├── .gitignore ├── LICENSE ├── README.md ├── chord_recognition.py ├── finetune.py ├── main.py ├── midi2remi.ipynb ├── model.py ├── modules.py ├── result ├── continuation.midi └── from_scratch.midi └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | __pycache__/ 2 | .vscode/ 3 | .ipynb_checkpoints/ 4 | .DS_Store 5 | miditoolkit.egg-info/ 6 | @eaDir 7 | *.pyc 8 | *.pypirc 9 | Thumbs.db 10 | *.gz -------------------------------------------------------------------------------- /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 | # REMI 2 | Authors: [Yu-Siang Huang](https://remyhuang.github.io/), [Yi-Hsuan Yang](http://mac.citi.sinica.edu.tw/~yang/) 3 | 4 | [**Paper (arXiv)**](https://arxiv.org/abs/2002.00212) | [**Blog**](https://ailabs.tw/human-interaction/pop-music-transformer/) | [**Audio demo (Google Drive)**](https://drive.google.com/open?id=1LzPBjHPip4S0CBOLquk5CNapvXSfys54) | [**Online interactive demo**](https://vibertthio.com/transformer/) 5 | 6 | REMI, which stands for `REvamped MIDI-derived events`, is a new event representation we propose for converting MIDI scores into text-like discrete tokens. Compared to the MIDI-like event representation adopted in exising Transformer-based music composition models, REMI provides sequence models a metrical context for modeling the rhythmic patterns of music. Using REMI as the event representation, we train a Transformer-XL model to generate minute-long Pop piano music with expressive, coherent and clear structure of rhythm and harmony, without needing any post-processing to refine the result. The model also provides controllability of local tempo changes and chord progression. 7 | 8 | ## Citation 9 | ``` 10 | @inproceedings{10.1145/3394171.3413671, 11 | author = {Huang, Yu-Siang and Yang, Yi-Hsuan}, 12 | title = {Pop Music Transformer: Beat-Based Modeling and Generation of Expressive Pop Piano Compositions}, 13 | year = {2020}, 14 | isbn = {9781450379885}, 15 | publisher = {Association for Computing Machinery}, 16 | address = {New York, NY, USA}, 17 | url = {https://doi.org/10.1145/3394171.3413671}, 18 | doi = {10.1145/3394171.3413671}, 19 | pages = {1180–1188}, 20 | numpages = {9}, 21 | location = {Seattle, WA, USA}, 22 | series = {MM '20} 23 | } 24 | ``` 25 | 26 | ## Getting Started 27 | ### Install Dependencies 28 | * python 3.6 (recommend using [Anaconda](https://www.anaconda.com/distribution/)) 29 | * tensorflow-gpu 1.14.0 (`pip install tensorflow-gpu==1.14.0`) 30 | * [miditoolkit](https://github.com/YatingMusic/miditoolkit) (`pip install miditoolkit`) 31 | 32 | ### Download Pre-trained Checkpoints 33 | We provide two pre-trained checkpoints for generating samples. 34 | * `REMI-tempo-checkpoint` [(428 MB)](https://drive.google.com/open?id=1gxuTSkF51NP04JZgTE46Pg4KQsbHQKGo) 35 | * `REMI-tempo-chord-checkpoint` [(429 MB)](https://drive.google.com/open?id=1nAKjaeahlzpVAX0F9wjQEG_hL4UosSbo) 36 | 37 | ### Obtain the MIDI Data 38 | We provide the MIDI files including local tempo changes and estimated chord. [(5 MB)](https://drive.google.com/open?id=1JUDHGrVYGyHtjkfI2vgR1xb2oU8unlI3) 39 | * `data/train`: 775 files used for training models 40 | * `data/evaluation`: 100 files (prompts) used for the continuation experiments 41 | 42 | ## Generate Samples 43 | See `main.py` as an example: 44 | ```python 45 | from model import PopMusicTransformer 46 | import os 47 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 48 | 49 | def main(): 50 | # declare model 51 | model = PopMusicTransformer( 52 | checkpoint='REMI-tempo-checkpoint', 53 | is_training=False) 54 | 55 | # generate from scratch 56 | model.generate( 57 | n_target_bar=16, 58 | temperature=1.2, 59 | topk=5, 60 | output_path='./result/from_scratch.midi', 61 | prompt=None) 62 | 63 | # generate continuation 64 | model.generate( 65 | n_target_bar=16, 66 | temperature=1.2, 67 | topk=5, 68 | output_path='./result/continuation.midi', 69 | prompt='./data/evaluation/000.midi') 70 | 71 | # close model 72 | model.close() 73 | 74 | if __name__ == '__main__': 75 | main() 76 | ``` 77 | 78 | ## Convert MIDI to REMI 79 | You can find out how to convert the MIDI messages into REMI events in the `midi2remi.ipynb`. 80 | 81 | ## FAQ 82 | #### 1. How to synthesize the audio files (e.g., mp3)? 83 | We strongly recommend using DAW (e.g., Logic Pro) to open/play the generated MIDI files. Or, you can use [FluidSynth](https://github.com/FluidSynth/fluidsynth) with a [SoundFont](https://sites.google.com/site/soundfonts4u/). However, it may not be able to correctly handle the tempo changes (see [fluidsynth/issues/141](https://github.com/FluidSynth/fluidsynth/issues/141)). 84 | 85 | #### 2. What is the function of the inputs "temperature" and "topk"? 86 | It is the temperature-controlled stochastic sampling methods are used for generating text from a trained language model. You can find out more details in the reference paper [CTRL: 4.1 Sampling](https://einstein.ai/presentations/ctrl.pdf). 87 | > It is worth noting that the sampling method used for generation is very critical to the quality of the output, which is a research topic worthy of further exploration. 88 | 89 | #### 3. How to finetune with my personal MIDI data? 90 | Please see [issue/Training on custom MIDI corpus](https://github.com/YatingMusic/remi/issues/2) 91 | 92 | ## Acknowledgement 93 | - The content of `modules.py` comes from the [kimiyoung/transformer-xl](https://github.com/kimiyoung/transformer-xl) repository. 94 | - Thanks [@vibertthio](https://github.com/vibertthio) for the awesome online interactive demo. 95 | -------------------------------------------------------------------------------- /chord_recognition.py: -------------------------------------------------------------------------------- 1 | import miditoolkit 2 | import numpy as np 3 | 4 | class MIDIChord(object): 5 | def __init__(self): 6 | # define pitch classes 7 | self.PITCH_CLASSES = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B'] 8 | # define chord maps (required) 9 | self.CHORD_MAPS = {'maj': [0, 4], 10 | 'min': [0, 3], 11 | 'dim': [0, 3, 6], 12 | 'aug': [0, 4, 8], 13 | 'dom': [0, 4, 7, 10]} 14 | # define chord insiders (+1) 15 | self.CHORD_INSIDERS = {'maj': [7], 16 | 'min': [7], 17 | 'dim': [9], 18 | 'aug': [], 19 | 'dom': []} 20 | # define chord outsiders (-1) 21 | self.CHORD_OUTSIDERS_1 = {'maj': [2, 5, 9], 22 | 'min': [2, 5, 8], 23 | 'dim': [2, 5, 10], 24 | 'aug': [2, 5, 9], 25 | 'dom': [2, 5, 9]} 26 | # define chord outsiders (-2) 27 | self.CHORD_OUTSIDERS_2 = {'maj': [1, 3, 6, 8, 10], 28 | 'min': [1, 4, 6, 9, 11], 29 | 'dim': [1, 4, 7, 8, 11], 30 | 'aug': [1, 3, 6, 7, 10], 31 | 'dom': [1, 3, 6, 8, 11]} 32 | 33 | def note2pianoroll(self, notes, max_tick, ticks_per_beat): 34 | return miditoolkit.pianoroll.parser.notes2pianoroll( 35 | note_stream_ori=notes, 36 | max_tick=max_tick, 37 | ticks_per_beat=ticks_per_beat) 38 | 39 | def sequencing(self, chroma): 40 | candidates = {} 41 | for index in range(len(chroma)): 42 | if chroma[index]: 43 | root_note = index 44 | _chroma = np.roll(chroma, -root_note) 45 | sequence = np.where(_chroma == 1)[0] 46 | candidates[root_note] = list(sequence) 47 | return candidates 48 | 49 | def scoring(self, candidates): 50 | scores = {} 51 | qualities = {} 52 | for root_note, sequence in candidates.items(): 53 | if 3 not in sequence and 4 not in sequence: 54 | scores[root_note] = -100 55 | qualities[root_note] = 'None' 56 | elif 3 in sequence and 4 in sequence: 57 | scores[root_note] = -100 58 | qualities[root_note] = 'None' 59 | else: 60 | # decide quality 61 | if 3 in sequence: 62 | if 6 in sequence: 63 | quality = 'dim' 64 | else: 65 | quality = 'min' 66 | elif 4 in sequence: 67 | if 8 in sequence: 