├── LICENSE ├── README.md ├── baseline.yaml ├── common.py ├── main.py ├── mixup_layer.py └── subcluster_adacos.py /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 | # sub-cluster-AdaCos 2 | Accompanying code for the paper "Sub-Cluster AdaCos: Learning Representations for Anomalous Sound Detection" (http://wilkinghoff.com/publications/ijcnn21_sub-cluster.pdf). 3 | 4 | You will need all datasets from task 2 "Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring" of the DCASE challenge 2020 (http://dcase.community/challenge2020/task-unsupervised-detection-of-anomalous-sounds) as well as Tensorflow 2. 5 | 6 | When finding this code helpful, or reusing parts of it, a citation would be appreciated: 7 | 8 | @inproceedings{wilkinghoff2021sub, 9 | title={Sub-Cluster {A}da{C}os: Learning Representations for Anomalous Sound Detection}, 10 | author={Wilkinghoff, Kevin}, 11 | booktitle={International Joint Conference on Neural Networks (IJCNN)}, 12 | year={2021}, 13 | publisher={IEEE} 14 | } 15 | -------------------------------------------------------------------------------- /baseline.yaml: -------------------------------------------------------------------------------- 1 | dev_directory : ./dev_data 2 | eval_directory : ./eval_data 3 | all_data_directory: ./all_data 4 | model_directory: ./model 5 | result_directory: ./result 6 | result_file: result.csv 7 | 8 | max_fpr : 0.1 9 | 10 | feature: 11 | n_mels: 128 12 | frames : 5 13 | n_fft: 1024 14 | hop_length: 512 15 | power: 2.0 16 | 17 | 18 | fit: 19 | compile: 20 | optimizer : adam 21 | loss : mean_squared_error 22 | epochs : 100 23 | batch_size : 512 24 | shuffle : True 25 | validation_split : 0.1 26 | verbose : 1 27 | -------------------------------------------------------------------------------- /common.py: -------------------------------------------------------------------------------- 1 | """ 2 | @file common.py 3 | @brief Commonly used script 4 | @author Toshiki Nakamura, Yuki Nikaido, and Yohei Kawaguchi (Hitachi Ltd.) 5 | Copyright (C) 2020 Hitachi, Ltd. All right reserved. 6 | """ 7 | 8 | ######################################################################## 9 | # import python-library 10 | ######################################################################## 11 | # default 12 | import glob 13 | import argparse 14 | import sys 15 | import os 16 | 17 | # additional 18 | import numpy 19 | import librosa 20 | import librosa.core 21 | import librosa.feature 22 | import yaml 23 | 24 | ######################################################################## 25 | 26 | 27 | ######################################################################## 28 | # setup STD I/O 29 | ######################################################################## 30 | """ 31 | Standard output is logged in "baseline.log". 32 | """ 33 | import logging 34 | 35 | logging.basicConfig(level=logging.DEBUG, filename="baseline.log") 36 | logger = logging.getLogger(' ') 37 | handler = logging.StreamHandler() 38 | formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s') 39 | handler.setFormatter(formatter) 40 | logger.addHandler(handler) 41 | 42 | 43 | ######################################################################## 44 | 45 | 46 | ######################################################################## 47 | # version 48 | ######################################################################## 49 | __versions__ = "1.0.0" 50 | ######################################################################## 51 | 52 | 53 | ######################################################################## 54 | # argparse 55 | ######################################################################## 56 | def command_line_chk(): 57 | parser = argparse.ArgumentParser(description='Without option argument, it will not run properly.') 58 | parser.add_argument('-v', '--version', action='store_true', help="show application version") 59 | parser.add_argument('-e', '--eval', action='store_true', help="run mode Evaluation") 60 | parser.add_argument('-d', '--dev', action='store_true', help="run mode Development") 61 | parser.add_argument('-a', '--all', action='store_true', help="run mode All_Data") 62 | args = parser.parse_args() 63 | if args.version: 64 | print("===============================") 65 | print("DCASE 2020 task 2 baseline\nversion {}".format(__versions__)) 66 | print("===============================\n") 67 | if args.eval ^ args.dev ^ args.all: 68 | if args.dev: 69 | flag = True 70 | elif args.all: 71 | flag = 2 72 | else: 73 | flag = False 74 | else: 75 | flag = None 76 | print("incorrect argument") 77 | print("please set option argument '--dev' or '--eval'") 78 | return flag 79 | ######################################################################## 80 | 81 | 82 | ######################################################################## 83 | # load parameter.yaml 84 | ######################################################################## 85 | def yaml_load(): 86 | with open("baseline.yaml") as stream: 87 | param = yaml.safe_load(stream) 88 | return param 89 | 90 | ######################################################################## 91 | 92 | 93 | ######################################################################## 94 | # file I/O 95 | ######################################################################## 96 | # wav file Input 97 | def file_load(wav_name, mono=False): 98 | """ 99 | load .wav file. 100 | 101 | wav_name : str 102 | target .wav file 103 | sampling_rate : int 104 | audio file sampling_rate 105 | mono : boolean 106 | When load a multi channels file and this param True, the returned data will be merged for mono data 107 | 108 | return : numpy.array( float ) 109 | """ 110 | try: 111 | return librosa.load(wav_name, sr=None, mono=mono) 112 | except: 113 | logger.error("file_broken or not exists!! : {}".format(wav_name)) 114 | 115 | 116 | ######################################################################## 117 | 118 | 119 | ######################################################################## 120 | # feature extractor 121 | ######################################################################## 122 | def file_to_vector_array(file_name, 123 | n_mels=64, 124 | frames=5, 125 | n_fft=1024, 126 | hop_length=512, 127 | power=2.0): 128 | """ 129 | convert file_name to a vector array. 130 | 131 | file_name : str 132 | target .wav file 133 | 134 | return : numpy.array( numpy.array( float ) ) 135 | vector array 136 | * dataset.shape = (dataset_size, feature_vector_length) 137 | """ 138 | # 01 calculate the number of dimensions 139 | dims = n_mels * frames 140 | 141 | # 02 generate melspectrogram using librosa 142 | y, sr = file_load(file_name) 143 | mel_spectrogram = librosa.feature.melspectrogram(y=y, 144 | sr=sr, 145 | n_fft=n_fft, 146 | hop_length=hop_length, 147 | n_mels=n_mels, 148 | power=power) 149 | 150 | # 03 convert melspectrogram to log mel energy 151 | log_mel_spectrogram = 20.0 / power * numpy.log10(mel_spectrogram + sys.float_info.epsilon) 152 | 153 | # 04 calculate total vector size 154 | vector_array_size = len(log_mel_spectrogram[0, :]) - frames + 1 155 | 156 | # 05 skip too short clips 157 | if vector_array_size < 1: 158 | return numpy.empty((0, dims)) 159 | 160 | # 06 generate feature vectors by concatenating multiframes 161 | vector_array = numpy.zeros((vector_array_size, dims)) 162 | for t in range(frames): 163 | vector_array[:, n_mels * t: n_mels * (t + 1)] = log_mel_spectrogram[:, t: t + vector_array_size].T 164 | 165 | return vector_array 166 | 167 | 168 | # load dataset 169 | def select_dirs(param, mode): 170 | """ 171 | param : dict 172 | baseline.yaml data 173 | 174 | return : 175 | if active type the development : 176 | dirs : list [ str ] 177 | load base directory list of dev_data 178 | if active type the evaluation : 179 | dirs : list [ str ] 180 | load base directory list of eval_data 181 | """ 182 | if mode == 1: 183 | logger.info("load_directory <- development") 184 | dir_path = os.path.abspath("{base}/*".format(base=param["dev_directory"])) 185 | dirs = sorted(glob.glob(dir_path)) 186 | elif mode == 2: 187 | logger.info("load_directory <- development + evaluation") 188 | dir_path = os.path.abspath("{base}/*".format(base=param["all_data_directory"])) 189 | dirs = sorted(glob.glob(dir_path)) 190 | else: 191 | logger.info("load_directory <- evaluation") 192 | dir_path = os.path.abspath("{base}/*".format(base=param["eval_directory"])) 193 | dirs = sorted(glob.glob(dir_path)) 194 | return dirs 195 | 196 | ######################################################################## 197 | 198 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import numpy as np 3 | import keras 4 | import os 5 | import soundfile as sf 6 | import tensorflow as tf 7 | import librosa 8 | from sklearn.metrics import roc_auc_score, roc_curve 9 | from sklearn.utils import class_weight 10 | import common as com 11 | from tqdm import tqdm 12 | from sklearn.preprocessing import LabelEncoder 13 | from mixup_layer import MixupLayer 14 | from subcluster_adacos import SCAdaCos 15 | from sklearn.mixture import GaussianMixture 16 | 17 | 18 | def mixupLoss(y_true, y_pred): 19 | return tf.keras.losses.categorical_crossentropy(y_true=y_pred[:, :, 1], y_pred=y_pred[:, :, 0]) 20 | 21 | 22 | def length_norm(mat): 23 | norm_mat = [] 24 | for line in mat: 25 | temp = line/np.math.sqrt(sum(np.power(line, 2))) 26 | norm_mat.append(temp) 27 | norm_mat = np.array(norm_mat) 28 | return norm_mat 29 | 30 | 31 | class LogMelExtractor(object): 32 | """ 33 | Original source code (before changes): https://github.com/qiuqiangkong/dcase2019_task1 34 | """ 35 | def __init__(self, sample_rate, window_size, hop_size, mel_bins, fmin, fmax): 36 | """Log mel feature extractor. 