├── LICENSE ├── README.md ├── data ├── README.md ├── convert_AITEX.py ├── convert_BrainMRI.py ├── convert_HeadCT.py ├── convert_MastCam.py ├── convert_SDD.py ├── convert_elpv.py ├── convert_hyperkvasir.py └── convert_optical.py ├── dataloaders ├── dataloader.py └── utlis.py ├── datasets ├── base_dataset.py ├── cutmix.py └── mvtecad.py ├── modeling ├── layers │ ├── __init__.py │ ├── binary_focal_loss.py │ └── deviation_loss.py ├── net.py └── networks │ ├── backbone.py │ ├── resnet.py │ └── resnet18.py ├── requirements.txt └── train.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU AFFERO GENERAL PUBLIC LICENSE 2 | Version 3, 19 November 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 Affero General Public License is a free, copyleft license for 11 | software and other kinds of works, specifically designed to ensure 12 | cooperation with the community in the case of network server software. 13 | 14 | The licenses for most software and other practical works are designed 15 | to take away your freedom to share and change the works. By contrast, 16 | our General Public Licenses are intended to guarantee your freedom to 17 | share and change all versions of a program--to make sure it remains free 18 | software for all its users. 19 | 20 | When we speak of free software, we are referring to freedom, not 21 | price. Our General Public Licenses are designed to make sure that you 22 | have the freedom to distribute copies of free software (and charge for 23 | them if you wish), that you receive source code or can get it if you 24 | want it, that you can change the software or use pieces of it in new 25 | free programs, and that you know you can do these things. 26 | 27 | Developers that use our General Public Licenses protect your rights 28 | with two steps: (1) assert copyright on the software, and (2) offer 29 | you this License which gives you legal permission to copy, distribute 30 | and/or modify the software. 31 | 32 | A secondary benefit of defending all users' freedom is that 33 | improvements made in alternate versions of the program, if they 34 | receive widespread use, become available for other developers to 35 | incorporate. Many developers of free software are heartened and 36 | encouraged by the resulting cooperation. However, in the case of 37 | software used on network servers, this result may fail to come about. 38 | The GNU General Public License permits making a modified version and 39 | letting the public access it on a server without ever releasing its 40 | source code to the public. 41 | 42 | The GNU Affero General Public License is designed specifically to 43 | ensure that, in such cases, the modified source code becomes available 44 | to the community. It requires the operator of a network server to 45 | provide the source code of the modified version running there to the 46 | users of that server. Therefore, public use of a modified version, on 47 | a publicly accessible server, gives the public access to the source 48 | code of the modified version. 49 | 50 | An older license, called the Affero General Public License and 51 | published by Affero, was designed to accomplish similar goals. This is 52 | a different license, not a version of the Affero GPL, but Affero has 53 | released a new version of the Affero GPL which permits relicensing under 54 | this license. 55 | 56 | The precise terms and conditions for copying, distribution and 57 | modification follow. 58 | 59 | TERMS AND CONDITIONS 60 | 61 | 0. Definitions. 62 | 63 | "This License" refers to version 3 of the GNU Affero General Public License. 64 | 65 | "Copyright" also means copyright-like laws that apply to other kinds of 66 | works, such as semiconductor masks. 67 | 68 | "The Program" refers to any copyrightable work licensed under this 69 | License. Each licensee is addressed as "you". "Licensees" and 70 | "recipients" may be individuals or organizations. 71 | 72 | To "modify" a work means to copy from or adapt all or part of the work 73 | in a fashion requiring copyright permission, other than the making of an 74 | exact copy. The resulting work is called a "modified version" of the 75 | earlier work or a work "based on" the earlier work. 76 | 77 | A "covered work" means either the unmodified Program or a work based 78 | on the Program. 79 | 80 | To "propagate" a work means to do anything with it that, without 81 | permission, would make you directly or secondarily liable for 82 | infringement under applicable copyright law, except executing it on a 83 | computer or modifying a private copy. Propagation includes copying, 84 | distribution (with or without modification), making available to the 85 | public, and in some countries other activities as well. 86 | 87 | To "convey" a work means any kind of propagation that enables other 88 | parties to make or receive copies. Mere interaction with a user through 89 | a computer network, with no transfer of a copy, is not conveying. 90 | 91 | An interactive user interface displays "Appropriate Legal Notices" 92 | to the extent that it includes a convenient and prominently visible 93 | feature that (1) displays an appropriate copyright notice, and (2) 94 | tells the user that there is no warranty for the work (except to the 95 | extent that warranties are provided), that licensees may convey the 96 | work under this License, and how to view a copy of this License. If 97 | the interface presents a list of user commands or options, such as a 98 | menu, a prominent item in the list meets this criterion. 99 | 100 | 1. Source Code. 101 | 102 | The "source code" for a work means the preferred form of the work 103 | for making modifications to it. "Object code" means any non-source 104 | form of a work. 105 | 106 | A "Standard Interface" means an interface that either is an official 107 | standard defined by a recognized standards body, or, in the case of 108 | interfaces specified for a particular programming language, one that 109 | is widely used among developers working in that language. 110 | 111 | The "System Libraries" of an executable work include anything, other 112 | than the work as a whole, that (a) is included in the normal form of 113 | packaging a Major Component, but which is not part of that Major 114 | Component, and (b) serves only to enable use of the work with that 115 | Major Component, or to implement a Standard Interface for which an 116 | implementation is available to the public in source code form. A 117 | "Major Component", in this context, means a major essential component 118 | (kernel, window system, and so on) of the specific operating system 119 | (if any) on which the executable work runs, or a compiler used to 120 | produce the work, or an object code interpreter used to run it. 121 | 122 | The "Corresponding Source" for a work in object code form means all 123 | the source code needed to generate, install, and (for an executable 124 | work) run the object code and to modify the work, including scripts to 125 | control those activities. However, it does not include the work's 126 | System Libraries, or general-purpose tools or generally available free 127 | programs which are used unmodified in performing those activities but 128 | which are not part of the work. For example, Corresponding Source 129 | includes interface definition files associated with source files for 130 | the work, and the source code for shared libraries and dynamically 131 | linked subprograms that the work is specifically designed to require, 132 | such as by intimate data communication or control flow between those 133 | subprograms and other parts of the work. 134 | 135 | The Corresponding Source need not include anything that users 136 | can regenerate automatically from other parts of the Corresponding 137 | Source. 138 | 139 | The Corresponding Source for a work in source code form is that 140 | same work. 141 | 142 | 2. Basic Permissions. 143 | 144 | All rights granted under this License are granted for the term of 145 | copyright on the Program, and are irrevocable provided the stated 146 | conditions are met. This License explicitly affirms your unlimited 147 | permission to run the unmodified Program. The output from running a 148 | covered work is covered by this License only if the output, given its 149 | content, constitutes a covered work. This License acknowledges your 150 | rights of fair use or other equivalent, as provided by copyright law. 151 | 152 | You may make, run and propagate covered works that you do not 153 | convey, without conditions so long as your license otherwise remains 154 | in force. You may convey covered works to others for the sole purpose 155 | of having them make modifications exclusively for you, or provide you 156 | with facilities for running those works, provided that you comply with 157 | the terms of this License in conveying all material for which you do 158 | not control copyright. Those thus making or running the covered works 159 | for you must do so exclusively on your behalf, under your direction 160 | and control, on terms that prohibit them from making any copies of 161 | your copyrighted material outside their relationship with you. 162 | 163 | Conveying under any other circumstances is permitted solely under 164 | the conditions stated below. Sublicensing is not allowed; section 10 165 | makes it unnecessary. 166 | 167 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 168 | 169 | No covered work shall be deemed part of an effective technological 170 | measure under any applicable law fulfilling obligations under article 171 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 172 | similar laws prohibiting or restricting circumvention of such 173 | measures. 174 | 175 | When you convey a covered work, you waive any legal power to forbid 176 | circumvention of technological measures to the extent such circumvention 177 | is effected by exercising rights under this License with respect to 178 | the covered work, and you disclaim any intention to limit operation or 179 | modification of the work as a means of enforcing, against the work's 180 | users, your or third parties' legal rights to forbid circumvention of 181 | technological measures. 182 | 183 | 4. Conveying Verbatim Copies. 184 | 185 | You may convey verbatim copies of the Program's source code as you 186 | receive it, in any medium, provided that you conspicuously and 187 | appropriately publish on each copy an appropriate copyright notice; 188 | keep intact all notices stating that this License and any 189 | non-permissive terms added in accord with section 7 apply to the code; 190 | keep intact all notices of the absence of any warranty; and give all 191 | recipients a copy of this License along with the Program. 192 | 193 | You may charge any price or no price for each copy that you convey, 194 | and you may offer support or warranty protection for a fee. 195 | 196 | 5. Conveying Modified Source Versions. 