├── .gitignore ├── LICENSE ├── README.md ├── images ├── atypical.png └── typical.png ├── main.py ├── models ├── build.py └── train.py ├── packages.py └── toolkits ├── evaluations.py ├── utils.py └── visualizations.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | *.egg-info/ 24 | .installed.cfg 25 | *.egg 26 | MANIFEST 27 | 28 | # PyInstaller 29 | # Usually these files are written by a python script from a template 30 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 31 | *.manifest 32 | *.spec 33 | 34 | # Installer logs 35 | pip-log.txt 36 | pip-delete-this-directory.txt 37 | 38 | # Unit test / coverage reports 39 | htmlcov/ 40 | .tox/ 41 | .coverage 42 | .coverage.* 43 | .cache 44 | nosetests.xml 45 | coverage.xml 46 | *.cover 47 | .hypothesis/ 48 | .pytest_cache/ 49 | 50 | # Translations 51 | *.mo 52 | *.pot 53 | 54 | # Django stuff: 55 | *.log 56 | local_settings.py 57 | db.sqlite3 58 | 59 | # Flask stuff: 60 | instance/ 61 | .webassets-cache 62 | 63 | # Scrapy stuff: 64 | .scrapy 65 | 66 | # Sphinx documentation 67 | docs/_build/ 68 | 69 | # PyBuilder 70 | target/ 71 | 72 | # Jupyter Notebook 73 | .ipynb_checkpoints 74 | 75 | # pyenv 76 | .python-version 77 | 78 | # celery beat schedule file 79 | celerybeat-schedule 80 | 81 | # SageMath parsed files 82 | *.sage.py 83 | 84 | # Environments 85 | .env 86 | .venv 87 | env/ 88 | venv/ 89 | ENV/ 90 | env.bak/ 91 | venv.bak/ 92 | 93 | # Spyder project settings 94 | .spyderproject 95 | .spyproject 96 | 97 | # Rope project settings 98 | .ropeproject 99 | 100 | # mkdocs documentation 101 | /site 102 | 103 | # mypy 104 | .mypy_cache/ 105 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Deep One-Class Classification Using Intra-Class Splitting 2 | 3 | This repository contains an implementation of a one-class classification method using deep learning. The method was proposed in our "Deep One-Class Classification Using Intra-Class Splitting" paper presented at the IEEE Data Science Workshop 2019. It is based on intra-class splitting, i.e. splitting given normal samples into typical and atypical subsets: 4 | 5 |

6 | 7 | 8 | 9 |

10 | 11 | ## Paper 12 | [Deep One-Class Classification Using Intra-Class Splitting](https://arxiv.org/abs/1902.01194) 13 | by Patrick Schlachter, Yiwen Liao and Bin Yang 14 | Institute of Signal Processing and System Theory, University of Stuttgart, Germany 15 | IEEE Data Science Workshop 2019 in Minneapolis, Minnesota 16 | 17 | If you use this work for your research, please cite our paper: 18 | ``` 19 | @inproceedings{schlachter2019, 20 | author={Patrick Schlachter, Yiwen Liao and Bin Yang}, 21 | booktitle={2019 IEEE Data Science Workshop (DSW)}, 22 | title={Deep One-Class Classification Using Intra-Class Splitting}, 23 | year={2019}, 24 | month={June}, 25 | } 26 | ``` 27 | 28 | ## Repository 29 | ### `models` 30 | Contains build and train functions of the underlying neural network models. 31 | 32 | ### `toolkits` 33 | Contains evaluation, visualization and util functions. 34 | 35 | ### `main.py` 36 | The main function to start training and evaluation. 37 | 38 | ### `packages.py` 39 | Imports necessary Python packages. 40 | -------------------------------------------------------------------------------- /images/atypical.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/patrickschlachter/deep-occ-using-ics/c09bdd248eebf565a615bde86ae350611f085afe/images/atypical.png -------------------------------------------------------------------------------- /images/typical.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/patrickschlachter/deep-occ-using-ics/c09bdd248eebf565a615bde86ae350611f085afe/images/typical.png -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 18/8/2 12:09 4 | # @Author : Yiwen Liao 5 | # @File : main.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | 11 | from models.train import * 12 | from toolkits.utils import set_seed 13 | 14 | 15 | def run_model(dataset=None, normal_class=None): 16 | """Run ICSNET models 17 | 18 | :param dataset: Name of a desired dataset: mnist, fmnist or cifar10. 19 | :param normal_class: An integer value standing for the desired known class. 20 | :return: None 21 | """ 22 | 23 | set_seed() 24 | 25 | data = get_data(dataset=dataset, normal_class=normal_class, data_format='tensor') 26 | name = dataset + '_%d' % normal_class 27 | 28 | train_autoencoder(data=data, 29 | epoch=50, 30 | batch_size=64, 31 | reg=1e-5, 32 | latent_fea=128, 33 | name=name) 34 | 35 | train_icsnet(data=data, 36 | thr=10, 37 | epoch=800, 38 | batch_size=64, 39 | reg=1e-3, 40 | latent_fea=128, 41 | name=name) 42 | 43 | 44 | if __name__ == '__main__': 45 | run_model('mnist', 6) 46 | -------------------------------------------------------------------------------- /models/build.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 18/7/31 10:06 4 | # @Author : Yiwen Liao 5 | # @File : build.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | 11 | from useful_packages import * 12 | 13 | 14 | # ==================== ICSNET ==================== 15 | def build_icsnet(img_shape=None, reg=None, latent_fea=None): 16 | """Create ICSNET model. 17 | 18 | :param img_shape: A tuple standing for image shape, i.e. (height, width, # channels). 19 | :param reg: Float value for weight decay. 20 | :param latent_fea: A desired integer number of latent features. 21 | :return: An ICSNET model and a joint-ICSNET model. 