├── LICENSE ├── README.md ├── data-visualization.ipynb ├── data └── WUSTL-IIoT │ └── README.md ├── models └── WUSTL-IIoT │ └── README.md ├── output └── WUSTL-IIoT │ └── README.md ├── processing.ipynb ├── raw-data └── README.md ├── test-main.py ├── train-main.py └── utils ├── __init__.py ├── __pycache__ ├── __init__.cpython-310.pyc ├── __init__.cpython-39.pyc ├── attention.cpython-310.pyc ├── attention.cpython-39.pyc ├── layers.cpython-310.pyc ├── layers.cpython-39.pyc ├── loading.cpython-310.pyc ├── loading.cpython-39.pyc ├── saving.cpython-310.pyc ├── saving.cpython-39.pyc ├── training.cpython-310.pyc ├── transformer.cpython-310.pyc └── transformer.cpython-39.pyc ├── attention.py ├── layers.py ├── loading.py ├── saving.py └── transformer.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # transformerAD 2 | Code for the paper "Anomaly-Based Intrusion Detection in IIoT Networks Using Transformer Models". 3 | 4 | The utils folder contains the implementation for the transformer class and all of its dependencies, along with some extra functions providing data loading and model saving support. The train-main.py and test-main.py files contain examples of model training and saving and AD execution, with saving of the results. The data-visualization.ipynb Jupyter Notebook includes the code for result evaluation and visualization. 5 | 6 | Models will be saved in a /models/ directory, while its outputs will be kept in an /output/ directory. A /data/ directory must contain the processed datasets in its respective subfolders. 7 | 8 | This repository has been tested on the WUSTL-IIoT-2021 datasets. The code for the processing of both datasets in order for them to be in a format compatible with the transformer model can be found in the /dataset-processing directory. In order to reproduce the experiments presented on the paper: 9 | 10 | 1) Follow the instructions in the README file in /raw-data/ to download the WUSTL-IIoT dataset. Extract the CSV file in this same directory and keep the original name. 11 | 2) Execute the processing.ipynb Jupyter Notebook from the root directory. Note: All the cells must be executed. 12 | 3) Run the train-main.py script from the root directory. Hyperparameters can be tweaked. 13 | 4) Run the test-main.py script from the root directory. Testing batch size can be adjusted to memory requirements. 14 | 5) Use the data-visualization.ipynb Jupyter Notebook from the root directory to analyze the produced results. 15 | 16 | Note that the train-main.py script will always generate a new numerated model subfolder. The 'model_name' variable in the test-main.py script must aim to the desired model. Similarly, running test-main.py will generate a new output subfolder relative to the used model. The 'output' variable in the data-visualization.ipynb Notebook must refer to the desired output as well. 17 | 18 | Different datasets can be used with minimal adjustment. 19 | 20 | -------------------------------------------------------------------------------- /data/WUSTL-IIoT/README.md: -------------------------------------------------------------------------------- 1 | ## Data directory 2 | Must be filled with the output of the processing.ipynb Jupyter Notebook from the root directory. 3 | -------------------------------------------------------------------------------- /models/WUSTL-IIoT/README.md: -------------------------------------------------------------------------------- 1 | ## Models directory 2 | Saved models from train-main.py execution will be stored in this directory. 3 | -------------------------------------------------------------------------------- /output/WUSTL-IIoT/README.md: -------------------------------------------------------------------------------- 1 | ## Output directory 2 | Results from executions of test-main.py will be stored in this directory. 3 | -------------------------------------------------------------------------------- /processing.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": null, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "import pandas as pd\n", 10 | "data_path = 'raw-data/'\n", 11 | "output_path = 'data/WUSTL-IIoT'\n", 12 | "df = pd.read_csv(data_path + 'wustl_iiot_2021.csv', low_memory=False)\n", 13 | "df = df.sort_values(by=['StartTime'])\n", 14 | "mask = list(df['Target'])\n", 15 | "quarter_mask = len(mask)//4" 16 | ] 17 | }, 18 | { 19 | "cell_type": "code", 20 | "execution_count": null, 21 | "metadata": {}, 22 | "outputs": [], 23 | "source": [ 24 | "recon_df = df[:quarter_mask+10000]\n", 25 | "recon_df['Traffic'].value_counts()" 26 | ] 27 | }, 28 | { 29 | "cell_type": "code", 30 | "execution_count": null, 31 | "metadata": {}, 32 | "outputs": [], 33 | "source": [ 34 | "dos_df = df[quarter_mask+10000:2*quarter_mask]\n", 35 | "dos_df['Traffic'].value_counts()" 36 | ] 37 | }, 38 | { 39 | "cell_type": "code", 40 | "execution_count": null, 41 | "metadata": {}, 42 | "outputs": [], 43 | "source": [ 44 | "comm_df = df[2*quarter_mask:3*quarter_mask]\n", 45 | "comm_df['Traffic'].value_counts()" 46 | ] 47 | }, 48 | { 49 | "cell_type": "code", 50 | "execution_count": null, 51 | "metadata": {}, 52 | "outputs": [], 53 | "source": [ 54 | "train_df = df[3*quarter_mask:]\n", 55 | "train_df['Traffic'].value_counts()" 56 | ] 57 | }, 58 | { 59 | "cell_type": "code", 60 | "execution_count": null, 61 | "metadata": {}, 62 | "outputs": [], 63 | "source": [ 64 | "train_df.to_pickle(output_path+'train_WUSTL-IIoT.pkl')\n", 65 | "recon_df.to_pickle(output_path+'recon_WUSTL-IIoT.pkl')\n", 66 | "dos_df.to_pickle(output_path+'dos_WUSTL-IIoT.pkl')\n", 67 | "comm_df.to_pickle(output_path+'comm_WUSTL-IIoT.pkl')" 68 | ] 69 | } 70 | ], 71 | "metadata": { 72 | "kernelspec": { 73 | "display_name": "Python 3", 74 | "language": "python", 75 | "name": "python3" 76 | }, 77 | "language_info": { 78 | "codemirror_mode": { 79 | "name": "ipython", 80 | "version": 3 81 | }, 82 | "file_extension": ".py", 83 | "mimetype": "text/x-python", 84 | "name": "python", 85 | "nbconvert_exporter": "python", 86 | "pygments_lexer": "ipython3", 87 | "version": "3.9.13" 88 | }, 89 | "orig_nbformat": 4, 90 | "vscode": { 91 | "interpreter": { 92 | "hash": "11938c6bc6919ae2720b4d5011047913343b08a43b18698fd82dedb0d4417594" 93 | } 94 | } 95 | }, 96 | "nbformat": 4, 97 | "nbformat_minor": 2 98 | } 99 | -------------------------------------------------------------------------------- /raw-data/README.md: -------------------------------------------------------------------------------- 1 | ## Raw data directory 2 | Raw data from the WUSTL-IIoT dataset must be pasted in this directory for the processing.ipynb Jupyter Notebook to process it and store it in the data directory for later usage by the '-main' programs. 3 | Data can be downloaded from https://www.cse.wustl.edu/~jain/iiot2/index.html, direct download link is https://www.cse.wustl.edu/~jain/iiot2/ftp/wustl_iiot_2021.zip. 4 | -------------------------------------------------------------------------------- /test-main.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import pandas as pd 3 | import utils 4 | from pickle import dump 5 | import os 6 | 7 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 8 | print(f"Using {device} device") 9 | 10 | test_batch_size = 256 11 | 12 | dataset = 'WUSTL-IIoT' 13 | model_name = 'model_1' 14 | model_path = f'models/{dataset}/{model_name}/' 15 | model, data_info = utils.load_model(model_path, device) 16 | assert(data_info['dataset'] == dataset) 17 | 18 | dos_data = pd.read_pickle(f'data/{dataset}/dos_{dataset}.pkl') 19 | test_dos = utils.prepare_data(dos_data, dataset, model_path) 20 | assert(data_info['data_columns'] == list(test_dos.columns)) 21 | 22 | recon_data = pd.read_pickle(f'data/{dataset}/recon_{dataset}.pkl') 23 | test_recon = utils.prepare_data(recon_data, dataset, model_path) 24 | assert(data_info['data_columns'] == list(test_recon.columns)) 25 | 26 | comm_data = pd.read_pickle(f'data/{dataset}/comm_{dataset}.pkl') 27 | test_comm = utils.prepare_data(comm_data, dataset, model_path) 28 | assert(data_info['data_columns'] == list(test_comm.columns)) 29 | 30 | if __name__ == '__main__': 31 | 32 | dos_loader = utils.get_dataLoader(test_dos, data_info['window_size'], device, batch_size = test_batch_size) 33 | dos_scores = model.detect(dos_loader) 34 | recon_loader = utils.get_dataLoader(test_recon, data_info['window_size'], device, batch_size = test_batch_size) 35 | recon_scores = model.detect(recon_loader) 36 | comm_loader = utils.get_dataLoader(test_comm, data_info['window_size'], device, batch_size = test_batch_size) 37 | comm_scores = model.detect(comm_loader) 38 | 39 | output_dir = f"output/{dataset}/output_{model_name}/" 40 | if not os.path.exists(output_dir): 41 | os.makedirs(output_dir) 42 | dump(dos_scores, open(f'{output_dir}dos_scores.pkl', 'wb')) 43 | dump(recon_scores, open(f'{output_dir}recon_scores.pkl', 'wb')) 44 | dump(comm_scores, open(f'{output_dir}comm_scores.pkl', 'wb')) 45 | dump(data_info, open(f'{output_dir}data_info.pkl', 'wb')) -------------------------------------------------------------------------------- /train-main.