├── .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 | [](https://www.travis-ci.org/zeruniverse/neural-colorization)
4 | 
5 | 
6 |
7 | GAN for image colorization based on [Johnson's network structure](https://github.com/jcjohnson/fast-neural-style).
8 |
9 | 
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 |
--------------------------------------------------------------------------------