore ├── .travis.yml ├── LICENSE ├── README.md ├── build_dataset_directory.py ├── colorize.py ├── model.py ├── resize_all_imgs.py └── train.py /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__/* 2 | *.pyc 3 | -------------------------------------------------------------------------------- /.travis.yml: -------------------------------------------------------------------------------- 1 | language: python 2 | python: 3 | - '3.6' 4 | before_script: 5 | - pip install opencv-python 6 | - pip install numpy 7 | - pip install pillow 8 | - pip install scikit-image 9 | - pip install http://download.pytorch.org/whl/cpu/torch-0.4.0-cp36-cp36m-linux_x86_64.whl 10 | - pip install torchvision 11 | - pip install scipy 12 | script: 13 | - python -m py_compile model.py 14 | - python -m py_compile resize_all_imgs.py 15 | - python -m py_compile colorize.py 16 | - python -m py_compile build_dataset_directory.py 17 | - python -m py_compile train.py 18 | - wget -O model.pth "https://github.com/zeruniverse/neural-colorization/releases/download/1.1/G.pth" 19 | - wget -O a.jpg "https://avatars1.githubusercontent.com/u/4648756" 20 | - mkdir -p train_raw_folder/random/arbitrary 21 | - cp a.jpg train_raw_folder/random/arbitrary/0.jpg 22 | - cp a.jpg train_raw_folder/random/arbitrary/1.jpeg 23 | - python build_dataset_directory.py -i train_raw_folder -o train 24 | - python resize_all_imgs.py -d train 25 | - ls train 26 | - python colorize.py -i a.jpg -o aa.jpg -m model.pth 27 | after_script: 28 | - ls -lh aa.jpg 29 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GPL 3.0 for personal or research use. COMMERCIAL USE PROHIBITED. 2 | 3 | Below is the GNU GPL 3.0 LICENSE. 4 | 5 | (Please note this license only applies for personal or research use. COMMERCIAL USE IS PROHIBITED) 6 | ----------------- 7 | 8 | GNU GENERAL PUBLIC LICENSE 9 | Version 3, 29 June 2007 10 | 11 | Copyright (C) 2007 Free Software Foundation, Inc. 12 | Everyone is permitted to copy and distribute verbatim copies 13 | of this license document, but changing it is not allowed. 14 | 15 | Preamble 16 | 17 | The GNU General Public License is a free, copyleft license for 18 | software and other kinds of works. 19 | 20 | The licenses for most software and other practical works are designed 21 | to take away your freedom to share and change the works. By contrast, 22 | the GNU General Public License is intended to guarantee your freedom to 23 | share and change all versions of a program--to make sure it remains free 24 | software for all its users. We, the Free Software Foundation, use the 25 | GNU General Public License for most of our software; it applies also to 26 | any other work released this way by its authors. You can apply it to 27 | your programs, too. 28 | 29 | When we speak of free software, we are referring to freedom, not 30 | price. Our General Public Licenses are designed to make sure that you 31 | have the freedom to distribute copies of free software (and charge for 32 | them if you wish), that you receive source code or can get it if you 33 | want it, that you can change the software or use pieces of it in new 34 | free programs, and that you know you can do these things. 35 | 36 | To protect your rights, we need to prevent others from denying you 37 | these rights or asking you to surrender the rights. Therefore, you have 38 | certain responsibilities if you distribute copies of the software, or if 39 | you modify it: responsibilities to respect the freedom of others. 40 | 41 | For example, if you distribute copies of such a program, whether 42 | gratis or for a fee, you must pass on to the recipients the same 43 | freedoms that you received. You must make sure that they, too, receive 44 | or can get the source code. And you must show them these terms so they 45 | know their rights. 46 | 47 | Developers that use the GNU GPL protect your rights with two steps: 48 | (1) assert copyright on the software, and (2) offer you this License 49 | giving you legal permission to copy, distribute and/or modify it. 50 | 51 | For the developers' and authors' protection, the GPL clearly explains 52 | that there is no warranty for this free software. For both users' and 53 | authors' sake, the GPL requires that modified versions be marked as 54 | changed, so that their problems will not be attributed erroneously to 55 | authors of previous versions. 56 | 57 | Some devices are designed to deny users access to install or run 58 | modified versions of the software inside them, although the manufacturer 59 | can do so. This is fundamentally incompatible with the aim of 60 | protecting users' freedom to change the software. The systematic 61 | pattern of such abuse occurs in the area of products for individuals to 62 | use, which is precisely where it is most unacceptable. Therefore, we 63 | have designed this version of the GPL to prohibit the practice for those 64 | products. If such problems arise substantially in other domains, we 65 | stand ready to extend this provision to those domains in future versions 66 | of the GPL, as needed to protect the freedom of users. 67 | 68 | Finally, every program is threatened constantly by software patents. 69 | States should not allow patents to restrict development and use of 70 | software on general-purpose computers, but in those that do, we wish to 71 | avoid the special danger that patents applied to a free program could 72 | make it effectively proprietary. To prevent this, the GPL assures that 73 | patents cannot be used to render the program non-free. 74 | 75 | The precise terms and conditions for copying, distribution and 76 | modification follow. 77 | 78 | TERMS AND CONDITIONS 79 | 80 | 0. Definitions. 81 | 82 | "This License" refers to version 3 of the GNU General Public License. 83 | 84 | "Copyright" also means copyright-like laws that apply to other kinds of 85 | works, such as semiconductor masks. 86 | 87 | "The Program" refers to any copyrightable work licensed under this 88 | License. Each licensee is addressed as "you". "Licensees" and 89 | "recipients" may be individuals or organizations. 90 | 91 | To "modify" a work means to copy from or adapt all or part of the work 92 | in a fashion requiring copyright permission, other than the making of an 93 | exact copy. The resulting work is called a "modified version" of the 94 | earlier work or a work "based on" the earlier work. 95 | 96 | A "covered work" means either the unmodified Program or a work based 97 | on the Program. 98 | 99 | To "propagate" a work means to do anything with it that, without 100 | permission, would make you directly or secondarily liable for 101 | infringement under applicable copyright law, except executing it on a 102 | computer or modifying a private copy. Propagation includes copying, 103 | distribution (with or without modification), making available to the 104 | public, and in some countries other activities as well. 105 | 106 | To "convey" a work means any kind of propagation that enables other 107 | parties to make or receive copies. Mere interaction with a user through 108 | a computer network, with no transfer of a copy, is not conveying. 