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import pandas as pd 3 | import utils 4 | 5 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 6 | print(f"Using {device} device") 7 | 8 | dataset = 'WUSTL-IIoT' 9 | normal_data = pd.read_pickle('data/'+dataset+'/train_'+dataset+'.pkl') 10 | train_data, train_scaler = utils.prepare_data(normal_data, dataset) 11 | 12 | d_model = int(len(train_data.columns)) 13 | 14 | attention = 'single' 15 | 16 | N_layers = 6 17 | window_size = 50 18 | batch_size = 1024 19 | epochs = 25 20 | dropout = 0 21 | ff_neurons = 512 22 | 23 | if __name__ == '__main__': 24 | 25 | train_loader = utils.get_dataLoader(train_data, window_size, device, batch_size = batch_size) 26 | 27 | model = utils.Transformer(d_model, N_layers, attention, window_size, device, dropout, ff_neurons) 28 | model.initialize() 29 | model.train_model(train_loader, epochs=epochs, print_every=1) 30 | 31 | if isinstance(attention, int): 32 | attention = f"multi-head {attention}" 33 | 34 | info_dict = { 35 | 'data_columns' : list(train_data.columns), 36 | 'attention' : attention, 37 | 'N_layers' : N_layers, 38 | 'window_size' : window_size, 39 | 'batch_size' : batch_size, 40 | 'model_info' : str(model), 41 | 'epochs' : epochs, 42 | 'dataset' : dataset, 43 | 'dropout' : dropout, 44 | 'ff_neurons' : ff_neurons 45 | } 46 | 47 | utils.save_model(model, info_dict, train_scaler) -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .loading import * 2 | from .transformer import * 3 | from .saving import * -------------------------------------------------------------------------------- /utils/__pycache__/__init__.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/__init__.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/__init__.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/__init__.cpython-39.pyc -------------------------------------------------------------------------------- /utils/__pycache__/attention.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/attention.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/attention.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/attention.cpython-39.pyc -------------------------------------------------------------------------------- /utils/__pycache__/layers.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/layers.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/layers.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/layers.cpython-39.pyc -------------------------------------------------------------------------------- /utils/__pycache__/loading.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/loading.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/loading.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/loading.cpython-39.pyc -------------------------------------------------------------------------------- /utils/__pycache__/saving.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/saving.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/saving.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/saving.cpython-39.pyc -------------------------------------------------------------------------------- /utils/__pycache__/training.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/training.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/transformer.cpython-310.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/transformer.cpython-310.pyc -------------------------------------------------------------------------------- /utils/__pycache__/transformer.cpython-39.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jogecodes/transformerAD/1fdb0d43feac99f7b9823695ae1b76aeb8457f7b/utils/__pycache__/transformer.cpython-39.pyc -------------------------------------------------------------------------------- /utils/attention.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import math 3 | from torch import nn 4 | 5 | # Clase para aplicar codificación posicional sobre un dato dado 6 | class PositionalEncoder(nn.Module): 7 | def __init__(self, d_model, window, device, encoding_length = 10000): 8 | super().__init__() 9 | # Dimensión de los datos de entrada, es decir, el número de atributos que lo describen 10 | self.d_model = d_model 11 | # Tamaño de la ventana deslizante bajo la que se definen las secuencias 12 | self.window = window 13 | # Creación de la matriz de codificación posicional, pe, cuyos elementos dependen de la posición y la dimensión 14 | pe = torch.zeros(window, d_model).to(torch.device(device)) 15 | for pos in range(window): 16 | # La codificación posicional se aplica a cada dimensión de los datos de entrada 17 | # Se aplica una codificación seno y coseno, con una frecuencia que depende de la dimensión 18 | # El resultado es un vector de dimensión d_model (la misma dimensión que los datos de entrada) 19 | # que es determinista para cada posición de la secuencia 20 | for i in