├── LICENSE ├── README.md ├── __init__.py ├── evaluate.py ├── log └── __init__.py ├── models.py ├── quantization ├── __init__.py ├── lsqplus_quantize_V1.py ├── lsqplus_quantize_V2.py ├── lsqquantize_V1.py └── lsqquantize_V2.py ├── seebnparam.py └── trains.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # LSQ and LSQ+
2 | LSQ+ net or LSQplus net and LSQ net
3 | 4 | ## commit log
5 | ` 6 | 2023-01-08 7 | ` 8 | Dorefa and Pact, [https://github.com/ZouJiu1/Dorefa_Pact](https://github.com/ZouJiu1/Dorefa_Pact)
9 | --------------------------------------------------------------------------------------------------------------
10 | add torch.nn.Parameter .data, retrain models 18-01-2022
11 | 12 | I'm not the author, I just complish an unofficial implementation of LSQ+ or LSQplus and LSQ，the origin paper you can find LSQ+ here [arxiv.org/abs/2004.09576](https://arxiv.org/abs/2004.09576) and LSQ here [arxiv.org/abs/1902.08153](https://arxiv.org/abs/1902.08153).
13 | 14 | pytorch==1.8.1
15 | 16 | You should train 32-bit float model firstly, then you can finetune a low bit-width quantization QAT model by loading the trained 32-bit float model
17 | 18 | Dataset used for training is CIFAR10 and model used is Resnet18 revised
19 | 20 | ## Version introduction 21 | lsqplus_quantize_V1.py: initialize s、beta of activation quantization according to LSQ+ [LSQ+: Improving low-bit quantization through learnable offsets and better initialization](https://arxiv.org/abs/2004.09576)

22 | lsqplus_quantize_V2.py: initialize s、beta of activation quantization according to min max values

23 | lsqquantize_V1.py：initialize s of activation quantization according to LSQ [Learned Step Size Quantization](https://arxiv.org/abs/1902.08153)

24 | lsqquantize_V2.py: initialize s of activation quantization = 1

25 | lsqplus_quantize_V2.py has the best result when use cifar10 dataset
26 | 27 | ## The Train Results 28 | ### For the below table all set a_bit=8, w_bit=8 29 | | version | weight per_channel | learning rate | A s initial | A beta initial | best epoch | Accuracy | models 30 | | ------ | --------- | ------ | ------ | ------ | ------ | ------ | ------ | 31 | | Float 32bit | - | <=66 0.1
<=86 0.01
<=99 0.001
<=112 0.0001 | - | - | 112 | 92.6 | [https://www.aliyundrive.com/s/6B2AZ45fFjx](https://www.aliyundrive.com/s/6B2AZ45fFjx) | 32 | | lsqplus_quantize_V1 | × | <=31 0.1
<=61 0.01
<=81 0.001
<112 0.0001 | 1 | -1e-9 | 90 | 90.3 | [https://www.aliyundrive.com/s/FNZRhoTe8uW](https://www.aliyundrive.com/s/FNZRhoTe8uW) | 33 | | lsqplus_quantize_V2 | × | as before | - | - | 87 | 92.8 | [https://www.aliyundrive.com/s/WDH3ZnEa7vy](https://www.aliyundrive.com/s/WDH3ZnEa7vy) | 34 | | lsqplus_quantize_V1 | ✔ | as before | - | - | 96 | 91.19 | [https://www.aliyundrive.com/s/JATsi4vdurp](https://www.aliyundrive.com/s/JATsi4vdurp) | 35 | | lsqplus_quantize_V2 | ✔ | as before | - | - | 69 | 92.8 | [https://www.aliyundrive.com/s/LRWHaBLQGWc](https://www.aliyundrive.com/s/LRWHaBLQGWc) | 36 | | lsqquantize_V1 | × | as before | - | - | 102 | 91.89 | [https://www.aliyundrive.com/s/nR1KZZRuB23](https://www.aliyundrive.com/s/nR1KZZRuB23) | 37 | | lsqquantize_V2 | × | as before | - | - | 69 | 91.82 | [https://www.aliyundrive.com/s/7fjmViqUvh4](https://www.aliyundrive.com/s/7fjmViqUvh4) | 38 | | lsqquantize_V1 | ✔ | as before | - | - | 108 | 91.29 | [https://www.aliyundrive.com/s/](https://www.aliyundrive.com/s/PX84qGorVxY) | 39 | | lsqquantize_V2 | ✔ | as before | - | - | 72 | 91.72 | [https://www.aliyundrive.com/s/7nGvMVZcKp7](https://www.aliyundrive.com/s/7nGvMVZcKp7) | 40 |
41 | all 42 | 43 | [https://www.aliyundrive.com/s/hng9XsvhYru](https://www.aliyundrive.com/s/hng9XsvhYru) 44 | 45 |
46 | A represent activation, I use moving average method to initialize s and beta.

47 | 48 | LEARNED STEP SIZE QUANTIZATION
49 | LSQ+: Improving low-bit quantization through learnable offsets and better initialization
50 | 51 | ### References
52 | https://github.com/666DZY666/micronet
53 | https://github.com/hustzxd/LSQuantization
54 | https://github.com/zhutmost/lsq-net
55 | https://github.com/Zhen-Dong/HAWQ
56 | https://github.com/KwangHoonAn/PACT
57 | https://github.com/Jermmy/pytorch-quantization-demo
58 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | 2 | 3 | -------------------------------------------------------------------------------- /evaluate.py: -------------------------------------------------------------------------------- 1 | #encoding=utf-8 2 | #Author: ZouJiu 3 | #Time: 2021-11-13 4 | 5 | import numpy as np 6 | import torch 7 | import os 8 | import time 9 | import torch 10 | import torchvision 11 | import torchvision.transforms as transforms 12 | from torch.utils.data import Dataset, DataLoader 13 | # from load_datas import TF, trainDataset, collate_fn 14 | import models #, resnet50 15 | from quantization.lsqquantize_V1 import prepare as lsqprepareV1 16 | from quantization.lsqquantize_V2 import prepare as lsqprepareV2 17 | from quantization.lsqplus_quantize_V1 import prepare as lsqplusprepareV1 18 | from quantization.lsqplus_quantize_V2 import prepare as lsqplusprepareV2 19 | from quantization.lsqplus_quantize_V1 import update_LSQplus_activation_Scalebeta 20 | import torch.optim as optim 21 | import datetime 22 | os.environ["CUDA_VISIBLE_DEVICES"] = '1' 23 | 24 | def adjust_lr(optimizer, stepiters, epoch): 25 | if epoch < 135: 26 | lr = 0.1 27 | elif epoch < 185: 28 | lr = 0.01 29 | elif epoch < 290: 30 | lr = 0.001 31 | else: 32 | import sys 33 | sys.exit(0) 34 | for param_group in optimizer.param_groups: 35 | param_group['lr'] = lr 36 | 37 | def evaluate(): 38 | config = {'a_bit':8, 'w_bit':8, "all_positive":False, "per_channel":True, 39 | "num_classes":10,"batch_init":20} 40 | pretrainedmodel = r'C:\Users\10696\Desktop\QAT\lsq+\log\model_108_42510_0.003_92.528_2021-11-27_17-49-47.pth' 41 | Resnet_pretrain = False #test 42 | batch_size = 128 43 | num_epochs = 290 44 | Floatmodel = True #QAT or float-32 train 45 | LSQplus = True #LSQ+ or LSQ 46 | scratch = True #从最开始训练，不是finetuning， 若=False就是finetuning 47 | tim = datetime.datetime.strftime(datetime.datetime.now(),"%Y-%m-%d %H-%M-%S").replace(' ', '_') 48 | 49 | test_transform = transforms.Compose([ 50 | # transforms.Resize((32, 32)), 51 | transforms.ToTensor(), 52 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))]) 53 | 54 | batch_size = 128 #Accuracy all is: 73.4 55 | 56 | testset = torchvision.datasets.CIFAR10(root='datas', train=False, 57 | download=True, transform=test_transform) 58 | testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, 59 | shuffle=False, num_workers=2, drop_last=True) 60 | 61 | classes = ('plane', 'car', 'bird', 'cat', 62 | 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') 63 | device = "cuda" if torch.cuda.is_available() else "cpu" 64 | 65 | model = models.resnet18(pretrained = Resnet_pretrain, num_classes=config['num_classes']) 66 | 67 | #LSQ+ 68 | if LSQplus and not Floatmodel: 69 | lsqplusprepare(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 70 | all_positive=config["all_positive"], per_channel=config["per_channel"], 71 | batch_init = config["batch_init"]) 72 | elif not LSQplus and not Floatmodel: 73 | #LSQ 74 | lsqprepare(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 75 | all_positive=config["all_positive"], per_channel=config["per_channel"], 76 | batch_init = config["batch_init"]) 77 | elif Floatmodel: 78 | pass 79 | 80 | if not Floatmodel: 81 | print(model) 82 | if not os.path.exists(pretrainedmodel): 83 | print('the pretrainedmodel do not exists %s'%pretrainedmodel) 84 | if pretrainedmodel and os.path.exists(pretrainedmodel): 85 | print('loading pretrained model: ', pretrainedmodel) 86 | if torch.cuda.is_available(): 87 | state_dict = torch.load(pretrainedmodel, map_location='cuda') 88 | else: 89 | state_dict = torch.load(pretrainedmodel, map_location='cpu') 90 | model.load_state_dict(state_dict['state_dict']) 91 | if not scratch: 92 | iteration = state_dict['iteration'] 93 | alliters = state_dict['alliters'] 94 | nowepoch = state_dict['nowepoch'] 95 | else: 96 | iteration = 0 97 | alliters = 0 98 | nowepoch = 0 99 | print('loading complete') 100 | else: 101 | print('no pretrained model') 102 | iteration = 0 103 | alliters = 0 104 | nowepoch = 0 105 | model = model.to(device) 106 | 107 | print('validation of testes') 108 | # prepare to count predictions for each class 109 | correct_pred = {classname: 0 for classname in classes} 110 | total_pred = {classname: 0 for classname in classes} 111 | 112 | model.eval() 113 | # again no gradients needed 114 | with torch.no_grad(): 115 | for data in testloader: 116 | images, labels = data 117 | images = images.to(device) 118 | labels = labels.to(device) 119 | outputs = model(images) 120 | _, predictions = torch.max(outputs, 1) 121 | # collect the correct predictions for each class 122 | for label, prediction in zip(labels, predictions): 123 | if label == prediction: 124 | correct_pred[classes[label]] += 1 125 | total_pred[classes[label]] += 1 126 | 127 | 128 | # print accuracy for each class 129 | correctall = 0 130 | alltest = 0 131 | for classname, correct_count in correct_pred.items(): 132 | accuracy = 100 * float(correct_count) / total_pred[classname] 133 | print("Accuracy for class {:5s} is: {:.1f} %".format(classname, 134 | accuracy)) 135 | correctall += correct_count 136 | alltest += total_pred[classname] 137 | print("Accuracy all is: {:.1f}".format(100 * float(correctall)/alltest)) 138 | 139 | 140 | if __name__ == '__main__': 141 | evaluate() 142 | -------------------------------------------------------------------------------- /log/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZouJiu1/LSQplus/2076e86479491f0e68ada31d36948596a1ee24f9/log/__init__.py -------------------------------------------------------------------------------- /models.