37 | 38 | Args: 39 | sample_rate: int 40 | window_size: int 41 | hop_size: int 42 | mel_bins: int 43 | fmin: int, minimum frequency of mel filter banks 44 | fmax: int, maximum frequency of mel filter banks 45 | """ 46 | 47 | self.window_size = window_size 48 | self.hop_size = hop_size 49 | self.window_func = np.hanning(window_size) 50 | 51 | self.melW = librosa.filters.mel( 52 | sr=sample_rate, 53 | n_fft=window_size, 54 | n_mels=mel_bins, 55 | fmin=fmin, 56 | fmax=fmax).T 57 | 58 | def transform(self, audio): 59 | """Extract feature of a singlechannel audio file. 60 | 61 | Args: 62 | audio: (samples,) 63 | 64 | Returns: 65 | feature: (frames_num, freq_bins) 66 | """ 67 | 68 | window_size = self.window_size 69 | hop_size = self.hop_size 70 | window_func = self.window_func 71 | 72 | # Compute short-time Fourier transform 73 | stft_matrix = librosa.core.stft( 74 | y=audio, 75 | n_fft=window_size, 76 | hop_length=hop_size, 77 | window=window_func, 78 | center=True, 79 | dtype=np.complex64, 80 | pad_mode='reflect').T 81 | '''(N, n_fft // 2 + 1)''' 82 | 83 | # Mel spectrogram 84 | mel_spectrogram = np.dot(np.abs(stft_matrix) ** 2, self.melW) 85 | 86 | # Log mel spectrogram 87 | logmel_spectrogram = librosa.core.power_to_db( 88 | mel_spectrogram, ref=1.0, amin=1e-10, 89 | top_db=None) 90 | 91 | logmel_spectrogram = logmel_spectrogram.astype(np.float32).transpose() 92 | return logmel_spectrogram 93 | 94 | 95 | def model_cnn(emb_size, num_classes, time_dim, min_val, n_subclusters): 96 | data_input = tf.keras.layers.Input(shape=(time_dim, emb_size, 1), dtype='float32') 97 | label_input = tf.keras.layers.Input(shape=(num_classes), dtype='float32') 98 | y = label_input 99 | x = data_input 100 | l2_weight_decay = tf.keras.regularizers.l2(1e-5) 101 | x, y = MixupLayer(prob=1)([x, y]) 102 | 103 | # first block 104 | x = tf.keras.layers.Conv2D(16, 7, strides=2, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(x) 105 | x = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 106 | x = tf.keras.layers.BatchNormalization()(x) 107 | x = tf.keras.layers.MaxPooling2D(3, strides=2)(x) 108 | 109 | # second block 110 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 111 | xr = tf.keras.layers.Conv2D(16, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 112 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 113 | xr = tf.keras.layers.BatchNormalization()(xr) 114 | xr = tf.keras.layers.Conv2D(16, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 115 | x = tf.keras.layers.Add()([x, xr]) 116 | x = tf.keras.layers.BatchNormalization()(x) 117 | 118 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 119 | xr = tf.keras.layers.Conv2D(16, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 120 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 121 | xr = tf.keras.layers.BatchNormalization()(xr) 122 | xr = tf.keras.layers.Conv2D(16, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 123 | x = tf.keras.layers.Add()([x, xr]) 124 | x = tf.keras.layers.BatchNormalization()(x) 125 | 126 | # third block 127 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 128 | xr = tf.keras.layers.Conv2D(32, 3, strides=(2, 2), activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 129 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 130 | xr = tf.keras.layers.BatchNormalization()(xr) 131 | xr = tf.keras.layers.Conv2D(32, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 132 | x = tf.keras.layers.MaxPooling2D((2, 2), padding='same')(x) 133 | x = tf.keras.layers.Conv2D(kernel_size=1, filters=32, strides=1, padding="same", kernel_regularizer=l2_weight_decay)(x) 134 | x = tf.keras.layers.Add()([x, xr]) 135 | x = tf.keras.layers.BatchNormalization()(x) 136 | 137 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 138 | xr = tf.keras.layers.Conv2D(32, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 139 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 140 | xr = tf.keras.layers.BatchNormalization()(xr) 141 | xr = tf.keras.layers.Conv2D(32, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 142 | x = tf.keras.layers.Add()([x, xr]) 143 | x = tf.keras.layers.BatchNormalization()(x) 144 | 145 | # fourth block 146 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 147 | xr = tf.keras.layers.Conv2D(64, 3, strides=(2, 2), activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 148 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 149 | xr = tf.keras.layers.BatchNormalization()(xr) 150 | xr = tf.keras.layers.Conv2D(64, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 151 | x = tf.keras.layers.MaxPooling2D((2, 2), padding='same')(x) 152 | x = tf.keras.layers.Conv2D(kernel_size=1, filters=64, strides=1, padding="same", kernel_regularizer=l2_weight_decay)(x) 153 | x = tf.keras.layers.Add()([x, xr]) 154 | x = tf.keras.layers.BatchNormalization()(x) 155 | 156 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 157 | xr = tf.keras.layers.Conv2D(64, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 158 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 159 | xr = tf.keras.layers.BatchNormalization()(xr) 160 | xr = tf.keras.layers.Conv2D(64, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 161 | x = tf.keras.layers.Add()([x, xr]) 162 | x = tf.keras.layers.BatchNormalization()(x) 163 | 164 | # fifth block 165 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 166 | xr = tf.keras.layers.Conv2D(128, 3, strides=(2, 2), activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 167 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 