range(0, d_model, 2): 21 | pe[pos, i] = math.sin(pos / (encoding_length ** ((2 * i)/d_model))) 22 | if i + 1 < d_model: 23 | pe[pos, i + 1] = math.cos(pos / (encoding_length ** ((2 * (i + 1))/d_model))) 24 | # Se formatea como tensor para poder sumarse a los datos de entrada 25 | pe = pe.unsqueeze(0) 26 | self.pe = pe 27 | 28 | def forward(self, x): 29 | # Se suma la codificación posicional a los datos de entrada 30 | # Sólo se toman las posiciones existentes en la secuencia de entrada 31 | x = x + self.pe[:, :x.size(1), :] 32 | return x 33 | 34 | def scaled_DPattention(q, k, v, d_attention, mask=None, dropout=None): 35 | # Scaled dot-product attention, definida en "Attention is all you need" (Vaswani et al., 2017) 36 | # Al trasponer k, su producto con q proporciona una matriz cuadrada de orden igual al número de 37 | # elementos de la secuencia para cada input del batch 38 | scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(d_attention) 39 | # Aplica una máscara, si es necesario, haciendo infinitamente baja la atención en los elementos 40 | # que no se deben considerar 41 | if mask is not None: 42 | # En el código original se hacía una operación de unsqueezing, pero la he quitado 43 | # para que me cuadren las dimensiones. 44 | # mask = mask.unsqueeze(1) 45 | scores = scores.masked_fill(mask == 0, -1e9) 46 | # Normalización de los scores de acuerdo a una capa softmax 47 | scores = nn.functional.softmax(scores, dim=-1) 48 | # Aplica una capa de dropout, si es necesario 49 | if dropout is not None: 50 | scores = dropout(scores) 51 | # Multiplica los scores por los valores para tasarlos por la atención 52 | output = torch.matmul(scores, v) 53 | return output 54 | 55 | class GeneralAttention(nn.Module): 56 | def __init__(self, d_model, device, dropout = 0.1): 57 | # La atención se inicializa como un modelo de NN 58 | super().__init__() 59 | # Dimensión de los datos de entrada 60 | self.d_model = d_model 61 | # Definición de la capa de dropout 62 | self.dropout = nn.Dropout(dropout).to(torch.device(device)) 63 | 64 | def forward(self, q, k, v, mask=None): 65 | # Se calcula la atención usando la función de atención producto interno reescalada 66 | scores = scaled_DPattention(q, k, v, self.d_model, mask, self.dropout) 67 | # Se aplica la capa lineal de salida y se devuelve la atención calculada 68 | output = scores 69 | return output 70 | 71 | class SingleHeadAttention(nn.Module): 72 | def __init__(self, d_model, device, dropout = 0.1): 73 | # La atención se inicializa como un modelo de NN 74 | super().__init__() 75 | # Dimensión de los datos de entrada 76 | self.d_model = d_model 77 | # Definición de las capas lineales para la atención, que aplican a 78 | # los datos de entrada una transformación lineal del tipo y = Ax + b 79 | self.q_linear = nn.Linear(d_model, d_model) 80 | self.v_linear = nn.Linear(d_model, d_model) 81 | self.k_linear = nn.Linear(d_model, d_model) 82 | # Definición de la capa de dropout 83 | self.dropout = nn.Dropout(dropout).to(torch.device(device)) 84 | # Definición de la capa lineal de salida 85 | self.out = nn.Linear(d_model, d_model).to(torch.device(device)) 86 | 87 | def forward(self, q, k, v, mask=None): 88 | # Se calcula la atención usando la función de atención producto interno reescalada 89 | q_linear = self.q_linear(q) 90 | k_linear = self.k_linear(k) 91 | v_linear = self.v_linear(v) 92 | scores = scaled_DPattention(q_linear, k_linear, v_linear, self.d_model, mask, self.dropout) 93 | # Se aplica la capa lineal de salida y se devuelve la atención calculada 94 | output = self.out(scores) 95 | return output 96 | 97 | # TODO: temrinar de programar, que por ahora no funciona 98 | class MultiHeadAttention(nn.Module): 99 | def __init__(self, heads, d_model, device, dropout = 0.1): 100 | # La atención multicabeza se inicializa como un modelo de NN 101 | super().__init__() 102 | # Dimensión de los datos de entrada 103 | self.d_model = d_model 104 | # Número de cabezas paralelas para la atención 105 | self.N_heads = heads 106 | # Dimensión de la vista por cada cabeza, calculada como la mínima necesaria 107 | # para reconstruir sin pérdida la dimensionalidad de los datos de entrada 108 | self.d_head = (d_model // heads)+1 109 | # Definición de las capas lineales para la atención, que aplican a 110 | # los datos de entrada una transformación lineal del tipo y = Ax + b 111 | # Pasan de la dimensión original de los datos de entrada a la dimensión 112 | # total de las vistas para cada cabeza 113 | self.q_linear = nn.Linear(d_model, self.N_heads * self.d_head) 114 | self.v_linear = nn.Linear(d_model, self.N_heads * self.d_head) 115 | self.k_linear = nn.Linear(d_model, self.N_heads * self.d_head) 116 | # Definición de la capa de dropout 117 | self.dropout = nn.Dropout(dropout).to(torch.device(device)) 118 | # Definición de la capa lineal de salida 119 | self.out = nn.Linear(self.N_heads * self.d_head, d_model).to(torch.device(device)) 120 | 121 | def forward(self, q, k, v, mask=None): 122 | # Tamaño del batch utilizado 123 | batch_size = q.size(0) 124 | q_len = q.size(1) 125 | k_len = k.size(1) 126 | v_len = v.size(1) 127 | 128 | # Se transforman los datos y se les aplica una vista, que reduce 129 | # la dimensionalidad para evaluarlos en cada cabeza 130 | # Llegados a este punto ya no tienen que ser la misma cosa porque 131 | # cada uno ha pasado por un transformador lineal distinto 132 | q_lin = self.q_linear(q) 133 | k_lin = self.k_linear(k) 134 | v_lin = self.v_linear(v) 135 | 136 | q_view = q_lin.view(batch_size, q_len, self.N_heads, self.d_head) 137 | k_view = k_lin.view(batch_size, k_len, self.N_heads, self.d_head) 138 | v_view = v_lin.view(batch_size, v_len, self.N_heads, self.d_head) 139 | 140 | # Se transponen q, k y v para que tengan dimensión 141 | # batch_size * N_heads * seq_len * d_head 142 | q_heads = q_view.transpose(1,2) 143 | k_heads = k_view.transpose(1,2) 144 | v_heads = v_view.transpose(1,2) 145 | 146 | # calculate attention using attention function 147 | scores = scaled_DPattention(q_heads, k_heads, v_heads, self.d_head, mask, self.dropout) 148 | 149 | # concatenate heads and put through final linear layer 150 | concat = scores.transpose(1,2).contiguous().view(batch_size, -1, self.N_heads * self.d_head) 151 | 152 | output = self.out(concat) 153 | 154 | return output -------------------------------------------------------------------------------- /utils/layers.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import copy 3 | from torch.autograd import Variable 4 | from torch import nn 5 | from .attention import SingleHeadAttention, MultiHeadAttention, GeneralAttention 6 | 7 | # Esta es el último módulo por el que pasan los datos en el decoder y encoder 8 | # Cnstituye una red profunda normal, con un mogollón de arcos de conexión 9 | class FeedForward(nn.Module): 10 | # La dimensión de las capas lineales del módulo ff (d_ff) es un hiperparámetro 11 | # que se puede ajustar, si bien se ha fijado a 512 neuronas por defecto 12 | def __init__(self, d_model, device, dropout, d_ff): 13 | super().__init__() 14 | # Se definen las capas lineales y de dropout 15 | self.linear_1 = nn.Linear(d_model, d_ff).to(torch.device(device)) 16 | self.dropout = nn.Dropout(dropout).to(torch.device(device)) 17 | self.linear_2 = nn.Linear(d_ff, d_model).to(torch.device(device)) 18 | self.device = device 19 | def forward(self, x): 20 | # Las capas se conectan mediante una función de activación ReLU 21 | x = self.dropout(nn.functional.relu(self.linear_1(x))) 22 | x = self.linear_2(x) 23 | return x 24 | 25 | # Capa de normalización de batch 26 | class Norm(nn.Module): 27 | def __init__(self, d_model, device, eps = 1e-6): 28 | super().__init__() 29 | self.size = d_model 30 | # Se definen los parámetros de normalización, alpha y bias, como variables de aprendizaje 31 | self.alpha = nn.Parameter(torch.ones(self.size).to(torch.device(device))) 32 | self.bias = nn.Parameter(torch.zeros(self.size).to(torch.device(device))) 33 | self.eps = eps 34 | def forward(self, x): 35 | # El input x se normaliza con respecto a las dimensiones 0 y 1, que corresponden al batch y a la secuencia, 36 | # de forma que para cada atributo quede calculada su media en la muestra dada por la secuencia 37 | x_mean = x.mean(dim=(0,1), keepdim=True) 38 | x_std = x.std(dim=(0,1), keepdim=True) 39 | norm = self.alpha * (x - x_mean) / (x_std + self.eps) + self.bias 40 | return norm 41 | 42 | # Construcción de un módulo (capa) encoder, que consta de una subcapa de atención y 43 | # una subcapa de feed-forward, también aplica normalización por batch entre cada subcapa 44 | class EncoderLayer(nn.Module): 45 | def __init__(self, d_model, attention, device, dropout, d_ff): 46 | super().