22 | """ 23 | # ==================== Constants Definition ==================== 24 | acti_func = 'linear' 25 | clf_acti = 'sigmoid' 26 | 27 | acti_alpha = 0.2 28 | set_bias = False 29 | 30 | weights_init = tn(mean=0, stddev=0.01) 31 | 32 | bn_eps = 1e-3 33 | bn_m = 0.99 34 | 35 | # ==================== ICSNET ==================== 36 | 37 | input_layer = Input(shape=img_shape, name='input_layer') 38 | 39 | conv_1 = Conv2D(filters=16, kernel_size=(3, 3), activation=acti_func, name='conv_1', 40 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 41 | kernel_initializer=weights_init)(input_layer) 42 | conv_11 = Conv2D(filters=16, kernel_size=(3, 3), activation=acti_func, name='conv_11', 43 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 44 | kernel_initializer=weights_init)(conv_1) 45 | conv_1 = Concatenate()([conv_1, conv_11]) # 32x32x64 46 | lrelu_1 = LeakyReLU(alpha=acti_alpha)(conv_1) 47 | pool_1 = AveragePooling2D(pool_size=(2, 2), name='pool_1')(lrelu_1) # 16x16 / 14x14 48 | bn_1 = BatchNormalization(momentum=bn_m, epsilon=bn_eps, name='bn_1')(pool_1) 49 | 50 | conv_2 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='conv_2', 51 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 52 | kernel_initializer=weights_init)(bn_1) 53 | conv_22 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='conv_22', 54 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 55 | kernel_initializer=weights_init)(conv_2) 56 | conv_2 = Concatenate()([conv_2, conv_22]) # 16x16x128 57 | lrelu_2 = LeakyReLU(alpha=acti_alpha)(conv_2) 58 | pool_2 = AveragePooling2D(pool_size=(2, 2), name='pool_2')(lrelu_2) # 8x8 / 7x7 59 | if img_shape[0] == 28: 60 | pool_2 = ZeroPadding2D(padding=(1, 1))(pool_2) # zero-padding if mnist or fashion-mnist 61 | bn_2 = BatchNormalization(momentum=bn_m, epsilon=bn_eps, name='bn_2')(pool_2) 62 | 63 | conv_3 = Conv2D(filters=64, kernel_size=(3, 3), activation=acti_func, name='conv_3', 64 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 65 | kernel_initializer=weights_init)(bn_2) 66 | conv_33 = Conv2D(filters=64, kernel_size=(3, 3), activation=acti_func, name='conv_33', 67 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 68 | kernel_initializer=weights_init)(conv_3) 69 | conv_3 = Concatenate()([conv_3, conv_33]) # 8x8x256 70 | lrelu_3 = LeakyReLU(alpha=acti_alpha)(conv_3) 71 | pool_3 = AveragePooling2D(pool_size=(2, 2), name='pool_3')(lrelu_3) # 4x4 72 | bn_3 = BatchNormalization(momentum=bn_m, epsilon=bn_eps, name='bn_3')(pool_3) 73 | 74 | conv_4 = Conv2D(filters=128, kernel_size=(3, 3), activation=acti_func, name='conv_4', 75 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 76 | kernel_initializer=weights_init)(bn_3) 77 | conv_44 = Conv2D(filters=128, kernel_size=(3, 3), activation=acti_func, name='conv_44', 78 | padding='same', kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 79 | kernel_initializer=weights_init)(conv_4) 80 | conv_4 = Concatenate()([conv_4, conv_44]) # 4x4x512 81 | lrelu_4 = LeakyReLU(alpha=acti_alpha)(conv_4) 82 | pool_4 = AveragePooling2D(pool_size=(2, 2), name='pool_4')(lrelu_4) # 2x2 83 | bn_4 = BatchNormalization(momentum=bn_m, epsilon=bn_eps, name='bn_4')(pool_4) 84 | 85 | conv_5 = Conv2D(filters=256, kernel_size=(1, 1), activation=acti_func, name='conv_5', 86 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 87 | kernel_initializer=weights_init)(bn_4) # 2x2 88 | conv_55 = Conv2D(filters=256, kernel_size=(1, 1), activation=acti_func, name='conv_55', 89 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias, 90 | kernel_initializer=weights_init)(bn_4) 91 | conv_5 = Concatenate()([conv_5, conv_55]) # 2x2x512 92 | lrelu_5 = LeakyReLU(alpha=acti_alpha)(conv_5) 93 | bn_5 = BatchNormalization(name='bn_5')(lrelu_5) 94 | 95 | flt_6 = Flatten()(bn_5) 96 | 97 | dense_7 = Dense(units=256, activation=acti_func, name='dense_7', 98 | kernel_regularizer=regularizers.l2(reg), use_bias=not set_bias, 99 | kernel_initializer=weights_init)(flt_6) 100 | lrelu_7 = LeakyReLU(alpha=acti_alpha)(dense_7) 101 | drop_7 = Dropout(rate=0.5)(lrelu_7) 102 | 103 | dense_8 = Dense(units=latent_fea, activation=acti_func, name='dense_8', 104 | kernel_regularizer=regularizers.l2(reg), use_bias=not set_bias, 105 | kernel_initializer=weights_init)(drop_7) 106 | 107 | lrelu_8 = LeakyReLU(alpha=acti_alpha)(dense_8) 108 | drop_8 = Dropout(rate=0.5)(lrelu_8) 109 | 110 | dense_9 = Dense(units=1, activation='sigmoid', name='dense_9', 111 | kernel_regularizer=regularizers.l2(reg), use_bias=not set_bias, 112 | kernel_initializer=weights_init)(drop_8) 113 | 114 | output_layer = Reshape(target_shape=(-1,), name='top_layer')(dense_9) 115 | latent_layer = Reshape(target_shape=(-1,), name='latent_layer')(lrelu_8) 116 | 117 | icsnet = Model(inputs=input_layer, outputs=output_layer, name='ICSNET') 118 | icsnet_latent = Model(inputs=input_layer, outputs=latent_layer, name='ICSNET_latent') 119 | 120 | # ==================== Subnetwork ==================== 121 | 122 | input_1 = Input(shape=img_shape, name='input_1') 123 | input_2 = Input(shape=img_shape, name='input_2') 124 | 125 | lat_1 = icsnet_latent(input_1) 126 | lat_2 = icsnet_latent(input_2) 127 | 128 | latent_dist = Subtract(name='latent_dist')([lat_1, lat_2]) 129 | 130 | dense_ly = Dense(units=1, activation=clf_acti, name='dense_ly', 131 | kernel_regularizer=regularizers.l2(reg))(latent_dist) 132 | 133 | ic_network = Model(inputs=[input_1, input_2], outputs=dense_ly) 134 | 135 | # ==================== Multiple Input Layers ==================== 136 | 137 | typical_input_1 = Input(shape=img_shape, name='typical_input_1') 138 | typical_input_2 = Input(shape=img_shape, name='typical_input_2') 139 | atypical_input_1 = Input(shape=img_shape, name='atypical_input_1') 140 | atypical_input_2 = Input(shape=img_shape, name='atypical_input_2') 141 | 142 | typical_dist = ic_network([typical_input_1, typical_input_2]) 143 | atypical_dist = ic_network([atypical_input_1, atypical_input_2]) 144 | 145 | # ==================== Final Model ==================== 146 | 147 | icsnet_joint = Model(inputs=[typical_input_1, typical_input_2, atypical_input_1, atypical_input_2], 148 | outputs=[typical_dist, atypical_dist], name='ICSNET_joint') 149 | icsnet.summary() 150 | 151 | return icsnet, icsnet_joint 152 | 153 | 154 | # ==================== Autoencoders ==================== 155 | def build_ae(img_shape=None, reg=None, latent_fea=None): 156 | """Create autoencoder models. 157 | 158 | :param img_shape: A tuple standing for image shape, i.e. (height, width, # channels). 159 | :param reg: Float value for weight decay. 160 | :param latent_fea: A desired integer number of latent features. 161 | :return: An autoencoder model. 162 | """ 163 | # ==================== Constants Definition ==================== 164 | acti_func = 'relu' 165 | set_bias = False 166 | weights_init = tn(mean=0, stddev=0.01) 167 | 168 | # ==================== Encoder ==================== 169 | 170 | input_layer = Input(shape=img_shape, name='input_layer') 171 | 172 | if img_shape[0] == 28: 173 | pad_layer = ZeroPadding2D(padding=(2, 2))(input_layer) 174 | else: 175 | pad_layer = input_layer 176 | 177 | conv_1 = Conv2D(filters=8, kernel_size=(3, 3), activation=acti_func, name='conv_1', 178 | kernel_initializer=weights_init, padding='same', 179 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(pad_layer) # 32x32 180 | conv_11 = Conv2D(filters=8, kernel_size=(3, 3), activation=acti_func, name='conv_11', 181 | kernel_initializer=weights_init, padding='same', 182 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(conv_1) 183 | conv_1 = Concatenate()([conv_11, conv_1]) 184 | 185 | pool_1 = MaxPooling2D(pool_size=(2, 2), name='pool_1')(conv_1) # 16x16 186 | 187 | bn_1 = BatchNormalization(name='bn_1')(pool_1) 188 | 189 | conv_2 = Conv2D(filters=16, kernel_size=(3, 3), activation=acti_func, name='conv_2', 190 | kernel_initializer=weights_init, padding='same', 191 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_1) 192 | conv_22 = Conv2D(filters=16, kernel_size=(3, 3), activation=acti_func, name='conv_22', 193 | kernel_initializer=weights_init, padding='same', 194 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(conv_2) 195 | conv_2 = Concatenate()([conv_22, conv_2]) 196 | 197 | pool_2 = MaxPooling2D(pool_size=(2, 2), name='pool_2')(conv_2) # 8x8 198 | 199 | bn_2 = BatchNormalization(name='bn_2')(pool_2) 200 | 201 | conv_3 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='conv_3', 202 | kernel_initializer=weights_init, padding='same', 203 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_2) 204 | conv_33 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='conv_33', 205 | kernel_initializer=weights_init, padding='same', 206 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(conv_3) 207 | conv_3 = Concatenate()([conv_33, conv_3]) 208 | 209 | pool_3 = MaxPooling2D(pool_size=(2, 2), name='pool_3')(conv_3) # 4x4 210 | 211 | bn_3 = BatchNormalization(name='bn_3')(pool_3) 212 | 213 | conv_4 = Conv2D(filters=64, kernel_size=(3, 3), activation=acti_func, name='conv_4', 214 | kernel_regularizer=regularizers.l2(reg), padding='same', 215 | kernel_initializer=weights_init, use_bias=set_bias)(bn_3) 216 | conv_44 = Conv2D(filters=64, kernel_size=(3, 3), activation=acti_func, name='conv_44', 217 | kernel_regularizer=regularizers.l2(reg), padding='same', 218 | kernel_initializer=weights_init, use_bias=set_bias)(conv_4) 219 | conv_4 = Concatenate()([conv_44, conv_4]) 220 | 221 | pool_4 = MaxPooling2D(pool_size=(2, 2), name='pool_4')(conv_4) # 2x2 222 | 223 | bn_4 = BatchNormalization(name='bn_4')(pool_4) 224 | 225 | conv_lat = Conv2D(filters=latent_fea, kernel_size=(1, 1), activation=acti_func, name='conv_lat', 226 | kernel_regularizer=regularizers.l2(reg), padding='same', 227 | kernel_initializer=weights_init, use_bias=not set_bias)(bn_4) 228 | 229 | # ==================== Decoder ==================== 230 | 231 | bn_5 = BatchNormalization(name='bn_lat')(conv_lat) 232 | 233 | convt_5 = Conv2DTranspose(filters=32, kernel_size=(3, 3), activation=acti_func, name='convt_5', 234 | kernel_initializer=weights_init, padding='same', strides=(2, 2), 235 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_5) 236 | convt_55 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='convt_55', 237 | kernel_regularizer=regularizers.l2(reg), padding='same', 238 | kernel_initializer=weights_init, use_bias=set_bias)(convt_5) 239 | convt_5 = Concatenate()([convt_55, convt_5]) 240 | 241 | bn_5 = BatchNormalization(name='bn_5')(convt_5) # 4x4 242 | 243 | convt_6 = Conv2DTranspose(filters=32, kernel_size=(3, 3), activation=acti_func, name='convt_6', 244 | padding='same', kernel_initializer=weights_init, strides=(2, 2), 245 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_5) 246 | convt_66 = Conv2D(filters=32, kernel_size=(3, 3), activation=acti_func, name='convt_66', 247 | kernel_regularizer=regularizers.l2(reg), padding='same', 248 | kernel_initializer=weights_init, use_bias=set_bias)(convt_6) 249 | convt_6 = Concatenate()([convt_66, convt_6]) 250 | 251 | bn_6 = BatchNormalization(name='bn_6')(convt_6) # 8x8 252 | 253 | convt_7 = Conv2DTranspose(filters=16, kernel_size=(3, 3), activation=acti_func, name='convt_7', 254 | padding='same', kernel_initializer=weights_init, strides=(2, 2), 255 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_6) 256 | convt_77 = Conv2D(filters=16, kernel_size=(3, 3), activation=acti_func, name='convt_77', 257 | kernel_regularizer=regularizers.l2(reg), padding='same', 258 | kernel_initializer=weights_init, use_bias=set_bias)(convt_7) 259 | convt_7 = Concatenate()([convt_77, convt_7]) 260 | 261 | bn_7 = BatchNormalization(name='bn_7')(convt_7) # 16x16 262 | 263 | convt_8 = Conv2DTranspose(filters=8, kernel_size=(3, 3), activation=acti_func, name='convt_8', 264 | padding='same', kernel_initializer=weights_init, strides=(2, 2), 265 | kernel_regularizer=regularizers.l2(reg), use_bias=set_bias)(bn_7) 266 | convt_88 = Conv2D(filters=8, kernel_size=(3, 3), activation=acti_func, name='convt_88', 267 | kernel_regularizer=regularizers.l2(reg), padding='same', 268 | kernel_initializer=weights_init, use_bias=set_bias)(convt_8) 269 | convt_8 = Concatenate()([convt_88, convt_8]) 270 | 271 | bn_8 = BatchNormalization(name='bn_8')(convt_8) # 32x32 272 | 273 | conv_9 = Conv2D(filters=img_shape[-1], kernel_size=(3, 3), activation='sigmoid', name='conv_9', 274 | kernel_initializer=weights_init, padding='same', 275 | kernel_regularizer=regularizers.l2(reg), use_bias=not set_bias)(bn_8) 276 | 277 | if img_shape[0] == 28: 278 | conv_9 = Cropping2D(cropping=((2, 2), (2, 2)))(conv_9) 279 | else: 280 | pass 281 | 282 | ae = Model(inputs=input_layer, outputs=conv_9, name='autoencoder') 283 | ae.summary() 284 | 285 | return ae 286 | -------------------------------------------------------------------------------- /models/train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 18/8/2 10:57 4 | # @Author : Yiwen Liao 5 | # @File : train.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | 11 | from useful_packages import * 12 | from toolkits.utils import * 13 | from toolkits.evaluations import * 14 | from toolkits.visualizations import * 15 | from .build import * 16 | 17 | 18 | INIT_SEED = 2018 19 | RANDOM_STATE = np.random.RandomState(INIT_SEED) 20 | RECORD_STEP = 100 21 | 22 | 23 | def train_icsnet(data=None, thr=None, epoch=None, batch_size=None, reg=None, latent_fea=None, name=''): 24 | """Train an ICSNET. 25 | 26 | :param data: A dictionary containing the training data. 27 | :param thr: Splitting ratio between 1 and 99. 28 | :param epoch: Number of desired training epochs. 29 | :param batch_size: Desired number of batch size. 30 | :param reg: Float value for weight decay. 31 | :param latent_fea: A desired integer number of latent features. 32 | :param name: Name for saving models and results. 33 | :return: None 34 | """ 35 | 36 | # ==================== Training data ==================== 37 | x_train = data['x_train_normal'] 38 | img_shape = x_train.shape[1:] 39 | num_train = x_train.shape[0] 40 | 41 | # ==================== Split the datasets ==================== 42 | print('\nLoading the trained autoencoder...') 43 | autoencoder_path = './trained_models/ae_%s.h5' % name 44 | autoencoder = load_model(autoencoder_path, compile=False) 45 | 46 | typical_index, atypical_index = split_data(model=autoencoder, data=x_train, tau=thr, split_method='ssim') 47 | x_train_typical = x_train[typical_index] 48 | x_train_atypical = x_train[atypical_index] 49 | 50 | y_train = np.ones(shape=(num_train,)) 51 | y_train[atypical_index] = 0 52 | y_ty = np.zeros(shape=(batch_size,)) 53 | y_aty = np.ones(shape=(batch_size,)) 54 | y_train_joint = np.ones(shape=(4*batch_size,)) 55 | y_train_joint[2*batch_size:] = 0 56 | 57 | # ==================== Build ICS-classifier ==================== 58 | customized_optimizer = optimizers.adam(lr=3e-4, beta_1=0.5, decay=1e-8) 59 | model_set = build_icsnet(img_shape=img_shape, reg=reg, latent_fea=latent_fea) 60 | 61 | icsnet = model_set[0] 62 | icsnet_joint = model_set[1] 63 | 64 | icsnet.compile(optimizer=customized_optimizer, loss='binary_crossentropy') 65 | icsnet_joint.compile(optimizer=customized_optimizer, loss=['binary_crossentropy', 'binary_crossentropy'], 66 | loss_weights=[1., 1.], metrics=['accuracy']) 67 | 68 | # ==================== Train ICS-classifier ==================== 69 | idx_ty = np.random.permutation(np.arange(0, len(x_train_typical))) 70 | idx_aty = np.random.permutation(np.arange(0, len(x_train_atypical))) 71 | idx_batch = np.random.permutation(np.arange(0, 4 * batch_size)) 72 | 73 | baccu_val = [] 74 | baccu_test = [] 75 | valid_res_ref = 0 76 | best_test_res = 0 77 | for e in range(epoch): 78 | 79 | if (e+1) % RECORD_STEP == 0 or e == 0: 80 | print('\nTraining for epoch ' + str(e + 1) + '...') 