├── .gitignore ├── ICPP_FORMAL.pdf ├── ICPP_Presentation.pdf ├── LICENSE.md ├── Makefile ├── README.md ├── core ├── __init__.py ├── optim │ ├── __init__.py │ └── gradient_sgd.py ├── server.py └── utils │ ├── GradualWarmupScheduler.py │ ├── __init__.py │ ├── constant.py │ ├── log.py │ ├── messaging.py │ └── serialization.py ├── example ├── ImageNet_dali_dataloader.py ├── Imagenet_dist.py ├── Imagenet_local.py ├── an4.py ├── cifar10.py ├── graph.py ├── main.py ├── models.py ├── pssh_script.py └── tinyimagenet.py ├── notebook ├── Async-result-analyze-cifar.ipynb ├── Async-result-analyze-imagenet.ipynb ├── Async-result-analyze.ipynb ├── Untitled.ipynb └── result_analyze.ipynb ├── requirements.txt └── setup.py /.gitignore: -------------------------------------------------------------------------------- 1 | data 2 | __pycache__/ 3 | .pyc* 4 | train.log 5 | env/ 6 | venv/ 7 | .idea/ 8 | .ipynb_checkpoints/* 9 | log/* 10 | build/* 11 | dist/* 12 | pytorch_distbelief.egg-info/* 13 | -------------------------------------------------------------------------------- /ICPP_FORMAL.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yanring/DGS/97e738995eccd771741f87af2251e929788e106e/ICPP_FORMAL.pdf -------------------------------------------------------------------------------- /ICPP_Presentation.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yanring/DGS/97e738995eccd771741f87af2251e929788e106e/ICPP_Presentation.pdf -------------------------------------------------------------------------------- /LICENSE.md: -------------------------------------------------------------------------------- 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 | -------------------------------------------------------------------------------- /Makefile: -------------------------------------------------------------------------------- 1 | setup: 2 | sudo apt-get -y virtualenv 3 | virtualenv -p python3 venv 4 | . venv/bin/activate && pip install -r requirements.txt && pip install . 5 | 6 | install: 7 | pip install . 8 | 9 | graph: 10 | python example/graph.py 11 | mv train_time.png test_time.png docs 12 | 13 | first: 14 | python example/main.py --rank 1 --world-size 3 15 | 16 | second: 17 | python example/main.py --rank 2 --world-size 3 18 | 19 | server: 20 | python example/main.py --rank 0 --world-size 3 --server 21 | 22 | single: 23 | python example/main.py --no-distributed 24 | 25 | gpu: 26 | python example/main.py --no-distributed --cuda 27 | 28 | dist: 29 | python3 setup.py sdist bdist_wheel 30 | 31 | upload: dist 32 | twine upload dist/* 33 | 34 | upload-test: dist 35 | twine upload --repository-url https://test.pypi.org/legacy/ dist/* 36 | 37 | install-test: 38 | python3 -m pip install --index-url https://test.pypi.org/simple/ pytorch-distbelief 39 | 40 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # DGS PyTorch 2 | [![License](https://img.shields.io/github/license/yanring/gradientserver)]() 3 | 4 | Modern large scale machine learning applications require stochastic optimization algorithms to be implemented with distributed computational architectures. A key bottleneck is the communication overhead for exchanging information, such as stochastic gradients, among different nodes. Recently, gradient sparsification techniques have been proposed to reduce communications cost and thus alleviate the network overhead. However, most of gradient sparsification techniques consider only synchronous parallelism and cannot be applied in asynchronous distributed training. 5 | 6 | In this project, we present a dual-way gradient sparsification approach (DGS) that is suitable for asynchronous distributed training. 7 | 8 | ![Architecture of PS Based ASGD and DGS](https://yanring-1252048839.cos.ap-guangzhou.myqcloud.com/img/20191211100612.png) 9 | 10 | We implemented a async parameter server based on PyTorch gloo backend. Our optimizer implemented 5 training methonds: DGS, ASGD, GradientDropping, DeepGradientCompression, single node momentum SGD. 11 | 12 | ## Features 13 | 1. Async parameter server on PyTorch 14 | 2. Support gradient sparsification, e.g. gradient dropping [1]. 15 | 3. Sparse communication. 16 | 4. GPU training. 17 | 5. DALI dataloader for imagenet. 18 | 6. Two distributed example script: cifar10 and ImageNet. 19 | 20 | ## Performance 21 | 22 | Experiments with different scales on ImageNet and CIFAR-10 show that: 23 | (1) compared with ASGD, Gradient Dropping and Deep Gradient Compression, DGS with SAMomentum consistently achieves better performance; 24 | (2) DGS improves the scalability of asynchronous training, especially with limited networking infrastructure. 25 | 26 | ![Speedups for DGS and ASGD with 10Gbps and 1Gbps Ethenet](https://yanring-1252048839.cos.ap-guangzhou.myqcloud.com/img/20191209143644.png) 27 | 28 | Figure above shows the training speedup with different network bandwidth values. 29 | As the number of workers increases, the acceleration of ASGD decreases to nearly zero due to the bottleneck of communication. In contrast, DGS achieves nearly linear speedup with 10Gbps. With 1Gbps network, ASGD only achieves $1\times$ speedup with 16 workers, while DGS achieves $12.6\times$ speedup, which proves the the superiority of our DGS under low bandwidth. 30 | 31 | 32 | ## Quick Start 33 | ``` 34 | # environ setup 35 | 36 | git clone https://github.com/yanring/GradientServer.git 37 | 38 | cd GradientServer 39 | 40 | pip install -r requirements.txt 41 | ``` 42 | 43 | ``` 44 | # distributed cifar-10 example 45 | # On the server (rank 0 is the server) 46 | python example/cifar10.py --world-size 3 --rank 0 --cuda 47 | # On the worker 1 48 | python example/cifar10.py --world-size 3 --rank 1 --cuda 49 | # On the worker 2 50 | python example/cifar10.py --world-size 3 --rank 2 --cuda 51 | ``` 52 | 53 | 54 | ## Use DGS in Your Code 55 | 56 | First, you should start a server. 57 | ``` 58 | init_server(args, net) 59 | ``` 60 | Second, replace your torch.optimizer.sgd with GradientSGD. 61 | ``` 62 | 63 | optimizer = GradientSGD(net.parameters(), lr=args.lr, model=net, momentum=args.momentum,weight_decay=args.weight_decay,args=args) 64 | ``` 65 | 66 | ## Limitations and Future Plans 67 | TODO 68 | 69 | ## Publications 70 | Zijie Yan, Danyang Xiao, Mengqiang Chen, Jieying Zhou, Weigang Wu: Dual-Way Gradient Sparsification for Asynchronous Distributed Deep Learning. ICPP 2020: 49:1-49:10 71 | 72 | ## Refrence 73 | [1] Alham Fikri Aji and Kenneth Heafield, ‘Sparse communication for distributed gradient descent’, arXiv preprint arXiv:1704.05021, (2017). 74 | [2] Jesse Cai and Rohan Varma, Implementation of Google's DistBelief paper. https://github.com/ucla-labx/distbelief 75 | -------------------------------------------------------------------------------- /core/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yanring/DGS/97e738995eccd771741f87af2251e929788e106e/core/__init__.py -------------------------------------------------------------------------------- /core/optim/__init__.py: -------------------------------------------------------------------------------- 1 | from .gradient_sgd import GradientSGD -------------------------------------------------------------------------------- /core/optim/gradient_sgd.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import sys 4 | import threading 5 | import time 6 | from queue import Queue 7 | 8 | import torch.distributed as dist 9 | from torch.optim.optimizer import Optimizer, required 10 | 11 | from core.utils.messaging import send_message, GSMessageCode, GradientMessageListener 12 | from core.utils.serialization import ravel_model_params, update_model_params, unravel_model_params, \ 13 | ravel_sparse_gradient, unravel_sparse_gradient, worker_gradient_executor, DGC, Aji 14 | 15 | WORKPATH = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) 16 | sys.path.append(WORKPATH) 17 | print(WORKPATH) 18 | _LOGGER = logging.getLogger(__name__) 19 | lock = threading.Lock() 20 | 21 | 22 | class GradientListener(GradientMessageListener): 23 | """DownpourListener""" 24 | 25 | def __init__(self, model, queue, args=None): 26 | super(GradientListener, self).__init__(ravel_model_params(model).numel(), source=0, args=args) 27 | self.lr = 0.05 28 | self.queue = queue 29 | self.version = 0 30 | self.model = model 31 | self.flag = False 32 | 33 | def receive(self, sender, message_code, gradient_version, lr, parameter): 34 | """receive parameter updates from the server and reflect them into the client's model.""" 35 | _LOGGER.info("Processing message: {}, version: {}, lr: {}".format(message_code.name, gradient_version, self.lr)) 36 | # print("Processing message: {}, version: {}, lr: {}".format(message_code.name, gradient_version, lr)) 37 | self.lr = lr 38 | if message_code == GSMessageCode.GradientUpdate: 39 | update_model_params(self.model, parameter, -1) 40 | self.version = gradient_version 41 | self.queue.put(gradient_version) 42 | elif message_code == GSMessageCode.SparseGradientUpdate: 43 | parameter = unravel_sparse_gradient(parameter).cuda().to_dense() 44 | update_model_params(self.model, parameter, -1) 45 | # print('4',parameter.sum()) 46 | 47 | self.version = gradient_version 48 | self.queue.put(gradient_version) 49 | elif message_code == GSMessageCode.ModelRequest: 50 | model = ravel_model_params(self.model, grads=False) 51 | send_message(GSMessageCode.ModelUpdate, model, dst=0, gradient_version=0) 52 | print('send model to server') 53 | elif message_code == GSMessageCode.ModelUpdate: 54 | # print('sync model!', gradient_version, ' ', datetime.now(), ' synced model :', parameter.sum()) 55 | unravel_model_params(self.model, parameter) 56 | self.version = gradient_version 57 | self.flag = True 58 | # TODO change back 59 | if self.version > 1: 60 | self.queue.put(gradient_version) 61 | # lock.release() 62 | 63 | 64 | class GradientSGD(Optimizer): 65 | """GradientSGD""" 66 | 67 | def __init__(self, params, lr=required, model=required, momentum=None, weight_decay=0, args=None): 68 | """ 69 | :param params: 70 | :param lr: 71 | :param n_push: 72 | :param n_pull: 73 | :param model: 74 | """ 75 | print('in my optimizer ') 76 | if lr is not required and lr < 0.0: 77 | raise ValueError("Invalid learning rate: {}".format(lr)) 78 | defaults = dict(lr=lr, ) 79 | self.model = model 80 | self.filter_gradient = ravel_model_params(model) 81 | self.momentum = momentum 82 | self.v_kt = self.filter_gradient.clone().zero_() 83 | self.u_kt = self.filter_gradient.clone().zero_() 84 | self.idx = 0 85 | self.version = 0 86 | self.queue = Queue(maxsize=1) 87 | if args.rank > 0: 88 | time.sleep(0.1 * int(args.rank)) 89 | dist.init_process_group('gloo', init_method='file://%s/sharedfile' % WORKPATH, group_name='mygroup', 90 | world_size=args.world_size, rank=args.rank) 91 | print('I am node rank:%d' % dist.get_rank()) 92 | self.listener = GradientListener(model, self.queue, args=args) 93 | self.listener.start() 94 | self.tmp = 0 95 | self.compress_ratio = None 96 | self.weight_decay = weight_decay 97 | print('weight_decay', self.weight_decay, 'lr', lr, 'momentum', self.momentum) 98 | self.args = args 99 | super(GradientSGD, self).__init__(params, defaults) 100 | 101 | def step(self, closure=None): 102 | """Performs a single optimization step. 103 | 104 | Arguments: 105 | closure (callable, optional): A closure that reevaluates the model 106 | and returns the loss. 107 | """ 108 | loss = None 109 | if closure is not None: 110 | loss = closure() 111 | if not self.listener.flag: 112 | while not self.args.no_distributed and not self.listener.flag: 113 | print('wait for server') 114 | time.sleep(1) 115 | return loss 116 | 117 | # get the lr 118 | if self.args.rank == 1: 119 | lr = self.param_groups[0]['lr'] 120 | # lr = 0.2 121 | else: 122 | if self.tmp != self.listener.lr: 123 | print('lr from %f to %f' % (self.tmp, self.listener.lr)) 124 | self.tmp = self.listener.lr 125 | self.param_groups[0]['lr'] = self.tmp 126 | lr = self.param_groups[0]['lr'] 127 | # print(lr) 128 | # lr = self.param_groups[0]['lr'] 129 | # keep track of accumulated gradients so that we can send 130 | # ASYNC 131 | if self.args.mode == 'asgd': 132 | # print('Running asgd') 133 | for param in self.model.parameters(): 134 | if self.weight_decay != 0: 135 | param.grad.data.add_(self.weight_decay, param.data) 136 | self.filter_gradient = ravel_model_params(self.model, grads=True, cuda=True).mul_(lr) 137 | 138 | send_message(GSMessageCode.GradientUpdate, self.filter_gradient, dst=0, 139 | gradient_version=self.listener.version + 1) 140 | self.version = self.queue.get() 141 | self.idx += 1 142 | return loss 143 | elif self.args.mode == 'gradient_sgd': 144 | # if self.version < 5: 145 | # print('Running gradient_sgd') 146 | 147 | raveled_gradients = worker_gradient_executor(self.model, self.filter_gradient, self.u_kt, self.v_kt, 148 | rate=0.01 * (lr / self.args.lr), 149 | lr=lr, momentum=self.momentum, weight_decay=self.weight_decay) 150 | # print(1,raveled_gradients.sum()) 151 | sparse_gradient = ravel_sparse_gradient(raveled_gradients) 152 | 153 | elif self.args.mode == 'dgc': 154 | # if self.version < 5: 155 | # print('Running dgc') 156 | raveled_gradients = DGC(self.model, self.filter_gradient, self.u_kt, self.v_kt, 