197 | 198 | You may convey a work based on the Program, or the modifications to 199 | produce it from the Program, in the form of source code under the 200 | terms of section 4, provided that you also meet all of these conditions: 201 | 202 | a) The work must carry prominent notices stating that you modified 203 | it, and giving a relevant date. 204 | 205 | b) The work must carry prominent notices stating that it is 206 | released under this License and any conditions added under section 207 | 7. This requirement modifies the requirement in section 4 to 208 | "keep intact all notices". 209 | 210 | c) You must license the entire work, as a whole, under this 211 | License to anyone who comes into possession of a copy. This 212 | License will therefore apply, along with any applicable section 7 213 | additional terms, to the whole of the work, and all its parts, 214 | regardless of how they are packaged. This License gives no 215 | permission to license the work in any other way, but it does not 216 | invalidate such permission if you have separately received it. 217 | 218 | d) If the work has interactive user interfaces, each must display 219 | Appropriate Legal Notices; however, if the Program has interactive 220 | interfaces that do not display Appropriate Legal Notices, your 221 | work need not make them do so. 222 | 223 | A compilation of a covered work with other separate and independent 224 | works, which are not by their nature extensions of the covered work, 225 | and which are not combined with it such as to form a larger program, 226 | in or on a volume of a storage or distribution medium, is called an 227 | "aggregate" if the compilation and its resulting copyright are not 228 | used to limit the access or legal rights of the compilation's users 229 | beyond what the individual works permit. Inclusion of a covered work 230 | in an aggregate does not cause this License to apply to the other 231 | parts of the aggregate. 232 | 233 | 6. Conveying Non-Source Forms. 234 | 235 | You may convey a covered work in object code form under the terms 236 | of sections 4 and 5, provided that you also convey the 237 | machine-readable Corresponding Source under the terms of this License, 238 | in one of these ways: 239 | 240 | a) Convey the object code in, or embodied in, a physical product 241 | (including a physical distribution medium), accompanied by the 242 | Corresponding Source fixed on a durable physical medium 243 | customarily used for software interchange. 244 | 245 | b) Convey the object code in, or embodied in, a physical product 246 | (including a physical distribution medium), accompanied by a 247 | written offer, valid for at least three years and valid for as 248 | long as you offer spare parts or customer support for that product 249 | model, to give anyone who possesses the object code either (1) a 250 | copy of the Corresponding Source for all the software in the 251 | product that is covered by this License, on a durable physical 252 | medium customarily used for software interchange, for a price no 253 | more than your reasonable cost of physically performing this 254 | conveying of source, or (2) access to copy the 255 | Corresponding Source from a network server at no charge. 256 | 257 | c) Convey individual copies of the object code with a copy of the 258 | written offer to provide the Corresponding Source. This 259 | alternative is allowed only occasionally and noncommercially, and 260 | only if you received the object code with such an offer, in accord 261 | with subsection 6b. 262 | 263 | d) Convey the object code by offering access from a designated 264 | place (gratis or for a charge), and offer equivalent access to the 265 | Corresponding Source in the same way through the same place at no 266 | further charge. You need not require recipients to copy the 267 | Corresponding Source along with the object code. If the place to 268 | copy the object code is a network server, the Corresponding Source 269 | may be on a different server (operated by you or a third party) 270 | that supports equivalent copying facilities, provided you maintain 271 | clear directions next to the object code saying where to find the 272 | Corresponding Source. Regardless of what server hosts the 273 | Corresponding Source, you remain obligated to ensure that it is 274 | available for as long as needed to satisfy these requirements. 275 | 276 | e) Convey the object code using peer-to-peer transmission, provided 277 | you inform other peers where the object code and Corresponding 278 | Source of the work are being offered to the general public at no 279 | charge under subsection 6d. 280 | 281 | A separable portion of the object code, whose source code is excluded 282 | from the Corresponding Source as a System Library, need not be 283 | included in conveying the object code work. 284 | 285 | A "User Product" is either (1) a "consumer product", which means any 286 | tangible personal property which is normally used for personal, family, 287 | or household purposes, or (2) anything designed or sold for incorporation 288 | into a dwelling. In determining whether a product is a consumer product, 289 | doubtful cases shall be resolved in favor of coverage. For a particular 290 | product received by a particular user, "normally used" refers to a 291 | typical or common use of that class of product, regardless of the status 292 | of the particular user or of the way in which the particular user 293 | actually uses, or expects or is expected to use, the product. A product 294 | is a consumer product regardless of whether the product has substantial 295 | commercial, industrial or non-consumer uses, unless such uses represent 296 | the only significant mode of use of the product. 297 | 298 | "Installation Information" for a User Product means any methods, 299 | procedures, authorization keys, or other information required to install 300 | and execute modified versions of a covered work in that User Product from 301 | a modified version of its Corresponding Source. The information must 302 | suffice to ensure that the continued functioning of the modified object 303 | code is in no case prevented or interfered with solely because 304 | modification has been made. 305 | 306 | If you convey an object code work under this section in, or with, or 307 | specifically for use in, a User Product, and the conveying occurs as 308 | part of a transaction in which the right of possession and use of the 309 | User Product is transferred to the recipient in perpetuity or for a 310 | fixed term (regardless of how the transaction is characterized), the 311 | Corresponding Source conveyed under this section must be accompanied 312 | by the Installation Information. But this requirement does not apply 313 | if neither you nor any third party retains the ability to install 314 | modified object code on the User Product (for example, the work has 315 | been installed in ROM). 316 | 317 | The requirement to provide Installation Information does not include a 318 | requirement to continue to provide support service, warranty, or updates 319 | for a work that has been modified or installed by the recipient, or for 320 | the User Product in which it has been modified or installed. Access to a 321 | network may be denied when the modification itself materially and 322 | adversely affects the operation of the network or violates the rules and 323 | protocols for communication across the network. 324 | 325 | Corresponding Source conveyed, and Installation Information provided, 326 | in accord with this section must be in a format that is publicly 327 | documented (and with an implementation available to the public in 328 | source code form), and must require no special password or key for 329 | unpacking, reading or copying. 330 | 331 | 7. Additional Terms. 332 | 333 | "Additional permissions" are terms that supplement the terms of this 334 | License by making exceptions from one or more of its conditions. 335 | Additional permissions that are applicable to the entire Program shall 336 | be treated as though they were included in this License, to the extent 337 | that they are valid under applicable law. If additional permissions 338 | apply only to part of the Program, that part may be used separately 339 | under those permissions, but the entire Program remains governed by 340 | this License without regard to the additional permissions. 341 | 342 | When you convey a copy of a covered work, you may at your option 343 | remove any additional permissions from that copy, or from any part of 344 | it. (Additional permissions may be written to require their own 345 | removal in certain cases when you modify the work.) You may place 346 | additional permissions on material, added by you to a covered work, 347 | for which you have or can give appropriate copyright permission. 348 | 349 | Notwithstanding any other provision of this License, for material you 350 | add to a covered work, you may (if authorized by the copyright holders of 351 | that material) supplement the terms of this License with terms: 352 | 353 | a) Disclaiming warranty or limiting liability differently from the 354 | terms of sections 15 and 16 of this License; or 355 | 356 | b) Requiring preservation of specified reasonable legal notices or 357 | author attributions in that material or in the Appropriate Legal 358 | Notices displayed by works containing it; or 359 | 360 | c) Prohibiting misrepresentation of the origin of that material, or 361 | requiring that modified versions of such material be marked in 362 | reasonable ways as different from the original version; or 363 | 364 | d) Limiting the use for publicity purposes of names of licensors or 365 | authors of the material; or 366 | 367 | e) Declining to grant rights under trademark law for use of some 368 | trade names, trademarks, or service marks; or 369 | 370 | f) Requiring indemnification of licensors and authors of that 371 | material by anyone who conveys the material (or modified versions of 372 | it) with contractual assumptions of liability to the recipient, for 373 | any liability that these contractual assumptions directly impose on 374 | those licensors and authors. 375 | 376 | All other non-permissive additional terms are considered "further 377 | restrictions" within the meaning of section 10. If the Program as you 378 | received it, or any part of it, contains a notice stating that it is 379 | governed by this License along with a term that is a further 380 | restriction, you may remove that term. If a license document contains 381 | a further restriction but permits relicensing or conveying under this 382 | License, you may add to a covered work material governed by the terms 383 | of that license document, provided that the further restriction does 384 | not survive such relicensing or conveying. 385 | 386 | If you add terms to a covered work in accord with this section, you 387 | must place, in the relevant source files, a statement of the 388 | additional terms that apply to those files, or a notice indicating 389 | where to find the applicable terms. 390 | 391 | Additional terms, permissive or non-permissive, may be stated in the 392 | form of a separately written license, or stated as exceptions; 393 | the above requirements apply either way. 394 | 395 | 8. Termination. 