├── Data
├── datasets
│ └── flowers
│ │ └── allclasses.txt
└── text.txt
├── LICENSE
├── README.md
├── Utils
├── image_processing.py
├── inception_score.py
├── ops.py
└── plot_msssim.py
├── _config.yml
├── create_dataset.py
├── dataprep.py
├── decoder.py
├── encode_text.py
├── encoder.py
├── generate_images.py
├── inception_score.py
├── model.py
├── msssim.py
├── requirements.txt
├── skipthoughts.py
├── t_interpolation.py
├── train.py
└── z_interpolation.py
/Data/datasets/flowers/allclasses.txt:
--------------------------------------------------------------------------------
1 | class_00001
2 | class_00002
3 | class_00003
4 | class_00004
5 | class_00005
6 | class_00006
7 | class_00007
8 | class_00008
9 | class_00009
10 | class_00010
11 | class_00011
12 | class_00012
13 | class_00013
14 | class_00014
15 | class_00015
16 | class_00016
17 | class_00017
18 | class_00018
19 | class_00019
20 | class_00020
21 | class_00021
22 | class_00022
23 | class_00023
24 | class_00024
25 | class_00025
26 | class_00026
27 | class_00027
28 | class_00028
29 | class_00029
30 | class_00030
31 | class_00031
32 | class_00032
33 | class_00033
34 | class_00034
35 | class_00035
36 | class_00036
37 | class_00037
38 | class_00038
39 | class_00039
40 | class_00040
41 | class_00041
42 | class_00042
43 | class_00043
44 | class_00044
45 | class_00045
46 | class_00046
47 | class_00047
48 | class_00048
49 | class_00049
50 | class_00050
51 | class_00051
52 | class_00052
53 | class_00053
54 | class_00054
55 | class_00055
56 | class_00056
57 | class_00057
58 | class_00058
59 | class_00059
60 | class_00060
61 | class_00061
62 | class_00062
63 | class_00063
64 | class_00064
65 | class_00065
66 | class_00066
67 | class_00067
68 | class_00068
69 | class_00069
70 | class_00070
71 | class_00071
72 | class_00072
73 | class_00073
74 | class_00074
75 | class_00075
76 | class_00076
77 | class_00077
78 | class_00078
79 | class_00079
80 | class_00080
81 | class_00081
82 | class_00082
83 | class_00083
84 | class_00084
85 | class_00085
86 | class_00086
87 | class_00087
88 | class_00088
89 | class_00089
90 | class_00090
91 | class_00091
92 | class_00092
93 | class_00093
94 | class_00094
95 | class_00095
96 | class_00096
97 | class_00097
98 | class_00098
99 | class_00099
100 | class_00100
101 | class_00101
102 | class_00102
--------------------------------------------------------------------------------
/Data/text.txt:
--------------------------------------------------------------------------------
1 | a flower with red petals which are pointed
2 | many pointed petals
3 | A yellow flower
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | GNU GENERAL PUBLIC LICENSE
2 | Version 3, 29 June 2007
3 |
4 | Copyright (C) 2007 Free Software Foundation, Inc.
5 | Everyone is permitted to copy and distribute verbatim copies
6 | of this license document, but changing it is not allowed.
7 |
8 | Preamble
9 |
10 | The GNU General Public License is a free, copyleft license for
11 | software and other kinds of works.
12 |
13 | The licenses for most software and other practical works are designed
14 | to take away your freedom to share and change the works. By contrast,
15 | the GNU General Public License is intended to guarantee your freedom to
16 | share and change all versions of a program--to make sure it remains free
17 | software for all its users. We, the Free Software Foundation, use the
18 | GNU General Public License for most of our software; it applies also to
19 | any other work released this way by its authors. You can apply it to
20 | your programs, too.
21 |
22 | When we speak of free software, we are referring to freedom, not
23 | price. Our General Public Licenses are designed to make sure that you
24 | have the freedom to distribute copies of free software (and charge for
25 | them if you wish), that you receive source code or can get it if you
26 | want it, that you can change the software or use pieces of it in new
27 | free programs, and that you know you can do these things.
28 |
29 | To protect your rights, we need to prevent others from denying you
30 | these rights or asking you to surrender the rights. Therefore, you have
31 | certain responsibilities if you distribute copies of the software, or if
32 | you modify it: responsibilities to respect the freedom of others.
33 |
34 | For example, if you distribute copies of such a program, whether
35 | gratis or for a fee, you must pass on to the recipients the same
36 | freedoms that you received. You must make sure that they, too, receive
37 | or can get the source code. And you must show them these terms so they
38 | know their rights.
39 |
40 | Developers that use the GNU GPL protect your rights with two steps:
41 | (1) assert copyright on the software, and (2) offer you this License
42 | giving you legal permission to copy, distribute and/or modify it.
43 |
44 | For the developers' and authors' protection, the GPL clearly explains
45 | that there is no warranty for this free software. For both users' and
46 | authors' sake, the GPL requires that modified versions be marked as
47 | changed, so that their problems will not be attributed erroneously to
48 | authors of previous versions.
49 |
50 | Some devices are designed to deny users access to install or run
51 | modified versions of the software inside them, although the manufacturer
52 | can do so. This is fundamentally incompatible with the aim of
53 | protecting users' freedom to change the software. The systematic
54 | pattern of such abuse occurs in the area of products for individuals to
55 | use, which is precisely where it is most unacceptable. Therefore, we
56 | have designed this version of the GPL to prohibit the practice for those
57 | products. If such problems arise substantially in other domains, we
58 | stand ready to extend this provision to those domains in future versions
59 | of the GPL, as needed to protect the freedom of users.
60 |
61 | Finally, every program is threatened constantly by software patents.
62 | States should not allow patents to restrict development and use of
63 | software on general-purpose computers, but in those that do, we wish to
64 | avoid the special danger that patents applied to a free program could
65 | make it effectively proprietary. To prevent this, the GPL assures that
66 | patents cannot be used to render the program non-free.
67 |
68 | The precise terms and conditions for copying, distribution and
69 | modification follow.
70 |
71 | TERMS AND CONDITIONS
72 |
73 | 0. Definitions.
74 |
75 | "This License" refers to version 3 of the GNU General Public License.
76 |
77 | "Copyright" also means copyright-like laws that apply to other kinds of
78 | works, such as semiconductor masks.
79 |
80 | "The Program" refers to any copyrightable work licensed under this
81 | License. Each licensee is addressed as "you". "Licensees" and
82 | "recipients" may be individuals or organizations.
83 |
84 | To "modify" a work means to copy from or adapt all or part of the work
85 | in a fashion requiring copyright permission, other than the making of an
86 | exact copy. The resulting work is called a "modified version" of the
87 | earlier work or a work "based on" the earlier work.
88 |
89 | A "covered work" means either the unmodified Program or a work based
90 | on the Program.
91 |
92 | To "propagate" a work means to do anything with it that, without
93 | permission, would make you directly or secondarily liable for
94 | infringement under applicable copyright law, except executing it on a
95 | computer or modifying a private copy. Propagation includes copying,
96 | distribution (with or without modification), making available to the
97 | public, and in some countries other activities as well.
98 |
99 | To "convey" a work means any kind of propagation that enables other
100 | parties to make or receive copies. Mere interaction with a user through
101 | a computer network, with no transfer of a copy, is not conveying.
102 |
103 | An interactive user interface displays "Appropriate Legal Notices"
104 | to the extent that it includes a convenient and prominently visible
105 | feature that (1) displays an appropriate copyright notice, and (2)
106 | tells the user that there is no warranty for the work (except to the
107 | extent that warranties are provided), that licensees may convey the
108 | work under this License, and how to view a copy of this License. If
109 | the interface presents a list of user commands or options, such as a
110 | menu, a prominent item in the list meets this criterion.
111 |
112 | 1. Source Code.
113 |
114 | The "source code" for a work means the preferred form of the work
115 | for making modifications to it. "Object code" means any non-source
116 | form of a work.
117 |
118 | A "Standard Interface" means an interface that either is an official
119 | standard defined by a recognized standards body, or, in the case of
120 | interfaces specified for a particular programming language, one that
121 | is widely used among developers working in that language.
122 |
123 | The "System Libraries" of an executable work include anything, other
124 | than the work as a whole, that (a) is included in the normal form of
125 | packaging a Major Component, but which is not part of that Major
126 | Component, and (b) serves only to enable use of the work with that
127 | Major Component, or to implement a Standard Interface for which an
128 | implementation is available to the public in source code form. A
129 | "Major Component", in this context, means a major essential component
130 | (kernel, window system, and so on) of the specific operating system
131 | (if any) on which the executable work runs, or a compiler used to
132 | produce the work, or an object code interpreter used to run it.
133 |
134 | The "Corresponding Source" for a work in object code form means all
135 | the source code needed to generate, install, and (for an executable
136 | work) run the object code and to modify the work, including scripts to
137 | control those activities. However, it does not include the work's
138 | System Libraries, or general-purpose tools or generally available free
139 | programs which are used unmodified in performing those activities but
140 | which are not part of the work. For example, Corresponding Source
141 | includes interface definition files associated with source files for
142 | the work, and the source code for shared libraries and dynamically
143 | linked subprograms that the work is specifically designed to require,
144 | such as by intimate data communication or control flow between those
145 | subprograms and other parts of the work.
146 |
147 | The Corresponding Source need not include anything that users
148 | can regenerate automatically from other parts of the Corresponding
149 | Source.
150 |
151 | The Corresponding Source for a work in source code form is that
152 | same work.
153 |
154 | 2. Basic Permissions.
155 |
156 | All rights granted under this License are granted for the term of
157 | copyright on the Program, and are irrevocable provided the stated
158 | conditions are met. This License explicitly affirms your unlimited
159 | permission to run the unmodified Program. The output from running a
160 | covered work is covered by this License only if the output, given its
161 | content, constitutes a covered work. This License acknowledges your
162 | rights of fair use or other equivalent, as provided by copyright law.
163 |
164 | You may make, run and propagate covered works that you do not
165 | convey, without conditions so long as your license otherwise remains
166 | in force. You may convey covered works to others for the sole purpose
167 | of having them make modifications exclusively for you, or provide you
168 | with facilities for running those works, provided that you comply with
169 | the terms of this License in conveying all material for which you do
170 | not control copyright. Those thus making or running the covered works
171 | for you must do so exclusively on your behalf, under your direction
172 | and control, on terms that prohibit them from making any copies of
173 | your copyrighted material outside their relationship with you.
174 |
175 | Conveying under any other circumstances is permitted solely under
176 | the conditions stated below. Sublicensing is not allowed; section 10
177 | makes it unnecessary.
178 |
179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180 |
181 | No covered work shall be deemed part of an effective technological
182 | measure under any applicable law fulfilling obligations under article
183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184 | similar laws prohibiting or restricting circumvention of such
185 | measures.
186 |
187 | When you convey a covered work, you waive any legal power to forbid
188 | circumvention of technological measures to the extent such circumvention
189 | is effected by exercising rights under this License with respect to
190 | the covered work, and you disclaim any intention to limit operation or
191 | modification of the work as a means of enforcing, against the work's
192 | users, your or third parties' legal rights to forbid circumvention of
193 | technological measures.
194 |
195 | 4. Conveying Verbatim Copies.
196 |
197 | You may convey verbatim copies of the Program's source code as you
198 | receive it, in any medium, provided that you conspicuously and
199 | appropriately publish on each copy an appropriate copyright notice;
200 | keep intact all notices stating that this License and any
201 | non-permissive terms added in accord with section 7 apply to the code;
202 | keep intact all notices of the absence of any warranty; and give all
203 | recipients a copy of this License along with the Program.
204 |
205 | You may charge any price or no price for each copy that you convey,
206 | and you may offer support or warranty protection for a fee.
207 |
208 | 5. Conveying Modified Source Versions.
209 |
210 | You may convey a work based on the Program, or the modifications to
211 | produce it from the Program, in the form of source code under the
212 | terms of section 4, provided that you also meet all of these conditions:
213 |
214 | a) The work must carry prominent notices stating that you modified
215 | it, and giving a relevant date.
216 |
217 | b) The work must carry prominent notices stating that it is
218 | released under this License and any conditions added under section
219 | 7. This requirement modifies the requirement in section 4 to
220 | "keep intact all notices".
221 |
222 | c) You must license the entire work, as a whole, under this
223 | License to anyone who comes into possession of a copy. This
224 | License will therefore apply, along with any applicable section 7
225 | additional terms, to the whole of the work, and all its parts,
226 | regardless of how they are packaged. This License gives no
227 | permission to license the work in any other way, but it does not
228 | invalidate such permission if you have separately received it.
229 |
230 | d) If the work has interactive user interfaces, each must display
231 | Appropriate Legal Notices; however, if the Program has interactive
232 | interfaces that do not display Appropriate Legal Notices, your
233 | work need not make them do so.
234 |
235 | A compilation of a covered work with other separate and independent
236 | works, which are not by their nature extensions of the covered work,
237 | and which are not combined with it such as to form a larger program,
238 | in or on a volume of a storage or distribution medium, is called an
239 | "aggregate" if the compilation and its resulting copyright are not
240 | used to limit the access or legal rights of the compilation's users
241 | beyond what the individual works permit. Inclusion of a covered work
242 | in an aggregate does not cause this License to apply to the other
243 | parts of the aggregate.
244 |
245 | 6. Conveying Non-Source Forms.
246 |
247 | You may convey a covered work in object code form under the terms
248 | of sections 4 and 5, provided that you also convey the
249 | machine-readable Corresponding Source under the terms of this License,
250 | in one of these ways:
251 |
252 | a) Convey the object code in, or embodied in, a physical product
253 | (including a physical distribution medium), accompanied by the
254 | Corresponding Source fixed on a durable physical medium
255 | customarily used for software interchange.
256 |
257 | b) Convey the object code in, or embodied in, a physical product
258 | (including a physical distribution medium), accompanied by a
259 | written offer, valid for at least three years and valid for as
260 | long as you offer spare parts or customer support for that product
261 | model, to give anyone who possesses the object code either (1) a
262 | copy of the Corresponding Source for all the software in the
263 | product that is covered by this License, on a durable physical
264 | medium customarily used for software interchange, for a price no
265 | more than your reasonable cost of physically performing this
266 | conveying of source, or (2) access to copy the
267 | Corresponding Source from a network server at no charge.
268 |
269 | c) Convey individual copies of the object code with a copy of the
270 | written offer to provide the Corresponding Source. This
271 | alternative is allowed only occasionally and noncommercially, and
272 | only if you received the object code with such an offer, in accord
273 | with subsection 6b.
274 |
275 | d) Convey the object code by offering access from a designated
276 | place (gratis or for a charge), and offer equivalent access to the
277 | Corresponding Source in the same way through the same place at no
278 | further charge. You need not require recipients to copy the
279 | Corresponding Source along with the object code. If the place to
280 | copy the object code is a network server, the Corresponding Source
281 | may be on a different server (operated by you or a third party)
282 | that supports equivalent copying facilities, provided you maintain
283 | clear directions next to the object code saying where to find the
284 | Corresponding Source. Regardless of what server hosts the
285 | Corresponding Source, you remain obligated to ensure that it is
286 | available for as long as needed to satisfy these requirements.
287 |
288 | e) Convey the object code using peer-to-peer transmission, provided
289 | you inform other peers where the object code and Corresponding
290 | Source of the work are being offered to the general public at no
291 | charge under subsection 6d.
292 |
293 | A separable portion of the object code, whose source code is excluded
294 | from the Corresponding Source as a System Library, need not be
295 | included in conveying the object code work.
296 |
297 | A "User Product" is either (1) a "consumer product", which means any
298 | tangible personal property which is normally used for personal, family,
299 | or household purposes, or (2) anything designed or sold for incorporation
300 | into a dwelling. In determining whether a product is a consumer product,
301 | doubtful cases shall be resolved in favor of coverage. For a particular
302 | product received by a particular user, "normally used" refers to a
303 | typical or common use of that class of product, regardless of the status
304 | of the particular user or of the way in which the particular user
305 | actually uses, or expects or is expected to use, the product. A product
306 | is a consumer product regardless of whether the product has substantial
307 | commercial, industrial or non-consumer uses, unless such uses represent
308 | the only significant mode of use of the product.
309 |
310 | "Installation Information" for a User Product means any methods,
311 | procedures, authorization keys, or other information required to install
312 | and execute modified versions of a covered work in that User Product from
313 | a modified version of its Corresponding Source. The information must
314 | suffice to ensure that the continued functioning of the modified object
315 | code is in no case prevented or interfered with solely because
316 | modification has been made.
317 |
318 | If you convey an object code work under this section in, or with, or
319 | specifically for use in, a User Product, and the conveying occurs as
320 | part of a transaction in which the right of possession and use of the
321 | User Product is transferred to the recipient in perpetuity or for a
322 | fixed term (regardless of how the transaction is characterized), the
323 | Corresponding Source conveyed under this section must be accompanied
324 | by the Installation Information. But this requirement does not apply
325 | if neither you nor any third party retains the ability to install
326 | modified object code on the User Product (for example, the work has
327 | been installed in ROM).
328 |
329 | The requirement to provide Installation Information does not include a
330 | requirement to continue to provide support service, warranty, or updates
331 | for a work that has been modified or installed by the recipient, or for
332 | the User Product in which it has been modified or installed. Access to a
333 | network may be denied when the modification itself materially and
334 | adversely affects the operation of the network or violates the rules and
335 | protocols for communication across the network.
336 |
337 | Corresponding Source conveyed, and Installation Information provided,
338 | in accord with this section must be in a format that is publicly
339 | documented (and with an implementation available to the public in
340 | source code form), and must require no special password or key for
341 | unpacking, reading or copying.
342 |
343 | 7. Additional Terms.
344 |
345 | "Additional permissions" are terms that supplement the terms of this
346 | License by making exceptions from one or more of its conditions.
347 | Additional permissions that are applicable to the entire Program shall
348 | be treated as though they were included in this License, to the extent
349 | that they are valid under applicable law. If additional permissions
350 | apply only to part of the Program, that part may be used separately
351 | under those permissions, but the entire Program remains governed by
352 | this License without regard to the additional permissions.
353 |
354 | When you convey a copy of a covered work, you may at your option
355 | remove any additional permissions from that copy, or from any part of
356 | it. (Additional permissions may be written to require their own
357 | removal in certain cases when you modify the work.) You may place
358 | additional permissions on material, added by you to a covered work,
359 | for which you have or can give appropriate copyright permission.
360 |
361 | Notwithstanding any other provision of this License, for material you
362 | add to a covered work, you may (if authorized by the copyright holders of
363 | that material) supplement the terms of this License with terms:
364 |
365 | a) Disclaiming warranty or limiting liability differently from the
366 | terms of sections 15 and 16 of this License; or
367 |
368 | b) Requiring preservation of specified reasonable legal notices or
369 | author attributions in that material or in the Appropriate Legal
370 | Notices displayed by works containing it; or
371 |
372 | c) Prohibiting misrepresentation of the origin of that material, or
373 | requiring that modified versions of such material be marked in
374 | reasonable ways as different from the original version; or
375 |
376 | d) Limiting the use for publicity purposes of names of licensors or
377 | authors of the material; or
378 |
379 | e) Declining to grant rights under trademark law for use of some
380 | trade names, trademarks, or service marks; or
381 |
382 | f) Requiring indemnification of licensors and authors of that
383 | material by anyone who conveys the material (or modified versions of
384 | it) with contractual assumptions of liability to the recipient, for
385 | any liability that these contractual assumptions directly impose on
386 | those licensors and authors.
387 |
388 | All other non-permissive additional terms are considered "further
389 | restrictions" within the meaning of section 10. If the Program as you
390 | received it, or any part of it, contains a notice stating that it is
391 | governed by this License along with a term that is a further
392 | restriction, you may remove that term. If a license document contains
393 | a further restriction but permits relicensing or conveying under this
394 | License, you may add to a covered work material governed by the terms
395 | of that license document, provided that the further restriction does
396 | not survive such relicensing or conveying.
397 |
398 | If you add terms to a covered work in accord with this section, you
399 | must place, in the relevant source files, a statement of the
400 | additional terms that apply to those files, or a notice indicating
401 | where to find the applicable terms.
402 |
403 | Additional terms, permissive or non-permissive, may be stated in the
404 | form of a separately written license, or stated as exceptions;
405 | the above requirements apply either way.
406 |
407 | 8. Termination.
408 |
409 | You may not propagate or modify a covered work except as expressly
410 | provided under this License. Any attempt otherwise to propagate or
411 | modify it is void, and will automatically terminate your rights under
412 | this License (including any patent licenses granted under the third
413 | paragraph of section 11).
414 |
415 | However, if you cease all violation of this License, then your
416 | license from a particular copyright holder is reinstated (a)
417 | provisionally, unless and until the copyright holder explicitly and
418 | finally terminates your license, and (b) permanently, if the copyright
419 | holder fails to notify you of the violation by some reasonable means
420 | prior to 60 days after the cessation.
421 |
422 | Moreover, your license from a particular copyright holder is
423 | reinstated permanently if the copyright holder notifies you of the
424 | violation by some reasonable means, this is the first time you have
425 | received notice of violation of this License (for any work) from that
426 | copyright holder, and you cure the violation prior to 30 days after
427 | your receipt of the notice.
428 |
429 | Termination of your rights under this section does not terminate the
430 | licenses of parties who have received copies or rights from you under
431 | this License. If your rights have been terminated and not permanently
432 | reinstated, you do not qualify to receive new licenses for the same
433 | material under section 10.
434 |
435 | 9. Acceptance Not Required for Having Copies.
436 |
437 | You are not required to accept this License in order to receive or
438 | run a copy of the Program. Ancillary propagation of a covered work
439 | occurring solely as a consequence of using peer-to-peer transmission
440 | to receive a copy likewise does not require acceptance. However,
441 | nothing other than this License grants you permission to propagate or
442 | modify any covered work. These actions infringe copyright if you do
443 | not accept this License. Therefore, by modifying or propagating a
444 | covered work, you indicate your acceptance of this License to do so.
445 |
446 | 10. Automatic Licensing of Downstream Recipients.
447 |
448 | Each time you convey a covered work, the recipient automatically
449 | receives a license from the original licensors, to run, modify and
450 | propagate that work, subject to this License. You are not responsible
451 | for enforcing compliance by third parties with this License.
452 |
453 | An "entity transaction" is a transaction transferring control of an
454 | organization, or substantially all assets of one, or subdividing an
455 | organization, or merging organizations. If propagation of a covered
456 | work results from an entity transaction, each party to that
457 | transaction who receives a copy of the work also receives whatever
458 | licenses to the work the party's predecessor in interest had or could
459 | give under the previous paragraph, plus a right to possession of the
460 | Corresponding Source of the work from the predecessor in interest, if
461 | the predecessor has it or can get it with reasonable efforts.
462 |
463 | You may not impose any further restrictions on the exercise of the
464 | rights granted or affirmed under this License. For example, you may
465 | not impose a license fee, royalty, or other charge for exercise of
466 | rights granted under this License, and you may not initiate litigation
467 | (including a cross-claim or counterclaim in a lawsuit) alleging that
468 | any patent claim is infringed by making, using, selling, offering for
469 | sale, or importing the Program or any portion of it.
470 |
471 | 11. Patents.
472 |
473 | A "contributor" is a copyright holder who authorizes use under this
474 | License of the Program or a work on which the Program is based. The
475 | work thus licensed is called the contributor's "contributor version".
476 |
477 | A contributor's "essential patent claims" are all patent claims
478 | owned or controlled by the contributor, whether already acquired or
479 | hereafter acquired, that would be infringed by some manner, permitted
480 | by this License, of making, using, or selling its contributor version,
481 | but do not include claims that would be infringed only as a
482 | consequence of further modification of the contributor version. For
483 | purposes of this definition, "control" includes the right to grant
484 | patent sublicenses in a manner consistent with the requirements of
485 | this License.
486 |
487 | Each contributor grants you a non-exclusive, worldwide, royalty-free
488 | patent license under the contributor's essential patent claims, to
489 | make, use, sell, offer for sale, import and otherwise run, modify and
490 | propagate the contents of its contributor version.