109 | 110 | An interactive user interface displays "Appropriate Legal Notices" 111 | to the extent that it includes a convenient and prominently visible 112 | feature that (1) displays an appropriate copyright notice, and (2) 113 | tells the user that there is no warranty for the work (except to the 114 | extent that warranties are provided), that licensees may convey the 115 | work under this License, and how to view a copy of this License. If 116 | the interface presents a list of user commands or options, such as a 117 | menu, a prominent item in the list meets this criterion. 118 | 119 | 1. Source Code. 120 | 121 | The "source code" for a work means the preferred form of the work 122 | for making modifications to it. "Object code" means any non-source 123 | form of a work. 124 | 125 | A "Standard Interface" means an interface that either is an official 126 | standard defined by a recognized standards body, or, in the case of 127 | interfaces specified for a particular programming language, one that 128 | is widely used among developers working in that language. 129 | 130 | The "System Libraries" of an executable work include anything, other 131 | than the work as a whole, that (a) is included in the normal form of 132 | packaging a Major Component, but which is not part of that Major 133 | Component, and (b) serves only to enable use of the work with that 134 | Major Component, or to implement a Standard Interface for which an 135 | implementation is available to the public in source code form. A 136 | "Major Component", in this context, means a major essential component 137 | (kernel, window system, and so on) of the specific operating system 138 | (if any) on which the executable work runs, or a compiler used to 139 | produce the work, or an object code interpreter used to run it. 140 | 141 | The "Corresponding Source" for a work in object code form means all 142 | the source code needed to generate, install, and (for an executable 143 | work) run the object code and to modify the work, including scripts to 144 | control those activities. However, it does not include the work's 145 | System Libraries, or general-purpose tools or generally available free 146 | programs which are used unmodified in performing those activities but 147 | which are not part of the work. For example, Corresponding Source 148 | includes interface definition files associated with source files for 149 | the work, and the source code for shared libraries and dynamically 150 | linked subprograms that the work is specifically designed to require, 151 | such as by intimate data communication or control flow between those 152 | subprograms and other parts of the work. 153 | 154 | The Corresponding Source need not include anything that users 155 | can regenerate automatically from other parts of the Corresponding 156 | Source. 157 | 158 | The Corresponding Source for a work in source code form is that 159 | same work. 160 | 161 | 2. Basic Permissions. 162 | 163 | All rights granted under this License are granted for the term of 164 | copyright on the Program, and are irrevocable provided the stated 165 | conditions are met. This License explicitly affirms your unlimited 166 | permission to run the unmodified Program. The output from running a 167 | covered work is covered by this License only if the output, given its 168 | content, constitutes a covered work. This License acknowledges your 169 | rights of fair use or other equivalent, as provided by copyright law. 170 | 171 | You may make, run and propagate covered works that you do not 172 | convey, without conditions so long as your license otherwise remains 173 | in force. You may convey covered works to others for the sole purpose 174 | of having them make modifications exclusively for you, or provide you 175 | with facilities for running those works, provided that you comply with 176 | the terms of this License in conveying all material for which you do 177 | not control copyright. Those thus making or running the covered works 178 | for you must do so exclusively on your behalf, under your direction 179 | and control, on terms that prohibit them from making any copies of 180 | your copyrighted material outside their relationship with you. 181 | 182 | Conveying under any other circumstances is permitted solely under 183 | the conditions stated below. Sublicensing is not allowed; section 10 184 | makes it unnecessary. 185 | 186 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 187 | 188 | No covered work shall be deemed part of an effective technological 189 | measure under any applicable law fulfilling obligations under article 190 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 191 | similar laws prohibiting or restricting circumvention of such 192 | measures. 193 | 194 | When you convey a covered work, you waive any legal power to forbid 195 | circumvention of technological measures to the extent such circumvention 196 | is effected by exercising rights under this License with respect to 197 | the covered work, and you disclaim any intention to limit operation or 198 | modification of the work as a means of enforcing, against the work's 199 | users, your or third parties' legal rights to forbid circumvention of 200 | technological measures. 201 | 202 | 4. Conveying Verbatim Copies. 203 | 204 | You may convey verbatim copies of the Program's source code as you 205 | receive it, in any medium, provided that you conspicuously and 206 | appropriately publish on each copy an appropriate copyright notice; 207 | keep intact all notices stating that this License and any 208 | non-permissive terms added in accord with section 7 apply to the code; 209 | keep intact all notices of the absence of any warranty; and give all 210 | recipients a copy of this License along with the Program. 211 | 212 | You may charge any price or no price for each copy that you convey, 213 | and you may offer support or warranty protection for a fee. 214 | 215 | 5. Conveying Modified Source Versions. 216 | 217 | You may convey a work based on the Program, or the modifications to 218 | produce it from the Program, in the form of source code under the 219 | terms of section 4, provided that you also meet all of these conditions: 220 | 221 | a) The work must carry prominent notices stating that you modified 222 | it, and giving a relevant date. 223 | 224 | b) The work must carry prominent notices stating that it is 225 | released under this License and any conditions added under section 226 | 7. This requirement modifies the requirement in section 4 to 227 | "keep intact all notices". 228 | 229 | c) You must license the entire work, as a whole, under this 230 | License to anyone who comes into possession of a copy. This 231 | License will therefore apply, along with any applicable section 7 232 | additional terms, to the whole of the work, and all its parts, 233 | regardless of how they are packaged. This License gives no 234 | permission to license the work in any other way, but it does not 235 | invalidate such permission if you have separately received it. 236 | 237 | d) If the work has interactive user interfaces, each must display 238 | Appropriate Legal Notices; however, if the Program has interactive 239 | interfaces that do not display Appropriate Legal Notices, your 240 | work need not make them do so. 241 | 242 | A compilation of a covered work with other separate and independent 243 | works, which are not by their nature extensions of the covered work, 244 | and which are not combined with it such as to form a larger program, 245 | in or on a volume of a storage or distribution medium, is called an 246 | "aggregate" if the compilation and its resulting copyright are not 247 | used to limit the access or legal rights of the compilation's users 248 | beyond what the individual works permit. Inclusion of a covered work 249 | in an aggregate does not cause this License to apply to the other 250 | parts of the aggregate. 