396 | 397 | You may not propagate or modify a covered work except as expressly 398 | provided under this License. Any attempt otherwise to propagate or 399 | modify it is void, and will automatically terminate your rights under 400 | this License (including any patent licenses granted under the third 401 | paragraph of section 11). 402 | 403 | However, if you cease all violation of this License, then your 404 | license from a particular copyright holder is reinstated (a) 405 | provisionally, unless and until the copyright holder explicitly and 406 | finally terminates your license, and (b) permanently, if the copyright 407 | holder fails to notify you of the violation by some reasonable means 408 | prior to 60 days after the cessation. 409 | 410 | Moreover, your license from a particular copyright holder is 411 | reinstated permanently if the copyright holder notifies you of the 412 | violation by some reasonable means, this is the first time you have 413 | received notice of violation of this License (for any work) from that 414 | copyright holder, and you cure the violation prior to 30 days after 415 | your receipt of the notice. 416 | 417 | Termination of your rights under this section does not terminate the 418 | licenses of parties who have received copies or rights from you under 419 | this License. If your rights have been terminated and not permanently 420 | reinstated, you do not qualify to receive new licenses for the same 421 | material under section 10. 422 | 423 | 9. Acceptance Not Required for Having Copies. 424 | 425 | You are not required to accept this License in order to receive or 426 | run a copy of the Program. Ancillary propagation of a covered work 427 | occurring solely as a consequence of using peer-to-peer transmission 428 | to receive a copy likewise does not require acceptance. However, 429 | nothing other than this License grants you permission to propagate or 430 | modify any covered work. These actions infringe copyright if you do 431 | not accept this License. Therefore, by modifying or propagating a 432 | covered work, you indicate your acceptance of this License to do so. 433 | 434 | 10. Automatic Licensing of Downstream Recipients. 435 | 436 | Each time you convey a covered work, the recipient automatically 437 | receives a license from the original licensors, to run, modify and 438 | propagate that work, subject to this License. You are not responsible 439 | for enforcing compliance by third parties with this License. 440 | 441 | An "entity transaction" is a transaction transferring control of an 442 | organization, or substantially all assets of one, or subdividing an 443 | organization, or merging organizations. If propagation of a covered 444 | work results from an entity transaction, each party to that 445 | transaction who receives a copy of the work also receives whatever 446 | licenses to the work the party's predecessor in interest had or could 447 | give under the previous paragraph, plus a right to possession of the 448 | Corresponding Source of the work from the predecessor in interest, if 449 | the predecessor has it or can get it with reasonable efforts. 450 | 451 | You may not impose any further restrictions on the exercise of the 452 | rights granted or affirmed under this License. For example, you may 453 | not impose a license fee, royalty, or other charge for exercise of 454 | rights granted under this License, and you may not initiate litigation 455 | (including a cross-claim or counterclaim in a lawsuit) alleging that 456 | any patent claim is infringed by making, using, selling, offering for 457 | sale, or importing the Program or any portion of it. 458 | 459 | 11. Patents. 460 | 461 | A "contributor" is a copyright holder who authorizes use under this 462 | License of the Program or a work on which the Program is based. The 463 | work thus licensed is called the contributor's "contributor version". 464 | 465 | A contributor's "essential patent claims" are all patent claims 466 | owned or controlled by the contributor, whether already acquired or 467 | hereafter acquired, that would be infringed by some manner, permitted 468 | by this License, of making, using, or selling its contributor version, 469 | but do not include claims that would be infringed only as a 470 | consequence of further modification of the contributor version. For 471 | purposes of this definition, "control" includes the right to grant 472 | patent sublicenses in a manner consistent with the requirements of 473 | this License. 474 | 475 | Each contributor grants you a non-exclusive, worldwide, royalty-free 476 | patent license under the contributor's essential patent claims, to 477 | make, use, sell, offer for sale, import and otherwise run, modify and 478 | propagate the contents of its contributor version. 479 | 480 | In the following three paragraphs, a "patent license" is any express 481 | agreement or commitment, however denominated, not to enforce a patent 482 | (such as an express permission to practice a patent or covenant not to 483 | sue for patent infringement). To "grant" such a patent license to a 484 | party means to make such an agreement or commitment not to enforce a 485 | patent against the party. 486 | 487 | If you convey a covered work, knowingly relying on a patent license, 488 | and the Corresponding Source of the work is not available for anyone 489 | to copy, free of charge and under the terms of this License, through a 490 | publicly available network server or other readily accessible means, 491 | then you must either (1) cause the Corresponding Source to be so 492 | available, or (2) arrange to deprive yourself of the benefit of the 493 | patent license for this particular work, or (3) arrange, in a manner 494 | consistent with the requirements of this License, to extend the patent 495 | license to downstream recipients. "Knowingly relying" means you have 496 | actual knowledge that, but for the patent license, your conveying the 497 | covered work in a country, or your recipient's use of the covered work 498 | in a country, would infringe one or more identifiable patents in that 499 | country that you have reason to believe are valid. 500 | 501 | If, pursuant to or in connection with a single transaction or 502 | arrangement, you convey, or propagate by procuring conveyance of, a 503 | covered work, and grant a patent license to some of the parties 504 | receiving the covered work authorizing them to use, propagate, modify 505 | or convey a specific copy of the covered work, then the patent license 506 | you grant is automatically extended to all recipients of the covered 507 | work and works based on it. 508 | 509 | A patent license is "discriminatory" if it does not include within 510 | the scope of its coverage, prohibits the exercise of, or is 511 | conditioned on the non-exercise of one or more of the rights that are 512 | specifically granted under this License. You may not convey a covered 513 | work if you are a party to an arrangement with a third party that is 514 | in the business of distributing software, under which you make payment 515 | to the third party based on the extent of your activity of conveying 516 | the work, and under which the third party grants, to any of the 517 | parties who would receive the covered work from you, a discriminatory 518 | patent license (a) in connection with copies of the covered work 519 | conveyed by you (or copies made from those copies), or (b) primarily 520 | for and in connection with specific products or compilations that 521 | contain the covered work, unless you entered into that arrangement, 522 | or that patent license was granted, prior to 28 March 2007. 523 | 524 | Nothing in this License shall be construed as excluding or limiting 525 | any implied license or other defenses to infringement that may 526 | otherwise be available to you under applicable patent law. 527 | 528 | 12. No Surrender of Others' Freedom. 529 | 530 | If conditions are imposed on you (whether by court order, agreement or 531 | otherwise) that contradict the conditions of this License, they do not 532 | excuse you from the conditions of this License. If you cannot convey a 533 | covered work so as to satisfy simultaneously your obligations under this 534 | License and any other pertinent obligations, then as a consequence you may 535 | not convey it at all. For example, if you agree to terms that obligate you 536 | to collect a royalty for further conveying from those to whom you convey 537 | the Program, the only way you could satisfy both those terms and this 538 | License would be to refrain entirely from conveying the Program. 539 | 540 | 13. Remote Network Interaction; Use with the GNU General Public License. 541 | 542 | Notwithstanding any other provision of this License, if you modify the 543 | Program, your modified version must prominently offer all users 544 | interacting with it remotely through a computer network (if your version 545 | supports such interaction) an opportunity to receive the Corresponding 546 | Source of your version by providing access to the Corresponding Source 547 | from a network server at no charge, through some standard or customary 548 | means of facilitating copying of software. This Corresponding Source 549 | shall include the Corresponding Source for any work covered by version 3 550 | of the GNU General Public License that is incorporated pursuant to the 551 | following paragraph. 552 | 553 | Notwithstanding any other provision of this License, you have 554 | permission to link or combine any covered work with a work licensed 555 | under version 3 of the GNU General Public License into a single 556 | combined work, and to convey the resulting work. The terms of this 557 | License will continue to apply to the part which is the covered work, 558 | but the work with which it is combined will remain governed by version 559 | 3 of the GNU General Public License. 560 | 561 | 14. Revised Versions of this License. 562 | 563 | The Free Software Foundation may publish revised and/or new versions of 564 | the GNU Affero General Public License from time to time. Such new versions 565 | will be similar in spirit to the present version, but may differ in detail to 566 | address new problems or concerns. 567 | 568 | Each version is given a distinguishing version number. If the 569 | Program specifies that a certain numbered version of the GNU Affero General 570 | Public License "or any later version" applies to it, you have the 571 | option of following the terms and conditions either of that numbered 572 | version or of any later version published by the Free Software 573 | Foundation. If the Program does not specify a version number of the 574 | GNU Affero General Public License, you may choose any version ever published 575 | by the Free Software Foundation. 576 | 577 | If the Program specifies that a proxy can decide which future 578 | versions of the GNU Affero General Public License can be used, that proxy's 579 | public statement of acceptance of a version permanently authorizes you 580 | to choose that version for the Program. 