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torch import Tensor 3 | import torch.nn as nn 4 | from torch.nn import functional as F 5 | from torch.hub import load_state_dict_from_url 6 | from typing import Type, Any, Callable, Union, List, Optional 7 | 8 | 9 | __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 10 | 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 11 | 'wide_resnet50_2', 'wide_resnet101_2'] 12 | 13 | 14 | model_urls = { 15 | 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 16 | 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 17 | 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 18 | 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 19 | 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', 20 | 'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth', 21 | 'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth', 22 | 'wide_resnet50_2': 'https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth', 23 | 'wide_resnet101_2': 'https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth', 24 | } 25 | 26 | 27 | def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d: 28 | """3x3 convolution with padding""" 29 | return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, 30 | padding=dilation, groups=groups, bias=False, dilation=dilation) 31 | 32 | 33 | def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d: 34 | """1x1 convolution""" 35 | return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) 36 | 37 | 38 | class BasicBlock(nn.Module): 39 | expansion: int = 1 40 | 41 | def __init__( 42 | self, 43 | inplanes: int, 44 | planes: int, 45 | stride: int = 1, 46 | downsample: Optional[nn.Module] = None, 47 | groups: int = 1, 48 | base_width: int = 64, 49 | dilation: int = 1, 50 | norm_layer: Optional[Callable[..., nn.Module]] = None 51 | ) -> None: 52 | super(BasicBlock, self).__init__() 53 | if norm_layer is None: 54 | norm_layer = nn.BatchNorm2d 55 | if groups != 1 or base_width != 64: 56 | raise ValueError('BasicBlock only supports groups=1 and base_width=64') 57 | if dilation > 1: 58 | raise NotImplementedError("Dilation > 1 not supported in BasicBlock") 59 | # Both self.conv1 and self.downsample layers downsample the input when stride != 1 60 | self.conv1 = conv3x3(inplanes, planes, stride) 61 | self.bn1 = norm_layer(planes) 62 | self.relu = nn.ReLU(inplace=True) 63 | self.conv2 = conv3x3(planes, planes) 64 | self.bn2 = norm_layer(planes) 65 | self.downsample = downsample 66 | self.stride = stride 67 | 68 | def forward(self, x: Tensor) -> Tensor: 69 | identity = x 70 | out = self.conv1(x) 71 | out = self.bn1(out) 72 | out = self.relu(out) 73 | 74 | out = self.conv2(out) 75 | out = self.bn2(out) 76 | 77 | if self.downsample is not None: 78 | identity = self.downsample(x) 79 | 80 | out += identity 81 | out = self.relu(out) 82 | 83 | return out 84 | 85 | 86 | class Bottleneck(nn.Module): 87 | # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2) 88 | # while original implementation places the stride at the first 1x1 convolution(self.conv1) 89 | # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385. 90 | # This variant is also known as ResNet V1.5 and improves accuracy according to 91 | # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch. 92 | 93 | expansion: int = 4 94 | 95 | def __init__( 96 | self, 97 | inplanes: int, 98 | planes: int, 99 | stride: int = 1, 100 | downsample: Optional[nn.Module] = None, 101 | groups: int = 1, 102 | base_width: int = 64, 103 | dilation: int = 1, 104 | norm_layer: Optional[Callable[..., nn.Module]] = None 105 | ) -> None: 106 | super(Bottleneck, self).__init__() 107 | if norm_layer is None: 108 | norm_layer = nn.BatchNorm2d 109 | width = int(planes * (base_width / 64.)) * groups 110 | # Both self.conv2 and self.downsample layers downsample the input when stride != 1 111 | self.conv1 = conv1x1(inplanes, width) 112 | self.bn1 = norm_layer(width) 113 | self.conv2 = conv3x3(width, width, stride, groups, dilation) 114 | self.bn2 = norm_layer(width) 115 | self.conv3 = conv1x1(width, planes * self.expansion) 116 | self.bn3 = norm_layer(planes * self.expansion) 117 | self.relu = nn.ReLU(inplace=True) 118 | self.downsample = downsample 119 | self.stride = stride 120 | 121 | def forward(self, x: Tensor) -> Tensor: 122 | identity = x 123 | 124 | out = self.conv1(x) 125 | out = self.bn1(out) 126 | out = self.relu(out) 127 | 128 | out = self.conv2(out) 129 | out = self.bn2(out) 130 | out = self.relu(out) 131 | 132 | out = self.conv3(out) 133 | out = self.bn3(out) 134 | 135 | if self.downsample is not None: 136 | identity = self.downsample(x) 137 | 138 | out += identity 139 | out = self.relu(out) 140 | 141 | return out 142 | 143 | 144 | class ResNet(nn.Module): 145 | 146 | def __init__( 147 | self, 148 | block: Type[Union[BasicBlock, Bottleneck]], 149 | layers: List[int], 150 | num_classes: int = 1000, 151 | zero_init_residual: bool = False, 152 | groups: int = 1, 153 | width_per_group: int = 64, 154 | replace_stride_with_dilation: Optional[List[bool]] = None, 155 | norm_layer: Optional[Callable[..., nn.Module]] = None 156 | ) -> None: 157 | super(ResNet, self).__init__() 158 | if norm_layer is None: 159 | norm_layer = nn.BatchNorm2d 160 | self._norm_layer = norm_layer 161 | 162 | self.inplanes = 64 163 | self.dilation = 1 164 | if replace_stride_with_dilation is None: 165 | # each element in the tuple indicates if we should replace 166 | # the 2x2 stride with a dilated convolution instead 167 | replace_stride_with_dilation = [False, False, False] 168 | if len(replace_stride_with_dilation) != 3: 169 | raise ValueError("replace_stride_with_dilation should be None " 170 | "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) 171 | self.groups = groups 172 | self.base_width = width_per_group 173 | self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, 174 | bias=False) 175 | self.bn1 = norm_layer(self.inplanes) 176 | self.relu = nn.ReLU(inplace=True) 177 | self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 178 | self.layer1 = self._make_layer(block, 64, layers[0]) 179 | self.layer2 = self._make_layer(block, 128, layers[1], stride=2, 180 | dilate=replace_stride_with_dilation[0]) 181 | self.layer3 = self._make_layer(block, 256, layers[2], stride=2, 182 | dilate=replace_stride_with_dilation[1]) 183 | self.layer4 = self._make_layer(block, 512, layers[3], stride=2, 184 | dilate=replace_stride_with_dilation[2]) 185 | # self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) 186 | self.poolavg = nn.AvgPool2d(4) 187 | self.fc = nn.Linear(512 * block.expansion, num_classes) 188 | 189 | for m in self.modules(): 190 | if isinstance(m, nn.Conv2d): 191 | nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 192 | elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): 193 | nn.init.constant_(m.weight, 1) 194 | nn.init.constant_(m.bias, 0) 195 | 196 | # Zero-initialize the last BN in each residual branch, 197 | # so that the residual branch starts with zeros, and each residual block behaves like an identity. 198 | # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 199 | if zero_init_residual: 200 | for m in self.modules(): 201 | if isinstance(m, Bottleneck): 202 | nn.init.constant_(m.bn3.weight, 0) # type: ignore[arg-type] 203 | elif isinstance(m, BasicBlock): 204 | nn.init.constant_(m.bn2.weight, 0) # type: ignore[arg-type] 205 | 206 | def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int, 207 | stride: int = 1, dilate: bool = False) -> nn.Sequential: 208 | norm_layer = self._norm_layer 209 | downsample = None 210 | previous_dilation = self.dilation 211 | if dilate: 212 | self.dilation *= stride 213 | stride = 1 214 | if stride != 1 or self.inplanes != planes * block.expansion: 215 | downsample = nn.Sequential( 216 | conv1x1(self.inplanes, planes * block.expansion, stride), 217 | norm_layer(planes * block.expansion), 218 | ) 219 | 220 | layers = [] 221 | layers.append(block(self.inplanes, planes, stride, downsample, self.groups, 222 | self.base_width, previous_dilation, norm_layer)) 223 | self.inplanes = planes * block.expansion 224 | for _ in range(1, blocks): 225 | layers.append(block(self.inplanes, planes, groups=self.groups, 226 | base_width=self.base_width, dilation=self.dilation, 227 | norm_layer=norm_layer)) 228 | 229 | return nn.Sequential(*layers) 230 | 231 | def _forward_impl(self, x: Tensor) -> Tensor: 232 | # See note [TorchScript super()] 233 | x = self.conv1(x) 234 | x = self.bn1(x) 235 | x = self.relu(x) 236 | # x = self.maxpool(x) 237 | 238 | x = self.layer1(x) 239 | x = self.layer2(x) 240 | x = self.layer3(x) 241 | x = self.layer4(x) 242 | 243 | # x = self.avgpool(x) 244 | # x = torch.flatten(x, 1) 245 | 246 | x = self.poolavg(x) 247 | x = x.view(x.size()[0], -1) 248 | x = self.fc(x) 249 | 250 | return x 251 | 252 | def forward(self, x: Tensor) -> Tensor: 253 | return self._forward_impl(x) 254 | 255 | 256 | def _resnet( 257 | arch: str, 258 | block: Type[Union[BasicBlock, Bottleneck]], 259 | layers: List[int], 260 | pretrained: bool, 261 | progress: bool, 262 | **kwargs: Any 263 | ) -> ResNet: 264 | model = ResNet(block, layers, **kwargs) 265 | # if pretrained: 266 | # state_dict = load_state_dict_from_url(model_urls[arch], 267 | # progress=progress) 268 | # delete = [] 269 | # for key, value in state_dict.items(): 270 | # if 'fc' in key: 271 | # delete.append(key) 272 | # print('detele: ', delete) 273 | # for key in delete: 274 | # state_dict.pop(key) 275 | # model.load_state_dict(state_dict, strict=False) 276 | return model 277 | 278 | 279 | def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 280 | r"""ResNet-18 model from 281 | `"Deep Residual Learning for Image Recognition" `_. 282 | 283 | Args: 284 | pretrained (bool): If True, returns a model pre-trained on ImageNet 285 | progress (bool): If True, displays a progress bar of the download to stderr 286 | """ 287 | return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress, 288 | **kwargs) 289 | 290 | 291 | def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 292 | r"""ResNet-34 model from 293 | `"Deep Residual Learning for Image Recognition" `_. 294 | 295 | Args: 296 | pretrained (bool): If True, returns a model pre-trained on ImageNet 297 | progress (bool): If True, displays a progress bar of the download to stderr 298 | """ 299 | return _resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress, 300 | **kwargs) 301 | 302 | 303 | def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 304 | r"""ResNet-50 model from 305 | `"Deep Residual Learning for Image Recognition" `_. 306 | 307 | Args: 308 | pretrained (bool): If True, returns a model pre-trained on ImageNet 309 | progress (bool): If True, displays a progress bar of the download to stderr 310 | """ 311 | return _resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress, 312 | **kwargs) 313 | 314 | 315 | def resnet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 316 | r"""ResNet-101 model from 317 | `"Deep Residual Learning for Image Recognition" `_. 318 | 319 | Args: 320 | pretrained (bool): If True, returns a model pre-trained on ImageNet 321 | progress (bool): If True, displays a progress bar of the download to stderr 322 | """ 323 | return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress, 324 | **kwargs) 325 | 326 | 327 | def resnet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 328 | r"""ResNet-152 model from 329 | `"Deep Residual Learning for Image Recognition" `_. 