491 |
492 | In the following three paragraphs, a "patent license" is any express
493 | agreement or commitment, however denominated, not to enforce a patent
494 | (such as an express permission to practice a patent or covenant not to
495 | sue for patent infringement). To "grant" such a patent license to a
496 | party means to make such an agreement or commitment not to enforce a
497 | patent against the party.
498 |
499 | If you convey a covered work, knowingly relying on a patent license,
500 | and the Corresponding Source of the work is not available for anyone
501 | to copy, free of charge and under the terms of this License, through a
502 | publicly available network server or other readily accessible means,
503 | then you must either (1) cause the Corresponding Source to be so
504 | available, or (2) arrange to deprive yourself of the benefit of the
505 | patent license for this particular work, or (3) arrange, in a manner
506 | consistent with the requirements of this License, to extend the patent
507 | license to downstream recipients. "Knowingly relying" means you have
508 | actual knowledge that, but for the patent license, your conveying the
509 | covered work in a country, or your recipient's use of the covered work
510 | in a country, would infringe one or more identifiable patents in that
511 | country that you have reason to believe are valid.
512 |
513 | If, pursuant to or in connection with a single transaction or
514 | arrangement, you convey, or propagate by procuring conveyance of, a
515 | covered work, and grant a patent license to some of the parties
516 | receiving the covered work authorizing them to use, propagate, modify
517 | or convey a specific copy of the covered work, then the patent license
518 | you grant is automatically extended to all recipients of the covered
519 | work and works based on it.
520 |
521 | A patent license is "discriminatory" if it does not include within
522 | the scope of its coverage, prohibits the exercise of, or is
523 | conditioned on the non-exercise of one or more of the rights that are
524 | specifically granted under this License. You may not convey a covered
525 | work if you are a party to an arrangement with a third party that is
526 | in the business of distributing software, under which you make payment
527 | to the third party based on the extent of your activity of conveying
528 | the work, and under which the third party grants, to any of the
529 | parties who would receive the covered work from you, a discriminatory
530 | patent license (a) in connection with copies of the covered work
531 | conveyed by you (or copies made from those copies), or (b) primarily
532 | for and in connection with specific products or compilations that
533 | contain the covered work, unless you entered into that arrangement,
534 | or that patent license was granted, prior to 28 March 2007.
535 |
536 | Nothing in this License shall be construed as excluding or limiting
537 | any implied license or other defenses to infringement that may
538 | otherwise be available to you under applicable patent law.
539 |
540 | 12. No Surrender of Others' Freedom.
541 |
542 | If conditions are imposed on you (whether by court order, agreement or
543 | otherwise) that contradict the conditions of this License, they do not
544 | excuse you from the conditions of this License. If you cannot convey a
545 | covered work so as to satisfy simultaneously your obligations under this
546 | License and any other pertinent obligations, then as a consequence you may
547 | not convey it at all. For example, if you agree to terms that obligate you
548 | to collect a royalty for further conveying from those to whom you convey
549 | the Program, the only way you could satisfy both those terms and this
550 | License would be to refrain entirely from conveying the Program.
551 |
552 | 13. Use with the GNU Affero General Public License.
553 |
554 | Notwithstanding any other provision of this License, you have
555 | permission to link or combine any covered work with a work licensed
556 | under version 3 of the GNU Affero General Public License into a single
557 | combined work, and to convey the resulting work. The terms of this
558 | License will continue to apply to the part which is the covered work,
559 | but the special requirements of the GNU Affero General Public License,
560 | section 13, concerning interaction through a network will apply to the
561 | combination as such.
562 |
563 | 14. Revised Versions of this License.
564 |
565 | The Free Software Foundation may publish revised and/or new versions of
566 | the GNU General Public License from time to time. Such new versions will
567 | be similar in spirit to the present version, but may differ in detail to
568 | address new problems or concerns.
569 |
570 | Each version is given a distinguishing version number. If the
571 | Program specifies that a certain numbered version of the GNU General
572 | Public License "or any later version" applies to it, you have the
573 | option of following the terms and conditions either of that numbered
574 | version or of any later version published by the Free Software
575 | Foundation. If the Program does not specify a version number of the
576 | GNU General Public License, you may choose any version ever published
577 | by the Free Software Foundation.
578 |
579 | If the Program specifies that a proxy can decide which future
580 | versions of the GNU General Public License can be used, that proxy's
581 | public statement of acceptance of a version permanently authorizes you
582 | to choose that version for the Program.
583 |
584 | Later license versions may give you additional or different
585 | permissions. However, no additional obligations are imposed on any
586 | author or copyright holder as a result of your choosing to follow a
587 | later version.
588 |
589 | 15. Disclaimer of Warranty.
590 |
591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599 |
600 | 16. Limitation of Liability.
601 |
602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610 | SUCH DAMAGES.
611 |
612 | 17. Interpretation of Sections 15 and 16.
613 |
614 | If the disclaimer of warranty and limitation of liability provided
615 | above cannot be given local legal effect according to their terms,
616 | reviewing courts shall apply local law that most closely approximates
617 | an absolute waiver of all civil liability in connection with the
618 | Program, unless a warranty or assumption of liability accompanies a
619 | copy of the Program in return for a fee.
620 |
621 | END OF TERMS AND CONDITIONS
622 |
623 | How to Apply These Terms to Your New Programs
624 |
625 | If you develop a new program, and you want it to be of the greatest
626 | possible use to the public, the best way to achieve this is to make it
627 | free software which everyone can redistribute and change under these terms.
628 |
629 | To do so, attach the following notices to the program. It is safest
630 | to attach them to the start of each source file to most effectively
631 | state the exclusion of warranty; and each file should have at least
632 | the "copyright" line and a pointer to where the full notice is found.
633 |
634 | {one line to give the program's name and a brief idea of what it does.}
635 | Copyright (C) {year} {name of author}
636 |
637 | This program is free software: you can redistribute it and/or modify
638 | it under the terms of the GNU General Public License as published by
639 | the Free Software Foundation, either version 3 of the License, or
640 | (at your option) any later version.
641 |
642 | This program is distributed in the hope that it will be useful,
643 | but WITHOUT ANY WARRANTY; without even the implied warranty of
644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645 | GNU General Public License for more details.
646 |
647 | You should have received a copy of the GNU General Public License
648 | along with this program. If not, see .
649 |
650 | Also add information on how to contact you by electronic and paper mail.
651 |
652 | If the program does terminal interaction, make it output a short
653 | notice like this when it starts in an interactive mode:
654 |
655 | {project} Copyright (C) {year} {fullname}
656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657 | This is free software, and you are welcome to redistribute it
658 | under certain conditions; type `show c' for details.
659 |
660 | The hypothetical commands `show w' and `show c' should show the appropriate
661 | parts of the General Public License. Of course, your program's commands
662 | might be different; for a GUI interface, you would use an "about box".
663 |
664 | You should also get your employer (if you work as a programmer) or school,
665 | if any, to sign a "copyright disclaimer" for the program, if necessary.
666 | For more information on this, and how to apply and follow the GNU GPL, see
667 | .
668 |
669 | The GNU General Public License does not permit incorporating your program
670 | into proprietary programs. If your program is a subroutine library, you
671 | may consider it more useful to permit linking proprietary applications with
672 | the library. If this is what you want to do, use the GNU Lesser General
673 | Public License instead of this License. But first, please read
674 | .
675 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | This is the Tensorflow implementation of TAC-GAN model presented in
2 | [https://arxiv.org/abs/1703.06412](https://arxiv.org/abs/1703.06412).
3 |
4 | Text Conditioned Auxiliary Classifier Generative Adversarial Network,
5 | (TAC-GAN) is a text to image Generative Adversarial Network (GAN) for
6 | synthesizing images from their text descriptions. TAC-GAN builds upon the
7 | [AC-GAN](https://arxiv.org/abs/1610.09585) by conditioning the generated images
8 | on a text description instead of on a class label. In the presented TAC-GAN
9 | model, the input vector of the Generative network is built based on a noise
10 | vector and another vector containing an embedded representation of the
11 | textual description. While the Discriminator is similar to that of
12 | the AC-GAN, it is also augmented to receive the text information as
13 | input before performing its classification.
14 |
15 | For embedding the textual descriptions of the images into vectors we used
16 | [skip-thought vectors](https://arxiv.org/abs/1506.06726)
17 |
18 | The following is the architecture of the TAC-GAN model
19 |
20 |
22 |
23 | # Prerequisites
24 | Some important dependencies are the following and the rest can be installed
25 | using the ```requirements.txt```
26 | 1. Python 3.5
27 | 2. [Tensorflow 1.2.0](https://github.com/tensorflow/tensorflow)
28 | 4. [Theano 0.9.0](https://github.com/Theano/Theano) : for skip thought vectors
29 | 5. [scikit-learn](http://scikit-learn.org/stable/index.html) : for skip thought vectors
30 | 6. [NLTK 3.2.1](http://www.nltk.org/) : for skip thought vectors
31 |
32 | It is recommended to use a virtual environment for running this project and
33 | installing the required dependencies in it by using the
34 | [***requirements.txt***](https://github.com/dashayushman/TAC-GAN/blob/master/requirements.txt) file.
35 |
36 | The project has been tested on a Ubuntu 14.04 machine with an 12 GB NVIDIA
37 | Titen X GPU
38 |
39 | # 1. Setup and Run
40 |
41 | ## 1.1. Clone the Repository
42 |
43 | ```
44 | git clone https://github.com/dashayushman/TAC-GAN.git
45 | cd TAC-GAN
46 | ```
47 |
48 | ## 1.2. Download the Dataset
49 |
50 | The model presented in the paper was trained on the
51 | [flowers dataset](http://www.robots.ox.ac.uk/~vgg/data/flowers/102/ ). This
52 | To train the TAC-GAN on the flowers dataset, first, download the dataset by
53 | doing the following,
54 |
55 | 1. **Download the flower images** from
56 | [here](http://www.robots.ox.ac.uk/~vgg/data/flowers/102/102flowers.tgz).
57 | Extract the ```102flowers.tgz``` file and copy the extracted ```jpg``` folder
58 | to ```Data/datasets/flowers```
59 |
60 | 2. **Download the captions** from
61 | [here](https://drive.google.com/file/d/0B0ywwgffWnLLcms2WWJQRFNSWXM/).
62 | Extract the downloaded file, copy the text_c10 folder and paste it in ```
63 | Data/datasets/flowers``` directory
64 |
65 | 3. **Download the pretrained skip-thought vectors model** from
66 | [here](https://github.com/ryankiros/skip-thoughts#getting-started) and copy
67 | the downloaded files to ```Data/skipthoughts```
68 |
69 | **NB:** *It is recommended to keep all the images in an SSD if available. This
70 | makes the batch loading and processing operation faster.*
71 |
72 | ## 1.3. Data Preprocessing
73 | Extract the skip-thought features for the captions and prepare the dataset
74 | for training by running the following script
75 |
76 | ```
77 | python dataprep.py --data_dir=Data --dataset=flowers
78 | ```
79 |
80 | This script will create a set of pickled files in the datet directory which
81 | will be used during training. The following are the available flags for data preparation:
82 |
83 | FLAG | VALUE TYPE | DEFAULT VALUE | DESCRIPTION
84 | --- | --- | --- | ---
85 | data_dir | str | Data | The data directory |
86 | dataset | str | flowers | Dataset to use. For Eg., "flowers" |
87 |
88 | ## 1.4. Training
89 |
90 | To train TAC-GAN with the default hyper parameters run the following script
91 |
92 | ```
93 | python train.py --dataset="flowers" --model_name=TAC-GAN
94 | ```
95 |
96 | While training, you can montor samples generated by the model in the ```Data/training/TAC_GAN/samples``` directory. Notice that a directory is created according to the ***"model_name"*** taht you provide. This directory contains all the data related to a particular experiment. This can also be considered as an ***"experiment name"*** too.
97 |
98 | The following flags can be set to change the hyperparameters of the network.
99 |
100 | FLAG | VALUE TYPE | DEFAULT VALUE | DESCRIPTION
101 | --- | --- | --- | ---
102 | z-dim | int | 100 | Number of dimensions of the Noise vector |
103 | t_dim | int | 256 | Number of dimensions for the latent representation of the text embedding.
104 | batch_size | int | 64 | Mini-Batch Size
105 | image_size | int | 128 | Batch size to use during training.
106 | gf_dim | int | 64 | Number of conv filters in the first layer of the generator.
107 | df_dim | int | 64 | Number of conv filters in the first layer of the discriminator.
108 | caption_vector_length | int | 4800 | Length of the caption vector embedding (vector generated using skip-thought vectors model).
109 | n_classes | int | 102 | Number of classes
110 | data_dir | String | Data | Data directory
111 | learning_rate | float | 0.0002 | Learning rate
112 | beta1 | float | 0.5 | Momentum for Adam Update
113 | epochs | int | 200 | Maximum number of epochs to train
114 | save_every | int | 30 | Save model and samples after this many number.of iterations
115 | resume_model | Boolean | False | To Load the pre-trained model
116 | data_set | String | flowers | Which dataset to use: "flowers"
117 | model_name | String | model_1 | Name of the model: Can be anything
118 | train | bool | True | This is True while training and false otherwise. Used for batch normalization
119 |
120 | We used the following script (hyper-parameters) in for the results that we show in our paper
121 |
122 | ```
123 | python train.py --t_dim=100 --image_size=128 --data_set=flowers --model_name=TAC_GAN --train=True --resume_model=True --z_dim=100 --n_classes=102 --epochs=400 --save_every=20 --caption_vector_length=4800 --batch_size=128
124 | ```
125 |
126 | ## 1.5. Monitoring
127 |
128 | While training, you can monitor the updates on the *terminal* as well as by using [*tensorboard*](https://www.tensorflow.org/get_started/summaries_and_tensorboard)
129 |
130 | ### 1.5.1 The Terminal:
131 | 
132 |
133 | ### 1.5.1 Tensorboard:
134 |
135 | You can use the following script to start tensorboard and visualize realtime changes:
136 |
137 | ```
138 | tensorboard --logdir=Data/training/TAC_GAN/summaries
139 | ```
140 |
141 | 
142 |
143 | # 2. Generating Images for the text in the dataset
144 |
145 | Once you have trained the model for certain epochs you can generate images for all the text descriptions in the dataset use the following script. This will create a synthetic dataset with images generated by the generator.
146 |
147 | ```
148 | python train.py --data_set=flowers --epochs=100 --output_dir=Data/synthetic_dataset --checkpoints_dir=Data/training/TAC_GAN/checkpoints
149 | ```
150 |
151 | Notice that the ***checkpoints*** directory is ls created automatically created inside the ***model*** directory after you run the training script.
152 |
153 | This script will create the following directory structure:
154 |
155 | ```
156 | Data
157 | |__synthetic_dataset
158 | |___ds
159 | |___train
160 | |___val
161 | ```
162 |
163 | the ***train*** directory will contain all the images generated from the text descriptions of the images in the training set and the same goes for the ***val*** directory.
164 |
165 | # 3. Generating Images from any Text
166 |
167 | To generate images from any text, do the following
168 |
169 | ## 3.1 Add Text Descriptions:
170 |
171 | Write your text descriptions in a file or use the example file ```Data/text.txt``` that we have provided in the Data directory.
172 | The text description file should contain one text description per line. For example,
173 |
174 | ```
175 | a flower with red petals which are pointed
176 | many pointed petals
177 | A yellow flower
178 | ```
179 |
180 | ## 3.2 Extract Skip-Thought Vectors:
181 |
182 | Run the following script for extracting the Skip-Thought vectors for the text descriptions
183 |
184 | ```
185 | python encode_text.py --caption_file=Data/text.txt --data_dir=Data
186 | ```
187 |
188 | This script will create a pickle file called ```Data/enc_text.pkl``` with features extracted from the text descriptions.
189 |
190 | ## 3.3 Generate Images:
191 |
192 | To generate images for the text descriptions, run the following script,
193 |
194 | ```
195 | python generate_images.py --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --images_per_caption=30 --data_dir=Data
196 | ```
197 |
198 | This will create a directory ```Data/images_generated_from_text/``` with a folder corresponding to every row of the ***text.txt*** file. Each of these folders will contain images for that text.
199 |
200 | The following are the parameters you need to set, in case you have used different parameters for training the model.
201 |
202 | FLAG | VALUE TYPE | DEFAULT VALUE | DESCRIPTION
203 | --- | --- | --- | ---
204 | z-dim | int | 100 | Number of dimensions of the Noise vector |
205 | t_dim | int | 256 | Number of dimensions for the latent representation of the text embedding.
206 | batch_size | int | 64 | Mini-Batch Size
207 | image_size | int | 128 | Batch size to use during training.
208 | gf_dim | int | 64 | Number of conv filters in the first layer of the generator.
209 | df_dim | int | 64 | Number of conv filters in the first layer of the discriminator.
210 | caption_vector_length | int | 4800 | Length of the caption vector embedding (vector generated using skip-thought vectors model).
211 | n_classes | int | 102 | Number of classes
212 | data_dir | String | Data | Data directory
213 | learning_rate | float | 0.0002 | Learning rate
214 | beta1 | float | 0.5 | Momentum for Adam Update
215 | images_per_caption | int | 30 | Maximum number of images that you want to generate for each of the text descriptions
216 | data_set | String | flowers | Which dataset to use: "flowers"
217 | checkpoints_dir | String | /tmp | Path to the checkpoints directory which will be used to generate the images
218 |
219 | # 4. Evaluation
220 |
221 | We have used two metrics for evaluating TAC-GAN,
222 |
223 | 1. [Inception-Scope](https://github.com/openai/improved-gan/tree/master/inception_score)
224 | 2. [MS-SSIM score](https://github.com/tensorflow/models/blob/master/compression/image_encoder/msssim.py)
225 |
226 | The links are from where we adapted the code for evaluating TAC-GAN. Before evaluating the model, generate a synthetic dataset by referring to [**Section 6**](#6-generating-images-for-text-in-the-dataset)
227 |
228 | ## 4.1 Inception Score
229 |
230 | To calculate the inception score, use the following script,
231 |
232 | ```
233 | python inception_score.py --output_dir=Data/synthetic_dataset --data_dir=Data --n_images=30000 --image_size=128
234 | ```
235 |
236 | This will create a collection of all the generated images in ```Data/synthetic_dataset/ds_inception``` and show the inception score on the terminal.
237 |
238 | The following are the set of available parameters/flags
239 |
240 | FLAG | VALUE TYPE | DEFAULT VALUE | DESCRIPTION
241 | --- | --- | --- | ---
242 | output_dir | str | Data/ds_inception | Directory to dump all the images for calculating the inception score |
243 | data_dir | str | Data/synthetic_dataset/ds | The root directory of the synthetic dataset |
244 | n_images | int | 30000 | Number of images to consider for calculating inception score |
245 | image_size | int | 128 | Size of the image to consider for calculating inception score |
246 |
247 |
248 | ## 4.2 MS-SSIM
249 |
250 | To calculate the MS-SSIM score, use the following script,
251 |
252 | ```
253 | python inception_score.py --output_dir=Data --data_dir=Data --dataset=flowers --syn_dataset_dir=Data/synthetic_datset/ds
254 | ```
255 |
256 | This will create a ```Data/msssim.tsv``` tab separated file. The data in this file is structured as follows
257 |
258 | ```
259 |
260 | ```
261 |
262 | Once you have generated the ***msssim.tsv*** file, you can use the following script to generate a figure to compare the MS-SSIM score of the images in the real dataset with the images in the synthetic dataset belonging to the same class,
263 |
264 | ```
265 | python utility/plot_msssim.py --input_file=Data/msssim.tsv --output_file=Data/msssim
266 | ```
267 |
268 | This will create ```Data/msssim.pdf```, which is the ***.pdf*** file of the generated figure.
269 |
270 | # 5. Generate Interpolated Images
271 |
272 | In our paper we show the effect of interpolating the noise and the text embedding vectors on the generated image. Images are randomply selected and their text descriptions are used to generate synthetic images. The following sub-sections will elaborate on how to do it and which scripts will help you in doing it.
273 |
274 | ## 5.1 Z (Noise) Interpolation
275 |
276 | For interpolating the noise vector and generating images, use the following scripts
277 |
278 | ```
279 | python z_interpolation.py --output_dir=Data/synthetic_dataset --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --n_images=500
280 | ```
281 |
282 | This will generate the interpolated images in ```Data/synthetic_dataset/z_interpolation/```.
283 |
284 | ## 5.1 T (Text Embedding) Interpolation
285 |
286 | For interpolating the text embedding vectors and generating images, use the following scripts
287 |
288 | ```
289 | python t_interpolation.py --output_dir=Data/synthetic_dataset --data_set=flowers --checkpoints_dir=Data/training/TAC_GAN/checkpoints --n_images=500
290 | ```
291 |
292 | This will generate the interpolated images in ```Data/synthetic_dataset/t_interpolation/```.
293 |
294 | ***NOTE:*** Both the above mentioned scripts have the same flags/arguments, which are the following,
295 |
296 | FLAG | VALUE TYPE | DEFAULT VALUE | DESCRIPTION
297 | --- | --- | --- | ---
298 | z-dim | int | 100 | Number of dimensions of the Noise vector |
299 | t_dim | int | 256 | Number of dimensions for the latent representation of the text embedding.
300 | batch_size | int | 64 | Mini-Batch Size
301 | image_size | int | 128 | Batch size to use during training.
302 | gf_dim | int | 64 | Number of conv filters in the first layer of the generator.
303 | df_dim | int | 64 | Number of conv filters in the first layer of the discriminator.
304 | caption_vector_length | int | 4800 | Length of the caption vector embedding (vector generated using skip-thought vectors model).
305 | n_classes | int | 102 | Number of classes
306 | data_dir | String | Data | Data directory
307 | learning_rate | float | 0.0002 | Learning rate
308 | beta1 | float | 0.5 | Momentum for Adam Update
309 | data_set | str | flowers | The dataset to use: "flowers"
310 | output_dir | String | Data/synthetic_dataset | The directory in which the t_interpolated images will be generated
311 | checkpoints_dir | String | /tmp | Path to the checkpoints directory which will be used to generate the images
312 | n_interp | int | 100 | The factor difference between each interpolation (Should ideally be a multiple of 10)
313 | n_images | int | 500 | Number of images to randomply sample for generating interpolation results
314 |
315 | # 6. References
316 |
317 | ### TAC-GAN
318 |
319 | If you find this code usefull, then please use the following BibTex to cite our work.
320 |
321 | ```
322 | @article{dash2017tac,
323 | title={TAC-GAN-Text Conditioned Auxiliary Classifier Generative Adversarial Network},
324 | author={Dash, Ayushman and Gamboa, John Cristian Borges and Ahmed, Sheraz and Afzal, Muhammad Zeshan and Liwicki, Marcus},
325 | journal={arXiv preprint arXiv:1703.06412},
326 | year={2017}
327 | }
328 | ```
329 |
330 | ### Oxford-102 Flowers Dataset
331 |
332 | If you use the Oxford-102 Flowers Dataset, then please cite their work using the following BibTex.
333 |
334 | ```
335 | @InProceedings{Nilsback08,
336 | author = "Nilsback, M-E. and Zisserman, A.",
337 | title = "Automated Flower Classification over a Large Number of Classes",
338 | booktitle = "Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing",
339 | year = "2008",
340 | month = "Dec"
341 | }
342 | ```
343 |
344 | ### Skip-Thought
345 |
346 | If you use the Skip-Thought model in your work like us, then please cite their work using the following BibTex
347 |
348 | ```
349 | @article{kiros2015skip,
350 | title={Skip-Thought Vectors},
351 | author={Kiros, Ryan and Zhu, Yukun and Salakhutdinov, Ruslan and Zemel, Richard S and Torralba, Antonio and Urtasun, Raquel and Fidler, Sanja},
352 | journal={arXiv preprint arXiv:1506.06726},
353 | year={2015}
354 | }
355 | ```
356 |
357 | ### Code
358 |
359 | We have referred to the [text-to-image](https://github.com/paarthneekhara/text-to-image) and [DCGAN-tensorflow](https://github.com/carpedm20/DCGAN-tensorflow) repositories for developing our code, and we are extremely thankful to them.