251 | 252 | 6. Conveying Non-Source Forms. 253 | 254 | You may convey a covered work in object code form under the terms 255 | of sections 4 and 5, provided that you also convey the 256 | machine-readable Corresponding Source under the terms of this License, 257 | in one of these ways: 258 | 259 | a) Convey the object code in, or embodied in, a physical product 260 | (including a physical distribution medium), accompanied by the 261 | Corresponding Source fixed on a durable physical medium 262 | customarily used for software interchange. 263 | 264 | b) Convey the object code in, or embodied in, a physical product 265 | (including a physical distribution medium), accompanied by a 266 | written offer, valid for at least three years and valid for as 267 | long as you offer spare parts or customer support for that product 268 | model, to give anyone who possesses the object code either (1) a 269 | copy of the Corresponding Source for all the software in the 270 | product that is covered by this License, on a durable physical 271 | medium customarily used for software interchange, for a price no 272 | more than your reasonable cost of physically performing this 273 | conveying of source, or (2) access to copy the 274 | Corresponding Source from a network server at no charge. 275 | 276 | c) Convey individual copies of the object code with a copy of the 277 | written offer to provide the Corresponding Source. This 278 | alternative is allowed only occasionally and noncommercially, and 279 | only if you received the object code with such an offer, in accord 280 | with subsection 6b. 281 | 282 | d) Convey the object code by offering access from a designated 283 | place (gratis or for a charge), and offer equivalent access to the 284 | Corresponding Source in the same way through the same place at no 285 | further charge. You need not require recipients to copy the 286 | Corresponding Source along with the object code. If the place to 287 | copy the object code is a network server, the Corresponding Source 288 | may be on a different server (operated by you or a third party) 289 | that supports equivalent copying facilities, provided you maintain 290 | clear directions next to the object code saying where to find the 291 | Corresponding Source. Regardless of what server hosts the 292 | Corresponding Source, you remain obligated to ensure that it is 293 | available for as long as needed to satisfy these requirements. 294 | 295 | e) Convey the object code using peer-to-peer transmission, provided 296 | you inform other peers where the object code and Corresponding 297 | Source of the work are being offered to the general public at no 298 | charge under subsection 6d. 299 | 300 | A separable portion of the object code, whose source code is excluded 301 | from the Corresponding Source as a System Library, need not be 302 | included in conveying the object code work. 303 | 304 | A "User Product" is either (1) a "consumer product", which means any 305 | tangible personal property which is normally used for personal, family, 306 | or household purposes, or (2) anything designed or sold for incorporation 307 | into a dwelling. In determining whether a product is a consumer product, 308 | doubtful cases shall be resolved in favor of coverage. For a particular 309 | product received by a particular user, "normally used" refers to a 310 | typical or common use of that class of product, regardless of the status 311 | of the particular user or of the way in which the particular user 312 | actually uses, or expects or is expected to use, the product. A product 313 | is a consumer product regardless of whether the product has substantial 314 | commercial, industrial or non-consumer uses, unless such uses represent 315 | the only significant mode of use of the product. 316 | 317 | "Installation Information" for a User Product means any methods, 318 | procedures, authorization keys, or other information required to install 319 | and execute modified versions of a covered work in that User Product from 320 | a modified version of its Corresponding Source. The information must 321 | suffice to ensure that the continued functioning of the modified object 322 | code is in no case prevented or interfered with solely because 323 | modification has been made. 324 | 325 | If you convey an object code work under this section in, or with, or 326 | specifically for use in, a User Product, and the conveying occurs as 327 | part of a transaction in which the right of possession and use of the 328 | User Product is transferred to the recipient in perpetuity or for a 329 | fixed term (regardless of how the transaction is characterized), the 330 | Corresponding Source conveyed under this section must be accompanied 331 | by the Installation Information. But this requirement does not apply 332 | if neither you nor any third party retains the ability to install 333 | modified object code on the User Product (for example, the work has 334 | been installed in ROM). 335 | 336 | The requirement to provide Installation Information does not include a 337 | requirement to continue to provide support service, warranty, or updates 338 | for a work that has been modified or installed by the recipient, or for 339 | the User Product in which it has been modified or installed. Access to a 340 | network may be denied when the modification itself materially and 341 | adversely affects the operation of the network or violates the rules and 342 | protocols for communication across the network. 343 | 344 | Corresponding Source conveyed, and Installation Information provided, 345 | in accord with this section must be in a format that is publicly 346 | documented (and with an implementation available to the public in 347 | source code form), and must require no special password or key for 348 | unpacking, reading or copying. 349 | 350 | 7. Additional Terms. 351 | 352 | "Additional permissions" are terms that supplement the terms of this 353 | License by making exceptions from one or more of its conditions. 354 | Additional permissions that are applicable to the entire Program shall 355 | be treated as though they were included in this License, to the extent 356 | that they are valid under applicable law. If additional permissions 357 | apply only to part of the Program, that part may be used separately 358 | under those permissions, but the entire Program remains governed by 359 | this License without regard to the additional permissions. 