330 | 331 | Args: 332 | pretrained (bool): If True, returns a model pre-trained on ImageNet 333 | progress (bool): If True, displays a progress bar of the download to stderr 334 | """ 335 | return _resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress, 336 | **kwargs) 337 | 338 | 339 | def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 340 | r"""ResNeXt-50 32x4d model from 341 | `"Aggregated Residual Transformation for Deep Neural Networks" `_. 342 | 343 | Args: 344 | pretrained (bool): If True, returns a model pre-trained on ImageNet 345 | progress (bool): If True, displays a progress bar of the download to stderr 346 | """ 347 | kwargs['groups'] = 32 348 | kwargs['width_per_group'] = 4 349 | return _resnet('resnext50_32x4d', Bottleneck, [3, 4, 6, 3], 350 | pretrained, progress, **kwargs) 351 | 352 | 353 | def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 354 | r"""ResNeXt-101 32x8d model from 355 | `"Aggregated Residual Transformation for Deep Neural Networks" `_. 356 | 357 | Args: 358 | pretrained (bool): If True, returns a model pre-trained on ImageNet 359 | progress (bool): If True, displays a progress bar of the download to stderr 360 | """ 361 | kwargs['groups'] = 32 362 | kwargs['width_per_group'] = 8 363 | return _resnet('resnext101_32x8d', Bottleneck, [3, 4, 23, 3], 364 | pretrained, progress, **kwargs) 365 | 366 | 367 | def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 368 | r"""Wide ResNet-50-2 model from 369 | `"Wide Residual Networks" `_. 370 | 371 | The model is the same as ResNet except for the bottleneck number of channels 372 | which is twice larger in every block. The number of channels in outer 1x1 373 | convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 374 | channels, and in Wide ResNet-50-2 has 2048-1024-2048. 375 | 376 | Args: 377 | pretrained (bool): If True, returns a model pre-trained on ImageNet 378 | progress (bool): If True, displays a progress bar of the download to stderr 379 | """ 380 | kwargs['width_per_group'] = 64 * 2 381 | return _resnet('wide_resnet50_2', Bottleneck, [3, 4, 6, 3], 382 | pretrained, progress, **kwargs) 383 | 384 | 385 | def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 386 | r"""Wide ResNet-101-2 model from 387 | `"Wide Residual Networks" `_. 388 | 389 | The model is the same as ResNet except for the bottleneck number of channels 390 | which is twice larger in every block. The number of channels in outer 1x1 391 | convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 392 | channels, and in Wide ResNet-50-2 has 2048-1024-2048. 393 | 394 | Args: 395 | pretrained (bool): If True, returns a model pre-trained on ImageNet 396 | progress (bool): If True, displays a progress bar of the download to stderr 397 | """ 398 | kwargs['width_per_group'] = 64 * 2 399 | return _resnet('wide_resnet101_2', Bottleneck, [3, 4, 23, 3], 400 | pretrained, progress, **kwargs) 401 | 402 | def wide_resnet1202(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: 403 | r""" 404 | The model is the same as ResNet except for the bottleneck number of channels 405 | which is twice larger in every block. The number of channels in outer 1x1 406 | convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 407 | channels, and in Wide ResNet-50-2 has 2048-1024-2048. 408 | 409 | Args: 410 | pretrained (bool): If True, returns a model pre-trained on ImageNet 411 | progress (bool): If True, displays a progress bar of the download to stderr 412 | """ 413 | return _resnet('resnet1202', BasicBlock, [200, 200, 200, 200], 414 | pretrained=False, progress=progress, **kwargs) 415 | -------------------------------------------------------------------------------- /quantization/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZouJiu1/LSQplus/2076e86479491f0e68ada31d36948596a1ee24f9/quantization/__init__.py -------------------------------------------------------------------------------- /quantization/lsqplus_quantize_V1.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from torch.autograd import Function 7 | from quantization.lsqquantize_V1 import Round 8 | 9 | class ALSQPlus(Function): 10 | @staticmethod 11 | def forward(ctx, weight, alpha, g, Qn, Qp, beta): 12 | # assert alpha > 0, "alpha={}".format(alpha) 13 | ctx.save_for_backward(weight, alpha, beta) 14 | ctx.other = g, Qn, Qp 15 | w_q = Round.apply(torch.div((weight - beta), alpha).clamp(Qn, Qp)) 16 | w_q = w_q * alpha + beta 17 | return w_q 18 | 19 | @staticmethod 20 | def backward(ctx, grad_weight): 21 | weight, alpha, beta = ctx.saved_tensors 22 | g, Qn, Qp = ctx.other 23 | q_w = (weight - beta) / alpha 24 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 25 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 26 | between = 1.0 - smaller -bigger #得到位于量化区间的index 27 | grad_alpha = ((smaller * Qn + bigger * Qp + 28 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) 29 | grad_beta = ((smaller + bigger) * grad_weight * g).sum().unsqueeze(dim=0) 30 | #在量化区间之外的值都是常数，故导数也是0 31 | grad_weight = between * grad_weight 32 | #返回的梯度要和forward的参数对应起来 33 | return grad_weight, grad_alpha, None, None, None, grad_beta 34 | 35 | class WLSQPlus(Function): 36 | @staticmethod 37 | def forward(ctx, weight, alpha, g, Qn, Qp, per_channel): 38 | # assert alpha > 0, "alpha={}".format(alpha) 39 | ctx.save_for_backward(weight, alpha) 40 | ctx.other = g, Qn, Qp, per_channel 41 | if per_channel: 42 | sizes = weight.size() 43 | weight = weight.contiguous().view(weight.size()[0], -1) 44 | weight = torch.transpose(weight, 0, 1) 45 | alpha = torch.broadcast_to(alpha, weight.size()) 46 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 47 | w_q = w_q * alpha 48 | w_q = torch.transpose(w_q, 0, 1) 49 | w_q = w_q.contiguous().view(sizes) 50 | else: 51 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 52 | w_q = w_q * alpha 53 | return w_q 54 | 55 | @staticmethod 56 | def backward(ctx, grad_weight): 57 | weight, alpha = ctx.saved_tensors 58 | g, Qn, Qp, per_channel = ctx.other 59 | if per_channel: 60 | sizes = weight.size() 61 | weight = weight.contiguous().view(weight.size()[0], -1) 62 | weight = torch.transpose(weight, 0, 1) 63 | alpha = torch.broadcast_to(alpha, weight.size()) 64 | q_w = weight / alpha 65 | q_w = torch.transpose(q_w, 0, 1) 66 | q_w = q_w.contiguous().view(sizes) 67 | else: 68 | q_w = weight / alpha 69 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 70 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 71 | between = 1.0 - smaller -bigger #得到位于量化区间的index 72 | if per_channel: 73 | grad_alpha = ((smaller * Qn + bigger * Qp + 74 | between * Round.apply(q_w) - between * q_w)*grad_weight * g) 75 | grad_alpha = grad_alpha.contiguous().view(grad_alpha.size()[0], -1).sum(dim=1) 76 | else: 77 | grad_alpha = ((smaller * Qn + bigger * Qp + 78 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) 79 | #在量化区间之外的值都是常数，故导数也是0 80 | grad_weight = between * grad_weight 81 | return grad_weight, grad_alpha, None, None, None, None 82 | 83 | def grad_scale(x, scale): 84 | y = x 85 | y_grad = x * scale 86 | return (y - y_grad).detach() + y_grad 87 | 88 | def round_pass(x): 89 | y = x.round() 90 | y_grad = x 91 | return (y - y_grad).detach() + y_grad 92 | 93 | def get_percentile_min_max(input, lower_percentile, uppper_percentile, output_tensor): 94 | batch_size = input.shape[0] 95 | lower_index = round(batch_size * (1 - lower_percentile*0.01)) 96 | upper_index = round(batch_size * (1 - uppper_percentile*0.01)) 97 | 98 | upper_bound = torch.kthvalue(input, k=upper_index).values 99 | 100 | if lower_percentile==0: 101 | lower_bound = upper_bound * 0 102 | else: 103 | low_bound = -torch.kthvalue(-input, k=lower_index).values 104 | 105 | 106 | # A(特征)量化 107 | class LSQPlusActivationQuantizer(nn.Module): 108 | def __init__(self, a_bits, all_positive=False,batch_init = 20): 109 | #activations 没有per-channel这个选项的 110 | super(LSQPlusActivationQuantizer, self).__init__() 111 | self.a_bits = a_bits 112 | self.all_positive = all_positive 113 | self.batch_init = batch_init 114 | if self.all_positive: 115 | # unsigned activation is quantized to [0, 2^b-1] 116 | self.Qn = 0 117 | self.Qp = 2 ** self.a_bits - 1 118 | else: 119 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 120 | self.Qn = - 2 ** (self.a_bits - 1) 121 | self.Qp = 2 ** (self.a_bits - 1) - 1 122 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) 123 | # self.beta = torch.nn.Parameter(torch.tensor([float(0)])) 124 | self.beta = torch.nn.Parameter(torch.tensor([float(-1e-9)]), requires_grad=True) 125 | self.init_state = 0 126 | 127 | # 量化/反量化 128 | def forward(self, activation): 129 | #V1 130 | # print(self.a_bits, self.batch_init) 131 | if self.a_bits == 32: 132 | q_a = activation 133 | elif self.a_bits == 1: 134 | print('！Binary quantization is not supported ！') 135 | assert self.a_bits != 1 136 | else: 137 | if self.init_state==0: 138 | self.g = 1.0/math.sqrt(activation.numel() * self.Qp) 139 | self.init_state += 1 140 | q_a = ALSQPlus.apply(activation, self.s, self.g, self.Qn, self.Qp, self.beta) 141 | # print(self.s, self.beta) 142 | return q_a 143 | 144 | # W(权重)量化 145 | class LSQPlusWeightQuantizer(nn.Module): 146 | def __init__(self, w_bits, all_positive=False, per_channel=False,batch_init = 20): 147 | super(LSQPlusWeightQuantizer, self).