360 |
--------------------------------------------------------------------------------
/Utils/image_processing.py:
--------------------------------------------------------------------------------
1 | import numpy as np
2 | from scipy import misc
3 | import random
4 | import skimage
5 | import skimage.io
6 | import skimage.transform
7 | import os
8 |
9 | def load_image_array_flowers(image_file, image_size):
10 | img = skimage.io.imread(image_file)
11 | # GRAYSCALE
12 | if len(img.shape) == 2:
13 | img_new = np.ndarray( (img.shape[0], img.shape[1], 3), dtype = 'uint8')
14 | img_new[:,:,0] = img
15 | img_new[:,:,1] = img
16 | img_new[:,:,2] = img
17 | img = img_new
18 |
19 | img_resized = skimage.transform.resize(img, (image_size, image_size))
20 |
21 | # FLIP HORIZONTAL WIRH A PROBABILITY 0.5
22 | if random.random() > 0.5:
23 | img_resized = np.fliplr(img_resized)
24 |
25 |
26 | return img_resized.astype('float32')
27 |
28 | def load_image_array(image_file, image_size,
29 | image_id, data_dir='Data/datasets/mscoco/train2014',
30 | mode='train'):
31 | img = None
32 | if os.path.exists(image_file):
33 | #print('found' + image_file)
34 | img = skimage.io.imread(image_file)
35 | else:
36 | print('notfound' + image_file)
37 | img = skimage.io.imread('http://mscoco.org/images/%d' % (image_id))
38 | img_path = os.path.join(data_dir, 'COCO_%s2014_%.12d.jpg' % ( mode,
39 | image_id))
40 | skimage.io.imsave(img_path, img)
41 |
42 | # GRAYSCALE
43 | if len(img.shape) == 2:
44 | img_new = np.ndarray( (img.shape[0], img.shape[1], 3), dtype = 'uint8')
45 | img_new[:,:,0] = img
46 | img_new[:,:,1] = img
47 | img_new[:,:,2] = img
48 | img = img_new
49 |
50 | img_resized = skimage.transform.resize(img, (image_size, image_size))
51 |
52 | # FLIP HORIZONTAL WIRH A PROBABILITY 0.5
53 | if random.random() > 0.5:
54 | img_resized = np.fliplr(img_resized)
55 |
56 | return img_resized.astype('float32')
57 |
58 | def load_image_inception(image_file, image_size=128):
59 | img = skimage.io.imread(image_file)
60 | # GRAYSCALE
61 | if len(img.shape) == 2:
62 | img_new = np.ndarray((img.shape[0], img.shape[1], 3), dtype='uint8')
63 | img_new[:, :, 0] = img
64 | img_new[:, :, 1] = img
65 | img_new[:, :, 2] = img
66 | img = img_new
67 |
68 | if image_size != 0:
69 | img = skimage.transform.resize(img, (image_size, image_size), mode='reflect')
70 |
71 | return img.astype('int32')
72 |
73 | if __name__ == '__main__':
74 | # TEST>>>
75 | arr = load_image_array('sample.jpg', 64)
76 | print(arr.mean())
77 | # rev = np.fliplr(arr)
78 | misc.imsave( 'rev.jpg', arr)
79 |
--------------------------------------------------------------------------------
/Utils/inception_score.py:
--------------------------------------------------------------------------------
1 | # Code derived from tensorflow/tensorflow/models/image/imagenet/classify_image.py
2 | from __future__ import absolute_import
3 | from __future__ import division
4 | from __future__ import print_function
5 |
6 | import os.path
7 | import sys
8 | import tarfile
9 |
10 | import numpy as np
11 | from six.moves import urllib
12 | import tensorflow as tf
13 | import glob
14 | import scipy.misc
15 | import math
16 | import sys
17 |
18 | MODEL_DIR = '/tmp/imagenet'
19 | DATA_URL = 'http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz'
20 | softmax = None
21 |
22 | # Call this function with list of images. Each of elements should be a
23 | # numpy array with values ranging from 0 to 255.
24 | def get_inception_score(images, splits=10):
25 | assert(type(images) == list)
26 | assert(type(images[0]) == np.ndarray)
27 | assert(len(images[0].shape) == 3)
28 | assert(np.max(images[0]) > 10)
29 | assert(np.min(images[0]) >= 0.0)
30 | inps = []
31 | for img in images:
32 | img = img.astype(np.float32)
33 | inps.append(np.expand_dims(img, 0))
34 | bs = 100
35 | with tf.Session() as sess:
36 | preds = []
37 | n_batches = int(math.ceil(float(len(inps)) / float(bs)))
38 | for i in range(n_batches):
39 | sys.stdout.write(".")
40 | sys.stdout.flush()
41 | inp = inps[(i * bs):min((i + 1) * bs, len(inps))]
42 | inp = np.concatenate(inp, 0)
43 | pred = sess.run(softmax, {'ExpandDims:0': inp})
44 | preds.append(pred)
45 | preds = np.concatenate(preds, 0)
46 | scores = []
47 | for i in range(splits):
48 | part = preds[(i * preds.shape[0] // splits):((i + 1) * preds.shape[0] // splits), :]
49 | kl = part * (np.log(part) - np.log(np.expand_dims(np.mean(part, 0), 0)))
50 | kl = np.mean(np.sum(kl, 1))
51 | scores.append(np.exp(kl))
52 | return np.mean(scores), np.std(scores)
53 |
54 | # This function is called automatically.
55 | def _init_inception():
56 | global softmax
57 | if not os.path.exists(MODEL_DIR):
58 | os.makedirs(MODEL_DIR)
59 | filename = DATA_URL.split('/')[-1]
60 | filepath = os.path.join(MODEL_DIR, filename)
61 | if not os.path.exists(filepath):
62 | def _progress(count, block_size, total_size):
63 | sys.stdout.write('\r>> Downloading %s %.1f%%' % (
64 | filename, float(count * block_size) / float(total_size) * 100.0))
65 | sys.stdout.flush()
66 | filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, _progress)
67 | print()
68 | statinfo = os.stat(filepath)
69 | print('Succesfully downloaded', filename, statinfo.st_size, 'bytes.')
70 | tarfile.open(filepath, 'r:gz').extractall(MODEL_DIR)
71 | with tf.gfile.FastGFile(os.path.join(
72 | MODEL_DIR, 'classify_image_graph_def.pb'), 'rb') as f:
73 | graph_def = tf.GraphDef()
74 | graph_def.ParseFromString(f.read())
75 | _ = tf.import_graph_def(graph_def, name='')
76 | # Works with an arbitrary minibatch size.
77 | with tf.Session() as sess:
78 | pool3 = sess.graph.get_tensor_by_name('pool_3:0')
79 | ops = pool3.graph.get_operations()
80 | for op_idx, op in enumerate(ops):
81 | for o in op.outputs:
82 | shape = o.get_shape()
83 | shape = [s.value for s in shape]
84 | new_shape = []
85 | for j, s in enumerate(shape):
86 | if s == 1 and j == 0:
87 | new_shape.append(None)
88 | else:
89 | new_shape.append(s)
90 | o._shape = tf.TensorShape(new_shape)
91 | w = sess.graph.get_operation_by_name("softmax/logits/MatMul").inputs[1]
92 | logits = tf.matmul(tf.squeeze(pool3), w)
93 | softmax = tf.nn.softmax(logits)
94 |
95 | if softmax is None:
96 | _init_inception()
97 |
--------------------------------------------------------------------------------
/Utils/ops.py:
--------------------------------------------------------------------------------
1 | # RESUED CODE FROM https://github.com/carpedm20/DCGAN-tensorflow/blob/master/ops.py
2 | import math
3 | import numpy as np
4 | import tensorflow as tf
5 |
6 | from tensorflow.python.framework import ops
7 |
8 |
9 | class batch_norm(object):
10 | """Code modification of http://stackoverflow.com/a/33950177"""
11 | def __init__(self, epsilon=1e-5, momentum = 0.9, name="batch_norm"):
12 | with tf.variable_scope(name):
13 | self.epsilon = epsilon
14 | self.momentum = momentum
15 |
16 | self.ema = tf.train.ExponentialMovingAverage(decay=self.momentum)
17 | self.name = name
18 |
19 | def __call__(self, x, train=True):
20 | shape = x.get_shape().as_list()
21 |
22 | if train:
23 | with tf.variable_scope(self.name) as scope:
24 | self.beta = tf.get_variable("beta", [shape[-1]],
25 | initializer=tf.constant_initializer(0.))
26 | self.gamma = tf.get_variable("gamma", [shape[-1]],
27 | initializer=tf.random_normal_initializer(1., 0.02))
28 |
29 | try:
30 | batch_mean, batch_var = tf.nn.moments(x, [0, 1, 2], name='moments')
31 | except:
32 | batch_mean, batch_var = tf.nn.moments(x, [0, 1], name='moments')
33 |
34 | ema_apply_op = self.ema.apply([batch_mean, batch_var])
35 | self.ema_mean, self.ema_var = self.ema.average(batch_mean), self.ema.average(batch_var)
36 |
37 | with tf.control_dependencies([ema_apply_op]):
38 | mean, var = tf.identity(batch_mean), tf.identity(batch_var)
39 | else:
40 | mean, var = self.ema_mean, self.ema_var
41 |
42 | normed = tf.nn.batch_norm_with_global_normalization(
43 | x, mean, var, self.beta, self.gamma, self.epsilon, scale_after_normalization=True)
44 |
45 | return normed
46 |
47 |
48 |
49 | def binary_cross_entropy(preds, targets, name=None):
50 | """Computes binary cross entropy given `preds`.
51 |
52 | For brevity, let `x = `, `z = targets`. The logistic loss is
53 |
54 | loss(x, z) = - sum_i (x[i] * log(z[i]) + (1 - x[i]) * log(1 - z[i]))
55 |
56 | Args:
57 | preds: A `Tensor` of type `float32` or `float64`.
58 | targets: A `Tensor` of the same type and shape as `preds`.
59 | """
60 | eps = 1e-12
61 | with ops.op_scope([preds, targets], name, "bce_loss") as name:
62 | preds = ops.convert_to_tensor(preds, name="preds")
63 | targets = ops.convert_to_tensor(targets, name="targets")
64 | return tf.reduce_mean(-(targets * tf.log(preds + eps) +
65 | (1. - targets) * tf.log(1. - preds + eps)))
66 |
67 | def conv_cond_concat(x, y):
68 | """Concatenate conditioning vector on feature map axis."""
69 | x_shapes = x.get_shape()
70 | y_shapes = y.get_shape()
71 | return tf.concat(3, [x, y*tf.ones([x_shapes[0], x_shapes[1], x_shapes[2], y_shapes[3]])])
72 |
73 | def conv2d(input_, output_dim, k_h=5, k_w=5, d_h=2, d_w=2, stddev=0.02,
74 | name="conv2d"):
75 | with tf.variable_scope(name):
76 | #w = tf.get_variable('w', [k_h, k_w, input_.get_shape()[-1], output_dim], initializer=tf.truncated_normal_initializer(stddev=stddev))
77 | w = tf.get_variable('w', [k_h, k_w, input_.get_shape()[-1], output_dim], initializer = tf.contrib.layers.xavier_initializer())
78 | conv = tf.nn.conv2d(input_, w, strides=[1, d_h, d_w, 1], padding='SAME')
79 |
80 | biases = tf.get_variable('biases', [output_dim], initializer=tf.constant_initializer(0.0))
81 | conv = tf.reshape(tf.nn.bias_add(conv, biases), conv.get_shape())
82 |
83 | return conv
84 |
85 | def deconv2d(input_, output_shape, k_h=5, k_w=5, d_h=2, d_w=2, stddev=0.02,
86 | name="deconv2d", with_w=False):
87 | with tf.variable_scope(name):
88 | # filter : [height, width, output_channels, in_channels]
89 | #w = tf.get_variable('w', [k_h, k_h, output_shape[-1], input_.get_shape()[-1]], initializer=tf.random_normal_initializer(stddev=stddev))
90 | w = tf.get_variable('w', [k_h, k_h, output_shape[-1], input_.get_shape()[-1]], initializer = tf.contrib.layers.xavier_initializer())
91 | try:
92 | deconv = tf.nn.conv2d_transpose(input_, w, output_shape=output_shape,
93 | strides=[1, d_h, d_w, 1])
94 |
95 | # Support for verisons of TensorFlow before 0.7.0
96 | except AttributeError:
97 | deconv = tf.nn.deconv2d(input_, w, output_shape=output_shape,
98 | strides=[1, d_h, d_w, 1])
99 |
100 | biases = tf.get_variable('biases', [output_shape[-1]], initializer=tf.constant_initializer(0.0))
101 | deconv = tf.reshape(tf.nn.bias_add(deconv, biases), deconv.get_shape())
102 |
103 | if with_w:
104 | return deconv, w, biases
105 | else:
106 | return deconv
107 |
108 | def lrelu(x, leak=0.2, name="lrelu"):
109 | return tf.maximum(x, leak*x)
110 |
111 | def linear(input_, output_size, scope=None, stddev=0.02, bias_start=0.0,
112 | with_w=False):
113 | shape = input_.get_shape().as_list()
114 |
115 | with tf.variable_scope(scope or "Linear"):
116 | #matrix = tf.get_variable("Matrix", [shape[1], output_size], tf.float32,
117 | # tf.random_normal_initializer(stddev=stddev))
118 | matrix = tf.get_variable("Matrix", [shape[1], output_size], tf.float32,
119 | tf.contrib.layers.xavier_initializer())
120 | bias = tf.get_variable("bias", [output_size],
121 | initializer=tf.constant_initializer(bias_start))
122 | if with_w:
123 | return tf.matmul(input_, matrix) + bias, matrix, bias
124 | else:
125 | return tf.matmul(input_, matrix) + bias
126 |
127 | def attention(decoder_output, seq_outputs, output_size, time_steps,
128 | name="attention"):
129 |
130 | with tf.variable_scope(name):
131 | ui = []
132 | w_1 = tf.get_variable("w1", [output_size, output_size],
133 | tf.float32,
134 | tf.contrib.layers.xavier_initializer())
135 | w_2 = tf.get_variable("w2", [output_size, output_size],
136 | tf.float32,
137 | tf.contrib.layers.xavier_initializer())
138 | v = tf.get_variable("v", [output_size, 1],
139 | tf.float32,
140 | tf.contrib.layers.xavier_initializer())
141 | for seq_out in seq_outputs:
142 | ui.append(tf.matmul(tf.nn.tanh(tf.matmul(seq_out, w_1) +
143 | tf.matmul(decoder_output, w_2)), v))
144 |
145 | return ui
146 |
147 | def get_gt(batch_size, classes, real=1, name="gt"):
148 |
149 | with tf.variable_scope(name, reuse=None):
150 | r_f = tf.get_variable("rf", [batch_size, 1],
151 | initializer = tf.constant_initializer(
152 | real))
153 | gt = tf.concat(1, [r_f, classes], name = 'gt_concat_classes')
154 | return gt
--------------------------------------------------------------------------------
/Utils/plot_msssim.py:
--------------------------------------------------------------------------------
1 | # Plots a .tsv file. This was a simple script used to generate
2 | # the graph of the MS-SSIM in the paper.
3 |
4 | import matplotlib.pyplot as plt
5 | import argparse
6 |
7 | parser = argparse.ArgumentParser()
8 |
9 | parser.add_argument('--input_file', type=str, default="Data/msssim.tsv",
10 | help='the .tsv file that contains the msssim scores '
11 | 'generated by using msssim.py')
12 |
13 | parser.add_argument('--output_file', type=str, default="Data/msssim",
14 | help='The name of the output figure file that you '
15 | 'want to generate.')
16 |
17 | args = parser.parse_args()
18 |
19 | def open_tsv(file_name):
20 | with open(file_name, 'r') as f:
21 | tsv = f.readlines()
22 |
23 | ret = []
24 | for l in tsv:
25 | ret.append(l.split('\t'))
26 |
27 | print(ret)
28 | return ret
29 |
30 | tsv = open_tsv(args.tsv_file)
31 |
32 | x = []
33 | y = []
34 | for i in tsv:
35 | x.append(i[1])
36 | y.append(i[4])
37 |
38 |
39 | plt.scatter(x, y)
40 | plt.plot([0,1], [0,1])
41 | plt.xlim(0, 1)
42 | plt.ylim(0, 1)
43 | plt.xlabel('training data MS-SSIM value')
44 | plt.ylabel('samples MS-SSIM value')
45 |
46 | output_file_name = args.output_file + '.pdf'
47 | plt.savefig(output_file_name, format='pdf')
48 |
--------------------------------------------------------------------------------
/_config.yml:
--------------------------------------------------------------------------------
1 | theme: jekyll-theme-cayman
--------------------------------------------------------------------------------
/create_dataset.py:
--------------------------------------------------------------------------------
1 | import tensorflow as tf
2 | import numpy as np
3 | import model
4 | import argparse
5 | import pickle
6 | from os.path import join
7 | import scipy.misc
8 | import random
9 | import os
10 | from Utils import image_processing
11 |
12 |
13 | def main():
14 | parser = argparse.ArgumentParser()
15 | parser.add_argument('--z_dim', type=int, default=100,
16 | help='Noise dimension')
17 |
18 | parser.add_argument('--t_dim', type=int, default=256,
19 | help='Text feature dimension')
20 |
21 | parser.add_argument('--batch_size', type=int, default=64,
22 | help='Batch Size')
23 |
24 | parser.add_argument('--image_size', type=int, default=128,
25 | help='Image Size a, a x a')
26 |
27 | parser.add_argument('--gf_dim', type=int, default=64,
28 | help='Number of conv in the first layer gen.')
29 |
30 | parser.add_argument('--df_dim', type=int, default=64,
31 | help='Number of conv in the first layer discr.')