81 | y_test_normal = icsnet.predict(data['x_test_normal'], batch_size=128) 82 | y_test_abnormal = icsnet.predict(data['x_test_abnormal'], batch_size=128) 83 | 84 | valid_baccu, test_baccu = one_class_evaluation(df_normal=y_test_normal, df_abnormal=y_test_abnormal) 85 | 86 | print('\nCurrent validation AUC: %.4f' % valid_baccu) 87 | print('Current test AUC: %.4f' % test_baccu) 88 | baccu_val.append(valid_baccu) 89 | baccu_test.append(test_baccu) 90 | 91 | if valid_baccu > valid_res_ref: 92 | icsnet.save('./trained_models/icsnet_%s_thr_%d_best.h5' % (name, thr)) 93 | valid_res_ref = valid_baccu 94 | best_test_res = test_baccu 95 | 96 | print('\nBest valid AUC till now: %.4f' % valid_res_ref) 97 | print('Best test AUC till now: %.4f' % best_test_res) 98 | 99 | x_ty = x_train_typical[idx_ty][:batch_size] 100 | x_aty = x_train_atypical[idx_aty][:batch_size] 101 | 102 | x_ref_ty = x_train_typical[idx_ty][batch_size:2 * batch_size] 103 | x_ref_aty = x_train_atypical[idx_aty][batch_size:2 * batch_size] 104 | 105 | icsnet_joint.fit(x=[x_ty, x_ref_ty, x_aty, x_ref_aty], 106 | y=[y_ty, y_aty], 107 | epochs=3, 108 | verbose=0, 109 | batch_size=batch_size) 110 | 111 | x_train_batch = np.vstack([x_ty, x_ref_ty, x_aty, x_ref_aty]) 112 | x_train_batch = x_train_batch[idx_batch] 113 | y_train_batch = y_train_joint[idx_batch] 114 | 115 | if (e + 1) % 1 == 0: 116 | icsnet.fit(x=x_train_batch, y=y_train_batch, batch_size=batch_size, epochs=1, verbose=0) 117 | 118 | np.random.shuffle(idx_ty) 119 | np.random.shuffle(idx_aty) 120 | np.random.shuffle(idx_batch) 121 | 122 | baccu_test = np.asarray(baccu_test) 123 | baccu_val = np.asarray(baccu_val) 124 | 125 | # Plot the balanced accuracy vs. epochs 126 | plt.figure(figsize=(15, 10)) 127 | plt.plot(np.arange(0, len(baccu_test)) * RECORD_STEP, 128 | baccu_test, label='BACCU Test: Last score is %.4f.\nBest Score is %.4f' % (baccu_test[-1], 129 | np.max(baccu_test))) 130 | plt.scatter(np.arange(0, len(baccu_test)) * RECORD_STEP, baccu_test) 131 | plt.plot(np.arange(0, len(baccu_val)) * RECORD_STEP, 132 | baccu_val, label='BACCU Valid: Last score is %.4f.\nBest Score is %.4f' % (baccu_val[-1], 133 | np.max(baccu_val))) 134 | plt.scatter(np.arange(0, len(baccu_val)) * RECORD_STEP, baccu_val) 135 | 136 | plt.xlabel('Epoch') 137 | plt.ylabel('Scores') 138 | plt.legend() 139 | plt.savefig('AUC_epoch_%s_thr_%d.png' % (name, thr)) 140 | plt.close() 141 | 142 | 143 | def train_autoencoder(data=None, epoch=None, batch_size=None, reg=None, latent_fea=None, name=''): 144 | """Training autoencoder for data splitting. 145 | 146 | :param data: A dictionary containing the training data. 147 | :param epoch: Number of desired training epochs. 148 | :param batch_size: Desired number of batch size. 149 | :param reg: Float value for weight decay. 150 | :param latent_fea: A desired integer number of latent features. 151 | :param name: Name for saving models and results. 152 | :return: None 153 | """ 154 | 155 | x_train = data['x_train_normal'] 156 | img_shape = x_train.shape[1:] 157 | 158 | ae = build_ae(img_shape=img_shape, reg=reg, latent_fea=latent_fea) 159 | customized_optimizer = optimizers.rmsprop(lr=1e-3, decay=1e-8) 160 | ae.compile(optimizer=customized_optimizer, loss='mse') 161 | 162 | for e in range(epoch): 163 | 164 | print('Training for epoch %d...' % e) 165 | ae.save('./trained_models/ae_%s.h5' % name) 166 | 167 | np.random.shuffle(data['x_train_normal']) 168 | 169 | ae.fit(x=data['x_train_normal'], 170 | y=data['x_train_normal'], 171 | batch_size=batch_size, 172 | epochs=1, 173 | verbose=1) 174 | 175 | ae.save('./trained_models/ae_%s.h5' % name) 176 | 177 | pred = ae.predict(data['x_train_normal'][:100]) 178 | img_visualize(data=pred, num_to_show=10, to_save=True, name='rec_train_normal_%s' % name) 179 | 180 | pred = ae.predict(data['x_test_normal'][:100]) 181 | img_visualize(data=pred, num_to_show=10, to_save=True, name='rec_test_normal_%s' % name) 182 | 183 | pred = ae.predict(data['x_test_abnormal'][:100]) 184 | img_visualize(data=pred, num_to_show=10, to_save=True, name='rec_test_abnormal_%s' % name) -------------------------------------------------------------------------------- /packages.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 2018/7/30 19:17 4 | # @Author : Yiwen Liao 5 | # @File : packages.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | import math, gc, os, glob 11 | import time as t 12 | import random as rn 13 | 14 | import numpy as np 15 | import seaborn as sns 16 | import tensorflow as tf 17 | import keras.backend as K 18 | import matplotlib.pyplot as plt 19 | 20 | from PIL import Image 21 | from matplotlib2tikz import save as tikz_save 22 | from skimage.measure import compare_ssim, compare_psnr 23 | from skimage.feature import hog 24 | from sklearn import svm, datasets 25 | from sklearn.ensemble import IsolationForest as iForest 26 | from sklearn.manifold import TSNE 27 | from sklearn.decomposition import PCA 28 | from sklearn.metrics import confusion_matrix, roc_curve, auc, roc_auc_score, accuracy_score 29 | 30 | from keras import losses, metrics, optimizers, regularizers 31 | from keras.datasets import fashion_mnist, mnist, cifar10, cifar100 32 | from keras.models import Model, load_model, Sequential 33 | from keras.preprocessing.image import ImageDataGenerator 34 | from keras.layers import Dense, Input, Add, Lambda, Reshape, Concatenate, Subtract, Multiply, ZeroPadding2D, Conv2DTranspose 35 | from keras.layers import LeakyReLU, BatchNormalization, Conv2D, UpSampling2D, Flatten, MaxPooling2D, Cropping2D, Dropout 36 | from keras.layers import AveragePooling2D, Activation, GlobalAveragePooling2D, PReLU 37 | from keras.utils import plot_model, to_categorical 38 | from keras.initializers import truncated_normal as tn 39 | from keras.backend.tensorflow_backend import set_session 40 | -------------------------------------------------------------------------------- /toolkits/evaluations.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 2018/8/4 22:40 4 | # @Author : Yiwen Liao 5 | # @File : evaluations.