__init__() 47 | self.norm_1 = Norm(d_model, device).to(torch.device(device)) 48 | self.norm_2 = Norm(d_model, device).to(torch.device(device)) 49 | 50 | if attention == 'general': 51 | self.attn = GeneralAttention(d_model, device).to(torch.device(device)) 52 | elif attention == 'single': 53 | self.attn = SingleHeadAttention(d_model, device).to(torch.device(device)) 54 | elif isinstance(attention, int): 55 | heads = attention 56 | self.attn = MultiHeadAttention(heads, d_model, device).to(torch.device(device)) 57 | else: 58 | raise ValueError('Attention type not recognized') 59 | 60 | self.ff = FeedForward(d_model, device, dropout, d_ff).to(torch.device(device)) 61 | self.dropout_1 = nn.Dropout(dropout).to(torch.device(device)) 62 | self.dropout_2 = nn.Dropout(dropout).to(torch.device(device)) 63 | 64 | def forward(self, x): 65 | # Se normalizan los inputs 66 | x_norm_1 = self.norm_1(x) 67 | # Se calcula la atención sobre la entrada normalizada (que es q, k y v) y se aplica dropout 68 | # Nótese que la entrada x se suma a la salida de la capa de atención, formando una conexión residual 69 | x_dropout_1 = x + self.dropout_1(self.attn(x_norm_1, x_norm_1, x_norm_1)) 70 | # Se renormalizan los inputs 71 | x_norm_2 = self.norm_2(x_dropout_1) 72 | # Se propaga sobre la entrada renormalizada y se aplica dropout 73 | # De nuevo, la entrada x se suma a la salida de la capa de atención 74 | x_dropout_2 = x_dropout_1 + self.dropout_2(self.ff(x_norm_2)) 75 | return x_dropout_2 76 | 77 | # Construcción de un módulo (capa) encoder, que consta de dos subcapas de atención y 78 | # una subcapa de feed-forward, también aplica normalización por batch entre cada subcapa 79 | class DecoderLayer(nn.Module): 80 | def __init__(self, d_model, attention, device, dropout, d_ff): 81 | super().__init__() 82 | self.norm_1 = Norm(d_model, device).to(torch.device(device)) 83 | self.norm_2 = Norm(d_model, device).to(torch.device(device)) 84 | self.norm_3 = Norm(d_model, device).to(torch.device(device)) 85 | 86 | self.dropout_1 = nn.Dropout(dropout).to(torch.device(device)) 87 | self.dropout_2 = nn.Dropout(dropout).to(torch.device(device)) 88 | self.dropout_3 = nn.Dropout(dropout).to(torch.device(device)) 89 | 90 | if attention == 'general': 91 | self.attn_1 = GeneralAttention(d_model, device).to(torch.device(device)) 92 | self.attn_2 = GeneralAttention(d_model, device).to(torch.device(device)) 93 | elif attention == 'single': 94 | self.attn_1 = SingleHeadAttention(d_model, device).to(torch.device(device)) 95 | self.attn_2 = SingleHeadAttention(d_model, device).to(torch.device(device)) 96 | elif isinstance(attention, int): 97 | heads = attention 98 | self.attn_1 = MultiHeadAttention(heads, d_model, device).to(torch.device(device)) 99 | self.attn_2 = MultiHeadAttention(heads, d_model, device).to(torch.device(device)) 100 | else: 101 | raise ValueError('Attention type not recognized') 102 | 103 | self.ff = FeedForward(d_model, device, dropout, d_ff).to(torch.device(device)) 104 | 105 | def forward(self, x, e_outputs, trg_mask): 106 | x2 = self.norm_1(x) 107 | # TODO: probar sin máscara 108 | # x = x + self.dropout_1(self.attn_1(x2, x2, x2, trg_mask)) 109 | x = x + self.dropout_1(self.attn_1(x2, x2, x2)) 110 | x2 = self.norm_2(x) 111 | # De acuerdo con "Attention is all you need", el input del decoder entra a esta atención 112 | # como q, y la salida del encoder como los valores v y las claves k. 113 | x = x + self.dropout_2(self.attn_2(x2, e_outputs, e_outputs)) 114 | x2 = self.norm_3(x) 115 | x = x + self.dropout_3(self.ff(x2)) 116 | return x 117 | 118 | # Función que genera múltiples copias del módulo que se indique 119 | def get_clones(module, N_layers): 120 | return nn.ModuleList([copy.deepcopy(module) for i in range(N_layers)]) -------------------------------------------------------------------------------- /utils/loading.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | import pandas as pd 4 | import os 5 | import pickle 6 | from sklearn.preprocessing import StandardScaler 7 | from .transformer import Transformer 8 | 9 | # Estructura de datos de pyTorch, es necesario para poder iterar sobre los datos 10 | class seq_Loader(torch.utils.data.Dataset): 11 | # Define la inicialización a partir de la del padre 12 | def __init__(self, init_data, window, device): 13 | super(seq_Loader, self).__init__() 14 | # Inicializa la clase padre 15 | self.dataset = init_data 16 | # Tamaño de la ventana deslizante bajo la que se definen las secuencias 17 | self.window = window 18 | # Dispositivo en el que se guarda el tensor 19 | self.device = device 20 | 21 | def __len__(self): 22 | # Número de secuencias que se pueden generar, teniendo en cuenta el tamaño de la ventana 23 | return (len(self.dataset)-(self.window-1)) 24 | 25 | def __getitem__(self, idx): 26 | # Obtiene una secuencia de datos desde idx y con tamaño window 27 | row = self.dataset.iloc[idx:idx+self.window] 28 | # Convierte los datos a un tensor de pyTorch 29 | data = torch.from_numpy(np.array(row)).float().to(torch.device(self.device)) 30 | return data 31 | 32 | def get_dataLoader(data, window_size, device, batch_size = 1): 33 | train_set = seq_Loader(data, window_size, device) 34 | train_loader = torch.utils.data.DataLoader( 35 | train_set, 36 | batch_size=batch_size, 