32 |
33 | parser.add_argument('--caption_vector_length', type=int, default=4800,
34 | help='Caption Vector Length')
35 |
36 | parser.add_argument('--n_classes', type=int, default=102,
37 | help='Number of classes/class labels')
38 |
39 | parser.add_argument('--data_dir', type=str, default="Data",
40 | help='Data Directory')
41 |
42 | parser.add_argument('--learning_rate', type=float, default=0.0002,
43 | help='Learning Rate')
44 |
45 | parser.add_argument('--beta1', type=float, default=0.5,
46 | help='Momentum for Adam Update')
47 |
48 | parser.add_argument('--epochs', type=int, default=200,
49 | help='Max number of epochs')
50 |
51 | parser.add_argument('--data_set', type=str, default="flowers",
52 | help='Dat set: flowers')
53 |
54 | parser.add_argument('--output_dir', type=str, default="Data/ds",
55 | help='The directory in which this dataset will be '
56 | 'created')
57 |
58 | parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
59 | help='Path to the checkpoints directory')
60 |
61 | args = parser.parse_args()
62 |
63 | model_stage_1_ds_tr, model_stage_1_ds_val, datasets_root_dir = \
64 | prepare_dirs(args)
65 |
66 | loaded_data = load_training_data(datasets_root_dir, args.data_set,
67 | args.caption_vector_length, args.n_classes)
68 |
69 | model_options = {
70 | 'z_dim': args.z_dim,
71 | 't_dim': args.t_dim,
72 | 'batch_size': args.batch_size,
73 | 'image_size': args.image_size,
74 | 'gf_dim': args.gf_dim,
75 | 'df_dim': args.df_dim,
76 | 'caption_vector_length': args.caption_vector_length,
77 | 'n_classes': loaded_data['n_classes']
78 | }
79 |
80 | gan = model.GAN(model_options)
81 | input_tensors, variables, loss, outputs, checks = gan.build_model()
82 |
83 | sess = tf.InteractiveSession()
84 | tf.initialize_all_variables().run()
85 |
86 | saver = tf.train.Saver(max_to_keep=10000)
87 | print('resuming model from checkpoint' +
88 | str(tf.train.latest_checkpoint(args.checkpoints_dir)))
89 | if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
90 | saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
91 | print('Successfully loaded model from ')
92 | else:
93 | print('Could not load checkpoints')
94 | exit()
95 |
96 | print('Generating images for the captions in the training set at ' +
97 | model_stage_1_ds_tr)
98 | for i in range(args.epochs):
99 | batch_no = 0
100 | while batch_no * args.batch_size + args.batch_size < \
101 | loaded_data['data_length']:
102 |
103 | real_images, wrong_images, caption_vectors, z_noise, image_files, \
104 | real_classes, wrong_classes, image_caps, image_ids, \
105 | image_caps_ids = get_training_batch(batch_no, args.batch_size,
106 | args.image_size, args.z_dim, datasets_root_dir,
107 | args.data_set, loaded_data)
108 |
109 | feed = {
110 | input_tensors['t_real_image'].name: real_images,
111 | input_tensors['t_wrong_image'].name: wrong_images,
112 | input_tensors['t_real_caption'].name: caption_vectors,
113 | input_tensors['t_z'].name: z_noise,
114 | input_tensors['t_real_classes'].name: real_classes,
115 | input_tensors['t_wrong_classes'].name: wrong_classes,
116 | input_tensors['t_training'].name: True
117 | }
118 |
119 | g_loss, gen = sess.run([loss['g_loss'], outputs['generator']],
120 | feed_dict=feed)
121 |
122 | print("LOSSES", g_loss, batch_no, i,
123 | len(loaded_data['image_list']) / args.batch_size)
124 | batch_no += 1
125 | save_distributed_image_batch(model_stage_1_ds_tr, gen, image_caps,
126 | image_ids, image_caps_ids)
127 |
128 | print('Finished generating images for the training set captions.\n\n')
129 | print('Generating images for the captions in the validation set at ' +
130 | model_stage_1_ds_val)
131 | for i in range(args.epochs):
132 | batch_no = 0
133 | while batch_no * args.batch_size + args.batch_size < \
134 | loaded_data['val_data_len']:
135 |
136 | val_captions, val_image_files, val_image_caps, val_image_ids, \
137 | val_image_caps_ids, val_z_noise = get_val_caps_batch(batch_no,
138 | args.batch_size, args.z_dim, loaded_data, args.data_set,
139 | datasets_root_dir)
140 |
141 | val_feed = {
142 | input_tensors['t_real_caption'].name: val_captions,
143 | input_tensors['t_z'].name: val_z_noise,
144 | input_tensors['t_training'].name: True
145 | }
146 |
147 | val_gen, val_attn_spn = sess.run(
148 | [outputs['generator'], checks['attn_span']],
149 | feed_dict=val_feed)
150 |
151 | print("LOSSES", batch_no, i, len(
152 | loaded_data['val_img_list']) / args.batch_size)
153 | batch_no += 1
154 | save_distributed_image_batch(model_stage_1_ds_val, val_gen,
155 | val_image_caps,
156 | val_image_ids,
157 | val_image_caps_ids, val_attn_spn)
158 | print('Finished generating images for the validation set captions.\n\n')
159 |
160 | def prepare_dirs(args):
161 |
162 | model_stage_1_ds_tr = join(args.output_dir, 'ds', 'train')
163 | if not os.path.exists(model_stage_1_ds_tr):
164 | os.makedirs(model_stage_1_ds_tr)
165 |
166 | model_stage_1_ds_val = join(args.output_dir, 'ds', 'val')
167 | if not os.path.exists(model_stage_1_ds_val):
168 | os.makedirs(model_stage_1_ds_val)
169 |
170 | datasets_root_dir = join(args.data_dir, 'datasets')
171 |
172 | return model_stage_1_ds_tr, model_stage_1_ds_val, datasets_root_dir
173 |
174 |
175 | def load_training_data(data_dir, data_set, caption_vector_length, n_classes):
176 | if data_set == 'flowers':
177 | flower_str_captions = pickle.load(
178 | open(join(data_dir, 'flowers', 'flowers_caps.pkl'), "rb"))
179 |
180 | img_classes = pickle.load(
181 | open(join(data_dir, 'flowers', 'flower_tc.pkl'), "rb"))
182 |
183 | flower_enc_captions = pickle.load(
184 | open(join(data_dir, 'flowers', 'flower_tv.pkl'), "rb"))
185 | # h1 = h5py.File(join(data_dir, 'flower_tc.hdf5'))
186 | tr_image_ids = pickle.load(
187 | open(join(data_dir, 'flowers', 'train_ids.pkl'), "rb"))
188 | val_image_ids = pickle.load(
189 | open(join(data_dir, 'flowers', 'val_ids.pkl'), "rb"))
190 |
191 | # n_classes = n_classes
192 | max_caps_len = caption_vector_length
193 |
194 | tr_n_imgs = len(tr_image_ids)
195 | val_n_imgs = len(val_image_ids)
196 |
197 | return {
198 | 'image_list': tr_image_ids,
199 | 'captions': flower_enc_captions,
200 | 'data_length': tr_n_imgs,
201 | 'classes': img_classes,
202 | 'n_classes': n_classes,
203 | 'max_caps_len': max_caps_len,
204 | 'val_img_list': val_image_ids,
205 | 'val_captions': flower_enc_captions,
206 | 'val_data_len': val_n_imgs,
207 | 'str_captions': flower_str_captions
208 | }
209 |
210 | else:
211 | raise Exception('Dataset not found')
212 |
213 |
214 | def save_distributed_image_batch(data_dir, generated_images, image_caps,
215 | image_ids, caps_ids):
216 | for i, (image_id, caps_id, image_cap) in enumerate(zip( image_ids, \
217 | caps_ids, image_caps)):
218 | image_dir = join(data_dir, str(image_id), str(caps_id))
219 | if not os.path.exists(image_dir):
220 | os.makedirs(image_dir)
221 | collection_dir = join(data_dir, 'collection')
222 | if not os.path.exists(collection_dir):
223 | os.makedirs(collection_dir)
224 | caps_dir = join(image_dir, "caps.txt")
225 | if not os.path.exists(caps_dir):
226 | with open(caps_dir, "w") as text_file:
227 | text_file.write(image_cap + "\n")
228 |
229 | fake_image_255 = (generated_images[i, :, :, :])
230 | if i == 0:
231 | scipy.misc.imsave(join(collection_dir, '{}.jpg'.format(image_id)),
232 | fake_image_255)
233 | num_files = len(os.walk(image_dir).__next__()[2])
234 | scipy.misc.imsave(join(image_dir, '{}.jpg'.format(num_files + 1)),
235 | fake_image_255)
236 |
237 |
238 | def get_training_batch(batch_no, batch_size, image_size, z_dim, data_dir,
239 | data_set, loaded_data=None):
240 |
241 | if data_set == 'flowers':
242 | real_images = np.zeros((batch_size, image_size, image_size, 3))
243 | wrong_images = np.zeros((batch_size, image_size, image_size, 3))
244 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
245 | real_classes = np.zeros((batch_size, loaded_data['n_classes']))
246 | wrong_classes = np.zeros((batch_size, loaded_data['n_classes']))
247 |
248 | cnt = 0
249 | image_files, image_caps, image_ids, image_caps_ids = [], [], [], []
250 |
251 | for i in range(batch_no * batch_size,
252 | batch_no * batch_size + batch_size):
253 |
254 | idx = i % len(loaded_data['image_list'])
255 | image_file = join(data_dir,
256 | 'flowers/jpg/' + loaded_data['image_list'][idx])
257 |
258 | image_ids.append(loaded_data['image_list'][idx])
259 |
260 | image_array = image_processing.load_image_array_flowers(image_file,
261 | image_size)
262 | real_images[cnt, :, :, :] = image_array
263 |
264 | # Improve this selection of wrong image
265 | wrong_image_id = random.randint(0,
266 | len(loaded_data['image_list']) - 1)
267 | wrong_image_file = join(data_dir,
268 | 'flowers/jpg/' + loaded_data['image_list'][
269 | wrong_image_id])
270 | wrong_image_array = image_processing.load_image_array_flowers(
271 | wrong_image_file,
272 | image_size)
273 | wrong_images[cnt, :, :, :] = wrong_image_array
274 |
275 | wrong_classes[cnt, :] = loaded_data['classes'][
276 | loaded_data['image_list'][
277 | wrong_image_id]][
278 | 0:loaded_data['n_classes']]
279 |
280 | random_caption = random.randint(0, 4)
281 | image_caps_ids.append(random_caption)
282 | captions[cnt, :] = \
283 | loaded_data['captions'][loaded_data['image_list'][idx]][
284 | random_caption][0:loaded_data['max_caps_len']]
285 |
286 | real_classes[cnt, :] = \
287 | loaded_data['classes'][loaded_data['image_list'][idx]][
288 | 0:loaded_data['n_classes']]
289 | str_cap = loaded_data['str_captions'][loaded_data['image_list']
290 | [idx]][random_caption]
291 |
292 | image_files.append(image_file)
293 | image_caps.append(str_cap)
294 | cnt += 1
295 |
296 | z_noise = np.random.uniform(-1, 1, [batch_size, z_dim])
297 | return real_images, wrong_images, captions, z_noise, image_files, \
298 | real_classes, wrong_classes, image_caps, image_ids, \
299 | image_caps_ids
300 | else:
301 | raise Exception('Dataset not found')
302 |
303 |
304 | def get_val_caps_batch(batch_no, batch_size, z_dim, loaded_data, data_set,
305 | data_dir):
306 |
307 | if data_set == 'flowers':
308 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
309 | batch_idx = range(batch_no * batch_size,
310 | batch_no * batch_size + batch_size)
311 | image_ids = np.take(loaded_data['val_img_list'], batch_idx)
312 |
313 | image_files = []
314 | image_caps = []
315 | image_caps_ids = []
316 |
317 | for idx, image_id in enumerate(image_ids) :
318 | image_file = join(data_dir,
319 | 'flowers/jpg/' + image_id)
320 | random_caption = random.randint(0, 4)
321 | image_caps_ids.append(random_caption)
322 | captions[idx, :] = \
323 | loaded_data['val_captions'][image_id][random_caption][
324 | 0 :loaded_data['max_caps_len']]
325 | str_cap = loaded_data['str_captions'][image_id][random_caption]
326 | image_caps.append(loaded_data['str_captions']
327 | [image_id][random_caption])
328 | image_files.append(image_file)
329 |
330 | z_noise = np.random.uniform(-1, 1, [batch_size, z_dim])
331 | return captions, image_files, image_caps, image_ids, image_caps_ids, \
332 | z_noise
333 |
334 |
335 | if __name__ == '__main__':
336 | main()
337 |
--------------------------------------------------------------------------------
/dataprep.py:
--------------------------------------------------------------------------------
1 | import os
2 | import argparse
3 | import skipthoughts
4 | import traceback
5 | import pickle
6 | import random
7 |
8 | import numpy as np
9 |
10 | from os.path import join
11 |
12 | def get_one_hot_targets(target_file_path):
13 | target = []
14 | one_hot_targets = []
15 | n_target = 0
16 | try :
17 | with open(target_file_path) as f :
18 | target = f.readlines()
19 | target = [t.strip('\n') for t in target]
20 | n_target = len(target)
21 | except IOError :
22 | print('Could not load the labels.txt file in the dataset. A '
23 | 'dataset folder is expected in the "data/datasets" '
24 | 'directory with the name that has been passed as an '
25 | 'argument to this method. This directory should contain a '
26 | 'file called labels.txt which contains a list of labels and '
27 | 'corresponding folders for the labels with the same name as '
28 | 'the labels.')
29 | traceback.print_stack()
30 |
31 | lbl_idxs = np.arange(n_target)
32 | one_hot_targets = np.zeros((n_target, n_target))
33 | one_hot_targets[np.arange(n_target), lbl_idxs] = 1
34 |
35 | return target, one_hot_targets, n_target
36 |
37 | def one_hot_encode_str_lbl(lbl, target, one_hot_targets):
38 | '''
39 | Encodes a string label into one-hot encoding
40 |
41 | Example:
42 | input: "window"
43 | output: [0 0 0 0 0 0 1 0 0 0 0 0]
44 | the length would depend on the number of classes in the dataset. The
45 | above is just a random example.
46 |
47 | :param lbl: The string label
48 | :return: one-hot encoding
49 | '''
50 | idx = target.index(lbl)
51 | return one_hot_targets[idx]
52 |
53 | def save_caption_vectors_flowers(data_dir, dt_range=(1, 103)) :
54 | import time
55 |
56 | img_dir = join(data_dir, 'flowers/jpg')
57 | all_caps_dir = join(data_dir, 'flowers/all_captions.txt')
58 | target_file_path = os.path.join(data_dir, "flowers/allclasses.txt")
59 | caption_dir = join(data_dir, 'flowers/text_c10')
60 | image_files = [f for f in os.listdir(img_dir) if 'jpg' in f]
61 | print(image_files[300 :400])
62 | image_captions = {}
63 | image_classes = {}
64 | class_dirs = []
65 | class_names = []
66 | img_ids = []
67 |
68 | target, one_hot_targets, n_target = get_one_hot_targets(target_file_path)
69 |
70 | for i in range(dt_range[0], dt_range[1]) :
71 | class_dir_name = 'class_%.5d' % (i)
72 | class_dir = join(caption_dir, class_dir_name)
73 | class_names.append(class_dir_name)
74 | class_dirs.append(class_dir)
75 | onlyimgfiles = [f[0 :11] + ".jpg" for f in os.listdir(class_dir)
76 | if 'txt' in f]
77 | for img_file in onlyimgfiles:
78 | image_classes[img_file] = None
79 |
80 | for img_file in onlyimgfiles:
81 | image_captions[img_file] = []
82 |
83 | for class_dir, class_name in zip(class_dirs, class_names) :
84 | caption_files = [f for f in os.listdir(class_dir) if 'txt' in f]
85 | for i, cap_file in enumerate(caption_files) :
86 | if i%50 == 0:
87 | print(str(i) + ' captions extracted from' + str(class_dir))
88 | with open(join(class_dir, cap_file)) as f :
89 | str_captions = f.read()
90 | captions = str_captions.split('\n')
91 | img_file = cap_file[0 :11] + ".jpg"
92 |
93 | # 5 captions per image
94 | image_captions[img_file] += [cap for cap in captions if len(cap) > 0][0 :5]
95 | image_classes[img_file] = one_hot_encode_str_lbl(class_name,
96 | target,
97 | one_hot_targets)
98 |
99 | model = skipthoughts.load_model()
100 | encoded_captions = {}
101 | for i, img in enumerate(image_captions) :
102 | st = time.time()
103 | encoded_captions[img] = skipthoughts.encode(model, image_captions[img])
104 | if i%20 == 0:
105 | print(i, len(image_captions), img)
106 | print("Seconds", time.time() - st)
107 |
108 | img_ids = list(image_captions.keys())
109 |
110 | random.shuffle(img_ids)
111 | n_train_instances = int(len(img_ids) * 0.9)
112 | tr_image_ids = img_ids[0 :n_train_instances]
113 | val_image_ids = img_ids[n_train_instances : -1]
114 |
115 | pickle.dump(image_captions,
116 | open(os.path.join(data_dir, 'flowers', 'flowers_caps.pkl'), "wb"))
117 |
118 | pickle.dump(tr_image_ids,
119 | open(os.path.join(data_dir, 'flowers', 'train_ids.pkl'), "wb"))
120 | pickle.dump(val_image_ids,
121 | open(os.path.join(data_dir, 'flowers', 'val_ids.pkl'), "wb"))
122 |
123 | ec_pkl_path = (join(data_dir, 'flowers', 'flower_tv.pkl'))
124 | pickle.dump(encoded_captions, open(ec_pkl_path, "wb"))
125 |
126 | fc_pkl_path = (join(data_dir, 'flowers', 'flower_tc.pkl'))
127 | pickle.dump(image_classes, open(fc_pkl_path, "wb"))
128 |
129 | def main() :
130 | parser = argparse.ArgumentParser()
131 | parser.add_argument('--data_dir', type = str, default = 'Data',
132 | help = 'Data directory')
133 | parser.add_argument('--dataset', type=str, default='flowers',
134 | help='Dataset to use. For Eg., "flowers"')
135 | args = parser.parse_args()
136 |
137 | dataset_dir = join(args.data_dir, "datasets")
138 | if args.dataset == 'flowers':
139 | save_caption_vectors_flowers(dataset_dir)
140 | else:
141 | print('Preprocessor for this dataset is not available.')
142 |
143 |
144 | if __name__ == '__main__' :
145 | main()
146 |
--------------------------------------------------------------------------------
/decoder.py:
--------------------------------------------------------------------------------
1 | #!/usr/bin/python
2 | #
3 | # Copyright 2016 The TensorFlow Authors. All Rights Reserved.
4 | # Licensed under the Apache License, Version 2.0 (the "License");
5 | # you may not use this file except in compliance with the License.
6 | # You may obtain a copy of the License at
7 | #
8 | # http://www.apache.org/licenses/LICENSE-2.0
9 | #
10 | # Unless required by applicable law or agreed to in writing, software
11 | # distributed under the License is distributed on an "AS IS" BASIS,
12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 | # See the License for the specific language governing permissions and
14 | # limitations under the License.
15 | # ==============================================================================
16 | r"""Neural Network Image Compression Decoder.
17 |
18 | Decompress an image from the numpy's npz format generated by the encoder.
19 |
20 | Example usage:
21 | python decoder.py --input_codes=output_codes.pkl --iteration=15 \
22 | --output_directory=/tmp/compression_output/ --model=residual_gru.pb
23 | """
24 | import io
25 | import os
26 |
27 | import numpy as np
28 | import tensorflow as tf
29 |
30 | tf.flags.DEFINE_string('input_codes', None, 'Location of binary code file.')
31 | tf.flags.DEFINE_integer('iteration', -1, 'The max quality level of '
32 | 'the images to output. Use -1 to infer from loaded '
33 | ' codes.')
34 | tf.flags.DEFINE_string('output_directory', None, 'Directory to save decoded '
35 | 'images.')
36 | tf.flags.DEFINE_string('model', None, 'Location of compression model.')
37 |
38 | FLAGS = tf.flags.FLAGS
39 |
40 |
41 | def get_input_tensor_names():
42 | name_list = ['GruBinarizer/SignBinarizer/Sign:0']
43 | for i in range(1, 16):
44 | name_list.append('GruBinarizer/SignBinarizer/Sign_{}:0'.format(i))
45 | return name_list
46 |
47 |
48 | def get_output_tensor_names():
49 | return ['loop_{0:02d}/add:0'.format(i) for i in range(0, 16)]
50 |
51 |
52 | def main(_):
53 | if (FLAGS.input_codes is None or FLAGS.output_directory is None or
54 | FLAGS.model is None):
55 | print('\nUsage: python decoder.py --input_codes=output_codes.pkl '
56 | '--iteration=15 --output_directory=/tmp/compression_output/ '
57 | '--model=residual_gru.pb\n\n')
58 | return
59 |
60 | if FLAGS.iteration < -1 or FLAGS.iteration > 15:
61 | print('\n--iteration must be between 0 and 15 inclusive, or -1 to infer '
62 | 'from file.\n')
63 | return
64 | iteration = FLAGS.iteration
65 |
66 | if not tf.gfile.Exists(FLAGS.output_directory):
67 | tf.gfile.MkDir(FLAGS.output_directory)
68 |
69 | if not tf.gfile.Exists(FLAGS.input_codes):
70 | print('\nInput codes not found.\n')
71 | return
72 |
73 | contents = ''
74 | with tf.gfile.FastGFile(FLAGS.input_codes, 'r') as code_file:
75 | contents = code_file.read()
76 | loaded_codes = np.load(io.BytesIO(contents))
77 | assert ['codes', 'shape'] not in loaded_codes.files
78 | loaded_shape = loaded_codes['shape']
79 | loaded_array = loaded_codes['codes']
80 |
81 | # Unpack and recover code shapes.
82 | unpacked_codes = np.reshape(np.unpackbits(loaded_array)
83 | [:np.prod(loaded_shape)],
84 | loaded_shape)
85 |
86 | numpy_int_codes = np.split(unpacked_codes, len(unpacked_codes))
87 | if iteration == -1:
88 | iteration = len(unpacked_codes) - 1
89 | # Convert back to float and recover scale.
90 | numpy_codes = [np.squeeze(x.astype(np.float32), 0) * 2 - 1 for x in
91 | numpy_int_codes]
92 |
93 | with tf.Graph().as_default() as graph:
94 | # Load the inference model for decoding.
95 | with tf.gfile.FastGFile(FLAGS.model, 'rb') as model_file:
96 | graph_def = tf.GraphDef()
97 | graph_def.ParseFromString(model_file.read())
98 | _ = tf.import_graph_def(graph_def, name='')
99 |
100 | # For encoding the tensors into PNGs.
101 | input_image = tf.placeholder(tf.uint8)
102 | encoded_image = tf.image.encode_png(input_image)
103 |
104 | input_tensors = [graph.get_tensor_by_name(name) for name in
105 | get_input_tensor_names()][0:iteration+1]
106 | outputs = [graph.get_tensor_by_name(name) for name in
107 | get_output_tensor_names()][0:iteration+1]
108 |
109 | feed_dict = {key: value for (key, value) in zip(input_tensors,
110 | numpy_codes)}
111 |
112 | with tf.Session(graph=graph) as sess:
113 | results = sess.run(outputs, feed_dict=feed_dict)
114 |
115 | for index, result in enumerate(results):
116 | img = np.uint8(np.clip(result + 0.5, 0, 255))
117 | img = img.squeeze()
118 | png_img = sess.run(encoded_image, feed_dict={input_image: img})
119 |
120 | with tf.gfile.FastGFile(os.path.join(FLAGS.output_directory,
121 | 'image_{0:02d}.png'.format(index)),
122 | 'w') as output_image:
123 | output_image.write(png_img)
124 |
125 |
126 | if __name__ == '__main__':
127 | tf.app.run()
128 |
--------------------------------------------------------------------------------
/encode_text.py:
--------------------------------------------------------------------------------
1 | import os
2 | import pickle
3 | import argparse
4 | import skipthoughts
5 | import sys
6 |
7 | def main():
8 | parser = argparse.ArgumentParser()
9 | parser.add_argument('--caption_file', type=str, default='Data/text.txt',
10 | help='caption file')
11 | parser.add_argument('--data_dir', type=str, default='Data',
12 | help='Data Directory')
13 |
14 | args = parser.parse_args()
15 |
16 | model = skipthoughts.load_model()
17 | encoded_captions = {}
18 | file_path = os.path.join(args.caption_file)
19 | dump_path = os.path.join(args.data_dir, 'enc_text.pkl')
20 | with open(file_path) as f:
21 | str_captions = f.read()
22 | captions = str_captions.split('\n')
23 | print(captions)
24 | encoded_captions['features'] = skipthoughts.encode(model, captions)
25 |
26 | pickle.dump(encoded_captions,
27 | open(dump_path, "wb"))
28 | print('Finished extracting Skip-Thought vectors of the given text '
29 | 'descriptions')
30 |
31 | if __name__ == '__main__':
32 | main()
--------------------------------------------------------------------------------
/encoder.py:
--------------------------------------------------------------------------------
1 | #!/usr/bin/python
2 | #
3 | # Copyright 2016 The TensorFlow Authors. All Rights Reserved.
4 | #
5 | # Licensed under the Apache License, Version 2.0 (the "License");
6 | # you may not use this file except in compliance with the License.
7 | # You may obtain a copy of the License at
8 | #
9 | # http://www.apache.org/licenses/LICENSE-2.0
10 | #
11 | # Unless required by applicable law or agreed to in writing, software
12 | # distributed under the License is distributed on an "AS IS" BASIS,
13 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 | # See the License for the specific language governing permissions and
15 | # limitations under the License.
16 | # ==============================================================================
17 | r"""Neural Network Image Compression Encoder.
18 |
19 | Compresses an image to a binarized numpy array. The image must be padded to a
20 | multiple of 32 pixels in height and width.
21 |
22 | Example usage:
23 | python encoder.py --input_image=/your/image/here.png \
24 | --output_codes=output_codes.pkl --iteration=15 --model=residual_gru.pb
25 | """
26 | import io
27 | import os
28 |
29 | import numpy as np
30 | import tensorflow as tf
31 |
32 | tf.flags.DEFINE_string('input_image', None, 'Location of input image. We rely '
33 | 'on tf.image to decode the image, so only PNG and JPEG '
34 | 'formats are currently supported.')
35 | tf.flags.DEFINE_integer('iteration', 15, 'Quality level for encoding image. '
36 | 'Must be between 0 and 15 inclusive.')
37 | tf.flags.DEFINE_string('output_codes', None, 'File to save output encoding.')
38 | tf.flags.DEFINE_string('model', None, 'Location of compression model.')
39 |
40 | FLAGS = tf.flags.FLAGS
41 |
42 |
43 | def get_output_tensor_names():
44 | name_list = ['GruBinarizer/SignBinarizer/Sign:0']
45 | for i in range(1, 16):
46 | name_list.append('GruBinarizer/SignBinarizer/Sign_{}:0'.format(i))
47 | return name_list
48 |
49 |
50 | def main(_):
51 | if (FLAGS.input_image is None or FLAGS.output_codes is None or
52 | FLAGS.model is None):
53 | print('\nUsage: python encoder.py --input_image=/your/image/here.png '
54 | '--output_codes=output_codes.pkl --iteration=15 '
55 | '--model=residual_gru.pb\n\n')
56 | return
57 |
58 | if FLAGS.iteration < 0 or FLAGS.iteration > 15:
59 | print('\n--iteration must be between 0 and 15 inclusive.\n')
60 | return
61 |
62 | with tf.gfile.FastGFile(FLAGS.input_image) as input_image:
63 | input_image_str = input_image.read()
64 |
65 | with tf.Graph().as_default() as graph:
66 | # Load the inference model for encoding.
67 | with tf.gfile.FastGFile(FLAGS.model, 'rb') as model_file:
68 | graph_def = tf.GraphDef()
69 | graph_def.ParseFromString(model_file.read())
70 | _ = tf.import_graph_def(graph_def, name='')
71 |
72 | input_tensor = graph.get_tensor_by_name('Placeholder:0')
73 | outputs = [graph.get_tensor_by_name(name) for name in
74 | get_output_tensor_names()]
75 |
76 | input_image = tf.placeholder(tf.string)
77 | _, ext = os.path.splitext(FLAGS.input_image)
78 | if ext == '.png':
79 | decoded_image = tf.image.decode_png(input_image, channels=3)
80 | elif ext == '.jpeg' or ext == '.jpg':
81 | decoded_image = tf.image.decode_jpeg(input_image, channels=3)
82 | else:
83 | assert False, 'Unsupported file format {}'.format(ext)
84 | decoded_image = tf.expand_dims(decoded_image, 0)
85 |
86 | with tf.Session(graph=graph) as sess:
87 | img_array = sess.run(decoded_image, feed_dict={input_image:
88 | input_image_str})
89 | results = sess.run(outputs, feed_dict={input_tensor: img_array})
90 |
91 | results = results[0:FLAGS.iteration + 1]
92 | int_codes = np.asarray([x.astype(np.int8) for x in results])
93 |
94 | # Convert int codes to binary.
95 | int_codes = (int_codes + 1)//2
96 | export = np.packbits(int_codes.reshape(-1))
97 |
98 | output = io.BytesIO()
99 | np.savez_compressed(output, shape=int_codes.shape, codes=export)
100 | with tf.gfile.FastGFile(FLAGS.output_codes, 'w') as code_file:
101 | code_file.write(output.getvalue())
102 |
103 |
104 | if __name__ == '__main__':
105 | tf.app.run()
106 |
--------------------------------------------------------------------------------
/generate_images.py:
--------------------------------------------------------------------------------
1 | import model
2 | import argparse
3 | import pickle
4 | import scipy.misc
5 | import random
6 | import os
7 |
8 | import tensorflow as tf
9 | import numpy as np
10 |
11 | from os.path import join
12 |
13 |
14 | def main():
15 | parser = argparse.ArgumentParser()
16 | parser.add_argument('--z_dim', type=int, default=100,
17 | help='Noise dimension')
18 |
19 | parser.add_argument('--t_dim', type=int, default=256,
20 | help='Text feature dimension')
21 |
22 | parser.add_argument('--batch_size', type=int, default=64,
23 | help='Batch Size')
24 |
25 | parser.add_argument('--image_size', type=int, default=128,
26 | help='Image Size a, a x a')
27 |
28 | parser.add_argument('--gf_dim', type=int, default=64,
29 | help='Number of conv in the first layer gen.')
30 |
31 | parser.add_argument('--df_dim', type=int, default=64,
32 | help='Number of conv in the first layer discr.')