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | 11 | import numpy as np 12 | from sklearn.metrics import roc_auc_score 13 | 14 | 15 | def one_class_evaluation(df_normal=None, df_abnormal=None): 16 | """Calculate AUC on test set. 17 | 18 | :param df_normal: Decision functions of normal samples in a test set. 19 | :param df_abnormal: Decision functions of abnormal samples in a test set. 20 | :return: Validation AUC and test AUC 21 | """ 22 | 23 | num_valid_normal = int(0.3 * len(df_normal)) 24 | num_valid_abnormal = int(0.3 * len(df_abnormal)) 25 | 26 | df_normal_valid = df_normal[:num_valid_normal, ...] 27 | df_abnormal_valid = df_abnormal[:num_valid_abnormal, ...] 28 | 29 | df_normal_test = df_normal[num_valid_normal:, ...] 30 | df_abnormal_test = df_abnormal[num_valid_abnormal:, ...] 31 | 32 | valid_label = np.vstack([np.ones(shape=(num_valid_normal, 1)), np.zeros(shape=(num_valid_abnormal, 1))]) 33 | valid_df = np.concatenate([df_normal_valid, df_abnormal_valid]) 34 | 35 | test_label = np.concatenate([np.ones_like(df_normal_test), np.zeros_like(df_abnormal_test)]) 36 | test_df = np.concatenate([df_normal_test, df_abnormal_test]) 37 | 38 | auc_valid = roc_auc_score(valid_label, valid_df) 39 | auc_test = roc_auc_score(test_label, test_df) 40 | 41 | return auc_valid, auc_test 42 | -------------------------------------------------------------------------------- /toolkits/utils.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 2018/7/30 19:18 4 | # @Author : Yiwen Liao 5 | # @File : utils.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | 11 | from useful_packages import * 12 | 13 | 14 | # ==================== Data Preprocessing ==================== 15 | def _extract_data(data=None, label=None, target_lb=None): 16 | """Extract dataset regarding given normal / abnormal labels- 17 | 18 | :param data: A numpy tensor. First axis should be the number of samples. 19 | :param label: The corresponding labels for the data. 20 | :param target_lb: An integer value standing for the only one known class. 21 | :return: normal data, abnormal data, normal labels, abnormal labels 22 | """ 23 | 24 | if isinstance(target_lb, int): 25 | idx_normal = np.where(label == target_lb)[0] 26 | idx_abnormal = np.where(label != target_lb)[0] 27 | else: 28 | raise ValueError('Target label should be a integer...') 29 | 30 | data_normal = data[idx_normal] 31 | data_abnormal = data[idx_abnormal] 32 | label_normal = label[idx_normal] 33 | label_abnormal = label[idx_abnormal] 34 | 35 | return data_normal, data_abnormal, label_normal, label_abnormal 36 | 37 | 38 | def _reshape_data(data=None, data_shape=None, num_channels=None): 39 | """Reshape image data into vectors / matrices / tensors. 40 | 41 | :param data: A numpy tensor. First axis should be the number of samples. 42 | :param data_shape: Desired data shape. It should be a string. 43 | :param num_channels: Number of the channels of the given data. 44 | :return: Reshaped data. 45 | """ 46 | 47 | num_samples = data.shape[0] 48 | data = data.reshape(num_samples, -1) 49 | num_features = data.shape[-1] 50 | height = int(np.sqrt(num_features / num_channels)) 51 | width = height 52 | 53 | if data_shape == 'vector': 54 | print('Images data are transformed into vectors...') 55 | elif data_shape == 'matrix': 56 | if num_channels == 1: 57 | data = data.reshape(num_samples, height, width) 58 | elif num_channels == 3: 59 | data = data.reshape(num_samples, height, width, num_channels) 60 | data = 0.2989 * data[:, :, :, 0] + 0.5870 * data[:, :, :, 1] + 0.1140 * data[:, :, :, 2] 61 | data = data.reshape(num_samples, height, width) 62 | else: 63 | raise ValueError('The input data is neither gray-scale images nor color images. Please choose other data...') 64 | elif data_shape == 'tensor': 65 | data = data.reshape(num_samples, height, width, num_channels) 66 | else: 67 | raise ValueError('No suitable data shape is found. Please enter a desired data shape...') 68 | 69 | return data 70 | 71 | 72 | def get_data(dataset=None, normal_class=None, data_format=None): 73 | """Obtain the dataset in a desired form stored in a dictionary. 74 | 75 | :param dataset: The name of desired dataset: mnist, fmnist or cifar10. 76 | :param normal_class: The class which is considered to be known during training. 77 | :param data_format: The desired data shape: vector, matrix or tensor. 78 | :return: A dictionary containing training and testing samples. 79 | """ 80 | 81 | if dataset == 'mnist': 82 | (x_train, y_train), (x_test, y_test) = mnist.load_data() 83 | num_channel = 1 84 | elif dataset == 'fmnist': 85 | (x_train, y_train), (x_test, y_test) = fashion_mnist.load_data() 86 | num_channel = 1 87 | elif dataset == 'cifar10': 88 | (x_train, y_train), (x_test, y_test) = cifar10.load_data() 89 | num_channel = 3 90 | else: 91 | raise ValueError('The dataset %s is not found...' % dataset) 92 | 93 | # Reshape data and its label into desired format 94 | y_train = np.reshape(y_train, newshape=(-1,)) 95 | y_test = np.reshape(y_test, newshape=(-1,)) 96 | 97 | x_train = _reshape_data(data=x_train, data_shape=data_format, num_channels=num_channel) 98 | x_test = _reshape_data(data=x_test, data_shape=data_format, num_channels=num_channel) 99 | 100 | # Image normalization 101 | x_train = (x_train / 255).astype('float32') 102 | x_test = (x_test / 255).astype('float32') 103 | 104 | x_train = (x_train - np.min(x_train))/(np.max(x_train) - np.min(x_train)) 105 | x_test = (x_test - np.min(x_test)) / (np.max(x_test) - np.min(x_test)) 106 | 107 | if normal_class is None: 108 | data = {'x_train_normal': x_train, 109 | 'y_train_normal': y_train, 110 | 'x_test_normal': x_test, 111 | 'y_test_normal': y_test} 112 | else: 113 | train_set = _extract_data(data=x_train, label=y_train, target_lb=normal_class) 114 | test_set = _extract_data(data=x_test, label=y_test, target_lb=normal_class) 115 | 116 | data = {'x_train_normal': train_set[0], 'x_train_abnormal': train_set[1], 117 | 'y_train_normal': train_set[2], 'y_train_abnormal': train_set[3], 118 | 'x_test_normal': test_set[0], 'x_test_abnormal': test_set[1], 119 | 'y_test_normal': test_set[2], 'y_test_abnormal': test_set[3]} 120 | return data 121 | 122 | 123 | # ==================== Image Processing ==================== 124 | def cal_ssim(x, x_rec): 125 | """Calculates SSIM between x and y. 126 | x should be a batch of original images. y should be a batch of reconstructed images. 127 | 128 | :param x: 4D tensor in form of batch_size x img_width x img_height x img_channels. 129 | :param x_rec: 4D tensor in form of batch_size x img_width x img_height x img_channels. 130 | :return: Numpy array with ssim score for each image in the given batch. 131 | """ 132 | 133 | res = [] 134 | num_img = x_rec.shape[0] 135 | print('Calucating SSIM...') 136 | for i in range(num_img): 137 | temp_x = x[i, ...] 138 | temp_rec = x_rec[i, ...] 139 | temp = compare_ssim(temp_x, temp_rec, multichannel=True, gaussian_weights=True) 140 | res.append(temp) 141 | print('SSIM is calculated...') 142 | print('_____________________') 143 | res = np.asarray(res) 144 | return res 145 | 146 | 147 | # ==================== Data Splitting ==================== 148 | def split_data(model=None, data=None, tau=None, split_method=None): 149 | """Split the given data into typical and atypical normal subsets. 150 | 151 | :param model: A trained autoencoder model. 152 | :param data: A numpy tensor standing for the training data. 153 | :param tau: Splitting ratio between 1 to 99. 154 | :param split_method: The name of splitting methods. Currently only 'ssim' is supported. 155 | :return: Indices of typical and atypical normal samples. 156 | """ 157 | 158 | print('\nSplitting data...') 159 | if split_method == 'ssim': 160 | reconstruction = model.predict(data, batch_size=128) 161 | similarity_score = cal_ssim(data, reconstruction) 162 | else: 163 | raise ValueError('\nNo valid splitting methods...') 164 | 165 | sim_thr = np.percentile(similarity_score, tau) 166 | 167 | typical_index = np.where(similarity_score > sim_thr)[0] 168 | atypical_index = np.where(similarity_score <= sim_thr)[0] 169 | 170 | return typical_index, atypical_index 171 | 172 | 173 | # ==================== Misc ==================== 174 | def set_seed(first_seed=2019): 175 | 176 | os.environ['PYTHONHASHSEED'] = '0' 177 | np.random.seed(first_seed) 178 | rn.seed(10) 179 | 180 | session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1) 181 | session_conf.gpu_options.per_process_gpu_memory_fraction = 0.6 182 | tf.set_random_seed(42) 183 | 184 | sess = tf.Session(graph=tf.get_default_graph(), config=session_conf) 185 | K.set_session(sess) -------------------------------------------------------------------------------- /toolkits/visualizations.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # @Time : 2018/11/21 16:37 4 | # @Author : Yiwen Liao 5 | # @File : visualizations.py 6 | # @Software : PyCharm 7 | # @License : Copyright(C), Yiwen Liao 8 | # @Contact : yiwen.liao93@gmail.com 9 | 10 | from useful_packages import * 11 | 12 | 13 | def latent_visualize(normal_encoded, abnormal_encoded, to_save=False, name='', kde_mode=True, hist_mode=False): 14 | 15 | """Visualize the 2D latent vectors with distributions using kernel density estimation. 16 | Visualize the latent representations of images. The input should only has 2 dimensions. If not, the inputs will be 17 | reshaped into vectors, which might lead information and correlation loss. 18 | :param normal_encoded: Numpy array for latent vectors of normal samples. 19 | :param abnormal_encoded: Numpy array for latent vectors of abnormal samples. 20 | :param to_save: Default False. If True, then all the plots will be saved. 21 | :param name: File name for saving plots. 22 | :param kde_mode: Default True. Visualize the distribution using kernel density estimation. 23 | :return: None 24 | """ 25 | 26 | num_se = normal_encoded.shape[0] 27 | num_re = abnormal_encoded.shape[0] 28 | 29 | fig_width = 15 30 | fig_height = 8 31 | 32 | if len(normal_encoded.shape) > 2 or len(abnormal_encoded.shape) > 2: 33 | print('The data has shape of ' + str(normal_encoded.shape) + '.\nTransforming data into vector form...') 34 | normal_encoded = normal_encoded.reshape(num_se, -1) 35 | abnormal_encoded = abnormal_encoded.reshape(num_re, -1) 36 | 37 | num_fea = normal_encoded.shape[-1] 38 | 39 | plt.figure(figsize=(fig_width, fig_height)) 40 | 41 | if num_fea > 16: 42 | print('The number of features is more than 16. Only shows the first 16 features...') 