360 | 361 | When you convey a copy of a covered work, you may at your option 362 | remove any additional permissions from that copy, or from any part of 363 | it. (Additional permissions may be written to require their own 364 | removal in certain cases when you modify the work.) You may place 365 | additional permissions on material, added by you to a covered work, 366 | for which you have or can give appropriate copyright permission. 367 | 368 | Notwithstanding any other provision of this License, for material you 369 | add to a covered work, you may (if authorized by the copyright holders of 370 | that material) supplement the terms of this License with terms: 371 | 372 | a) Disclaiming warranty or limiting liability differently from the 373 | terms of sections 15 and 16 of this License; or 374 | 375 | b) Requiring preservation of specified reasonable legal notices or 376 | author attributions in that material or in the Appropriate Legal 377 | Notices displayed by works containing it; or 378 | 379 | c) Prohibiting misrepresentation of the origin of that material, or 380 | requiring that modified versions of such material be marked in 381 | reasonable ways as different from the original version; or 382 | 383 | d) Limiting the use for publicity purposes of names of licensors or 384 | authors of the material; or 385 | 386 | e) Declining to grant rights under trademark law for use of some 387 | trade names, trademarks, or service marks; or 388 | 389 | f) Requiring indemnification of licensors and authors of that 390 | material by anyone who conveys the material (or modified versions of 391 | it) with contractual assumptions of liability to the recipient, for 392 | any liability that these contractual assumptions directly impose on 393 | those licensors and authors. 394 | 395 | All other non-permissive additional terms are considered "further 396 | restrictions" within the meaning of section 10. If the Program as you 397 | received it, or any part of it, contains a notice stating that it is 398 | governed by this License along with a term that is a further 399 | restriction, you may remove that term. If a license document contains 400 | a further restriction but permits relicensing or conveying under this 401 | License, you may add to a covered work material governed by the terms 402 | of that license document, provided that the further restriction does 403 | not survive such relicensing or conveying. 404 | 405 | If you add terms to a covered work in accord with this section, you 406 | must place, in the relevant source files, a statement of the 407 | additional terms that apply to those files, or a notice indicating 408 | where to find the applicable terms. 409 | 410 | Additional terms, permissive or non-permissive, may be stated in the 411 | form of a separately written license, or stated as exceptions; 412 | the above requirements apply either way. 413 | 414 | 8. Termination. 415 | 416 | You may not propagate or modify a covered work except as expressly 417 | provided under this License. Any attempt otherwise to propagate or 418 | modify it is void, and will automatically terminate your rights under 419 | this License (including any patent licenses granted under the third 420 | paragraph of section 11). 421 | 422 | However, if you cease all violation of this License, then your 423 | license from a particular copyright holder is reinstated (a) 424 | provisionally, unless and until the copyright holder explicitly and 425 | finally terminates your license, and (b) permanently, if the copyright 426 | holder fails to notify you of the violation by some reasonable means 427 | prior to 60 days after the cessation. 428 | 429 | Moreover, your license from a particular copyright holder is 430 | reinstated permanently if the copyright holder notifies you of the 431 | violation by some reasonable means, this is the first time you have 432 | received notice of violation of this License (for any work) from that 433 | copyright holder, and you cure the violation prior to 30 days after 434 | your receipt of the notice. 435 | 436 | Termination of your rights under this section does not terminate the 437 | licenses of parties who have received copies or rights from you under 438 | this License. If your rights have been terminated and not permanently 439 | reinstated, you do not qualify to receive new licenses for the same 440 | material under section 10. 441 | 442 | 9. Acceptance Not Required for Having Copies. 443 | 444 | You are not required to accept this License in order to receive or 445 | run a copy of the Program. Ancillary propagation of a covered work 446 | occurring solely as a consequence of using peer-to-peer transmission 447 | to receive a copy likewise does not require acceptance. However, 448 | nothing other than this License grants you permission to propagate or 449 | modify any covered work. These actions infringe copyright if you do 450 | not accept this License. Therefore, by modifying or propagating a 451 | covered work, you indicate your acceptance of this License to do so. 452 | 453 | 10. Automatic Licensing of Downstream Recipients. 454 | 455 | Each time you convey a covered work, the recipient automatically 456 | receives a license from the original licensors, to run, modify and 457 | propagate that work, subject to this License. You are not responsible 458 | for enforcing compliance by third parties with this License. 459 | 460 | An "entity transaction" is a transaction transferring control of an 461 | organization, or substantially all assets of one, or subdividing an 462 | organization, or merging organizations. If propagation of a covered 463 | work results from an entity transaction, each party to that 464 | transaction who receives a copy of the work also receives whatever 465 | licenses to the work the party's predecessor in interest had or could 466 | give under the previous paragraph, plus a right to possession of the 467 | Corresponding Source of the work from the predecessor in interest, if 468 | the predecessor has it or can get it with reasonable efforts. 469 | 470 | You may not impose any further restrictions on the exercise of the 471 | rights granted or affirmed under this License. For example, you may 472 | not impose a license fee, royalty, or other charge for exercise of 473 | rights granted under this License, and you may not initiate litigation 474 | (including a cross-claim or counterclaim in a lawsuit) alleging that 475 | any patent claim is infringed by making, using, selling, offering for 476 | sale, or importing the Program or any portion of it. 477 | 478 | 11. Patents. 479 | 480 | A "contributor" is a copyright holder who authorizes use under this 481 | License of the Program or a work on which the Program is based. The 482 | work thus licensed is called the contributor's "contributor version". 483 | 484 | A contributor's "essential patent claims" are all patent claims 485 | owned or controlled by the contributor, whether already acquired or 486 | hereafter acquired, that would be infringed by some manner, permitted 487 | by this License, of making, using, or selling its contributor version, 488 | but do not include claims that would be infringed only as a 489 | consequence of further modification of the contributor version. For 490 | purposes of this definition, "control" includes the right to grant 491 | patent sublicenses in a manner consistent with the requirements of 492 | this License. 