157 | rate=0.01, 158 | # rate=self.compress_ratio, 159 | lr=lr, momentum=self.momentum, weight_decay=self.weight_decay) 160 | sparse_gradient = ravel_sparse_gradient(raveled_gradients) 161 | elif self.args.mode == 'aji': 162 | # if self.version < 5: 163 | # print('Running aji ', self.version) 164 | raveled_gradients = Aji(self.model, self.filter_gradient, self.u_kt, self.v_kt, 165 | rate=0.01, 166 | lr=lr, weight_decay=self.weight_decay) 167 | sparse_gradient = ravel_sparse_gradient(raveled_gradients) 168 | elif self.args.mode == 'sgd': 169 | # if self.version < 5: 170 | # print('Running sgd') 171 | weight_decay = self.weight_decay 172 | momentum = self.momentum 173 | dampening = 0 174 | nesterov = 0 175 | g = ravel_model_params(self.model, grads=True) 176 | p = ravel_model_params(self.model, grads=False) 177 | g.add_(weight_decay, p) 178 | self.u_kt.mul_(momentum).add_(g) 179 | raveled_gradients = self.u_kt.mul(lr) 180 | else: 181 | raise Exception('no optimizer') 182 | # reset gradient version 183 | if self.args.no_distributed: 184 | # parameter = unravel_sparse_gradient(sparse_gradient).cuda() 185 | update_model_params(self.model, raveled_gradients, 1) 186 | # self.version = gradient_version 187 | self.queue.put(self.idx) 188 | else: 189 | send_message(GSMessageCode.SparseGradientUpdate, sparse_gradient, dst=0, 190 | gradient_version=self.listener.version + 1, lr=lr) 191 | self.version = self.queue.get() 192 | self.idx += 1 193 | return loss -------------------------------------------------------------------------------- /core/server.py: -------------------------------------------------------------------------------- 1 | # 2 | """ 3 | Parameter server for DGS 4 | """ 5 | import logging 6 | import threading 7 | 8 | import torch 9 | import torch.optim 10 | 11 | from core.utils.messaging import MessageCode, MessageListener, send_message, GSMessageCode, \ 12 | GradientMessageListener 13 | from core.utils.serialization import ravel_model_params, ravel_sparse_gradient, unravel_sparse_gradient, \ 14 | server_gradient_filter 15 | 16 | _LOGGER = logging.getLogger(__name__) 17 | cond = threading.Condition() 18 | 19 | 20 | class ParameterServer(MessageListener): 21 | """ParameterServer""" 22 | 23 | def __init__(self, model): 24 | _LOGGER.info("Creating ParameterServer") 25 | print("Creating ParameterServer") 26 | self.parameter_shard = torch.rand(ravel_model_params(model).numel()) 27 | self.model = model 28 | # init superclass 29 | super(ParameterServer, self).__init__(model) 30 | 31 | def receive(self, sender, message_code, parameter): 32 | print("Processing message: {} from sender {}".format(message_code.name, sender)) 33 | 34 | if message_code == MessageCode.ParameterUpdate: 35 | # be sure to clone here 36 | self.parameter_shard = parameter.clone() 37 | 38 | elif message_code == MessageCode.ParameterRequest: 39 | send_message(MessageCode.ParameterUpdate, self.parameter_shard, dst=sender) 40 | 41 | elif message_code == MessageCode.GradientUpdate: 42 | self.parameter_shard.add_(parameter) 43 | 44 | 45 | 46 | un_synced_worker = set() 47 | 48 | global_lr = 0.001 49 | 50 | 51 | class GradientServer(GradientMessageListener): 52 | """GradientServer""" 53 | 54 | def __init__(self, model, rank=0, worker_num=None, global_model=None, synced_model=None, size_list=None, args=None): 55 | _LOGGER.info("Creating GradientServer") 56 | print("Creating GradientServer") 57 | # self.gradient_warehouse = gradient_warehouse 58 | # self.net = model 59 | self.max_version = 0 60 | self.worker_count = 0 61 | self.worker_num = worker_num 62 | self.global_model = global_model 63 | self.global_model.share_memory_() 64 | super(GradientServer, self).__init__(model_size=global_model.numel(), source=rank, args=args) 65 | self.synced_model = synced_model 66 | self.synced_model.share_memory_() 67 | self.synced_version = 0 68 | self.acc_send_grad = synced_model.clone().zero_() 69 | self.acc_send_grad.share_memory_() 70 | self.agg_gradient = None 71 | self.size_list = size_list 72 | self.send_grad = self.acc_send_grad.clone() 73 | self.cuda = self.synced_model.is_cuda 74 | if rank == 1: 75 | for i in range(1, self.worker_num): 76 | self.sync_worker_model(i, 1) 77 | self.node_gradient = {} 78 | 79 | def sync_worker_model(self, sender, version): 80 | send_message(GSMessageCode.ModelUpdate, self.synced_model, dst=sender, gradient_version=version, lr=global_lr) 81 | 82 | def sync_model(self): 83 | self.synced_model.copy_(self.global_model) 84 | # self.synced_version = self 85 | return self.synced_model 86 | 87 | def update(self, rank, version, gradient_update): 88 | """ 89 | :param rank: rank of worker node 90 | :param version: version of gradient 91 | :param gradient_update: tensor, gradient update tensor 92 | :return: 93 | """ 94 | # print("update gradient from rank%d,version%d" % (rank, version)) 95 | self.global_model.add_(-1, gradient_update) 96 | 97 | self.agg_gradient = self.global_model.add(-1, self.synced_model) 98 | 99 | return self.agg_gradient, version 100 | 101 | def receive(self, sender, message_code, gradient_version, lr, parameter): 102 | global un_synced_worker, global_lr 103 | # print("rank {} Processing message: {} from sender {} gradient version {}".format(self.source, message_code.name, 104 | # sender, 105 | # gradient_version)) 106 | self.max_version = max(self.max_version, gradient_version) 107 | if sender == 1: 108 | global_lr = lr 109 | 110 | if message_code == GSMessageCode.GradientUpdate: 111 | if self.cuda: 112 | self.update(sender, gradient_version, parameter.cuda().float()) 113 | else: 114 | self.update(sender, gradient_version, parameter.float()) 115 | 116 | send_message(GSMessageCode.ModelUpdate, self.global_model, dst=sender, 117 | gradient_version=gradient_version) 118 | elif message_code == GSMessageCode.SparseGradientUpdate: 119 | if self.cuda: 120 | send_grad = self.update(sender, gradient_version, unravel_sparse_gradient(parameter).cuda()) 121 | else: 122 | send_grad = self.update(sender, gradient_version, unravel_sparse_gradient(parameter)) 123 | 124 | if sender == 1 and self.max_version % 150 is 1 and gradient_version > 20: 125 | self.sync_model() 126 | un_synced_worker = set(range(1, self.worker_num)) 127 | if sender in un_synced_worker: 128 | self.acc_send_grad.zero_() 129 | self.sync_worker_model(sender, gradient_version) 130 | un_synced_worker.remove(sender) 131 | else: 132 | self.send_grad = self.agg_gradient.add(-1, self.acc_send_grad) 133 | server_gradient_filter(self.size_list, self.send_grad, rate=0.01) 134 | # end = time.time() 135 | 136 | # print(abs(self.send_grad).sum()) 137 | # print('server cal cost time : %f' % (end - start)) 138 | send_message(GSMessageCode.SparseGradientUpdate, ravel_sparse_gradient(self.send_grad), sender, 139 | gradient_version, lr=global_lr) 140 | 141 | self.acc_send_grad.add_(self.send_grad) 142 | 143 | else: 144 | raise Exception('GSMessageCode not implemented') 145 | -------------------------------------------------------------------------------- /core/utils/GradualWarmupScheduler.py: -------------------------------------------------------------------------------- 1 | from torch.optim.lr_scheduler import ReduceLROnPlateau 2 | from torch.optim.lr_scheduler import _LRScheduler 3 | 4 | 5 | class GradualWarmupScheduler(_LRScheduler): 6 | """ Gradually warm-up(increasing) learning rate in optimizer. 7 | Proposed in 'Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour'. 8 | Args: 9 | optimizer (Optimizer): Wrapped optimizer. 10 | multiplier: target learning rate = base lr * multiplier 11 | total_epoch: target learning rate is reached at total_epoch, gradually 12 | after_scheduler: after target_epoch, use this scheduler(eg. ReduceLROnPlateau) 13 | """ 14 | 15 | def __init__(self, optimizer, multiplier, total_epoch, after_scheduler=None): 16 | self.multiplier = multiplier 17 | print('Using warm up lr scheduler') 18 | if self.multiplier <= 1.: 19 | raise ValueError('multiplier should be greater than 1.') 20 | self.total_epoch = total_epoch 21 | self.after_scheduler = after_scheduler 22 | self.finished = False 23 | super().__init__(optimizer) 24 | 25 | def get_lr(self): 26 | if self.last_epoch > self.total_epoch: 27 | if self.after_scheduler: 28 | if not self.finished: 29 | self.after_scheduler.base_lrs = [base_lr * self.multiplier for base_lr in self.base_lrs] 30 | self.finished = True 31 | return self.after_scheduler.get_lr() 32 | return [base_lr * self.multiplier for base_lr in self.base_lrs] 33 | 34 | return [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in 35 | self.base_lrs] 36 | 37 | def step_ReduceLROnPlateau(self, metrics, epoch=None): 38 | if epoch is None: 39 | epoch = self.last_epoch + 1 40 | self.last_epoch = epoch if epoch != 0 else 1 # ReduceLROnPlateau is called at the end of epoch, whereas others are called at beginning 41 | if self.last_epoch <= self.total_epoch: 42 | warmup_lr = [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in 43 | self.base_lrs] 44 | for param_group, lr in zip(self.optimizer.param_groups, warmup_lr): 45 | param_group['lr'] = lr 46 | else: 47 | self.after_scheduler.step(metrics, epoch) 48 | 49 | def step(self, epoch=None, metrics=None): 50 | if type(self.after_scheduler) != ReduceLROnPlateau: 51 | if self.finished and self.after_scheduler: 52 | return self.after_scheduler.step(epoch) 53 | else: 54 | return super(GradualWarmupScheduler, self).step(epoch) 55 | else: 56 | self.step_ReduceLROnPlateau(metrics, epoch) 57 | -------------------------------------------------------------------------------- /core/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yanring/DGS/97e738995eccd771741f87af2251e929788e106e/core/utils/__init__.py -------------------------------------------------------------------------------- /core/utils/constant.py: -------------------------------------------------------------------------------- 1 | MODEL_SIZE = None 2 | -------------------------------------------------------------------------------- /core/utils/log.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # Last Update: 2015/06/30 10:53:52 4 | '''Implements a simple log library. 5 | 6 | This module is a simple encapsulation of logging module to provide a more 7 | convenient interface to write log. The log will both print to stdout and 8 | write to log file. It provides a more flexible way to set the log actions, 9 | and also very simple. See examples showed below: 10 | 11 | Example 1: Use default settings 12 | 13 | import log 14 | log = log.Log(cmdlevel='info') 15 | log.debug('hello, world') 16 | log.info('hello, world') 17 | log.error('hello, world') 18 | log.critical('hello, world') 19 | 20 | Result: 21 | Print all log messages to file, and only print log whose level is greater 22 | than ERROR to stdout. The log file is located in 'xxx.log' if the module 23 | name is xxx.py. The default log file handler is size-rotated, if the log 24 | file's size is greater than 20M, then it will be rotated. 25 | 26 | Example 2: Use set_logger to change settings 27 | 28 | # Change limit size in bytes of default rotating action 29 | log.set_logger(limit = 10240) # 10M 30 | 31 | # Use time-rotated file handler, each day has a different log file, see 32 | # logging.handlers.TimedRotatingFileHandler for more help about 'when' 33 | log.set_logger(when = 'D', limit = 1) 34 | 35 | # Use normal file handler (not rotated) 36 | log.set_logger(backup_count = 0) 37 | 38 | # File log level set to INFO, and stdout log level set to DEBUG 39 | log.set_logger(cmdlevel = 'DEBUG', filelevel = 'INFO') 40 | 41 | # Change default log file name and log mode 42 | log.set_logger(filename = 'yyy.log', mode = 'w') 43 | 44 | # Change default log formatter 45 | log.set_logger(cmdfmt = '[%(levelname)s] %(message)s') 46 | ''' 47 | 48 | __author__ = "Mingo " 49 | __status__ = "Development" 50 | 51 | import logging 52 | import logging.handlers 53 | import os 54 | import sys 55 | import traceback 56 | 57 | 58 | class ColoredFormatter(logging.Formatter): 59 | '''A colorful formatter.''' 60 | 61 | def __init__(self, fmt=None, datefmt=None): 62 | logging.Formatter.