43 | num_fea = 16 44 | 45 | fea_index = np.arange(0, num_fea) 46 | 47 | cols = int(np.sqrt(num_fea)) 48 | rows = int(math.ceil(num_fea / cols)) 49 | 50 | for i in range(num_fea): 51 | plt.subplot(rows, cols, i + 1) 52 | sns.distplot(normal_encoded[:, fea_index[i]], label="normal", kde=kde_mode, hist=hist_mode) 53 | sns.distplot(abnormal_encoded[:, fea_index[i]], label="abnormal", kde=kde_mode, hist=hist_mode) 54 | plt.legend() 55 | plt.tight_layout() 56 | if to_save: 57 | plt.savefig(name) 58 | plt.close() 59 | else: 60 | plt.show() 61 | 62 | 63 | def img_visualize(data, interpolation=None, num_to_show=None, shuffle_img=False, to_save=False, name=''): 64 | 65 | """Visualize a batch of images with desired amount. 66 | Visualize a batch of images. Used as internal functions of other visualization functions. 67 | Input should only have 3 dimensions with (samples, width, height). 68 | :param data: Numpy array for a batch of images. 69 | :param interpolation: Boolean. If use interpolation for kernels. 70 | :param num_to_show: The number of images to show. 71 | :param shuffle_img:Boolean. If randomly select the images to be shown. 72 | :param to_save: Default False. If True, then all the plots will be saved. 73 | :param name: File name for saving plots. 74 | :return: None 75 | """ 76 | 77 | fig_width = 10 78 | fig_height = 10 79 | 80 | num_sample = data.shape[0] 81 | n = min(num_to_show, int(np.sqrt(num_sample))) 82 | 83 | digit_size = data.shape[1] 84 | if len(data.shape) > 3 and data.shape[-1] == 3: 85 | figure = np.zeros((digit_size * n, digit_size * n, 3)) 86 | else: 87 | figure = np.zeros((digit_size * n, digit_size * n)) 88 | 89 | img_order = np.arange(0, n * n) 90 | 91 | if shuffle_img: 92 | np.random.shuffle(img_order) 93 | 94 | for i in range(n): 95 | for j in range(n): 96 | index = img_order[n * i + j] 97 | digit = data[index, ...] 98 | if len(digit.shape) > 2 and digit.shape[-1] == 1: 99 | digit = np.squeeze(digit, axis=-1) 100 | if len(data.shape) > 3 and data.shape[-1] == 3: 101 | figure[i * digit_size: (i + 1) * digit_size, 102 | j * digit_size: (j + 1) * digit_size, :] = digit 103 | else: 104 | figure[i * digit_size: (i + 1) * digit_size, 105 | j * digit_size: (j + 1) * digit_size] = digit 106 | 107 | plt.figure(figsize=(fig_width, fig_height)) 108 | if len(data.shape) > 3 and data.shape[-1] == 3: 109 | plt.imshow(figure, interpolation=interpolation) 110 | else: 111 | plt.imshow(figure, cmap='Greys_r', interpolation=interpolation) 112 | 113 | plt.axis('off') 114 | 115 | if to_save: 116 | plt.savefig(name, dpi=500) 117 | plt.close() 118 | else: 119 | plt.show() 120 | 121 | 122 | def feature_visualize(data, num_to_show=None, to_save=False, name=''): 123 | 124 | num_fea = data.shape[-1] 125 | width = int(np.sqrt(num_fea)) 126 | height = int(num_fea / width) 127 | img_data = np.reshape(data, newshape=(-1, width, height)) 128 | img_visualize(data=img_data, num_to_show=num_to_show, to_save=to_save, name=name) 129 | 130 | 131 | def conv_fea_visualize(data, interpolation='bilinear', num_to_show=None, to_save=False, name=''): 132 | 133 | num_channel = data.shape[-1] 134 | 135 | for ch_index in range(5): 136 | 137 | temp_name = name + '_ch_' + str(ch_index+1) 138 | img_visualize(data=data[:, :, :, ch_index], interpolation=interpolation, 139 | num_to_show=num_to_show, to_save=to_save, name=temp_name) 140 | 141 | 142 | def single_conv_fea_visualize(data, interpolation='bilinear', to_save=False, name=''): 143 | 144 | num_channel = data.shape[-1] 145 | width = data.shape[1] 146 | height = data.shape[2] 147 | img_index = 324 148 | 149 | feature_set = np.zeros(shape=(num_channel, width, height)) 150 | for ch_index in range(num_channel): 151 | temp_img = data[img_index, :, :, ch_index].reshape((width, height)) 152 | feature_set[ch_index, ...] = temp_img 153 | 154 | img_visualize(data=feature_set, interpolation=interpolation, 155 | num_to_show=num_channel, to_save=to_save, name=name) 156 | 157 | 158 | def dim2_visualize(data, label, use_tsne=False, to_save=False, name=''): 159 | """Visualize the latent representations with scatter plots. 160 | Scatter plot only illustrate the latent representations of abnormal samples in order to show 161 | whether the abnormal samples can be well clustered in the latent space. 162 | :param data: Numpy array for latent representations of samples. 163 | :param label: Numpy array for corresponding labels of samples, which should have shape of (samples, 1) 164 | :param use_tsne: Use T-SNE to visualize latent representations. 165 | :param to_save: Default False. If True, then all the plots will be saved. 166 | :param name: File name for saving plots. 167 | :return: None 168 | """ 169 | 170 | if len(data.shape) > 2: 171 | print('The data has shape of ' + str(data.shape) + '. \nTransforming data into vector form...') 172 | data = data.reshape(data.shape[0], -1) 173 | 174 | if use_tsne: 175 | start_time = t.clock() 176 | data = TSNE().fit_transform(data) 177 | num_fea = 2 178 | end_time = t.clock() 179 | tsne_time = end_time - start_time 180 | print('It took %.2f seconds to perform T-SNE...' % tsne_time) 181 | else: 182 | num_fea = data.shape[-1] 183 | 184 | index = np.arange(0, num_fea) 185 | 186 | plt.figure(figsize=(20, 10)) 187 | plt.scatter(data[:, index[0]], data[:, index[1]], c=label, cmap='Set3', alpha=0.9) 188 | plt.xticks([]) 189 | plt.yticks([]) 190 | plt.xlabel(xlabel='Latent Feature 1', fontsize=22) 191 | plt.ylabel(ylabel='Latent Feature 2', fontsize=22) 192 | cb = plt.colorbar() 193 | cb.ax.tick_params(labelsize=22) 194 | # plt.legend() 195 | 196 | if to_save: 197 | #tikz_save(name + '.tex') 198 | plt.savefig(name, dpi=200) 199 | plt.close() 200 | else: 201 | plt.show() 202 | --------------------------------------------------------------------------------