├── LICENSE ├── README.md └── code ├── main.py ├── mean_final.npy ├── requirements.txt ├── std_final.npy ├── subspectralnet.py ├── tensorboard-logs └── events.out.tfevents.1550619217.xgpc3 └── utils.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 | # SubSpectralNet-PyTorch 2 | 3 | This repository contains the PyTorch Implementation of SubSpectralNets introduced in the following paper: 4 | 5 | [SubSpectralNet - Using Sub-Spectrogram based Convolutional Neural Networks for Acoustic Scene Classification](https://arxiv.org/abs/1810.12642) (Accepted in ICASSP 2019) 6 | 7 | [Sai Samarth R Phaye](https://ssrp.github.io), [Emmanouil Benetos](http://www.eecs.qmul.ac.uk/~emmanouilb/), and [Ye Wang](https://www.smcnus.org/profile/ye-wang/). 8 | 9 | [Click here for the presentation!](https://docs.google.com/presentation/d/1xyvpgGPkdrxgbBbEWvup5sPiajiWRdbQ7CZGd9nW0jY/) 10 | 11 | We introduce a novel approach of using spectrograms in Convolutional Neural Networks in the context of acoustic scene classification. First, we show from the statistical analysis that some specific bands of mel-spectrograms carry discriminative information than other bands, which is specific to every soundscape. From the inferences taken by this, we propose SubSpectralNets in which we first design a new convolutional layer that splits the time-frequency features into sub-spectrograms, then merges the band-level features on a later stage for the global classification. The effectiveness of SubSpectralNet is demonstrated by a relative improvement of +14% accuracy over the DCASE 2018 baseline model. The detailed architecture of SubSpectralNet is shown below. 12 | 13 |

14 | 15 |

16 | 17 | If you have any queries regarding the code, please contact us on the following email: phaye.samarth@gmail.com (Sai Samarth R Phaye) 18 | 19 | ## Usage 20 | 21 | Following are the steps to follow for using this implementation: 22 | 23 | Before anything, it is expected that you download and extract the [DCASE 2018 ASC Development Dataset](https://zenodo.org/record/1228142). You should obtain a folder named "TUT-urban-acoustic-scenes-2018-development", which contains various subfolders like "audio", "evaluation_setup" etc. Once you get this folder, following are the steps to execute the code: 24 | 25 | **Step 1. Clone the repository to local** 26 | ``` 27 | git clone https://github.com/ssrp/SubSpectralNet-PyTorch.git SubSpectralNets 28 | cd SubSpectralNets/code 29 | ``` 30 | 31 | **Step 2. Install the prerequisites** 32 | ``` 33 | pip install -r requirements.txt 34 | ``` 35 | 36 | **Step 3. Train a SubSpectralNet** 37 | 38 | Train with default settings: 39 | ``` 40 | python main.py --root-dir /TUT-urban-acoustic-scenes-2018-development/ 41 | ``` 42 | For more settings, the code is well-commented and it's easy to change the parameters looking at the comments. 43 | 44 | ## Results 45 | 46 | You should be able to see the losses for every iteration on the console. Additionally, you could also visualize the losses on TensorBoard by creating a tensorboard session using the following command (run on a new console, in the same directory): 47 | ``` 48 | tensorboard --logdir=. 49 | ``` 50 | For your convenience, there is one log file in the code/tensorboard-logs directory. 51 | 52 | The statistical results of the network are shared in Section 4 of the paper. Following is the accuracy curve obtained after training the model on 40 mel-bin magnitude spectrograms, 20 sub-spectrogram size and 10 mel-bin hop-size (72.18%, average-best accuracy in three runs). 53 | 54 |

55 | 56 |