37 | shuffle=False, 38 | # num_workers=os.cpu_count(), # Uses all the CPU cores, but doesn't work on Jupyter running on Windows 39 | num_workers=0, 40 | drop_last=False # Ignores the last batch when it is not complete 41 | ) 42 | return train_loader 43 | 44 | def prepare_data(data, dataset, scaling = 'gaussian', columns = 'normal'): 45 | if dataset == 'UNSW-NB15': 46 | pred_columns = ['dur', 'sbytes', 47 | 'dbytes', 'sttl', 'dttl', 'sloss', 'dloss', 'sload', 'dload', 48 | 'spkts', 'dpkts', 'swin', 'dwin', 'stcpb', 'dtcpb', 'smeansz', 49 | 'dmeansz', 'sjit', 'djit', 'stime', 50 | 'ltime', 'sintpkt', 'dintpkt', 'tcprtt', 'synack', 'ackdat'] 51 | extra_columns = ['is_sm_ips_ports', 'ct_state_ttl', 'ct_flw_http_mthd', 'is_ftp_login', 52 | 'ct_ftp_cmd', 'ct_srv_src', 'ct_srv_dst', 'ct_dst_ltm', 'ct_src_ ltm', 53 | 'ct_src_dport_ltm', 'ct_dst_sport_ltm', 'ct_dst_src_ltm', 'trans_depth', 'res_bdy_len'] 54 | min_columns = ['dur', 'sbytes', 'dbytes', 'sload', 'dload'] 55 | if columns == 'normal': 56 | columned_data = data[pred_columns] 57 | elif columns == 'minimal': 58 | columned_data = data[min_columns] 59 | elif columns == 'extended': 60 | columned_data = data[pred_columns+extra_columns] 61 | else: 62 | columned_data = data 63 | elif dataset == 'WUSTL-IIoT': 64 | pred_columns = ['Mean', 'Sport', 'Dport', 65 | 'SrcPkts', 'DstPkts', 'TotPkts', 'DstBytes', 'SrcBytes', 'TotBytes', 66 | 'SrcLoad', 'DstLoad', 'Load', 'SrcRate', 'DstRate', 'Rate', 'SrcLoss', 67 | 'DstLoss', 'Loss', 'pLoss', 'SrcJitter', 'DstJitter', 'SIntPkt', 68 | 'DIntPkt', 'Proto', 'Dur', 'TcpRtt', 'IdleTime', 'Sum', 'Min', 'Max', 69 | 'sDSb', 'sTtl', 'dTtl', 'sIpId', 'dIpId', 'SAppBytes', 'DAppBytes', 70 | 'TotAppByte', 'SynAck', 'RunTime', 'sTos', 'SrcJitAct', 'DstJitAct'] 71 | columned_data = data[pred_columns] 72 | else: 73 | raise ValueError('Dataset not implemented') 74 | if scaling == 'gaussian': 75 | scaler = StandardScaler() 76 | scaled_data = pd.DataFrame(scaler.fit_transform(columned_data.values), columns=columned_data.columns, index=columned_data.index) 77 | return scaled_data, scaler 78 | elif os.path.isfile(scaling+'scaler.pkl'): 79 | scaler = pickle.load(open(scaling+'scaler.pkl', 'rb')) 80 | scaled_data = pd.DataFrame(scaler.transform(columned_data.values), columns=columned_data.columns, index=columned_data.index) 81 | return scaled_data 82 | else: 83 | raise ValueError('Scaling method not implemented') 84 | 85 | def load_model(model_path, device, return_info = True): 86 | model_info = pickle.load(open(model_path+'model_info.pkl', 'rb')) 87 | data_columns = model_info['data_columns'] 88 | d_model = len(data_columns) 89 | attention = model_info['attention'] 90 | if attention[:10] == 'multi-head': 91 | attention = int(attention.split(' ')[1]) 92 | model = Transformer(d_model, model_info['N_layers'], attention, model_info['window_size'], device, model_info['dropout'], model_info['ff_neurons']) 93 | model.load_state_dict(torch.load(model_path+'model_state.pt', map_location=torch.device(device))) 94 | model_info = { 95 | 'data_columns': data_columns, 96 | 'window_size': model_info['window_size'], 97 | 'batch_size': model_info['batch_size'], 98 | 'dataset': model_info['dataset'], 99 | 'dropout' : model_info['dropout'], 100 | 'ff_neurons' : model_info['ff_neurons'] 101 | } 102 | if return_info: 103 | return model, model_info 104 | else: 105 | return model -------------------------------------------------------------------------------- /utils/saving.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import os 3 | import json 4 | from pickle import dump 5 | 6 | def save_model(model, info_dict, scaler, path = 'models/', loss = None): 7 | 8 | models_path = path 9 | dataset = info_dict['dataset'] 10 | models_path = f'{models_path}{dataset}/' 11 | if not os.path.exists(models_path): 12 | os.makedirs(models_path) 13 | 14 | arr = os.listdir(models_path) 15 | if len(arr) != 0: 16 | folder_nums = [] 17 | for folder_name in arr: 18 | if folder_name.split("_")[0] == "model": 19 | folder_nums.append(int(folder_name.split("_")[1])) 20 | if len(folder_nums) == 0: 21 | new_folder_num = 1 22 | else: 23 | try: 24 | new_folder_num = max(folder_nums) + 1 25 | except: 26 | new_folder_num = 666 27 | else: 28 | new_folder_num = 1 29 | model_dir = f"{models_path}/model_{new_folder_num}" 30 | os.makedirs(model_dir) 31 | torch.save(model.state_dict(), model_dir+'/model_state.pt') 32 | 33 | dump(info_dict, open(f'{model_dir}/model_info.pkl', 'wb')) 34 | dump(scaler, open(f'{model_dir}/scaler.pkl', 'wb')) 35 | if loss is not None: 36 | dump(loss, open(f'{model_dir}/loss.pkl', 'wb')) 37 | 38 | return True -------------------------------------------------------------------------------- /utils/transformer.