493 | 494 | Each contributor grants you a non-exclusive, worldwide, royalty-free 495 | patent license under the contributor's essential patent claims, to 496 | make, use, sell, offer for sale, import and otherwise run, modify and 497 | propagate the contents of its contributor version. 498 | 499 | In the following three paragraphs, a "patent license" is any express 500 | agreement or commitment, however denominated, not to enforce a patent 501 | (such as an express permission to practice a patent or covenant not to 502 | sue for patent infringement). To "grant" such a patent license to a 503 | party means to make such an agreement or commitment not to enforce a 504 | patent against the party. 505 | 506 | If you convey a covered work, knowingly relying on a patent license, 507 | and the Corresponding Source of the work is not available for anyone 508 | to copy, free of charge and under the terms of this License, through a 509 | publicly available network server or other readily accessible means, 510 | then you must either (1) cause the Corresponding Source to be so 511 | available, or (2) arrange to deprive yourself of the benefit of the 512 | patent license for this particular work, or (3) arrange, in a manner 513 | consistent with the requirements of this License, to extend the patent 514 | license to downstream recipients. "Knowingly relying" means you have 515 | actual knowledge that, but for the patent license, your conveying the 516 | covered work in a country, or your recipient's use of the covered work 517 | in a country, would infringe one or more identifiable patents in that 518 | country that you have reason to believe are valid. 519 | 520 | If, pursuant to or in connection with a single transaction or 521 | arrangement, you convey, or propagate by procuring conveyance of, a 522 | covered work, and grant a patent license to some of the parties 523 | receiving the covered work authorizing them to use, propagate, modify 524 | or convey a specific copy of the covered work, then the patent license 525 | you grant is automatically extended to all recipients of the covered 526 | work and works based on it. 527 | 528 | A patent license is "discriminatory" if it does not include within 529 | the scope of its coverage, prohibits the exercise of, or is 530 | conditioned on the non-exercise of one or more of the rights that are 531 | specifically granted under this License. You may not convey a covered 532 | work if you are a party to an arrangement with a third party that is 533 | in the business of distributing software, under which you make payment 534 | to the third party based on the extent of your activity of conveying 535 | the work, and under which the third party grants, to any of the 536 | parties who would receive the covered work from you, a discriminatory 537 | patent license (a) in connection with copies of the covered work 538 | conveyed by you (or copies made from those copies), or (b) primarily 539 | for and in connection with specific products or compilations that 540 | contain the covered work, unless you entered into that arrangement, 541 | or that patent license was granted, prior to 28 March 2007. 542 | 543 | Nothing in this License shall be construed as excluding or limiting 544 | any implied license or other defenses to infringement that may 545 | otherwise be available to you under applicable patent law. 546 | 547 | 12. No Surrender of Others' Freedom. 548 | 549 | If conditions are imposed on you (whether by court order, agreement or 550 | otherwise) that contradict the conditions of this License, they do not 551 | excuse you from the conditions of this License. If you cannot convey a 552 | covered work so as to satisfy simultaneously your obligations under this 553 | License and any other pertinent obligations, then as a consequence you may 554 | not convey it at all. For example, if you agree to terms that obligate you 555 | to collect a royalty for further conveying from those to whom you convey 556 | the Program, the only way you could satisfy both those terms and this 557 | License would be to refrain entirely from conveying the Program. 558 | 559 | 13. Use with the GNU Affero General Public License. 560 | 561 | Notwithstanding any other provision of this License, you have 562 | permission to link or combine any covered work with a work licensed 563 | under version 3 of the GNU Affero General Public License into a single 564 | combined work, and to convey the resulting work. The terms of this 565 | License will continue to apply to the part which is the covered work, 566 | but the special requirements of the GNU Affero General Public License, 567 | section 13, concerning interaction through a network will apply to the 568 | combination as such. 569 | 570 | 14. Revised Versions of this License. 571 | 572 | The Free Software Foundation may publish revised and/or new versions of 573 | the GNU General Public License from time to time. Such new versions will 574 | be similar in spirit to the present version, but may differ in detail to 575 | address new problems or concerns. 576 | 577 | Each version is given a distinguishing version number. If the 578 | Program specifies that a certain numbered version of the GNU General 579 | Public License "or any later version" applies to it, you have the 580 | option of following the terms and conditions either of that numbered 581 | version or of any later version published by the Free Software 582 | Foundation. If the Program does not specify a version number of the 583 | GNU General Public License, you may choose any version ever published 584 | by the Free Software Foundation. 585 | 586 | If the Program specifies that a proxy can decide which future 587 | versions of the GNU General Public License can be used, that proxy's 588 | public statement of acceptance of a version permanently authorizes you 589 | to choose that version for the Program. 590 | 591 | Later license versions may give you additional or different 592 | permissions. However, no additional obligations are imposed on any 593 | author or copyright holder as a result of your choosing to follow a 594 | later version. 595 | 596 | 15. Disclaimer of Warranty. 597 | 598 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 599 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 600 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 601 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 602 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 603 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 604 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 605 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 606 | 607 | 16. Limitation of Liability. 608 | 609 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 610 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 611 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 612 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 613 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 614 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 615 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 616 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 617 | SUCH DAMAGES. 