57 | -------------------------------------------------------------------------------- /code/main.py: -------------------------------------------------------------------------------- 1 | """ 2 | This script runs the SubSpectralNet training in PyTorch. 3 | 4 | Updated February 2019 5 | Sai Samarth R Phaye 6 | """ 7 | 8 | from __future__ import print_function, division 9 | 10 | # Basics 11 | import argparse 12 | import os 13 | import numpy as np 14 | 15 | from subspectralnet import * 16 | 17 | from utils import * 18 | 19 | # Ignore warnings 20 | import warnings 21 | warnings.filterwarnings("ignore") 22 | 23 | # Ignite Framework 24 | import torch 25 | from ignite.engine import Engine, Events, create_supervised_trainer, create_supervised_evaluator 26 | from ignite.metrics import Accuracy, Loss 27 | 28 | def run(train_batch_size, test_batch_size, epochs, lr, log_interval, log_dir, no_cuda, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, seed, root_dir, train_dir, eval_dir): 29 | """ 30 | Model runner 31 | 32 | Parameters 33 | ---------- 34 | train_batch_size : int 35 | Size of the training batch. Default: 16 36 | 37 | test_batch_size : int 38 | size of the testing batch. Default: 16 39 | 40 | epochs : int 41 | Number of training epochs. Default: 200 42 | 43 | lr : float 44 | Learning rate for the ADAM optimizer. Default: 0.001 45 | 46 | log_interval : int 47 | Interval for logging data: Default: 10 48 | 49 | log_dir : str 50 | Directory to save the logs 51 | 52 | no_cuda : Bool 53 | Should you NOT use cuda? Default: False 54 | 55 | sub_spectrogram_size : int 56 | Size of the SubSpectrogram. Default 20 57 | 58 | sub_spectrogram_mel_hop : int 59 | Mel-bin hop size of the SubSpectrogram. Default 10 60 | 61 | n_mel_bins : int 62 | Number of mel-bins of the Spectrogram extracted. Default: 40. 63 | 64 | seed : int 65 | Torch random seed value, for reproducable results. Default: 1 66 | 67 | root_dir : str 68 | Directory of the folder which contains the dataset (has 'audio' and 'evaluation_setup' folders inside) 69 | 70 | train_dir : str 71 | Set as default: 'evaluation_setup/train_fold1.txt' 72 | 73 | eval_dir : str 74 | Set as default: 'evaluation_setup/evaluate_fold1.txt' 75 | """ 76 | 77 | # check if possible to use CUDA 78 | use_cuda = not no_cuda and torch.cuda.is_available() 79 | 80 | # set seed 81 | torch.manual_seed(seed) 82 | 83 | # Map to GPU 84 | device = torch.device("cuda" if use_cuda else "cpu") 85 | 86 | # Load the data loaders 87 | train_loader, val_loader = get_data_loaders(train_batch_size, test_batch_size, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, use_cuda, root_dir, train_dir, eval_dir) 88 | 89 | # Get the model 90 | model = SubSpectralNet(sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, use_cuda).to(device) 91 | 92 | # Init the TensorBoard summary writer 93 | writer = create_summary_writer(model, train_loader, log_dir) 94 | 95 | # Init the optimizer 96 | optimizer = optim.Adam(model.parameters(), lr=lr) 97 | 98 | # Use GPU if possible 99 | if device: 100 | model.to(device) 101 | 102 | def update_model(engine, batch): 103 | """Prepare batch for training: pass to a device with options. 104 | 105 | """ 106 | model.train() 107 | optimizer.zero_grad() 108 | 109 | inputs, label = prepare_batch(batch, device=device) 110 | output = model(inputs) 111 | losses = [] 112 | for ite in range(output.shape[1]): 113 | losses.append(F.nll_loss(output[:,ite,:], label)) 114 | loss = sum(losses) 115 | loss.backward() 116 | optimizer.step() 117 | return losses, output 118 | 119 | # get the trainer module 120 | trainer = Engine(update_model) 121 | 122 | def evaluate(engine, batch): 123 | """Prepare batch for training: pass to a device with options. 124 | """ 125 | model.eval() 126 | with torch.no_grad(): 127 | inputs, label = prepare_batch(batch, device=device) 128 | output = model(inputs) 129 | losses = [] 130 | correct = [] 131 | for ite in range(output.shape[1]): 132 | losses.append(F.nll_loss(output[:,ite,:], label, reduction='sum').item()) 133 | return losses, output, label 134 | 135 | # get the evaluator module 136 | evaluator = Engine(evaluate) 137 | 138 | # define output transforms for multiple outputs. 