__init__() 148 | self.w_bits = w_bits 149 | self.all_positive = all_positive 150 | self.batch_init = batch_init 151 | if self.all_positive: 152 | # unsigned activation is quantized to [0, 2^b-1] 153 | self.Qn = 0 154 | self.Qp = 2 ** w_bits - 1 155 | else: 156 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 157 | self.Qn = - 2 ** (w_bits - 1) 158 | self.Qp = 2 ** (w_bits - 1) - 1 159 | self.per_channel = per_channel 160 | self.init_state = 0 161 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) 162 | # self.beta = torch.nn.Parameter(torch.ones(0), requires_grad=True) 163 | 164 | # 量化/反量化 165 | def forward(self, weight): 166 | if self.init_state==0: 167 | self.g = 1.0/math.sqrt(weight.numel() * self.Qp) 168 | self.div = 2**self.w_bits - 1 169 | if self.per_channel: 170 | weight_tmp = weight.detach().contiguous().view(weight.size()[0], -1) 171 | mean = torch.mean(weight_tmp, dim=1) 172 | std = torch.std(weight_tmp, dim=1) 173 | self.s.data, _ = torch.max(torch.stack([torch.abs(mean-3*std), torch.abs(mean + 3*std)]), dim=0) 174 | self.s.data = self.s.data/self.div 175 | else: 176 | mean = torch.mean(weight.detach()) 177 | std = torch.std(weight.detach()) 178 | self.s.data = max([torch.abs(mean-3*std), torch.abs(mean + 3*std)])/self.div 179 | self.init_state += 1 180 | elif self.init_state Qp).float() #bool值转浮点值，1.0或者0.0 223 | between = 1.0 - smaller -bigger #得到位于量化区间的index 224 | grad_alpha = ((smaller * Qn + bigger * Qp + 225 | between * Round.apply(q_input) - between * q_input) * g).sum().unsqueeze(dim=0) 226 | grad_beta = ((smaller + bigger) * g).sum().unsqueeze(dim=0) 227 | # print('grad_beta: ',grad_beta,g, smaller.sum(), bigger.sum(), between.sum(),Qn, Qp) 228 | child.activation_quantizer.s.grad.data.add_(g*(2*(child.quant_input-child.input)*grad_alpha).sum().unsqueeze(dim=0)) 229 | child.activation_quantizer.beta.grad.data.add_(g*(2*(child.quant_input-child.input)*grad_beta).sum().unsqueeze(dim=0)) 230 | 231 | model._modules[name] = child 232 | # print('after: ', model._modules[name].activation_quantizer.s.grad.data, model._modules[name].activation_quantizer.beta.grad.data, s, beta, 233 | # torch.square(child.quant_input-child.input).sum()) 234 | else: 235 | child = update_LSQplus_activation_Scalebeta(child) 236 | model._modules[name] = child 237 | return model 238 | 239 | 240 | 241 | class QuantConv2d(nn.Conv2d): 242 | def __init__(self, 243 | in_channels, 244 | out_channels, 245 | kernel_size, 246 | stride=1, 247 | padding=0, 248 | dilation=1, 249 | groups=1, 250 | bias=True, 251 | padding_mode='zeros', 252 | a_bits=8, 253 | w_bits=8, 254 | quant_inference=False, 255 | all_positive=False, 256 | per_channel=False, 257 | batch_init = 20): 258 | super(QuantConv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups, 259 | bias, padding_mode) 260 | self.quant_inference = quant_inference 261 | self.activation_quantizer = LSQPlusActivationQuantizer(a_bits=a_bits, all_positive=all_positive,batch_init = batch_init) 262 | self.weight_quantizer = LSQPlusWeightQuantizer(w_bits=w_bits, all_positive=all_positive, per_channel=per_channel,batch_init = batch_init) 263 | 264 | def forward(self, input): 265 | self.input = input 266 | self.quant_input = self.activation_quantizer(self.input) 267 | if not self.quant_inference: 268 | self.quant_weight = self.weight_quantizer(self.weight) 269 | else: 270 | self.quant_weight = self.weight 271 | 272 | output = F.conv2d(self.quant_input, self.quant_weight, self.bias, self.stride, self.padding, self.dilation, 273 | self.groups) 274 | return output 275 | 276 | 277 | class QuantConvTranspose2d(nn.ConvTranspose2d): 278 | def __init__(self, 279 | in_channels, 280 | out_channels, 281 | kernel_size, 282 | stride=1, 283 | padding=0, 284 | output_padding=0, 285 | dilation=1, 286 | groups=1, 287 | bias=True, 288 | padding_mode='zeros', 289 | a_bits=8, 290 | w_bits=8, 291 | quant_inference=False, 292 | all_positive=False, 293 | per_channel=False, 294 | batch_init = 20): 295 | super(QuantConvTranspose2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, output_padding, 296 | dilation, groups, bias, padding_mode) 297 | self.quant_inference = quant_inference 298 | self.activation_quantizer = LSQPlusActivationQuantizer(a_bits=a_bits, all_positive=all_positive,batch_init = batch_init) 299 | self.weight_quantizer = LSQPlusWeightQuantizer(w_bits=w_bits, all_positive=all_positive, per_channel=per_channel,batch_init = batch_init) 300 | 301 | def forward(self, input): 302 | self.input = input 303 | self.quant_input = self.activation_quantizer(self.input) 304 | if not self.quant_inference: 305 | self.quant_weight = self.weight_quantizer(self.weight) 306 | else: 307 | self.quant_weight = self.weight 308 | output = F.conv_transpose2d(self.quant_input, self.quant_weight, self.bias, self.stride, self.padding, self.output_padding, 309 | self.groups, self.dilation) 310 | return output 311 | 312 | 313 | class QuantLinear(nn.Linear): 314 | def __init__(self, 315 | in_features, 316 | out_features, 317 | bias=True, 318 | a_bits=8, 319 | w_bits=8, 320 | quant_inference=False, 321 | all_positive=False, 322 | per_channel=False, 323 | batch_init = 20): 324 | super(QuantLinear, self).__init__(in_features, out_features, bias) 325 | self.quant_inference = quant_inference 326 | self.activation_quantizer = LSQPlusActivationQuantizer(a_bits=a_bits, all_positive=all_positive,batch_init = batch_init) 327 | self.weight_quantizer = LSQPlusWeightQuantizer(w_bits=w_bits, all_positive=all_positive, per_channel=per_channel,batch_init = batch_init) 328 | 329 | def forward(self, input): 330 | self.input = input 331 | self.quant_input = self.activation_quantizer(self.input) 332 | if not self.quant_inference: 333 | self.quant_weight = self.weight_quantizer(self.weight) 334 | else: 335 | self.quant_weight = self.weight 336 | output = F.linear(self.quant_input, self.quant_weight, self.bias) 337 | return output 338 | 339 | 340 | def add_quant_op(module, layer_counter, a_bits=8, w_bits=8, 341 | quant_inference=False, all_positive=False, per_channel=False, batch_init = 20): 342 | for name, child in module.named_children(): 343 | if isinstance(child, nn.Conv2d): 344 | layer_counter[0] += 1 345 | if layer_counter[0] >= 1: #第一层也量化 346 | if child.bias is not None: 347 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 348 | child.kernel_size, stride=child.stride, 349 | padding=child.padding, dilation=child.dilation, 350 | groups=child.groups, bias=True, padding_mode=child.padding_mode, 351 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 352 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 353 | quant_conv.bias.data = child.bias 354 | else: 355 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 356 | child.kernel_size, stride=child.stride, 357 | padding=child.padding, dilation=child.dilation, 358 | groups=child.groups, bias=False, padding_mode=child.padding_mode, 359 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 360 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 361 | quant_conv.weight.data = child.weight 362 | module._modules[name] = quant_conv 363 | elif isinstance(child, nn.ConvTranspose2d): 364 | layer_counter[0] += 1 365 | if layer_counter[0] >= 1: #第一层也量化 366 | if child.bias is not None: 367 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 368 | child.out_channels, 369 | child.kernel_size, 370 | stride=child.stride, 371 | padding=child.padding, 372 | output_padding=child.output_padding, 373 | dilation=child.dilation, 374 | groups=child.groups, 375 | bias=True, 376 | padding_mode=child.padding_mode, 377 | a_bits=a_bits, 378 | w_bits=w_bits, 379 | quant_inference=quant_inference, 380 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 381 | quant_conv_transpose.bias.data = child.bias 382 | else: 383 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 384 | child.out_channels, 385 | child.kernel_size, 386 | stride=child.stride, 387 | padding=child.padding, 388 | output_padding=child.output_padding, 389 | dilation=child.dilation, 390 | groups=child.groups, bias=False, 391 | padding_mode=child.padding_mode, 392 | a_bits=a_bits, 393 | w_bits=w_bits, 394 | quant_inference=quant_inference, 395 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 396 | quant_conv_transpose.weight.data = child.weight 397 | module._modules[name] = quant_conv_transpose 398 | elif isinstance(child, nn.Linear): 399 | layer_counter[0] += 1 400 | if layer_counter[0] >= 1: #第一层也量化 401 | if child.bias is not None: 402 | quant_linear = QuantLinear(child.in_features, child.out_features, 403 | bias=True, a_bits=a_bits, w_bits=w_bits, 404 | quant_inference=quant_inference, 405 | all_positive=all_positive, per_channel=per_channel, 406 | batch_init = batch_init) 407 | quant_linear.bias.data = child.bias 408 | else: 409 | quant_linear = QuantLinear(child.in_features, child.out_features, 410 | bias=False, a_bits=a_bits, w_bits=w_bits, 411 | quant_inference=quant_inference, 412 | all_positive=all_positive, per_channel=per_channel, 413 | batch_init = batch_init) 414 | quant_linear.weight.data = child.weight 415 | module._modules[name] = quant_linear 416 | else: 417 | add_quant_op(child, layer_counter, a_bits=a_bits, w_bits=w_bits, 418 | quant_inference=quant_inference, all_positive=all_positive, 419 | per_channel=per_channel, batch_init = batch_init) 420 | 421 | 422 | def prepare(model, inplace=False, a_bits=8, w_bits=8, quant_inference=False, 423 | all_positive=False, per_channel=False, batch_init = 20): 424 | if not inplace: 425 | model = copy.deepcopy(model) 426 | layer_counter = [0] 427 | add_quant_op(model, layer_counter, a_bits=a_bits, w_bits=w_bits, 428 | quant_inference=quant_inference, 429 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 430 | return model 431 | -------------------------------------------------------------------------------- /quantization/lsqplus_quantize_V2.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from torch.autograd import Function 7 | from quantization.lsqquantize_V1 import Round 8 | 9 | class ALSQPlus(Function): 10 | @staticmethod 11 | def forward(ctx, weight, alpha, g, Qn, Qp, beta): 12 | # assert alpha > 0, "alpha={}".format(alpha) 13 | ctx.save_for_backward(weight, alpha, beta) 14 | ctx.other = g, Qn, Qp 15 | w_q = Round.apply(torch.div((weight - beta), alpha).clamp(Qn, Qp)) 16 | w_q = w_q * alpha + beta 17 | return w_q 18 | 19 | @staticmethod 20 | def backward(ctx, grad_weight): 21 | weight, alpha, beta = ctx.saved_tensors 22 | g, Qn, Qp = ctx.other 23 | q_w = (weight - beta) / alpha 24 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 25 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 26 | between = 1.0 - smaller -bigger #得到位于量化区间的index 27 | grad_alpha = ((smaller * Qn + bigger * Qp + 28 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) 29 | grad_beta = ((smaller + bigger) * grad_weight * g).sum().unsqueeze(dim=0) 30 | #在量化区间之外的值都是常数，故导数也是0 31 | grad_weight = between * grad_weight 32 | #返回的梯度要和forward的参数对应起来 33 | return grad_weight, grad_alpha, None, None, None, grad_beta 34 | 35 | class WLSQPlus(Function): 36 | @staticmethod 37 | def forward(ctx, weight, alpha, g, Qn, Qp, per_channel): 38 | # assert alpha > 0, "alpha={}".format(alpha) 39 | ctx.save_for_backward(weight, alpha) 40 | ctx.other = g, Qn, Qp, per_channel 41 | if per_channel: 42 | sizes = weight.size() 43 | weight = weight.contiguous().view(weight.size()[0], -1) 44 | weight = torch.transpose(weight, 0, 1) 45 | alpha = torch.broadcast_to(alpha, weight.size()) 46 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 47 | w_q = w_q * alpha 48 | w_q = torch.transpose(w_q, 0, 1) 49 | w_q = w_q.contiguous().view(sizes) 50 | else: 51 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 52 | w_q = w_q * alpha 53 | return w_q 54 | 55 | @staticmethod 56 | def backward(ctx, grad_weight): 57 | weight, alpha = ctx.saved_tensors 58 | g, Qn, Qp, per_channel = ctx.other 59 | if per_channel: 60 | sizes = weight.size() 61 | weight = weight.contiguous().view(weight.size()[0], -1) 