581 | 582 | Later license versions may give you additional or different 583 | permissions. However, no additional obligations are imposed on any 584 | author or copyright holder as a result of your choosing to follow a 585 | later version. 586 | 587 | 15. Disclaimer of Warranty. 588 | 589 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 590 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 591 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 592 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 593 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 594 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 595 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 596 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 597 | 598 | 16. Limitation of Liability. 599 | 600 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 601 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 602 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 603 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 604 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 605 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 606 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 607 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 608 | SUCH DAMAGES. 609 | 610 | 17. Interpretation of Sections 15 and 16. 611 | 612 | If the disclaimer of warranty and limitation of liability provided 613 | above cannot be given local legal effect according to their terms, 614 | reviewing courts shall apply local law that most closely approximates 615 | an absolute waiver of all civil liability in connection with the 616 | Program, unless a warranty or assumption of liability accompanies a 617 | copy of the Program in return for a fee. 618 | 619 | END OF TERMS AND CONDITIONS 620 | 621 | How to Apply These Terms to Your New Programs 622 | 623 | If you develop a new program, and you want it to be of the greatest 624 | possible use to the public, the best way to achieve this is to make it 625 | free software which everyone can redistribute and change under these terms. 626 | 627 | To do so, attach the following notices to the program. It is safest 628 | to attach them to the start of each source file to most effectively 629 | state the exclusion of warranty; and each file should have at least 630 | the "copyright" line and a pointer to where the full notice is found. 631 | 632 | 633 | Copyright (C) 634 | 635 | This program is free software: you can redistribute it and/or modify 636 | it under the terms of the GNU Affero General Public License as published 637 | by the Free Software Foundation, either version 3 of the License, or 638 | (at your option) any later version. 639 | 640 | This program is distributed in the hope that it will be useful, 641 | but WITHOUT ANY WARRANTY; without even the implied warranty of 642 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 643 | GNU Affero General Public License for more details. 644 | 645 | You should have received a copy of the GNU Affero General Public License 646 | along with this program. If not, see . 647 | 648 | Also add information on how to contact you by electronic and paper mail. 649 | 650 | If your software can interact with users remotely through a computer 651 | network, you should also make sure that it provides a way for users to 652 | get its source. For example, if your program is a web application, its 653 | interface could display a "Source" link that leads users to an archive 654 | of the code. There are many ways you could offer source, and different 655 | solutions will be better for different programs; see section 13 for the 656 | specific requirements. 657 | 658 | You should also get your employer (if you work as a programmer) or school, 659 | if any, to sign a "copyright disclaimer" for the program, if necessary. 660 | For more information on this, and how to apply and follow the GNU AGPL, see 661 | . 662 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection (CVPR2022) 2 | By Choubo Ding, Guansong Pang, Chunhua Shen 3 | 4 | Official PyTorch implementation of ["Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection"](https://arxiv.org/abs/2203.14506). 5 | 6 | ## Prerequisites 7 | This code is written in `Python 3.7` and requires the packages listed in `requirements.txt`. Install with `pip install -r 8 | requirements.txt` preferably in a virtualenv. 9 | 10 | ## Run 11 | 12 | #### Step 1. Setup the Anomaly Detection Dataset 13 | Download the Anomaly Detection Dataset and convert it to MVTec AD format. (For datasets we used in the paper, we provided the [convert script](https://github.com/Choubo/DRA/tree/main/data).) 14 | The dataset folder structure should look like: 15 | ``` 16 | DATA_PATH/ 17 | subset_1/ 18 | train/ 19 | good/ 20 | test/ 21 | good/ 22 | defect_class_1/ 23 | defect_class_2/ 24 | defect_class_3/ 25 | ... 26 | ... 27 | ``` 28 | 29 | #### Step 2. Running DRA 30 | ```bash 31 | python train.py --dataset_root=./data/mvtec_anomaly_detection \ 32 | --classname=carpet \ 33 | --experiment_dir=./experiment 34 | ``` 35 | - `dataset_root` denotes the path of the dataset. 36 | - `classname` denotes the subset name of the dataset. 37 | - `experiment_dir` denotes the path to store the experiment setting and model weight. 38 | - `outlier_root` (*optional) given the path of the outlier dataset to disable pseudo augmentation and enable external data for pseudo head. 39 | - `know_class` (*optional) specify the anomaly class in the training set to experiment within the hard setting. 40 | 41 | ## Citation 42 | ```bibtex 43 | @inproceedings{ding2022catching, 44 | title={Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection}, 45 | author={Ding, Choubo and Pang, Guansong and Shen, Chunhua}, 46 | booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 47 | year={2022} 48 | } 49 | ``` -------------------------------------------------------------------------------- /data/README.md: -------------------------------------------------------------------------------- 1 | ## Convert Script for Anomaly Detection Dataset 2 | The convert script of the data set used in the paper. 3 | 4 | ### Usage 5 | #### Step 1. Download the Anomaly Detection Dataset 6 | Download and unzip the required data set. (The download link of all datasets are provided in the appendix of paper.) 7 | #### Step 2. Running the Convert Script 8 | Running the corresponding convert script with the argument `dataset_root` as the root of the dataset. 9 | 10 | e.g: 11 | ```bash 12 | python convert_AITEX.py --dataset_root=./AITEX 13 | ``` 14 | -------------------------------------------------------------------------------- /data/convert_AITEX.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | from sklearn.model_selection import train_test_split 4 | import cv2 5 | import argparse 6 | 7 | parser = argparse.ArgumentParser() 8 | parser.add_argument('--dataset_root', type=str, help="dataset root") 9 | args = parser.parse_args() 10 | 11 | DEFEAT_CLASS = {'002': "Broken_end", '006': "Broken_yarn", '010': "Broken_pick", 12 | '016': "Weft_curling", '019': "Fuzzyball", '022': "Cut_selvage", 13 | '023': "Crease", '025': "Warp_ball", '027': "Knots", 14 | '029': "Contamination", '030': "Nep", '036': "Weft_crack"} 15 | 16 | normal_images = list() 17 | normal_fname = list() 18 | outlier_images = list() 19 | outlier_labels = list() 20 | outlier_fname = list() 21 | 22 | 23 | normal_root = os.path.join(args.dataset_root, 'NODefect_images') 24 | normal_dirs = os.listdir(normal_root) 25 | for dir in normal_dirs: 26 | files = os.listdir(os.path.join(normal_root, dir)) 27 | for image in files: 28 | image_name = image.split('.')[0] 29 | image_data = cv2.imread(os.path.join(normal_root, dir, image)) 30 | for i in range(16): 31 | normal_images.append(image_data[:, i*256:(i+1)*256 ,:]) 32 | normal_fname.append(dir + '_' + image_name + '_' + str(i)) 33 | 34 | outlier_root = os.path.join(args.dataset_root, 'Defect_images/Defect_images') 35 | label_root = os.path.join(args.dataset_root, 'Mask_images/Mask_images') 36 | files = os.listdir(os.path.join(outlier_root)) 37 | for image in files: 38 | split_images = list() 39 | split_labels = list() 40 | image_name = image.split('.')[0] 41 | image_data = cv2.imread(os.path.join(outlier_root, image)) 42 | label_data = cv2.imread(os.path.join(label_root, image_name + '_mask.png')) 43 | if image_data.shape[1] % image_data.shape[0] == 0: 44 | count = image_data.shape[1]//image_data.shape[0] 45 | else: 46 | count = image_data.shape[1] // image_data.shape[0] + 1 47 | for i in range(count): 48 | split_images.append(image_data[:, i * 256:(i + 1) * 256, :]) 49 | split_labels.append(label_data[:, i * 256:(i + 1) * 256, :]) 50 | for i, (im, la) in enumerate(zip(split_images, split_labels)): 51 | if np.max(la) != 0: 52 | outlier_images.append(im) 53 | outlier_labels.append(la) 54 | outlier_fname.append(image_name + '_' + str(i)) 55 | 56 | normal_train, normal_test, normal_name_train, normal_name_test = train_test_split(normal_images, normal_fname, test_size=0.25, random_state=42) 57 | 58 | target_root = './AITEX_anomaly_detection/AITEX' 59 | train_root = os.path.join(target_root, 'train/good') 60 | if not os.path.exists(train_root): 61 | os.makedirs(train_root) 62 | for image, name in zip(normal_train, normal_name_train): 63 | cv2.imwrite(os.path.join(train_root, name + '.png'), image) 64 | 65 | test_root = os.path.join(target_root, 'test/good') 66 | if not os.path.exists(test_root): 67 | os.makedirs(test_root) 68 | for image, name in zip(normal_test, normal_name_test): 69 | cv2.imwrite(os.path.join(test_root, name + '.png'), image) 70 | 71 | for image, label, name in zip(outlier_images, outlier_labels, outlier_fname): 72 | defect_class = DEFEAT_CLASS[name.split('_')[1]] 73 | defect_root = os.path.join(target_root, 'test', defect_class) 74 | label_root = os.path.join(target_root, 'ground_truth', defect_class) 75 | if not os.path.exists(defect_root): 76 | os.makedirs(defect_root) 77 | if not os.path.exists(label_root): 78 | os.makedirs(label_root) 79 | cv2.imwrite(os.path.join(defect_root, name + '.png'), image) 80 | cv2.imwrite(os.path.join(label_root, name + '_mask.png'), label) 81 | 82 | print("Done") -------------------------------------------------------------------------------- /data/convert_BrainMRI.py: -------------------------------------------------------------------------------- 1 | import os 2 | from sklearn.model_selection import train_test_split 3 | import shutil 4 | import argparse 5 | 6 | parser = argparse.ArgumentParser() 7 | parser.add_argument('--dataset_root', type=str, help="dataset root") 8 | args = parser.parse_args() 9 | 10 | normal_root = os.path.join(args.dataset_root, 'no') 11 | outlier_root = os.path.join(args.dataset_root, 'yes') 12 | 13 | normal_fnames = os.listdir(normal_root) 14 | outlier_fnames = os.listdir(outlier_root) 15 | 16 | normal_train, normal_test, _, _ = train_test_split(normal_fnames, normal_fnames, test_size=0.25, random_state=42) 17 | 18 | target_root = './BrainMRI_anomaly_detection/brainmri' 19 | train_root = os.path.join(target_root, 'train/good') 20 | if not os.path.exists(train_root): 21 | os.makedirs(train_root) 22 | for f in normal_train: 23 | source = os.path.join(normal_root, f) 24 | shutil.copy(source, train_root) 25 | 26 | test_normal_root = os.path.join(target_root, 'test/good') 27 | if not os.path.exists(test_normal_root): 28 | os.makedirs(test_normal_root) 29 | for f in normal_test: 30 | source = os.path.join(normal_root, f) 31 | shutil.copy(source, test_normal_root) 32 | 33 | test_outlier_root = os.path.join(target_root, 'test/defect') 34 | if not os.path.exists(test_outlier_root): 35 | os.makedirs(test_outlier_root) 36 | for f in outlier_fnames: 37 | source = os.path.join(outlier_root, f) 38 | shutil.copy(source, test_outlier_root) 39 | 40 | print("Done") -------------------------------------------------------------------------------- /data/convert_HeadCT.