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import nn 3 | from torch.autograd import Variable 4 | import numpy as np 5 | from .layers import Norm, EncoderLayer, DecoderLayer, get_clones 6 | from .attention import PositionalEncoder 7 | from tqdm import tqdm 8 | 9 | class Encoder(nn.Module): 10 | def __init__(self, d_model, N_layers, attention, window, device, dropout, d_ff): 11 | super().__init__() 12 | self.N_layers = N_layers 13 | self.pe = PositionalEncoder(d_model, window, device) 14 | self.layers = get_clones(EncoderLayer(d_model, attention, device, dropout, d_ff), N_layers) 15 | self.norm = Norm(d_model, device) 16 | def forward(self, src): 17 | # En la variable x se almacena src, el batch de datos de entrada en cada iteración, 18 | # pero con el positional encoding aplicado mediante una suma 19 | x = self.pe(src) 20 | # Seguidamente se pasa x por las N capas del encoder (que son todas idénticas) 21 | for i in range(self.N_layers): 22 | x = self.layers[i](x) 23 | return self.norm(x) 24 | 25 | class Decoder(nn.Module): 26 | def __init__(self, d_model, N_layers, attention, window, device, dropout, d_ff): 27 | super().__init__() 28 | self.N_layers = N_layers 29 | self.pe = PositionalEncoder(d_model, window, device) 30 | self.layers = get_clones(DecoderLayer(d_model, attention, device, dropout, d_ff), N_layers) 31 | self.norm = Norm(d_model, device) 32 | def forward(self, trg, e_outputs, mask): 33 | # x = self.embed(trg) 34 | x = self.pe(trg) 35 | for i in range(self.N_layers): 36 | x = self.layers[i](x, e_outputs, mask) 37 | return self.norm(x) 38 | 39 | class Transformer(nn.Module): 40 | def __init__(self, d_model, N_layers, attention, window, device, dropout=0.1, d_ff = 512): 41 | super().__init__() 42 | self.encoder = Encoder(d_model, N_layers, attention, window, device, dropout, d_ff) 43 | self.decoder = Decoder(d_model, N_layers, attention, window, device, dropout, d_ff) 44 | self.out = nn.Linear(d_model*(window-1), d_model).to(torch.device(device)) 45 | self.window = window 46 | self.device = device 47 | def nopeak_mask(self, size): 48 | np_mask = np.triu(np.ones((1, size, size)), k=1).astype('uint8') 49 | np_mask = Variable(torch.from_numpy(np_mask) == 0).to(torch.device(self.device)) 50 | return np_mask 51 | def initialize(self): 52 | for p in self.parameters(): 53 | if p.dim() > 1: 54 | nn.init.xavier_uniform_(p) 55 | # this code is very important! It initialises the parameters with a 56 | # range of values that stops the signal fading or getting too big. 57 | # See https://andyljones.tumblr.com/post/110998971763/an-explanation-of-xavier-initialization 58 | # for a mathematical explanation. 59 | def forward(self, src, trg): 60 | e_outputs = self.encoder(src) 61 | d_output = self.decoder(trg, e_outputs, self.nopeak_mask(trg.size(1))) 62 | output = self.out(d_output.view(d_output.size(0), -1)) 63 | # No se hace softmax a la salida porque no se buscan probabilidades, sino los valores predichos 64 | return output 65 | 66 | def train_model(self, data_loader, epochs=10, print_every=1, return_evo=False): 67 | optim = torch.optim.Adam(self.parameters(), lr=0.0001, betas=(0.9, 0.98), eps=1e-9) 68 | self.train() 69 | total_loss = 0 70 | loss_evo = [] 71 | for epoch in range(epochs): 72 | print(f'\nEpoch: {epoch+1} of {epochs}') 73 | with tqdm(total = data_loader.__len__()) as pbar: 74 | for i, batch in enumerate(data_loader): 75 | pbar.update(1) 76 | # Al decoder se le pasa como entrada el batch de datos de entrada menos el último dato, 77 | # que es el que se quiere predecir 78 | d_input = batch[:, :-1] 79 | trg_output = batch[:, -1] 80 | preds = self.forward(batch, d_input) 81 | optim.zero_grad() 82 | criterion = nn.MSELoss() 83 | loss = criterion(preds, trg_output) 84 | loss.backward() 85 | optim.step() 86 | 87 | total_loss += loss.item() 88 | loss_evo.append(loss.item()) 89 | 90 | if (i + 1) % print_every == 0: 91 | loss_avg = total_loss / print_every 92 | pbar.set_postfix({'Last mean loss': loss_avg}) 93 | total_loss = 0 94 | if return_evo: 95 | return loss_evo 96 | 97 | def detect(self, data_loader): 98 | self.eval() 99 | ano_scores = [] 100 | criterion = nn.MSELoss(reduction = 'none') 101 | with tqdm(total = data_loader.__len__()) as pbar: 102 | for batch in data_loader: 103 | pbar.update(1) 104 | d_input = batch[:, :-1] 105 | trg_output = batch[:, -1] 106 | preds = self.forward(batch, d_input) 107 | loss = criterion(preds, trg_output) 108 | ano_scores = ano_scores + loss.mean(dim=-1).tolist() 109 | return ano_scores --------------------------------------------------------------------------------