618 | 619 | 17. Interpretation of Sections 15 and 16. 620 | 621 | If the disclaimer of warranty and limitation of liability provided 622 | above cannot be given local legal effect according to their terms, 623 | reviewing courts shall apply local law that most closely approximates 624 | an absolute waiver of all civil liability in connection with the 625 | Program, unless a warranty or assumption of liability accompanies a 626 | copy of the Program in return for a fee. 627 | 628 | END OF TERMS AND CONDITIONS 629 | 630 | How to Apply These Terms to Your New Programs 631 | 632 | If you develop a new program, and you want it to be of the greatest 633 | possible use to the public, the best way to achieve this is to make it 634 | free software which everyone can redistribute and change under these terms. 635 | 636 | To do so, attach the following notices to the program. It is safest 637 | to attach them to the start of each source file to most effectively 638 | state the exclusion of warranty; and each file should have at least 639 | the "copyright" line and a pointer to where the full notice is found. 640 | 641 | {one line to give the program's name and a brief idea of what it does.} 642 | Copyright (C) {year} {name of author} 643 | 644 | This program is free software: you can redistribute it and/or modify 645 | it under the terms of the GNU General Public License as published by 646 | the Free Software Foundation, either version 3 of the License, or 647 | (at your option) any later version. 648 | 649 | This program is distributed in the hope that it will be useful, 650 | but WITHOUT ANY WARRANTY; without even the implied warranty of 651 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 652 | GNU General Public License for more details. 653 | 654 | You should have received a copy of the GNU General Public License 655 | along with this program. If not, see . 656 | 657 | Also add information on how to contact you by electronic and paper mail. 658 | 659 | If the program does terminal interaction, make it output a short 660 | notice like this when it starts in an interactive mode: 661 | 662 | {project} Copyright (C) {year} {fullname} 663 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 664 | This is free software, and you are welcome to redistribute it 665 | under certain conditions; type `show c' for details. 666 | 667 | The hypothetical commands `show w' and `show c' should show the appropriate 668 | parts of the General Public License. Of course, your program's commands 669 | might be different; for a GUI interface, you would use an "about box". 670 | 671 | You should also get your employer (if you work as a programmer) or school, 672 | if any, to sign a "copyright disclaimer" for the program, if necessary. 673 | For more information on this, and how to apply and follow the GNU GPL, see 674 | . 675 | 676 | The GNU General Public License does not permit incorporating your program 677 | into proprietary programs. If your program is a subroutine library, you 678 | may consider it more useful to permit linking proprietary applications with 679 | the library. If this is what you want to do, use the GNU Lesser General 680 | Public License instead of this License. But first, please read 681 | . 682 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # neural-colorization 2 | 3 | [![Build Status](https://www.travis-ci.org/zeruniverse/neural-colorization.svg?branch=pytorch)](https://www.travis-ci.org/zeruniverse/neural-colorization) 4 | ![Environment](https://img.shields.io/badge/python-3.6-blue.svg) 5 | ![License](https://img.shields.io/github/license/zeruniverse/QQRobot.svg) 6 | 7 | GAN for image colorization based on [Johnson's network structure](https://github.com/jcjohnson/fast-neural-style). 8 | 9 | ![Result](https://cloud.githubusercontent.com/assets/4648756/20504440/4067e0f6-affc-11e6-88e7-26de6f5c1cce.jpg) 10 | 11 | ## Setup 12 | 13 | Install the following Python libraries: 14 | + numpy 15 | + scipy 16 | + Pytorch 17 | + scikit-image 18 | + Pillow 19 | + opencv-python 20 | 21 | 22 | ## Colorize images 23 | 24 | ```bash 25 | #Download pre-trained model 26 | wget -O model.pth "https://github.com/zeruniverse/neural-colorization/releases/download/1.1/G.pth" 27 | 28 | #Colorize an image with CPU 29 | python colorize.py -m model.pth -i input.jpg -o output.jpg --gpu -1 30 | 31 | # If you want to colorize all images in a folder with GPU 32 | python colorize.py -m model.pth -i input -o output --gpu 0 33 | ``` 34 | 35 | ## Train your own model 36 | 37 | Note: Training is only supported with GPU (CUDA). 38 | 39 | ### Prepare dataset 40 | 41 | + Download some datasets and unzip them into a same folder (saying `train_raw_dataset`). If the images are not in `.jpg` format, you should convert them all in `.jpg`s. 42 | + run `python build_dataset_directory.py -i train_raw_dataset -o train` (you can skip this if all your images are **directly** under the `train_raw_dataset`, in which case, just rename the folder as `train`) 43 | + run `python resize_all_imgs.py -d train` to resize all training images into `256*256` (you can skip this if your images are already in `256*256`) 44 | 45 | ### Optional preparation 46 | 47 | It's highly recommended to train from my pretrained models. You can get both generator model and discriminator model from the GitHub Release: 48 | 49 | ```bash 50 | wget "https://github.com/zeruniverse/neural-colorization/releases/download/1.1/G.pth" 51 | wget "https://github.com/zeruniverse/neural-colorization/releases/download/1.1/D.pth" 52 | ``` 53 | 54 | It's also recommended to have a test image (the script will generate a colorization for the test image you give at every checkpoint so you can see how the model works during training). 55 | 56 | 57 | ### Training 58 | 59 | The required arguments are training image directory (e.g. `train`) and path to save checkpoints (e.g. `checkpoints`) 60 | 61 | ```bash 62 | python train.py -d train -c chekpoints 63 | ``` 64 | 65 | To add initial weights and test images: 66 | 67 | ```bash 68 | python train.py -d train -c chekpoints --d_init D.pth --g_init G.pth -t test.jpg 69 | ``` 70 | 71 | More options are available and you can run `python train.py --help` to print all options. 72 | 73 | For torch equivalent (no GAN), you can set option `-p 1e9` (set a very large weight for pixel loss). 74 | 75 | ## Reference 76 | [Perceptual Losses for Real-Time Style Transfer and Super-Resolution](https://github.com/jcjohnson/fast-neural-style) 77 | 78 | ## License 79 | 80 | GNU GPL 3.0 for personal or research use. COMMERCIAL USE PROHIBITED. 81 | 82 | Model weights are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) 83 | -------------------------------------------------------------------------------- /build_dataset_directory.py: -------------------------------------------------------------------------------- 1 | import os 2 | import shutil 3 | import argparse 4 | image_extensions = {'.jpg', '.jpeg', '.JPG', '.JPEG'} 5 | 6 | def parse_args(): 7 | parser = argparse.ArgumentParser(description="Put all places 365 images in single folder.") 