68 | quality = 'aug' 69 | else: 70 | if 7 in sequence and 10 in sequence: 71 | quality = 'dom' 72 | else: 73 | quality = 'maj' 74 | # decide score 75 | maps = self.CHORD_MAPS.get(quality) 76 | _notes = [n for n in sequence if n not in maps] 77 | score = 0 78 | for n in _notes: 79 | if n in self.CHORD_OUTSIDERS_1.get(quality): 80 | score -= 1 81 | elif n in self.CHORD_OUTSIDERS_2.get(quality): 82 | score -= 2 83 | elif n in self.CHORD_INSIDERS.get(quality): 84 | score += 1 85 | scores[root_note] = score 86 | qualities[root_note] = quality 87 | return scores, qualities 88 | 89 | def find_chord(self, pianoroll): 90 | chroma = miditoolkit.pianoroll.utils.tochroma(pianoroll=pianoroll) 91 | chroma = np.sum(chroma, axis=0) 92 | chroma = np.array([1 if c else 0 for c in chroma]) 93 | if np.sum(chroma) == 0: 94 | return 'N', 'N', 'N', 0 95 | else: 96 | candidates = self.sequencing(chroma=chroma) 97 | scores, qualities = self.scoring(candidates=candidates) 98 | # bass note 99 | sorted_notes = [] 100 | for i, v in enumerate(np.sum(pianoroll, axis=0)): 101 | if v > 0: 102 | sorted_notes.append(int(i%12)) 103 | bass_note = sorted_notes[0] 104 | # root note 105 | __root_note = [] 106 | _max = max(scores.values()) 107 | for _root_note, score in scores.items(): 108 | if score == _max: 109 | __root_note.append(_root_note) 110 | if len(__root_note) == 1: 111 | root_note = __root_note[0] 112 | else: 113 | #TODO: what should i do 114 | for n in sorted_notes: 115 | if n in __root_note: 116 | root_note = n 117 | break 118 | # quality 119 | quality = qualities.get(root_note) 120 | sequence = candidates.get(root_note) 121 | # score 122 | score = scores.get(root_note) 123 | return self.PITCH_CLASSES[root_note], quality, self.PITCH_CLASSES[bass_note], score 124 | 125 | def greedy(self, candidates, max_tick, min_length): 126 | chords = [] 127 | # start from 0 128 | start_tick = 0 129 | while start_tick < max_tick: 130 | _candidates = candidates.get(start_tick) 131 | _candidates = sorted(_candidates.items(), key=lambda x: (x[1][-1], x[0])) 132 | # choose 133 | end_tick, (root_note, quality, bass_note, _) = _candidates[-1] 134 | if root_note == bass_note: 135 | chord = '{}:{}'.format(root_note, quality) 136 | else: 137 | chord = '{}:{}/{}'.format(root_note, quality, bass_note) 138 | chords.append([start_tick, end_tick, chord]) 139 | start_tick = end_tick 140 | # remove :None 141 | temp = chords 142 | while ':None' in temp[0][-1]: 143 | try: 144 | temp[1][0] = temp[0][0] 145 | del temp[0] 146 | except: 147 | print('NO CHORD') 148 | return [] 149 | temp2 = [] 150 | for chord in temp: 151 | if ':None' not in chord[-1]: 152 | temp2.append(chord) 153 | else: 154 | temp2[-1][1] = chord[1] 155 | return temp2 156 | 157 | def extract(self, notes): 158 | # read 159 | max_tick = max([n.end for n in notes]) 160 | ticks_per_beat = 480 161 | pianoroll = self.note2pianoroll( 162 | notes=notes, 163 | max_tick=max_tick, 164 | ticks_per_beat=ticks_per_beat) 165 | # get lots of candidates 166 | candidates = {} 167 | # the shortest: 2 beat, longest: 4 beat 168 | for interval in [4, 2]: 169 | for start_tick in range(0, max_tick, ticks_per_beat): 170 | # set target pianoroll 171 | end_tick = int(ticks_per_beat * interval + start_tick) 172 | if end_tick > max_tick: 173 | end_tick = max_tick 174 | _pianoroll = pianoroll[start_tick:end_tick, :] 175 | # find chord 176 | root_note, quality, bass_note, score = self.find_chord(pianoroll=_pianoroll) 177 | # save 178 | if start_tick not in candidates: 179 | candidates[start_tick] = {} 180 | candidates[start_tick][end_tick] = (root_note, quality, bass_note, score) 181 | else: 182 | if end_tick not in candidates[start_tick]: 183 | candidates[start_tick][end_tick] = (root_note, quality, bass_note, score) 184 | # greedy 185 | chords = self.greedy(candidates=candidates, 186 | max_tick=max_tick, 187 | min_length=ticks_per_beat) 188 | return chords 189 | -------------------------------------------------------------------------------- /finetune.py: -------------------------------------------------------------------------------- 1 | from model import PopMusicTransformer 2 | from glob import glob 3 | import os 4 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 5 | 6 | def main(): 7 | # declare model 8 | model = PopMusicTransformer( 9 | checkpoint='REMI-tempo-checkpoint', 10 | is_training=True) 11 | # prepare data 12 | midi_paths = glob('YOUR PERSOANL FOLDER/*.midi') # you need to revise it 13 | training_data = model.prepare_data(midi_paths=midi_paths) 14 | 15 | # check output checkpoint folder 16 | #################################### 17 | # if you use "REMI-tempo-chord-checkpoint" for the pre-trained checkpoint 18 | # please name your output folder as something with "chord" 19 | # for example: my-love-chord, cute-doggy-chord, ... 20 | # if use "REMI-tempo-checkpoint" 21 | # for example: my-love, cute-doggy, ... 22 | #################################### 23 | output_checkpoint_folder = 'REMI-finetune' # your decision 24 | if not os.path.exists(output_checkpoint_folder): 25 | os.mkdir(output_checkpoint_folder) 26 | 27 | # finetune 28 | model.finetune( 29 | training_data=training_data, 30 | output_checkpoint_folder=output_checkpoint_folder) 31 | 32 | #################################### 33 | # after finetuning, please choose which checkpoint you want to try 34 | # and change the checkpoint names you choose into "model" 35 | # and copy the "dictionary.pkl" into the your output_checkpoint_folder 36 | # ***** the same as the content format in "REMI-tempo-checkpoint" ***** 37 | # and then, you can use "main.py" to generate your own music! 38 | # (do not forget to revise the checkpoint path to your own in "main.py") 39 | #################################### 40 | 41 | # close 42 | model.close() 43 | 44 | if __name__ == '__main__': 45 | main() 46 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | from model import PopMusicTransformer 2 | import os 3 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 4 | 5 | def main(): 6 | # declare model 7 | model = PopMusicTransformer( 8 | checkpoint='REMI-tempo-checkpoint', 9 | is_training=False) 10 | 11 | # generate from scratch 12 | model.generate( 13 | n_target_bar=16, 14 | temperature=1.2, 15 | topk=5, 16 | output_path='./result/from_scratch.midi', 17 | prompt=None) 18 | 19 | # generate continuation 20 | model.generate( 21 | n_target_bar=16, 22 | temperature=1.2, 23 | topk=5 24 | output_path='./result/continuation.midi', 25 | prompt='./data/evaluation/000.midi') 26 | 27 | # close model 28 | model.close() 29 | 30 | if __name__ == '__main__': 31 | main() 32 | -------------------------------------------------------------------------------- /midi2remi.