__init__(self, fmt, datefmt) 63 | 64 | def format(self, record): 65 | # Color escape string 66 | COLOR_RED = '\033[1;31m' 67 | COLOR_GREEN = '\033[1;32m' 68 | COLOR_YELLOW = '\033[1;33m' 69 | COLOR_BLUE = '\033[1;34m' 70 | COLOR_PURPLE = '\033[1;35m' 71 | COLOR_CYAN = '\033[1;36m' 72 | COLOR_GRAY = '\033[1;37m' 73 | COLOR_WHITE = '\033[1;38m' 74 | COLOR_RESET = '\033[1;0m' 75 | # Define log color 76 | LOG_COLORS = { 77 | 'DEBUG': '%s', 78 | 'INFO': COLOR_GREEN + '%s' + COLOR_RESET, 79 | 'WARNING': COLOR_YELLOW + '%s' + COLOR_RESET, 80 | 'ERROR': COLOR_RED + '%s' + COLOR_RESET, 81 | 'CRITICAL': COLOR_RED + '%s' + COLOR_RESET, 82 | 'EXCEPTION': COLOR_RED + '%s' + COLOR_RESET, 83 | } 84 | level_name = record.levelname 85 | msg = logging.Formatter.format(self, record) 86 | return LOG_COLORS.get(level_name, '%s') % msg 87 | 88 | 89 | class Log(): 90 | def __init__(self, loggername='', filename=None, mode='a', 91 | cmdlevel='DEBUG', 92 | filelevel='INFO', 93 | cmdfmt='[%(asctime)s] %(filename)s line:%(lineno)d %(levelname)-8s%(message)s', 94 | filefmt='[%(asctime)s] %(levelname)-8s%(message)s', 95 | cmddatefmt='%H:%M:%S', 96 | filedatefmt='%Y-%m-%d %H:%M:%S', 97 | backup_count=0, limit=20480, when=None, colorful=False): 98 | self.filename = filename 99 | self.loggername = loggername 100 | if self.filename is None: 101 | self.filename = getattr(sys.modules['__main__'], '__file__', 'log.py') 102 | self.filename = os.path.basename(self.filename.replace('.py', '.log')) 103 | # self.filename = os.path.join('/tmp', self.filename) 104 | if not os.path.exists(os.path.abspath(os.path.dirname(self.filename))): 105 | os.makedirs(os.path.abspath(os.path.dirname(self.filename))) 106 | self.mode = mode 107 | self.cmdlevel = cmdlevel 108 | self.filelevel = filelevel 109 | if isinstance(self.cmdlevel, str): 110 | self.cmdlevel = getattr(logging, self.cmdlevel.upper(), logging.DEBUG) 111 | if isinstance(self.filelevel, str): 112 | self.filelevel = getattr(logging, self.filelevel.upper(), logging.DEBUG) 113 | self.filefmt = filefmt 114 | self.cmdfmt = cmdfmt 115 | self.filedatefmt = filedatefmt 116 | self.cmddatefmt = cmddatefmt 117 | self.backup_count = backup_count 118 | self.limit = limit 119 | self.when = when 120 | self.colorful = colorful 121 | self.logger = None 122 | self.streamhandler = None 123 | self.filehandler = None 124 | if self.cmdlevel > 10: 125 | self.filefmt = '[%(asctime)s] %(levelname)-8s%(message)s' 126 | self.cmdfmt = '[%(asctime)s] %(levelname)-8s%(message)s' 127 | self.cmddatefmt = '%Y-%m-%d %H:%M:%S' 128 | self.set_logger(cmdlevel=self.cmdlevel) 129 | 130 | def init_logger(self): 131 | '''Reload the logger.''' 132 | if self.logger is None: 133 | self.logger = logging.getLogger(self.loggername) 134 | else: 135 | logging.shutdown() 136 | self.logger.handlers = [] 137 | self.streamhandler = None 138 | self.filehandler = None 139 | self.logger.setLevel(logging.DEBUG) 140 | 141 | def add_streamhandler(self): 142 | '''Add a stream handler to the logger.''' 143 | self.streamhandler = logging.StreamHandler() 144 | self.streamhandler.setLevel(self.cmdlevel) 145 | if self.colorful: 146 | formatter = ColoredFormatter(self.cmdfmt, self.cmddatefmt) 147 | else: 148 | formatter = logging.Formatter(self.cmdfmt, self.cmddatefmt, ) 149 | self.streamhandler.setFormatter(formatter) 150 | self.logger.addHandler(self.streamhandler) 151 | 152 | def add_filehandler(self): 153 | '''Add a file handler to the logger.''' 154 | # Choose the filehandler based on the passed arguments 155 | if self.backup_count == 0: # Use FileHandler 156 | self.filehandler = logging.FileHandler(self.filename, self.mode) 157 | elif self.when is None: # Use RotatingFileHandler 158 | self.filehandler = logging.handlers.RotatingFileHandler(self.filename, 159 | self.mode, self.limit, self.backup_count) 160 | else: # Use TimedRotatingFileHandler 161 | self.filehandler = logging.handlers.TimedRotatingFileHandler(self.filename, 162 | self.when, 1, self.backup_count) 163 | self.filehandler.setLevel(self.filelevel) 164 | formatter = logging.Formatter(self.filefmt, self.filedatefmt, ) 165 | self.filehandler.setFormatter(formatter) 166 | self.logger.addHandler(self.filehandler) 167 | 168 | def set_logger(self, **kwargs): 169 | '''Configure the logger.''' 170 | keys = ['mode', 'cmdlevel', 'filelevel', 'filefmt', 'cmdfmt', \ 171 | 'filedatefmt', 'cmddatefmt', 'backup_count', 'limit', \ 172 | 'when', 'colorful'] 173 | for (key, value) in kwargs.items(): 174 | if not (key in keys): 175 | return False 176 | setattr(self, key, value) 177 | if isinstance(self.cmdlevel, str): 178 | self.cmdlevel = getattr(logging, self.cmdlevel.upper(), logging.DEBUG) 179 | if isinstance(self.filelevel, str): 180 | self.filelevel = getattr(logging, self.filelevel.upper(), logging.DEBUG) 181 | if not "cmdfmt" in kwargs: 182 | self.filefmt = '[%(asctime)s] %(filename)s line:%(lineno)d %(levelname)-8s%(message)s' 183 | self.filedatefmt = '%Y-%m-%d %H:%M:%S' 184 | self.cmdfmt = '[%(asctime)s] %(filename)s line:%(lineno)d %(levelname)-8s%(message)s' 185 | self.cmddatefmt = '%H:%M:%S' 186 | if self.cmdlevel > 10: 187 | self.filefmt = '[%(asctime)s] %(levelname)-8s%(message)s' 188 | self.cmdfmt = '[%(asctime)s] %(levelname)-8s%(message)s' 189 | self.cmddatefmt = '%Y-%m-%d %H:%M:%S' 190 | self.init_logger() 191 | self.add_streamhandler() 192 | self.add_filehandler() 193 | # Import the common log functions for convenient 194 | self.import_log_funcs() 195 | return True 196 | 197 | def addFileLog(self, log): 198 | self.logger.addHandler(log.filehandler) 199 | return self 200 | 201 | def import_log_funcs(self): 202 | '''Import the common log functions from the logger to the class''' 203 | log_funcs = ['debug', 'info', 'warning', 'error', 'critical', 204 | 'exception'] 205 | for func_name in log_funcs: 206 | func = getattr(self.logger, func_name) 207 | setattr(self, func_name, func) 208 | 209 | def trace(self): 210 | info = sys.exc_info() 211 | for file, lineno, function, text in traceback.extract_tb(info[2]): 212 | self.error('%s line:%s in %s:%s' % (file, lineno, function, text)) 213 | self.error('%s: %s' % info[:2]) 214 | 215 | 216 | if __name__ == '__main__': 217 | log = Log(cmdlevel='info', colorful=True) 218 | err_log = Log('error', cmdlevel='info', filename='../../err.log', backup_count=1, when='D') 219 | # log = Log(cmdlevel='debug') 220 | log.set_logger(cmdlevel='debug') 221 | # log = log.addFileLog(err_log) 222 | for i in range(100000): 223 | log.debug('debug') 224 | err_log.debug('debug') 225 | log.info('debug%s' % 'haha') 226 | err_log.info('info%s' % 'haha') 227 | log.error((1, 2)) 228 | log.error('debug') 229 | log.info({'a': 1, 'b': 2}) 230 | os.system("pause") 231 | 232 | 233 | class A(): 234 | def __init__(self, log): 235 | self.log = log 236 | 237 | def a(self, a): 238 | self.log.info(a) 239 | 240 | 241 | class B(): 242 | def __init__(self, log): 243 | self.log = log 244 | 245 | def b(self, a): 246 | self.log.info(a) 247 | 248 | 249 | a = A(log) 250 | a.a("test a") 251 | b = B(log) 252 | b.b(5) 253 | 254 | 255 | def fun(a): 256 | return 10 / a 257 | 258 | 259 | try: 260 | a = fun(0) 261 | except: 262 | log.trace() 263 | -------------------------------------------------------------------------------- /core/utils/messaging.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import queue 4 | import socket 5 | import time 6 | from enum import Enum 7 | from multiprocessing.managers import BaseManager 8 | from threading import Thread 9 | 10 | import torch 11 | import torch.distributed as dist 12 | 13 | from core.utils.serialization import ravel_model_params 14 | 15 | _LOGGER = logging.getLogger(__name__) 16 | 17 | isCUDA = 0 18 | manager = None 19 | 20 | 21 | def tail(filename): 22 | with open(filename, 'r') as f: 23 | while True: 24 | line = f.readline() 25 | if not line: 26 | time.sleep(0.01) 27 | continue 28 | if dist.get_rank() == 0: 29 | print('delay:', time.time() - os.stat(filename).st_mtime) 30 | yield int(line) 31 | 32 | 33 | class MessageCode(Enum): 34 | """Different types of messages between client and server that we support go here.""" 35 | ParameterRequest = 0 36 | GradientUpdate = 1 37 | ParameterUpdate = 2 38 | EvaluateParams = 3 39 | 40 | 41 | class GSMessageCode(Enum): 42 | """Different types of messages between client and server that we support go here.""" 43 | GradientRequest = 0 44 | GradientUpdate = 1 45 | # ParameterUpdate = 2 46 | EvaluateParams = 3 47 | ModelRequest = 4 48 | ModelUpdate = 5 49 | SparseGradientUpdate = 6 50 | 51 | 52 | class ModelSize(Enum): 53 | """Different model size""" 54 | AlexNet = 2472266 55 | ResNet18 = 11173962 56 | 57 | 58 | class MessageListener(Thread): 59 | """MessageListener 60 | 61 | base class for message listeners, extends pythons threading Thread 62 | """ 63 | 64 | def __init__(self, model): 65 | """__init__ 66 | 67 | :param model: nn.Module to be defined by the user 68 | """ 69 | self.running = True 70 | self.model = model 71 | _LOGGER.info("Setting m_parameter") 72 | self.m_parameter = torch.zeros(ravel_model_params(model).numel() + 2) 73 | super(MessageListener, self).__init__() 74 | 75 | def receive(self, sender, message_code, parameter): 76 | """receive 77 | 78 | :param sender: rank id of the sender 79 | :param message_code: Enum code 80 | :param parameter: the data payload 81 | """ 82 | raise NotImplementedError() 83 | 84 | def run(self): 85 | _LOGGER.info("Started Running!") 86 | while self.running: 87 | _LOGGER.info("Polling for message...") 88 | dist.recv(tensor=self.m_parameter) 89 | self.receive(int(self.m_parameter[0].item()), 90 | MessageCode(self.m_parameter[1].item()), 91 | self.m_parameter[2:]) 92 | 93 | 94 | a1 = queue.Queue() 95 | b1 = queue.Queue() 96 | a2 = queue.Queue() 97 | b2 = queue.Queue() 98 | a3 = queue.Queue() 99 | b3 = queue.Queue() 100 | a4 = queue.Queue() 101 | b4 = queue.Queue() 102 | a5 = queue.Queue() 103 | b5 = queue.Queue() 104 | a6 = queue.Queue() 105 | b6 = queue.Queue() 106 | a7 = queue.Queue() 107 | b7 = queue.Queue() 108 | a8 = queue.Queue() 109 | b8 = queue.Queue() 110 | a9 = queue.Queue() 111 | b9 = queue.Queue() 112 | a10 = queue.Queue() 113 | b10 = queue.Queue() 114 | a11 = queue.Queue() 115 | b11 = queue.Queue() 116 | a12 = queue.Queue() 117 | b12 = queue.Queue() 118 | a13 = queue.Queue() 119 | b13 = queue.Queue() 120 | a14 = queue.Queue() 121 | b14 = queue.Queue() 122 | a15 = queue.Queue() 123 | b15 = queue.Queue() 124 | a16 = queue.Queue() 125 | b16 = queue.Queue() 126 | a17 = queue.Queue() 127 | b17 = queue.Queue() 128 | a18 = queue.Queue() 129 | b18 = queue.Queue() 130 | a19 = queue.Queue() 131 | b19 = queue.Queue() 132 | a20 = queue.Queue() 133 | b20 = queue.Queue() 134 | a21 = queue.Queue() 135 | b21 = queue.Queue() 136 | a22 = queue.Queue() 137 | b22 = queue.Queue() 138 | a23 = queue.Queue() 139 | b23 = queue.Queue() 140 | a24 = queue.Queue() 141 | b24 = queue.Queue() 142 | a25 = queue.Queue() 143 | b25 = queue.Queue() 144 | a26 = queue.Queue() 145 | b26 = queue.Queue() 146 | a27 = queue.Queue() 147 | b27 = queue.Queue() 148 | a28 = queue.Queue() 149 | b28 = queue.Queue() 150 | a29 = queue.Queue() 151 | b29 = queue.Queue() 152 | a30 = queue.Queue() 153 | b30 = queue.Queue() 154 | a31 = queue.Queue() 155 | b31 = queue.Queue() 156 | a32 = queue.Queue() 157 | b32 = queue.Queue() 158 | 159 | 160 | def rta1(): 161 | return a1 162 | 163 | 164 | def rtb1(): 165 | return b1 166 | 167 | 168 | def rta2(): 169 | return a2 170 | 171 | 172 | def rtb2(): 173 | return b2 174 | 175 | 176 | def rta3(): 177 | return a3 178 | 179 | 180 | def rtb3(): 181 | return b3 182 | 183 | 184 | def rta4(): 185 | return a4 186 | 187 | 188 | def rtb4(): 189 | return b4 190 | 191 | 192 | def rta5(): 193 | return a5 194 | 195 | 196 | def rtb5(): 197 | return b5 198 | 199 | 200 | def rta6(): 201 | return a6 202 | 203 | 204 | def rtb6(): 205 | return b6 206 | 207 | 208 | def rta7(): 209 | return a7 210 | 211 | 212 | def rtb7(): 213 | return b7 214 | 215 | 216 | def rta8(): 217 | return a8 218 | 219 | 220 | def rtb8(): 221 | return b8 222 | 223 | 224 | def rta9(): 225 | return a9 226 | 227 | 228 | def rtb9(): 229 | return b9 230 | 231 | 232 | def rta10(): 233 | return a10 234 | 235 | 236 | def rtb10(): 237 | return b10 238 | 239 | 240 | def rta11(): 241 | return a11 242 | 243 | 244 | def rtb11(): 245 | return b11 246 | 247 | 248 | def rta12(): 249 | return a12 250 | 251 | 252 | def rtb12(): 253 | return b12 254 | 255 | 256 | def rta13(): 257 | return a13 258 | 259 | 260 | def rtb13(): 261 | return b13 262 | 263 | 264 | def rta14(): 265 | return a14 266 | 267 | 268 | def rtb14(): 269 | return b14 270 | 271 | 272 | def rta15(): 273 | return a15 274 | 275 | 276 | def rtb15(): 277 | return b15 278 | 279 | 280 | def rta16(): 281 | return a16 282 | 283 | 284 | def rtb16(): 285 | return b16 286 | 287 | 288 | def rta17(): 289 | return a17 290 | 291 | 292 | def rtb17(): 293 | return b17 294 | 295 | 296 | def rta18(): 297 | return a18 298 | 299 | 300 | def rtb18(): 301 | return b18 302 | 303 | 304 | def rta19(): 305 | return a19 306 | 307 | 308 | def rtb19(): 309 | return b19 310 | 311 | 312 | def rta20(): 313 | return a20 314 | 315 | 316 | def rtb20(): 317 | return b20 318 | 319 | 320 | def rta21(): 321 | return a21 322 | 323 | 324 | def rtb21(): 325 | return b21 326 | 327 | 328 | def rta22(): 329 | return a22 330 | 331 | 332 | def rtb22(): 333 | return b22 334 | 335 | 336 | def rta23(): 337 | return a23 338 | 339 | 340 | def rtb23(): 341 | return b23 342 | 343 | 344 | def rta24(): 345 | return a24 346 | 347 | 348 | def rtb24(): 349 | return b24 350 | 351 | 352 | def rta25(): 353 | return a25 354 | 355 | 356 | def rtb25(): 357 | return b25 358 | 359 | 360 | def rta26(): 361 | return a26 362 | 363 | 364 | def rtb26(): 365 | return b26 366 | 367 | 368 | def rta27(): 369 | return a27 370 | 371 | 372 | def rtb27(): 373 | return b27 374 | 375 | 376 | def rta28(): 377 | return a28 378 | 379 | 380 | def rtb28(): 381 | return b28 382 | 383 | 384 | def rta29(): 385 | return a29 386 | 387 | 388 | def rtb29(): 389 | return b29 390 | 391 | 392 | def rta30(): 393 | return a30 394 | 395 | 396 | def rtb30(): 397 | return b30 398 | 399 | 400 | def rta31(): 401 | return a31 402 | 403 | 404 | def rtb31(): 405 | return b31 406 | 407 | 408 | def rta32(): 409 | return a32 410 | 411 | 412 | def rtb32(): 413 | return b32 414 | 415 | 416 | class GradientMessageListener(Thread): 417 | """MessageListener 418 | 419 | base class for message listeners, extends pythons threading Thread 420 | """ 421 | 422 | def __init__(self, model_size, source=0, args=None): 423 | """__init__ 424 | 425 | :param model: nn.Module to be defined by the user 426 | """ 427 | # self.model = model 428 | self.source = source 429 | _LOGGER.info("Setting m_parameter") 430 | self.m_parameter = torch.zeros(model_size + 4).double() 431 | self.cached_stamp = 0 432 | self.size_filename = None 433 | self.manager = None 434 | self.args = args 435 | if dist.get_rank() == 0 and self.source == 1: 436 | self.init_server_queue_manager() 437 | elif dist.get_rank() > 0: 438 | self.recv_queue, self.send_queue = self.init_worker_queue_manager() 439 | super(GradientMessageListener, self).