62 | weight = torch.transpose(weight, 0, 1) 63 | alpha = torch.broadcast_to(alpha, weight.size()) 64 | q_w = weight / alpha 65 | q_w = torch.transpose(q_w, 0, 1) 66 | q_w = q_w.contiguous().view(sizes) 67 | else: 68 | q_w = weight / alpha 69 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 70 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 71 | between = 1.0 - smaller -bigger #得到位于量化区间的index 72 | if per_channel: 73 | grad_alpha = ((smaller * Qn + bigger * Qp + 74 | between * Round.apply(q_w) - between * q_w)*grad_weight * g) 75 | grad_alpha = grad_alpha.contiguous().view(grad_alpha.size()[0], -1).sum(dim=1) 76 | else: 77 | grad_alpha = ((smaller * Qn + bigger * Qp + 78 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) 79 | #在量化区间之外的值都是常数，故导数也是0 80 | grad_weight = between * grad_weight 81 | return grad_weight, grad_alpha, None, None, None, None 82 | 83 | def grad_scale(x, scale): 84 | y = x 85 | y_grad = x * scale 86 | return (y - y_grad).detach() + y_grad 87 | 88 | def round_pass(x): 89 | y = x.round() 90 | y_grad = x 91 | return (y - y_grad).detach() + y_grad 92 | 93 | def get_percentile_min_max(input, lower_percentile, uppper_percentile, output_tensor): 94 | batch_size = input.shape[0] 95 | lower_index = round(batch_size * (1 - lower_percentile*0.01)) 96 | upper_index = round(batch_size * (1 - uppper_percentile*0.01)) 97 | 98 | upper_bound = torch.kthvalue(input, k=upper_index).values 99 | 100 | if lower_percentile==0: 101 | lower_bound = upper_bound * 0 102 | else: 103 | low_bound = -torch.kthvalue(-input, k=lower_index).values 104 | 105 | # def update_scale_betas(): 106 | # for m in model.modules(): 107 | # if isinstance(m, nn.) 108 | 109 | # A(特征)量化 110 | class LSQPlusActivationQuantizer(nn.Module): 111 | def __init__(self, a_bits, all_positive=False,batch_init = 20): 112 | #activations 没有per-channel这个选项的 113 | super(LSQPlusActivationQuantizer, self).__init__() 114 | self.a_bits = a_bits 115 | self.all_positive = all_positive 116 | self.batch_init = batch_init 117 | if self.all_positive: 118 | # unsigned activation is quantized to [0, 2^b-1] 119 | self.Qn = 0 120 | self.Qp = 2 ** self.a_bits - 1 121 | else: 122 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 123 | self.Qn = - 2 ** (self.a_bits - 1) 124 | self.Qp = 2 ** (self.a_bits - 1) - 1 125 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) 126 | self.beta = torch.nn.Parameter(torch.ones(0), requires_grad=True) 127 | self.init_state = 0 128 | 129 | # 量化/反量化 130 | def forward(self, activation): 131 | if self.init_state==0: 132 | self.g = 1.0/math.sqrt(activation.numel() * self.Qp) 133 | mina = torch.min(activation.detach()) 134 | self.s.data = (torch.max(activation.detach()) - mina)/(self.Qp-self.Qn) 135 | self.beta.data = mina - self.s.data *self.Qn 136 | self.init_state += 1 137 | elif self.init_state= 1: #第一层也量化 329 | if child.bias is not None: 330 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 331 | child.kernel_size, stride=child.stride, 332 | padding=child.padding, dilation=child.dilation, 333 | groups=child.groups, bias=True, padding_mode=child.padding_mode, 334 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 335 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 336 | quant_conv.bias.data = child.bias 337 | else: 338 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 339 | child.kernel_size, stride=child.stride, 340 | padding=child.padding, dilation=child.dilation, 341 | groups=child.groups, bias=False, padding_mode=child.padding_mode, 342 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 343 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 344 | quant_conv.weight.data = child.weight 345 | module._modules[name] = quant_conv 346 | elif isinstance(child, nn.ConvTranspose2d): 347 | layer_counter[0] += 1 348 | if layer_counter[0] >= 1: #第一层也量化 349 | if child.bias is not None: 350 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 351 | child.out_channels, 352 | child.kernel_size, 353 | stride=child.stride, 354 | padding=child.padding, 355 | output_padding=child.output_padding, 356 | dilation=child.dilation, 357 | groups=child.groups, 358 | bias=True, 359 | padding_mode=child.padding_mode, 360 | a_bits=a_bits, 361 | w_bits=w_bits, 362 | quant_inference=quant_inference, 363 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 364 | quant_conv_transpose.bias.data = child.bias 365 | else: 366 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 367 | child.out_channels, 368 | child.kernel_size, 369 | stride=child.stride, 370 | padding=child.padding, 371 | output_padding=child.output_padding, 372 | dilation=child.dilation, 373 | groups=child.groups, bias=False, 374 | padding_mode=child.padding_mode, 375 | a_bits=a_bits, 376 | w_bits=w_bits, 377 | quant_inference=quant_inference, 378 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 379 | quant_conv_transpose.weight.data = child.weight 380 | module._modules[name] = quant_conv_transpose 381 | elif isinstance(child, nn.Linear): 382 | layer_counter[0] += 1 383 | if layer_counter[0] >= 1: #第一层也量化 384 | if child.bias is not None: 385 | quant_linear = QuantLinear(child.in_features, child.out_features, 386 | bias=True, a_bits=a_bits, w_bits=w_bits, 387 | quant_inference=quant_inference, 388 | all_positive=all_positive, per_channel=per_channel, 389 | batch_init = batch_init) 390 | quant_linear.bias.data = child.bias 391 | else: 392 | quant_linear = QuantLinear(child.in_features, child.out_features, 393 | bias=False, a_bits=a_bits, w_bits=w_bits, 394 | quant_inference=quant_inference, 395 | all_positive=all_positive, per_channel=per_channel, 396 | batch_init = batch_init) 397 | quant_linear.weight.data = child.weight 398 | module._modules[name] = quant_linear 399 | else: 400 | add_quant_op(child, layer_counter, a_bits=a_bits, w_bits=w_bits, 401 | quant_inference=quant_inference, all_positive=all_positive, 402 | per_channel=per_channel, batch_init = batch_init) 403 | 404 | 405 | def prepare(model, inplace=False, a_bits=8, w_bits=8, quant_inference=False, 406 | all_positive=False, per_channel=False, batch_init = 20): 407 | if not inplace: 408 | model = copy.deepcopy(model) 409 | layer_counter = [0] 410 | add_quant_op(model, layer_counter, a_bits=a_bits, w_bits=w_bits, 411 | quant_inference=quant_inference, 412 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 413 | return model -------------------------------------------------------------------------------- /quantization/lsqquantize_V1.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from torch.autograd import Function 7 | 8 | 9 | # ********************* quantizers（量化器，量化） ********************* 10 | # 取整(ste) 11 | class Round(Function): 12 | @staticmethod 13 | def forward(self, input): 14 | sign = torch.sign(input) 15 | output = sign * torch.floor(torch.abs(input) + 0.5) 16 | return output 17 | 18 | @staticmethod 19 | def backward(self, grad_output): 20 | grad_input = grad_output.clone() 21 | return grad_input 22 | 23 | class FunLSQ(Function): 24 | @staticmethod 25 | def forward(ctx, weight, alpha, g, Qn, Qp, per_channel=False): 26 | #根据论文里LEARNED STEP SIZE QUANTIZATION第2节的公式 27 | # assert alpha > 0, "alpha={}".format(alpha) 28 | ctx.save_for_backward(weight, alpha) 29 | ctx.other = g, Qn, Qp, per_channel 30 | if per_channel: 31 | sizes = weight.size() 32 | weight = weight.contiguous().view(weight.size()[0], -1) 33 | weight = torch.transpose(weight, 0, 1) 34 | alpha = torch.broadcast_to(alpha, weight.size()) 35 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 36 | w_q = w_q * alpha 37 | w_q = torch.transpose(w_q, 0, 1) 38 | w_q = w_q.contiguous().view(sizes) 39 | else: 40 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 41 | w_q = w_q * alpha 42 | return w_q 43 | 44 | @staticmethod 45 | def backward(ctx, grad_weight): 46 | #根据论文里LEARNED STEP SIZE QUANTIZATION第2.1节 47 | #分为三部分：位于量化区间的、小于下界的、大于上界的 48 | weight, alpha = ctx.saved_tensors 49 | g, Qn, Qp, per_channel = ctx.other 50 | if per_channel: 51 | sizes = weight.size() 52 | weight = weight.contiguous().view(weight.size()[0], -1) 53 | weight = torch.transpose(weight, 0, 1) 54 | alpha = torch.broadcast_to(alpha, weight.size()) 55 | q_w = weight / alpha 56 | q_w = torch.transpose(q_w, 0, 1) 57 | q_w = q_w.contiguous().view(sizes) 58 | else: 59 | q_w = weight / alpha 60 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 61 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 62 | between = 1.0 - smaller -bigger #得到位于量化区间的index 63 | if per_channel: 64 | grad_alpha = ((smaller * Qn + bigger * Qp + 65 | between * Round.apply(q_w) - between * q_w)*grad_weight * g) 66 | grad_alpha = grad_alpha.contiguous().view(grad_alpha.size()[0], -1).sum(dim=1) 67 | else: 68 | grad_alpha = ((smaller * Qn + bigger * Qp + 69 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) #? 70 | #在量化区间之外的值都是常数，故导数也是0 71 | grad_weight = between * grad_weight 72 | return grad_weight, grad_alpha, None, None, None, None 73 | 74 | def grad_scale(x, scale): 75 | y = x 76 | y_grad = x * scale 77 | return (y - y_grad).detach() + y_grad 78 | 79 | def round_pass(x): 80 | y = x.round() 81 | y_grad = x 82 | return (y - y_grad).detach() + y_grad 83 | 84 | # A(特征)量化 85 | class LSQActivationQuantizer(nn.Module): 86 | def __init__(self, a_bits, all_positive=False, batch_init = 20): 87 | #activations 没有per-channel这个选项的 88 | super(LSQActivationQuantizer, self).