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 3, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "import miditoolkit\n", 10 | "import utils" 11 | ] 12 | }, 13 | { 14 | "cell_type": "markdown", 15 | "metadata": {}, 16 | "source": [ 17 | "# Read MIDI (example)" 18 | ] 19 | }, 20 | { 21 | "cell_type": "code", 22 | "execution_count": 4, 23 | "metadata": {}, 24 | "outputs": [], 25 | "source": [ 26 | "midi_obj = miditoolkit.midi.parser.MidiFile('./data/evaluation/000.midi')" 27 | ] 28 | }, 29 | { 30 | "cell_type": "code", 31 | "execution_count": 6, 32 | "metadata": {}, 33 | "outputs": [ 34 | { 35 | "name": "stdout", 36 | "output_type": "stream", 37 | "text": [ 38 | "Note(start=956, end=1530, pitch=59, velocity=55)\n", 39 | "Note(start=1420, end=1998, pitch=60, velocity=57)\n", 40 | "Note(start=1921, end=2519, pitch=43, velocity=58)\n", 41 | "Note(start=3372, end=3848, pitch=67, velocity=64)\n", 42 | "Note(start=1885, end=3960, pitch=62, velocity=71)\n", 43 | "Note(start=2410, end=4109, pitch=50, velocity=62)\n", 44 | "Note(start=2886, end=5285, pitch=59, velocity=69)\n", 45 | "Note(start=5285, end=5872, pitch=59, velocity=68)\n", 46 | "Note(start=3848, end=5910, pitch=67, velocity=71)\n", 47 | "Note(start=5761, end=6723, pitch=60, velocity=72)\n", 48 | "Note(start=6247, end=6795, pitch=52, velocity=57)\n", 49 | "Note(start=5761, end=6904, pitch=43, velocity=66)\n", 50 | "Note(start=7198, end=7673, pitch=69, velocity=68)\n", 51 | "Note(start=6723, end=8710, pitch=60, velocity=66)\n" 52 | ] 53 | } 54 | ], 55 | "source": [ 56 | "print(*midi_obj.instruments[0].notes, sep='\\n')" 57 | ] 58 | }, 59 | { 60 | "cell_type": "code", 61 | "execution_count": 9, 62 | "metadata": {}, 63 | "outputs": [ 64 | { 65 | "name": "stdout", 66 | "output_type": "stream", 67 | "text": [ 68 | "120.0 BPM at 0 ticks\n", 69 | "23.529411764705884 BPM at 480 ticks\n", 70 | "146.34146341463415 BPM at 960 ticks\n", 71 | "139.53488372093022 BPM at 1440 ticks\n", 72 | "146.34146341463415 BPM at 1920 ticks\n", 73 | "142.85714285714286 BPM at 2400 ticks\n", 74 | "146.34146341463415 BPM at 2880 ticks\n", 75 | "142.85714285714286 BPM at 3360 ticks\n", 76 | "146.34146341463415 BPM at 3840 ticks\n", 77 | "142.85714285714286 BPM at 4320 ticks\n" 78 | ] 79 | } 80 | ], 81 | "source": [ 82 | "print(*midi_obj.tempo_changes[:10], sep='\\n')" 83 | ] 84 | }, 85 | { 86 | "cell_type": "markdown", 87 | "metadata": {}, 88 | "source": [ 89 | "# Convert to REMI events" 90 | ] 91 | }, 92 | { 93 | "cell_type": "markdown", 94 | "metadata": {}, 95 | "source": [ 96 | "## 1. Read midi into \"Item\"" 97 | ] 98 | }, 99 | { 100 | "cell_type": "code", 101 | "execution_count": 10, 102 | "metadata": {}, 103 | "outputs": [], 104 | "source": [ 105 | "note_items, tempo_items = utils.read_items('./data/evaluation/000.midi')" 106 | ] 107 | }, 108 | { 109 | "cell_type": "code", 110 | "execution_count": 11, 111 | "metadata": {}, 112 | "outputs": [ 113 | { 114 | "name": "stdout", 115 | "output_type": "stream", 116 | "text": [ 117 | "Item(name=Note, start=956, end=1530, velocity=55, pitch=59)\n", 118 | "Item(name=Note, start=1420, end=1998, velocity=57, pitch=60)\n", 119 | "Item(name=Note, start=1885, end=3960, velocity=71, pitch=62)\n", 120 | "Item(name=Note, start=1921, end=2519, velocity=58, pitch=43)\n", 121 | "Item(name=Note, start=2410, end=4109, velocity=62, pitch=50)\n", 122 | "Item(name=Note, start=2886, end=5285, velocity=69, pitch=59)\n", 123 | "Item(name=Note, start=3372, end=3848, velocity=64, pitch=67)\n", 124 | "Item(name=Note, start=3848, end=5910, velocity=71, pitch=67)\n", 125 | "Item(name=Note, start=5285, end=5872, velocity=68, pitch=59)\n", 126 | "Item(name=Note, start=5761, end=6904, velocity=66, pitch=43)\n", 127 | "Item(name=Note, start=5761, end=6723, velocity=72, pitch=60)\n", 128 | "Item(name=Note, start=6247, end=6795, velocity=57, pitch=52)\n", 129 | "Item(name=Note, start=6723, end=8710, velocity=66, pitch=60)\n", 130 | "Item(name=Note, start=7198, end=7673, velocity=68, pitch=69)\n" 131 | ] 132 | } 133 | ], 134 | "source": [ 135 | "print(*note_items, sep='\\n')" 136 | ] 137 | }, 138 | { 139 | "cell_type": "code", 140 | "execution_count": 13, 141 | "metadata": {}, 142 | "outputs": [ 143 | { 144 | "name": "stdout", 145 | "output_type": "stream", 146 | "text": [ 147 | "Item(name=Tempo, start=0, end=None, velocity=None, pitch=120)\n", 148 | "Item(name=Tempo, start=480, end=None, velocity=None, pitch=23)\n", 149 | "Item(name=Tempo, start=960, end=None, velocity=None, pitch=146)\n", 150 | "Item(name=Tempo, start=1440, end=None, velocity=None, pitch=139)\n", 151 | "Item(name=Tempo, start=1920, end=None, velocity=None, pitch=146)\n", 152 | "Item(name=Tempo, start=2400, end=None, velocity=None, pitch=142)\n", 153 | "Item(name=Tempo, start=2880, end=None, velocity=None, pitch=146)\n", 154 | "Item(name=Tempo, start=3360, end=None, velocity=None, pitch=142)\n", 155 | "Item(name=Tempo, start=3840, end=None, velocity=None, pitch=146)\n", 156 | "Item(name=Tempo, start=4320, end=None, velocity=None, pitch=142)\n" 157 | ] 158 | } 159 | ], 160 | "source": [ 161 | "print(*tempo_items[:10], sep='\\n')" 162 | ] 163 | }, 164 | { 165 | "cell_type": "markdown", 166 | "metadata": {}, 167 | "source": [ 168 | "## 2. Quantize note items" 169 | ] 170 | }, 171 | { 172 | "cell_type": "code", 173 | "execution_count": 14, 174 | "metadata": {}, 175 | "outputs": [], 176 | "source": [ 177 | "note_items = utils.quantize_items(note_items)" 178 | ] 179 | }, 180 | { 181 | "cell_type": "code", 182 | "execution_count": 15, 183 | "metadata": {}, 184 | "outputs": [ 185 | { 186 | "name": "stdout", 187 | "output_type": "stream", 188 | "text": [ 189 | "Item(name=Note, start=960, end=1534, velocity=55, pitch=59)\n", 190 | "Item(name=Note, start=1440, end=2018, velocity=57, pitch=60)\n", 191 | "Item(name=Note, start=1920, end=3995, velocity=71, pitch=62)\n", 192 | "Item(name=Note, start=1920, end=2518, velocity=58, pitch=43)\n", 193 | "Item(name=Note, start=2400, end=4099, velocity=62, pitch=50)\n", 194 | "Item(name=Note, start=2880, end=5279, velocity=69, pitch=59)\n", 195 | "Item(name=Note, start=3360, end=3836, velocity=64, pitch=67)\n", 196 | "Item(name=Note, start=3840, end=5902, velocity=71, pitch=67)\n", 197 | "Item(name=Note, start=5280, end=5867, velocity=68, pitch=59)\n", 198 | "Item(name=Note, start=5760, end=6903, velocity=66, pitch=43)\n", 199 | "Item(name=Note, start=5760, end=6722, velocity=72, pitch=60)\n", 200 | "Item(name=Note, start=6240, end=6788, velocity=57, pitch=52)\n", 201 | "Item(name=Note, start=6720, end=8707, velocity=66, pitch=60)\n", 202 | "Item(name=Note, start=7080, end=7555, velocity=68, pitch=69)\n" 203 | ] 204 | } 205 | ], 206 | "source": [ 207 | "print(*note_items, sep='\\n')" 208 | ] 209 | }, 210 | { 211 | "cell_type": "markdown", 212 | "metadata": {}, 213 | "source": [ 214 | "## 3. extract chord (if needed)" 215 | ] 216 | }, 217 | { 218 | "cell_type": "code", 219 | "execution_count": 16, 220 | "metadata": {}, 221 | "outputs": [], 222 | "source": [ 223 | "chord_items = utils.extract_chords(note_items)" 224 | ] 225 | }, 226 | { 227 | "cell_type": "code", 228 | "execution_count": 17, 229 | "metadata": {}, 230 | "outputs": [ 231 | { 232 | "name": "stdout", 233 | "output_type": "stream", 234 | "text": [ 235 | "Item(name=Chord, start=0, end=960, velocity=None, pitch=N:N)\n", 236 | "Item(name=Chord, start=960, end=2880, velocity=None, pitch=G:maj)\n", 237 | "Item(name=Chord, start=2880, end=4800, velocity=None, pitch=G:maj)\n", 238 | "Item(name=Chord, start=4800, end=6720, velocity=None, pitch=C:maj)\n", 239 | "Item(name=Chord, start=6720, end=8707, velocity=None, pitch=A:min)\n" 240 | ] 241 | } 242 | ], 243 | "source": [ 244 | "print(*chord_items, sep='\\n')" 245 | ] 246 | }, 247 | { 248 | "cell_type": "markdown", 249 | "metadata": {}, 250 | "source": [ 251 | "## 4. group items" 252 | ] 253 | }, 254 | { 255 | "cell_type": "code", 256 | "execution_count": 18, 257 | "metadata": {}, 258 | "outputs": [], 259 | "source": [ 260 | "items = chord_items + tempo_items + note_items\n", 261 | "max_time = note_items[-1].end\n", 262 | "groups = utils.group_items(items, max_time)" 263 | ] 264 | }, 265 | { 266 | "cell_type": "code", 267 | "execution_count": 19, 268 | "metadata": {}, 269 | "outputs": [ 270 | { 271 | "name": "stdout", 272 | "output_type": "stream", 273 | "text": [ 274 | "0\n", 275 | "Item(name=Chord, start=0, end=960, velocity=None, pitch=N:N)\n", 276 | "Item(name=Tempo, start=0, end=None, velocity=None, pitch=120)\n", 277 | "Item(name=Tempo, start=480, end=None, velocity=None, pitch=23)\n", 278 | "Item(name=Chord, start=960, end=2880, velocity=None, pitch=G:maj)\n", 279 | "Item(name=Tempo, start=960, end=None, velocity=None, pitch=146)\n", 280 | "Item(name=Note, start=960, end=1534, velocity=55, pitch=59)\n", 281 | "Item(name=Tempo, start=1440, end=None, velocity=None, pitch=139)\n", 282 | "Item(name=Note, start=1440, end=2018, velocity=57, pitch=60)\n", 283 | "1920\n", 284 | "\n", 285 | "1920\n", 286 | "Item(name=Tempo, start=1920, end=None, velocity=None, pitch=146)\n", 287 | "Item(name=Note, start=1920, end=3995, velocity=71, pitch=62)\n", 288 | "Item(name=Note, start=1920, end=2518, velocity=58, pitch=43)\n", 289 | "Item(name=Tempo, start=2400, end=None, velocity=None, pitch=142)\n", 290 | "Item(name=Note, start=2400, end=4099, velocity=62, pitch=50)\n", 291 | "Item(name=Chord, start=2880, end=4800, velocity=None, pitch=G:maj)\n", 292 | "Item(name=Tempo, start=2880, end=None, velocity=None, pitch=146)\n", 293 | "Item(name=Note, start=2880, end=5279, velocity=69, pitch=59)\n", 294 | "Item(name=Tempo, start=3360, end=None, velocity=None, pitch=142)\n", 295 | "Item(name=Note, start=3360, end=3836, velocity=64, pitch=67)\n", 296 | "3840\n", 297 | "\n", 298 | "3840\n", 299 | "Item(name=Tempo, start=3840, end=None, velocity=None, pitch=146)\n", 300 | "Item(name=Note, start=3840, end=5902, velocity=71, pitch=67)\n", 301 | "Item(name=Tempo, start=4320, end=None, velocity=None, pitch=142)\n", 302 | "Item(name=Chord, start=4800, end=6720, velocity=None, pitch=C:maj)\n", 303 | "Item(name=Tempo, start=4800, end=None, velocity=None, pitch=142)\n", 304 | "Item(name=Tempo, start=5280, end=None, velocity=None, pitch=142)\n", 305 | "Item(name=Note, start=5280, end=5867, velocity=68, pitch=59)\n", 306 | "5760\n", 307 | "\n", 308 | "5760\n", 309 | "Item(name=Tempo, start=5760, end=None, velocity=None, pitch=146)\n", 310 | "Item(name=Note, start=5760, end=6903, velocity=66, pitch=43)\n", 311 | "Item(name=Note, start=5760, end=6722, velocity=72, pitch=60)\n", 312 | "Item(name=Tempo, start=6240, end=None, velocity=None, pitch=142)\n", 313 | "Item(name=Note, start=6240, end=6788, velocity=57, pitch=52)\n", 314 | "Item(name=Chord, start=6720, end=8707, velocity=None, pitch=A:min)\n", 315 | "Item(name=Tempo, start=6720, end=None, velocity=None, pitch=142)\n", 316 | "Item(name=Note, start=6720, end=8707, velocity=66, pitch=60)\n", 317 | "Item(name=Note, start=7080, end=7555, velocity=68, pitch=69)\n", 318 | "Item(name=Tempo, start=7200, end=None, velocity=None, pitch=142)\n", 319 | "7680\n", 320 | "\n" 321 | ] 322 | } 323 | ], 324 | "source": [ 325 | "for g in groups:\n", 326 | " print(*g, sep='\\n')\n", 327 | " print()" 328 | ] 329 | }, 330 | { 331 | "cell_type": "markdown", 332 | "metadata": {}, 333 | "source": [ 334 | "## 5. \"Item\" to \"Event\"" 335 | ] 336 | }, 337 | { 338 | "cell_type": "code", 339 | "execution_count": 20, 340 | "metadata": {}, 341 | "outputs": [], 342 | "source": [ 343 | "events = utils.item2event(groups)" 344 | ] 345 | }, 346 | { 347 | "cell_type": "code", 348 | "execution_count": 23, 349 | "metadata": {}, 350 | "outputs": [ 351 | { 352 | "name": "stdout", 353 | "output_type": "stream", 354 | "text": [ 355 | "Event(name=Bar, time=None, value=None, text=1)\n", 356 | "Event(name=Position, time=0, value=1/16, text=0)\n", 357 | "Event(name=Chord, time=0, value=N:N, text=N:N)\n", 358 | "Event(name=Position, time=0, value=1/16, text=0)\n", 359 | "Event(name=Tempo Class, time=0, value=mid, text=None)\n", 360 | "Event(name=Tempo Value, time=0, value=30, text=None)\n", 361 | "Event(name=Position, time=480, value=5/16, text=480)\n", 362 | "Event(name=Tempo Class, time=480, value=slow, text=None)\n", 363 | "Event(name=Tempo Value, time=480, value=0, text=None)\n", 364 | "Event(name=Position, time=960, value=9/16, text=960)\n", 365 | "Event(name=Chord, time=960, value=G:maj, text=G:maj)\n", 366 | "Event(name=Position, time=960, value=9/16, text=960)\n", 367 | "Event(name=Tempo Class, time=960, value=mid, text=None)\n", 368 | "Event(name=Tempo Value, time=960, value=56, text=None)\n", 369 | "Event(name=Position, time=960, value=9/16, text=960)\n", 370 | "Event(name=Note Velocity, time=960, value=13, text=55/52)\n", 371 | "Event(name=Note On, time=960, value=59, text=59)\n", 372 | "Event(name=Note Duration, time=960, value=9, text=574/600)\n", 373 | "Event(name=Position, time=1440, value=13/16, text=1440)\n", 374 | "Event(name=Tempo Class, time=1440, value=mid, text=None)\n", 375 | "Event(name=Tempo Value, time=1440, value=49, text=None)\n", 376 | "Event(name=Position, time=1440, value=13/16, text=1440)\n", 377 | "Event(name=Note Velocity, time=1440, value=14, text=57/56)\n", 378 | "Event(name=Note On, time=1440, value=60, text=60)\n", 379 | "Event(name=Note Duration, time=1440, value=9, text=578/600)\n", 380 | "Event(name=Bar, time=None, value=None, text=2)\n", 381 | "Event(name=Position, time=1920, value=1/16, text=1920)\n", 382 | "Event(name=Tempo Class, time=1920, value=mid, text=None)\n", 383 | "Event(name=Tempo Value, time=1920, value=56, text=None)\n", 384 | "Event(name=Position, time=1920, value=1/16, text=1920)\n" 385 | ] 386 | } 387 | ], 388 | "source": [ 389 | "print(*events[:30], sep='\\n')" 390 | ] 391 | } 392 | ], 393 | "metadata": { 394 | "kernelspec": { 395 | "display_name": "Python 3", 396 | "language": "python", 397 | "name": "python3" 398 | }, 399 | "language_info": { 400 | "codemirror_mode": { 401 | "name": "ipython", 402 | "version": 3 403 | }, 404 | "file_extension": ".py", 405 | "mimetype": "text/x-python", 406 | "name": "python", 407 | "nbconvert_exporter": "python", 408 | "pygments_lexer": "ipython3", 409 | "version": "3.7.6" 410 | } 411 | }, 412 | "nbformat": 4, 413 | "nbformat_minor": 4 414 | } 415 | -------------------------------------------------------------------------------- /model.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | import numpy as np 3 | import miditoolkit 4 | import modules 5 | import pickle 6 | import utils 7 | import time 8 | 9 | class PopMusicTransformer(object): 10 | ######################################## 11 | # initialize 12 | ######################################## 13 | def __init__(self, checkpoint, is_training=False): 14 | # load dictionary 15 | self.dictionary_path = '{}/dictionary.pkl'.format(checkpoint) 16 | self.event2word, self.word2event = pickle.load(open(self.dictionary_path, 'rb')) 17 | # model settings 18 | self.x_len = 512 19 | self.mem_len = 512 20 | self.n_layer = 12 21 | self.d_embed = 512 22 | self.d_model = 512 23 | self.dropout = 0.1 24 | self.n_head = 8 25 | self.d_head = self.d_model // self.n_head 26 | self.d_ff = 2048 27 | self.n_token = len(self.event2word) 28 | self.learning_rate = 0.0002 29 | # load model 30 | self.is_training = is_training 31 | if self.is_training: 32 | self.batch_size = 4 33 | else: 34 | self.batch_size = 1 35 | self.checkpoint_path = '{}/model'.format(checkpoint) 36 | self.load_model() 37 | 38 | ######################################## 39 | # load model 40 | ######################################## 41 | def load_model(self): 42 | # placeholders 43 | self.x = tf.compat.v1.placeholder(tf.int32, shape=[self.batch_size, None]) 44 | self.y = tf.compat.v1.placeholder(tf.int32, shape=[self.batch_size, None]) 45 | self.mems_i = [tf.compat.v1.placeholder(tf.float32, [self.mem_len, self.batch_size, self.d_model]) for _ in range(self.n_layer)] 46 | # model 47 | self.global_step = tf.compat.v1.train.get_or_create_global_step() 48 | initializer = tf.compat.v1.initializers.random_normal(stddev=0.02, seed=None) 49 | proj_initializer = tf.compat.v1.initializers.random_normal(stddev=0.01, seed=None) 50 | with tf.compat.v1.variable_scope(tf.compat.v1.get_variable_scope()): 51 | xx = tf.transpose(self.x, [1, 0]) 52 | yy = tf.transpose(self.y, [1, 0]) 53 | loss, self.logits, self.new_mem = modules.transformer( 54 | dec_inp=xx, 55 | target=yy, 56 | mems=self.mems_i, 57 | n_token=self.n_token, 58 | n_layer=self.n_layer, 59 | d_model=self.d_model, 60 | d_embed=self.d_embed, 61 | n_head=self.n_head, 62 | d_head=self.d_head, 63 | d_inner=self.d_ff, 64 | dropout=self.dropout, 65 | dropatt=self.dropout, 66 | initializer=initializer, 67 | proj_initializer=proj_initializer, 68 | is_training=self.is_training, 69 | mem_len=self.mem_len, 70 | cutoffs=[], 71 | div_val=-1, 72 | tie_projs=[], 73 | same_length=False, 74 | clamp_len=-1, 75 | input_perms=None, 76 | target_perms=None, 77 | head_target=None, 78 | untie_r=False, 79 | proj_same_dim=True) 80 | self.avg_loss = tf.reduce_mean(loss) 81 | # vars 82 | all_vars = tf.compat.v1.trainable_variables() 83 | grads = tf.gradients(self.avg_loss, all_vars) 84 | grads_and_vars = list(zip(grads, all_vars)) 85 | all_trainable_vars = tf.reduce_sum([tf.reduce_prod(v.shape) for v in tf.compat.v1.trainable_variables()]) 86 | # optimizer 87 | decay_lr = tf.compat.v1.train.cosine_decay( 88 | self.learning_rate, 89 | global_step=self.global_step, 90 | decay_steps=400000, 91 | alpha=0.004) 92 | optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=decay_lr) 93 | self.train_op = optimizer.apply_gradients(grads_and_vars, self.global_step) 94 | # saver 95 | self.saver = tf.compat.v1.train.Saver() 96 | config = tf.compat.v1.ConfigProto(allow_soft_placement=True) 97 | config.gpu_options.allow_growth = True 98 | self.sess = tf.compat.v1.Session(config=config) 99 | self.saver.restore(self.sess, self.checkpoint_path) 100 | 101 | ######################################## 102 | # temperature sampling 103 | ######################################## 104 | def temperature_sampling(self, logits, temperature, topk): 105 | probs = np.exp(logits / temperature) / np.sum(np.exp(logits / temperature)) 106 | if topk == 1: 107 | prediction = np.argmax(probs) 108 | else: 109 | sorted_index = np.argsort(probs)[::-1] 110 | candi_index = sorted_index[:topk] 111 | candi_probs = [probs[i] for i in candi_index] 112 | # normalize probs 113 | candi_probs /= sum(candi_probs) 114 | # choose by predicted probs 115 | prediction = np.random.choice(candi_index, size=1, p=candi_probs)[0] 116 | return prediction 117 | 118 | ######################################## 119 | # extract events for prompt continuation 120 | ######################################## 121 | def extract_events(self, input_path): 122 | note_items, tempo_items = utils.read_items(input_path) 123 | note_items = utils.quantize_items(note_items) 124 | max_time = note_items[-1].end 125 | if 'chord' in self.checkpoint_path: 126 | chord_items = utils.extract_chords(note_items) 127 | items = chord_items + tempo_items + note_items 128 | else: 129 | items = tempo_items + note_items 130 | groups = utils.group_items(items, max_time) 131 | events = utils.item2event(groups) 132 | return events 133 | 134 | ######################################## 135 | # generate 136 | ######################################## 137 | def generate(self, n_target_bar, temperature, topk, output_path, prompt=None): 138 | # if prompt, load it. Or, random start 139 | if prompt: 140 | events = self.extract_events(prompt) 141 | words = [[self.event2word['{}_{}'.format(e.name, e.value)] for e in events]] 142 | words[0].append(self.event2word['Bar_None']) 143 | else: 144 | words = [] 145 | for _ in range(self.batch_size): 146 | ws = [self.event2word['Bar_None']] 147 | if 'chord' in self.checkpoint_path: 148 | tempo_classes = [v for k, v in self.event2word.items() if 'Tempo Class' in k] 149 | tempo_values = [v for k, v in self.event2word.items() if 'Tempo Value' in k] 150 | chords = [v for k, v in self.event2word.items() if 'Chord' in k] 151 | ws.append(self.event2word['Position_1/16']) 152 | ws.append(np.random.choice(chords)) 153 | ws.append(self.event2word['Position_1/16']) 154 | ws.append(np.random.choice(tempo_classes)) 155 | ws.append(np.random.choice(tempo_values)) 156 | else: 157 | tempo_classes = [v for k, v in self.event2word.items() if 'Tempo Class' in k] 158 | tempo_values = [v for k, v in self.event2word.items() if 'Tempo Value' in k] 159 | ws.append(self.event2word['Position_1/16']) 160 | ws.append(np.random.choice(tempo_classes)) 161 | ws.append(np.random.choice(tempo_values)) 162 | words.append(ws) 163 | # initialize mem 164 | batch_m = [np.zeros((self.mem_len, self.batch_size, self.d_model), dtype=np.float32) for _ in range(self.n_layer)] 165 | # generate 166 | original_length = len(words[0]) 167 | initial_flag = 1 168 | current_generated_bar = 0 169 | while current_generated_bar < n_target_bar: 170 | # input 171 | if initial_flag: 172 | temp_x = np.zeros((self.batch_size, original_length)) 173 | for b in range(self.batch_size): 174 | for z, t in enumerate(words[b]): 175 | temp_x[b][z] = t 176 | initial_flag = 0 177 | else: 178 | temp_x = np.zeros((self.batch_size, 1)) 179 | for b in range(self.batch_size): 180 | temp_x[b][0] = words[b][-1] 181 | # prepare feed dict 182 | feed_dict = {self.x: temp_x} 183 | for m, m_np in zip(self.mems_i, batch_m): 184 | feed_dict[m] = m_np 185 | # model (prediction) 186 | _logits, _new_mem = self.sess.run([self.logits, self.new_mem], feed_dict=feed_dict) 187 | # sampling 188 | _logit = _logits[-1, 0] 189 | word = self.temperature_sampling( 190 | logits=_logit, 191 | temperature=temperature, 192 | topk=topk) 193 | words[0].append(word) 194 | # if bar event (only work for batch_size=1) 195 | if word == self.event2word['Bar_None']: 196 | current_generated_bar += 1 197 | # re-new mem 198 | batch_m = _new_mem 199 | # write 200 | if prompt: 201 | utils.write_midi( 202 | words=words[0][original_length:], 203 | word2event=self.word2event, 204 | output_path=output_path, 205 | prompt_path=prompt) 206 | else: 207 | utils.write_midi( 208 | words=words[0], 209 | word2event=self.word2event, 210 | output_path=output_path, 211 | prompt_path=None) 212 | 213 | ######################################## 214 | # prepare training data 215 | ######################################## 216 | def prepare_data(self, midi_paths): 217 | # extract events 218 | all_events = [] 219 | for path in midi_paths: 220 | events = self.extract_events(path) 221 | all_events.append(events) 222 | # event to word 223 | all_words = [] 224 | for events in all_events: 225 | words = [] 226 | for event in events: 227 | e = '{}_{}'.format(event.name, event.value) 228 | if e in self.event2word: 229 | words.append(self.event2word[e]) 230 | else: 231 | # OOV 232 | if event.name == 'Note Velocity': 233 | # replace with max velocity based on our training data 234 | words.append(self.event2word['Note Velocity_21']) 235 | else: 236 | # something is wrong 237 | # you should handle it for your own purpose 238 | print('something is wrong! {}'.format(e)) 239 | all_words.append(words) 240 | # to training data 241 | self.group_size = 5 242 | segments = [] 243 | for words in all_words: 244 | pairs = [] 245 | for i in range(0, len(words)-self.x_len-1, self.x_len): 246 | x = words[i:i+self.x_len] 247 | y = words[i+1:i+self.x_len+1] 248 | pairs.append([x, y]) 249 | pairs = np.array(pairs) 250 | # abandon the last 251 | for i in np.arange(0, len(pairs)-self.group_size, self.group_size*2): 252 | data = pairs[i:i+self.group_size] 253 | if len(data) == self.group_size: 254 | segments.append(data) 255 | segments = np.array(segments) 256 | return segments 257 | 258 | ######################################## 259 | # finetune 260 | ######################################## 261 | def finetune(self, training_data, output_checkpoint_folder): 262 | # shuffle 263 | index = np.arange(len(training_data)) 264 | np.random.shuffle(index) 