__init__() 440 | 441 | def receive(self, sender, message_code, gradient_version, lr, parameter): 442 | """receive 443 | 444 | :param lr: 445 | :param gradient_version: 446 | :param sender: rank id of the sender 447 | :param message_code: Enum code 448 | :param parameter: the data payload 449 | """ 450 | raise NotImplementedError() 451 | 452 | def run(self): 453 | # for sparse gradient transmission 454 | _LOGGER.info("Started Running!") 455 | self.running = True 456 | while self.running: 457 | _LOGGER.info("Polling for sparse message...") 458 | # for size in tail(self.size_filename): 459 | while True: 460 | size = QueueManager.get_size(self.source) 461 | # if dist.get_rank() == 0: 462 | # print('RECEIVING MESSAGE %dto%d.size:%d,' % ( 463 | # self.source, dist.get_rank(), size)) 464 | self.m_parameter = torch.zeros(size + 4).double() 465 | try: 466 | sender = dist.recv(tensor=self.m_parameter, src=self.source) 467 | except Exception as e: 468 | # print('Exception :', e) 469 | raise e 470 | time.sleep(0.5) 471 | continue 472 | self.m_parameter = self.m_parameter 473 | # if dist.get_rank() == 0: 474 | # print('run',self.m_parameter[int(len(self.m_parameter) / 2)-3:int(len(self.m_parameter) / 2)+2],self.m_parameter[int(len(self.m_parameter) / 2)-3:int(len(self.m_parameter) / 2)+2].long()) 475 | self.receive(int(self.m_parameter[0].item()), 476 | GSMessageCode(self.m_parameter[1].item()), 477 | int(self.m_parameter[2].item()), 478 | float(self.m_parameter[3].item()), 479 | self.m_parameter[4:]) 480 | 481 | def init_server_queue_manager(self): 482 | 483 | QueueManager.register('from0to1', callable=rta1) 484 | QueueManager.register('from1to0', callable=rtb1) 485 | QueueManager.register('from0to2', callable=rta2) 486 | QueueManager.register('from2to0', callable=rtb2) 487 | QueueManager.register('from0to3', callable=rta3) 488 | QueueManager.register('from3to0', callable=rtb3) 489 | QueueManager.register('from0to4', callable=rta4) 490 | QueueManager.register('from4to0', callable=rtb4) 491 | QueueManager.register('from0to5', callable=rta5) 492 | QueueManager.register('from5to0', callable=rtb5) 493 | QueueManager.register('from0to6', callable=rta6) 494 | QueueManager.register('from6to0', callable=rtb6) 495 | QueueManager.register('from0to7', callable=rta7) 496 | QueueManager.register('from7to0', callable=rtb7) 497 | QueueManager.register('from0to8', callable=rta8) 498 | QueueManager.register('from8to0', callable=rtb8) 499 | QueueManager.register('from0to9', callable=rta9) 500 | QueueManager.register('from9to0', callable=rtb9) 501 | QueueManager.register('from0to10', callable=rta10) 502 | QueueManager.register('from10to0', callable=rtb10) 503 | QueueManager.register('from0to11', callable=rta11) 504 | QueueManager.register('from11to0', callable=rtb11) 505 | QueueManager.register('from0to12', callable=rta12) 506 | QueueManager.register('from12to0', callable=rtb12) 507 | QueueManager.register('from0to13', callable=rta13) 508 | QueueManager.register('from13to0', callable=rtb13) 509 | QueueManager.register('from0to14', callable=rta14) 510 | QueueManager.register('from14to0', callable=rtb14) 511 | QueueManager.register('from0to15', callable=rta15) 512 | QueueManager.register('from15to0', callable=rtb15) 513 | QueueManager.register('from0to16', callable=rta16) 514 | QueueManager.register('from16to0', callable=rtb16) 515 | QueueManager.register('from0to17', callable=rta17) 516 | QueueManager.register('from17to0', callable=rtb17) 517 | QueueManager.register('from0to18', callable=rta18) 518 | QueueManager.register('from18to0', callable=rtb18) 519 | QueueManager.register('from0to19', callable=rta19) 520 | QueueManager.register('from19to0', callable=rtb19) 521 | QueueManager.register('from0to20', callable=rta20) 522 | QueueManager.register('from20to0', callable=rtb20) 523 | QueueManager.register('from0to21', callable=rta21) 524 | QueueManager.register('from21to0', callable=rtb21) 525 | QueueManager.register('from0to22', callable=rta22) 526 | QueueManager.register('from22to0', callable=rtb22) 527 | QueueManager.register('from0to23', callable=rta23) 528 | QueueManager.register('from23to0', callable=rtb23) 529 | QueueManager.register('from0to24', callable=rta24) 530 | QueueManager.register('from24to0', callable=rtb24) 531 | QueueManager.register('from0to25', callable=rta25) 532 | QueueManager.register('from25to0', callable=rtb25) 533 | QueueManager.register('from0to26', callable=rta26) 534 | QueueManager.register('from26to0', callable=rtb26) 535 | QueueManager.register('from0to27', callable=rta27) 536 | QueueManager.register('from27to0', callable=rtb27) 537 | QueueManager.register('from0to28', callable=rta28) 538 | QueueManager.register('from28to0', callable=rtb28) 539 | QueueManager.register('from0to29', callable=rta29) 540 | QueueManager.register('from29to0', callable=rtb29) 541 | QueueManager.register('from0to30', callable=rta30) 542 | QueueManager.register('from30to0', callable=rtb30) 543 | QueueManager.register('from0to31', callable=rta31) 544 | QueueManager.register('from31to0', callable=rtb31) 545 | QueueManager.register('from0to32', callable=rta32) 546 | QueueManager.register('from32to0', callable=rtb32) 547 | 548 | self.manager = QueueManager(address=('', 5000), authkey=b'abc') 549 | QueueManager.send_queue_list.append(0) 550 | QueueManager.recv_queue_list.append(0) 551 | QueueManager.manager = self.manager 552 | self.manager.start() 553 | for i in range(1, dist.get_world_size()): 554 | send_queue = eval('self.manager.from0to%d' % i)() 555 | QueueManager.send_queue_list.append(send_queue) 556 | recv_queue = eval('self.manager.from%dto0' % i)() 557 | QueueManager.recv_queue_list.append(recv_queue) 558 | pass 559 | 560 | def init_worker_queue_manager(self): 561 | # time.sleep(1) 562 | QueueManager.register('from0to%d' % dist.get_rank()) 563 | QueueManager.register('from%dto0' % dist.get_rank()) 564 | time.sleep(10) 565 | # self.manager = QueueManager(address=(socket.gethostbyname('localhost'), 5000), authkey=b'abc') 566 | if socket.gethostname() == 'yan-pc' or socket.gethostname() == 'yrx-MS-7A93' or 'ubuntu' in socket.gethostname(): 567 | print('queue init in 522') 568 | # self.manager = QueueManager(address=('172.18.166.108', 5000), authkey=b'abc') 569 | self.manager = QueueManager(address=(self.args.master, 5000), authkey=b'abc') 570 | else: 571 | time.sleep(10) 572 | print('queue init in th') 573 | self.manager = QueueManager(address=(self.args.master, 5000), authkey=b'abc') 574 | try: 575 | self.manager.connect() 576 | except Exception as e: 577 | print(e) 578 | time.sleep(10) 579 | self.manager.connect() 580 | send_queue = eval('self.manager.from%dto0' % dist.get_rank())() 581 | QueueManager.send_queue_list.append(send_queue) 582 | recv_queue = eval('self.manager.from0to%d' % dist.get_rank())() 583 | QueueManager.recv_queue_list.append(recv_queue) 584 | QueueManager.manager = self.manager 585 | return recv_queue, send_queue 586 | 587 | 588 | class QueueManager(BaseManager): 589 | manager = None 590 | send_queue_list = [] 591 | recv_queue_list = [] 592 | 593 | @classmethod 594 | def get_manager(cls): 595 | return cls.manager 596 | 597 | @classmethod 598 | def get_size(cls, opposite): 599 | recv_queue = cls.recv_queue_list[opposite] 600 | # exec('recv_queue = cls.manager.from%dto%d()' % (source, target)) 601 | res = None 602 | try: 603 | res = recv_queue.get(timeout=4000) 604 | except queue.Empty: 605 | print('task queue is empty') 606 | # print('RECV ', res, type(recv_queue), recv_queue) 607 | return int(res) 608 | 609 | @classmethod 610 | def put_size(cls, opposite, size): 611 | # send_queue = None 612 | # exec('send_queue = cls.manager.from%dto%d()' % (source, target)) 613 | send_queue = cls.send_queue_list[opposite] 614 | # print('SEND ', type(send_queue), size, send_queue) 615 | send_queue.put(size) 616 | 617 | 618 | def send_message(message_code, payload, dst=0, gradient_version=None, lr=0.1): 619 | """Sends a message to a destination 620 | Concatenates both the message code and destination with the payload into a single tensor and then sends that as a tensor 621 | """ 622 | # _LOGGER.info("SENDING MESSAGE: {} RANK: {}".format(message_code, dist.get_rank())) 623 | m_parameter = torch.Tensor([dist.get_rank(), message_code.value, gradient_version, lr]) 624 | # print(m_parameter.size(), payload.size()) 625 | if payload.is_cuda: 626 | payload = payload.cpu() 627 | size = str(payload.numel()) 628 | payload = torch.cat((m_parameter.double(), payload.double())) 629 | if dist.get_rank() == 0: 630 | print('%s SENDING MESSAGE %s gradient_version %d, %dto%d.size:%d' % ( 631 | str(time.time()), message_code, gradient_version, dist.get_rank(), dst, payload.numel())) 632 | # with open('%dto%d.size' % (dist.get_rank(), dst), 'a') as f: 633 | # f.write(size) 634 | QueueManager.put_size(dst, size) 635 | dist.send(tensor=payload, dst=dst) 636 | -------------------------------------------------------------------------------- /core/utils/serialization.py: -------------------------------------------------------------------------------- 1 | import time 2 | 3 | import torch 4 | 5 | from core.utils import constant 6 | 7 | current_model_size = None 8 | 9 | 10 | def ravel_model_params(model, grads=False, cuda=False): 11 | """ 12 | Squash model parameters or gradients into a single tensor. 13 | """ 14 | if next(model.parameters()).is_cuda: 15 | m_parameter = torch.Tensor([0]).cuda() 16 | else: 17 | m_parameter = torch.Tensor([0]) 18 | for parameter in list(model.parameters()): 19 | if grads: 20 | m_parameter = torch.cat((m_parameter, parameter.grad.view(-1))) 21 | else: 22 | m_parameter = torch.cat((m_parameter, parameter.data.view(-1))) 23 | return m_parameter[1:] 24 | 25 | 26 | def unravel_model_params(model, parameter_update): 27 | """ 28 | Assigns parameter_update params to model.parameters. 29 | This is done by iterating through model.parameters() and assigning the relevant params in parameter_update. 30 | NOTE: this function manipulates model.parameters. 31 | """ 32 | current_index = 0 # keep track of where to read from parameter_update 33 | for parameter in model.parameters(): 34 | numel = parameter.data.numel() 35 | size = parameter.data.size() 36 | parameter.data.copy_(parameter_update[current_index:current_index + numel].view(size)) 37 | current_index += numel 38 | 39 | 40 | def update_model_params(model, parameter_update, lr): 41 | """ 42 | Assigns parameter_update params to model.parameters. 43 | This is done by iterating through model.parameters() and adding the gradient in parameter_update. 44 | NOTE: this function manipulates model.parameters. 