33 |
34 | parser.add_argument('--caption_vector_length', type=int, default=4800,
35 | help='Caption Vector Length')
36 |
37 | parser.add_argument('--n_classes', type=int, default=102,
38 | help='Number of classes/class labels')
39 |
40 | parser.add_argument('--data_dir', type=str, default="Data",
41 | help='Data Directory')
42 |
43 | parser.add_argument('--learning_rate', type=float, default=0.0002,
44 | help='Learning Rate')
45 |
46 | parser.add_argument('--beta1', type=float, default=0.5,
47 | help='Momentum for Adam Update')
48 |
49 | parser.add_argument('--images_per_caption', type=int, default=30,
50 | help='The number of images that you want to generate '
51 | 'per text description')
52 |
53 | parser.add_argument('--data_set', type=str, default="flowers",
54 | help='Dat set: MS-COCO, flowers')
55 |
56 | parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
57 | help='Path to the checkpoints directory')
58 |
59 |
60 | args = parser.parse_args()
61 |
62 | datasets_root_dir = join(args.data_dir, 'datasets')
63 |
64 | loaded_data = load_training_data(datasets_root_dir, args.data_set,
65 | args.caption_vector_length,
66 | args.n_classes)
67 | model_options = {
68 | 'z_dim': args.z_dim,
69 | 't_dim': args.t_dim,
70 | 'batch_size': args.batch_size,
71 | 'image_size': args.image_size,
72 | 'gf_dim': args.gf_dim,
73 | 'df_dim': args.df_dim,
74 | 'caption_vector_length': args.caption_vector_length,
75 | 'n_classes': loaded_data['n_classes']
76 | }
77 |
78 | gan = model.GAN(model_options)
79 | input_tensors, variables, loss, outputs, checks = gan.build_model()
80 |
81 | sess = tf.InteractiveSession()
82 | tf.initialize_all_variables().run()
83 |
84 | saver = tf.train.Saver(max_to_keep=10000)
85 | print('Trying to resume model from ' +
86 | str(tf.train.latest_checkpoint(args.checkpoints_dir)))
87 | if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
88 | saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
89 | print('Successfully loaded model from ')
90 | else:
91 | print('Could not load checkpoints. Please provide a valid path to'
92 | ' your checkpoints directory')
93 | exit()
94 |
95 | print('Starting to generate images from text descriptions.')
96 | for sel_i, text_cap in enumerate(loaded_data['text_caps']['features']):
97 |
98 | print('Text idx: {}\nRaw Text: {}\n'.format(sel_i, text_cap))
99 | captions_1, image_files_1, image_caps_1, image_ids_1,\
100 | image_caps_ids_1 = get_caption_batch(loaded_data, datasets_root_dir,
101 | dataset=args.data_set, batch_size=args.batch_size)
102 |
103 | captions_1[args.batch_size-1, :] = text_cap
104 |
105 | for z_i in range(args.images_per_caption):
106 | z_noise = np.random.uniform(-1, 1, [args.batch_size, args.z_dim])
107 | val_feed = {
108 | input_tensors['t_real_caption'].name: captions_1,
109 | input_tensors['t_z'].name: z_noise,
110 | input_tensors['t_training'].name: True
111 | }
112 |
113 | val_gen = sess.run(
114 | [outputs['generator']],
115 | feed_dict=val_feed)
116 | dump_dir = os.path.join(args.data_dir,
117 | 'images_generated_from_text')
118 | save_distributed_image_batch(dump_dir, val_gen, sel_i, z_i,
119 | args.batch_size)
120 | print('Finished generating images from text description')
121 |
122 |
123 | def load_training_data(data_dir, data_set, caption_vector_length, n_classes):
124 | if data_set == 'flowers':
125 | flower_str_captions = pickle.load(
126 | open(join(data_dir, 'flowers', 'flowers_caps.pkl'), "rb"))
127 |
128 | img_classes = pickle.load(
129 | open(join(data_dir, 'flowers', 'flower_tc.pkl'), "rb"))
130 |
131 | flower_enc_captions = pickle.load(
132 | open(join(data_dir, 'flowers', 'flower_tv.pkl'), "rb"))
133 | # h1 = h5py.File(join(data_dir, 'flower_tc.hdf5'))
134 | tr_image_ids = pickle.load(
135 | open(join(data_dir, 'flowers', 'train_ids.pkl'), "rb"))
136 | val_image_ids = pickle.load(
137 | open(join(data_dir, 'flowers', 'val_ids.pkl'), "rb"))
138 | caps_new = pickle.load(
139 | open(join('Data', 'enc_text.pkl'), "rb"))
140 |
141 | # n_classes = n_classes
142 | max_caps_len = caption_vector_length
143 |
144 | tr_n_imgs = len(tr_image_ids)
145 | val_n_imgs = len(val_image_ids)
146 |
147 | return {
148 | 'image_list': tr_image_ids,
149 | 'captions': flower_enc_captions,
150 | 'data_length': tr_n_imgs,
151 | 'classes': img_classes,
152 | 'n_classes': n_classes,
153 | 'max_caps_len': max_caps_len,
154 | 'val_img_list': val_image_ids,
155 | 'val_captions': flower_enc_captions,
156 | 'val_data_len': val_n_imgs,
157 | 'str_captions': flower_str_captions,
158 | 'text_caps': caps_new
159 | }
160 |
161 | else:
162 | raise Exception('This dataset has not been handeled yet. '
163 | 'Contributions are welcome.')
164 |
165 |
166 | def save_distributed_image_batch(data_dir, generated_images, sel_i, z_i,
167 | batch_size=64):
168 | generated_images = np.squeeze(generated_images)
169 | folder_name = str(sel_i)
170 | image_dir = join(data_dir, folder_name)
171 | if not os.path.exists(image_dir):
172 | os.makedirs(image_dir)
173 | fake_image_255 = generated_images[batch_size-1]
174 | scipy.misc.imsave(join(image_dir, '{}.jpg'.format(z_i)),
175 | fake_image_255)
176 |
177 |
178 | def get_caption_batch(loaded_data, data_dir, dataset='flowers', batch_size=64):
179 |
180 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
181 | batch_idx = np.random.randint(0, loaded_data['data_length'],
182 | size=batch_size)
183 | image_ids = np.take(loaded_data['image_list'], batch_idx)
184 | image_files = []
185 | image_caps = []
186 | image_caps_ids = []
187 | for idx, image_id in enumerate(image_ids):
188 | image_file = join(data_dir, dataset, 'jpg' + image_id)
189 | random_caption = random.randint(0, 4)
190 | image_caps_ids.append(random_caption)
191 | captions[idx, :] = \
192 | loaded_data['captions'][image_id][random_caption][
193 | 0:loaded_data['max_caps_len']]
194 |
195 | image_caps.append(loaded_data['captions']
196 | [image_id][random_caption])
197 | image_files.append(image_file)
198 |
199 | return captions, image_files, image_caps, image_ids, image_caps_ids
200 |
201 | if __name__ == '__main__':
202 | main()
--------------------------------------------------------------------------------
/inception_score.py:
--------------------------------------------------------------------------------
1 | import os
2 | import argparse
3 | import progressbar
4 |
5 | from shutil import copy
6 | from Utils import inception_score as ins
7 | from Utils import image_processing as ip
8 |
9 |
10 |
11 | def prepare_inception_data(o_dir, i_dir):
12 | if not os.path.exists(o_dir):
13 | os.makedirs(o_dir)
14 | cnt = 0
15 | bar = progressbar.ProgressBar(redirect_stdout=True,
16 | max_value=progressbar.UnknownLength)
17 | for root, subFolders, files in os.walk(i_dir):
18 | if files:
19 | for f in files:
20 | if 'jpg' in f:
21 | f_name = str(cnt) + '_ins.' + f.split('.')[-1]
22 | cnt += 1
23 | file_dir = os.path.join(root, f)
24 | dest_path = os.path.join(o_dir, f)
25 | dest_new_name = os.path.join(o_dir, f_name)
26 | copy(file_dir, o_dir)
27 | os.rename(dest_path, dest_new_name)
28 | bar.update(cnt)
29 | bar.finish()
30 | print('Total number of files: {}'.format(cnt))
31 |
32 | def load_images(o_dir, i_dir, n_images=3000, size=128):
33 | prepare_inception_data(o_dir, i_dir)
34 | image_list = []
35 | done = False
36 | cnt = 0
37 | bar = progressbar.ProgressBar(redirect_stdout=True,
38 | max_value=progressbar.UnknownLength)
39 | for root, dirs, files in os.walk(o_dir):
40 | if files:
41 | for f in files:
42 | cnt += 1
43 | file_dir = os.path.join(root, f)
44 | image_list.append(ip.load_image_inception(file_dir, 0))
45 | bar.update(cnt)
46 | if len(image_list) == n_images:
47 | done = True
48 | break
49 | if done:
50 | break
51 | bar.finish()
52 | print('Finished Loading Files')
53 | return image_list
54 |
55 |
56 | if __name__ == '__main__':
57 | parser = argparse.ArgumentParser()
58 |
59 | parser.add_argument('--output_dir', type=str, default="Data/ds_inception",
60 | help='directory to dump all the images for '
61 | 'calculating the inception score')
62 |
63 | parser.add_argument('--data_dir', type=str,
64 | default="Data/synthetic_dataset/ds",
65 | help='The root directory of the synthetic dataset')
66 |
67 | parser.add_argument('--n_images', type=int, default=30000,
68 | help='Number of images to consider for calculating '
69 | 'inception score')
70 |
71 | parser.add_argument('--image_size', type=int, default=128,
72 | help='Size of the image to consider for calculating '
73 | 'inception score')
74 |
75 | args = parser.parse_args()
76 |
77 | imgs_list = load_images(args.output_dir, args.data_dir,
78 | n_images=args.n_images, size=args.image_size)
79 |
80 | print('Extracting Inception Score')
81 | mean, std = ins.get_inception_score(imgs_list)
82 | print('Mean Inception Score: {}\nStandard Deviation: {}'.format(mean, std))
--------------------------------------------------------------------------------
/model.py:
--------------------------------------------------------------------------------
1 | import tensorflow as tf
2 | import tensorflow.contrib.slim as slim
3 | from Utils import ops
4 |
5 |
6 | class GAN :
7 | '''
8 | OPTIONS
9 | z_dim : Noise dimension 100
10 | t_dim : Text feature dimension 256
11 | image_size : Image Dimension 64
12 | gf_dim : Number of conv in the first layer generator 64
13 | df_dim : Number of conv in the first layer discriminator 64
14 | gfc_dim : Dimension of gen untis for for fully connected layer 1024
15 | caption_vector_length : Caption Vector Length 2400
16 | batch_size : Batch Size 64
17 | '''
18 |
19 | def __init__(self, options) :
20 | self.options = options
21 |
22 | def build_model(self) :
23 |
24 | print('Initializing placeholder')
25 | img_size = self.options['image_size']
26 | t_real_image = tf.placeholder('float32', [self.options['batch_size'],
27 | img_size, img_size, 3],
28 | name = 'real_image')
29 | t_wrong_image = tf.placeholder('float32', [self.options['batch_size'],
30 | img_size, img_size, 3],
31 | name = 'wrong_image')
32 |
33 | t_real_caption = tf.placeholder('float32', [self.options['batch_size'],
34 | self.options['caption_vector_length']],
35 | name='real_captions')
36 |
37 | t_z = tf.placeholder('float32', [self.options['batch_size'],
38 | self.options['z_dim']], name='input_noise')
39 |
40 | t_real_classes = tf.placeholder('float32', [self.options['batch_size'],
41 | self.options['n_classes']],
42 | name='real_classes')
43 |
44 | t_wrong_classes = tf.placeholder('float32', [self.options['batch_size'],
45 | self.options['n_classes']],
46 | name='wrong_classes')
47 |
48 | t_training = tf.placeholder(tf.bool, name='training')
49 |
50 | print('Building the Generator')
51 | fake_image = self.generator(t_z, t_real_caption,
52 | t_training)
53 |
54 | print('Building the Discriminator')
55 | disc_real_image, disc_real_image_logits, disc_real_image_aux, \
56 | disc_real_image_aux_logits = self.discriminator(
57 | t_real_image, t_real_caption, self.options['n_classes'],
58 | t_training)
59 |
60 | disc_wrong_image, disc_wrong_image_logits, disc_wrong_image_aux, \
61 | disc_wrong_image_aux_logits = self.discriminator(
62 | t_wrong_image, t_real_caption, self.options['n_classes'],
63 | t_training, reuse = True)
64 |
65 | disc_fake_image, disc_fake_image_logits, disc_fake_image_aux, \
66 | disc_fake_image_aux_logits = self.discriminator(
67 | fake_image, t_real_caption, self.options['n_classes'],
68 | t_training, reuse = True)
69 |
70 | d_right_predictions = tf.equal(tf.argmax(disc_real_image_aux, 1),
71 | tf.argmax(t_real_classes, 1))
72 | d_right_accuracy = tf.reduce_mean(tf.cast(d_right_predictions,
73 | tf.float32))
74 |
75 | d_wrong_predictions = tf.equal(tf.argmax(disc_wrong_image_aux, 1),
76 | tf.argmax(t_wrong_classes, 1))
77 | d_wrong_accuracy = tf.reduce_mean(tf.cast(d_wrong_predictions,
78 | tf.float32))
79 |
80 | d_fake_predictions = tf.equal(tf.argmax(disc_fake_image_aux_logits, 1),
81 | tf.argmax(t_real_classes, 1))
82 | d_fake_accuracy = tf.reduce_mean(tf.cast(d_fake_predictions,
83 | tf.float32))
84 |
85 | tf.get_variable_scope()._reuse = False
86 |
87 | print('Building the Loss Function')
88 | g_loss_1 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
89 | logits=disc_fake_image_logits,
90 | labels=tf.ones_like(disc_fake_image)))
91 |
92 | g_loss_2 = tf.reduce_mean(
93 | tf.nn.sigmoid_cross_entropy_with_logits(
94 | logits=disc_fake_image_aux_logits,
95 | labels=t_real_classes))
96 |
97 | d_loss1 = tf.reduce_mean(
98 | tf.nn.sigmoid_cross_entropy_with_logits(
99 | logits=disc_real_image_logits,
100 | labels=tf.ones_like(disc_real_image)))
101 | d_loss1_1 = tf.reduce_mean(
102 | tf.nn.sigmoid_cross_entropy_with_logits(
103 | logits=disc_real_image_aux_logits,
104 | labels=t_real_classes))
105 | d_loss2 = tf.reduce_mean(
106 | tf.nn.sigmoid_cross_entropy_with_logits(
107 | logits=disc_wrong_image_logits,
108 | labels=tf.zeros_like(disc_wrong_image)))
109 | d_loss2_1 = tf.reduce_mean(
110 | tf.nn.sigmoid_cross_entropy_with_logits(
111 | logits=disc_wrong_image_aux_logits,
112 | labels=t_wrong_classes))
113 | d_loss3 = tf.reduce_mean(
114 | tf.nn.sigmoid_cross_entropy_with_logits(
115 | logits=disc_fake_image_logits,
116 | labels=tf.zeros_like(disc_fake_image)))
117 |
118 | d_loss = d_loss1 + d_loss1_1 + d_loss2 + d_loss2_1 + d_loss3 + g_loss_2
119 |
120 | g_loss = g_loss_1 + g_loss_2
121 |
122 | t_vars = tf.trainable_variables()
123 | print('List of all variables')
124 | for v in t_vars:
125 | print(v.name)
126 | print(v)
127 | self.add_histogram_summary(v.name, v)
128 |
129 | self.add_tb_scalar_summaries(d_loss, g_loss, d_loss1, d_loss2, d_loss3,
130 | d_loss1_1, d_loss2_1, g_loss_1, g_loss_2, d_right_accuracy,
131 | d_wrong_accuracy, d_fake_accuracy)
132 |
133 | self.add_image_summary('Generated Images', fake_image,
134 | self.options['batch_size'])
135 |
136 | d_vars = [var for var in t_vars if 'd_' in var.name]
137 | g_vars = [var for var in t_vars if 'g_' in var.name]
138 |
139 | input_tensors = {
140 | 't_real_image' : t_real_image,
141 | 't_wrong_image' : t_wrong_image,
142 | 't_real_caption' : t_real_caption,
143 | 't_z' : t_z,
144 | 't_real_classes' : t_real_classes,
145 | 't_wrong_classes' : t_wrong_classes,
146 | 't_training' : t_training,
147 |
148 | }
149 |
150 | variables = {
151 | 'd_vars' : d_vars,
152 | 'g_vars' : g_vars
153 | }
154 |
155 | loss = {
156 | 'g_loss' : g_loss,
157 | 'd_loss' : d_loss
158 | }
159 |
160 | outputs = {
161 | 'generator' : fake_image
162 | }
163 |
164 | checks = {
165 | 'd_loss1': d_loss1,
166 | 'd_loss2': d_loss2,
167 | 'd_loss3': d_loss3,
168 | 'g_loss_1': g_loss_1,
169 | 'g_loss_2': g_loss_2,
170 | 'd_loss1_1': d_loss1_1,
171 | 'd_loss2_1': d_loss2_1,
172 | 'disc_real_image_logits': disc_real_image_logits,
173 | 'disc_wrong_image_logits': disc_wrong_image,
174 | 'disc_fake_image_logits': disc_fake_image_logits
175 | }
176 |
177 | return input_tensors, variables, loss, outputs, checks
178 |
179 | def add_tb_scalar_summaries(self, d_loss, g_loss, d_loss1, d_loss2,
180 | d_loss3, d_loss1_1, d_loss2_1, g_loss_1,
181 | g_loss_2, d_right_accuracy,
182 | d_wrong_accuracy, d_fake_accuracy):
183 |
184 | self.add_scalar_summary("D_Loss", d_loss)
185 | self.add_scalar_summary("G_Loss", g_loss)
186 | self.add_scalar_summary("D loss-1 [Real/Fake loss for real images]",
187 | d_loss1)
188 | self.add_scalar_summary("D loss-2 [Real/Fake loss for wrong images]",
189 | d_loss2)
190 | self.add_scalar_summary("D loss-3 [Real/Fake loss for fake images]",
191 | d_loss3)
192 | self.add_scalar_summary(
193 | "D loss-4 [Aux Classifier loss for real images]", d_loss1_1)
194 | self.add_scalar_summary(
195 | "D loss-5 [Aux Classifier loss for wrong images]", d_loss2_1)
196 | self.add_scalar_summary("G loss-1 [Real/Fake loss for fake images]",
197 | g_loss_1)
198 | self.add_scalar_summary(
199 | "G loss-2 [Aux Classifier loss for fake images]", g_loss_2)
200 | self.add_scalar_summary("Discriminator Real Image Accuracy",
201 | d_right_accuracy)
202 | self.add_scalar_summary("Discriminator Wrong Image Accuracy",
203 | d_wrong_accuracy)
204 | self.add_scalar_summary("Discriminator Fake Image Accuracy",
205 | d_fake_accuracy)
206 |
207 | def add_scalar_summary(self, name, var):
208 | with tf.name_scope('summaries'):
209 | tf.summary.scalar(name, var)
210 |
211 | def add_histogram_summary(self, name, var):
212 | with tf.name_scope('summaries'):
213 | tf.summary.histogram(name, var)
214 |
215 | def add_image_summary(self, name, var, max_outputs=1):
216 | with tf.name_scope('summaries'):
217 | tf.summary.image(name, var, max_outputs=max_outputs)
218 |
219 | # GENERATOR IMPLEMENTATION based on :
220 | # https://github.com/carpedm20/DCGAN-tensorflow/blob/master/model.py
221 | def generator(self, t_z, t_text_embedding, t_training):
222 |
223 | s = self.options['image_size']
224 | s2, s4, s8, s16 = int(s / 2), int(s / 4), int(s / 8), int(s / 16)
225 |
226 | reduced_text_embedding = ops.lrelu(
227 | ops.linear(t_text_embedding, self.options['t_dim'], 'g_embedding'))
228 | z_concat = tf.concat([t_z, reduced_text_embedding], -1)
229 | z_ = ops.linear(z_concat, self.options['gf_dim'] * 8 * s16 * s16,
230 | 'g_h0_lin')
231 | h0 = tf.reshape(z_, [-1, s16, s16, self.options['gf_dim'] * 8])
232 | h0 = tf.nn.relu(slim.batch_norm(h0, is_training = t_training,
233 | scope="g_bn0"))
234 |
235 | h1 = ops.deconv2d(h0, [self.options['batch_size'], s8, s8,
236 | self.options['gf_dim'] * 4], name = 'g_h1')
237 | h1 = tf.nn.relu(slim.batch_norm(h1, is_training = t_training,
238 | scope="g_bn1"))
239 |
240 | h2 = ops.deconv2d(h1, [self.options['batch_size'], s4, s4,
241 | self.options['gf_dim'] * 2], name = 'g_h2')
242 | h2 = tf.nn.relu(slim.batch_norm(h2, is_training = t_training,
243 | scope="g_bn2"))
244 |
245 | h3 = ops.deconv2d(h2, [self.options['batch_size'], s2, s2,
246 | self.options['gf_dim'] * 1], name = 'g_h3')
247 | h3 = tf.nn.relu(slim.batch_norm(h3, is_training = t_training,
248 | scope="g_bn3"))
249 |
250 | h4 = ops.deconv2d(h3, [self.options['batch_size'], s, s, 3],
251 | name = 'g_h4')
252 | return (tf.tanh(h4) / 2. + 0.5)
253 |
254 |
255 | # DISCRIMINATOR IMPLEMENTATION based on :
256 | # https://github.com/carpedm20/DCGAN-tensorflow/blob/master/model.py
257 | def discriminator(self, image, t_text_embedding, n_classes, t_training,
258 | reuse = False) :
259 | if reuse :
260 | tf.get_variable_scope().reuse_variables()
261 |
262 | h0 = ops.lrelu(
263 | ops.conv2d(image, self.options['df_dim'], name = 'd_h0_conv')) # 64
264 |
265 | h1 = ops.lrelu(slim.batch_norm(ops.conv2d(h0,
266 | self.options['df_dim'] * 2,
267 | name = 'd_h1_conv'),
268 | reuse=reuse,
269 | is_training = t_training,
270 | scope = 'd_bn1')) # 32
271 |
272 | h2 = ops.lrelu(slim.batch_norm(ops.conv2d(h1,
273 | self.options['df_dim'] * 4,
274 | name = 'd_h2_conv'),
275 | reuse=reuse,
276 | is_training = t_training,
277 | scope = 'd_bn2')) # 16
278 | h3 = ops.lrelu(slim.batch_norm(ops.conv2d(h2,
279 | self.options['df_dim'] * 8,
280 | name = 'd_h3_conv'),
281 | reuse=reuse,
282 | is_training = t_training,
283 | scope = 'd_bn3')) # 8
284 | h3_shape = h3.get_shape().as_list()
285 | # ADD TEXT EMBEDDING TO THE NETWORK
286 | reduced_text_embeddings = ops.lrelu(ops.linear(t_text_embedding,
287 | self.options['t_dim'],
288 | 'd_embedding'))
289 | reduced_text_embeddings = tf.expand_dims(reduced_text_embeddings, 1)
290 | reduced_text_embeddings = tf.expand_dims(reduced_text_embeddings, 2)
291 | tiled_embeddings = tf.tile(reduced_text_embeddings,
292 | [1, h3_shape[1], h3_shape[1], 1],
293 | name = 'tiled_embeddings')
294 |
295 | h3_concat = tf.concat([h3, tiled_embeddings], 3, name = 'h3_concat')
296 | h3_new = ops.lrelu(slim.batch_norm(ops.conv2d(h3_concat,
297 | self.options['df_dim'] * 8,
298 | 1, 1, 1, 1,
299 | name = 'd_h3_conv_new'),
300 | reuse=reuse,
301 | is_training = t_training,
302 | scope = 'd_bn4')) # 4
303 |
304 | h3_flat = tf.reshape(h3_new, [self.options['batch_size'], -1])
305 |
306 | h4 = ops.linear(h3_flat, 1, 'd_h4_lin_rw')
307 | h4_aux = ops.linear(h3_flat, n_classes, 'd_h4_lin_ac')
308 |
309 | return tf.nn.sigmoid(h4), h4, tf.nn.sigmoid(h4_aux), h4_aux
310 |
311 | # This has not been used used yet but can be used
312 | def attention(self, decoder_output, seq_outputs, output_size, time_steps,
313 | reuse=False) :
314 | if reuse:
315 | tf.get_variable_scope().reuse_variables()
316 | ui = ops.attention(decoder_output, seq_outputs, output_size,
317 | time_steps, name = "g_a_attention")
318 |
319 | with tf.variable_scope('g_a_attention'):
320 | ui = tf.transpose(ui, [1, 0, 2])
321 | ai = tf.nn.softmax(ui, dim=1)
322 | seq_outputs = tf.transpose(seq_outputs, [1, 0, 2])
323 | d_dash = tf.reduce_sum(tf.mul(seq_outputs, ai), axis=1)
324 | return d_dash, ai
325 |
--------------------------------------------------------------------------------
/msssim.py:
--------------------------------------------------------------------------------
1 | #!/usr/bin/python
2 | #
3 | # Copyright 2016 The TensorFlow Authors. All Rights Reserved.