168 | xr = tf.keras.layers.BatchNormalization()(xr) 169 | xr = tf.keras.layers.Conv2D(128, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 170 | x = tf.keras.layers.MaxPooling2D((2, 2), padding='same')(x) 171 | x = tf.keras.layers.Conv2D(kernel_size=1, filters=128, strides=1, padding="same", kernel_regularizer=l2_weight_decay)(x) 172 | x = tf.keras.layers.Add()([x, xr]) 173 | x = tf.keras.layers.BatchNormalization()(x) 174 | 175 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(x) 176 | xr = tf.keras.layers.Conv2D(128, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 177 | xr = tf.keras.layers.LeakyReLU(alpha=0.1)(xr) 178 | xr = tf.keras.layers.BatchNormalization()(xr) 179 | xr = tf.keras.layers.Conv2D(128, 3, activation='linear', padding='same', kernel_regularizer=l2_weight_decay)(xr) 180 | x = tf.keras.layers.Add()([x, xr]) 181 | x = tf.keras.layers.BatchNormalization()(x) 182 | 183 | # get embeddings and classify 184 | x = tf.keras.layers.MaxPooling2D((10, 1), padding='same')(x) 185 | x = tf.keras.layers.Flatten(name='flat')(x) 186 | x = tf.keras.layers.Dense(128, kernel_regularizer=l2_weight_decay, name='emb')(x) 187 | output = SCAdaCos(n_classes=num_classes, n_subclusters=n_subclusters)([x, y, label_input]) 188 | loss_output = tf.keras.layers.Lambda(lambda x: tf.stack(x, axis=-1))([output, y]) 189 | 190 | return data_input, label_input, loss_output 191 | 192 | 193 | ######################################################################################################################## 194 | # Load data and compute embeddings 195 | ######################################################################################################################## 196 | # emb_size = 6144 197 | n_log_mel = 128 198 | target_sr = 16000 199 | param = com.yaml_load() 200 | extractor = LogMelExtractor(target_sr, 1024, 512, mel_bins=n_log_mel, fmin=0, fmax=target_sr/2) 201 | 202 | # load train data 203 | print('Loading train data') 204 | categories = os.listdir("./dev_data") 205 | 206 | if os.path.isfile(str(n_log_mel) + '_train_log_mel.npy'): 207 | train_log_mel = np.load(str(n_log_mel) + '_train_log_mel.npy') 208 | train_ids = np.load('train_ids.npy') 209 | train_files = np.load('train_files.npy') 210 | else: 211 | train_log_mel = [] 212 | train_ids = [] 213 | train_files = [] 214 | dicts = ['./dev_data/', './eval_data/'] 215 | #dicts = ['./eval_data/'] 216 | #dicts = ['./dev_data/'] 217 | eps=1e-12 218 | for label, category in enumerate(categories): 219 | print(category) 220 | for dict in dicts: 221 | for count, file in tqdm(enumerate(os.listdir(dict + category + "/train")), total=len(os.listdir(dict + category + "/train"))): 222 | file_path = dict + category + "/train/" + file 223 | wav, fs = sf.read(file_path) 224 | wav = librosa.core.to_mono(wav.transpose()).transpose() 225 | # extract log_mels 226 | log_mel = extractor.transform(wav).transpose() 227 | if log_mel.shape[0] > 313: 228 | log_mel = log_mel[log_mel.shape[0]-313:, :] 229 | train_log_mel.append(log_mel) 230 | train_ids.append(category + '_' + file.split('_')[-2]) 231 | train_files.append(file_path) 232 | # reshape arrays and store 233 | train_ids = np.array(train_ids) 234 | train_files = np.array(train_files) 235 | train_log_mel = np.expand_dims(np.array(train_log_mel, dtype=np.float32), axis=-1) 236 | np.save('train_ids.npy', train_ids) 237 | np.save('train_files.npy', train_files) 238 | np.save(str(n_log_mel) + '_train_log_mel.npy', train_log_mel) 239 | 240 | # load evaluation data 241 | print('Loading evaluation data') 242 | if os.path.isfile(str(n_log_mel) + '_eval_log_mel.npy'): 243 | eval_log_mel = np.load(str(n_log_mel) + '_eval_log_mel.npy') 244 | eval_ids = np.load('eval_ids.npy') 245 | eval_normal = np.load('eval_normal.npy') 246 | eval_files = np.load('eval_files.npy') 247 | else: 248 | eval_log_mel = [] 249 | eval_ids = [] 250 | eval_normal = [] 251 | eval_files = [] 252 | eps=1e-12 253 | for label, category in enumerate(categories): 254 | print(category) 255 | for count, file in tqdm(enumerate(os.listdir("./dev_data/" + category + "/test")), total=len(os.listdir("./dev_data/" + category + "/test"))): 256 | file_path = "./dev_data/" + category + "/test/" + file 257 | wav, fs = sf.read(file_path) 258 | wav = librosa.core.to_mono(wav.transpose()).transpose() 259 | # extract log_mels 260 | log_mel = extractor.transform(wav).transpose() 261 | if log_mel.shape[0] > 313: 262 | log_mel = log_mel[log_mel.shape[0]-313:, :] 263 | eval_log_mel.append(log_mel) 264 | eval_ids.append(category + '_' + file.split('_')[-2]) 265 | eval_normal.append(file.split('_')[0] == 'normal') 266 | eval_files.append(file_path) 267 | # reshape arrays and store 268 | eval_ids = np.array(eval_ids) 269 | eval_normal = np.array(eval_normal) 270 | eval_files = np.array(eval_files) 271 | eval_log_mel = np.expand_dims(np.array(eval_log_mel, dtype=np.float32), axis=-1) 272 | np.save('eval_ids.npy', eval_ids) 273 | np.save('eval_normal.npy', eval_normal) 274 | np.save('eval_files.npy', eval_files) 275 | np.save(str(n_log_mel) + '_eval_log_mel.npy', eval_log_mel) 