__init__() 89 | self.a_bits = a_bits 90 | self.all_positive = all_positive 91 | self.batch_init = batch_init 92 | if self.all_positive: 93 | # unsigned activation is quantized to [0, 2^b-1] 94 | self.Qn = 0 95 | self.Qp = 2 ** self.a_bits - 1 96 | else: 97 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 98 | self.Qn = - 2 ** (self.a_bits - 1) 99 | self.Qp = 2 ** (self.a_bits - 1) - 1 100 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) 101 | # self.s = torch.nn.Parameter(torch.ones(0.01), requires_grad=True) 102 | # self.register_parameter('Ascale', self.s) 103 | self.init_state = 0 104 | 105 | # 量化/反量化 106 | def forward(self, activation): 107 | ''' 108 | For this work, each layer of weights and each layer of activations has a distinct step size, represented 109 | as an fp32 value, initialized to 2h|v|i/√OP , computed on either the initial weights values or the first 110 | batch of activations, respectively 111 | ''' 112 | #V1 113 | if self.init_state==0: 114 | self.g = 1.0/math.sqrt(activation.numel() * self.Qp) 115 | self.s.data = torch.mean(torch.abs(activation.detach()))*2/(math.sqrt(self.Qp)) 116 | self.init_state += 1 117 | elif self.init_state= 1: #第一层也量化 293 | if child.bias is not None: 294 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 295 | child.kernel_size, stride=child.stride, 296 | padding=child.padding, dilation=child.dilation, 297 | groups=child.groups, bias=True, padding_mode=child.padding_mode, 298 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 299 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 300 | quant_conv.bias.data = child.bias 301 | else: 302 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 303 | child.kernel_size, stride=child.stride, 304 | padding=child.padding, dilation=child.dilation, 305 | groups=child.groups, bias=False, padding_mode=child.padding_mode, 306 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 307 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 308 | quant_conv.weight.data = child.weight 309 | module._modules[name] = quant_conv 310 | elif isinstance(child, nn.ConvTranspose2d): 311 | layer_counter[0] += 1 312 | if layer_counter[0] >= 1: #第一层也量化 313 | if child.bias is not None: 314 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 315 | child.out_channels, 316 | child.kernel_size, 317 | stride=child.stride, 318 | padding=child.padding, 319 | output_padding=child.output_padding, 320 | dilation=child.dilation, 321 | groups=child.groups, 322 | bias=True, 323 | padding_mode=child.padding_mode, 324 | a_bits=a_bits, 325 | w_bits=w_bits, 326 | quant_inference=quant_inference, 327 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 328 | quant_conv_transpose.bias.data = child.bias 329 | else: 330 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 331 | child.out_channels, 332 | child.kernel_size, 333 | stride=child.stride, 334 | padding=child.padding, 335 | output_padding=child.output_padding, 336 | dilation=child.dilation, 337 | groups=child.groups, bias=False, 338 | padding_mode=child.padding_mode, 339 | a_bits=a_bits, 340 | w_bits=w_bits, 341 | quant_inference=quant_inference, 342 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 343 | quant_conv_transpose.weight.data = child.weight 344 | module._modules[name] = quant_conv_transpose 345 | elif isinstance(child, nn.Linear): 346 | layer_counter[0] += 1 347 | if layer_counter[0] >= 1: #第一层也量化 348 | if child.bias is not None: 349 | quant_linear = QuantLinear(child.in_features, child.out_features, 350 | bias=True, a_bits=a_bits, w_bits=w_bits, 351 | quant_inference=quant_inference, 352 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 353 | quant_linear.bias.data = child.bias 354 | else: 355 | quant_linear = QuantLinear(child.in_features, child.out_features, 356 | bias=False, a_bits=a_bits, w_bits=w_bits, 357 | quant_inference=quant_inference, 358 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 359 | quant_linear.weight.data = child.weight 360 | module._modules[name] = quant_linear 361 | else: 362 | add_quant_op(child, layer_counter, a_bits=a_bits, w_bits=w_bits, 363 | quant_inference=quant_inference, all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 364 | 365 | 366 | def prepare(model, inplace=False, a_bits=8, w_bits=8, quant_inference=False, 367 | all_positive=False, per_channel=False, batch_init = 20): 368 | if not inplace: 369 | model = copy.deepcopy(model) 370 | layer_counter = [0] 371 | add_quant_op(model, layer_counter, a_bits=a_bits, w_bits=w_bits, 372 | quant_inference=quant_inference, all_positive=all_positive, 373 | per_channel=per_channel, batch_init = batch_init) 374 | return model 375 | -------------------------------------------------------------------------------- /quantization/lsqquantize_V2.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from torch.autograd import Function 7 | ''' 8 | self.s = torch.nn.Parameter(torch.ones(1)) #V2 9 | 激活值量化参数s初始化使用了常数1 10 | ''' 11 | 12 | # ********************* quantizers（量化器，量化） ********************* 13 | # 取整(ste) 14 | class Round(Function): 15 | @staticmethod 16 | def forward(self, input): 17 | sign = torch.sign(input) 18 | output = sign * torch.floor(torch.abs(input) + 0.5) 19 | return output 20 | 21 | @staticmethod 22 | def backward(self, grad_output): 23 | grad_input = grad_output.clone() 24 | return grad_input 25 | 26 | class FunLSQ(Function): 27 | @staticmethod 28 | def forward(ctx, weight, alpha, g, Qn, Qp, per_channel=False): 29 | #根据论文里LEARNED STEP SIZE QUANTIZATION第2节的公式 30 | # assert alpha > 0, "alpha={}".format(alpha) 31 | ctx.save_for_backward(weight, alpha) 32 | ctx.other = g, Qn, Qp, per_channel 33 | if per_channel: 34 | sizes = weight.size() 35 | weight = weight.contiguous().view(weight.size()[0], -1) 36 | weight = torch.transpose(weight, 0, 1) 37 | alpha = torch.broadcast_to(alpha, weight.size()) 38 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 39 | w_q = w_q * alpha 40 | w_q = torch.transpose(w_q, 0, 1) 41 | w_q = w_q.contiguous().view(sizes) 42 | else: 43 | w_q = Round.apply(torch.div(weight, alpha).clamp(Qn, Qp)) 44 | w_q = w_q * alpha 45 | return w_q 46 | 47 | @staticmethod 48 | def backward(ctx, grad_weight): 49 | #根据论文里LEARNED STEP SIZE QUANTIZATION第2.1节 50 | #分为三部分：位于量化区间的、小于下界的、大于上界的 51 | weight, alpha = ctx.saved_tensors 52 | g, Qn, Qp, per_channel = ctx.other 53 | if per_channel: 54 | sizes = weight.size() 55 | weight = weight.contiguous().view(weight.size()[0], -1) 56 | weight = torch.transpose(weight, 0, 1) 57 | alpha = torch.broadcast_to(alpha, weight.size()) 58 | q_w = weight / alpha 59 | q_w = torch.transpose(q_w, 0, 1) 60 | q_w = q_w.contiguous().view(sizes) 61 | else: 62 | q_w = weight / alpha 63 | smaller = (q_w < Qn).float() #bool值转浮点值，1.0或者0.0 64 | bigger = (q_w > Qp).float() #bool值转浮点值，1.0或者0.0 65 | between = 1.0 - smaller -bigger #得到位于量化区间的index 66 | if per_channel: 67 | grad_alpha = ((smaller * Qn + bigger * Qp + 68 | between * Round.apply(q_w) - between * q_w)*grad_weight * g) 69 | grad_alpha = grad_alpha.contiguous().view(grad_alpha.size()[0], -1).sum(dim=1) 70 | else: 71 | grad_alpha = ((smaller * Qn + bigger * Qp + 72 | between * Round.apply(q_w) - between * q_w)*grad_weight * g).sum().unsqueeze(dim=0) #? 73 | #在量化区间之外的值都是常数，故导数也是0 74 | grad_weight = between * grad_weight 75 | return grad_weight, grad_alpha, None, None, None, None 76 | 77 | def grad_scale(x, scale): 78 | y = x 79 | y_grad = x * scale 80 | return (y - y_grad).detach() + y_grad 81 | 82 | def round_pass(x): 83 | y = x.round() 84 | y_grad = x 85 | return (y - y_grad).detach() + y_grad 86 | 87 | # A(特征)量化 88 | class LSQActivationQuantizer(nn.Module): 89 | def __init__(self, a_bits, all_positive=False, batch_init = 20): 90 | #activations 没有per-channel这个选项的 91 | super(LSQActivationQuantizer, self).__init__() 92 | self.a_bits = a_bits 93 | self.all_positive = all_positive 94 | self.batch_init = batch_init 95 | if self.all_positive: 96 | # unsigned activation is quantized to [0, 2^b-1] 97 | self.Qn = 0 98 | self.Qp = 2 ** self.a_bits - 1 99 | else: 100 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 101 | self.Qn = - 2 ** (self.a_bits - 1) 102 | self.Qp = 2 ** (self.a_bits - 1) - 1 103 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) #V2 104 | # self.register_parameter('Ascale', self.s) 105 | self.init_state = 0 106 | 107 | # 量化/反量化 108 | def forward(self, activation): 109 | if self.a_bits == 32: 110 | output = activation 111 | elif self.a_bits == 1: 112 | print('！Binary quantization is not supported ！') 113 | assert self.a_bits != 1 114 | else: 115 | if self.init_state==0: 116 | self.g = 1.0/math.sqrt(activation.numel() * self.Qp) 117 | self.init_state += 1 118 | # print(self.s, self.g) 119 | q_a = FunLSQ.apply(activation, self.s, self.g, self.Qn, self.Qp) 120 | 121 | # alpha = grad_scale(self.s, g) 122 | # q_a = Round.apply((activation/alpha).clamp(Qn, Qp)) * alpha 123 | return q_a 124 | 125 | # W(权重)量化 126 | class LSQWeightQuantizer(nn.Module): 127 | def __init__(self, w_bits, all_positive=False, per_channel=False, batch_init = 20): 128 | super(LSQWeightQuantizer, self).__init__() 129 | self.w_bits = w_bits 130 | self.all_positive = all_positive 131 | self.batch_init = batch_init 132 | if self.all_positive: 133 | # unsigned activation is quantized to [0, 2^b-1] 134 | self.Qn = 0 135 | self.Qp = 2 ** w_bits - 1 136 | else: 137 | # signed weight/activation is quantized to [-2^(b-1), 2^(b-1)-1] 138 | self.Qn = - 2 ** (w_bits - 1) 139 | self.Qp = 2 ** (w_bits - 1) - 1 140 | self.per_channel = per_channel 141 | self.s = torch.nn.Parameter(torch.ones(1), requires_grad=True) 142 | # self.register_parameter('Wscale', self.s) 143 | self.init_state = 0 144 | 145 | # 量化/反量化 146 | def forward(self, weight): 147 | if self.init_state==0: 148 | self.g = 1.0/math.sqrt(weight.numel() * self.Qp) 149 | if self.per_channel: 150 | weight_tmp = weight.detach().contiguous().view(weight.size()[0], -1) 151 | self.s.data = torch.mean(torch.abs(weight_tmp), dim=1)*2/(math.sqrt(self.Qp)) 152 | else: 153 | self.s.data = torch.mean(torch.abs(weight.detach()))*2/(math.sqrt(self.Qp)) 154 | self.init_state += 1 155 | elif self.init_state= 1: #第一层也量化 282 | if child.bias is not None: 283 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 284 | child.kernel_size, stride=child.stride, 285 | padding=child.padding, dilation=child.dilation, 286 | groups=child.groups, bias=True, padding_mode=child.padding_mode, 287 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 288 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 289 | quant_conv.bias.data = child.bias 290 | else: 291 | quant_conv = QuantConv2d(child.in_channels, child.out_channels, 292 | child.kernel_size, stride=child.stride, 293 | padding=child.padding, dilation=child.dilation, 294 | groups=child.groups, bias=False, padding_mode=child.padding_mode, 295 | a_bits=a_bits, w_bits=w_bits, quant_inference=quant_inference, 296 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 297 | quant_conv.weight.data = child.weight 298 | module._modules[name] = quant_conv 299 | elif isinstance(child, nn.ConvTranspose2d): 300 | layer_counter[0] += 1 301 | if layer_counter[0] >= 1: #第一层也量化 302 | if child.bias is not None: 303 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 304 | child.out_channels, 305 | child.kernel_size, 306 | stride=child.stride, 307 | padding=child.padding, 308 | output_padding=child.output_padding, 309 | dilation=child.dilation, 310 | groups=child.groups, 311 | bias=True, 312 | padding_mode=child.padding_mode, 313 | a_bits=a_bits, 314 | w_bits=w_bits, 315 | quant_inference=quant_inference, 316 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 317 | quant_conv_transpose.bias.data = child.bias 318 | else: 319 | quant_conv_transpose = QuantConvTranspose2d(child.in_channels, 320 | child.out_channels, 321 | child.kernel_size, 322 | stride=child.stride, 323 | padding=child.padding, 324 | output_padding=child.output_padding, 325 | dilation=child.dilation, 326 | groups=child.groups, bias=False, 327 | padding_mode=child.padding_mode, 328 | a_bits=a_bits, 329 | w_bits=w_bits, 330 | quant_inference=quant_inference, 331 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 332 | quant_conv_transpose.weight.data = child.weight 333 | module._modules[name] = quant_conv_transpose 334 | elif isinstance(child, nn.Linear): 335 | layer_counter[0] += 1 336 | if layer_counter[0] >= 1: #第一层也量化 337 | if child.bias is not None: 338 | quant_linear = QuantLinear(child.in_features, child.out_features, 339 | bias=True, a_bits=a_bits, w_bits=w_bits, 340 | quant_inference=quant_inference, 341 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 342 | quant_linear.bias.data = child.bias 343 | else: 344 | quant_linear = QuantLinear(child.in_features, child.out_features, 345 | bias=False, a_bits=a_bits, w_bits=w_bits, 346 | quant_inference=quant_inference, 347 | all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 348 | quant_linear.weight.data = child.weight 349 | module._modules[name] = quant_linear 350 | else: 351 | add_quant_op(child, layer_counter, a_bits=a_bits, w_bits=w_bits, 352 | quant_inference=quant_inference, all_positive=all_positive, per_channel=per_channel, batch_init = batch_init) 353 | 354 | 355 | def prepare(model, inplace=False, a_bits=8, w_bits=8, quant_inference=False, 356 | all_positive=False, per_channel=False, batch_init = 20): 357 | if not inplace: 358 | model = copy.deepcopy(model) 359 | layer_counter = [0] 360 | add_quant_op(model, layer_counter, a_bits=a_bits, w_bits=w_bits, 361 | quant_inference=quant_inference, all_positive=all_positive, 362 | per_channel=per_channel, batch_init = batch_init) 363 | return model 364 | -------------------------------------------------------------------------------- /seebnparam.py: -------------------------------------------------------------------------------- 1 | #encoding=utf-8 2 | #Author: ZouJiu 3 | #Time: 2021-11-13 4 | 5 | import numpy as np 6 | import torch 7 | import os 8 | import time 9 | import torch 10 | import torchvision 11 | import torchvision.transforms as transforms 12 | from torch.utils.data import Dataset, DataLoader 13 | # from load_datas import TF, trainDataset, collate_fn 14 | import models #, resnet50 15 | from quantization.lsqquantize_V1 import prepare as lsqprepareV1 16 | from quantization.lsqquantize_V2 import prepare as lsqprepareV2 17 | from quantization.lsqplus_quantize_V1 import prepare as lsqplusprepareV1 18 | from quantization.lsqplus_quantize_V2 import prepare as lsqplusprepareV2 19 | from quantization.lsqplus_quantize_V1 import update_LSQplus_activation_Scalebeta 20 | import torch.optim as optim 21 | import datetime 22 | import matplotlib.pyplot as plt 23 | # os.environ["CUDA_VISIBLE_DEVICES"] = '0' 24 | 25 | def adjust_lr(optimizer, stepiters, epoch): 26 | # if stepiters < 100: #2warmup start 27 | # lr = stepiters*0.01/100 28 | # elif stepiters < 2000: 29 | # lr = 0.001 30 | # elif stepiters < 3000: 31 | # lr = 0.001 32 | if epoch <= 30: 33 | lr = 0.1 34 | elif epoch <= 46: 35 | lr = 0.01 36 | elif epoch <= 55: 37 | lr = 0.001 38 | else: 39 | lr = 0.0001 40 | for param_group in optimizer.param_groups: 41 | param_group['lr'] = lr 42 | return lr 43 | 44 | def trainer(): 45 | #batch_init 使用预训练模型对量化参数进行初始化的iters or steps 46 | config = {'a_bit':8, 'w_bit':8, "all_positive":False, "per_channel":True, 47 | "num_classes":10,"batch_init":20} 48 | pretrainedmodel = r'C:\Users\10696\Desktop\QAT\lsq+\log\model_108_42510_0.003_92.528_2021-11-27_17-49-47.pth' 49 | # Resnet_pretrain = False 50 | batch_size = 128 51 | num_epochs = 112 52 | Floatmodel = True #QAT or float-32 train False or True 53 | LSQplus = False #LSQ+ or LSQ True or False 54 | version = 'V1' 55 | scratch = False #从最开始训练，不是finetuning， 若=False就是finetuning 56 | showstep = 31 57 | #LSQPlusActivationQuantizer里的self.beta初始值要关注 58 | plusV1_inititers = 30 #update激活层的量化参数s和beta 59 | assert showstep > 0 60 | assert isinstance(showstep, int) 61 | assert isinstance(batch_size, int) 62 | assert isinstance(num_epochs, int) 63 | if Floatmodel: 64 | prefix = 'float32' 65 | elif LSQplus and not Floatmodel and version=='V1': 66 | if not config['per_channel']: 67 | prefix = 'LSQplus_V1' 68 | else: 69 | prefix = 'LSQplus_V1_pcl' 70 | elif LSQplus and not Floatmodel and version=='V2': 71 | if not config['per_channel']: 72 | prefix = 'LSQplus_V2' 73 | else: 74 | prefix = 'LSQplus_V2_pcl' 75 | elif not LSQplus and not Floatmodel and version=='V1': 76 | if not config['per_channel']: 77 | prefix = 'LSQ_V1' 78 | else: 79 | prefix = 'LSQ_V1_pcl' 80 | elif not LSQplus and not Floatmodel and version=='V2': 81 | if not config['per_channel']: 82 | prefix = 'LSQ_V2' 83 | else: 84 | prefix = 'LSQ_V2_pcl' 85 | else: 86 | print('setting is wrong......, please check it') 87 | exit(-1) 88 | 89 | tim = datetime.datetime.strftime(datetime.datetime.now(),"%Y-%m-%d %H-%M-%S").replace(' ', '_') 90 | logfile = r'log'+os.sep+prefix+'_log_%s.txt'%tim 91 | savepath = r'log' 92 | flogs = open(logfile, 'w') 93 | 94 | train_transform = transforms.Compose([ 95 | transforms.RandomCrop(32, padding=4), 96 | transforms.RandomHorizontalFlip(p=0.5), 97 | # transforms.Resize((32, 32)), 98 | transforms.ToTensor(), 99 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))]) 100 | test_transform = transforms.Compose([ 101 | # transforms.Resize((32, 32)), 102 | transforms.ToTensor(), 103 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))]) 104 | 105 | trainset = torchvision.datasets.CIFAR10(root='datas', train=True, 106 | download=True, transform=train_transform) 107 | trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, 108 | shuffle=True, num_workers=2, drop_last=True) 109 | 110 | testset = torchvision.datasets.CIFAR10(root='datas', train=False, 111 | download=True, transform=test_transform) 112 | testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, 113 | shuffle=False, num_workers=2, drop_last=True) 114 | 115 | classes = ('plane', 'car', 'bird', 'cat', 116 | 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') 117 | device = "cuda" if torch.cuda.is_available() else "cpu" 118 | 119 | model = models.resnet18(num_classes=config['num_classes']) 120 | 121 | #LSQ+ 122 | if LSQplus and not Floatmodel and version=='V1': 123 | #LSQplus V1 124 | lsqplusprepareV1(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 125 | all_positive=config["all_positive"], per_channel=config["per_channel"], 126 | batch_init = config["batch_init"]) 127 | print(model, '\npreparing lsqplus V1 models') 128 | elif LSQplus and not Floatmodel and version=='V2': 129 | #LSQplus V2 130 | lsqplusprepareV2(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 131 | all_positive=config["all_positive"], per_channel=config["per_channel"], 132 | batch_init = config["batch_init"]) 133 | print(model, '\npreparing lsqplus V2 models') 134 | elif not LSQplus and not Floatmodel and version=='V1': 135 | #LSQ V1 136 | lsqprepareV1(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 137 | all_positive=config["all_positive"], per_channel=config["per_channel"], 138 | batch_init = config["batch_init"]) 139 | print(model, '\npreparing lsq V1 models') 140 | elif not LSQplus and not Floatmodel and version=='V2': 141 | #LSQ V2 142 | lsqprepareV2(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 143 | all_positive=config["all_positive"], per_channel=config["per_channel"], 144 | batch_init = config["batch_init"]) 145 | print(model, '\npreparing lsq V2 models') 146 | elif Floatmodel: 147 | print(model, '\npreparing float models') 148 | pass 149 | # if not Floatmodel: 150 | # print(model) 151 | flogs.write(str(model)+'\n') 152 | if not os.path.exists(pretrainedmodel): 153 | print('the pretrainedmodel do not exists %s'%pretrainedmodel) 154 | if pretrainedmodel and os.path.exists(pretrainedmodel): 155 | print('loading pretrained model: ', pretrainedmodel) 156 | if torch.cuda.is_available(): 157 | state_dict = torch.load(pretrainedmodel, map_location='cuda') 158 | else: 159 | state_dict = torch.load(pretrainedmodel, map_location='cpu') 160 | missingkeys, unexpected_keys = model.load_state_dict(state_dict['state_dict'], strict=False) 161 | print('missingkeys: ', missingkeys) 162 | print('unexpected_keys: ', unexpected_keys) 163 | if not scratch: 164 | iteration = state_dict['iteration'] 165 | alliters = state_dict['alliters'] 166 | nowepoch = state_dict['nowepoch'] 167 | else: 168 | iteration = 0 169 | alliters = 0 170 | nowepoch = 0 171 | print('loading complete') 172 | else: 173 | print('no pretrained model') 174 | iteration = 0 175 | alliters = 0 176 | nowepoch = 0 177 | model = model.to(device) 178 | 179 | weight = [] 180 | count = 0 181 | weightsepa = [] 182 | for m in model.modules(): 183 | if isinstance(m, torch.nn.Conv2d): 184 | w = m.weight.data.clone().detach().numpy() #out channel, in channel, h, w 185 | out_channel = w.shape[0] 186 | w_per_channel = np.reshape(w, (out_channel, -1)) 187 | w_per_layer = np.reshape(w, (-1)) 188 | print(w_per_channel.shape, w_per_layer.shape) 189 | weight.append(w_per_layer) 190 | weightsepa.extend(w_per_channel) 191 | print(len(weightsepa[-1])) 192 | count += 1 193 | 194 | print(len(weightsepa[11])) 195 | plt.hist(weightsepa[11], bins=100) 196 | plt.title("all weights parameters") 197 | plt.ylabel('numbers') 198 | plt.xlabel("weights") 199 | plt.show() 200 | 201 | # plt.figure(figsize=(1620,1620)) 202 | fig, axs = plt.subplots(3, 3) 203 | for i in range(3): 204 | for j in range(3): 205 | axs[i, j].hist(weight[i+j], bins=100) 206 | # axs[i, j].set_title("weights of layer %d"%(i+j+1)) 207 | 208 | plt.show() 209 | 210 | bn = [] 211 | count = 0 212 | bnsepa = [] 213 | for m in model.modules(): 214 | if isinstance(m, torch.nn.BatchNorm2d): 215 | size = m.weight.data.shape[0] 216 | gammas = list(m.weight.data.clone().detach().numpy()) 217 | bn.extend(gammas) 218 | bnsepa.append(gammas) 219 | print(len(bnsepa[-1])) 220 | count += 1 221 | 222 | plt.hist(bn, bins=100) 223 | plt.ylabel('numbers') 224 | plt.xlabel("γ") 225 | plt.show() 226 | 227 | # bn.sort() 228 | plt.plot(np.arange(len(bn)), bn) 229 | plt.title("resnet18 BN γ parameters γ*x+β") 230 | plt.ylabel("no sorted γ") 231 | plt.xlabel("indexs") 232 | plt.show() 233 | 234 | bn.sort() 235 | plt.plot(np.arange(len(bn)), bn) 236 | plt.title("resnet18 BN γ parameters γ*x+β") 237 | plt.ylabel("sorted γ") 238 | plt.xlabel("indexs") 239 | plt.show() 240 | 241 | tmp9 = bnsepa[2] 242 | plt.plot(np.arange(len(tmp9)), tmp9) 243 | plt.title("resnet18 BN γ parameters γ*x+β") 244 | plt.ylabel("no sorted γ") 245 | plt.xlabel("3th