45 | """ 46 | current_index = 0 # keep track of where to read from parameter_update 47 | for parameter in model.parameters(): 48 | numel = parameter.data.numel() 49 | size = parameter.data.size() 50 | # print(parameter.data.device,parameter_update.device) 51 | parameter.data.add_(-lr, parameter_update[current_index:current_index + numel].view(size)) 52 | current_index += numel 53 | 54 | 55 | def worker_gradient_executor(net, payload, u_kt, v_kt, rate=0.01, lr=0.1, momentum=None, weight_decay=0): 56 | """ 57 | :param momentum: 58 | :param lr: 59 | :param v_kt: 60 | :param payload: 61 | :param u_kt: 62 | :param net: model 63 | :param rate: compression rate 64 | :return: gradients which lager than threshold 65 | """ 66 | # start = time.time() 67 | current_index = 0 68 | u_kt.mul_(momentum) 69 | for param in net.parameters(): 70 | numel = param.data.numel() 71 | layer_u_kt = u_kt[current_index:current_index + numel] 72 | if weight_decay != 0: 73 | param.grad.data.add_(weight_decay, param.data) 74 | layer_u_kt.add_(param.grad.data.view(-1).mul(lr)) 75 | k = int(numel * rate) if int(numel * rate) != 0 else 1 76 | k = numel - k 77 | abs_layer_u_kt = layer_u_kt.abs() 78 | threshold = torch.kthvalue(abs_layer_u_kt, k).values 79 | mask = abs_layer_u_kt.gt(threshold).float() 80 | # print(mask.sum()-len(layer_u_kt)) 81 | payload[current_index:current_index + numel].copy_(layer_u_kt.mul(mask)) 82 | layer_u_kt.add_(layer_u_kt.mul(1 - mask).mul(1 / momentum - 1)) 83 | # print(layer_u_kt.sum()) 84 | current_index += numel 85 | # end = time.time() 86 | return payload 87 | 88 | 89 | def DGC(net, payload, u_kt, v_kt, rate=0.01, lr=0.1, momentum=None, weight_decay=None): 90 | """ 91 | :param momentum: 92 | :param lr: 93 | :param v_kt: 94 | :param payload: 95 | :param u_kt: 96 | :param net: model 97 | :param rate: compression rate 98 | :return: gradients which lager than threshold 99 | """ 100 | start = time.time() 101 | current_index = 0 102 | u_kt.mul_(momentum) 103 | sum = 0 104 | for param in net.parameters(): 105 | numel = param.data.numel() 106 | layer_u_kt = u_kt[current_index:current_index + numel] 107 | layer_v_kt = v_kt[current_index:current_index + numel] 108 | if weight_decay != 0: 109 | param.grad.data.add_(weight_decay, param.data) 110 | layer_u_kt.add_(param.grad.data.view(-1)) 111 | layer_v_kt.add_(layer_u_kt) 112 | k = int(numel * rate) if int(numel * rate) != 0 else 1 113 | topn = [[1.0]] 114 | try: 115 | topn = torch.topk(abs(layer_v_kt), k) 116 | except Exception as e: 117 | print(e) 118 | print(k, layer_v_kt.nelement()) 119 | # print(layer_v_kt) 120 | threshold = float(topn[0][-1]) 121 | mask = (abs(layer_v_kt) > threshold).float() 122 | payload[current_index:current_index + numel].copy_(layer_v_kt.mul(mask).mul(lr)) 123 | layer_v_kt.mul_(1 - mask) 124 | layer_u_kt.mul_(1 - mask) 125 | current_index += numel 126 | return payload 127 | 128 | 129 | def Aji(net, payload, u_kt, v_kt, rate=0.01, lr=0.1, momentum=None, weight_decay=0): 130 | """ 131 | :param momentum: 132 | :param lr: 133 | :param v_kt: 134 | :param payload: 135 | :param u_kt: 136 | :param net: model 137 | :param rate: compression rate 138 | :return: gradients which lager than threshold 139 | """ 140 | start = time.time() 141 | current_index = 0 142 | # u_kt.mul_(momentum) 143 | for param in net.parameters(): 144 | numel = param.data.numel() 145 | # layer_u_kt = u_kt[current_index:current_index + numel] 146 | layer_v_kt = v_kt[current_index:current_index + numel] 147 | # layer_u_kt.add_(param.grad.data.view(-1)) 148 | if weight_decay != 0: 149 | param.grad.data.add_(weight_decay, param.data) 150 | layer_v_kt.add_(param.grad.data.view(-1).mul(lr)) 151 | k = int(numel * rate) if int(numel * rate) != 0 else 1 152 | topn = [[1.0]] 153 | try: 154 | topn = torch.topk(abs(layer_v_kt), k) 155 | except Exception as e: 156 | print(e) 157 | print(k, layer_v_kt.nelement()) 158 | threshold = float(topn[0][-1]) 159 | mask = (abs(layer_v_kt) > threshold).float() 160 | payload[current_index:current_index + numel].copy_(layer_v_kt.mul(mask)) 161 | layer_v_kt.mul_(1 - mask) 162 | # layer_u_kt.mul_(1 - mask) 163 | current_index += numel 164 | return payload 165 | 166 | 167 | 168 | def server_gradient_filter(size_list, gradients, rate=0.01): 169 | # print('gradients', gradients) 170 | current_index = 0 171 | for size in size_list: 172 | numel = size 173 | temp = gradients[current_index:current_index + numel] 174 | current_index += numel 175 | k = int(numel * rate) if int(numel * rate) != 0 else 1 176 | k = numel - k 177 | abs_temp = temp.abs() 178 | threshold = torch.kthvalue(abs_temp, k).values 179 | mask = abs_temp.gt(threshold) 180 | temp.mul_(mask) 181 | return gradients 182 | 183 | 184 | def ravel_sparse_gradient(temp_param): 185 | indices = temp_param.nonzero() 186 | values = temp_param[indices] 187 | sparse_gradient = torch.cat((indices.double(), values.double())).view(-1) 188 | return sparse_gradient 189 | 190 | 191 | def unravel_sparse_gradient(sparse_gradient): 192 | # len is 2472266 11173962 2400w 193 | split = int(len(sparse_gradient) / 2) 194 | i = sparse_gradient[:split] 195 | v = sparse_gradient[split:] 196 | size = torch.Size([constant.MODEL_SIZE]) 197 | # print('3',v.sum()) 198 | try: 199 | dense_gradient = torch.sparse_coo_tensor(i.reshape(1, -1).long(), v.float(), size, device=torch.device('cuda')) 200 | except Exception as e: 201 | print(i, v) 202 | print('sum indice', sum(i)) 203 | raise (e) 204 | print(size, constant.MODEL_SIZE, i[-5:], v[-5:]) 205 | dense_gradient = torch.FloatTensor(size).zero_() 206 | 207 | # print(dense_gradient.sum()) 208 | return dense_gradient 209 | -------------------------------------------------------------------------------- /example/ImageNet_dali_dataloader.py: -------------------------------------------------------------------------------- 1 | import time 2 | 3 | import nvidia.dali.ops as ops 4 | import nvidia.dali.types as types 5 | import torch.utils.data 6 | import torchvision.datasets as datasets 7 | import torchvision.transforms as transforms 8 | from nvidia.dali.pipeline import Pipeline 9 | from nvidia.dali.plugin.pytorch import DALIClassificationIterator 10 | 11 | 12 | class HybridTrainPipe(Pipeline): 13 | def __init__(self, batch_size, num_threads, device_id, data_dir, crop, dali_cpu=False, local_rank=0, world_size=1): 14 | super(HybridTrainPipe, self).__init__(batch_size, num_threads, device_id, seed=12 + device_id) 15 | dali_device = "gpu" 16 | self.input = ops.FileReader(file_root=data_dir, shard_id=local_rank, num_shards=world_size, random_shuffle=True) 17 | self.decode = ops.ImageDecoder(device="mixed", output_type=types.RGB) 18 | self.res = ops.RandomResizedCrop(device="gpu", size=crop, random_area=[0.08, 1.25]) 19 | self.cmnp = ops.CropMirrorNormalize(device="gpu", 20 | output_dtype=types.FLOAT, 21 | output_layout=types.NCHW, 22 | image_type=types.RGB, 23 | mean=[0.485 * 255, 0.456 * 255, 0.406 * 255], 24 | std=[0.229 * 255, 0.224 * 255, 0.225 * 255]) 25 | self.coin = ops.CoinFlip(probability=0.5) 26 | print('DALI "{0}" variant'.format(dali_device)) 27 | 28 | def define_graph(self): 29 | rng = self.coin() 30 | self.jpegs, self.labels = self.input(name="Reader") 31 | images = self.decode(self.jpegs) 32 | images = self.res(images) 33 | output = self.cmnp(images, mirror=rng) 34 | return [output, self.labels] 35 | 36 | 37 | class HybridValPipe(Pipeline): 38 | def __init__(self, batch_size, num_threads, device_id, data_dir, crop, size, local_rank=0, world_size=1): 39 | super(HybridValPipe, self).__init__(batch_size, num_threads, device_id, seed=12 + device_id) 40 | self.input = ops.FileReader(file_root=data_dir, shard_id=local_rank, num_shards=world_size, 41 | random_shuffle=False) 42 | self.decode = ops.ImageDecoder(device="mixed", output_type=types.RGB) 43 | self.res = ops.Resize(device="gpu", resize_shorter=size, interp_type=types.INTERP_TRIANGULAR) 44 | self.cmnp = ops.CropMirrorNormalize(device="gpu", 45 | output_dtype=types.FLOAT, 46 | output_layout=types.NCHW, 47 | crop=(crop, crop), 48 | image_type=types.RGB, 49 | mean=[0.485 * 255, 0.456 * 255, 0.406 * 255], 50 | std=[0.229 * 255, 0.224 * 255, 0.225 * 255]) 51 | 52 | def define_graph(self): 53 | self.jpegs, self.labels = self.input(name="Reader") 54 | images = self.decode(self.jpegs) 55 | images = self.res(images) 56 | output = self.cmnp(images) 57 | return [output, self.labels] 58 | 59 | 60 | def get_imagenet_iter_dali(type, image_dir, batch_size, num_threads, device_id, num_gpus, crop, val_size=256, 61 | world_size=1, 62 | local_rank=0): 63 | if type == 'train': 64 | pip_train = HybridTrainPipe(batch_size=batch_size, num_threads=num_threads, device_id=device_id, 65 | data_dir=image_dir + '/train', 66 | crop=crop, world_size=world_size, local_rank=local_rank) 67 | pip_train.build() 68 | dali_iter_train = DALIClassificationIterator(pip_train, size=pip_train.epoch_size("Reader") // world_size) 69 | return dali_iter_train 70 | elif type == 'val': 71 | pip_val = HybridValPipe(batch_size=batch_size, num_threads=num_threads, device_id=device_id, 72 | data_dir=image_dir + '/val', 73 | crop=crop, size=val_size, world_size=world_size, local_rank=local_rank) 74 | pip_val.build() 75 | dali_iter_val = DALIClassificationIterator(pip_val, size=pip_val.epoch_size("Reader") // world_size) 76 | return dali_iter_val 77 | 78 | 79 | def get_imagenet_iter_torch(type, image_dir, batch_size, num_threads, device_id, num_gpus, crop, val_size=256, 80 | world_size=1, local_rank=0): 81 | if type == 'train': 82 | transform = transforms.Compose([ 83 | transforms.RandomResizedCrop(crop, scale=(0.08, 1.25)), 84 | transforms.RandomHorizontalFlip(), 85 | transforms.ToTensor(), 86 | transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), 87 | ]) 88 | dataset = datasets.ImageFolder(image_dir + '/train', transform) 89 | dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=num_threads, 90 | pin_memory=True) 91 | else: 92 | transform = transforms.Compose([ 93 | transforms.Resize(val_size), 94 | transforms.CenterCrop(crop), 95 | transforms.ToTensor(), 96 | transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), 97 | ]) 98 | dataset = datasets.ImageFolder(image_dir + '/val', transform) 99 | dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=num_threads, 100 | pin_memory=True) 101 | return dataloader 102 | 103 | 104 | if __name__ == '__main__': 105 | train_loader = get_imagenet_iter_dali(type='train', image_dir='/userhome/memory_data/imagenet', batch_size=256, 106 | num_threads=4, crop=224, device_id=0, num_gpus=1) 107 | print('start iterate') 108 | start = time.time() 109 | for i, data in enumerate(train_loader): 110 | images = data[0]["data"].cuda(non_blocking=True) 111 | labels = data[0]["label"].squeeze().long().cuda(non_blocking=True) 112 | end = time.time() 113 | print('end iterate') 114 | print('dali iterate time: %fs' % (end - start)) 115 | 116 | train_loader = get_imagenet_iter_torch(type='train', image_dir='/userhome/data/imagenet', batch_size=256, 117 | num_threads=4, crop=224, device_id=0, num_gpus=1) 118 | print('start iterate') 119 | start = time.time() 120 | for i, data in enumerate(train_loader): 121 | images = data[0].cuda(non_blocking=True) 122 | labels = data[1].cuda(non_blocking=True) 123 | end = time.time() 124 | print('end iterate') 125 | print('torch iterate time: %fs' % (end - start)) 126 | -------------------------------------------------------------------------------- /example/Imagenet_dist.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | import random 4 | import shutil 5 | import socket 6 | import sys 7 | import time 8 | import warnings 9 | from datetime import datetime 10 | 11 | import pandas 12 | import torch 13 | import torch.backends.cudnn as cudnn 14 | import torch.distributed as dist 15 | import torch.multiprocessing as mp 16 | import torch.nn as nn 17 | import torch.nn.parallel 18 | import torch.optim 19 | import torch.utils.data 20 | import torch.utils.data.distributed 21 | import torchvision.datasets as datasets 22 | import torchvision.models as models 23 | import torchvision.transforms as transforms 24 | from PIL import ImageFile 25 | 26 | # if 'gpu' in socket.gethostname(): 27 | # print('network in th v100') 28 | # os.environ['GLOO_SOCKET_IFNAME'] = 'enp183s0f0' 29 | # else: 30 | # os.environ['GLOO_SOCKET_IFNAME'] = 'enp3s0' 31 | 32 | WORKPATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) 33 | sys.path.append(WORKPATH) 34 | from core.utils import constant 35 | from core.utils.serialization import ravel_model_params 36 | 37 | from core.optim import GradientSGD 38 | from example.main import init_server 39 | import _pickle as pickle 40 | 41 | ImageFile.LOAD_TRUNCATED_IMAGES = True 42 | model_names = sorted(name for name in models.__dict__ 43 | if name.islower() and not name.startswith("__") 44 | and callable(models.__dict__[name])) 45 | 46 | parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') 47 | parser.add_argument('-data', default='/home/yan/tmpfs/data/', 48 | help='path to dataset') 49 | parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet18', 50 | choices=model_names, 51 | help='model architecture: ' + 52 | ' | '.join(model_names) + 53 | ' (default: mobilenet_v2/resnet18)') 54 | parser.add_argument('-j', '--workers', default=8, type=int, metavar='N', 55 | help='number of data loading workers (default: 4)') 56 | parser.add_argument('--epochs', default=90, type=int, metavar='N', 57 | help='number of total epochs to run') 58 | parser.add_argument('--start-epoch', default=0, type=int, metavar='N', 59 | help='manual epoch number (useful on restarts)') 60 | parser.add_argument('-b', '--batch-size', default=256, type=int, 61 | metavar='N', 62 | help='mini-batch size (default: 256), this is the total ' 63 | 'batch size of all GPUs on the current node when ' 64 | 'using Data Parallel or Distributed Data Parallel') 65 | parser.add_argument('--lr', '--learning-rate', default=0.1, type=float, 66 | metavar='LR', help='initial learning rate', dest='lr') 67 | parser.add_argument('--momentum', deNonefault=0.7, type=float, metavar='M', 68 | help='momentum') 69 | parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float, 70 | metavar='W', help='weight decay (default: 1e-4)', 71 | dest='weight_decay') 72 | parser.add_argument('-p', '--print-freq', default=100, type=int, 73 | metavar='N', help='print frequency (default: 10)') 74 | parser.add_argument('--resume', default='', type=str, metavar='PATH', 75 | help='path to latest checkpoint (default: none)') 76 | parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true', 77 | help='evaluate model on validation set') 78 | parser.add_argument('--pretrained', dest='pretrained', action='store_true', 79 | help='use pre-trained model') 80 | parser.add_argument('--world-size', default=-1, type=int, 81 | help='number of nodes for distributed training') 82 | parser.add_argument('--rank', default=-1, type=int, 83 | help='node rank for distributed training') 84 | parser.add_argument('--dist-url', default='tcp://224.66.41.62:23456', type=str, 85 | help='url used to set up distributed training') 86 | parser.add_argument('--dist-backend', default='nccl', type=str, 87 | help='distributed backend') 88 | parser.add_argument('--seed', default=None, type=int, 89 | help='seed for initializing training. ') 90 | parser.add_argument('--multiprocessing-distributed', action='store_true', 91 | help='Use multi-processing distributed training to launch ' 92 | 'N processes per node, which has N GPUs. This is the ' 93 | 'fastest way to use PyTorch for either single node or ' 94 | 'multi node data parallel training') 95 | parser.add_argument('--network-interface', type=str, default='enp3s0', 96 | help='By default, Gloo backends will try to find the right network interface to use. ' 97 | 'If the automatically detected interface is not correct, you can override it ') 98 | 99 | # my settings 100 | parser.add_argument('--mode', type=str, default='gradient_sgd', help='gradient_sgd, dgc, Aji or asgd') 101 | parser.add_argument('--cuda', action='store_true', default=True, help='use CUDA for training') 102 | parser.add_argument('--no-distributed', action='store_true', default=False, 103 | help='distributed or local') 104 | 105 | logs = [] 106 | 107 | best_acc1 = 0 108 | 109 | 110 | def main(): 111 | args = parser.parse_args() 112 | os.environ['GLOO_SOCKET_IFNAME'] = args.network_interface 113 | if socket.gethostname() == 'yan-pc': 114 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 1) 115 | elif 'gn' in socket.gethostname(): 116 | print('init in K80') 117 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 4) 118 | elif 'gpu' in socket.gethostname(): 119 | print('init in V100') 120 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 4) 121 | else: 122 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 1) 123 | # os.environ['CUDA_VISIBLE_DEVICES'] = '1' 124 | print('Using device%s, device count:%d' % (os.environ['CUDA_VISIBLE_DEVICES'], torch.cuda.device_count())) 125 | if args.seed is not None: 126 | random.seed(args.seed) 127 | torch.manual_seed(args.seed) 128 | cudnn.deterministic = True 129 | warnings.warn('You have chosen to seed training. ' 130 | 'This will turn on the CUDNN deterministic setting, ' 131 | 'which can slow down your training considerably! ' 132 | 'You may see unexpected behavior when restarting ' 133 | 'from checkpoints.') 134 | 135 | if args.rank == 0: 136 | print("=> creating server '{}'".format(args.arch)) 137 | model = models.__dict__[args.arch]() 138 | init_server(args, model) 139 | exit(999) 140 | 141 | 142 | args.gpu = args.rank % torch.cuda.device_count() 143 | if args.gpu is not None: 144 | warnings.warn('You have chosen a specific GPU. This will completely ' 145 | 'disable data parallelism.') 146 | 147 | if args.dist_url == "env://" and args.world_size == -1: 148 | args.world_size = int(os.environ["WORLD_SIZE"]) 149 | 150 | args.distributed = False 151 | # args.distributed = args.world_size > 1 or args.multiprocessing_distributed 152 | ngpus_per_node = torch.cuda.device_count() 153 | if args.multiprocessing_distributed: 154 | # Since we have ngpus_per_node processes per node, the total world_size 155 | # needs to be adjusted accordingly 156 | args.world_size = ngpus_per_node * args.world_size 157 | # Use torch.multiprocessing.spawn to launch distributed processes: the 158 | # main_worker process function 159 | mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) 160 | else: 161 | # Simply call main_worker function 162 | main_worker(args.gpu, ngpus_per_node, args) 163 | 164 | 165 | def main_worker(gpu, ngpus_per_node, args): 166 | global best_acc1, logs 167 | args.gpu = gpu 168 | 169 | if args.gpu is not None: 170 | print("Use GPU: {} for training".format(args.gpu)) 171 | 172 | if args.distributed: 173 | if args.dist_url == "env://" and args.rank == -1: 174 | args.rank = int(os.environ["RANK"]) 175 | if args.multiprocessing_distributed: 176 | # For multiprocessing distributed training, rank needs to be the 177 | # global rank among all the processes 178 | args.rank = args.rank * ngpus_per_node + gpu 179 | dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, 180 | world_size=args.world_size, rank=args.rank) 181 | # create model 182 | if args.pretrained: 183 | print("=> using pre-trained model '{}'".format(args.arch)) 184 | model = models.__dict__[args.arch](pretrained=True) 185 | else: 186 | print("=> creating model '{}'".format(args.arch)) 187 | model = models.__dict__[args.arch]() 188 | constant.MODEL_SIZE = ravel_model_params(model).numel() 189 | 190 | if args.distributed: 191 | # For multiprocessing distributed, DistributedDataParallel constructor 192 | # should always set the single device scope, otherwise, 193 | # DistributedDataParallel will use all available devices. 194 | if args.gpu is not None: 195 | torch.cuda.set_device(args.gpu) 196 | model.cuda(args.gpu) 197 | # When using a single GPU per process and per 198 | # DistributedDataParallel, we need to divide the batch size 199 | # ourselves based on the total number of GPUs we have 200 | args.batch_size = int(args.batch_size / ngpus_per_node) 201 | args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node) 202 | model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) 203 | else: 204 | model.cuda() 205 | # DistributedDataParallel will divide and allocate batch_size to all 206 | # available GPUs if device_ids are not set 207 | model = torch.nn.parallel.DistributedDataParallel(model) 208 | elif args.gpu is not None: 209 | torch.cuda.set_device(args.gpu) 210 | model = model.cuda(args.gpu) 211 | else: 212 | # DataParallel will divide and allocate batch_size to all available GPUs 213 | if args.arch.startswith('alexnet') or args.arch.startswith('vgg'): 214 | model.features = torch.nn.DataParallel(model.features) 215 | model.cuda() 216 | else: 217 | # model = torch.nn.DataParallel(model).cuda() 218 | model = model.cuda() 219 | 220 | # define loss function (criterion) and optimizer 221 | criterion = nn.CrossEntropyLoss().cuda(args.gpu) 222 | 223 | # optimizer = torch.optim.SGD(model.parameters(), args.lr, 224 | # momentum=args.momentum, 225 | # weight_decay=args.weight_decay) 226 | optimizer = GradientSGD(model.parameters(), lr=args.lr, 227 | model=model, momentum=args.momentum, 228 | weight_decay=args.weight_decay, 229 | args=args) 230 | 231 | # optionally resume from a checkpoint 232 | if args.resume: 233 | if os.path.isfile(args.resume): 234 | print("=> loading checkpoint '{}'".format(args.resume)) 235 | if args.gpu is None: 236 | checkpoint = torch.load(args.resume) 237 | else: 238 | # Map model to be loaded to specified single gpu. 239 | loc = 'cuda:{}'.format(args.gpu) 240 | checkpoint = torch.load(args.resume, map_location=loc) 241 | args.start_epoch = checkpoint['epoch'] 242 | best_acc1 = checkpoint['best_acc1'] 243 | logs = checkpoint['logs'] 244 | if args.gpu is not None: 245 | # best_acc1 may be from a checkpoint from a different GPU 246 | best_acc1 = best_acc1.to(args.gpu) 247 | model.load_state_dict(checkpoint['state_dict']) 248 | optimizer.load_state_dict(checkpoint['optimizer']) 249 | print("=> loaded checkpoint '{}' (epoch {})" 250 | .format(args.resume, checkpoint['epoch'])) 251 | else: 252 | print("=> no checkpoint found at '{}'".format(args.resume)) 253 | 254 | cudnn.benchmark = True 255 | 256 | # Data loading code 257 | traindir = os.path.join(args.data, 'train') 258 | valdir = os.path.join(args.data, 'val') 259 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 260 | std=[0.229, 0.224, 0.225]) 261 | 262 | train_dataset = datasets.ImageFolder( 263 | traindir, 264 | transforms.Compose([ 265 | transforms.RandomResizedCrop(224), 266 | transforms.RandomHorizontalFlip(), 267 | transforms.ToTensor(), 268 | normalize, 269 | ])) 270 | val_dataset = datasets.ImageFolder(valdir, transforms.Compose([ 271 | transforms.Resize(256), 272 | transforms.CenterCrop(224), 273 | transforms.ToTensor(), 274 | normalize, 275 | ])) 276 | 277 | if args.distributed: 278 | train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) 279 | else: 280 | # train_sampler = None 281 | train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, args.world_size - 1, 282 | args.rank - 1) 283 | 284 | train_loader = torch.utils.data.DataLoader( 285 | train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None), 286 | num_workers=args.workers, pin_memory=True, sampler=train_sampler) 287 | val_loader = torch.utils.data.DataLoader( 288 | val_dataset, 289 | batch_size=args.batch_size, shuffle=False, 290 | num_workers=args.workers, pin_memory=True) 291 | if 'gpu' in socket.gethostname(): 292 | print('Init Dataloader in V100 cluster, using memory dataloader') 293 | # train_loader = MemoryDataLoader(args, train_loader,'train').load() 294 | val_loader = MemoryDataLoader(args, val_loader, 'val').load() 295 | 296 | if args.evaluate: 297 | validate(val_loader, model, criterion, args) 298 | return 299 | 300 | for epoch in range(args.start_epoch, args.epochs): 301 | if args.distributed: 302 | train_sampler.set_epoch(epoch) 303 | adjust_learning_rate(optimizer, epoch, args) 304 | 305 | # train for one epoch 306 | train(train_loader, model, criterion, optimizer, epoch, args) 307 | 308 | # evaluate on validation set 309 | acc1 = validate(val_loader, model, criterion, args) 310 | 311 | # remember best acc@1 and save checkpoint 312 | is_best = acc1 > best_acc1 313 | best_acc1 = max(acc1, best_acc1) 314 | 315 | # if not args.multiprocessing_distributed or (args.multiprocessing_distributed 316 | # and args.rank % ngpus_per_node == 0): 317 | # save_checkpoint({ 318 | # 'epoch': epoch + 1, 319 | # 'arch': args.arch, 320 | # 'state_dict': model.state_dict(), 321 | # 'best_acc1': best_acc1, 322 | # 'optimizer': optimizer.state_dict(), 323 | # 'logs': logs 324 | # }, is_best) 325 | 326 | # running log 327 | 328 | with open(WORKPATH + '/running.log', 'a+') as f: 329 | running_log = '{},node{}_{}_{}_m{}_e{}_{}.csv'.format(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), 330 | args.rank - 1, args.mode, 331 | args.arch, args.momentum, 332 | epoch, 333 | logs[-1]['test_accuracy']) 334 | f.write(running_log + '\n') 335 | 336 | df = pandas.DataFrame(logs) 337 | df.to_csv(WORKPATH + '/log/node{}_{}_{}_m{}_e{}_b{}_{}worker.csv'.format(args.rank - 1, args.mode, 338 | args.arch, args.momentum, 339 | args.epochs, 340 | args.batch_size, 341 | args.world_size - 1, ), 342 | index_label='index') 343 | 344 | 345 | def train(train_loader, model, criterion, optimizer, epoch, args): 346 | batch_time = AverageMeter('Time', ':6.3f') 347 | data_time = AverageMeter('Data', ':6.3f') 348 | losses = AverageMeter('Loss', ':.4e') 349 | top1 = AverageMeter('Acc@1', ':6.2f') 350 | top5 = AverageMeter('Acc@5', ':6.2f') 351 | progress = ProgressMeter( 352 | len(train_loader), 353 | [batch_time, data_time, losses, top1, top5], 354 | prefix="Epoch: [{}]".format(epoch)) 355 | 356 | # switch to train mode 357 | model.train() 358 | 359 | end = time.time() 360 | for i, (images, target) in enumerate(train_loader): 361 | # measure data loading time 362 | data_time.update(time.time() - end) 363 | 364 | if args.gpu is not None: 365 | images = images.cuda(args.gpu, non_blocking=True) 366 | target = target.cuda(args.gpu, non_blocking=True) 367 | 368 | # compute output 369 | output = model(images) 370 | loss = criterion(output, target) 371 | 372 | # measure accuracy and record loss 373 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 374 | losses.update(loss.item(), images.size(0)) 375 | top1.update(acc1[0], images.size(0)) 376 | top5.update(acc5[0], images.size(0)) 377 | log_obj = { 378 | 'timestamp': datetime.now(), 379 | 'iteration': i, 380 | 'training_loss': loss.item(), 381 | 'training_accuracy': acc1.tolist()[0], 382 | 'training_accuracy5': acc5.tolist()[0], 383 | } 384 | logs.append(log_obj) 385 | # compute gradient and do SGD step 386 | optimizer.zero_grad() 387 | loss.backward() 388 | optimizer.step() 389 | 390 | # measure elapsed time 391 | batch_time.update(time.time() - end) 392 | end = time.time() 393 | 394 | if i % args.print_freq == 0: 395 | progress.display(i) 396 | 397 | 398 | def validate(val_loader, model, criterion, args): 399 | batch_time = AverageMeter('Time', ':6.3f') 400 | losses = AverageMeter('Loss', ':.4e') 401 | top1 = AverageMeter('Acc@1', ':6.2f') 402 | top5 = AverageMeter('Acc@5', ':6.2f') 403 | progress = ProgressMeter( 404 | len(val_loader), 405 | [batch_time, losses, top1, top5], 406 | prefix='Test: ') 407 | 408 | # switch to evaluate mode 409 | model.eval() 410 | if 'gpu' not in socket.gethostname(): 411 | val_loader = enumerate(val_loader) 412 | with torch.no_grad(): 413 | end = time.time() 414 | # check 415 | for i, (images, target) in val_loader: 416 | if args.gpu is not None: 417 | images = images.cuda(args.gpu, non_blocking=True) 418 | target = target.cuda(args.gpu, non_blocking=True) 419 | 420 | # compute output 421 | output = model(images) 422 | loss = criterion(output, target) 423 | 424 | # measure accuracy and record loss 425 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 426 | losses.update(loss.item(), images.size(0)) 427 | top1.update(acc1[0], images.size(0)) 428 | top5.update(acc5[0], images.size(0)) 429 | 430 | # measure elapsed time 431 | batch_time.update(time.time() - end) 432 | end = time.time() 433 | 434 | if i % args.print_freq == 0: 435 | progress.display(i) 436 | 437 | # TODO: this should also be done with the ProgressMeter 438 | print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' 439 | .format(top1=top1, top5=top5)) 440 | logs[-1]['test_loss'], logs[-1]['test_accuracy'], logs[-1]['test_accuracy5'] = ( 441 | loss.item(), top1.avg.item(), top5.avg.item()) 442 | 443 | return top1.avg 444 | 445 | 446 | def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'): 447 | torch.save(state, filename) 448 | if is_best: 449 | shutil.copyfile(filename, 'model_best.pth.tar') 450 | 451 | 452 | class AverageMeter(object): 453 | """Computes and stores the average and current value""" 454 | 455 | def __init__(self, name, fmt=':f'): 456 | self.name = name 457 | self.fmt = fmt 458 | self.reset() 459 | 460 | def reset(self): 461 | self.val = 0 462 | self.avg = 0 463 | self.sum = 0 464 | self.count = 0 465 | 466 | def update(self, val, n=1): 467 | self.val = val 468 | self.sum += val * n 469 | self.count += n 470 | self.avg = self.sum / self.count 471 | 472 | def __str__(self): 473 | fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})' 474 | return fmtstr.format(**self.