4 | #
5 | # Licensed under the Apache License, Version 2.0 (the "License");
6 | # you may not use this file except in compliance with the License.
7 | # You may obtain a copy of the License at
8 | #
9 | # http://www.apache.org/licenses/LICENSE-2.0
10 | #
11 | # Unless required by applicable law or agreed to in writing, software
12 | # distributed under the License is distributed on an "AS IS" BASIS,
13 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 | # See the License for the specific language governing permissions and
15 | # limitations under the License.
16 | # ==============================================================================
17 |
18 | """Python implementation of MS-SSIM.
19 |
20 | Usage:
21 |
22 | python msssim.py --original_image=original.png --compared_image=distorted.png
23 | """
24 | import os
25 | import argparse
26 | import sys
27 |
28 | import tensorflow as tf
29 | import numpy as np
30 |
31 | from skimage.transform import resize
32 | from scipy import signal
33 | from scipy.ndimage.filters import convolve
34 |
35 |
36 | def _FSpecialGauss(size, sigma):
37 | """Function to mimic the 'fspecial' gaussian MATLAB function."""
38 | radius = size // 2
39 | offset = 0.0
40 | start, stop = -radius, radius + 1
41 | if size % 2 == 0:
42 | offset = 0.5
43 | stop -= 1
44 | x, y = np.mgrid[offset + start:stop, offset + start:stop]
45 | assert len(x) == size
46 | g = np.exp(-((x ** 2 + y ** 2) / (2.0 * sigma ** 2)))
47 | return g / g.sum()
48 |
49 |
50 | def _SSIMForMultiScale(img1, img2, max_val=255, filter_size=11,
51 | filter_sigma=1.5, k1=0.01, k2=0.03):
52 | """Return the Structural Similarity Map between `img1` and `img2`.
53 |
54 | This function attempts to match the functionality of ssim_index_new.m by
55 | Zhou Wang: http://www.cns.nyu.edu/~lcv/ssim/msssim.zip
56 |
57 | Arguments:
58 | img1: Numpy array holding the first RGB image batch.
59 | img2: Numpy array holding the second RGB image batch.
60 | max_val: the dynamic range of the images (i.e., the difference between the
61 | maximum the and minimum allowed values).
62 | filter_size: Size of blur kernel to use (will be reduced for small
63 | images).
64 | filter_sigma: Standard deviation for Gaussian blur kernel (will be reduced
65 | for small images).
66 | k1: Constant used to maintain stability in the SSIM calculation (0.01 in
67 | the original paper).
68 | k2: Constant used to maintain stability in the SSIM calculation (0.03 in
69 | the original paper).
70 |
71 | Returns:
72 | Pair containing the mean SSIM and contrast sensitivity between `img1` and
73 | `img2`.
74 |
75 | Raises:
76 | RuntimeError: If input images don't have the same shape or don't have four
77 | dimensions: [batch_size, height, width, depth].
78 | """
79 | if img1.shape != img2.shape:
80 | raise RuntimeError('Input images must have the same shape (%s vs. %s).',
81 | img1.shape, img2.shape)
82 | if img1.ndim != 4:
83 | raise RuntimeError('Input images must have four dimensions, not %d',
84 | img1.ndim)
85 |
86 | img1 = img1.astype(np.float64)
87 | img2 = img2.astype(np.float64)
88 | _, height, width, _ = img1.shape
89 |
90 | # Filter size can't be larger than height or width of images.
91 | size = min(filter_size, height, width)
92 |
93 | # Scale down sigma if a smaller filter size is used.
94 | sigma = size * filter_sigma / filter_size if filter_size else 0
95 |
96 | if filter_size:
97 | window = np.reshape(_FSpecialGauss(size, sigma), (1, size, size, 1))
98 | mu1 = signal.fftconvolve(img1, window, mode='valid')
99 | mu2 = signal.fftconvolve(img2, window, mode='valid')
100 | sigma11 = signal.fftconvolve(img1 * img1, window, mode='valid')
101 | sigma22 = signal.fftconvolve(img2 * img2, window, mode='valid')
102 | sigma12 = signal.fftconvolve(img1 * img2, window, mode='valid')
103 | else:
104 | # Empty blur kernel so no need to convolve.
105 | mu1, mu2 = img1, img2
106 | sigma11 = img1 * img1
107 | sigma22 = img2 * img2
108 | sigma12 = img1 * img2
109 |
110 | mu11 = mu1 * mu1
111 | mu22 = mu2 * mu2
112 | mu12 = mu1 * mu2
113 | sigma11 -= mu11
114 | sigma22 -= mu22
115 | sigma12 -= mu12
116 |
117 | # Calculate intermediate values used by both ssim and cs_map.
118 | c1 = (k1 * max_val) ** 2
119 | c2 = (k2 * max_val) ** 2
120 | v1 = 2.0 * sigma12 + c2
121 | v2 = sigma11 + sigma22 + c2
122 | ssim = np.mean((((2.0 * mu12 + c1) * v1) / ((mu11 + mu22 + c1) * v2)))
123 | cs = np.mean(v1 / v2)
124 | return ssim, cs
125 |
126 |
127 | def MultiScaleSSIM(img1, img2, max_val=255, filter_size=11, filter_sigma=1.5,
128 | k1=0.01, k2=0.03, weights=None):
129 | """Return the MS-SSIM score between `img1` and `img2`.
130 |
131 | This function implements Multi-Scale Structural Similarity (MS-SSIM) Image
132 | Quality Assessment according to Zhou Wang's paper, "Multi-scale structural
133 | similarity for image quality assessment" (2003).
134 | Link: https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf
135 |
136 | Author's MATLAB implementation:
137 | http://www.cns.nyu.edu/~lcv/ssim/msssim.zip
138 |
139 | Arguments:
140 | img1: Numpy array holding the first RGB image batch.
141 | img2: Numpy array holding the second RGB image batch.
142 | max_val: the dynamic range of the images (i.e., the difference between the
143 | maximum the and minimum allowed values).
144 | filter_size: Size of blur kernel to use (will be reduced for small
145 | images).
146 | filter_sigma: Standard deviation for Gaussian blur kernel (will be reduced
147 | for small images).
148 | k1: Constant used to maintain stability in the SSIM calculation (0.01 in
149 | the original paper).
150 | k2: Constant used to maintain stability in the SSIM calculation (0.03 in
151 | the original paper).
152 | weights: List of weights for each level; if none, use five levels and the
153 | weights from the original paper.
154 |
155 | Returns:
156 | MS-SSIM score between `img1` and `img2`.
157 |
158 | Raises:
159 | RuntimeError: If input images don't have the same shape or don't have four
160 | dimensions: [batch_size, height, width, depth].
161 | """
162 | if img1.shape != img2.shape:
163 | raise RuntimeError('Input images must have the same shape (%s vs. %s).',
164 | img1.shape, img2.shape)
165 | if img1.ndim != 4:
166 | raise RuntimeError('Input images must have four dimensions, not %d',
167 | img1.ndim)
168 |
169 | # Note: default weights don't sum to 1.0 but do match the paper / matlab
170 | # code.
171 | weights = np.array(weights if weights else
172 | [0.0448, 0.2856, 0.3001, 0.2363, 0.1333])
173 | levels = weights.size
174 | downsample_filter = np.ones((1, 2, 2, 1)) / 4.0
175 | im1, im2 = [x.astype(np.float64) for x in [img1, img2]]
176 | mssim = np.array([])
177 | mcs = np.array([])
178 | for _ in range(levels):
179 | ssim, cs = _SSIMForMultiScale(
180 | im1, im2, max_val=max_val, filter_size=filter_size,
181 | filter_sigma=filter_sigma, k1=k1, k2=k2)
182 | mssim = np.append(mssim, ssim)
183 | mcs = np.append(mcs, cs)
184 | filtered = [convolve(im, downsample_filter, mode='reflect')
185 | for im in [im1, im2]]
186 | im1, im2 = [x[:, ::2, ::2, :] for x in filtered]
187 | return (np.prod(mcs[0:levels - 1] ** weights[0:levels - 1]) *
188 | (mssim[levels - 1] ** weights[levels - 1]))
189 |
190 |
191 | def calculate_msssim(img_dir, gen_img_dir, caption_dir, output_dir):
192 |
193 | image_files = [f for f in os.listdir(img_dir) if 'jpg' in f]
194 | image_captions = {}
195 | image_classes = {}
196 | class_dirs = []
197 | class_names = []
198 | img_ids = []
199 | class_dict = {}
200 | gen_class_dict = {}
201 |
202 | print('Initializing objects for calculating MS-SSIM')
203 | for i in range(1, 103):
204 | class_dir_name = 'class_%.5d' % (i)
205 | class_dir = os.path.join(caption_dir, class_dir_name)
206 | class_names.append(class_dir_name)
207 | class_dirs.append(class_dir)
208 | onlyimgfiles = [f[0:11] + ".jpg" for f in os.listdir(class_dir)
209 | if 'txt' in f]
210 | for img_file in onlyimgfiles:
211 | image_classes[img_file] = None
212 |
213 | for img_file in onlyimgfiles:
214 | image_captions[img_file] = []
215 |
216 | for class_dir, class_name in zip(class_dirs, class_names):
217 | caption_files = [f for f in os.listdir(class_dir) if 'txt' in f]
218 | class_imgs = []
219 | gen_class_imgs = []
220 | for i, cap_file in enumerate(caption_files):
221 | if i % 50 == 0:
222 | print(str(i) + ' captions extracted from' + str(class_dir))
223 |
224 | class_imgs.append(cap_file[0:11] + ".jpg")
225 | image1_tr_path = os.path.join(gen_img_dir, 'train',
226 | cap_file[0:11] + ".jpg")
227 | if os.path.exists(image1_tr_path):
228 | for root, subFolders, files in os.walk(image1_tr_path):
229 | if files:
230 | for f in files:
231 | if 'jpg' in f:
232 | gen_class_imgs.append(os.path.join(root, f))
233 |
234 | class_dict[class_name] = class_imgs
235 | gen_class_dict[class_name] = gen_class_imgs
236 | with tf.Session() as sess:
237 | for class_name in class_dict.keys():
238 | img_list = class_dict[class_name]
239 | gen_img_list = gen_class_dict[class_name]
240 | real_msssim = []
241 | fake_msssim = []
242 | print('calculating MS-SSIM for real images of class : ' + str(
243 | class_name))
244 | for i in range(0, len(img_list)):
245 | for j in range(i, len(img_list)):
246 | if (i == j):
247 | continue
248 | image1_path = os.path.join(img_dir, img_list[i])
249 | image2_path = os.path.join(img_dir, img_list[j])
250 | with open(image1_path, 'rb') as image_file:
251 | img1_str = image_file.read()
252 | with open(image2_path, 'rb') as image_file:
253 | img2_str = image_file.read()
254 | input_img = tf.placeholder(tf.string)
255 | decoded_image = tf.expand_dims(
256 | tf.image.decode_png(input_img, channels=3), 0)
257 |
258 | img1 = np.squeeze(sess.run(decoded_image,
259 | feed_dict={input_img: img1_str}))
260 | img2 = np.squeeze(sess.run(decoded_image,
261 | feed_dict={input_img: img2_str}))
262 | img1 = resize(img1, (128, 128, 3), mode='reflect')
263 | img2 = resize(img2, (128, 128, 3), mode='reflect')
264 |
265 | img1 = np.expand_dims(img1, axis=0)
266 | img2 = np.expand_dims(img2, axis=0)
267 |
268 | real_msssim.append(MultiScaleSSIM(img1, img2, max_val=255))
269 |
270 | for i in range(0, len(gen_img_list)):
271 | for j in range(i, len(gen_img_list)):
272 | if (i == j):
273 | continue
274 | image1_path = os.path.join('', gen_img_list[i])
275 | image2_path = os.path.join('', gen_img_list[j])
276 | with open(image1_path, 'rb') as image_file:
277 | img1_str = image_file.read()
278 | with open(image2_path, 'rb') as image_file:
279 | img2_str = image_file.read()
280 | input_img = tf.placeholder(tf.string)
281 | decoded_image = tf.expand_dims(
282 | tf.image.decode_png(input_img, channels=3), 0)
283 | # with tf.Session() as sess:
284 | img1 = sess.run(decoded_image,
285 | feed_dict={input_img: img1_str})
286 | img2 = sess.run(decoded_image,
287 | feed_dict={input_img: img2_str})
288 | fake_msssim.append(MultiScaleSSIM(img1, img2, max_val=255))
289 |
290 | mean_real_msssim = np.mean(real_msssim)
291 | mean_fake_msssim = np.mean(fake_msssim)
292 |
293 | tsv_dir = os.path.join(output_dir, 'msssim')
294 | tsv_path = os.path.join(tsv_dir, 'msssim.tsv')
295 | if not os.path.exists(tsv_dir):
296 | os.makedirs(tsv_dir)
297 |
298 | if os.path.exists(tsv_path):
299 | os.remove(tsv_path)
300 |
301 | with open(tsv_path, 'a') as f:
302 | str_real_mean = "%.9f" % mean_real_msssim
303 | str_fake_mean = "%.9f" % mean_fake_msssim
304 | f.write(
305 | class_name + '\t' + str_real_mean + '\t' + str_fake_mean +
306 | '\n')
307 |
308 |
309 | if __name__ == '__main__':
310 | parser = argparse.ArgumentParser()
311 |
312 | parser.add_argument('--output_dir', type=str, default="Data/ms-ssim",
313 | help='directory to dump all the images for '
314 | 'calculating inception score')
315 |
316 | parser.add_argument('--data_dir', type=str, default="Data",
317 | help='Root directory of the data')
318 |
319 | parser.add_argument('--dataset', type=str, default="flowers",
320 | help='The root directory of the synthetic dataset')
321 |
322 | parser.add_argument('--syn_dataset_dir', type=str, default="flowers",
323 | help='The root directory of the synthetic dataset')
324 |
325 | args = parser.parse_args()
326 |
327 | if args.dataset != 'flowers':
328 | print('Dataset Not Found')
329 | sys.exit()
330 |
331 | img_dir = os.path.join(args.data_dir, 'datasets', args.dataset, 'jpg')
332 | gen_img_dir = args.syn_dataset_dir
333 | caption_dir = os.path.join(args.data_dir, 'datasets', 'flowers',
334 | 'text_c10')
335 |
336 | calculate_msssim(img_dir, gen_img_dir, caption_dir, args.output_dir)
337 |
--------------------------------------------------------------------------------
/requirements.txt:
--------------------------------------------------------------------------------
1 | backports-abc==0.5
2 | backports.weakref==1.0rc1
3 | bleach==1.5.0
4 | cycler==0.10.0
5 | decorator==4.0.11
6 | entrypoints==0.2.3
7 | html5lib==0.9999999
8 | ipykernel==4.6.1
9 | ipython==6.1.0
10 | ipython-genutils==0.2.0
11 | ipywidgets==6.0.0
12 | jedi==0.10.2
13 | Jinja2==2.9.6
14 | jsonschema==2.6.0
15 | jupyter==1.0.0
16 | jupyter-client==5.1.0
17 | jupyter-console==5.1.0
18 | jupyter-core==4.3.0
19 | Markdown==2.2.0
20 | MarkupSafe==1.0
21 | matplotlib==2.0.2
22 | mistune==0.7.4
23 | mpmath==0.19
24 | nbconvert==5.2.1
25 | nbformat==4.3.0
26 | nose==1.3.7
27 | notebook==5.0.0
28 | numpy==1.13.0
29 | pandas==0.20.2
30 | pandocfilters==1.4.1
31 | pexpect==4.2.1
32 | pickleshare==0.7.4
33 | progressbar2==3.30.2
34 | prompt-toolkit==1.0.14
35 | protobuf==3.3.0
36 | ptyprocess==0.5.2
37 | Pygments==2.2.0
38 | pyparsing==2.2.0
39 | python-dateutil==2.6.0
40 | python-utils==2.1.0
41 | pytz==2017.2
42 | pyzmq==16.0.2
43 | qtconsole==4.3.0
44 | scipy==0.19.1
45 | simplegeneric==0.8.1
46 | six==1.10.0
47 | sympy==1.0
48 | tensorflow-gpu==1.2.0
49 | terminado==0.6
50 | testpath==0.3.1
51 | Theano==0.9.0
52 | tornado==4.5.1
53 | traitlets==4.3.2
54 | typing==3.6.1
55 | wcwidth==0.1.7
56 | Werkzeug==0.12.2
57 | widgetsnbextension==2.0.0
58 |
--------------------------------------------------------------------------------
/skipthoughts.py:
--------------------------------------------------------------------------------
1 | '''
2 | Skip-thought vectors
3 | https://github.com/ryankiros/skip-thoughts
4 | '''
5 | import os
6 |
7 | import theano
8 | import theano.tensor as tensor
9 |
10 | import _pickle as pkl
11 | import numpy
12 | import copy
13 | import nltk
14 |
15 | from collections import OrderedDict, defaultdict
16 | from scipy.linalg import norm
17 | from nltk.tokenize import word_tokenize
18 |
19 | profile = False
20 |
21 | #-----------------------------------------------------------------------------#
22 | # Specify model and table locations here
23 | #-----------------------------------------------------------------------------#
24 | path_to_models = 'Data/skipthoughts/'
25 | path_to_tables = 'Data/skipthoughts/'
26 | #-----------------------------------------------------------------------------#
27 |
28 | path_to_umodel = path_to_models + 'uni_skip.npz'
29 | path_to_bmodel = path_to_models + 'bi_skip.npz'
30 |
31 |
32 | def load_model():
33 | """
34 | Load the model with saved tables
35 | """
36 | # Load model options
37 | print('Loading model parameters...')
38 | with open('%s.pkl'%path_to_umodel, 'rb') as f:
39 | uoptions = pkl.load(f)
40 | with open('%s.pkl'%path_to_bmodel, 'rb') as f:
41 | boptions = pkl.load(f)
42 |
43 | # Load parameters
44 | uparams = init_params(uoptions)
45 | uparams = load_params(path_to_umodel, uparams)
46 | utparams = init_tparams(uparams)
47 | bparams = init_params_bi(boptions)
48 | bparams = load_params(path_to_bmodel, bparams)
49 | btparams = init_tparams(bparams)
50 |
51 | # Extractor functions
52 | print('Compiling encoders...')
53 | embedding, x_mask, ctxw2v = build_encoder(utparams, uoptions)
54 | f_w2v = theano.function([embedding, x_mask], ctxw2v, name='f_w2v')
55 | embedding, x_mask, ctxw2v = build_encoder_bi(btparams, boptions)
56 | f_w2v2 = theano.function([embedding, x_mask], ctxw2v, name='f_w2v2')
57 |
58 | # Tables
59 | print('Loading tables...')
60 | utable, btable = load_tables()
61 |
62 | # Store everything we need in a dictionary
63 | print('Packing up...')