layer") 246 | plt.show() 247 | 248 | tmp9 = bnsepa[2] 249 | tmp9.sort() 250 | plt.plot(np.arange(len(tmp9)), tmp9) 251 | plt.title("resnet18 BN γ parameters γ*x+β") 252 | plt.ylabel("sorted γ") 253 | plt.xlabel("3th layer") 254 | plt.show() 255 | 256 | tmp9 = bnsepa[17] 257 | plt.plot(np.arange(len(tmp9)), tmp9) 258 | plt.title("resnet18 BN γ parameters γ*x+β") 259 | plt.ylabel("no sorted γ") 260 | plt.xlabel("18th layer") 261 | plt.show() 262 | 263 | tmp9 = bnsepa[17] 264 | tmp9.sort() 265 | plt.plot(np.arange(len(tmp9)), tmp9) 266 | plt.title("resnet18 BN γ parameters γ*x+β") 267 | plt.ylabel("sorted γ") 268 | plt.xlabel("18th layer") 269 | plt.show() 270 | 271 | plt.close() 272 | print(len(bnsepa)) 273 | print(bn[:30]) 274 | print(bn[-30:]) 275 | 276 | if __name__ == '__main__': 277 | trainer() 278 | -------------------------------------------------------------------------------- /trains.py: -------------------------------------------------------------------------------- 1 | #encoding=utf-8 2 | #Author: ZouJiu 3 | #Time: 2021-11-13 4 | 5 | import numpy as np 6 | import torch 7 | import os 8 | import time 9 | import torch 10 | import torchvision 11 | import torchvision.transforms as transforms 12 | from torch.utils.data import Dataset, DataLoader 13 | # from load_datas import TF, trainDataset, collate_fn 14 | import models #, resnet50 15 | from quantization.lsqquantize_V1 import prepare as lsqprepareV1 16 | from quantization.lsqquantize_V2 import prepare as lsqprepareV2 17 | from quantization.lsqplus_quantize_V1 import prepare as lsqplusprepareV1 18 | from quantization.lsqplus_quantize_V2 import prepare as lsqplusprepareV2 19 | from quantization.lsqplus_quantize_V1 import update_LSQplus_activation_Scalebeta 20 | import torch.optim as optim 21 | import datetime 22 | # os.environ["CUDA_VISIBLE_DEVICES"] = '0' 23 | 24 | def adjust_lr(optimizer, stepiters, epoch): 25 | # if stepiters < 100: #2warmup start 26 | # lr = stepiters*0.01/100 27 | # elif stepiters < 2000: 28 | # lr = 0.001 29 | # elif stepiters < 3000: 30 | # lr = 0.001 31 | if epoch <= 31: 32 | lr = 0.1 33 | elif epoch <= 61: 34 | lr = 0.01 35 | elif epoch <= 81: 36 | lr = 0.001 37 | else: 38 | lr = 0.0001 39 | for param_group in optimizer.param_groups: 40 | param_group['lr'] = lr 41 | return lr 42 | 43 | def trainer(): 44 | #batch_init 使用预训练模型对量化参数进行初始化的iters or steps 45 | config = {'a_bit':8, 'w_bit':8, "all_positive":False, "per_channel":False, 46 | "num_classes":10,"batch_init":20} 47 | pretrainedmodel = r'C:\Users\10696\Desktop\QAT\lsq+\log\model_108_42510_0.003_92.528_2021-11-27_17-49-47.pth' 48 | # Resnet_pretrain = False 49 | batch_size = 128 50 | num_epochs = 112 51 | Floatmodel = False #QAT or float-32 train False or True 52 | LSQplus = False #LSQ+ or LSQ True or False 53 | version = 'V1' 54 | scratch = True #从最开始训练，不是finetuning， 若=False就是finetuning 55 | showstep = 31 56 | #LSQPlusActivationQuantizer里的self.beta初始值要关注 57 | plusV1_inititers = 30 #update激活层的量化参数s和beta 58 | assert showstep > 0 59 | assert isinstance(showstep, int) 60 | assert isinstance(batch_size, int) 61 | assert isinstance(num_epochs, int) 62 | if Floatmodel: 63 | prefix = 'float32' 64 | elif LSQplus and not Floatmodel and version=='V1': 65 | if not config['per_channel']: 66 | prefix = 'LSQplus_V1' 67 | else: 68 | prefix = 'LSQplus_V1_pcl' 69 | elif LSQplus and not Floatmodel and version=='V2': 70 | if not config['per_channel']: 71 | prefix = 'LSQplus_V2' 72 | else: 73 | prefix = 'LSQplus_V2_pcl' 74 | elif not LSQplus and not Floatmodel and version=='V1': 75 | if not config['per_channel']: 76 | prefix = 'LSQ_V1' 77 | else: 78 | prefix = 'LSQ_V1_pcl' 79 | elif not LSQplus and not Floatmodel and version=='V2': 80 | if not config['per_channel']: 81 | prefix = 'LSQ_V2' 82 | else: 83 | prefix = 'LSQ_V2_pcl' 84 | else: 85 | print('setting is wrong......, please check it') 86 | exit(-1) 87 | 88 | tim = datetime.datetime.strftime(datetime.datetime.now(),"%Y-%m-%d %H-%M-%S").replace(' ', '_') 89 | logfile = r'log'+os.sep+prefix+'_log_%s.txt'%tim 90 | savepath = r'log' 91 | flogs = open(logfile, 'w') 92 | 93 | train_transform = transforms.Compose([ 94 | transforms.RandomCrop(32, padding=4), 95 | transforms.RandomHorizontalFlip(p=0.5), 96 | # transforms.Resize((32, 32)), 97 | transforms.ToTensor(), 98 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))]) 99 | test_transform = transforms.Compose([ 100 | # transforms.Resize((32, 32)), 101 | transforms.ToTensor(), 102 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.201))]) 103 | 104 | trainset = torchvision.datasets.CIFAR10(root='datas', train=True, 105 | download=True, transform=train_transform) 106 | trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, 107 | shuffle=True, num_workers=2, drop_last=True) 108 | 109 | testset = torchvision.datasets.CIFAR10(root='datas', train=False, 110 | download=True, transform=test_transform) 111 | testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, 112 | shuffle=False, num_workers=2, drop_last=True) 113 | 114 | classes = ('plane', 'car', 'bird', 'cat', 115 | 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') 116 | device = "cuda" if torch.cuda.is_available() else "cpu" 117 | 118 | model = models.resnet18(num_classes=config['num_classes']) 119 | 120 | #LSQ+ 121 | if LSQplus and not Floatmodel and version=='V1': 122 | #LSQplus V1 123 | lsqplusprepareV1(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 124 | all_positive=config["all_positive"], per_channel=config["per_channel"], 125 | batch_init = config["batch_init"]) 126 | print(model, '\npreparing lsqplus V1 models') 127 | elif LSQplus and not Floatmodel and version=='V2': 128 | #LSQplus V2 129 | lsqplusprepareV2(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 130 | all_positive=config["all_positive"], per_channel=config["per_channel"], 131 | batch_init = config["batch_init"]) 132 | print(model, '\npreparing lsqplus V2 models') 133 | elif not LSQplus and not Floatmodel and version=='V1': 134 | #LSQ V1 135 | lsqprepareV1(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 136 | all_positive=config["all_positive"], per_channel=config["per_channel"], 137 | batch_init = config["batch_init"]) 138 | print(model, '\npreparing lsq V1 models') 139 | elif not LSQplus and not Floatmodel and version=='V2': 140 | #LSQ V2 141 | lsqprepareV2(model, inplace=True, a_bits=config["a_bit"], w_bits=config["w_bit"], 142 | all_positive=config["all_positive"], per_channel=config["per_channel"], 143 | batch_init = config["batch_init"]) 144 | print(model, '\npreparing lsq V2 models') 145 | elif Floatmodel: 146 | print(model, '\npreparing float models') 147 | pass 148 | # if not Floatmodel: 149 | # print(model) 150 | flogs.write(str(model)+'\n') 151 | if not os.path.exists(pretrainedmodel): 152 | print('the pretrainedmodel do not exists %s'%pretrainedmodel) 153 | if pretrainedmodel and os.path.exists(pretrainedmodel): 154 | print('loading pretrained model: ', pretrainedmodel) 155 | if torch.cuda.is_available(): 156 | state_dict = torch.load(pretrainedmodel, map_location='cuda') 157 | else: 158 | state_dict = torch.load(pretrainedmodel, map_location='cpu') 159 | missingkeys, unexpected_keys = model.load_state_dict(state_dict['state_dict'], strict=False) 160 | print('missingkeys: ', missingkeys) 161 | print('unexpected_keys: ', unexpected_keys) 162 | if not scratch: 163 | iteration = state_dict['iteration'] 164 | alliters = state_dict['alliters'] 165 | nowepoch = state_dict['nowepoch'] 166 | else: 167 | iteration = 0 168 | alliters = 0 169 | nowepoch = 0 170 | print('loading complete') 171 | else: 172 | print('no pretrained model') 173 | iteration = 0 174 | alliters = 0 175 | nowepoch = 0 176 | model = model.to(device) 177 | # print(torch.__version__) 178 | time.sleep(3) 179 | adam = False 180 | lr = 0.001 # initial learning rate (SGD=1E-2, Adam=1E-3) 181 | momnetum=0.9 182 | params = [p for p in model.parameters() if p.requires_grad] 183 | # if adam: 184 | # optimizer = optim.Adam(params, lr=lr, betas=(momnetum, 0.999)) # adjust beta1 to momentum 185 | # else: 186 | optimizer = optim.SGD(params, lr=lr, momentum=momnetum, weight_decay=5e-4) 187 | # and a learning rate scheduler 188 | # lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, 189 | # step_size=7, 190 | # gamma=0.1) 191 | torch.manual_seed(999999) 192 | start = time.time() 193 | print('Using {} device'.format(device)) 194 | flogs.write('Using {} device'.format(device)+'\n') 195 | stepiters = 0 196 | criterion = torch.nn.CrossEntropyLoss() 197 | pre = -999999 198 | for epoch in range(num_epochs): 199 | print('\nEpoch {}/{}'.format(epoch, num_epochs)) 200 | flogs.write('Epoch {}/{}'.format(epoch, num_epochs)+'\n') 201 | print('-'*100) 202 | running_loss = 0 203 | if epochnowepoch: 254 | print('validation of testes') 255 | with torch.no_grad(): 256 | count = 0 257 | print('length of testloader: ', len(testloader)) 258 | for data in testloader: 259 | count += 1 260 | images, labels = data 261 | images = images.to(device) 262 | labels = labels.to(device) 263 | outputs = model(images) 264 | # if count==100: 265 | # break 266 | _, predictions = torch.max(outputs, 1) 267 | # collect the correct predictions for each class 268 | for label, prediction in zip(labels, predictions): 269 | if label == prediction: 270 | correct_pred[classes[label]] += 1 271 | total_pred[classes[label]] += 1 272 | 273 | # print accuracy for each class 274 | correctall = 0 275 | alltest = 0 276 | for classname, correct_count in correct_pred.items(): 277 | accuracy = 100 * float(correct_count) / total_pred[classname] 278 | print("Validation Accuracy for class {:5s} is: {:.1f} %".format(classname, 279 | accuracy)) 280 | correctall += correct_count 281 | alltest += total_pred[classname] 282 | flogs.write("Accuracy for class {:5s} is: {:.1f} %".format(classname, accuracy)+'\n') 283 | flogs.flush() 284 | Accuracy = round(100 * float(correctall)/alltest, 3) 285 | print("Accuracy all is: {:.1f}".format(Accuracy)) 286 | 287 | # lr_scheduler.step() 288 | iteration=0 289 | try: 290 | if epoch>nowepoch and Accuracy>pre: 291 | torch.save(savestate, os.path.join(savepath, prefix+'_models_{}_{}_{}_{:.3f}_{}_{}.pth'.format( 292 | lr, epoch, stepiters, loss.item(),Accuracy,tim))) 293 | pre = Accuracy 294 | except: 295 | pass 296 | # evaluate(model, dataloader_test, device = device) 297 | timeused = time.time() - start 298 | print('Training complete in {:.0f}m {:.0f}s'.format(timeused//60, timeused%60)) 299 | flogs.close() 300 | 301 | if __name__ == '__main__': 302 | trainer() 303 | --------------------------------------------------------------------------------