139 | def output_transform1(output): 140 | # `output` variable is returned by above `process_function` 141 | losses, correct, label = output 142 | return correct[:,0,:], label 143 | 144 | metric = Accuracy(output_transform=output_transform1) 145 | metric.attach(evaluator, "acc_highband") 146 | metric = Loss(F.nll_loss, output_transform=output_transform1) 147 | metric.attach(evaluator, "loss_highband") 148 | 149 | def output_transform2(output): 150 | # `output` variable is returned by above `process_function` 151 | losses, correct, label = output 152 | return correct[:,1,:], label 153 | 154 | metric = Accuracy(output_transform=output_transform2) 155 | metric.attach(evaluator, "acc_midband") 156 | metric = Loss(F.nll_loss, output_transform=output_transform2) 157 | metric.attach(evaluator, "loss_midband") 158 | 159 | def output_transform3(output): 160 | # `output` variable is returned by above `process_function` 161 | losses, correct, label = output 162 | return correct[:,2,:], label 163 | 164 | metric = Accuracy(output_transform=output_transform3) 165 | metric.attach(evaluator, "acc_lowband") 166 | metric = Loss(F.nll_loss, output_transform=output_transform3) 167 | metric.attach(evaluator, "loss_lowband") 168 | 169 | def output_transform(output): 170 | # `output` variable is returned by above `process_function` 171 | losses, correct, label = output 172 | return correct[:,3,:], label 173 | 174 | metric = Accuracy(output_transform=output_transform) 175 | metric.attach(evaluator, "acc_globalclassifier") 176 | metric = Loss(F.nll_loss, output_transform=output_transform) 177 | metric.attach(evaluator, "loss_globalclassifier") 178 | 179 | # Log the events in Ignite: EVERY ITERATION 180 | @trainer.on(Events.ITERATION_COMPLETED) 181 | def log_training_loss(engine): 182 | iter = (engine.state.iteration - 1) % len(train_loader) + 1 183 | if iter % log_interval == 0: 184 | losses, output = engine.state.output 185 | epoch = engine.state.epoch 186 | print('Train Epoch: {} [{}/{}]\tLosses: {:.6f} (Top Band), {:.6f} (Mid Band), {:.6f} (Low Band), {:.6f} (Global Classifier)'.format( 187 | epoch, iter, len(train_loader), losses[0].item(), losses[1].item(), losses[2].item(), losses[3].item())) 188 | # TensorBoard Logs 189 | writer.add_scalar("training/loss_topband_itr", losses[0].item(), engine.state.iteration) 190 | writer.add_scalar("training/loss_midband_itr", losses[1].item(), engine.state.iteration) 191 | writer.add_scalar("training/loss_lowband_itr", losses[2].item(), engine.state.iteration) 192 | writer.add_scalar("training/loss_global_itr", losses[3].item(), engine.state.iteration) 193 | 194 | 195 | # Log the events in Ignite: Test the training data on EVERY EPOCH 196 | @trainer.on(Events.EPOCH_COMPLETED) 197 | def log_training_results(engine): 198 | evaluator.run(train_loader) 199 | print("Training Results - Epoch: {} Global accuracy: {:.2f} Avg loss: {:.2f}" 200 | .format(engine.state.epoch, evaluator.state.metrics['acc_globalclassifier'], evaluator.state.metrics['loss_globalclassifier'])) 201 | # TensorBoard Logs 202 | writer.add_scalar("training/global_loss", evaluator.state.metrics['loss_globalclassifier'], engine.state.epoch) 203 | writer.add_scalar("training/lowband_loss", evaluator.state.metrics['loss_lowband'], engine.state.epoch) 204 | writer.add_scalar("training/midband_loss", evaluator.state.metrics['loss_midband'], engine.state.epoch) 205 | writer.add_scalar("training/highband_loss", evaluator.state.metrics['loss_highband'], engine.state.epoch) 206 | writer.add_scalar("training/global_acc", evaluator.state.metrics['acc_globalclassifier'], engine.state.epoch) 207 | writer.add_scalar("training/lowband_acc", evaluator.state.metrics['acc_lowband'], engine.state.epoch) 208 | writer.add_scalar("training/midband_acc", evaluator.state.metrics['acc_midband'], engine.state.epoch) 209 | writer.add_scalar("training/highband_acc", evaluator.state.metrics['acc_highband'], engine.state.epoch) 210 | 211 | 212 | # Log the events in Ignite: Test the validation data on EVERY EPOCH 213 | @trainer.on(Events.EPOCH_COMPLETED) 214 | def log_validation_results(engine): 215 | evaluator.run(val_loader) 216 | print("Validation Results - Epoch: {} Global accuracy: {:.2f} Avg loss: {:.2f}" 217 | .format(engine.state.epoch, evaluator.state.metrics['acc_globalclassifier'], evaluator.state.metrics['loss_globalclassifier'])) 