265 | training_data = training_data[index] 266 | num_batches = len(training_data) // self.batch_size 267 | st = time.time() 268 | for e in range(200): 269 | total_loss = [] 270 | for i in range(num_batches): 271 | segments = training_data[self.batch_size*i:self.batch_size*(i+1)] 272 | batch_m = [np.zeros((self.mem_len, self.batch_size, self.d_model), dtype=np.float32) for _ in range(self.n_layer)] 273 | for j in range(self.group_size): 274 | batch_x = segments[:, j, 0, :] 275 | batch_y = segments[:, j, 1, :] 276 | # prepare feed dict 277 | feed_dict = {self.x: batch_x, self.y: batch_y} 278 | for m, m_np in zip(self.mems_i, batch_m): 279 | feed_dict[m] = m_np 280 | # run 281 | _, gs_, loss_, new_mem_ = self.sess.run([self.train_op, self.global_step, self.avg_loss, self.new_mem], feed_dict=feed_dict) 282 | batch_m = new_mem_ 283 | total_loss.append(loss_) 284 | print('>>> Epoch: {}, Step: {}, Loss: {:.5f}, Time: {:.2f}'.format(e, gs_, loss_, time.time()-st)) 285 | self.saver.save(self.sess, '{}/model-{:03d}-{:.3f}'.format(output_checkpoint_folder, e, np.mean(total_loss))) 286 | # stop 287 | if np.mean(total_loss) <= 0.1: 288 | break 289 | 290 | ######################################## 291 | # close 292 | ######################################## 293 | def close(self): 294 | self.sess.close() 295 | -------------------------------------------------------------------------------- /modules.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | 3 | def embedding_lookup(lookup_table, x): 4 | return tf.compat.v1.nn.embedding_lookup(lookup_table, x) 5 | 6 | 7 | def normal_embedding_lookup(x, n_token, d_embed, d_proj, initializer, 8 | proj_initializer, scope='normal_embed', **kwargs): 9 | emb_scale = d_proj ** 0.5 10 | with tf.compat.v1.variable_scope(scope): 11 | lookup_table = tf.compat.v1.get_variable('lookup_table', [n_token, d_embed], initializer=initializer) 12 | y = embedding_lookup(lookup_table, x) 13 | if d_proj != d_embed: 14 | proj_W = tf.compat.v1.get_variable('proj_W', [d_embed, d_proj], initializer=proj_initializer) 15 | y = tf.einsum('ibe,ed->ibd', y, proj_W) 16 | else: 17 | proj_W = None 18 | ret_params = [lookup_table, proj_W] 19 | y *= emb_scale 20 | return y, ret_params 21 | 22 | 23 | def normal_softmax(hidden, target, n_token, params, scope='normal_softmax', **kwargs): 24 | def _logit(x, W, b, proj): 25 | y = x 26 | if proj is not None: 27 | y = tf.einsum('ibd,ed->ibe', y, proj) 28 | return tf.einsum('ibd,nd->ibn', y, W) + b 29 | 30 | params_W, params_projs = params[0], params[1] 31 | 32 | with tf.compat.v1.variable_scope(scope): 33 | softmax_b = tf.compat.v1.get_variable('bias', [n_token], initializer=tf.zeros_initializer()) 34 | output = _logit(hidden, params_W, softmax_b, params_projs) 35 | nll = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=target, logits=output) 36 | return nll, output 37 | 38 | 39 | def positional_embedding(pos_seq, inv_freq, bsz=None): 40 | sinusoid_inp = tf.einsum('i,j->ij', pos_seq, inv_freq) 41 | pos_emb = tf.concat([tf.sin(sinusoid_inp), tf.cos(sinusoid_inp)], -1) 42 | if bsz is not None: 43 | return tf.tile(pos_emb[:, None, :], [1, bsz, 1]) 44 | else: 45 | return pos_emb[:, None, :] 46 | 47 | 48 | def positionwise_FF(inp, d_model, d_inner, dropout, kernel_initializer, 49 | scope='ff', is_training=True): 50 | output = inp 51 | with tf.compat.v1.variable_scope(scope): 52 | output = tf.keras.layers.Dense(d_inner, activation=tf.nn.relu, 53 | kernel_initializer=kernel_initializer, name='layer_1')(inp) 54 | output = tf.keras.layers.Dropout(dropout, name='drop_1')(output, training=is_training) 55 | output = tf.keras.layers.Dense(d_model, activation=tf.nn.relu, 56 | kernel_initializer=kernel_initializer, name='layer_2')(output) 57 | output = tf.keras.layers.Dropout(dropout, name='drop_2')(output, training=is_training) 58 | output = tf.keras.layers.LayerNormalization(axis=-1)(output + inp) 59 | return output 60 | 61 | 62 | def _create_mask(qlen, mlen, same_length=False): 63 | attn_mask = tf.ones([qlen, qlen]) 64 | mask_u = tf.linalg.band_part(attn_mask, 0, -1) 65 | mask_dia = tf.linalg.band_part(attn_mask, 0, 0) 66 | attn_mask_pad = tf.zeros([qlen, mlen]) 67 | ret = tf.concat([attn_mask_pad, mask_u - mask_dia], 1) 68 | if same_length: 69 | mask_l = tf.matrix_band_part(attn_mask, -1, 0) 70 | ret = tf.concat([ret[:, :qlen] + mask_l - mask_dia, ret[:, qlen:]], 1) 71 | return ret 72 | 73 | 74 | def _cache_mem(curr_out, prev_mem, mem_len=None): 75 | if mem_len is None or prev_mem is None: 76 | new_mem = curr_out 77 | elif mem_len == 0: 78 | return prev_mem 79 | else: 80 | new_mem = tf.concat([prev_mem, curr_out], 0)[-mem_len:] 81 | return tf.stop_gradient(new_mem) 82 | 83 | 84 | def rel_shift(x): 85 | x_size = tf.shape(x) 86 | x = tf.pad(x, [[0, 0], [1, 0], [0, 0], [0, 0]]) 87 | x = tf.reshape(x, [x_size[1] + 1, x_size[0], x_size[2], x_size[3]]) 88 | x = tf.slice(x, [1, 0, 0, 0], [-1, -1, -1, -1]) 89 | x = tf.reshape(x, x_size) 90 | return x 91 | 92 | 93 | def rel_multihead_attn(w, r, r_w_bias, r_r_bias, attn_mask, mems, d_model, 94 | n_head, d_head, dropout, dropatt, is_training, 95 | kernel_initializer, scope='rel_attn'): 96 | scale = 1 / (d_head ** 0.5) 97 | with tf.compat.v1.variable_scope(scope): 98 | qlen = tf.shape(w)[0] 99 | rlen = tf.shape(r)[0] 100 | bsz = tf.shape(w)[1] 101 | 102 | cat = tf.concat([mems, w], 0) if mems is not None and mems.shape.ndims > 1 else w 103 | 104 | w_heads = tf.keras.layers.Dense(3 * n_head * d_head, use_bias=False, 105 | kernel_initializer=kernel_initializer, name='qkv')(cat) 106 | r_head_k = tf.keras.layers.Dense(n_head * d_head, use_bias=False, 107 | kernel_initializer=kernel_initializer, name='r')(r) 108 | 109 | w_head_q, w_head_k, w_head_v = tf.split(w_heads, 3, -1) 110 | w_head_q = w_head_q[-qlen:] 111 | 112 | klen = tf.shape(w_head_k)[0] 113 | 114 | w_head_q = tf.reshape(w_head_q, [qlen, bsz, n_head, d_head]) 115 | w_head_k = tf.reshape(w_head_k, [klen, bsz, n_head, d_head]) 116 | w_head_v = tf.reshape(w_head_v, [klen, bsz, n_head, d_head]) 117 | 118 | r_head_k = tf.reshape(r_head_k, [rlen, n_head, d_head]) 119 | 120 | rw_head_q = w_head_q + r_w_bias 121 | rr_head_q = w_head_q + r_r_bias 122 | 123 | AC = tf.einsum('ibnd,jbnd->ijbn', rw_head_q, w_head_k) 124 | BD = tf.einsum('ibnd,jnd->ijbn', rr_head_q, r_head_k) 125 | BD = rel_shift(BD) 126 | 127 | attn_score = (AC + BD) * scale 128 | attn_mask_t = attn_mask[:, :, None, None] 129 | attn_score = attn_score * (1 - attn_mask_t) - 1e30 * attn_mask_t 130 | 131 | attn_prob = tf.nn.softmax(attn_score, 1) 132 | attn_prob = tf.keras.layers.Dropout(dropatt)(attn_prob, training=is_training) 133 | 134 | attn_vec = tf.einsum('ijbn,jbnd->ibnd', attn_prob, w_head_v) 135 | size_t = tf.shape(attn_vec) 136 | attn_vec = tf.reshape(attn_vec, [size_t[0], size_t[1], n_head * d_head]) 137 | 138 | attn_out = tf.keras.layers.Dense(d_model, use_bias=False, 139 | kernel_initializer=kernel_initializer, name='o')(attn_vec) 140 | attn_out = tf.keras.layers.Dropout(dropout)(attn_out, training=is_training) 141 | output = tf.keras.layers.LayerNormalization(axis=-1)(attn_out + w) 142 | return output 143 | 144 | 145 | def transformer(dec_inp, target, mems, n_token, n_layer, d_model, d_embed, 146 | n_head, d_head, d_inner, dropout, dropatt, 147 | initializer, is_training, proj_initializer=None, 148 | mem_len=None, cutoffs=[], div_val=1, tie_projs=[], 149 | same_length=False, clamp_len=-1, 150 | input_perms=None, target_perms=None, head_target=None, 151 | untie_r=False, proj_same_dim=True, 152 | scope='transformer'): 153 | """ 154 | cutoffs: a list of python int. Cutoffs for adaptive softmax. 155 | tie_projs: a list of python bools. Whether to tie the projections. 