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | from sklearn.model_selection import train_test_split 4 | import shutil 5 | import argparse 6 | 7 | parser = argparse.ArgumentParser() 8 | parser.add_argument('--dataset_root', type=str, help="dataset root") 9 | args = parser.parse_args() 10 | 11 | label_file = os.path.join(args.dataset_root, 'labels.csv') 12 | 13 | data = np.loadtxt(label_file, dtype=int, delimiter=',', skiprows=1) 14 | 15 | fnames = data[:, 0] 16 | label = data[:, 1] 17 | 18 | normal_fnames = fnames[label==0] 19 | outlier_fnames = fnames[label==1] 20 | 21 | normal_train, normal_test, _, _ = train_test_split(normal_fnames, normal_fnames, test_size=0.25, random_state=42) 22 | 23 | target_root = './HeadCT_anomaly_detection/headct' 24 | train_root = os.path.join(target_root, 'train/good') 25 | if not os.path.exists(train_root): 26 | os.makedirs(train_root) 27 | for f in normal_train: 28 | source = os.path.join(args.dataset_root, 'head_ct/head_ct/', '{:0>3d}.png'.format(f)) 29 | shutil.copy(source, train_root) 30 | 31 | test_normal_root = os.path.join(target_root, 'test/good') 32 | if not os.path.exists(test_normal_root): 33 | os.makedirs(test_normal_root) 34 | for f in normal_test: 35 | source = os.path.join(args.dataset_root, 'head_ct/head_ct/', '{:0>3d}.png'.format(f)) 36 | shutil.copy(source, test_normal_root) 37 | 38 | test_outlier_root = os.path.join(target_root, 'test/defect') 39 | if not os.path.exists(test_outlier_root): 40 | os.makedirs(test_outlier_root) 41 | for f in outlier_fnames: 42 | source = os.path.join(args.dataset_root, 'head_ct/head_ct/', '{:0>3d}.png'.format(f)) 43 | shutil.copy(source, test_outlier_root) 44 | 45 | print('Done') -------------------------------------------------------------------------------- /data/convert_MastCam.py: -------------------------------------------------------------------------------- 1 | import os 2 | import shutil 3 | import argparse 4 | 5 | parser = argparse.ArgumentParser() 6 | parser.add_argument('--dataset_root', type=str, help="dataset root") 7 | args = parser.parse_args() 8 | 9 | normal_train = list() 10 | normal_test = list() 11 | outlier_fnames = list() 12 | normal_root = os.path.join(args.dataset_root, 'train_typical') 13 | for file in os.listdir(normal_root): 14 | normal_train.append(os.path.join(normal_root, file)) 15 | 16 | test_normal_root = os.path.join(args.dataset_root, 'test_typical') 17 | for file in os.listdir(test_normal_root): 18 | normal_test.append(os.path.join(test_normal_root, file)) 19 | 20 | outlier_root = os.path.join(args.dataset_root, 'test_novel') 21 | for dir in os.listdir(outlier_root): 22 | class_root = os.path.join(outlier_root, dir) 23 | for file in os.listdir(class_root): 24 | outlier_fnames.append(os.path.join(class_root, file)) 25 | 26 | target_root = './MastCam_anomaly_detection/mastcam' 27 | train_root = os.path.join(target_root, 'train/good') 28 | if not os.path.exists(train_root): 29 | os.makedirs(train_root) 30 | for f in normal_train: 31 | shutil.copy(f, train_root) 32 | 33 | test_normal_root = os.path.join(target_root, 'test/good') 34 | if not os.path.exists(test_normal_root): 35 | os.makedirs(test_normal_root) 36 | for f in normal_test: 37 | shutil.copy(f, test_normal_root) 38 | 39 | test_outlier_root = os.path.join(target_root, 'test') 40 | if not os.path.exists(test_outlier_root): 41 | os.makedirs(test_outlier_root) 42 | for f in outlier_fnames: 43 | class_name = f.split('/')[-2] 44 | target_root = os.path.join(test_outlier_root,class_name) 45 | if not os.path.exists(target_root): 46 | os.makedirs(target_root) 47 | shutil.copy(f, target_root) 48 | 49 | print('Done') -------------------------------------------------------------------------------- /data/convert_SDD.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | from sklearn.model_selection import train_test_split 4 | import cv2 5 | import argparse 6 | 7 | parser = argparse.ArgumentParser() 8 | parser.add_argument('--dataset_root', type=str, help="dataset root") 9 | args = parser.parse_args() 10 | 11 | dirs = os.listdir(args.dataset_root) 12 | normal_images = list() 13 | normal_labels = list() 14 | normal_fname = list() 15 | outlier_images = list() 16 | outlier_labels = list() 17 | outlier_fname = list() 18 | for d in dirs: 19 | files = os.listdir(os.path.join(args.dataset_root, d)) 20 | images = list() 21 | for f in files: 22 | if 'jpg' in f[-3:]: 23 | images.append(f) 24 | 25 | for image in images: 26 | split_images = list() 27 | split_labels = list() 28 | image_name = image.split('.')[0] 29 | image_data = cv2.imread(os.path.join(args.dataset_root, d, image)) 30 | label_data = cv2.imread(os.path.join(args.dataset_root, d, image_name + '_label.bmp')) 31 | if image_data.shape != label_data.shape: 32 | raise ValueError 33 | image_length = image_data.shape[0] 34 | split_images.append(image_data[:image_length // 3, :, :]) 35 | split_images.append(image_data[image_length // 3:image_length * 2 // 3, :, :]) 36 | split_images.append(image_data[image_length * 2 // 3:, :, :]) 37 | split_labels.append(label_data[:image_length // 3, :, :]) 38 | split_labels.append(label_data[image_length // 3:image_length * 2 // 3, :, :]) 39 | split_labels.append(label_data[image_length * 2 // 3:, :, :]) 40 | for i, (im, la) in enumerate(zip(split_images, split_labels)): 41 | if np.max(la) != 0: 42 | outlier_images.append(im) 43 | outlier_labels.append(la) 44 | outlier_fname.append(d + '_' + image_name + '_' + str(i)) 45 | else: 46 | normal_images.append(im) 47 | normal_labels.append(la) 48 | normal_fname.append(d + '_' + image_name + '_' + str(i)) 49 | 50 | normal_train, normal_test, normal_name_train, normal_name_test = train_test_split(normal_images, normal_fname, test_size=0.25, random_state=42) 51 | 52 | target_root = './SDD_anomaly_detection/SDD' 53 | train_root = os.path.join(target_root, 'train/good') 54 | if not os.path.exists(train_root): 55 | os.makedirs(train_root) 56 | for image, name in zip(normal_train, normal_name_train): 57 | cv2.imwrite(os.path.join(train_root, name + '.png'), image) 58 | 59 | test_root = os.path.join(target_root, 'test/good') 60 | if not os.path.exists(test_root): 61 | os.makedirs(test_root) 62 | for image, name in zip(normal_test, normal_name_test): 63 | cv2.imwrite(os.path.join(test_root, name + '.png'), image) 64 | 65 | defect_root = os.path.join(target_root, 'test/defect') 66 | label_root = os.path.join(target_root, 'ground_truth/defect') 67 | if not os.path.exists(defect_root): 68 | os.makedirs(defect_root) 69 | if not os.path.exists(label_root): 70 | os.makedirs(label_root) 71 | for image, label, name in zip(outlier_images, outlier_labels, outlier_fname): 72 | cv2.imwrite(os.path.join(defect_root, name + '.png'), image) 73 | cv2.imwrite(os.path.join(label_root, name + '_mask.png'), label) 74 | 75 | print("Done") 76 | -------------------------------------------------------------------------------- /data/convert_elpv.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | from sklearn.model_selection import train_test_split 4 | import shutil 5 | import argparse 6 | 7 | parser = argparse.ArgumentParser() 8 | parser.add_argument('--dataset_root', type=str, help="dataset root") 9 | args = parser.parse_args() 10 | 11 | label_file = os.path.join(args.dataset_root, 'labels.csv') 12 | 13 | data = np.genfromtxt(label_file, dtype=['|S19', '0: 36 | self.test_threshold = int((len(normal_data)/(1-self.args.test_rate)) * self.args.test_rate) + self.args.nAnomaly 37 | 38 | self.ood_data = self.get_ood_data() 39 | 40 | if self.train is False: 41 | normal_data = list() 42 | split = 'test' 43 | normal_files = os.listdir(os.path.join(self.root, split, 'good')) 44 | for file in normal_files: 45 | if 'png' in file[-3:] or 'PNG' in file[-3:] or 'jpg' in file[-3:] or 'npy' in file[-3:]: 46 | normal_data.append(split + '/good/' + file) 47 | 48 | outlier_data, pollution_data = self.split_outlier() 49 | outlier_data.sort() 50 | 51 | normal_data = normal_data + pollution_data 52 | 53 | normal_label = np.zeros(len(normal_data)).tolist() 54 | outlier_label = np.ones(len(outlier_data)).tolist() 55 | 56 | self.images = normal_data + outlier_data 57 | self.labels = np.array(normal_label + outlier_label) 58 | self.normal_idx = np.argwhere(self.labels == 0).flatten() 59 | self.outlier_idx = np.argwhere(self.labels == 1).flatten() 60 | 61 | def get_ood_data(self): 62 | ood_data = list() 63 | if self.args.outlier_root is None: 64 | return None 65 | dataset_classes = os.listdir(self.args.outlier_root) 66 | for cl in dataset_classes: 67 | if cl == self.args.classname: 68 | continue 69 | cl_root = os.path.join(self.args.outlier_root, cl, 'train', 'good') 70 | ood_file = os.listdir(cl_root) 71 | for file in ood_file: 72 | if 'png' in file[-3:] or 'PNG' in file[-3:] or 'jpg' in file[-3:] or 'npy' in file[-3:]: 73 | ood_data.append(os.path.join(cl_root, file)) 74 | return ood_data 75 | 76 | def split_outlier(self): 77 | outlier_data_dir = os.path.join(self.root, 'test') 78 | outlier_classes = os.listdir(outlier_data_dir) 79 | if self.know_class in outlier_classes: 80 | print("Know outlier class: " + self.know_class) 81 | outlier_data = list() 82 | know_class_data = list() 83 | for cl in outlier_classes: 84 | if cl == 'good': 85 | continue 86 | outlier_file = os.listdir(os.path.join(outlier_data_dir, cl)) 87 | for file in outlier_file: 88 | if 'png' in file[-3:] or 'PNG' in file[-3:] or 'jpg' in file[-3:] or 'npy' in file[-3:]: 89 | if cl == self.know_class: 90 | know_class_data.append('test/' + cl + '/' + file) 91 | else: 92 | outlier_data.append('test/' + cl + '/' + file) 93 | np.random.RandomState(self.args.ramdn_seed).shuffle(know_class_data) 94 | know_outlier = know_class_data[0:self.args.nAnomaly] 95 | unknow_outlier = outlier_data 96 | if self.train: 97 | return know_outlier, list() 98 | else: 99 | return unknow_outlier, list() 100 | 101 | 102 | outlier_data = list() 103 | for cl in outlier_classes: 104 | if cl == 'good': 105 | continue 106 | outlier_file = os.listdir(os.path.join(outlier_data_dir, cl)) 107 | for file in outlier_file: 108 | if 'png' in file[-3:] or 'PNG' in file[-3:] or 'jpg' in file[-3:] or 'npy' in file[-3:]: 109 | outlier_data.append('test/' + cl + '/' + file) 110 | np.random.RandomState(self.args.ramdn_seed).shuffle(outlier_data) 111 | if self.train: 112 | return outlier_data[0:self.args.nAnomaly], outlier_data[self.args.nAnomaly:self.args.nAnomaly + self.nPollution] 113 | else: 114 | return outlier_data[self.test_threshold:], list() 115 | 116 | def load_image(self, path): 117 | if 'npy' in path[-3:]: 118 | img = np.load(path).astype(np.uint8) 119 | img = img[:, :, :3] 120 | return Image.fromarray(img) 121 | return Image.open(path).convert('RGB') 122 | 123 | def transform_train(self): 124 | composed_transforms = transforms.Compose([ 125 | transforms.Resize((self.args.img_size,self.args.img_size)), 126 | transforms.RandomRotation(180), 127 | transforms.ToTensor(), 128 | transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) 129 | return composed_transforms 130 | 131 | def transform_pseudo(self): 132 | composed_transforms = transforms.Compose([ 133 | transforms.Resize((self.args.img_size,self.args.img_size)), 134 | CutMix(), 135 | transforms.RandomRotation(180), 136 | transforms.ToTensor(), 137 | transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) 138 | return composed_transforms 139 | 140 | def transform_test(self): 141 | composed_transforms = transforms.Compose([ 142 | transforms.Resize((self.args.img_size, self.args.img_size)), 143 | transforms.ToTensor(), 144 | transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) 145 | return composed_transforms 146 | 147 | def __len__(self): 148 | return len(self.images) 149 | 150 | def __getitem__(self, index): 151 | rnd = random.randint(0, 1) 152 | if index in self.normal_idx and rnd == 0 and self.train: 153 | if self.ood_data is None: 154 | index = random.choice(self.normal_idx) 155 | image = self.load_image(os.path.join(self.root, self.images[index])) 156 | transform = self.transform_pseudo 157 | else: 158 | image = self.load_image(random.choice(self.ood_data)) 159 | transform = self.transform 160 | label = 2 161 | else: 162 | image = self.load_image(os.path.join(self.root, self.images[index])) 163 | transform = self.transform 164 | label = self.labels[index] 165 | sample = {'image': transform(image), 'label': label} 166 | return sample 167 | -------------------------------------------------------------------------------- /modeling/layers/__init__.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from modeling.layers.deviation_loss import DeviationLoss 3 | from modeling.layers.binary_focal_loss import BinaryFocalLoss 4 | 5 | def build_criterion(criterion): 6 | if criterion == "deviation": 7 | print("Loss : Deviation") 8 | return DeviationLoss() 9 | elif criterion == "BCE": 10 | print("Loss : Binary Cross Entropy") 11 | return torch.nn.BCEWithLogitsLoss() 12 | elif criterion == "focal": 13 | print("Loss : Focal") 14 | return BinaryFocalLoss() 15 | elif criterion == "CE": 16 | print("Loss : CE") 17 | return torch.nn.CrossEntropyLoss() 18 | else: 19 | raise NotImplementedError -------------------------------------------------------------------------------- /modeling/layers/binary_focal_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | class BinaryFocalLoss(nn.Module): 6 | def __init__(self, alpha=1, gamma=2, logits=True, reduce=True): 7 | super(BinaryFocalLoss, self).__init__() 8 | self.alpha = alpha 9 | self.gamma = gamma 10 | self.logits = logits 11 | self.reduce = reduce 12 | 13 | def forward(self, inputs, targets): 14 | if self.logits: 15 | BCE_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction='none') 16 | else: 17 | BCE_loss = F.binary_cross_entropy(inputs, targets, reduction='none') 18 | pt = torch.exp(-BCE_loss) 19 | F_loss = self.alpha * (1-pt)**self.gamma * BCE_loss 20 | 21 | if self.reduce: 22 | return torch.mean(F_loss) 23 | else: 24 | return F_loss -------------------------------------------------------------------------------- /modeling/layers/deviation_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | class DeviationLoss(nn.Module): 5 | 6 | def __init__(self): 7 | super().__init__() 8 | 9 | def forward(self, y_pred, y_true): 10 | confidence_margin = 5. 11 | ref = torch.normal(mean=0., std=torch.full([5000], 1.)).cuda() 12 | dev = (y_pred - torch.mean(ref)) / torch.std(ref) 13 | inlier_loss = torch.abs(dev) 14 | outlier_loss = torch.abs((confidence_margin - dev).clamp_(min=0.)) 15 | dev_loss = (1 - y_true) * inlier_loss + y_true * outlier_loss 16 | return torch.mean(dev_loss) 17 | -------------------------------------------------------------------------------- /modeling/net.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from modeling.networks.backbone import build_feature_extractor, NET_OUT_DIM 5 | 6 | 7 | class HolisticHead(nn.Module): 8 | def __init__(self, in_dim, dropout=0): 9 | super(HolisticHead, self).__init__() 10 | self.fc1 = nn.Linear(in_dim, 256) 11 | self.fc2 = nn.Linear(256, 1) 12 | self.drop = nn.Dropout(dropout) 13 | 14 | def forward(self, x): 15 | x = F.adaptive_avg_pool2d(x, (1, 1)) 16 | x = x.view(x.size(0), -1) 17 | x = self.drop(F.relu(self.fc1(x))) 18 | x = self.fc2(x) 19 | return torch.abs(x) 20 | 21 | 22 | class PlainHead(nn.Module): 23 | def __init__(self, in_dim, topk_rate=0.1): 24 | super(PlainHead, self).__init__() 25 | self.scoring = nn.Conv2d(in_channels=in_dim, out_channels=1, kernel_size=1, padding=0) 26 | self.topk_rate = topk_rate 27 | 28 | def forward(self, x): 29 | x = self.scoring(x) 30 | x = x.view(int(x.size(0)), -1) 31 | topk = max(int(x.size(1) * self.topk_rate), 1) 32 | x = torch.topk(torch.abs(x), topk, dim=1)[0] 33 | x = torch.mean(x, dim=1).view(-1, 1) 34 | return x 35 | 36 | 37 | class CompositeHead(PlainHead): 38 | def __init__(self, in_dim, topk=0.1): 39 | super(CompositeHead, self).__init__(in_dim, topk) 40 | self.conv = nn.Sequential(nn.Conv2d(in_dim, in_dim, 3, padding=1), 41 | nn.BatchNorm2d(in_dim), 42 | nn.ReLU()) 43 | 44 | def forward(self, x, ref): 45 | ref = torch.mean(ref, dim=0).repeat([x.size(0), 1, 1, 1]) 46 | x = ref - x 47 | x = self.conv(x) 48 | x = super().forward(x) 49 | return x 50 | 51 | 52 | class DRA(nn.Module): 53 | def __init__(self, cfg, backbone="resnet18"): 54 | super(DRA, self).__init__() 55 | self.cfg = cfg 56 | self.feature_extractor = build_feature_extractor(backbone, cfg) 57 | self.in_c = NET_OUT_DIM[backbone] 58 | self.holistic_head = HolisticHead(self.in_c) 59 | self.seen_head = PlainHead(self.in_c, self.cfg.topk) 60 | self.pseudo_head = PlainHead(self.in_c, self.cfg.topk) 61 | self.composite_head = CompositeHead(self.in_c, self.cfg.topk) 62 | 63 | def forward(self, image, label): 64 | image_pyramid = list() 65 | for i in range(self.cfg.total_heads): 66 | image_pyramid.append(list()) 67 | for s in range(self.cfg.n_scales): 68 | image_scaled = F.interpolate(image, size=self.cfg.img_size // (2 ** s)) if s > 0 else image 69 | feature = self.feature_extractor(image_scaled) 70 | 71 | ref_feature = feature[:self.cfg.nRef, :, :, :] 72 | feature = feature[self.cfg.nRef:, :, :, :] 73 | 74 | if self.training: 75 | normal_scores = self.holistic_head(feature) 76 | abnormal_scores = self.seen_head(feature[label != 2]) 77 | dummy_scores = self.pseudo_head(feature[label != 1]) 78 | comparison_scores = self.composite_head(feature, ref_feature) 79 | else: 80 | normal_scores = self.holistic_head(feature) 81 | abnormal_scores = self.seen_head(feature) 82 | dummy_scores = self.pseudo_head(feature) 83 | comparison_scores = self.composite_head(feature, ref_feature) 84 | for i, scores in enumerate([normal_scores, abnormal_scores, dummy_scores, comparison_scores]): 85 | image_pyramid[i].append(scores) 86 | for i in range(self.cfg.total_heads): 87 | image_pyramid[i] = torch.cat(image_pyramid[i], dim=1) 88 | image_pyramid[i] = torch.mean(image_pyramid[i], dim=1) 89 | return image_pyramid 90 | 91 | 92 | -------------------------------------------------------------------------------- /modeling/networks/backbone.py: -------------------------------------------------------------------------------- 1 | from torchvision.models import alexnet 2 | from modeling.networks.resnet18 import FeatureRESNET18 3 | 4 | NET_OUT_DIM = {'alexnet': 256, 'resnet18': 512} 5 | 6 | def build_feature_extractor(backbone, cfg): 7 | if backbone == "alexnet": 8 | print("Feature extractor: AlexNet") 9 | return alexnet(pretrained=True).features 10 | elif backbone == "resnet18": 11 | print("Feature extractor: ResNet-18") 12 | return FeatureRESNET18() 13 | else: 14 | raise NotImplementedError -------------------------------------------------------------------------------- /modeling/networks/resnet.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import Tensor 3 | import torch.nn as nn 4 | from torchvision.models.utils import load_state_dict_from_url 5 | from typing import Type, Any, Callable, Union, List, Optional 6 | 7 | 8 | __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 9 | 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 10 | 'wide_resnet50_2', 'wide_resnet101_2'] 11 | 12 | 13 | model_urls = { 14 | 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 15 | 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 16 | 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 17 | 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 18 | 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', 19 | 'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth', 20 | 'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth', 21 | 'wide_resnet50_2': 'https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth', 22 | 'wide_resnet101_2': 'https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth', 23 | } 24 | 25 | 26 | def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d: 27 | """3x3 convolution with padding""" 28 | return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, 29 | padding=dilation, groups=groups, bias=False, dilation=dilation) 30 | 31 | 32 | def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d: 33 | """1x1 convolution""" 34 | return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) 35 | 36 | 37 | class BasicBlock(nn.Module): 38 | expansion: int = 1 39 | 40 | def __init__( 41 | self, 42 | inplanes: int, 43 | planes: int, 44 | stride: int = 1, 45 | downsample: Optional[nn.Module] = None, 46 | groups: int = 1, 47 | base_width: int = 64, 48 | dilation: int = 1, 49 | norm_layer: Optional[Callable[..., nn.Module]] = None 50 | ) -> None: 51 | super(BasicBlock, self).__init__() 52 | if norm_layer is None: 53 | norm_layer = nn.BatchNorm2d 54 | if groups != 1 or base_width != 64: 55 | raise ValueError('BasicBlock only supports groups=1 and base_width=64') 56 | if dilation > 1: 57 | raise NotImplementedError("Dilation > 1 not supported in BasicBlock") 58 | # Both self.conv1 and self.downsample layers downsample the input when stride != 1 59 | self.conv1 = conv3x3(inplanes, planes, stride) 60 | self.bn1 = norm_layer(planes) 61 | self.relu = nn.ReLU(inplace=True) 62 | self.conv2 = conv3x3(planes, planes) 63 | self.bn2 = norm_layer(planes) 64 | self.downsample = downsample 65 | self.stride = stride 66 | 67 | def forward(self, x: Tensor) -> Tensor: 68 | identity = x 69 | 70 | out = self.conv1(x) 71 | out = self.bn1(out) 72 | out = self.relu(out) 73 | 74 | out = self.conv2(out) 75 | out = self.bn2(out) 76 | 77 | if self.downsample is not None: 78 | identity = self.downsample(x) 79 | 80 | out += identity 81 | out = self.relu(out) 82 | 83 | return out 84 | 85 | 86 | class Bottleneck(nn.Module): 87 | # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2) 88 | # while original implementation places the stride at the first 1x1 convolution(self.conv1) 89 | # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385. 