218 | # TensorBoard Logs 219 | writer.add_scalar("validation/global_loss", evaluator.state.metrics['loss_globalclassifier'], engine.state.epoch) 220 | writer.add_scalar("validation/lowband_loss", evaluator.state.metrics['loss_lowband'], engine.state.epoch) 221 | writer.add_scalar("validation/midband_loss", evaluator.state.metrics['loss_midband'], engine.state.epoch) 222 | writer.add_scalar("validation/highband_loss", evaluator.state.metrics['loss_highband'], engine.state.epoch) 223 | writer.add_scalar("validation/global_acc", evaluator.state.metrics['acc_globalclassifier'], engine.state.epoch) 224 | writer.add_scalar("validation/lowband_acc", evaluator.state.metrics['acc_lowband'], engine.state.epoch) 225 | writer.add_scalar("validation/midband_acc", evaluator.state.metrics['acc_midband'], engine.state.epoch) 226 | writer.add_scalar("validation/highband_acc", evaluator.state.metrics['acc_highband'], engine.state.epoch) 227 | 228 | # kick everything off 229 | trainer.run(train_loader, max_epochs=epochs) 230 | 231 | # close the writer 232 | writer.close() 233 | 234 | # return the model 235 | return model 236 | 237 | def main(): 238 | """ 239 | Main method. Initializes the parser, default variables and calls the run function. 240 | """ 241 | # Training settings 242 | parser = argparse.ArgumentParser(description='PyTorch code for SubSpectralNets') 243 | 244 | parser.add_argument('--batch-size', type=int, default=16, metavar='BS', 245 | help='input batch size for training (default: 16)') 246 | 247 | parser.add_argument('--test-batch-size', type=int, default=16, metavar='TBS', 248 | help='input batch size for testing (default: 16)') 249 | 250 | parser.add_argument('--epochs', type=int, default=200, metavar='E', 251 | help='number of epochs to train (default: 200)') 252 | 253 | parser.add_argument('--sub-spectrogram-size', type=int, default=20, metavar='SSS', 254 | help='sub-spectrogram size (default: 16)') 255 | 256 | parser.add_argument('--sub-spectrogram-mel-hop', type=int, default=10, metavar='MH', 257 | help='sub-spectrogram mel-hop value (default: 10)') 258 | 259 | parser.add_argument('--time-indices', type=int, default=500, metavar='T', 260 | help='temporal dimension size of the spectrogram') 261 | 262 | parser.add_argument('--channels', type=int, default=1, metavar='C', 263 | help='number of audio channels (default: 1)') 264 | 265 | parser.add_argument('--mel-bins', type=int, default=40, metavar='MB', 266 | help='number of audio channels (default: 40)') 267 | 268 | parser.add_argument('--lr', type=float, default=0.01, metavar='LR', 269 | help='learning rate (default: 0.001)') 270 | 271 | parser.add_argument('--no-cuda', action='store_true', default=False, 272 | help='disables CUDA training') 273 | 274 | parser.add_argument('--seed', type=int, default=1, metavar='S', 275 | help='random seed (default: 1)') 276 | 277 | parser.add_argument('--log-interval', type=int, default=10, metavar='I', 278 | help='how many batches to wait before logging training status') 279 | 280 | parser.add_argument('--save-model', action='store_true', default=False, 281 | help='For Saving the current Model') 282 | 283 | parser.add_argument('--root-dir', type=str, default="../../TUT-urban-acoustic-scenes-2018-development/", metavar='RD', 284 | help='Root directory for the dataset: must contain \'audio\' folder') 285 | 286 | parser.add_argument('--train-dir', type=str, default="evaluation_setup/fold1_train.txt", metavar='TD', 287 | help='Link to train data labels file') 288 | 289 | parser.add_argument('--eval-dir', type=str, default="evaluation_setup/fold1_evaluate.txt", metavar='ED', 290 | help='Link to evaluate data labels file') 291 | 292 | parser.add_argument("--log_dir", type=str, default="tensorboard_logs", 293 | help="log directory for Tensorboard log output") 294 | 295 | args = parser.parse_args() 296 | 297 | # Sub-Spectrogram Size 298 | sub_spectrogram_size = args.sub_spectrogram_size 299 | 300 | # Init Directories 301 | root_dir = args.root_dir 302 | train_dir = args.train_dir 303 | eval_dir = args.eval_dir 304 | 305 | # Mel-bins 306 | n_mel_bins = args.mel_bins 307 | 308 | # Mel-bins sub_spectrogram_mel_hop 309 | sub_spectrogram_mel_hop = args.sub_spectrogram_mel_hop 310 | 311 | # Time Indices 312 | timeInd = args.time_indices 313 | 314 | # Channels used 315 | channels = args.channels 316 | 317 | # get number of classifiers (number of sub-spectrograms + 1 for global classifier) 318 | numClassifiers = 0 319 | while(sub_spectrogram_mel_hop*numClassifiers <= n_mel_bins - sub_spectrogram_size): 320 | numClassifiers = numClassifiers + 1 321 | # + 1 for global classifier 322 | numClassifiers = numClassifiers + 1 323 | 324 | # Run the model 325 | model = run(args.batch_size, args.test_batch_size, args.epochs, args.lr, args.log_interval, args.log_dir, args.no_cuda, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, args.seed, root_dir, train_dir, eval_dir) 326 | 327 | # save the model 328 | if (args.save_model): 329 | torch.save(model.state_dict(),"subspectralnet_cnn.pt") 330 | 331 | if __name__ == '__main__': 332 | """ 333 | Python main function. 334 | Calls another main() function because using this main function is too 'main'stream. 335 | """ 336 | main() 337 | -------------------------------------------------------------------------------- /code/mean_final.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ssrp/SubSpectralNet-PyTorch/ebc9fa752c356eaf798d4e353bf390e3e5dd7060/code/mean_final.npy -------------------------------------------------------------------------------- /code/requirements.txt: -------------------------------------------------------------------------------- 1 | matplotlib==2.1.0 2 | torchvision==0.2.1 3 | dcase_util==0.2.5 4 | torch==1.0.1.post2 5 | pytorch-ignite==0.1.2 6 | numpy==1.15.4 7 | librosa==0.6.2 8 | tensorboardX==1.6 9 | -------------------------------------------------------------------------------- /code/std_final.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ssrp/SubSpectralNet-PyTorch/ebc9fa752c356eaf798d4e353bf390e3e5dd7060/code/std_final.npy -------------------------------------------------------------------------------- /code/subspectralnet.py: -------------------------------------------------------------------------------- 1 | """ 2 | This script contains the heart of the code: the SubSpectralNet architecture. 3 | 4 | Updated February 2019 5 | Sai Samarth R Phaye 6 | """ 7 | 8 | # PyTorch 9 | import torch 10 | import torch.nn.functional as F 11 | import torch.nn as nn 12 | import torch.optim as optim 13 | from torchvision import transforms, utils 14 | 15 | # Math 16 | import math 17 | 18 | class SubSpectralNet(nn.Module): 19 | """ SubSpectralNet architecture """ 20 | def __init__(self, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, use_gpu): 21 | """ 22 | Init the model layers 23 | 24 | Parameters 25 | ---------- 26 | sub_spectrogram_size : int 27 | Size of the SubSpectrogram. Default: 20 28 | 29 | sub_spectrogram_mel_hop : int 30 | Mel-bin hop size of the SubSpectrogram. Default 10 31 | 32 | n_mel_bins : int 33 | Number of mel-bins of the Spectrogram extracted. Default: 40. 34 | 35 | use_gpu : Bool 36 | Use GPU or not. Default: True 37 | """ 38 | super(SubSpectralNet, self).__init__() 39 | self.sub_spectrogram_size, self.sub_spectrogram_mel_hop, self.n_mel_bins, self.use_gpu = sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, use_gpu 40 | 41 | # For max-pool after the second conv-layer 42 | self.n_max_pool = int(self.sub_spectrogram_size / 10) 43 | 44 | # Number of SubSpectrograms: used for defining the number of conv-layers 45 | self.n_sub_spectrograms = 0 46 | 47 | while(self.sub_spectrogram_mel_hop*self.n_sub_spectrograms <= self.n_mel_bins - self.sub_spectrogram_size): 48 | self.n_sub_spectrograms = self.n_sub_spectrograms + 1 49 | 50 | # init the layers 51 | self.conv1 = nn.ModuleList([nn.Conv2d(in_channels=1, out_channels=32, kernel_size=7, stride=1, padding=3) for _ in range(self.n_sub_spectrograms)]) 52 | self.conv1_bn = nn.ModuleList([nn.BatchNorm2d(32) for _ in range(self.n_sub_spectrograms)]) 