8 | parser.add_argument("-i", 9 | "--input_dir", 10 | required=True, 11 | type=str, 12 | help="input folder: the folder containing unzipped places 365 files") 13 | parser.add_argument("-o", 14 | "--output_dir", 15 | required=True, 16 | type=str, 17 | help="output folder: the folder to put all images") 18 | args = parser.parse_args() 19 | return args 20 | 21 | def genlist(image_dir): 22 | image_list = [] 23 | for filename in os.listdir(image_dir): 24 | path = os.path.join(image_dir,filename) 25 | if os.path.isdir(path): 26 | image_list = image_list + genlist(path) 27 | else: 28 | ext = os.path.splitext(filename)[1] 29 | if ext in image_extensions: 30 | image_list.append(os.path.join(image_dir, filename)) 31 | return image_list 32 | 33 | 34 | args = parse_args() 35 | if not os.path.exists(args.output_dir): 36 | os.makedirs(args.output_dir) 37 | flist = genlist(args.input_dir) 38 | for i,p in enumerate(flist): 39 | if os.path.getsize(p) != 0: 40 | os.rename(p,os.path.join(args.output_dir,str(i)+'.jpg')) 41 | shutil.rmtree(args.input_dir) 42 | print('done') -------------------------------------------------------------------------------- /colorize.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from model import generator 3 | from torch.autograd import Variable 4 | from scipy.ndimage import zoom 5 | import cv2 6 | import os 7 | from PIL import Image 8 | import argparse 9 | import numpy as np 10 | from skimage.color import rgb2yuv,yuv2rgb 11 | 12 | def parse_args(): 13 | parser = argparse.ArgumentParser(description="Colorize images") 14 | parser.add_argument("-i", 15 | "--input", 16 | type=str, 17 | required=True, 18 | help="input image/input dir") 19 | parser.add_argument("-o", 20 | "--output", 21 | type=str, 22 | required=True, 23 | help="output image/output dir") 24 | parser.add_argument("-m", 25 | "--model", 26 | type=str, 27 | required=True, 28 | help="location for model (Generator)") 29 | parser.add_argument("--gpu", 30 | type=int, 31 | default=-1, 32 | help="which GPU to use? [-1 for cpu]") 33 | args = parser.parse_args() 34 | return args 35 | 36 | args = parse_args() 37 | 38 | G = generator() 39 | 40 | if args.gpu>=0: 41 | G=G.cuda(args.gpu) 42 | G.load_state_dict(torch.load(args.model)) 43 | else: 44 | G.load_state_dict(torch.load(args.model,map_location={'cuda:0': 'cpu'})) 45 | 46 | def inference(G,in_path,out_path): 47 | p=Image.open(in_path).convert('RGB') 48 | img_yuv = rgb2yuv(p) 49 | H,W,_ = img_yuv.shape 50 | infimg = np.expand_dims(np.expand_dims(img_yuv[...,0], axis=0), axis=0) 51 | img_variable = Variable(torch.Tensor(infimg-0.5)) 52 | if args.gpu>=0: 53 | img_variable=img_variable.cuda(args.gpu) 54 | res = G(img_variable) 55 | uv=res.cpu().detach().numpy() 56 | uv[:,0,:,:] *= 0.436 57 | uv[:,1,:,:] *= 0.615 58 | (_,_,H1,W1) = uv.shape 59 | uv = zoom(uv,(1,1,H/H1,W/W1)) 60 | yuv = np.concatenate([infimg,uv],axis=1)[0] 61 | rgb=yuv2rgb(yuv.transpose(1,2,0)) 62 | cv2.imwrite(out_path,(rgb.clip(min=0,max=1)*256)[:,:,[2,1,0]]) 63 | 64 | 65 | if not os.path.isdir(args.input): 66 | inference(G,args.input,args.output) 67 | else: 68 | if not os.path.exists(args.output): 69 | os.makedirs(args.output) 70 | for f in os.listdir(args.input): 71 | inference(G,os.path.join(args.input,f),os.path.join(args.output,f)) -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from functools import reduce 4 | from torch.autograd import Variable 5 | 6 | 7 | class shave_block(nn.Module): 8 | def __init__(self, s): 9 | super(shave_block, self).__init__() 10 | self.s=s 11 | def forward(self,x): 12 | return x[:,:,self.s:-self.s,self.s:-self.s] 13 | 14 | class LambdaBase(nn.Sequential): 15 | def __init__(self, fn, *args): 16 | super(LambdaBase, self).__init__(*args) 17 | self.lambda_func = fn 18 | 19 | def forward_prepare(self, input): 20 | output = [] 21 | for module in self._modules.values(): 22 | output.append(module(input)) 23 | return output if output else input 24 | 25 | class Lambda(LambdaBase): 26 | def forward(self, input): 27 | return self.lambda_func(self.forward_prepare(input)) 28 | 29 | class LambdaMap(LambdaBase): 30 | def forward(self, input): 31 | return list(map(self.lambda_func,self.forward_prepare(input))) 32 | 33 | class LambdaReduce(LambdaBase): 34 | def forward(self, input): 35 | return reduce(self.lambda_func,self.forward_prepare(input)) 36 | 37 | def generator(): 38 | G = nn.Sequential( # Sequential, 39 | nn.ReflectionPad2d((40, 40, 40, 40)), 40 | nn.Conv2d(1,32,(9, 9),(1, 1),(4, 4)), 41 | nn.BatchNorm2d(32), 42 | nn.ReLU(), 43 | nn.Conv2d(32,64,(3, 3),(2, 2),(1, 1)), 44 | nn.BatchNorm2d(64), 45 | nn.ReLU(), 46 | nn.Conv2d(64,128,(3, 3),(2, 2),(1, 1)), 47 | nn.BatchNorm2d(128), 48 | nn.ReLU(), 49 | nn.Sequential( # Sequential, 50 | LambdaMap(lambda x: x, # ConcatTable, 51 | nn.Sequential( # Sequential, 52 | nn.Conv2d(128,128,(3, 3)), 53 | nn.BatchNorm2d(128), 54 | nn.ReLU(), 55 | nn.Conv2d(128,128,(3, 3)), 56 | nn.BatchNorm2d(128), 57 | ), 58 | shave_block(2), 59 | ), 60 | LambdaReduce(lambda x,y: x+y), # CAddTable, 61 | ), 62 | nn.Sequential( # Sequential, 63 | LambdaMap(lambda x: x, # ConcatTable, 64 | nn.Sequential( # Sequential, 65 | nn.Conv2d(128,128,(3, 3)), 66 | nn.BatchNorm2d(128), 67 | nn.ReLU(), 68 | nn.Conv2d(128,128,(3, 3)), 69 | nn.BatchNorm2d(128), 70 | ), 71 | shave_block(2), 72 | ), 73 | LambdaReduce(lambda x,y: x+y), # CAddTable, 74 | ), 75 | nn.Sequential( # Sequential, 76 | LambdaMap(lambda x: x, # ConcatTable, 77 | nn.Sequential( # Sequential, 78 | nn.Conv2d(128,128,(3, 3)), 79 | nn.BatchNorm2d(128), 80 | nn.ReLU(), 81 | nn.Conv2d(128,128,(3, 3)), 82 | nn.BatchNorm2d(128), 83 | ), 84 | shave_block(2), 85 | ), 86 | LambdaReduce(lambda x,y: x+y), # CAddTable, 87 | ), 88 | nn.Sequential( # Sequential, 89 | LambdaMap(lambda x: x, # ConcatTable, 90 | nn.Sequential( # Sequential, 91 | nn.Conv2d(128,128,(3, 3)), 92 | nn.BatchNorm2d(128), 93 | nn.ReLU(), 94 | nn.Conv2d(128,128,(3, 3)), 95 | nn.BatchNorm2d(128), 96 | ), 97 | shave_block(2), 98 | ), 99 | LambdaReduce(lambda x,y: x+y), # CAddTable, 100 | ), 101 | nn.Sequential( # Sequential, 102 | LambdaMap(lambda x: x, # ConcatTable, 103 | nn.Sequential( # Sequential, 104 | nn.Conv2d(128,128,(3, 3)), 105 | nn.BatchNorm2d(128), 106 | nn.ReLU(), 107 | nn.Conv2d(128,128,(3, 3)), 108 | nn.BatchNorm2d(128), 109 | ), 110 | shave_block(2), 111 | ), 112 | LambdaReduce(lambda x,y: x+y), # CAddTable, 113 | ), 114 | nn.ConvTranspose2d(128,64,(3, 3),(2, 2),(1, 1),(1, 1)), 115 | nn.BatchNorm2d(64), 116 | nn.ReLU(), 117 | nn.ConvTranspose2d(64,32,(3, 3),(2, 2),(1, 1),(1, 1)), 118 | nn.BatchNorm2d(32), 119 | nn.ReLU(), 120 | nn.Conv2d(32,2,(9, 9),(1, 1),(4, 4)), 121 | nn.Tanh(), 122 | ) 123 | return G -------------------------------------------------------------------------------- /resize_all_imgs.py: -------------------------------------------------------------------------------- 1 | from multiprocessing import Pool 2 | from PIL import Image 3 | import os 4 | import argparse 5 | 6 | def parse_args(): 7 | parser = argparse.ArgumentParser(description="Resize all colorful imgs to 256*256 for training") 8 | parser.add_argument("-d", 9 | "--dir", 10 | required=True, 11 | type=str, 12 | help="The directory includes all jpg images") 13 | parser.add_argument("-n", 14 | "--nprocesses", 15 | default=10, 16 | type=int, 17 | help="Using how many processes") 18 | args = parser.parse_args() 19 | return args 20 | 21 | def doit(x): 22 | a=Image.open(x) 23 | if a.getbands()!