156 | perms: a list of tensors. Each tensor should of size [len, bsz, bin_size]. 157 | Only used in the adaptive setting. 158 | """ 159 | new_mems = [] 160 | with tf.compat.v1.variable_scope(scope): 161 | if untie_r: 162 | r_w_bias = tf.compat.v1.get_variable('r_w_bias', [n_layer, n_head, d_head], initializer=initializer) 163 | r_r_bias = tf.compat.v1.get_variable('r_r_bias', [n_layer, n_head, d_head], initializer=initializer) 164 | else: 165 | r_w_bias = tf.compat.v1.get_variable('r_w_bias', [n_head, d_head], initializer=initializer) 166 | r_r_bias = tf.compat.v1.get_variable('r_r_bias', [n_head, d_head], initializer=initializer) 167 | 168 | qlen = tf.shape(dec_inp)[0] 169 | mlen = tf.shape(mems[0])[0] if mems is not None else 0 170 | klen = qlen + mlen 171 | 172 | if proj_initializer is None: 173 | proj_initializer = initializer 174 | 175 | embeddings, shared_params = normal_embedding_lookup( 176 | x=dec_inp, 177 | n_token=n_token, 178 | d_embed=d_embed, 179 | d_proj=d_model, 180 | initializer=initializer, 181 | proj_initializer=proj_initializer) 182 | 183 | attn_mask = _create_mask(qlen, mlen, same_length) 184 | 185 | pos_seq = tf.range(klen - 1, -1, -1.0) 186 | if clamp_len > 0: 187 | pos_seq = tf.minimum(pos_seq, clamp_len) 188 | inv_freq = 1 / (10000 ** (tf.range(0, d_model, 2.0) / d_model)) 189 | pos_emb = positional_embedding(pos_seq, inv_freq) 190 | 191 | output = tf.keras.layers.Dropout(rate=dropout)(embeddings, training=is_training) 192 | pos_emb = tf.keras.layers.Dropout(rate=dropout)(pos_emb, training=is_training) 193 | 194 | if mems is None: 195 | mems = [None] * n_layer 196 | 197 | for i in range(n_layer): 198 | # cache new mems 199 | new_mems.append(_cache_mem(output, mems[i], mem_len)) 200 | 201 | with tf.compat.v1.variable_scope('layer_{}'.format(i)): 202 | output = rel_multihead_attn( 203 | w=output, 204 | r=pos_emb, 205 | r_w_bias=r_w_bias if not untie_r else r_w_bias[i], 206 | r_r_bias=r_r_bias if not untie_r else r_r_bias[i], 207 | attn_mask=attn_mask, 208 | mems=mems[i], 209 | d_model=d_model, 210 | n_head=n_head, 211 | d_head=d_head, 212 | dropout=dropout, 213 | dropatt=dropatt, 214 | is_training=is_training, 215 | kernel_initializer=initializer) 216 | 217 | output = positionwise_FF( 218 | inp=output, 219 | d_model=d_model, 220 | d_inner=d_inner, 221 | dropout=dropout, 222 | kernel_initializer=initializer, 223 | is_training=is_training) 224 | 225 | output = tf.keras.layers.Dropout(dropout)(output, training=is_training) 226 | 227 | loss, logits = normal_softmax( 228 | hidden=output, 229 | target=target, 230 | n_token=n_token, 231 | params=shared_params) 232 | 233 | return loss, logits, new_mems -------------------------------------------------------------------------------- /result/continuation.midi: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/YatingMusic/remi/6d407258fa5828600a5474354862353ef4e4e8ae/result/continuation.midi -------------------------------------------------------------------------------- /result/from_scratch.midi: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/YatingMusic/remi/6d407258fa5828600a5474354862353ef4e4e8ae/result/from_scratch.midi -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | import chord_recognition 2 | import numpy as np 3 | import miditoolkit 4 | import copy 5 | 6 | # parameters for input 7 | DEFAULT_VELOCITY_BINS = np.linspace(0, 128, 32+1, dtype=np.int) 8 | DEFAULT_FRACTION = 16 9 | DEFAULT_DURATION_BINS = np.arange(60, 3841, 60, dtype=int) 10 | DEFAULT_TEMPO_INTERVALS = [range(30, 90), range(90, 150), range(150, 210)] 11 | 12 | # parameters for output 13 | DEFAULT_RESOLUTION = 480 14 | 15 | # define "Item" for general storage 16 | class Item(object): 17 | def __init__(self, name, start, end, velocity, pitch): 18 | self.name = name 19 | self.start = start 20 | self.end = end 21 | self.velocity = velocity 22 | self.pitch = pitch 23 | 24 | def __repr__(self): 25 | return 'Item(name={}, start={}, end={}, velocity={}, pitch={})'.format( 26 | self.name, self.start, self.end, self.velocity, self.pitch) 27 | 28 | # read notes and tempo changes from midi (assume there is only one track) 29 | def read_items(file_path): 30 | midi_obj = miditoolkit.midi.parser.MidiFile(file_path) 31 | # note 32 | note_items = [] 33 | notes = midi_obj.instruments[0].notes 34 | notes.sort(key=lambda x: (x.start, x.pitch)) 35 | for note in notes: 36 | note_items.append(Item( 37 | name='Note', 38 | start=note.start, 39 | end=note.end, 40 | velocity=note.velocity, 41 | pitch=note.pitch)) 42 | note_items.sort(key=lambda x: x.start) 43 | # tempo 44 | tempo_items = [] 45 | for tempo in midi_obj.tempo_changes: 46 | tempo_items.append(Item( 47 | name='Tempo', 48 | start=tempo.time, 49 | end=None, 50 | velocity=None, 51 | pitch=int(tempo.tempo))) 52 | tempo_items.sort(key=lambda x: x.start) 53 | # expand to all beat 54 | max_tick = tempo_items[-1].start 55 | existing_ticks = {item.start: item.pitch for item in tempo_items} 56 | wanted_ticks = np.arange(0, max_tick+1, DEFAULT_RESOLUTION) 57 | output = [] 58 | for tick in wanted_ticks: 59 | if tick in existing_ticks: 60 | output.append(Item( 61 | name='Tempo', 62 | start=tick, 63 | end=None, 64 | velocity=None, 65 | pitch=existing_ticks[tick])) 66 | else: 67 | output.append(Item( 68 | name='Tempo', 69 | start=tick, 70 | end=None, 71 | velocity=None, 72 | pitch=output[-1].pitch)) 73 | tempo_items = output 74 | return note_items, tempo_items 75 | 76 | # quantize items 77 | def quantize_items(items, ticks=120): 78 | # grid 79 | grids = np.arange(0, items[-1].start, ticks, dtype=int) 80 | # process 81 | for item in items: 82 | index = np.argmin(abs(grids - item.start)) 83 | shift = grids[index] - item.start 84 | item.start += shift 85 | item.end += shift 86 | return items 87 | 88 | # extract chord 89 | def extract_chords(items): 90 | method = chord_recognition.MIDIChord() 91 | chords = method.extract(notes=items) 92 | output = [] 93 | for chord in chords: 94 | output.append(Item( 95 | name='Chord', 96 | start=chord[0], 97 | end=chord[1], 98 | velocity=None, 99 | pitch=chord[2].split('/')[0])) 100 | return output 101 | 102 | # group items 103 | def group_items(items, max_time, ticks_per_bar=DEFAULT_RESOLUTION*4): 104 | items.sort(key=lambda x: x.start) 105 | downbeats = np.arange(0, max_time+ticks_per_bar, ticks_per_bar) 106 | groups = [] 107 | for db1, db2 in zip(downbeats[:-1], downbeats[1:]): 108 | insiders = [] 109 | for item in items: 110 | if (item.start >= db1) and (item.start < db2): 111 | insiders.append(item) 112 | overall = [db1] + insiders + [db2] 113 | groups.append(overall) 114 | return groups 115 | 116 | # define "Event" for event storage 117 | class Event(object): 118 | def __init__(self, name, time, value, text): 119 | self.name = name 120 | self.time = time 121 | self.value = value 122 | self.text = text 123 | 124 | def __repr__(self): 125 | return 'Event(name={}, time={}, value={}, text={})'.format( 126 | self.name, self.time, self.value, self.text) 127 | 128 | # item to event 129 | def item2event(groups): 130 | events = [] 131 | n_downbeat = 0 132 | for i in range(len(groups)): 133 | if 'Note' not in [item.name for item in groups[i][1:-1]]: 134 | continue 135 | bar_st, bar_et = groups[i][0], groups[i][-1] 136 | n_downbeat += 1 137 | events.append(Event( 138 | name='Bar', 139 | time=None, 140 | value=None, 141 | text='{}'.format(n_downbeat))) 142 | for item in groups[i][1:-1]: 143 | # position 144 | flags = np.linspace(bar_st, bar_et, DEFAULT_FRACTION, endpoint=False) 145 | index = np.argmin(abs(flags-item.start)) 146 | events.append(Event( 147 | name='Position', 148 | time=item.start, 149 | value='{}/{}'.format(index+1, DEFAULT_FRACTION), 150 | text='{}'.format(item.start))) 151 | if item.name == 'Note': 152 | # velocity 153 | velocity_index = np.searchsorted( 154 | DEFAULT_VELOCITY_BINS, 155 | item.velocity, 156 | side='right') - 1 157 | events.append(Event( 158 | name='Note Velocity', 159 | time=item.start, 160 | value=velocity_index, 161 | text='{}/{}'.format(item.velocity, DEFAULT_VELOCITY_BINS[velocity_index]))) 162 | # pitch 163 | events.append(Event( 164 | name='Note On', 165 | time=item.start, 166 | value=item.pitch, 167 | text='{}'.format(item.pitch))) 168 | # duration 169 | duration = item.end - item.start 170 | index = np.argmin(abs(DEFAULT_DURATION_BINS-duration)) 171 | events.append(Event( 172 | name='Note Duration', 173 | time=item.start, 174 | value=index, 175 | text='{}/{}'.format(duration, DEFAULT_DURATION_BINS[index]))) 176 | elif item.name == 'Chord': 177 | events.append(Event( 178 | name='Chord', 179 | time=item.start, 180 | value=item.pitch, 181 | text='{}'.format(item.pitch))) 182 | elif item.name == 'Tempo': 183 | tempo = item.pitch 184 | if tempo in DEFAULT_TEMPO_INTERVALS[0]: 185 | tempo_style = Event('Tempo Class', item.start, 'slow', None) 186 | tempo_value = Event('Tempo Value', item.start, 187 | tempo-DEFAULT_TEMPO_INTERVALS[0].start, None) 188 | elif tempo in DEFAULT_TEMPO_INTERVALS[1]: 189 | tempo_style = Event('Tempo Class', item.start, 'mid', None) 190 | tempo_value = Event('Tempo Value', item.start, 191 | tempo-DEFAULT_TEMPO_INTERVALS[1].start, None) 192 | elif tempo in DEFAULT_TEMPO_INTERVALS[2]: 193 | tempo_style = Event('Tempo Class', item.start, 'fast', None) 194 | tempo_value = Event('Tempo Value', item.start, 195 | tempo-DEFAULT_TEMPO_INTERVALS[2].start, None) 196 | elif tempo < DEFAULT_TEMPO_INTERVALS[0].start: 197 | tempo_style = Event('Tempo Class', item.start, 'slow', None) 198 | tempo_value = Event('Tempo Value', item.start, 0, None) 199 | elif tempo > DEFAULT_TEMPO_INTERVALS[2].stop: 200 | tempo_style = Event('Tempo Class', item.start, 'fast', None) 201 | tempo_value = Event('Tempo Value', item.start, 59, None) 202 | events.append(tempo_style) 203 | events.append(tempo_value) 204 | return events 205 | 206 | ############################################################################################# 207 | # WRITE MIDI 208 | ############################################################################################# 209 | def word_to_event(words, word2event): 210 | events = [] 211 | for word in words: 212 | event_name, event_value = word2event.get(word).split('_') 213 | events.append(Event(event_name, None, event_value, None)) 214 | return events 215 | 216 | def write_midi(words, word2event, output_path, prompt_path=None): 217 | events = word_to_event(words, word2event) 218 | # get downbeat and note (no time) 219 | temp_notes = [] 220 | temp_chords = [] 221 | temp_tempos = [] 222 | for i in range(len(events)-3): 223 | if events[i].name == 'Bar' and i > 0: 224 | temp_notes.append('Bar') 225 | temp_chords.append('Bar') 226 | temp_tempos.append('Bar') 227 | elif events[i].name == 'Position' and \ 228 | events[i+1].name == 'Note Velocity' and \ 229 | events[i+2].name == 'Note On' and \ 230 | events[i+3].name == 'Note Duration': 231 | # start time and end time from position 232 | position = int(events[i].value.split('/')[0]) - 1 233 | # velocity 234 | index = int(events[i+1].value) 235 | velocity = int(DEFAULT_VELOCITY_BINS[index]) 236 | # pitch 237 | pitch = int(events[i+2].value) 238 | # duration 239 | index = int(events[i+3].value) 240 | duration = DEFAULT_DURATION_BINS[index] 241 | # adding 242 | temp_notes.append([position, velocity, pitch, duration]) 243 | elif events[i].name == 'Position' and events[i+1].name == 'Chord': 244 | position = int(events[i].value.split('/')[0]) - 1 245 | temp_chords.append([position, events[i+1].value]) 246 | elif events[i].name == 'Position' and \ 247 | events[i+1].name == 'Tempo Class' and \ 248 | events[i+2].name == 'Tempo Value': 249 | position = int(events[i].value.split('/')[0]) - 1 250 | if events[i+1].value == 'slow': 251 | tempo = DEFAULT_TEMPO_INTERVALS[0].start + int(events[i+2].value) 252 | elif events[i+1].value == 'mid': 253 | tempo = DEFAULT_TEMPO_INTERVALS[1].start + int(events[i+2].value) 254 | elif events[i+1].value == 'fast': 255 | tempo = DEFAULT_TEMPO_INTERVALS[2].start + int(events[i+2].value) 256 | temp_tempos.append([position, tempo]) 257 | # get specific time for notes 258 | ticks_per_beat = DEFAULT_RESOLUTION 259 | ticks_per_bar = DEFAULT_RESOLUTION * 4 # assume 4/4 260 | notes = [] 261 | current_bar = 0 262 | for note in temp_notes: 263 | if note == 'Bar': 264 | current_bar += 1 265 | else: 266 | position, velocity, pitch, duration = note 267 | # position (start time) 268 | current_bar_st = current_bar * ticks_per_bar 269 | current_bar_et = (current_bar + 1) * ticks_per_bar 270 | flags = np.linspace(current_bar_st, current_bar_et, DEFAULT_FRACTION, endpoint=False, dtype=int) 271 | st = flags[position] 272 | # duration (end time) 273 | et = st + duration 274 | notes.append(miditoolkit.Note(velocity, pitch, st, et)) 275 | # get specific time for chords 276 | if len(temp_chords) > 0: 277 | chords = [] 278 | current_bar = 0 279 | for chord in temp_chords: 280 | if chord == 'Bar': 281 | current_bar += 1 282 | else: 283 | position, value = chord 284 | # position (start time) 285 | current_bar_st = current_bar * ticks_per_bar 286 | current_bar_et = (current_bar + 1) * ticks_per_bar 287 | flags = np.linspace(current_bar_st, current_bar_et, DEFAULT_FRACTION, endpoint=False, dtype=int) 288 | st = flags[position] 289 | chords.append([st, value]) 290 | # get specific time for tempos 291 | tempos = [] 292 | current_bar = 0 293 | for tempo in temp_tempos: 294 | if tempo == 'Bar': 295 | current_bar += 1 296 | else: 297 | position, value = tempo 298 | # position (start time) 299 | current_bar_st = current_bar * ticks_per_bar 300 | current_bar_et = (current_bar + 1) * ticks_per_bar 301 | flags = np.linspace(current_bar_st, current_bar_et, DEFAULT_FRACTION, endpoint=False, dtype=int) 302 | st = flags[position] 303 | tempos.append([int(st), value]) 304 | # write 305 | if prompt_path: 306 | midi = miditoolkit.midi.parser.MidiFile(prompt_path) 307 | # 308 | last_time = DEFAULT_RESOLUTION * 4 * 4 309 | # note shift 310 | for note in notes: 311 | note.start += last_time 312 | note.end += last_time 313 | midi.instruments[0].notes.extend(notes) 314 | # tempo changes 315 | temp_tempos = [] 316 | for tempo in midi.tempo_changes: 317 | if tempo.time < DEFAULT_RESOLUTION*4*4: 318 | temp_tempos.append(tempo) 319 | else: 320 | break 321 | for st, bpm in tempos: 322 | st += last_time 323 | temp_tempos.append(miditoolkit.midi.containers.TempoChange(bpm, st)) 324 | midi.tempo_changes = temp_tempos 325 | # write chord into marker 326 | if len(temp_chords) > 0: 327 | for c in chords: 328 | midi.markers.append( 329 | miditoolkit.midi.containers.Marker(text=c[1], time=c[0]+last_time)) 330 | else: 331 | midi = miditoolkit.midi.parser.MidiFile() 332 | midi.ticks_per_beat = DEFAULT_RESOLUTION 333 | # write instrument 334 | inst = miditoolkit.midi.containers.Instrument(0, is_drum=False) 335 | inst.notes = notes 336 | midi.instruments.append(inst) 337 | # write tempo 338 | tempo_changes = [] 339 | for st, bpm in tempos: 340 | tempo_changes.append(miditoolkit.midi.containers.TempoChange(bpm, st)) 341 | midi.tempo_changes = tempo_changes 342 | # write chord into marker 343 | if len(temp_chords) > 0: 344 | for c in chords: 345 | midi.markers.append( 346 | miditoolkit.midi.containers.Marker(text=c[1], time=c[0])) 347 | # write 348 | midi.dump(output_path) 349 | --------------------------------------------------------------------------------