53 | self.mp1 = nn.ModuleList([nn.MaxPool2d((self.n_max_pool,5)) for _ in range(self.n_sub_spectrograms)]) 54 | self.drop1 = nn.ModuleList([nn.Dropout(0.3) for _ in range(self.n_sub_spectrograms)]) 55 | self.conv2 = nn.ModuleList([nn.Conv2d(in_channels=32, out_channels=64, kernel_size=7, stride=1, padding=3) for _ in range(self.n_sub_spectrograms)]) 56 | self.conv2_bn = nn.ModuleList([nn.BatchNorm2d(64) for _ in range(self.n_sub_spectrograms)]) 57 | self.mp2 = nn.ModuleList([nn.MaxPool2d((4,100)) for _ in range(self.n_sub_spectrograms)]) 58 | self.drop2 = nn.ModuleList([nn.Dropout(0.3) for _ in range(self.n_sub_spectrograms)]) 59 | 60 | self.fc1 = nn.ModuleList([nn.Linear(1*2*64, 32) for _ in range(self.n_sub_spectrograms)]) 61 | self.drop3 = nn.ModuleList([nn.Dropout(0.3) for _ in range(self.n_sub_spectrograms)]) 62 | self.fc2 = nn.ModuleList([nn.Linear(32, 10) for _ in range(self.n_sub_spectrograms)]) 63 | 64 | numFCs = int(math.log(self.n_sub_spectrograms*32, 2)) 65 | neurons = int(math.pow(2, numFCs)) 66 | 67 | self.fcGlobal = [] 68 | tempNeurons =int(32*self.n_sub_spectrograms) 69 | while(neurons >= 64): 70 | self.fcGlobal.append(nn.Linear(tempNeurons, neurons)) 71 | self.fcGlobal.append(nn.ReLU(0.3)) 72 | self.fcGlobal.append(nn.Dropout(0.3)) 73 | tempNeurons = neurons 74 | neurons = int(neurons / 2) 75 | self.fcGlobal.append(nn.Linear(tempNeurons, 10)) 76 | self.fcGlobal = nn.ModuleList(self.fcGlobal) 77 | 78 | def forward(self, x): 79 | """ 80 | Feed-forward pass 81 | 82 | Parameters 83 | ---------- 84 | x : tensor 85 | Input batch. Size: [batch_size, channels, sub_spectrogram_size, n_time_indices, n_sub_spectrograms]. Default [16, 1, 20, 500, 3] 86 | 87 | Returns 88 | ------- 89 | outputs: tensor 90 | final output of the model. Size: [batch_size, n_sub_spectrograms, n_labels]. Default: [16, 4, 10] 91 | """ 92 | outputs = [] 93 | intermediate = [] 94 | x = x.float() 95 | if self.use_gpu: 96 | x = x.cuda() 97 | input_var = x 98 | 99 | # for every sub-spectrogram 100 | for i in range(x.shape[4]): 101 | x = input_var 102 | x = self.conv1[i](x[:, :, :, :, i]) 103 | x = self.conv1_bn[i](x) 104 | x = F.relu(x) 105 | x = self.mp1[i](x) 106 | x = self.drop1[i](x) 107 | x = self.conv2[i](x) 108 | x = self.conv2_bn[i](x) 109 | x = F.relu(x) 110 | x = self.mp2[i](x) 111 | x = self.drop2[i](x) 112 | x = x.view(-1, 1*2*64) 113 | x = self.fc1[i](x) 114 | x = F.relu(x) 115 | intermediate.append(x) 116 | x = self.drop3[i](x) 117 | x = self.fc2[i](x) 118 | x = x.view(-1, 1, 10) 119 | outputs.append(x) 120 | 121 | # extracted intermediate layers 122 | x = torch.cat((intermediate), 1) 123 | 124 | # global classification 125 | for i in range(len(self.fcGlobal)): 126 | x = self.fcGlobal[i](x) 127 | x = x.view(-1, 1, 10) 128 | outputs.append(x) 129 | 130 | # all the outputs (low, mid and high band + global classifier) 131 | outputs = torch.cat((outputs), 1) 132 | outputs = F.log_softmax(outputs, dim=2) 133 | return outputs 134 | -------------------------------------------------------------------------------- /code/tensorboard-logs/events.out.tfevents.1550619217.xgpc3: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ssrp/SubSpectralNet-PyTorch/ebc9fa752c356eaf798d4e353bf390e3e5dd7060/code/tensorboard-logs/events.out.tfevents.1550619217.xgpc3 -------------------------------------------------------------------------------- /code/utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | This script contains the basic building blocks of the DataLoader and Transforms for preprocessing the data in batches. 3 | 4 | Updated February 2019 5 | Sai Samarth R Phaye 6 | """ 7 | 8 | # Basics 9 | import numpy as np 10 | import os 11 | 12 | # Audio Processing 13 | import librosa 14 | import dcase_util 15 | 16 | # PyTorch 17 | import torch 18 | from torchvision import transforms, utils 19 | from torch.utils.data import Dataset, DataLoader 20 | from ignite._utils import convert_tensor 21 | 22 | # Tensorboard 23 | try: 24 | from tensorboardX import SummaryWriter 25 | except ImportError: 26 | raise RuntimeError("No tensorboardX package is found. Please install with the command: \npip install tensorboardX") 27 | 28 | class ToTensor(object): 29 | """ Convert ndarrays in sample to Tensors.""" 