90 | # This variant is also known as ResNet V1.5 and improves accuracy according to 91 | # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch. 92 | 93 | expansion: int = 4 94 | 95 | def __init__( 96 | self, 97 | inplanes: int, 98 | planes: int, 99 | stride: int = 1, 100 | downsample: Optional[nn.Module] = None, 101 | groups: int = 1, 102 | base_width: int = 64, 103 | dilation: int = 1, 104 | norm_layer: Optional[Callable[..., nn.Module]] = None 105 | ) -> None: 106 | super(Bottleneck, self).__init__() 107 | if norm_layer is None: 108 | norm_layer = nn.BatchNorm2d 109 | width = int(planes * (base_width / 64.)) * groups 110 | # Both self.conv2 and self.downsample layers downsample the input when stride != 1 111 | self.conv1 = conv1x1(inplanes, width) 112 | self.bn1 = norm_layer(width) 113 | self.conv2 = conv3x3(width, width, stride, groups, dilation) 114 | self.bn2 = norm_layer(width) 115 | self.conv3 = conv1x1(width, planes * self.expansion) 116 | self.bn3 = norm_layer(planes * self.expansion) 117 | self.relu = nn.ReLU(inplace=True) 118 | self.downsample = downsample 119 | self.stride = stride 120 | 121 | def forward(self, x: Tensor) -> Tensor: 122 | identity = x 123 | 124 | out = self.conv1(x) 125 | out = self.bn1(out) 126 | out = self.relu(out) 127 | 128 | out = self.conv2(out) 129 | out = self.bn2(out) 130 | out = self.relu(out) 131 | 132 | out = self.conv3(out) 133 | out = self.bn3(out) 134 | 135 | if self.downsample is not None: 136 | identity = self.downsample(x) 137 | 138 | out += identity 139 | out = self.relu(out) 140 | 141 | return out 142 | 143 | 144 | class ResNet(nn.Module): 145 | 146 | def __init__( 147 | self, 148 | block: Type[Union[BasicBlock, Bottleneck]], 149 | layers: List[int], 150 | num_classes: int = 1000, 151 | zero_init_residual: bool = False, 152 | groups: int = 1, 153 | width_per_group: int = 64, 154 | replace_stride_with_dilation: Optional[List[bool]] = None, 155 | norm_layer: Optional[Callable[..., nn.Module]] = None 156 | ) -> None: 157 | super(ResNet, self).__init__() 158 | if norm_layer is None: 159 | norm_layer = nn.BatchNorm2d 160 | self._norm_layer = norm_layer 161 | 162 | self.inplanes = 64 163 | self.dilation = 1 164 | if replace_stride_with_dilation is None: 165 | # each element in the tuple indicates if we should replace 166 | # the 2x2 stride with a dilated convolution instead 167 | replace_stride_with_dilation = [False, False, False] 168 | if len(replace_stride_with_dilation) != 3: 169 | raise ValueError("replace_stride_with_dilation should be None " 170 | "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) 171 | self.groups = groups 172 | self.base_width = width_per_group 173 | self.conv1 = nn.Conv2d(6, self.inplanes, kernel_size=7, stride=2, padding=3, 174 | bias=False) 175 | self.bn1 = norm_layer(self.inplanes) 176 | self.relu = nn.ReLU(inplace=True) 177 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 178 | self.layer1 = self._make_layer(block, 64, layers[0]) 179 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2, 180 | dilate=replace_stride_with_dilation[0]) 181 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2, 182 | dilate=replace_stride_with_dilation[1]) 183 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2, 184 | dilate=replace_stride_with_dilation[2]) 185 | self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) 186 | self.fc = nn.Linear(512 * block.expansion, num_classes) 187 | 188 | for m in self.modules(): 189 | if isinstance(m, nn.Conv2d): 190 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 191 | elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): 192 | nn.init.constant_(m.weight, 1) 193 | nn.init.constant_(m.bias, 0) 194 | 195 | # Zero-initialize the last BN in each residual branch, 196 | # so that the residual branch starts with zeros, and each residual block behaves like an identity. 197 | # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 198 | if zero_init_residual: 199 | for m in self.modules(): 200 | if isinstance(m, Bottleneck): 201 | nn.init.constant_(m.bn3.weight, 0) # type: ignore[arg-type] 202 | elif isinstance(m, BasicBlock): 203 | nn.init.constant_(m.bn2.weight, 0) # type: ignore[arg-type] 204 | 205 | def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int, 206 | stride: int = 1, dilate: bool = False) -> nn.Sequential: 207 | norm_layer = self._norm_layer 208 | downsample = None 209 | previous_dilation = self.dilation 210 | if dilate: 211 | self.dilation *= stride 212 | stride = 1 213 | if stride != 1 or self.inplanes != planes * block.expansion: 214 | downsample = nn.Sequential( 215 | conv1x1(self.inplanes, planes * block.expansion, stride), 216 | norm_layer(planes * block.expansion), 217 | ) 218 | 219 | layers = [] 220 | layers.append(block(self.inplanes, planes, stride, downsample, self.groups, 221 | self.base_width, previous_dilation, norm_layer)) 222 | self.inplanes = planes * block.expansion 223 | for _ in range(1, blocks): 224 | layers.append(block(self.inplanes, planes, groups=self.groups, 225 | base_width=self.base_width, dilation=self.dilation, 226 | norm_layer=norm_layer)) 227 | 228 | return nn.Sequential(*layers) 229 | 230 | def _forward_impl(self, x: Tensor) -> Tensor: 231 | # See note [TorchScript super()] 232 | x = self.conv1(x) 233 | x = self.bn1(x) 234 | x = self.relu(x) 235 | x = self.maxpool(x) 236 | 237 | x = self.layer1(x) 238 | x = self.layer2(x) 239 | x = self.layer3(x) 240 | x = self.layer4(x) 241 | 242 | x = self.avgpool(x) 243 | x = torch.flatten(x, 1) 244 | x = self.fc(x) 245 | 246 | return x 247 | 248 | def forward(self, x: Tensor) -> Tensor: 249 | return self._forward_impl(x) 250 | 251 | 252 | def _resnet( 253 | arch: str, 254 | block: Type[Union[BasicBlock, Bottleneck]], 255 | layers: List[int], 256 | pretrained: bool, 257 | progress: bool, 258 | **kwargs: Any 259 | ) -> ResNet: 260 | model = ResNet(block, layers, **kwargs) 261 | if pretrained: 262 | state_dict = load_state_dict_from_url(model_urls[arch], 263 | progress=progress) 264 | model.load_state_dict(state_dict) 265 | return model 266 | 267 | 268 | def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 269 | r"""ResNet-18 model from 270 | `"Deep Residual Learning for Image Recognition" `_. 271 | Args: 272 | pretrained (bool): If True, returns a model pre-trained on ImageNet 273 | progress (bool): If True, displays a progress bar of the download to stderr 274 | """ 275 | return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress, 276 | **kwargs) 277 | 278 | 279 | def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 280 | r"""ResNet-34 model from 281 | `"Deep Residual Learning for Image Recognition" `_. 282 | Args: 283 | pretrained (bool): If True, returns a model pre-trained on ImageNet 284 | progress (bool): If True, displays a progress bar of the download to stderr 285 | """ 286 | return _resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress, 287 | **kwargs) 288 | 289 | 290 | def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 291 | r"""ResNet-50 model from 292 | `"Deep Residual Learning for Image Recognition" `_. 293 | Args: 294 | pretrained (bool): If True, returns a model pre-trained on ImageNet 295 | progress (bool): If True, displays a progress bar of the download to stderr 296 | """ 297 | return _resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress, 298 | **kwargs) 299 | 300 | 301 | def resnet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 302 | r"""ResNet-101 model from 303 | `"Deep Residual Learning for Image Recognition" `_. 304 | Args: 305 | pretrained (bool): If True, returns a model pre-trained on ImageNet 306 | progress (bool): If True, displays a progress bar of the download to stderr 307 | """ 308 | return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress, 309 | **kwargs) 310 | 311 | 312 | def resnet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 313 | r"""ResNet-152 model from 314 | `"Deep Residual Learning for Image Recognition" `_. 315 | Args: 316 | pretrained (bool): If True, returns a model pre-trained on ImageNet 317 | progress (bool): If True, displays a progress bar of the download to stderr 318 | """ 319 | return _resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress, 320 | **kwargs) 321 | 322 | 323 | def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 324 | r"""ResNeXt-50 32x4d model from 325 | `"Aggregated Residual Transformation for Deep Neural Networks" `_. 326 | Args: 327 | pretrained (bool): If True, returns a model pre-trained on ImageNet 328 | progress (bool): If True, displays a progress bar of the download to stderr 329 | """ 330 | kwargs['groups'] = 32 331 | kwargs['width_per_group'] = 4 332 | return _resnet('resnext50_32x4d', Bottleneck, [3, 4, 6, 3], 333 | pretrained, progress, **kwargs) 334 | 335 | 336 | def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 337 | r"""ResNeXt-101 32x8d model from 338 | `"Aggregated Residual Transformation for Deep Neural Networks" `_. 339 | Args: 340 | pretrained (bool): If True, returns a model pre-trained on ImageNet 341 | progress (bool): If True, displays a progress bar of the download to stderr 342 | """ 343 | kwargs['groups'] = 32 344 | kwargs['width_per_group'] = 8 345 | return _resnet('resnext101_32x8d', Bottleneck, [3, 4, 23, 3], 346 | pretrained, progress, **kwargs) 347 | 348 | 349 | def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 350 | r"""Wide ResNet-50-2 model from 351 | `"Wide Residual Networks" `_. 352 | The model is the same as ResNet except for the bottleneck number of channels 353 | which is twice larger in every block. The number of channels in outer 1x1 354 | convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 355 | channels, and in Wide ResNet-50-2 has 2048-1024-2048. 356 | Args: 357 | pretrained (bool): If True, returns a model pre-trained on ImageNet 358 | progress (bool): If True, displays a progress bar of the download to stderr 359 | """ 360 | kwargs['width_per_group'] = 64 * 2 361 | return _resnet('wide_resnet50_2', Bottleneck, [3, 4, 6, 3], 362 | pretrained, progress, **kwargs) 363 | 364 | 365 | def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 366 | r"""Wide ResNet-101-2 model from 367 | `"Wide Residual Networks" `_. 368 | The model is the same as ResNet except for the bottleneck number of channels 369 | which is twice larger in every block. The number of channels in outer 1x1 370 | convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 371 | channels, and in Wide ResNet-50-2 has 2048-1024-2048. 372 | Args: 373 | pretrained (bool): If True, returns a model pre-trained on ImageNet 374 | progress (bool): If True, displays a progress bar of the download to stderr 375 | """ 376 | kwargs['width_per_group'] = 64 * 2 377 | return _resnet('wide_resnet101_2', Bottleneck, [3, 4, 23, 3], 378 | pretrained, progress, **kwargs) -------------------------------------------------------------------------------- /modeling/networks/resnet18.