64 | model = {}
65 | model['uoptions'] = uoptions
66 | model['boptions'] = boptions
67 | model['utable'] = utable
68 | model['btable'] = btable
69 | model['f_w2v'] = f_w2v
70 | model['f_w2v2'] = f_w2v2
71 |
72 | return model
73 |
74 |
75 | def load_tables():
76 | """
77 | Load the tables
78 | """
79 | words = []
80 | utable = numpy.load(path_to_tables + 'utable.npy', encoding='bytes')
81 | btable = numpy.load(path_to_tables + 'btable.npy', encoding='bytes')
82 | f = open(path_to_tables + 'dictionary.txt', 'rb')
83 | for line in f:
84 | words.append(line.decode('utf-8').strip())
85 | f.close()
86 | utable = OrderedDict(zip(words, utable))
87 | btable = OrderedDict(zip(words, btable))
88 | return utable, btable
89 |
90 |
91 | def encode(model, X, use_norm=True, verbose=True, batch_size=128, use_eos=False):
92 | """
93 | Encode sentences in the list X. Each entry will return a vector
94 | """
95 | # first, do preprocessing
96 | X = preprocess(X)
97 |
98 | # word dictionary and init
99 | d = defaultdict(lambda : 0)
100 | for w in model['utable'].keys():
101 | d[w] = 1
102 | ufeatures = numpy.zeros((len(X), model['uoptions']['dim']), dtype='float32')
103 | bfeatures = numpy.zeros((len(X), 2 * model['boptions']['dim']), dtype='float32')
104 |
105 | # length dictionary
106 | ds = defaultdict(list)
107 | captions = [s.split() for s in X]
108 | for i,s in enumerate(captions):
109 | ds[len(s)].append(i)
110 |
111 | # Get features. This encodes by length, in order to avoid wasting computation
112 | for k in ds.keys():
113 | if verbose:
114 | print(k)
115 | numbatches = len(ds[k]) / batch_size + 1
116 | for minibatch in range(int(numbatches)):
117 | caps = ds[k][int(minibatch)::int(numbatches)]
118 |
119 | if use_eos:
120 | uembedding = numpy.zeros((k+1, len(caps), model['uoptions']['dim_word']), dtype='float32')
121 | bembedding = numpy.zeros((k+1, len(caps), model['boptions']['dim_word']), dtype='float32')
122 | else:
123 | uembedding = numpy.zeros((k, len(caps), model['uoptions']['dim_word']), dtype='float32')
124 | bembedding = numpy.zeros((k, len(caps), model['boptions']['dim_word']), dtype='float32')
125 | for ind, c in enumerate(caps):
126 | caption = captions[c]
127 | for j in range(len(caption)):
128 | if d[caption[j]] > 0:
129 | uembedding[j,ind] = model['utable'][caption[j]]
130 | bembedding[j,ind] = model['btable'][caption[j]]
131 | else:
132 | uembedding[j,ind] = model['utable']['UNK']
133 | bembedding[j,ind] = model['btable']['UNK']
134 | if use_eos:
135 | uembedding[-1,ind] = model['utable']['']
136 | bembedding[-1,ind] = model['btable']['']
137 | if use_eos:
138 | uff = model['f_w2v'](uembedding, numpy.ones((len(caption)+1,len(caps)), dtype='float32'))
139 | bff = model['f_w2v2'](bembedding, numpy.ones((len(caption)+1,len(caps)), dtype='float32'))
140 | else:
141 | uff = model['f_w2v'](uembedding, numpy.ones((len(caption),len(caps)), dtype='float32'))
142 | bff = model['f_w2v2'](bembedding, numpy.ones((len(caption),len(caps)), dtype='float32'))
143 | if use_norm:
144 | for j in range(len(uff)):
145 | uff[j] /= norm(uff[j])
146 | bff[j] /= norm(bff[j])
147 | for ind, c in enumerate(caps):
148 | ufeatures[c] = uff[ind]
149 | bfeatures[c] = bff[ind]
150 |
151 | features = numpy.c_[ufeatures, bfeatures]
152 | return features
153 |
154 |
155 | def preprocess(text):
156 | """
157 | Preprocess text for encoder
158 | """
159 | X = []
160 | sent_detector = nltk.data.load('tokenizers/punkt/english.pickle')
161 | for t in text:
162 | sents = sent_detector.tokenize(t)
163 | result = ''
164 | for s in sents:
165 | tokens = word_tokenize(s)
166 | result += ' ' + ' '.join(tokens)
167 | X.append(result)
168 | return X
169 |
170 |
171 | def nn(model, text, vectors, query, k=5):
172 | """
173 | Return the nearest neighbour sentences to query
174 | text: list of sentences
175 | vectors: the corresponding representations for text
176 | query: a string to search
177 | """
178 | qf = encode(model, [query])
179 | qf /= norm(qf)
180 | scores = numpy.dot(qf, vectors.T).flatten()
181 | sorted_args = numpy.argsort(scores)[::-1]
182 | sentences = [text[a] for a in sorted_args[:k]]
183 | print('QUERY: ' + query)
184 | print('NEAREST: ')
185 | for i, s in enumerate(sentences):
186 | print(s, sorted_args[i])
187 |
188 |
189 | def word_features(table):
190 | """
191 | Extract word features into a normalized matrix
192 | """
193 | features = numpy.zeros((len(table), 620), dtype='float32')
194 | keys = table.keys()
195 | for i in range(len(table)):
196 | f = table[keys[i]]
197 | features[i] = f / norm(f)
198 | return features
199 |
200 |
201 | def nn_words(table, wordvecs, query, k=10):
202 | """
203 | Get the nearest neighbour words
204 | """
205 | keys = table.keys()
206 | qf = table[query]
207 | scores = numpy.dot(qf, wordvecs.T).flatten()
208 | sorted_args = numpy.argsort(scores)[::-1]
209 | words = [keys[a] for a in sorted_args[:k]]
210 | print('QUERY: ' + query)
211 | print('NEAREST: ')
212 | for i, w in enumerate(words):
213 | print(w)
214 |
215 |
216 | def _p(pp, name):
217 | """
218 | make prefix-appended name
219 | """
220 | return '%s_%s'%(pp, name)
221 |
222 |
223 | def init_tparams(params):
224 | """
225 | initialize Theano shared variables according to the initial parameters
226 | """
227 | tparams = OrderedDict()
228 | for kk, pp in params.items():
229 | tparams[kk] = theano.shared(params[kk], name=kk)
230 | return tparams
231 |
232 |
233 | def load_params(path, params):
234 | """
235 | load parameters
236 | """
237 | pp = numpy.load(path)
238 | for kk, vv in params.items():
239 | if kk not in pp:
240 | warnings.warn('%s is not in the archive'%kk)
241 | continue
242 | params[kk] = pp[kk]
243 | return params
244 |
245 |
246 | # layers: 'name': ('parameter initializer', 'feedforward')
247 | layers = {'gru': ('param_init_gru', 'gru_layer')}
248 |
249 | def get_layer(name):
250 | fns = layers[name]
251 | return (eval(fns[0]), eval(fns[1]))
252 |
253 |
254 | def init_params(options):
255 | """
256 | initialize all parameters needed for the encoder
257 | """
258 | params = OrderedDict()
259 |
260 | # embedding
261 | params['Wemb'] = norm_weight(options['n_words_src'], options['dim_word'])
262 |
263 | # encoder: GRU
264 | params = get_layer(options['encoder'])[0](options, params, prefix='encoder',
265 | nin=options['dim_word'], dim=options['dim'])
266 | return params
267 |
268 |
269 | def init_params_bi(options):
270 | """
271 | initialize all paramters needed for bidirectional encoder
272 | """
273 | params = OrderedDict()
274 |
275 | # embedding
276 | params['Wemb'] = norm_weight(options['n_words_src'], options['dim_word'])
277 |
278 | # encoder: GRU
279 | params = get_layer(options['encoder'])[0](options, params, prefix='encoder',
280 | nin=options['dim_word'], dim=options['dim'])
281 | params = get_layer(options['encoder'])[0](options, params, prefix='encoder_r',
282 | nin=options['dim_word'], dim=options['dim'])
283 | return params
284 |
285 |
286 | def build_encoder(tparams, options):
287 | """
288 | build an encoder, given pre-computed word embeddings
289 | """
290 | # word embedding (source)
291 | embedding = tensor.tensor3('embedding', dtype='float32')
292 | x_mask = tensor.matrix('x_mask', dtype='float32')
293 |
294 | # encoder
295 | proj = get_layer(options['encoder'])[1](tparams, embedding, options,
296 | prefix='encoder',
297 | mask=x_mask)
298 | ctx = proj[0][-1]
299 |
300 | return embedding, x_mask, ctx
301 |
302 |
303 | def build_encoder_bi(tparams, options):
304 | """
305 | build bidirectional encoder, given pre-computed word embeddings
306 | """
307 | # word embedding (source)
308 | embedding = tensor.tensor3('embedding', dtype='float32')
309 | embeddingr = embedding[::-1]
310 | x_mask = tensor.matrix('x_mask', dtype='float32')
311 | xr_mask = x_mask[::-1]
312 |
313 | # encoder
314 | proj = get_layer(options['encoder'])[1](tparams, embedding, options,
315 | prefix='encoder',
316 | mask=x_mask)
317 | projr = get_layer(options['encoder'])[1](tparams, embeddingr, options,
318 | prefix='encoder_r',
319 | mask=xr_mask)
320 |
321 | ctx = tensor.concatenate([proj[0][-1], projr[0][-1]], axis=1)
322 |
323 | return embedding, x_mask, ctx
324 |
325 |
326 | # some utilities
327 | def ortho_weight(ndim):
328 | W = numpy.random.randn(ndim, ndim)
329 | u, s, v = numpy.linalg.svd(W)
330 | return u.astype('float32')
331 |
332 |
333 | def norm_weight(nin,nout=None, scale=0.1, ortho=True):
334 | if nout == None:
335 | nout = nin
336 | if nout == nin and ortho:
337 | W = ortho_weight(nin)
338 | else:
339 | W = numpy.random.uniform(low=-scale, high=scale, size=(nin, nout))
340 | return W.astype('float32')
341 |
342 |
343 | def param_init_gru(options, params, prefix='gru', nin=None, dim=None):
344 | """
345 | parameter init for GRU
346 | """
347 | if nin == None:
348 | nin = options['dim_proj']
349 | if dim == None:
350 | dim = options['dim_proj']
351 | W = numpy.concatenate([norm_weight(nin,dim),
352 | norm_weight(nin,dim)], axis=1)
353 | params[_p(prefix,'W')] = W
354 | params[_p(prefix,'b')] = numpy.zeros((2 * dim,)).astype('float32')
355 | U = numpy.concatenate([ortho_weight(dim),
356 | ortho_weight(dim)], axis=1)
357 | params[_p(prefix,'U')] = U
358 |
359 | Wx = norm_weight(nin, dim)
360 | params[_p(prefix,'Wx')] = Wx
361 | Ux = ortho_weight(dim)
362 | params[_p(prefix,'Ux')] = Ux
363 | params[_p(prefix,'bx')] = numpy.zeros((dim,)).astype('float32')
364 |
365 | return params
366 |
367 |
368 | def gru_layer(tparams, state_below, options, prefix='gru', mask=None, **kwargs):
369 | """
370 | Forward pass through GRU layer
371 | """
372 | nsteps = state_below.shape[0]
373 | if state_below.ndim == 3:
374 | n_samples = state_below.shape[1]
375 | else:
376 | n_samples = 1
377 |
378 | dim = tparams[_p(prefix,'Ux')].shape[1]
379 |
380 | if mask == None:
381 | mask = tensor.alloc(1., state_below.shape[0], 1)
382 |
383 | def _slice(_x, n, dim):
384 | if _x.ndim == 3:
385 | return _x[:, :, n*dim:(n+1)*dim]
386 | return _x[:, n*dim:(n+1)*dim]
387 |
388 | state_below_ = tensor.dot(state_below, tparams[_p(prefix, 'W')]) + tparams[_p(prefix, 'b')]
389 | state_belowx = tensor.dot(state_below, tparams[_p(prefix, 'Wx')]) + tparams[_p(prefix, 'bx')]
390 | U = tparams[_p(prefix, 'U')]
391 | Ux = tparams[_p(prefix, 'Ux')]
392 |
393 | def _step_slice(m_, x_, xx_, h_, U, Ux):
394 | preact = tensor.dot(h_, U)
395 | preact += x_
396 |
397 | r = tensor.nnet.sigmoid(_slice(preact, 0, dim))
398 | u = tensor.nnet.sigmoid(_slice(preact, 1, dim))
399 |
400 | preactx = tensor.dot(h_, Ux)
401 | preactx = preactx * r
402 | preactx = preactx + xx_
403 |
404 | h = tensor.tanh(preactx)
405 |
406 | h = u * h_ + (1. - u) * h
407 | h = m_[:,None] * h + (1. - m_)[:,None] * h_
408 |
409 | return h
410 |
411 | seqs = [mask, state_below_, state_belowx]
412 | _step = _step_slice
413 |
414 | rval, updates = theano.scan(_step,
415 | sequences=seqs,
416 | outputs_info = [tensor.alloc(0., n_samples, dim)],
417 | non_sequences = [tparams[_p(prefix, 'U')],
418 | tparams[_p(prefix, 'Ux')]],
419 | name=_p(prefix, '_layers'),
420 | n_steps=nsteps,
421 | profile=profile,
422 | strict=True)
423 | rval = [rval]
424 | return rval
425 |
426 |
--------------------------------------------------------------------------------
/t_interpolation.py:
--------------------------------------------------------------------------------
1 | import model
2 | import argparse
3 | import pickle
4 | import scipy.misc
5 | import random
6 | import os
7 | import progressbar
8 |
9 | import tensorflow as tf
10 | import numpy as np
11 |
12 | from os.path import join
13 |
14 |
15 | def main():
16 | parser = argparse.ArgumentParser()
17 | parser.add_argument('--z_dim', type=int, default=100,
18 | help='Noise dimension')
19 |
20 | parser.add_argument('--t_dim', type=int, default=256,
21 | help='Text feature dimension')
22 |
23 | parser.add_argument('--batch_size', type=int, default=64,
24 | help='Batch Size')
25 |
26 | parser.add_argument('--image_size', type=int, default=128,
27 | help='Image Size a, a x a')
28 |
29 | parser.add_argument('--gf_dim', type=int, default=64,
30 | help='Number of conv in the first layer gen.')
31 |
32 | parser.add_argument('--df_dim', type=int, default=64,
33 | help='Number of conv in the first layer discr.')
34 |
35 | parser.add_argument('--caption_vector_length', type=int, default=4800,
36 | help='Caption Vector Length')
37 |
38 | parser.add_argument('--n_classes', type=int, default=102,
39 | help='Number of classes/class labels')
40 |
41 | parser.add_argument('--data_dir', type=str, default="Data",
42 | help='Data Directory')
43 |
44 | parser.add_argument('--learning_rate', type=float, default=0.0002,
45 | help='Learning Rate')
46 |
47 | parser.add_argument('--beta1', type=float, default=0.5,
48 | help='Momentum for Adam Update')
49 |
50 | parser.add_argument('--data_set', type=str, default="flowers",
51 | help='Dats set: flowers')
52 |
53 | parser.add_argument('--output_dir', type=str, default="Data/synthetic_dataset",
54 | help='The directory in which the t_interpolated '
55 | 'images will be generated')
56 |
57 | parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
58 | help='Path to the checkpoints directory')
59 |
60 | parser.add_argument('--n_interp', type=int, default=100,
61 | help='The difference between each interpolation. '
62 | 'Should ideally be a multiple of 10')
63 |
64 | parser.add_argument('--n_images', type=int, default=500,
65 | help='Number of images to randomply sample for '
66 | 'generating interpolation results')
67 |
68 | args = parser.parse_args()
69 |
70 | datasets_root_dir = join(args.data_dir, 'datasets')
71 |
72 | loaded_data = load_training_data(datasets_root_dir, args.data_set,
73 | args.caption_vector_length,
74 | args.n_classes)
75 | model_options = {
76 | 'z_dim': args.z_dim,
77 | 't_dim': args.t_dim,
78 | 'batch_size': args.batch_size,
79 | 'image_size': args.image_size,
80 | 'gf_dim': args.gf_dim,
81 | 'df_dim': args.df_dim,
82 | 'caption_vector_length': args.caption_vector_length,
83 | 'n_classes': loaded_data['n_classes']
84 | }
85 |
86 | gan = model.GAN(model_options)
87 | input_tensors, variables, loss, outputs, checks = gan.build_model()
88 |
89 | sess = tf.InteractiveSession()
90 | tf.initialize_all_variables().run()
91 |
92 | saver = tf.train.Saver(max_to_keep=10000)
93 | print('resuming model from checkpoint' +
94 | str(tf.train.latest_checkpoint(args.checkpoints_dir)))
95 | if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
96 | saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
97 | print('Successfully loaded model')
98 | else:
99 | print('Could not load checkpoints')
100 | exit()
101 |
102 | random.shuffle(loaded_data['image_list'])
103 | selected_images = loaded_data['image_list'][:args.n_images]
104 | cap_id = [np.random.randint(0, 4) for cap_i in range(len(selected_images))]
105 |
106 | print('Generating Images by interpolating the text features and z')
107 | bar = progressbar.ProgressBar(redirect_stdout=True,
108 | max_value=args.n_images)
109 |
110 | for sel_i, (sel_img, sel_cap) in enumerate(zip(selected_images, cap_id)):
111 | for sel_j in range(sel_i, len(cap_id)):
112 | print(str(sel_i) + '\t->\t' + str(sel_j))
113 | if sel_i == sel_j:
114 | continue
115 | sel_img_2 = selected_images[sel_j]
116 | sel_cap_2 = cap_id[sel_j]
117 |
118 | captions_1, image_files_1, image_caps_1, image_ids_1, \
119 | image_caps_ids_1 = get_images_z_intr(sel_img, sel_cap,
120 | loaded_data, datasets_root_dir)
121 |
122 | captions_2, image_files_2, image_caps_2, image_ids_2, \
123 | image_caps_ids_2 = get_images_z_intr(sel_img_2, sel_cap_2,
124 | loaded_data, datasets_root_dir)
125 |
126 | z_noise_1 = np.random.uniform(-1, 1, [args.batch_size, args.z_dim])
127 | z_noise_2 = np.random.uniform(-1, 1, [args.batch_size, args.z_dim])
128 | intr_z_list = get_interp_vec(z_noise_1, z_noise_2, args.z_dim,
129 | args.n_interp, args.batch_size)
130 |
131 | intr_t_list = get_interp_vec(captions_1, captions_2,
132 | args.caption_vector_length,
133 | args.n_interp, args.batch_size)
134 |
135 | for z_i, z_noise in enumerate(intr_z_list):
136 | for t_i, captions in enumerate(intr_t_list):
137 | val_feed = {
138 | input_tensors['t_real_caption'].name: captions,
139 | input_tensors['t_z'].name: z_noise,
140 | input_tensors['t_training'].name: True
141 | }
142 | val_gen = sess.run([outputs['generator']],
143 | feed_dict=val_feed)
144 |
145 | save_distributed_image_batch(args.output_dir, val_gen,
146 | sel_i, sel_j, z_i, t_i, sel_img, sel_cap,
147 | sel_img_2, sel_cap_2, args.batch_size)
148 | bar.update(sel_i)
149 | bar.finish()
150 | print('Finished generating interpolated images')
151 |
152 | def load_training_data(data_dir, data_set, caption_vector_length, n_classes):
153 | if data_set == 'flowers':
154 | flower_str_captions = pickle.load(
155 | open(join(data_dir, 'flowers', 'flowers_caps.pkl'), "rb"))
156 |
157 | img_classes = pickle.load(
158 | open(join(data_dir, 'flowers', 'flower_tc.pkl'), "rb"))
159 |
160 | flower_enc_captions = pickle.load(
161 | open(join(data_dir, 'flowers', 'flower_tv.pkl'), "rb"))
162 | # h1 = h5py.File(join(data_dir, 'flower_tc.hdf5'))
163 | tr_image_ids = pickle.load(
164 | open(join(data_dir, 'flowers', 'train_ids.pkl'), "rb"))
165 | val_image_ids = pickle.load(
166 | open(join(data_dir, 'flowers', 'val_ids.pkl'), "rb"))
167 |
168 | # n_classes = n_classes
169 | max_caps_len = caption_vector_length
170 |
171 | tr_n_imgs = len(tr_image_ids)
172 | val_n_imgs = len(val_image_ids)
173 |
174 | return {
175 | 'image_list': tr_image_ids,
176 | 'captions': flower_enc_captions,
177 | 'data_length': tr_n_imgs,
178 | 'classes': img_classes,
179 | 'n_classes': n_classes,
180 | 'max_caps_len': max_caps_len,
181 | 'val_img_list': val_image_ids,
182 | 'val_captions': flower_enc_captions,
183 | 'val_data_len': val_n_imgs,
184 | 'str_captions': flower_str_captions
185 | }
186 |
187 | else:
188 | raise Exception('Dataset Not Found')
189 |
190 |
191 | def save_distributed_image_batch(data_dir, generated_images, sel_i, sel_2, z_i,
192 | t_i, sel_img, sel_cap, sel_img_2, sel_cap_2,
193 | batch_size):
194 |
195 | generated_images = np.squeeze(generated_images)
196 | folder_name = str(sel_i) + '_' + str(sel_2)
197 |
198 | image_dir = join(data_dir, 't_interpolation', folder_name, str(z_i))
199 | if not os.path.exists(image_dir):
200 | os.makedirs(image_dir)
201 |
202 | meta_path = os.path.join(image_dir, "meta.txt")
203 | with open(meta_path, "w") as text_file:
204 | text_file.write(str(sel_img) + "\t" + str(sel_cap) +
205 | str(sel_img_2) + "\t" + str(sel_cap_2))
206 | fake_image_255 = (generated_images[batch_size-1])
207 | scipy.misc.imsave(join(image_dir, '{}.jpg'.format(t_i)),
208 | fake_image_255)
209 |
210 |
211 | def get_images_z_intr(sel_img, sel_cap, loaded_data, data_dir, batch_size=64):
212 |
213 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
214 | batch_idx = np.random.randint(0, loaded_data['data_length'],
215 | size=batch_size - 1)
216 | image_ids = np.take(loaded_data['image_list'], batch_idx)
217 | image_files = []
218 | image_caps = []
219 | image_caps_ids = []
220 | for idx, image_id in enumerate(image_ids):
221 | image_file = join(data_dir,
222 | 'flowers/jpg/' + image_id)
223 | random_caption = random.randint(0, 4)
224 | image_caps_ids.append(random_caption)
225 | captions[idx, :] = \
226 | loaded_data['captions'][image_id][random_caption][
227 | 0:loaded_data['max_caps_len']]
228 | str_cap = loaded_data['str_captions'][image_id][random_caption]
229 | image_caps.append(loaded_data['captions']
230 | [image_id][random_caption])
231 | image_files.append(image_file)
232 | if idx == batch_size-2:
233 | idx = idx + 1
234 | image_id = sel_img
235 | image_file = join(data_dir,
236 | 'flowers/jpg/' + sel_img)
237 | random_caption = sel_cap
238 | image_caps_ids.append(random_caption)
239 | captions[idx, :] = \
240 | loaded_data['captions'][image_id][random_caption][
241 | 0:loaded_data['max_caps_len']]
242 | str_cap = loaded_data['str_captions'][image_id][random_caption]
243 | image_caps.append(loaded_data['str_captions']
244 | [image_id][random_caption])
245 | image_files.append(image_file)
246 | break
247 |
248 | return captions, image_files, image_caps, image_ids, image_caps_ids
249 |
250 |
251 | def get_interp_vec(vec_1, vec_2, dim, n_interp, batch_size):
252 |
253 | intrip_list = []
254 | bals = np.arange(0, 1, 1 / n_interp)
255 | for bal in bals:
256 | left = np.full((batch_size, dim), bal)
257 | right = np.full((batch_size, dim), 1.0 - bal)
258 | intrip_vec = np.multiply(vec_1, left) + np.multiply(vec_2, right)
259 | intrip_list.append(intrip_vec)
260 | return intrip_list
261 |
262 | if __name__ == '__main__':
263 | main()
264 |
--------------------------------------------------------------------------------
/train.py:
--------------------------------------------------------------------------------
1 | import model
2 | import argparse
3 | import pickle
4 | import scipy.misc
5 | import random
6 | import os
7 | import shutil
8 |
9 | import tensorflow as tf
10 | import numpy as np
11 |
12 | from os.path import join
13 | from Utils import image_processing
14 |
15 | def main():
16 | parser = argparse.ArgumentParser()
17 | parser.add_argument('--z_dim', type=int, default=100,
18 | help='Noise dimension')
19 |
20 | parser.add_argument('--t_dim', type=int, default=256,
21 | help='Text feature dimension')
22 |
23 | parser.add_argument('--batch_size', type=int, default=64,
24 | help='Batch Size')
25 |
26 | parser.add_argument('--image_size', type=int, default=128,
27 | help='Image Size a, a x a')
28 |
29 | parser.add_argument('--gf_dim', type=int, default=64,
30 | help='Number of conv in the first layer gen.')
31 |
32 | parser.add_argument('--df_dim', type=int, default=64,
33 | help='Number of conv in the first layer discr.')
34 |
35 | parser.add_argument('--caption_vector_length', type=int, default=4800,
36 | help='Caption Vector Length')
37 |
38 | parser.add_argument('--n_classes', type = int, default = 102,
39 | help = 'Number of classes/class labels')
40 |
41 | parser.add_argument('--data_dir', type=str, default="Data",
42 | help='Data Directory')
43 |
44 | parser.add_argument('--learning_rate', type=float, default=0.0002,
45 | help='Learning Rate')
46 |
47 | parser.add_argument('--beta1', type=float, default=0.5,
48 | help='Momentum for Adam Update')
49 |
50 | parser.add_argument('--epochs', type=int, default=200,
51 | help='Max number of epochs')
52 |
53 | parser.add_argument('--save_every', type=int, default=30,
54 | help='Save Model/Samples every x iterations over '
55 | 'batches')
56 |
57 | parser.add_argument('--resume_model', type=bool, default=False,
58 | help='Pre-Trained Model load or not')
59 |
60 | parser.add_argument('--data_set', type=str, default="flowers",
61 | help='Dat set: MS-COCO, flowers')
62 |
63 | parser.add_argument('--model_name', type=str, default="TAC_GAN",
64 | help='model_1 or model_2')
65 |
66 | parser.add_argument('--train', type = bool, default = True,
67 | help = 'True while training and otherwise')
68 |
69 | args = parser.parse_args()
70 |
71 | model_dir, model_chkpnts_dir, model_samples_dir, model_val_samples_dir,\
72 | model_summaries_dir = initialize_directories(args)
73 |
74 | datasets_root_dir = join(args.data_dir, 'datasets')
75 | loaded_data = load_training_data(datasets_root_dir, args.data_set,
76 | args.caption_vector_length,
77 | args.n_classes)
78 | model_options = {
79 | 'z_dim': args.z_dim,
80 | 't_dim': args.t_dim,
81 | 'batch_size': args.batch_size,
82 | 'image_size': args.image_size,
83 | 'gf_dim': args.gf_dim,
84 | 'df_dim': args.df_dim,
85 | 'caption_vector_length': args.caption_vector_length,
86 | 'n_classes': loaded_data['n_classes']
87 | }
88 |
89 | # Initialize and build the GAN model
90 | gan = model.GAN(model_options)
91 | input_tensors, variables, loss, outputs, checks = gan.build_model()
92 |
93 | d_optim = tf.train.AdamOptimizer(args.learning_rate,
94 | beta1=args.beta1).minimize(loss['d_loss'],
95 | var_list=variables['d_vars'])
96 | g_optim = tf.train.AdamOptimizer(args.learning_rate,
97 | beta1=args.beta1).minimize(loss['g_loss'],
98 | var_list=variables['g_vars'])
99 |
100 | global_step_tensor = tf.Variable(1, trainable=False, name='global_step')
101 | merged = tf.summary.merge_all()
102 | sess = tf.InteractiveSession()
103 |
104 | summary_writer = tf.summary.FileWriter(model_summaries_dir, sess.graph)
105 |
106 | tf.global_variables_initializer().run()
107 | saver = tf.train.Saver(max_to_keep=10000)
108 |
109 | if args.resume_model:
110 | print('Trying to resume training from a previous checkpoint' +
111 | str(tf.train.latest_checkpoint(model_chkpnts_dir)))
112 | if tf.train.latest_checkpoint(model_chkpnts_dir) is not None:
113 | saver.restore(sess, tf.train.latest_checkpoint(model_chkpnts_dir))
114 | print('Successfully loaded model. Resuming training.')