30 | 31 | def __call__(self, sample): 32 | data, label = sample['data'], sample['label'] 33 | 34 | # swap color axis (not required) 35 | data = data.transpose((0, 1, 2, 3)) 36 | 37 | return {'data': torch.from_numpy(data), 38 | 'label': torch.from_numpy(label)} 39 | 40 | 41 | 42 | class Normalize(object): 43 | """Bin-wise Normalization of the Mel-Spectrograms.""" 44 | 45 | def __init__(self): 46 | # Use the pre-calculated mean and standard deviations 47 | self.mean = np.load('mean_final.npy') 48 | self.std = np.load('std_final.npy') 49 | self.mean = torch.from_numpy(self.mean) 50 | self.std = torch.from_numpy(self.std) 51 | self.mean = torch.reshape(self.mean, [40,1]) 52 | self.std = torch.reshape(self.std, [40, 1]) 53 | 54 | def __call__(self, sample): 55 | data, label = sample['data'], sample['label'] 56 | 57 | data[:,:,:,0] = (data[:,:,:,0] - self.mean)/self.std 58 | 59 | return {'data': data, 60 | 'label': label} 61 | 62 | class ToSubSpectrograms(object): 63 | """ Generate Sub-Spectrogram Tensors """ 64 | def __init__(self, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins): 65 | """ 66 | Parameters 67 | ---------- 68 | sub_spectrogram_size : int 69 | Size of the SubSpectrogram. Default: 20 70 | 71 | sub_spectrogram_mel_hop : int 72 | Mel-bin hop size of the SubSpectrogram. Default 10 73 | 74 | n_mel_bins : int 75 | Number of mel-bins of the Spectrogram extracted. Default: 40. 76 | """ 77 | self.sub_spectrogram_size, self.sub_spectrogram_mel_hop, self.n_mel_bins = sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins 78 | 79 | def __call__(self, sample): 80 | """ 81 | Parameters 82 | ---------- 83 | sample : PyTorch tensor 84 | The input tensor data and label 85 | Returns 86 | ------- 87 | sub_spectrograms: tensor 88 | A list of sub-spectrograms. Default size [channels, sub_spectrogram_size, time_indices, n_sub_spectrograms] 89 | label: tensor 90 | Corresponding label 91 | """ 92 | spectrogram, label = sample['data'], sample['label'] 93 | 94 | i = 0 95 | sub_spectrograms = torch.from_numpy(np.asarray([])) 96 | while(self.sub_spectrogram_mel_hop*i <= self.n_mel_bins - self.sub_spectrogram_size): 97 | 98 | # Extract a Sub-Spectrogram 99 | subspectrogram = spectrogram[:,i*self.sub_spectrogram_mel_hop:i*self.sub_spectrogram_mel_hop+self.sub_spectrogram_size,:, :] 100 | 101 | if i == 0: 102 | sub_spectrograms = subspectrogram 103 | else: 104 | sub_spectrograms = torch.cat((subspectrogram, sub_spectrograms), 3) 105 | 106 | i = i + 1 107 | 108 | return sub_spectrograms, label 109 | 110 | class DCASEDataset(Dataset): 111 | """ DCASE 2018 Dataset extraction """ 112 | 113 | def __init__(self, csv_file, root_dir, transform=None): 114 | """ 115 | Parameters 116 | ---------- 117 | csv_file : str 118 | Location of the CSV file, with respect to the root_dir (should be something like 'evaluation_setup/fold1_train.txt') 119 | 120 | root_dir : int 121 | Root directory of the dataset folder (which contains 'audio' and 'evaluation_setup' folders). 122 | 123 | transform : PyTorch transforms, optional 124 | Used for transforming the data 125 | """ 126 | 127 | list1 = [] 128 | list2 = [] 129 | with open(root_dir + csv_file, 'r') as f: 130 | content = f.readlines() 131 | for x in content: 132 | row = x.split() 133 | list1.append(row[0]) 134 | list2.append(row[1]) 135 | self.root_dir = root_dir 136 | self.transform = transform 137 | self.datalist = list1 138 | self.labels = list2 139 | self.default_labels = ['airport','bus','metro','metro_station','park','public_square','shopping_mall','street_pedestrian','street_traffic','tram'] 140 | 141 | def __len__(self): 142 | """ set the len(object) funciton """ 143 | return len(self.datalist) 144 | 145 | def __getitem__(self, idx): 146 | """ 147 | Function to extract the spectrogram samples and labels from the audio dataset. 