276 | 277 | # load test data 278 | print('Loading test data') 279 | if os.path.isfile(str(n_log_mel) + '_test_log_mel.npy'): 280 | test_log_mel = np.load(str(n_log_mel) + '_test_log_mel.npy') 281 | test_ids = np.load('test_ids.npy') 282 | test_files = np.load('test_files.npy') 283 | else: 284 | test_log_mel = [] 285 | test_ids = [] 286 | test_files = [] 287 | eps = 1e-12 288 | for label, category in enumerate(categories): 289 | print(category) 290 | for count, file in tqdm(enumerate(os.listdir("./eval_data/" + category + "/test")), total=len(os.listdir("./eval_data/" + category + "/test"))): 291 | file_path = "./eval_data/" + category + "/test/" + file 292 | wav, fs = sf.read(file_path) 293 | wav = librosa.core.to_mono(wav.transpose()).transpose() 294 | # extract log_mels 295 | log_mel = extractor.transform(wav).transpose() 296 | if log_mel.shape[0] > 313: 297 | log_mel = log_mel[log_mel.shape[0]-313:, :] 298 | test_log_mel.append(log_mel) 299 | test_ids.append(category + '_' + file.split('_')[-2]) 300 | test_files.append(file_path) 301 | # reshape arrays and store 302 | test_ids = np.array(test_ids) 303 | test_files = np.array(test_files) 304 | test_log_mel = np.expand_dims(np.array(test_log_mel, dtype=np.float32), axis=-1) 305 | np.save('test_ids.npy', test_ids) 306 | np.save('test_files.npy', test_files) 307 | np.save(str(n_log_mel) + '_test_log_mel.npy', test_log_mel) 308 | 309 | # encode ids as labels 310 | le = LabelEncoder() 311 | train_labels = le.fit_transform(train_ids) 312 | eval_labels = le.transform(eval_ids) 313 | test_labels = le.transform(test_ids) 314 | 315 | # distinguish between normal and anomalous samples 316 | unknown_log_mel = eval_log_mel[~eval_normal] 317 | unknown_labels = eval_labels[~eval_normal] 318 | unknown_files = eval_files[~eval_normal] 319 | unknown_ids = eval_ids[~eval_normal] 320 | eval_log_mel = eval_log_mel[eval_normal] 321 | eval_labels = eval_labels[eval_normal] 322 | eval_files = eval_files[eval_normal] 323 | eval_ids = eval_ids[eval_normal] 324 | 325 | # set up dict to convert machine type into a vector indicating all ids that belong to that type 326 | type_dict = {'ToyCar': np.array([1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]), 327 | 'ToyConveyor': np.array([0,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]), 328 | 'fan': np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]), 329 | 'pump': np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0]), 330 | 'slider': np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,0,0,0,0,0,0,0]), 331 | 'valve': np.array([0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1])} 332 | type_dict_s = {'ToyCar': np.array([1,0,0,0,0,0]), 333 | 'ToyConveyor': np.array([0,1,0,0,0,0]), 334 | 'fan': np.array([0,0,1,0,0,0]), 335 | 'pump': np.array([0,0,0,1,0,0]), 336 | 'slider': np.array([0,0,0,0,1,0]), 337 | 'valve': np.array([0,0,0,0,0,1])} 338 | 339 | type_labels_train = np.array([type_dict[train_id.split('_')[0]] for train_id in train_ids]) 340 | type_labels_eval = np.array([type_dict[eval_id.split('_')[0]] for eval_id in eval_ids]) 341 | type_labels_unknown = np.array([type_dict[unknown_id.split('_')[0]] for unknown_id in unknown_ids]) 342 | type_labels_train_s = np.array([type_dict_s[train_id.split('_')[0]] for train_id in train_ids]) 343 | type_labels_eval_s = np.array([type_dict_s[eval_id.split('_')[0]] for eval_id in eval_ids]) 344 | type_labels_unknown_s = np.array([type_dict_s[unknown_id.split('_')[0]] for unknown_id in unknown_ids]) 345 | 346 | ######################################################################################################################## 347 | # Preprocessing 348 | ######################################################################################################################## 349 | 350 | # feature normalization 351 | print('Normalizing data') 352 | eps = 1e-12 353 | mean_log_mel = np.expand_dims(np.repeat(np.expand_dims(np.mean(train_log_mel.reshape(train_log_mel.shape[0]*train_log_mel.shape[1], train_log_mel.shape[2], 1), axis=0), axis=0), 354 | repeats=train_log_mel.shape[1], axis=0), axis=0) 355 | std_log_mel = np.expand_dims(np.repeat(np.expand_dims(np.std(train_log_mel.reshape(train_log_mel.shape[0]*train_log_mel.shape[1], train_log_mel.shape[2], 1), axis=0), axis=0), 356 | repeats=train_log_mel.shape[1], axis=0), axis=0) 357 | train_log_mel = (train_log_mel-mean_log_mel)/(std_log_mel+eps) 358 | eval_log_mel = (eval_log_mel-mean_log_mel)/(std_log_mel+eps) 359 | unknown_log_mel = (unknown_log_mel-mean_log_mel)/(std_log_mel+eps) 360 | test_log_mel = (test_log_mel-mean_log_mel)/(std_log_mel+eps) 361 | 362 | ######################################################################################################################## 363 | # train cnn 364 | ######################################################################################################################## 365 | batch_size = 64 366 | batch_size_test = 64 367 | epochs = 100 368 | aeons = 4 369 | alpha = 1 370 | 371 | # predicting with GMMs 372 | pred_eval = np.zeros((eval_log_mel.shape[0], np.unique(train_labels).shape[0], 3)) 373 | pred_unknown = np.zeros((unknown_log_mel.shape[0], np.unique(train_labels).shape[0], 3)) 374 | pred_test = np.zeros((test_log_mel.shape[0], np.unique(train_labels).shape[0], 3)) 375 | pred_train = np.zeros((train_log_mel.shape[0], np.unique(train_labels).shape[0], 3)) 376 | for n_subclusters in 2**np.arange(7): 377 | y_train_cat = keras.utils.np_utils.to_categorical(train_labels, num_classes=len(np.unique(train_labels))) 378 | y_eval_cat = keras.utils.np_utils.to_categorical(eval_labels, num_classes=len(np.unique(train_labels))) 379 | y_unknown_cat = keras.utils.np_utils.to_categorical(unknown_labels, num_classes=len(np.unique(train_labels))) 380 | 381 | # compile model 382 | data_input, label_input, loss_output = model_cnn(emb_size=train_log_mel.shape[2], 383 | num_classes=len(np.unique(train_labels)), 384 | time_dim=train_log_mel.shape[1], min_val=np.min(train_log_mel), n_subclusters=n_subclusters) 385 | model = tf.keras.Model(inputs=[data_input, label_input], outputs=[loss_output]) 386 | model.compile(loss=[mixupLoss], optimizer=tf.keras.optimizers.Adam()) 387 | print(model.summary()) 388 | callbacks = [ 389 | tf.keras.callbacks.TensorBoard(log_dir=os.path.join("logs"), histogram_freq=0, write_graph=True, 390 | write_images=False) 391 | ] 392 | 393 | for k in np.arange(aeons): 394 | print('subclusters: ' + str(n_subclusters)) 395 | print('aeon: ' + str(k)) 396 | # fit model 397 | weight_path = './models/wts_log_mel_' + str(k + 1) + 'k_' + str(n_log_mel) + '_' + str(n_subclusters) + '.h5' 398 | if not os.path.isfile(weight_path): 399 | class_weights = class_weight.compute_class_weight('balanced', np.unique(train_labels), train_labels) 400 | class_weights = {i: class_weights[i] for i in range(class_weights.shape[0])} 401 | model.fit([train_log_mel, y_train_cat], y_train_cat, verbose=1, 402 | batch_size= batch_size, epochs=epochs, callbacks=callbacks, 403 | validation_data=([eval_log_mel, y_eval_cat], y_eval_cat), class_weight=class_weights) 404 | model.save(weight_path) 405 | else: 406 | model = tf.keras.models.load_model(weight_path, 407 | custom_objects={'MixupLayer': MixupLayer, 'mixupLoss': mixupLoss, 'SCAdaCos': SCAdaCos}) 408 | 409 | emb_model = tf.keras.Model(model.input, model.get_layer('emb').output) 410 | eval_embs = emb_model.predict([eval_log_mel, y_eval_cat], batch_size=batch_size) 411 | train_embs = emb_model.predict([train_log_mel, y_train_cat], batch_size=batch_size) 412 | unknown_embs = emb_model.predict([unknown_log_mel, np.zeros((unknown_log_mel.shape[0], len(np.unique(train_labels))))], batch_size=batch_size) 413 | test_embs = emb_model.predict([test_log_mel, np.zeros((test_log_mel.shape[0], len(np.unique(train_labels))))], batch_size=batch_size) 414 | 415 | # length normalization 416 | print('normalizing lengths') 417 | x_train_ln = length_norm(train_embs) 418 | x_eval_ln = length_norm(eval_embs) 419 | x_test_ln = length_norm(test_embs) 420 | x_unknown_ln = length_norm(unknown_embs) 421 | 422 | model_means = model.layers[-2].get_weights()[0].transpose() 423 | model_means_ln = length_norm(model_means) 424 | 425 | x_train_ln = np.concatenate([x_train_ln, np.mean(train_log_mel, axis=1)[:, :, 0], np.max(train_log_mel, axis=1)[:, :, 0]], axis=-1) 426 | x_eval_ln = np.concatenate([x_eval_ln, np.mean(eval_log_mel, axis=1)[:, :, 0], np.max(eval_log_mel, axis=1)[:, :, 0]], axis=-1) 427 | x_test_ln = np.concatenate([x_test_ln, np.mean(test_log_mel, axis=1)[:, :, 0], np.max(test_log_mel, axis=1)[:, :, 0]], axis=-1) 428 | x_unknown_ln = np.concatenate([x_unknown_ln, np.mean(unknown_log_mel, axis=1)[:, :, 0], np.max(unknown_log_mel, axis=1)[:, :, 0]], axis=-1) 429 | for j, lab in tqdm(enumerate(np.unique(train_labels)), total=len(np.unique(train_labels))): 430 | clf1 = GaussianMixture(n_components=n_subclusters, covariance_type='full', reg_covar=1e-3, means_init=model_means_ln[j * n_subclusters:(j + 1) * n_subclusters]).fit( 431 | x_train_ln[train_labels == lab, :train_embs.shape[1]]) 432 | clf2 = GaussianMixture(n_components=1, covariance_type='full', reg_covar=1e-3).fit( 433 | x_train_ln[train_labels == lab, train_embs.shape[1]:train_embs.shape[1] + train_log_mel.shape[2]]) 434 | clf3 = GaussianMixture(n_components=1, covariance_type='full', reg_covar=1e-3).fit( 435 | x_train_ln[train_labels == lab, train_embs.shape[1] + train_log_mel.shape[2]:]) 436 | 437 | pred_eval[:, j, 0] += -np.max(clf1._estimate_log_prob(x_eval_ln[:, :eval_embs.shape[1]]), axis=-1) 438 | pred_eval[:, j, 1] += -clf2.score_samples( 439 | x_eval_ln[:, eval_embs.shape[1]:eval_embs.shape[1] + eval_log_mel.shape[2]]) 440 | pred_eval[:, j, 2] += -clf3.score_samples(x_eval_ln[:, eval_embs.shape[1] + eval_log_mel.shape[2]:]) 441 | 442 | pred_unknown[:, j, 0] += -np.max(clf1._estimate_log_prob(x_unknown_ln[:, :unknown_embs.shape[1]]), axis=-1) 443 | pred_unknown[:, j, 1] += -clf2.score_samples( 444 | x_unknown_ln[:, unknown_embs.shape[1]:unknown_embs.shape[1] + unknown_log_mel.shape[2]]) 445 | pred_unknown[:, j, 2] += -clf3.score_samples( 446 | x_unknown_ln[:, unknown_embs.shape[1] + unknown_log_mel.shape[2]:]) 447 | 448 | pred_test[:, j, 0] += -np.max(clf1._estimate_log_prob(x_test_ln[:, :test_embs.shape[1]]), axis=-1) 449 | pred_test[:, j, 1] += -clf2.score_samples( 450 | x_test_ln[:, test_embs.shape[1]:test_embs.shape[1] + test_log_mel.shape[2]]) 451 | pred_test[:, j, 2] += -clf3.score_samples(x_test_ln[:, test_embs.shape[1] + test_log_mel.shape[2]:]) 452 | 453 | # use mean for machine type ToyConveyor 454 | pred_eval_final = pred_eval[:, :, 0] 455 | pred_unknown_final = pred_unknown[:, :, 0] 456 | pred_test_final = pred_test[:, :, 0] 457 | for lab in np.unique(train_labels): 458 | if le.inverse_transform([lab])[0].split('_')[0] == 'ToyConveyor': 459 | pred_eval_final[:, lab] = pred_eval[:, lab, 1] 460 | pred_unknown_final[:, lab] = pred_unknown[:, lab, 1] 461 | pred_test_final[:, lab] = pred_test[:, lab, 1] 462 | 463 | # output performance 464 | print('performance on evaluation set') 465 | y_pred_eval = np.argmin(pred_eval_final, axis=1) 466 | y_pred_unknown = np.argmin(pred_unknown_final, axis=1) 467 | print('####################') 468 | print('closed-set performance by machine id:') 469 | print('evaluation files: ' + str(np.mean(y_pred_eval == eval_labels))) 470 | print('unknown files: ' + str(np.mean(y_pred_unknown == unknown_labels))) 471 | print('all files: ' + str( 472 | np.mean(np.hstack([y_pred_unknown, y_pred_eval]) == np.hstack([unknown_labels, eval_labels])))) 473 | print('####################') 474 | type_labels_eval1 = np.array([eval_id.split('_')[0] for eval_id in eval_ids]) 475 | type_labels_unknown1 = np.array([unknown_id.split('_')[0] for unknown_id in unknown_ids]) 476 | type_pred_eval1 = np.array([pred_id.split('_')[0] for pred_id in le.inverse_transform(y_pred_eval)]) 477 | type_pred_unknown1 = np.array([pred_id.split('_')[0] for pred_id in le.inverse_transform(y_pred_unknown)]) 478 | print('closed-set performance by machine type:') 479 | print('evaluation files: ' + str(np.mean(type_pred_eval1 == type_labels_eval1))) 480 | print('unknown files: ' + str(np.mean(type_pred_unknown1 == type_labels_unknown1))) 481 | print('all files: ' + str(np.mean( 482 | np.hstack([type_pred_unknown1, type_pred_eval1]) == np.hstack([type_labels_unknown1, type_labels_eval1])))) 483 | print('####################') 484 | print('closed-set performance on test data') 485 | y_pred_test = np.argmin(pred_test_final, axis=1) 486 | type_labels_test1 = np.array([test_id.split('_')[0] for test_id in test_ids]) 487 | type_pred_test1 = np.array([pred_id.split('_')[0] for pred_id in le.inverse_transform(y_pred_test)]) 488 | print('for machine id: ' + str(np.mean(y_pred_test == test_labels))) 489 | print('for machine type: ' + str(np.mean(type_pred_test1 == type_labels_test1))) 490 | print('####################') 491 | aucs = [] 492 | p_aucs = [] 493 | for j, cat in enumerate(np.unique(eval_ids)): 494 | y_pred = np.concatenate([pred_eval_final[eval_labels == le.transform([cat]), le.transform([cat])], 495 | pred_unknown_final[unknown_labels == le.transform([cat]), le.transform([cat])]], 496 | axis=0) 497 | y_true = np.concatenate([np.zeros(np.sum(eval_labels == le.transform([cat]))), 498 | np.ones(np.sum(unknown_labels == le.transform([cat])))], axis=0) 499 | auc = roc_auc_score(y_true, y_pred) 500 | aucs.append(auc) 501 | p_auc = roc_auc_score(y_true, y_pred, max_fpr=param["max_fpr"]) 502 | p_aucs.append(p_auc) 503 | print('AUC for category ' + str(cat) + ': ' + str(auc * 100)) 504 | print('pAUC for category ' + str(cat) + ': ' + str(p_auc * 100)) 505 | print('####################') 506 | aucs = np.array(aucs) 507 | p_aucs = np.array(p_aucs) 508 | for cat in categories: 509 | mean_auc = np.mean(aucs[np.array([eval_id.split('_')[0] for eval_id in np.unique(eval_ids)]) == cat]) 510 | print('mean AUC for category ' + str(cat) + ': ' + str(mean_auc * 100)) 511 | mean_p_auc = np.mean(p_aucs[np.array([eval_id.split('_')[0] for eval_id in np.unique(eval_ids)]) == cat]) 512 | print('mean pAUC for category ' + str(cat) + ': ' + str(mean_p_auc * 100)) 513 | print('####################') 514 | for cat in categories: 515 | mean_auc = np.mean(aucs[np.array([eval_id.split('_')[0] for eval_id in np.unique(eval_ids)]) == cat]) 516 | mean_p_auc = np.mean(p_aucs[np.array([eval_id.split('_')[0] for eval_id in np.unique(eval_ids)]) == cat]) 517 | print('mean of AUC and pAUC for category ' + str(cat) + ': ' + str((mean_p_auc + mean_auc) * 50)) 518 | print('####################') 519 | mean_auc = np.mean(aucs) 520 | print('mean AUC: ' + str(mean_auc * 100)) 521 | mean_p_auc = np.mean(p_aucs) 522 | print('mean pAUC: ' + str(mean_p_auc * 100)) 523 | 524 | # create challenge submission files 525 | print('creating submission files') 526 | for j, cat in enumerate(np.unique(test_ids)): 527 | file_idx = test_labels == le.transform([cat]) 528 | results = pd.DataFrame() 529 | results['output1'], results['output2'] = [[f.split('/')[-1] for f in test_files[file_idx]], 530 | [str(s) for s in pred_test_final[file_idx, le.transform([cat])]]] 531 | results.to_csv('teams/mfcc_emb/anomaly_score_' + cat.split('_')[0] + '_id_' + cat.split('_')[-1] + '.csv', 532 | encoding='utf-8', index=False, header=False) 533 | print('####################') 534 | print('>>>> finished! <<<<<') 535 | print('####################') 536 | -------------------------------------------------------------------------------- /mixup_layer.