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from torchvision import models 3 | 4 | 5 | class FeatureRESNET18(nn.Module): 6 | def __init__(self): 7 | super(FeatureRESNET18, self).__init__() 8 | self.net = models.resnet18(pretrained=True) 9 | 10 | def forward(self, x): 11 | x = self.net.conv1(x) 12 | x = self.net.bn1(x) 13 | x = self.net.relu(x) 14 | x = self.net.maxpool(x) 15 | x = self.net.layer1(x) 16 | x = self.net.layer2(x) 17 | x = self.net.layer3(x) 18 | x = self.net.layer4(x) 19 | return x 20 | 21 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | matplotlib==3.3.3 2 | numpy==1.18.5 3 | pandas==0.25.3 4 | Pillow==8.4.0 5 | scikit_learn==1.0.1 6 | torch==1.1.0 7 | torchvision==0.3.0 8 | tqdm==4.54.0 9 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | import argparse 6 | import os 7 | 8 | from dataloaders.dataloader import initDataloader 9 | from modeling.net import DRA 10 | from tqdm import tqdm 11 | from sklearn.metrics import average_precision_score, roc_auc_score 12 | from modeling.layers import build_criterion 13 | import random 14 | 15 | import matplotlib.pyplot as plt 16 | 17 | WEIGHT_DIR = './weights' 18 | 19 | 20 | class Trainer(object): 21 | 22 | def __init__(self, args): 23 | self.args = args 24 | # Define Dataloader 25 | kwargs = {'num_workers': args.workers} 26 | self.train_loader, self.test_loader= initDataloader.build(args, **kwargs) 27 | if self.args.total_heads == 4: 28 | temp_args = args 29 | temp_args.batch_size = self.args.nRef 30 | temp_args.nAnomaly = 0 31 | self.ref_loader, _ = initDataloader.build(temp_args, **kwargs) 32 | self.ref = iter(self.ref_loader) 33 | 34 | self.model = DRA(args, backbone=self.args.backbone) 35 | 36 | if self.args.pretrain_dir != None: 37 | self.model.load_state_dict(torch.load(self.args.pretrain_dir)) 38 | print('Load pretrain weight from: ' + self.args.pretrain_dir) 39 | 40 | self.criterion = build_criterion(args.criterion) 41 | 42 | self.optimizer = torch.optim.Adam(self.model.parameters(), lr=0.0002, weight_decay=1e-5) 43 | self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=10, gamma=0.1) 44 | 45 | def generate_target(self, target, eval=False): 46 | targets = list() 47 | if eval: 48 | targets.append(target==0) 49 | targets.append(target) 50 | targets.append(target) 51 | targets.append(target) 52 | return targets 53 | else: 54 | temp_t = target != 0 55 | targets.append(target == 0) 56 | targets.append(temp_t[target != 2]) 57 | targets.append(temp_t[target != 1]) 58 | targets.append(target != 0) 59 | return targets 60 | 61 | def training(self, epoch): 62 | train_loss = 0.0 63 | class_loss = list() 64 | for j in range(self.args.total_heads): 65 | class_loss.append(0.0) 66 | self.model.train() 67 | self.scheduler.step() 68 | tbar = tqdm(self.train_loader) 69 | for idx, sample in enumerate(tbar): 70 | image, target = sample['image'], sample['label'] 71 | if self.args.cuda: 72 | image, target = image.cuda(), target.cuda() 73 | if self.args.total_heads == 4: 74 | try: 75 | ref_image = next(self.ref)['image'] 76 | except StopIteration: 77 | self.ref = iter(self.ref_loader) 78 | ref_image = next(self.ref)['image'] 79 | ref_image = ref_image.cuda() 80 | image = torch.cat([ref_image, image], dim=0) 81 | 82 | outputs = self.model(image, target) 83 | targets = self.generate_target(target) 84 | 85 | losses = list() 86 | for i in range(self.args.total_heads): 87 | if self.args.criterion == 'CE': 88 | prob = F.softmax(outputs[i], dim=1) 89 | losses.append(self.criterion(prob, targets[i].long()).view(-1, 1)) 90 | else: 91 | losses.append(self.criterion(outputs[i], targets[i].float()).view(-1, 1)) 92 | 93 | loss = torch.cat(losses) 94 | loss = torch.sum(loss) 95 | 96 | self.optimizer.zero_grad() 97 | loss.backward() 98 | 99 | torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0) 100 | self.optimizer.step() 101 | train_loss += loss.item() 102 | for i in range(self.args.total_heads): 103 | class_loss[i] += losses[i].item() 104 | 105 | tbar.set_description('Epoch:%d, Train loss: %.3f' % (epoch, train_loss / (idx + 1))) 106 | 107 | 108 | def normalization(self, data): 109 | return data 110 | 111 | def eval(self): 112 | self.model.eval() 113 | tbar = tqdm(self.test_loader, desc='\r') 114 | test_loss = 0.0 115 | class_pred = list() 116 | for i in range(self.args.total_heads): 117 | class_pred.append(np.array([])) 118 | total_target = np.array([]) 119 | for i, sample in enumerate(tbar): 120 | image, target = sample['image'], sample['label'] 121 | if self.args.cuda: 122 | image, target = image.cuda(), target.cuda() 123 | 124 | if self.args.total_heads == 4: 125 | try: 126 | ref_image = next(self.ref)['image'] 127 | except StopIteration: 128 | self.ref = iter(self.ref_loader) 129 | ref_image = next(self.ref)['image'] 130 | ref_image = ref_image.cuda() 131 | image = torch.cat([ref_image, image], dim=0) 132 | 133 | with torch.no_grad(): 134 | outputs = self.model(image, target) 135 | targets = self.generate_target(target, eval=True) 136 | 137 | losses = list() 138 | for i in range(self.args.total_heads): 139 | if self.args.criterion == 'CE': 140 | prob = F.softmax(outputs[i], dim=1) 141 | losses.append(self.criterion(prob, targets[i].long())) 142 | else: 143 | losses.append(self.criterion(outputs[i], targets[i].float())) 144 | 145 | loss = losses[0] 146 | for i in range(1, self.args.total_heads): 147 | loss += losses[i] 148 | 149 | test_loss += loss.item() 150 | tbar.set_description('Test loss: %.3f' % (test_loss / (i + 1))) 151 | 152 | for i in range(self.args.total_heads): 153 | if i == 0: 154 | data = -1 * outputs[i].data.cpu().numpy() 155 | else: 156 | data = outputs[i].data.cpu().numpy() 157 | class_pred[i] = np.append(class_pred[i], data) 158 | total_target = np.append(total_target, target.cpu().numpy()) 159 | 160 | total_pred = self.normalization(class_pred[0]) 161 | for i in range(1, self.args.total_heads): 162 | total_pred = total_pred + self.normalization(class_pred[i]) 163 | 164 | with open(self.args.experiment_dir + '/result.txt', mode='a+', encoding="utf-8") as w: 165 | for label, score in zip(total_target, total_pred): 166 | w.write(str(label) + ' ' + str(score) + "\n") 167 | 168 | total_roc, total_pr = aucPerformance(total_pred, total_target) 169 | 170 | normal_mask = total_target == 0 171 | outlier_mask = total_target == 1 172 | plt.clf() 173 | plt.bar(np.arange(total_pred.size)[normal_mask], total_pred[normal_mask], color='green') 174 | plt.bar(np.arange(total_pred.size)[outlier_mask], total_pred[outlier_mask], color='red') 175 | plt.ylabel("Anomaly score") 176 | plt.savefig(args.experiment_dir + "/vis.png") 177 | return total_roc, total_pr 178 | 179 | def save_weights(self, filename): 180 | # if not os.path.exists(WEIGHT_DIR): 181 | # os.makedirs(WEIGHT_DIR) 182 | torch.save(self.model.state_dict(), os.path.join(args.experiment_dir, filename)) 183 | 184 | def load_weights(self, filename): 185 | path = os.path.join(WEIGHT_DIR, filename) 186 | self.model.load_state_dict(torch.load(path)) 187 | 188 | def init_network_weights_from_pretraining(self): 189 | 190 | net_dict = self.model.state_dict() 191 | ae_net_dict = self.ae_model.state_dict() 192 | 193 | ae_net_dict = {k: v for k, v in ae_net_dict.items() if k in net_dict} 194 | net_dict.update(ae_net_dict) 195 | self.model.load_state_dict(net_dict) 196 | 197 | def aucPerformance(mse, labels, prt=True): 198 | roc_auc = roc_auc_score(labels, mse) 199 | ap = average_precision_score(labels, mse) 200 | if prt: 201 | print("AUC-ROC: %.4f, AUC-PR: %.4f" % (roc_auc, ap)) 202 | return roc_auc, ap; 203 | 204 | if __name__ == '__main__': 205 | parser = argparse.ArgumentParser() 206 | parser.add_argument("--batch_size", type=int, default=48, help="batch size used in SGD") 207 | parser.add_argument("--steps_per_epoch", type=int, default=20, help="the number of batches per epoch") 208 | parser.add_argument("--epochs", type=int, default=30, help="the number of epochs") 209 | parser.add_argument("--cont_rate", type=float, default=0.0, help="the outlier contamination rate in the training data") 210 | parser.add_argument("--test_threshold", type=int, default=0, 211 | help="the outlier contamination rate in the training data") 212 | parser.add_argument("--test_rate", type=float, default=0.0, 213 | help="the outlier contamination rate in the training data") 214 | parser.add_argument("--dataset", type=str, default='mvtecad', help="a list of data set names") 215 | parser.add_argument("--ramdn_seed", type=int, default=42, help="the random seed number") 216 | parser.add_argument('--workers', type=int, default=4, metavar='N', help='dataloader threads') 217 | parser.add_argument('--no-cuda', action='store_true', default=False, help='disables CUDA training') 218 | parser.add_argument('--savename', type=str, default='model.pkl', help="save modeling") 219 | parser.add_argument('--dataset_root', type=str, default='./data/mvtec_anomaly_detection', help="dataset root") 220 | parser.add_argument('--experiment_dir', type=str, default='./experiment/experiment_14', help="dataset root") 221 | parser.add_argument('--classname', type=str, default='capsule', help="dataset class") 222 | parser.add_argument('--img_size', type=int, default=448, help="dataset root") 223 | parser.add_argument("--nAnomaly", type=int, default=10, help="the number of anomaly data in training set") 224 | parser.add_argument("--n_scales", type=int, default=2, help="number of scales at which features are extracted") 225 | parser.add_argument('--backbone', type=str, default='resnet18', help="backbone") 226 | parser.add_argument('--criterion', type=str, default='deviation', help="loss") 227 | parser.add_argument("--topk", type=float, default=0.1, help="topk in MIL") 228 | parser.add_argument('--know_class', type=str, default=None, help="set the know class for hard setting") 229 | parser.add_argument('--pretrain_dir', type=str, default=None, help="root of pretrain weight") 230 | parser.add_argument("--total_heads", type=int, default=4, help="number of head in training") 231 | parser.add_argument("--nRef", type=int, default=5, help="number of reference set") 232 | parser.add_argument('--outlier_root', type=str, default=None, help="OOD dataset root") 233 | args = parser.parse_args() 234 | 235 | args.cuda = not args.no_cuda and torch.cuda.is_available() 236 | trainer = Trainer(args) 237 | 238 | 239 | argsDict = args.__dict__ 240 | if not os.path.exists(args.experiment_dir): 241 | os.makedirs(args.experiment_dir) 242 | with open(args.experiment_dir + '/setting.txt', 'w') as f: 243 | f.writelines('------------------ start ------------------' + '\n') 244 | for eachArg, value in argsDict.items(): 245 | f.writelines(eachArg + ' : ' + str(value) + '\n') 246 | f.writelines('------------------- end -------------------') 247 | 248 | print('Total Epoches:', trainer.args.epochs) 249 | trainer.model = trainer.model.to('cuda') 250 | trainer.criterion = trainer.criterion.to('cuda') 251 | for epoch in range(0, trainer.args.epochs): 252 | trainer.training(epoch) 253 | trainer.eval() 254 | trainer.save_weights(args.savename) 255 | 256 | --------------------------------------------------------------------------------