148 | """ 149 | wav_name = os.path.join(self.root_dir, 150 | self.datalist[idx]) 151 | 152 | # extracting with 22050 sampling rate by default 153 | audioContainer = dcase_util.containers.AudioContainer().load(filename=wav_name, fs=22050) 154 | # use only one channel (NOTE: In the paper, both channels are used) 155 | audio = audioContainer.data[0] 156 | sr = audioContainer.fs 157 | 158 | # extract mel-spectrogram. results in a time-frequency matrix of 40x500 size. 159 | spec = librosa.feature.melspectrogram(y=audio, sr=sr, S=None, n_fft=883, hop_length=441, n_mels=40) 160 | logmel = librosa.core.amplitude_to_db(spec) 161 | logmel = np.reshape(logmel, [1, logmel.shape[0], logmel.shape[1], 1]) 162 | 163 | label = np.asarray(self.default_labels.index(self.labels[idx])) 164 | sample = {'data': logmel, 'label': label} 165 | if self.transform: 166 | sample = self.transform(sample) 167 | return sample 168 | 169 | def get_data_loaders(train_batch_size, test_batch_size, sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins, use_cuda, root_dir, train_dir, eval_dir): 170 | """ 171 | Function to return the data loaders 172 | 173 | Parameters 174 | ---------- 175 | train_batch_size : int 176 | Size of the training batch. Default: 16 177 | 178 | test_batch_size : int 179 | size of the testing batch. Default: 16 180 | 181 | sub_spectrogram_size : int 182 | Size of the SubSpectrogram. Default 20 183 | 184 | sub_spectrogram_mel_hop : int 185 | Mel-bin hop size of the SubSpectrogram. Default 10 186 | 187 | n_mel_bins : int 188 | Number of mel-bins of the Spectrogram extracted. Default: 40. 189 | 190 | use_gpu : Bool 191 | Use GPU or not. Default: True 192 | 193 | root_dir : str 194 | Directory of the folder which contains the dataset (has 'audio' and 'evaluation_setup' folders inside) 195 | 196 | train_dir : str 197 | Set as default: 'evaluation_setup/train_fold1.txt' 198 | 199 | eval_dir : str 200 | Set as default: 'evaluation_setup/evaluate_fold1.txt' 201 | 202 | Returns 203 | ------- 204 | train_loader and val_loader 205 | data loading objects 206 | """ 207 | 208 | kwargs = {'num_workers': 16, 'pin_memory': True} if use_cuda else {} 209 | 210 | data_transform = transforms.Compose([ 211 | ToTensor(), Normalize(), ToSubSpectrograms(sub_spectrogram_size, sub_spectrogram_mel_hop, n_mel_bins) 212 | ]) 213 | 214 | dcase_train = DCASEDataset(csv_file=train_dir, 215 | root_dir=root_dir, transform=data_transform) 216 | dcase_test = DCASEDataset(csv_file=eval_dir, 217 | root_dir=root_dir, transform=data_transform) 218 | 219 | train_loader = torch.utils.data.DataLoader(dcase_train, 220 | batch_size=train_batch_size, shuffle=True, **kwargs) 221 | 222 | val_loader = torch.utils.data.DataLoader(dcase_test, 223 | batch_size=test_batch_size, shuffle=True, **kwargs) 224 | 225 | return train_loader, val_loader 226 | 227 | def create_summary_writer(model, data_loader, log_dir): 228 | """ 229 | Create the summary writer for TensorBoard 230 | 231 | Parameters 232 | ---------- 233 | model : PyTorch model object 234 | Size of the training batch. 235 | 236 | data_loader : data_loader 237 | Data loader object to create the graph 238 | 239 | log_dir : str 240 | Directory to save the logs 241 | 242 | Returns 243 | ------- 244 | train_loader and val_loader 245 | data loading objects 246 | """ 247 | writer = SummaryWriter(log_dir=log_dir) 248 | data_loader_iter = iter(data_loader) 249 | x, y = next(data_loader_iter) 250 | try: 251 | writer.add_graph(model, x) 252 | except Exception as e: 253 | print("Failed to save model graph: {}".format(e)) 254 | return writer 255 | 256 | def prepare_batch(batch, device=None, non_blocking=False): 257 | """ 258 | Inbuilt function in the ignite._utils, for converting the data to tensors. 259 | Returns the tensors of the input data, using convert_tensor function. 260 | """ 261 | x, y = batch 262 | return (convert_tensor(x, device=device, non_blocking=non_blocking), 263 | convert_tensor(y, device=device, non_blocking=non_blocking)) 264 | --------------------------------------------------------------------------------