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras import backend as K 2 | from tensorflow.keras import layers 3 | import tensorflow as tf 4 | import tensorflow_probability as tfp 5 | 6 | class MixupLayer(layers.Layer): 7 | def __init__(self, prob, alpha=1, **kwargs): 8 | super(MixupLayer, self).__init__(**kwargs) 9 | self.prob = prob 10 | self.alpha = alpha 11 | 12 | def build(self, input_shape): 13 | self.built = True 14 | 15 | def call(self, inputs, training=None): 16 | # get mixup weights 17 | if self.alpha == 1: 18 | #dist = tfp.distributions.Beta(0.5, 0.5) 19 | #l = dist.sample([tf.shape(inputs[0])[0]]) 20 | l = tf.random.uniform(shape=[tf.shape(inputs[0])[0]]) 21 | X_l = tf.reshape(l, [-1]+[1]*(len(inputs[0].shape)-1)) 22 | y_l = tf.reshape(l, [-1]+[1]*(len(inputs[1].shape)-1)) 23 | 24 | # mixup data 25 | X1 = inputs[0] 26 | X2 = tf.reverse(inputs[0], axis=[0]) 27 | X = X1 * X_l + X2 * (1 - X_l) 28 | 29 | # mixup labels 30 | y1 = inputs[1] 31 | y2 = tf.reverse(inputs[1], axis=[0]) 32 | y = y1 * y_l + y2 * (1 - y_l) 33 | 34 | # apply mixup or not 35 | dec = tf.dtypes.cast(tf.random.uniform(shape=[tf.shape(inputs[0])[0]]) < self.prob, tf.dtypes.float32) 36 | dec1 = tf.reshape(dec, [-1] + [1] * (len(inputs[0].shape) - 1)) 37 | out1 = dec1 * X + (1 - dec1) * inputs[0] 38 | dec2 = tf.reshape(dec, [-1] + [1] * (len(inputs[1].shape) - 1)) 39 | out2 = dec2 * y + (1 - dec2) * inputs[1] 40 | outputs = [out1, out2] 41 | 42 | # pick output corresponding to training phase 43 | return K.in_train_phase(outputs, inputs, training=training) 44 | 45 | def get_config(self): 46 | config = { 47 | 'prob': self.prob, 48 | 'alpha': self.alpha 49 | } 50 | base_config = super(MixupLayer, self).get_config() 51 | return dict(list(base_config.items()) + list(config.items())) 52 | 53 | -------------------------------------------------------------------------------- /subcluster_adacos.py: -------------------------------------------------------------------------------- 1 | import math 2 | import tensorflow as tf 3 | import tensorflow_probability as tfp 4 | from tensorflow.keras import backend as K 5 | 6 | 7 | class SCAdaCos(tf.keras.layers.Layer): 8 | def __init__(self, n_classes=10, n_subclusters=1, regularizer=None, **kwargs): 9 | super(SCAdaCos, self).__init__(**kwargs) 10 | self.n_classes = n_classes 11 | self.n_subclusters = n_subclusters 12 | self.s_init = math.sqrt(2) * math.log(n_classes*n_subclusters - 1) 13 | self.regularizer = tf.keras.regularizers.get(regularizer) 14 | 15 | def build(self, input_shape): 16 | super(SCAdaCos, self).build(input_shape[0]) 17 | self.W = self.add_weight(name='W_AdaCos' + str(self.n_classes) + '_' + str(self.n_subclusters), 18 | shape=(input_shape[0][-1], self.n_classes*self.n_subclusters), 19 | initializer='glorot_uniform', 20 | trainable=True, 21 | regularizer=self.regularizer) 22 | self.s = self.add_weight(name='s' + str(self.n_classes) + '_' + str(self.n_subclusters), 23 | shape=(), 24 | initializer=tf.keras.initializers.Constant(self.s_init), 25 | trainable=False, 26 | aggregation=tf.VariableAggregation.MEAN) 27 | 28 | def call(self, inputs, training=None): 29 | x, y1, y2 = inputs 30 | y1_orig = y1 31 | y1 = tf.repeat(y1, repeats=self.n_subclusters, axis=-1) 32 | y2 = tf.repeat(y2, repeats=self.n_subclusters, axis=-1) 33 | # normalize feature 34 | x = tf.nn.l2_normalize(x, axis=1) 35 | # normalize weights 36 | W = tf.nn.l2_normalize(self.W, axis=0) 37 | # dot product 38 | logits = x @ W # same as cos theta 39 | theta = tf.acos(K.clip(logits, -1.0 + K.epsilon(), 1.0 - K.epsilon())) 40 | 41 | if training: 42 | max_s_logits = tf.reduce_max(self.s * logits) 43 | B_avg = tf.exp(self.s*logits-max_s_logits) 44 | B_avg = tf.reduce_mean(tf.reduce_sum(B_avg, axis=1)) 45 | theta_class = tf.reduce_sum(y1 * theta, axis=1) * tf.math.count_nonzero(y1_orig, axis=1, dtype=tf.dtypes.float32) # take mix-upped angle of mix-upped classes 46 | theta_med = tfp.stats.percentile(theta_class, q=50) # computes median 47 | self.s.assign( 48 | (max_s_logits + tf.math.log(B_avg)) / 49 | tf.math.cos(tf.minimum(math.pi / 4, theta_med)) + K.epsilon()) 50 | logits *= self.s 51 | out = tf.keras.activations.softmax(logits) 52 | out = tf.reshape(out, (-1, self.n_classes, self.n_subclusters)) 53 | out = tf.math.reduce_sum(out, axis=2) 54 | return out 55 | 56 | def compute_output_shape(self, input_shape): 57 | return (None, self.n_classes) 58 | 59 | def get_config(self): 60 | config = { 61 | 'n_classes': self.n_classes, 62 | 'regularizer': self.regularizer, 63 | 'n_subclusters': self.n_subclusters 64 | } 65 | base_config = super(SCAdaCos, self).get_config() 66 | return dict(list(base_config.items()) + list(config.items())) --------------------------------------------------------------------------------