=('R','G','B'): 24 | os.remove(x) 25 | return 26 | a.resize((256,256),Image.BICUBIC).save(x) 27 | return 28 | 29 | args=parse_args() 30 | pool = Pool(processes=args.nprocesses) 31 | jpgs = [] 32 | flist = os.listdir(args.dir) 33 | full_flist = [os.path.join(args.dir,x) for x in flist] 34 | pool.map(doit, full_flist) 35 | print('done') -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import argparse 4 | from torch.autograd import Variable 5 | import torchvision.models as models 6 | import os 7 | from torch.utils import data 8 | from model import generator 9 | import numpy as np 10 | from PIL import Image 11 | from skimage.color import rgb2yuv,yuv2rgb 12 | import cv2 13 | 14 | def parse_args(): 15 | parser = argparse.ArgumentParser(description="Train a GAN based model") 16 | parser.add_argument("-d", 17 | "--training_dir", 18 | type=str, 19 | required=True, 20 | help="Training directory (folder contains all 256*256 images)") 21 | parser.add_argument("-t", 22 | "--test_image", 23 | type=str, 24 | default=None, 25 | help="Test image location") 26 | parser.add_argument("-c", 27 | "--checkpoint_location", 28 | type=str, 29 | required=True, 30 | help="Place to save checkpoints") 31 | parser.add_argument("-e", 32 | "--epoch", 33 | type=int, 34 | default=120, 35 | help="Epoches to run training") 36 | parser.add_argument("--gpu", 37 | type=int, 38 | default=0, 39 | help="which GPU to use?") 40 | parser.add_argument("-b", 41 | "--batch_size", 42 | type=int, 43 | default=20, 44 | help="batch size") 45 | parser.add_argument("-w", 46 | "--num_workers", 47 | type=int, 48 | default=6, 49 | help="Number of workers to fetch data") 50 | parser.add_argument("-p", 51 | "--pixel_loss_weights", 52 | type=float, 53 | default=1000.0, 54 | help="Pixel-wise loss weights") 55 | parser.add_argument("--g_every", 56 | type=int, 57 | default=1, 58 | help="Training generator every k iteration") 59 | parser.add_argument("--g_lr", 60 | type=float, 61 | default=1e-4, 62 | help="learning rate for generator") 63 | parser.add_argument("--d_lr", 64 | type=float, 65 | default=1e-4, 66 | help="learning rate for discriminator") 67 | parser.add_argument("-i", 68 | "--checkpoint_every", 69 | type=int, 70 | default=100, 71 | help="Save checkpoint every k iteration (checkpoints for same epoch will overwrite)") 72 | parser.add_argument("--d_init", 73 | type=str, 74 | default=None, 75 | help="Init weights for discriminator") 76 | parser.add_argument("--g_init", 77 | type=str, 78 | default=None, 79 | help="Init weights for generator") 80 | args = parser.parse_args() 81 | return args 82 | 83 | # define data generator 84 | class img_data(data.Dataset): 85 | def __init__(self, path): 86 | files = os.listdir(path) 87 | self.files = [os.path.join(path,x) for x in files] 88 | def __len__(self): 89 | return len(self.files) 90 | 91 | def __getitem__(self, index): 92 | img = Image.open(self.files[index]) 93 | yuv = rgb2yuv(img) 94 | y = yuv[...,0]-0.5 95 | u_t = yuv[...,1] / 0.43601035 96 | v_t = yuv[...,2] / 0.61497538 97 | return torch.Tensor(np.expand_dims(y,axis=0)),torch.Tensor(np.stack([u_t,v_t],axis=0)) 98 | 99 | 100 | args = parse_args() 101 | if not os.path.exists(os.path.join(args.checkpoint_location,'weights')): 102 | os.makedirs(os.path.join(args.checkpoint_location,'weights')) 103 | 104 | # Define G, same as torch version 105 | G = generator().cuda(args.gpu) 106 | 107 | # define D 108 | D = models.resnet18(pretrained=False,num_classes=2) 109 | D.fc = nn.Sequential(nn.Linear(512, 1), nn.Sigmoid()) 110 | D = D.cuda(args.gpu) 111 | 112 | trainset = img_data(args.training_dir) 113 | params = {'batch_size': args.batch_size, 114 | 'shuffle': True, 115 | 'num_workers': args.num_workers} 116 | training_generator = data.DataLoader(trainset, **params) 117 | if args.test_image is not None: 118 | test_img = Image.open(args.test_image).convert('RGB').resize((256,256)) 119 | test_yuv = rgb2yuv(test_img) 120 | test_inf = test_yuv[...,0].reshape(1,1,256,256) 121 | test_var = Variable(torch.Tensor(test_inf-0.5)).cuda(args.gpu) 122 | if args.d_init is not None: 123 | D.load_state_dict(torch.load(args.d_init)) 124 | if args.g_init is not None: 125 | G.load_state_dict(torch.load(args.g_init)) 126 | 127 | # save test image for beginning 128 | if args.test_image is not None: 129 | test_res = G(test_var) 130 | uv=test_res.cpu().detach().numpy() 131 | uv[:,0,:,:] *= 0.436 132 | uv[:,1,:,:] *= 0.615 133 | test_yuv = np.concatenate([test_inf,uv],axis=1).reshape(3,256,256) 134 | test_rgb = yuv2rgb(test_yuv.transpose(1,2,0)) 135 | cv2.imwrite(os.path.join(args.checkpoint_location,'test_init.jpg'),(test_rgb.clip(min=0,max=1)*256)[:,:,[2,1,0]]) 136 | 137 | i=0 138 | adversarial_loss = torch.nn.BCELoss() 139 | optimizer_G = torch.optim.Adam(G.parameters(), lr=args.g_lr, betas=(0.5, 0.999)) 140 | optimizer_D = torch.optim.Adam(D.parameters(), lr=args.d_lr, betas=(0.5, 0.999)) 141 | for epoch in range(args.epoch): 142 | for y, uv in training_generator: 143 | # Adversarial ground truths 144 | valid = Variable(torch.Tensor(y.size(0), 1).fill_(1.0), requires_grad=False).cuda(args.gpu) 145 | fake = Variable(torch.Tensor(y.size(0), 1).fill_(0.0), requires_grad=False).cuda(args.gpu) 146 | 147 | yvar = Variable(y).cuda(args.gpu) 148 | uvvar = Variable(uv).cuda(args.gpu) 149 | real_imgs = torch.cat([yvar,uvvar],dim=1) 150 | 151 | optimizer_G.zero_grad() 152 | uvgen = G(yvar) 153 | # Generate a batch of images 154 | gen_imgs = torch.cat([yvar.detach(),uvgen],dim=1) 155 | 156 | # Loss measures generator's ability to fool the discriminator 157 | g_loss_gan = adversarial_loss(D(gen_imgs), valid) 158 | g_loss = g_loss_gan + args.pixel_loss_weights * torch.mean((uvvar-uvgen)**2) 159 | if i%args.g_every==0: 160 | g_loss.backward() 161 | optimizer_G.step() 162 | 163 | optimizer_D.zero_grad() 164 | 165 | # Measure discriminator's ability to classify real from generated samples 166 | real_loss = adversarial_loss(D(real_imgs), valid) 167 | fake_loss = adversarial_loss(D(gen_imgs.detach()), fake) 168 | d_loss = (real_loss + fake_loss) / 2 169 | d_loss.backward() 170 | optimizer_D.step() 171 | i+=1 172 | if i%args.checkpoint_every==0: 173 | print ("Epoch: %d: [D loss: %f] [G total loss: %f] [G GAN Loss: %f]" % (epoch, d_loss.item(), g_loss.item(), g_loss_gan.item())) 174 | 175 | torch.save(D.state_dict(), os.path.join(args.checkpoint_location,'weights','D'+str(epoch)+'.pth')) 176 | torch.save(G.state_dict(), os.path.join(args.checkpoint_location,'weights','G'+str(epoch)+'.pth')) 177 | if args.test_image is not None: 178 | test_res = G(test_var) 179 | uv=test_res.cpu().detach().numpy() 180 | uv[:,0,:,:] *= 0.436 181 | uv[:,1,:,:] *= 0.615 182 | test_yuv = np.concatenate([test_inf,uv],axis=1).reshape(3,256,256) 183 | test_rgb = yuv2rgb(test_yuv.transpose(1,2,0)) 184 | cv2.imwrite(os.path.join(args.checkpoint_location,'test_epoch_'+str(epoch)+'.jpg'),(test_rgb.clip(min=0,max=1)*256)[:,:,[2,1,0]]) 185 | torch.save(D.state_dict(), os.path.join(args.checkpoint_location,'D_final.pth')) 186 | torch.save(G.state_dict(), os.path.join(args.checkpoint_location,'G_final.pth')) 187 | --------------------------------------------------------------------------------