115 | else:
116 | print('Could not load checkpoints. Training a new model')
117 | global_step = global_step_tensor.eval()
118 | gs_assign_op = global_step_tensor.assign(global_step)
119 | for i in range(args.epochs):
120 | batch_no = 0
121 | while batch_no * args.batch_size + args.batch_size < \
122 | loaded_data['data_length']:
123 |
124 | real_images, wrong_images, caption_vectors, z_noise, image_files, \
125 | real_classes, wrong_classes, image_caps, image_ids = \
126 | get_training_batch(batch_no, args.batch_size,
127 | args.image_size, args.z_dim,
128 | 'train', datasets_root_dir,
129 | args.data_set, loaded_data)
130 |
131 | # DISCR UPDATE
132 | check_ts = [checks['d_loss1'], checks['d_loss2'],
133 | checks['d_loss3'], checks['d_loss1_1'],
134 | checks['d_loss2_1']]
135 |
136 | feed = {
137 | input_tensors['t_real_image'].name : real_images,
138 | input_tensors['t_wrong_image'].name : wrong_images,
139 | input_tensors['t_real_caption'].name : caption_vectors,
140 | input_tensors['t_z'].name : z_noise,
141 | input_tensors['t_real_classes'].name : real_classes,
142 | input_tensors['t_wrong_classes'].name : wrong_classes,
143 | input_tensors['t_training'].name : args.train
144 | }
145 |
146 | _, d_loss, gen, d1, d2, d3, d4, d5= sess.run([d_optim,
147 | loss['d_loss'],outputs['generator']] + check_ts,
148 | feed_dict=feed)
149 |
150 | print("D total loss: {}\n"
151 | "D loss-1 [Real/Fake loss for real images] : {} \n"
152 | "D loss-2 [Real/Fake loss for wrong images]: {} \n"
153 | "D loss-3 [Real/Fake loss for fake images]: {} \n"
154 | "D loss-4 [Aux Classifier loss for real images]: {} \n"
155 | "D loss-5 [Aux Classifier loss for wrong images]: {}"
156 | " ".format(d_loss, d1, d2, d3, d4, d5))
157 |
158 | # GEN UPDATE
159 | _, g_loss, gen = sess.run([g_optim, loss['g_loss'],
160 | outputs['generator']], feed_dict=feed)
161 |
162 | # GEN UPDATE TWICE
163 | _, summary, g_loss, gen, g1, g2 = sess.run([g_optim, merged,
164 | loss['g_loss'], outputs['generator'], checks['g_loss_1'],
165 | checks['g_loss_2']], feed_dict=feed)
166 | summary_writer.add_summary(summary, global_step)
167 | print("\n\nLOSSES\nDiscriminator Loss: {}\nGenerator Loss: {"
168 | "}\nBatch Number: {}\nEpoch: {},\nTotal Batches per "
169 | "epoch: {}\n".format( d_loss, g_loss, batch_no, i,
170 | int(len(loaded_data['image_list']) / args.batch_size)))
171 | print("\nG loss-1 [Real/Fake loss for fake images] : {} \n"
172 | "G loss-2 [Aux Classifier loss for fake images]: {} \n"
173 | " ".format(g1, g2))
174 | global_step += 1
175 | sess.run(gs_assign_op)
176 | batch_no += 1
177 | if (batch_no % args.save_every) == 0 and batch_no != 0:
178 | print("Saving Images and the Model\n\n")
179 |
180 | save_for_vis(model_samples_dir, real_images, gen, image_files,
181 | image_caps, image_ids)
182 | save_path = saver.save(sess,
183 | join(model_chkpnts_dir,
184 | "latest_model_{}_temp.ckpt".format(
185 | args.data_set)))
186 |
187 | # Getting a batch for validation
188 | val_captions, val_image_files, val_image_caps, val_image_ids = \
189 | get_val_caps_batch(args.batch_size, loaded_data,
190 | args.data_set, datasets_root_dir)
191 |
192 | shutil.rmtree(model_val_samples_dir)
193 | os.makedirs(model_val_samples_dir)
194 |
195 | for val_viz_cnt in range(0, 4):
196 | val_z_noise = np.random.uniform(-1, 1, [args.batch_size,
197 | args.z_dim])
198 |
199 | val_feed = {
200 | input_tensors['t_real_caption'].name : val_captions,
201 | input_tensors['t_z'].name : val_z_noise,
202 | input_tensors['t_training'].name : True
203 | }
204 |
205 | val_gen = sess.run([outputs['generator']],
206 | feed_dict=val_feed)
207 | save_for_viz_val(model_val_samples_dir, val_gen,
208 | val_image_files, val_image_caps,
209 | val_image_ids, args.image_size,
210 | val_viz_cnt)
211 |
212 | # Save the model after every epoch
213 | if i % 1 == 0:
214 | epoch_dir = join(model_chkpnts_dir, str(i))
215 | if not os.path.exists(epoch_dir):
216 | os.makedirs(epoch_dir)
217 |
218 | save_path = saver.save(sess,
219 | join(epoch_dir,
220 | "model_after_{}_epoch_{}.ckpt".
221 | format(args.data_set, i)))
222 | val_captions, val_image_files, val_image_caps, val_image_ids = \
223 | get_val_caps_batch(args.batch_size, loaded_data,
224 | args.data_set, datasets_root_dir)
225 |
226 | shutil.rmtree(model_val_samples_dir)
227 | os.makedirs(model_val_samples_dir)
228 |
229 | for val_viz_cnt in range(0, 10):
230 | val_z_noise = np.random.uniform(-1, 1, [args.batch_size,
231 | args.z_dim])
232 | val_feed = {
233 | input_tensors['t_real_caption'].name : val_captions,
234 | input_tensors['t_z'].name : val_z_noise,
235 | input_tensors['t_training'].name : True
236 | }
237 | val_gen = sess.run([outputs['generator']], feed_dict=val_feed)
238 | save_for_viz_val(model_val_samples_dir, val_gen,
239 | val_image_files, val_image_caps,
240 | val_image_ids, args.image_size,
241 | val_viz_cnt)
242 |
243 |
244 | def load_training_data(data_dir, data_set, caption_vector_length, n_classes) :
245 | if data_set == 'flowers' :
246 | flower_str_captions = pickle.load(
247 | open(join(data_dir, 'flowers', 'flowers_caps.pkl'), "rb"))
248 |
249 | img_classes = pickle.load(
250 | open(join(data_dir, 'flowers', 'flower_tc.pkl'), "rb"))
251 |
252 | flower_enc_captions = pickle.load(
253 | open(join(data_dir, 'flowers', 'flower_tv.pkl'), "rb"))
254 | tr_image_ids = pickle.load(
255 | open(join(data_dir, 'flowers', 'train_ids.pkl'), "rb"))
256 | val_image_ids = pickle.load(
257 | open(join(data_dir, 'flowers', 'val_ids.pkl'), "rb"))
258 |
259 | max_caps_len = caption_vector_length
260 | tr_n_imgs = len(tr_image_ids)
261 | val_n_imgs = len(val_image_ids)
262 |
263 | return {
264 | 'image_list' : tr_image_ids,
265 | 'captions' : flower_enc_captions,
266 | 'data_length' : tr_n_imgs,
267 | 'classes' : img_classes,
268 | 'n_classes' : n_classes,
269 | 'max_caps_len' : max_caps_len,
270 | 'val_img_list' : val_image_ids,
271 | 'val_captions' : flower_enc_captions,
272 | 'val_data_len' : val_n_imgs,
273 | 'str_captions' : flower_str_captions
274 | }
275 |
276 | else :
277 | raise Exception('No Dataset Found')
278 |
279 |
280 | def initialize_directories(args):
281 | model_dir = join(args.data_dir, 'training', args.model_name)
282 | if not os.path.exists(model_dir):
283 | os.makedirs(model_dir)
284 |
285 | model_chkpnts_dir = join(model_dir, 'checkpoints')
286 | if not os.path.exists(model_chkpnts_dir):
287 | os.makedirs(model_chkpnts_dir)
288 |
289 | model_summaries_dir = join(model_dir, 'summaries')
290 | if not os.path.exists(model_summaries_dir):
291 | os.makedirs(model_summaries_dir)
292 |
293 | model_samples_dir = join(model_dir, 'samples')
294 | if not os.path.exists(model_samples_dir):
295 | os.makedirs(model_samples_dir)
296 |
297 | model_val_samples_dir = join(model_dir, 'val_samples')
298 | if not os.path.exists(model_val_samples_dir):
299 | os.makedirs(model_val_samples_dir)
300 |
301 | return model_dir, model_chkpnts_dir, model_samples_dir, \
302 | model_val_samples_dir, model_summaries_dir
303 |
304 |
305 | def save_for_viz_val(data_dir, generated_images, image_files, image_caps,
306 | image_ids, image_size, id):
307 |
308 | generated_images = np.squeeze(np.array(generated_images))
309 | for i in range(0, generated_images.shape[0]) :
310 | image_dir = join(data_dir, str(image_ids[i]))
311 | if not os.path.exists(image_dir):
312 | os.makedirs(image_dir)
313 |
314 | real_image_path = join(image_dir,
315 | '{}.jpg'.format(image_ids[i]))
316 | if os.path.exists(image_dir):
317 | real_images_255 = image_processing.load_image_array(image_files[i],
318 | image_size, image_ids[i], mode='val')
319 | scipy.misc.imsave(real_image_path, real_images_255)
320 |
321 | caps_dir = join(image_dir, "caps.txt")
322 | if not os.path.exists(caps_dir):
323 | with open(caps_dir, "w") as text_file:
324 | text_file.write(image_caps[i]+"\n")
325 |
326 | fake_images_255 = generated_images[i]
327 | scipy.misc.imsave(join(image_dir, 'fake_image_{}.jpg'.format(id)),
328 | fake_images_255)
329 |
330 |
331 | def save_for_vis(data_dir, real_images, generated_images, image_files,
332 | image_caps, image_ids) :
333 |
334 | shutil.rmtree(data_dir)
335 | os.makedirs(data_dir)
336 |
337 | for i in range(0, real_images.shape[0]) :
338 | real_images_255 = (real_images[i, :, :, :])
339 | scipy.misc.imsave(join(data_dir,
340 | '{}_{}.jpg'.format(i, image_files[i].split('/')[-1])),
341 | real_images_255)
342 |
343 | fake_images_255 = (generated_images[i, :, :, :])
344 | scipy.misc.imsave(join(data_dir, 'fake_image_{}.jpg'.format(
345 | i)), fake_images_255)
346 |
347 | str_caps = '\n'.join(image_caps)
348 | str_image_ids = '\n'.join([str(image_id) for image_id in image_ids])
349 | with open(join(data_dir, "caps.txt"), "w") as text_file:
350 | text_file.write(str_caps)
351 | with open(join(data_dir, "ids.txt"), "w") as text_file:
352 | text_file.write(str_image_ids)
353 |
354 |
355 | def get_val_caps_batch(batch_size, loaded_data, data_set, data_dir):
356 |
357 | if data_set == 'flowers':
358 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
359 |
360 | batch_idx = np.random.randint(0, loaded_data['val_data_len'],
361 | size = batch_size)
362 | image_ids = np.take(loaded_data['val_img_list'], batch_idx)
363 | image_files = []
364 | image_caps = []
365 | for idx, image_id in enumerate(image_ids) :
366 | image_file = join(data_dir,
367 | 'flowers/jpg/' + image_id)
368 | random_caption = random.randint(0, 4)
369 | captions[idx, :] = \
370 | loaded_data['val_captions'][image_id][random_caption][
371 | 0 :loaded_data['max_caps_len']]
372 |
373 | image_caps.append(loaded_data['str_captions']
374 | [image_id][random_caption])
375 | image_files.append(image_file)
376 |
377 | return captions, image_files, image_caps, image_ids
378 | else:
379 | raise Exception('Dataset not found')
380 |
381 |
382 | def get_training_batch(batch_no, batch_size, image_size, z_dim, split,
383 | data_dir, data_set, loaded_data = None) :
384 | if data_set == 'flowers':
385 | real_images = np.zeros((batch_size, image_size, image_size, 3))
386 | wrong_images = np.zeros((batch_size, image_size, image_size, 3))
387 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
388 | real_classes = np.zeros((batch_size, loaded_data['n_classes']))
389 | wrong_classes = np.zeros((batch_size, loaded_data['n_classes']))
390 |
391 | cnt = 0
392 | image_files = []
393 | image_caps = []
394 | image_ids = []
395 | for i in range(batch_no * batch_size,
396 | batch_no * batch_size + batch_size) :
397 | idx = i % len(loaded_data['image_list'])
398 | image_file = join(data_dir,
399 | 'flowers/jpg/' + loaded_data['image_list'][idx])
400 |
401 | image_ids.append(loaded_data['image_list'][idx])
402 |
403 | image_array = image_processing.load_image_array_flowers(image_file,
404 | image_size)
405 | real_images[cnt, :, :, :] = image_array
406 |
407 | # Improve this selection of wrong image
408 | wrong_image_id = random.randint(0,
409 | len(loaded_data['image_list']) - 1)
410 | wrong_image_file = join(data_dir,
411 | 'flowers/jpg/' + loaded_data['image_list'][
412 | wrong_image_id])
413 | wrong_image_array = image_processing.load_image_array_flowers(wrong_image_file,
414 | image_size)
415 | wrong_images[cnt, :, :, :] = wrong_image_array
416 |
417 | wrong_classes[cnt, :] = loaded_data['classes'][loaded_data['image_list'][
418 | wrong_image_id]][0 :loaded_data['n_classes']]
419 |
420 | random_caption = random.randint(0, 4)
421 | captions[cnt, :] = \
422 | loaded_data['captions'][loaded_data['image_list'][idx]][
423 | random_caption][0 :loaded_data['max_caps_len']]
424 |
425 | real_classes[cnt, :] = \
426 | loaded_data['classes'][loaded_data['image_list'][idx]][
427 | 0 :loaded_data['n_classes']]
428 | str_cap = loaded_data['str_captions'][loaded_data['image_list']
429 | [idx]][random_caption]
430 |
431 | image_files.append(image_file)
432 | image_caps.append(str_cap)
433 | cnt += 1
434 |
435 | z_noise = np.random.uniform(-1, 1, [batch_size, z_dim])
436 | return real_images, wrong_images, captions, z_noise, image_files, \
437 | real_classes, wrong_classes, image_caps, image_ids
438 | else:
439 | raise Exception('Dataset not found')
440 |
441 |
442 | if __name__ == '__main__' :
443 | main()
444 |
--------------------------------------------------------------------------------
/z_interpolation.py:
--------------------------------------------------------------------------------
1 | import model
2 | import argparse
3 | import pickle
4 | import scipy.misc
5 | import random
6 | import os
7 | import progressbar
8 |
9 | import tensorflow as tf
10 | import numpy as np
11 |
12 | from os.path import join
13 |
14 | def main():
15 | parser = argparse.ArgumentParser()
16 | parser.add_argument('--z_dim', type=int, default=100,
17 | help='Noise dimension')
18 |
19 | parser.add_argument('--t_dim', type=int, default=256,
20 | help='Text feature dimension')
21 |
22 | parser.add_argument('--batch_size', type=int, default=64,
23 | help='Batch Size')
24 |
25 | parser.add_argument('--image_size', type=int, default=128,
26 | help='Image Size a, a x a')
27 |
28 | parser.add_argument('--gf_dim', type=int, default=64,
29 | help='Number of conv in the first layer gen.')
30 |
31 | parser.add_argument('--df_dim', type=int, default=64,
32 | help='Number of conv in the first layer discr.')
33 |
34 | parser.add_argument('--caption_vector_length', type=int, default=4800,
35 | help='Caption Vector Length')
36 |
37 | parser.add_argument('--n_classes', type=int, default=102,
38 | help='Number of classes/class labels')
39 |
40 | parser.add_argument('--data_dir', type=str, default="Data",
41 | help='Data Directory')
42 |
43 | parser.add_argument('--learning_rate', type=float, default=0.0002,
44 | help='Learning Rate')
45 |
46 | parser.add_argument('--beta1', type=float, default=0.5,
47 | help='Momentum for Adam Update')
48 |
49 | parser.add_argument('--data_set', type=str, default="flowers",
50 | help='Dat set: flowers')
51 |
52 | parser.add_argument('--output_dir', type=str,
53 | default="Data/synthetic_dataset",
54 | help='The directory in which this dataset will be '
55 | 'created')
56 |
57 | parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
58 | help='Path to the checkpoints directory')
59 |
60 | parser.add_argument('--n_interp', type=int, default=100,
61 | help='The difference between each interpolation. '
62 | 'Should ideally be a multiple of 10')
63 |
64 | parser.add_argument('--n_images', type=int, default=500,
65 | help='Number of images to randomply sample for '
66 | 'generating interpolation results')
67 |
68 | args = parser.parse_args()
69 | datasets_root_dir = join(args.data_dir, 'datasets')
70 |
71 | loaded_data = load_training_data(datasets_root_dir, args.data_set,
72 | args.caption_vector_length, args.n_classes)
73 |
74 | model_options = {
75 | 'z_dim': args.z_dim,
76 | 't_dim': args.t_dim,
77 | 'batch_size': args.batch_size,
78 | 'image_size': args.image_size,
79 | 'gf_dim': args.gf_dim,
80 | 'df_dim': args.df_dim,
81 | 'caption_vector_length': args.caption_vector_length,
82 | 'n_classes': loaded_data['n_classes']
83 | }
84 |
85 | gan = model.GAN(model_options)
86 | input_tensors, variables, loss, outputs, checks = gan.build_model()
87 |
88 | sess = tf.InteractiveSession()
89 | tf.initialize_all_variables().run()
90 |
91 | saver = tf.train.Saver(max_to_keep=10000)
92 | print('resuming model from checkpoint' +
93 | str(tf.train.latest_checkpoint(args.checkpoints_dir)))
94 | if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
95 | saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
96 | print('Successfully loaded model')
97 | else:
98 | print('Could not load checkpoints')
99 | exit()
100 |
101 | random.shuffle(loaded_data['image_list'])
102 | selected_images = loaded_data['image_list'][:args.n_images]
103 | cap_id = [np.random.randint(0, 4) for cap_i in range(len(selected_images))]
104 |
105 | print('Generating Images by interpolating z')
106 | bar = progressbar.ProgressBar(redirect_stdout=True,
107 | max_value=args.n_images)
108 | for sel_i, (sel_img, sel_cap) in enumerate(zip(selected_images, cap_id)):
109 | captions, image_files, image_caps, image_ids, image_caps_ids = \
110 | get_images_z_intr(sel_img, sel_cap, loaded_data,
111 | datasets_root_dir, args.batch_size)
112 |
113 | z_noise_1 = np.full((args.batch_size, args.z_dim), -1.0)
114 | z_noise_2 = np.full((args.batch_size, args.z_dim), 1.0)
115 | intr_z_list = get_interp_vec(z_noise_1, z_noise_2, args.z_dim,
116 | args.n_interp, args.batch_size)
117 |
118 | for z_i, z_noise in enumerate(intr_z_list):
119 | val_feed = {
120 | input_tensors['t_real_caption'].name: captions,
121 | input_tensors['t_z'].name: z_noise,
122 | input_tensors['t_training'].name: True
123 | }
124 |
125 | val_gen = sess.run([outputs['generator']], feed_dict=val_feed)
126 |
127 | save_distributed_image_batch(args.output_dir, val_gen, sel_i, z_i,
128 | sel_img, sel_cap, args.batch_size)
129 | bar.update(sel_i)
130 | bar.finish()
131 | print('Finished generating interpolated images')
132 |
133 |
134 | def load_training_data(data_dir, data_set, caption_vector_length, n_classes):
135 | if data_set == 'flowers':
136 | flower_str_captions = pickle.load(
137 | open(join(data_dir, 'flowers', 'flowers_caps.pkl'), "rb"))
138 |
139 | img_classes = pickle.load(
140 | open(join(data_dir, 'flowers', 'flower_tc.pkl'), "rb"))
141 |
142 | flower_enc_captions = pickle.load(
143 | open(join(data_dir, 'flowers', 'flower_tv.pkl'), "rb"))
144 | # h1 = h5py.File(join(data_dir, 'flower_tc.hdf5'))
145 | tr_image_ids = pickle.load(
146 | open(join(data_dir, 'flowers', 'train_ids.pkl'), "rb"))
147 | val_image_ids = pickle.load(
148 | open(join(data_dir, 'flowers', 'val_ids.pkl'), "rb"))
149 |
150 | max_caps_len = caption_vector_length
151 |
152 | tr_n_imgs = len(tr_image_ids)
153 | val_n_imgs = len(val_image_ids)
154 |
155 | return {
156 | 'image_list': tr_image_ids,
157 | 'captions': flower_enc_captions,
158 | 'data_length': tr_n_imgs,
159 | 'classes': img_classes,
160 | 'n_classes': n_classes,
161 | 'max_caps_len': max_caps_len,
162 | 'val_img_list': val_image_ids,
163 | 'val_captions': flower_enc_captions,
164 | 'val_data_len': val_n_imgs,
165 | 'str_captions': flower_str_captions
166 | }
167 |
168 | else:
169 | raise Exception('Dataset Not Found!!')
170 |
171 | def save_distributed_image_batch(data_dir, generated_images, sel_i, z_i,
172 | sel_img, sel_cap, batch_size):
173 |
174 | generated_images = np.squeeze(generated_images)
175 | image_dir = join(data_dir, 'z_interpolation', str(sel_i))
176 | if not os.path.exists(image_dir):
177 | os.makedirs(image_dir)
178 | meta_path = os.path.join(image_dir, "meta.txt")
179 | with open(meta_path, "w") as text_file:
180 | text_file.write(str(sel_img) + "\t" + str(sel_cap))
181 | fake_image_255 = generated_images[batch_size - 1]
182 | scipy.misc.imsave(join(image_dir, '{}.jpg'.format(z_i)),
183 | fake_image_255)
184 |
185 |
186 | def get_images_z_intr(sel_img, sel_cap, loaded_data, data_dir, batch_size=64):
187 |
188 | captions = np.zeros((batch_size, loaded_data['max_caps_len']))
189 | batch_idx = np.random.randint(0, loaded_data['data_length'],
190 | size = batch_size-1)
191 |
192 | image_ids = np.take(loaded_data['image_list'], batch_idx)
193 | image_files = []
194 | image_caps = []
195 | image_caps_ids = []
196 |
197 | for idx, image_id in enumerate(image_ids):
198 | image_file = join(data_dir,
199 | 'flowers/jpg/' + image_id)
200 | random_caption = random.randint(0, 4)
201 | image_caps_ids.append(random_caption)
202 | captions[idx, :] = \
203 | loaded_data['captions'][image_id][random_caption][
204 | 0:loaded_data['max_caps_len']]
205 | str_cap = loaded_data['str_captions'][image_id][random_caption]
206 |
207 | image_caps.append(loaded_data['captions']
208 | [image_id][random_caption])
209 | image_files.append(image_file)
210 | if idx == batch_size-2:
211 | idx = idx+1
212 | image_id = sel_img
213 | image_file = join(data_dir,
214 | 'flowers/jpg/' + sel_img)
215 | random_caption = sel_cap
216 | image_caps_ids.append(random_caption)
217 | captions[idx, :] = \
218 | loaded_data['captions'][image_id][random_caption][
219 | 0:loaded_data['max_caps_len']]
220 | str_cap = loaded_data['str_captions'][image_id][random_caption]
221 | image_caps.append(loaded_data['str_captions']
222 | [image_id][random_caption])
223 | image_files.append(image_file)
224 | break
225 |
226 | return captions, image_files, image_caps, image_ids, image_caps_ids
227 |
228 |
229 | def get_interp_vec(vec_1, vec_2, dim, n_interp, batch_size):
230 |
231 | intrip_list = []
232 | bals = np.arange(0, 1, 1/n_interp)
233 | for bal in bals:
234 | left = np.full((batch_size, dim), bal)
235 | right = np.full((batch_size, dim), 1.0 - bal)
236 | intrip_vec = np.multiply(vec_1, left) + np.multiply(vec_2, right)
237 | intrip_list.append(intrip_vec)
238 | return intrip_list
239 |
240 |
241 | if __name__ == '__main__':
242 | main()
243 |
--------------------------------------------------------------------------------