__dict__) 475 | 476 | 477 | class ProgressMeter(object): 478 | def __init__(self, num_batches, meters, prefix=""): 479 | self.batch_fmtstr = self._get_batch_fmtstr(num_batches) 480 | self.meters = meters 481 | self.prefix = prefix 482 | 483 | def display(self, batch): 484 | entries = [self.prefix + self.batch_fmtstr.format(batch)] 485 | entries += [str(meter) for meter in self.meters] 486 | print('\t'.join(entries)) 487 | 488 | def _get_batch_fmtstr(self, num_batches): 489 | num_digits = len(str(num_batches // 1)) 490 | fmt = '{:' + str(num_digits) + 'd}' 491 | return '[' + fmt + '/' + fmt.format(num_batches) + ']' 492 | 493 | 494 | def adjust_learning_rate(optimizer, epoch, args): 495 | """Sets the learning rate to the initial LR decayed by 10 every 30 epochs""" 496 | lr = args.lr * (0.1 ** (epoch // 30)) 497 | for param_group in optimizer.param_groups: 498 | param_group['lr'] = lr 499 | 500 | 501 | def accuracy(output, target, topk=(1,)): 502 | """Computes the accuracy over the k top predictions for the specified values of k""" 503 | with torch.no_grad(): 504 | maxk = max(topk) 505 | batch_size = target.size(0) 506 | 507 | _, pred = output.topk(maxk, 1, True, True) 508 | pred = pred.t() 509 | correct = pred.eq(target.view(1, -1).expand_as(pred)) 510 | 511 | res = [] 512 | for k in topk: 513 | correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) 514 | res.append(correct_k.mul_(100.0 / batch_size)) 515 | return res 516 | 517 | 518 | class MemoryDataLoader(object): 519 | 520 | def __init__(self, args, dataloader, type, path='.', dataset='imagenet'): 521 | self.pickle_name = '_'.join([dataset, str(args.batch_size), type]) 522 | self.path = os.path.join(path, self.pickle_name) 523 | if not os.path.exists(self.path): 524 | self.save(dataloader) 525 | 526 | def load(self): 527 | print('Loading pickle to memory') 528 | with open(self.path, 'rb') as file: 529 | data_loader_list = pickle.load(file) 530 | return data_loader_list 531 | 532 | def save(self, dataloader): 533 | loader_list = [] 534 | for i, (images, target) in enumerate(dataloader): 535 | print(self.pickle_name, ' Processing dataloader: ', str(i) + '/' + str(len(dataloader))) 536 | t = (i, (images, target)) 537 | loader_list.append(t) 538 | 539 | with open(self.path, 'wb') as file: 540 | print('Saving dataloader to pickle') 541 | pickle.dump(loader_list, file) 542 | print('Saving dataloader Success') 543 | 544 | 545 | if __name__ == '__main__': 546 | main() 547 | -------------------------------------------------------------------------------- /example/Imagenet_local.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | import random 4 | import shutil 5 | import time 6 | import warnings 7 | from datetime import datetime 8 | 9 | import pandas 10 | import torch 11 | import torch.backends.cudnn as cudnn 12 | import torch.distributed as dist 13 | import torch.multiprocessing as mp 14 | import torch.nn as nn 15 | import torch.nn.parallel 16 | import torch.optim 17 | import torch.utils.data 18 | import torch.utils.data.distributed 19 | import torchvision.datasets as datasets 20 | import torchvision.models as models 21 | import torchvision.transforms as transforms 22 | 23 | WORKPATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) 24 | model_names = sorted(name for name in models.__dict__ 25 | if name.islower() and not name.startswith("__") 26 | and callable(models.__dict__[name])) 27 | from PIL import ImageFile 28 | 29 | ImageFile.LOAD_TRUNCATED_IMAGES = True 30 | parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') 31 | parser.add_argument('data', metavar='DIR', 32 | help='path to dataset') 33 | parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet18', 34 | choices=model_names, 35 | help='model architecture: ' + 36 | ' | '.join(model_names) + 37 | ' (default: resnet18)') 38 | parser.add_argument('-j', '--workers', default=8, type=int, metavar='N', 39 | help='number of data loading workers (default: 4)') 40 | parser.add_argument('--epochs', default=90, type=int, metavar='N', 41 | help='number of total epochs to run') 42 | parser.add_argument('--start-epoch', default=0, type=int, metavar='N', 43 | help='manual epoch number (useful on restarts)') 44 | parser.add_argument('-b', '--batch-size', default=256, type=int, 45 | metavar='N', 46 | help='mini-batch size (default: 256), this is the total ' 47 | 'batch size of all GPUs on the current node when ' 48 | 'using Data Parallel or Distributed Data Parallel') 49 | parser.add_argument('--lr', '--learning-rate', default=0.1, type=float, 50 | metavar='LR', help='initial learning rate', dest='lr') 51 | parser.add_argument('--momentum', default=0.7, type=float, metavar='M', 52 | help='momentum') 53 | parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float, 54 | metavar='W', help='weight decay (default: 1e-4)', 55 | dest='weight_decay') 56 | parser.add_argument('-p', '--print-freq', default=10, type=int, 57 | metavar='N', help='print frequency (default: 10)') 58 | parser.add_argument('--resume', default='', type=str, metavar='PATH', 59 | help='path to latest checkpoint (default: none)') 60 | parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true', 61 | help='evaluate model on validation set') 62 | parser.add_argument('--pretrained', dest='pretrained', action='store_true', 63 | help='use pre-trained model') 64 | parser.add_argument('--world-size', default=-1, type=int, 65 | help='number of nodes for distributed training') 66 | parser.add_argument('--rank', default=-1, type=int, 67 | help='node rank for distributed training') 68 | parser.add_argument('--dist-url', default='tcp://224.66.41.62:23456', type=str, 69 | help='url used to set up distributed training') 70 | parser.add_argument('--dist-backend', default='nccl', type=str, 71 | help='distributed backend') 72 | parser.add_argument('--seed', default=None, type=int, 73 | help='seed for initializing training. ') 74 | parser.add_argument('--gpu', default=None, type=int, 75 | help='GPU id to use.') 76 | parser.add_argument('--multiprocessing-distributed', action='store_true', 77 | help='Use multi-processing distributed training to launch ' 78 | 'N processes per node, which has N GPUs. This is the ' 79 | 'fastest way to use PyTorch for either single node or ' 80 | 'multi node data parallel training') 81 | logs = [] 82 | best_acc1 = 0 83 | 84 | 85 | def main(): 86 | args = parser.parse_args() 87 | 88 | if args.seed is not None: 89 | random.seed(args.seed) 90 | torch.manual_seed(args.seed) 91 | cudnn.deterministic = True 92 | warnings.warn('You have chosen to seed training. ' 93 | 'This will turn on the CUDNN deterministic setting, ' 94 | 'which can slow down your training considerably! ' 95 | 'You may see unexpected behavior when restarting ' 96 | 'from checkpoints.') 97 | 98 | if args.gpu is not None: 99 | warnings.warn('You have chosen a specific GPU. This will completely ' 100 | 'disable data parallelism.') 101 | 102 | if args.dist_url == "env://" and args.world_size == -1: 103 | args.world_size = int(os.environ["WORLD_SIZE"]) 104 | 105 | args.distributed = args.world_size > 1 or args.multiprocessing_distributed 106 | 107 | ngpus_per_node = torch.cuda.device_count() 108 | if args.multiprocessing_distributed: 109 | # Since we have ngpus_per_node processes per node, the total world_size 110 | # needs to be adjusted accordingly 111 | args.world_size = ngpus_per_node * args.world_size 112 | # Use torch.multiprocessing.spawn to launch distributed processes: the 113 | # main_worker process function 114 | mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) 115 | else: 116 | # Simply call main_worker function 117 | main_worker(args.gpu, ngpus_per_node, args) 118 | 119 | 120 | def main_worker(gpu, ngpus_per_node, args): 121 | global best_acc1, logs 122 | args.gpu = gpu 123 | 124 | if args.gpu is not None: 125 | print("Use GPU: {} for training".format(args.gpu)) 126 | 127 | if args.distributed: 128 | if args.dist_url == "env://" and args.rank == -1: 129 | args.rank = int(os.environ["RANK"]) 130 | if args.multiprocessing_distributed: 131 | # For multiprocessing distributed training, rank needs to be the 132 | # global rank among all the processes 133 | args.rank = args.rank * ngpus_per_node + gpu 134 | dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, 135 | world_size=args.world_size, rank=args.rank) 136 | # create model 137 | if args.pretrained: 138 | print("=> using pre-trained model '{}'".format(args.arch)) 139 | model = models.__dict__[args.arch](pretrained=True) 140 | else: 141 | print("=> creating model '{}'".format(args.arch)) 142 | model = models.__dict__[args.arch]() 143 | 144 | if args.distributed: 145 | # For multiprocessing distributed, DistributedDataParallel constructor 146 | # should always set the single device scope, otherwise, 147 | # DistributedDataParallel will use all available devices. 148 | if args.gpu is not None: 149 | torch.cuda.set_device(args.gpu) 150 | model.cuda(args.gpu) 151 | # When using a single GPU per process and per 152 | # DistributedDataParallel, we need to divide the batch size 153 | # ourselves based on the total number of GPUs we have 154 | args.batch_size = int(args.batch_size / ngpus_per_node) 155 | args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node) 156 | model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) 157 | else: 158 | model.cuda() 159 | # DistributedDataParallel will divide and allocate batch_size to all 160 | # available GPUs if device_ids are not set 161 | model = torch.nn.parallel.DistributedDataParallel(model) 162 | elif args.gpu is not None: 163 | torch.cuda.set_device(args.gpu) 164 | model = model.cuda(args.gpu) 165 | else: 166 | # DataParallel will divide and allocate batch_size to all available GPUs 167 | if args.arch.startswith('alexnet') or args.arch.startswith('vgg'): 168 | model.features = torch.nn.DataParallel(model.features) 169 | model.cuda() 170 | else: 171 | model = torch.nn.DataParallel(model).cuda() 172 | 173 | # define loss function (criterion) and optimizer 174 | criterion = nn.CrossEntropyLoss().cuda(args.gpu) 175 | 176 | optimizer = torch.optim.SGD(model.parameters(), args.lr, 177 | momentum=args.momentum, 178 | weight_decay=args.weight_decay) 179 | 180 | # optionally resume from a checkpoint 181 | if args.resume: 182 | if os.path.isfile(args.resume): 183 | print("=> loading checkpoint '{}'".format(args.resume)) 184 | if args.gpu is None: 185 | checkpoint = torch.load(args.resume) 186 | else: 187 | # Map model to be loaded to specified single gpu. 188 | loc = 'cuda:{}'.format(args.gpu) 189 | checkpoint = torch.load(args.resume, map_location=loc) 190 | args.start_epoch = checkpoint['epoch'] 191 | best_acc1 = checkpoint['best_acc1'] 192 | if args.gpu is not None: 193 | # best_acc1 may be from a checkpoint from a different GPU 194 | best_acc1 = best_acc1.to(args.gpu) 195 | model.load_state_dict(checkpoint['state_dict']) 196 | optimizer.load_state_dict(checkpoint['optimizer']) 197 | print("=> loaded checkpoint '{}' (epoch {})" 198 | .format(args.resume, checkpoint['epoch'])) 199 | else: 200 | print("=> no checkpoint found at '{}'".format(args.resume)) 201 | 202 | cudnn.benchmark = True 203 | 204 | # Data loading code 205 | traindir = os.path.join(args.data, 'train') 206 | valdir = os.path.join(args.data, 'val') 207 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 208 | std=[0.229, 0.224, 0.225]) 209 | 210 | train_dataset = datasets.ImageFolder( 211 | traindir, 212 | transforms.Compose([ 213 | transforms.RandomResizedCrop(224), 214 | transforms.RandomHorizontalFlip(), 215 | transforms.ToTensor(), 216 | normalize, 217 | ])) 218 | 219 | if args.distributed: 220 | train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) 221 | else: 222 | train_sampler = None 223 | 224 | train_loader = torch.utils.data.DataLoader( 225 | train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None), 226 | num_workers=args.workers, pin_memory=True, sampler=train_sampler) 227 | 228 | val_loader = torch.utils.data.DataLoader( 229 | datasets.ImageFolder(valdir, transforms.Compose([ 230 | transforms.Resize(256), 231 | transforms.CenterCrop(224), 232 | transforms.ToTensor(), 233 | normalize, 234 | ])), 235 | batch_size=args.batch_size, shuffle=False, 236 | num_workers=args.workers, pin_memory=True) 237 | 238 | if args.evaluate: 239 | validate(val_loader, model, criterion, args) 240 | return 241 | 242 | for epoch in range(args.start_epoch, args.epochs): 243 | if args.distributed: 244 | train_sampler.set_epoch(epoch) 245 | adjust_learning_rate(optimizer, epoch, args) 246 | 247 | # train for one epoch 248 | train(train_loader, model, criterion, optimizer, epoch, args) 249 | 250 | # evaluate on validation set 251 | acc1 = validate(val_loader, model, criterion, args) 252 | 253 | # remember best acc@1 and save checkpoint 254 | is_best = acc1 > best_acc1 255 | best_acc1 = max(acc1, best_acc1) 256 | 257 | if not args.multiprocessing_distributed or (args.multiprocessing_distributed 258 | and args.rank % ngpus_per_node == 0): 259 | save_checkpoint({ 260 | 'epoch': epoch + 1, 261 | 'arch': args.arch, 262 | 'state_dict': model.state_dict(), 263 | 'best_acc1': best_acc1, 264 | 'optimizer': optimizer.state_dict(), 265 | }, is_best) 266 | 267 | with open(WORKPATH + '/running.log', 'a+') as f: 268 | running_log = '{},node{}_{}_{}_m{}_e{}_{}.csv'.format(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), 269 | args.rank - 1, 'single', 270 | args.arch, args.momentum, 271 | epoch, 272 | logs[-1]['test_accuracy']) 273 | f.write(running_log + '\n') 274 | 275 | df = pandas.DataFrame(logs) 276 | df.to_csv(WORKPATH + '/log/node{}_{}_{}_m{}_e{}_b{}_{}worker.csv'.format(args.rank - 1, 'single', 277 | args.arch, args.momentum, 278 | args.epochs, 279 | args.batch_size, 280 | args.world_size - 1, ), 281 | index_label='index') 282 | 283 | 284 | def train(train_loader, model, criterion, optimizer, epoch, args): 285 | batch_time = AverageMeter('Time', ':6.3f') 286 | data_time = AverageMeter('Data', ':6.3f') 287 | losses = AverageMeter('Loss', ':.4e') 288 | top1 = AverageMeter('Acc@1', ':6.2f') 289 | top5 = AverageMeter('Acc@5', ':6.2f') 290 | progress = ProgressMeter( 291 | len(train_loader), 292 | [batch_time, data_time, losses, top1, top5], 293 | prefix="Epoch: [{}]".format(epoch)) 294 | 295 | # switch to train mode 296 | model.train() 297 | 298 | end = time.time() 299 | for i, (images, target) in enumerate(train_loader): 300 | # measure data loading time 301 | data_time.update(time.time() - end) 302 | 303 | if args.gpu is not None: 304 | images = images.cuda(args.gpu, non_blocking=True) 305 | target = target.cuda(args.gpu, non_blocking=True) 306 | 307 | # compute output 308 | output = model(images) 309 | loss = criterion(output, target) 310 | 311 | # measure accuracy and record loss 312 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 313 | losses.update(loss.item(), images.size(0)) 314 | top1.update(acc1[0], images.size(0)) 315 | top5.update(acc5[0], images.size(0)) 316 | log_obj = { 317 | 'timestamp': datetime.now(), 318 | 'iteration': i, 319 | 'training_loss': loss.item(), 320 | 'training_accuracy': acc1.tolist()[0], 321 | 'training_accuracy5': acc5.tolist()[0], 322 | } 323 | logs.append(log_obj) 324 | 325 | # compute gradient and do SGD step 326 | optimizer.zero_grad() 327 | loss.backward() 328 | optimizer.step() 329 | 330 | # measure elapsed time 331 | batch_time.update(time.time() - end) 332 | end = time.time() 333 | 334 | if i % args.print_freq == 0: 335 | progress.display(i) 336 | 337 | 338 | def validate(val_loader, model, criterion, args): 339 | batch_time = AverageMeter('Time', ':6.3f') 340 | losses = AverageMeter('Loss', ':.4e') 341 | top1 = AverageMeter('Acc@1', ':6.2f') 342 | top5 = AverageMeter('Acc@5', ':6.2f') 343 | progress = ProgressMeter( 344 | len(val_loader), 345 | [batch_time, losses, top1, top5], 346 | prefix='Test: ') 347 | 348 | # switch to evaluate mode 349 | model.eval() 350 | 351 | with torch.no_grad(): 352 | end = time.time() 353 | for i, (images, target) in enumerate(val_loader): 354 | if args.gpu is not None: 355 | images = images.cuda(args.gpu, non_blocking=True) 356 | target = target.cuda(args.gpu, non_blocking=True) 357 | 358 | # compute output 359 | output = model(images) 360 | loss = criterion(output, target) 361 | 362 | # measure accuracy and record loss 363 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 364 | losses.update(loss.item(), images.size(0)) 365 | top1.update(acc1[0], images.size(0)) 366 | top5.update(acc5[0], images.size(0)) 367 | 368 | # measure elapsed time 369 | batch_time.update(time.time() - end) 370 | end = time.time() 371 | 372 | if i % args.print_freq == 0: 373 | progress.display(i) 374 | 375 | # TODO: this should also be done with the ProgressMeter 376 | logs[-1]['test_loss'], logs[-1]['test_accuracy'], logs[-1]['test_accuracy5'] = ( 377 | loss.item(), top1.avg.item(), top5.avg.item()) 378 | print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' 379 | .format(top1=top1, top5=top5)) 380 | 381 | return top1.avg 382 | 383 | 384 | def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'): 385 | torch.save(state, filename) 386 | if is_best: 387 | shutil.copyfile(filename, 'model_best.pth.tar') 388 | 389 | 390 | class AverageMeter(object): 391 | """Computes and stores the average and current value""" 392 | 393 | def __init__(self, name, fmt=':f'): 394 | self.name = name 395 | self.fmt = fmt 396 | self.reset() 397 | 398 | def reset(self): 399 | self.val = 0 400 | self.avg = 0 401 | self.sum = 0 402 | self.count = 0 403 | 404 | def update(self, val, n=1): 405 | self.val = val 406 | self.sum += val * n 407 | self.count += n 408 | self.avg = self.sum / self.count 409 | 410 | def __str__(self): 411 | fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})' 412 | return fmtstr.format(**self.__dict__) 413 | 414 | 415 | class ProgressMeter(object): 416 | def __init__(self, num_batches, meters, prefix=""): 417 | self.batch_fmtstr = self._get_batch_fmtstr(num_batches) 418 | self.meters = meters 419 | self.prefix = prefix 420 | 421 | def display(self, batch): 422 | entries = [self.prefix + self.batch_fmtstr.format(batch)] 423 | entries += [str(meter) for meter in self.meters] 424 | print('\t'.join(entries)) 425 | 426 | def _get_batch_fmtstr(self, num_batches): 427 | num_digits = len(str(num_batches // 1)) 428 | fmt = '{:' + str(num_digits) + 'd}' 429 | return '[' + fmt + '/' + fmt.format(num_batches) + ']' 430 | 431 | 432 | def adjust_learning_rate(optimizer, epoch, args): 433 | """Sets the learning rate to the initial LR decayed by 10 every 30 epochs""" 434 | lr = args.lr * (0.1 ** (epoch // 30)) 435 | for param_group in optimizer.param_groups: 436 | param_group['lr'] = lr 437 | 438 | 439 | def accuracy(output, target, topk=(1,)): 440 | """Computes the accuracy over the k top predictions for the specified values of k""" 441 | with torch.no_grad(): 442 | maxk = max(topk) 443 | batch_size = target.size(0) 444 | 445 | _, pred = output.topk(maxk, 1, True, True) 446 | pred = pred.t() 447 | correct = pred.eq(target.view(1, -1).expand_as(pred)) 448 | 449 | res = [] 450 | for k in topk: 451 | correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) 452 | res.append(correct_k.mul_(100.0 / batch_size)) 453 | return res 454 | 455 | 456 | if __name__ == '__main__': 457 | main() 458 | -------------------------------------------------------------------------------- /example/an4.py: -------------------------------------------------------------------------------- 1 | import json 2 | import os 3 | import sys 4 | import time 5 | 6 | import torch.optim as optim 7 | import torch.utils.data.distributed 8 | from torch_baidu_ctc import CTCLoss 9 | from tqdm import tqdm 10 | 11 | WORKPATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) 12 | print(WORKPATH) 13 | sys.path.append(WORKPATH + '/deepspeech') 14 | 15 | from data.data_loader import AudioDataLoader, SpectrogramDataset, DistributedBucketingSampler 16 | from decoder import GreedyDecoder 17 | from model import DeepSpeech, supported_rnns 18 | from datetime import datetime 19 | import pandas as pd 20 | 21 | torch.manual_seed(123456) 22 | torch.cuda.manual_seed_all(123456) 23 | device = torch.device("cuda") 24 | 25 | 26 | def to_np(x): 27 | return x.data.cpu().numpy() 28 | 29 | 30 | class AverageMeter(object): 31 | """Computes and stores the average and current value""" 32 | 33 | def __init__(self): 34 | self.reset() 35 | 36 | def reset(self): 37 | self.val = 0 38 | self.avg = 0 39 | self.sum = 0 40 | self.count = 0 41 | 42 | def update(self, val, n=1): 43 | self.val = val 44 | self.sum += val * n 45 | self.count += n 46 | self.avg = self.sum / self.count 47 | 48 | 49 | Total_param_num = 0 50 | Sparse_param_num = 0 51 | criterion = CTCLoss() 52 | best_wer = None 53 | decoder = None 54 | audio_conf = None 55 | labels = None 56 | 57 | 58 | # 获取数据 59 | 60 | 61 | def init_net(args): 62 | # Model 63 | global decoder, audio_conf, labels 64 | with open(args.labels_path) as label_file: 65 | labels = str(''.join(json.load(label_file))) 66 | audio_conf = dict(sample_rate=args.sample_rate, 67 | window_size=args.window_size, 68 | window_stride=args.window_stride, 69 | window=args.window, 70 | noise_dir=args.noise_dir, 71 | noise_prob=args.noise_prob, 72 | noise_levels=(args.noise_min, args.noise_max)) 73 | 74 | rnn_type = args.rnn_type.lower() 75 | decoder = GreedyDecoder(labels) 76 | assert rnn_type in supported_rnns, "rnn_type should be either lstm, rnn or gru" 77 | net = DeepSpeech(rnn_hidden_size=args.hidden_size, 78 | nb_layers=args.hidden_layers, 79 | labels=labels, 80 | rnn_type=supported_rnns[rnn_type], 81 | audio_conf=audio_conf, 82 | bidirectional=args.bidirectional) 83 | net = net.cuda() 84 | return net 85 | 86 | 87 | def an4(args, optimizer, net): 88 | avg_loss, start_epoch, start_iter = 0, 0, 0 89 | 90 | # Training Data 91 | train_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.train_manifest, labels=labels, 92 | normalize=True, augment=True) 93 | train_sampler = DistributedBucketingSampler(train_dataset, batch_size=args.batch_size, 94 | num_replicas=args.world_size - 1, rank=args.rank - 1) 95 | train_loader = AudioDataLoader(train_dataset, num_workers=args.num_workers, batch_sampler=train_sampler) 96 | 97 | # Testing Data 98 | test_dataset = SpectrogramDataset(audio_conf=audio_conf, manifest_filepath=args.val_manifest, labels=labels, 99 | normalize=True, augment=False) 100 | test_loader = AudioDataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers) 101 | 102 | # Optimizer and scheduler of Training 103 | # optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, nesterov=True) 104 | scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=2, verbose=True, factor=args.learning_anneal) 105 | 106 | logs = [] 107 | print("Training Start") 108 | losses = AverageMeter() 109 | 110 | for epoch in range(args.epochs): 111 | 112 | print("Training for epoch {}".format(epoch)) 113 | net.train() 114 | 115 | for i, (data) in enumerate(train_loader): 116 | batch_start_time = time.time() 117 | 118 | inputs, targets, input_percentages, target_sizes = data 119 | input_sizes = input_percentages.mul_(int(inputs.size(3))).int() 120 | 121 | inputs = inputs.to(device) 122 | 123 | optimizer.zero_grad() 124 | 125 | out, output_sizes = net(inputs, input_sizes) 126 | out = out.transpose(0, 1) # TxNxH 127 | 128 | # Loss Operation 129 | loss = criterion(out, targets, output_sizes, target_sizes).to(device) 130 | loss = loss / inputs.size(0) # average the loss by minibatch 131 | 132 | inf = float("inf") 133 | 134 | loss_value = loss.item() 135 | 136 | if loss_value == inf or loss_value == -inf: 137 | print("WARNING: received an inf loss, setting loss value to 0") 138 | loss_value = 0 139 | 140 | avg_loss += loss_value 141 | losses.update(loss_value, inputs.size(0)) 142 | 143 | # compute gradient 144 | loss.backward() 145 | 146 | # Gradient Clip 147 | torch.nn.utils.clip_grad_norm_(net.parameters(), args.max_norm) 148 | 149 | # paralist = gradient_execute(net) 150 | 151 | # SGD step 152 | optimizer.step() 153 | 154 | # for para1, para2 in zip(paralist, net.parameters()): 155 | # para2.grad.data = para1 156 | 157 | log_obj = { 158 | 'timestamp': datetime.now(), 159 | 'iteration': i, 160 | 'training_loss': losses.avg, 161 | 'total_param': Total_param_num, 162 | 'sparse_param': Sparse_param_num, 163 | 'mini_batch_time': (time.time() - batch_start_time) 164 | } 165 | logs.append(log_obj) 166 | 167 | 168 | if i % 5 == 0: 169 | print("Timestamp: {timestamp} | " 170 | "Iteration: {iteration:6} | " 171 | "Loss: {training_loss:6.4f} | " 172 | "Total_param: {total_param:6} | " 173 | "Sparse_param: {sparse_param:6} | " 174 | "Mini_Batch_Time: {mini_batch_time:6.4f} | ".format(**log_obj)) 175 | 176 | 177 | # if True: 178 | test_wer, test_cer = evaluate(net, test_loader) 179 | logs[-1]['test_wer'], logs[-1]['test_cer'] = test_wer, test_cer 180 | 181 | print("Timestamp: {timestamp} | " 182 | "Iteration: {iteration:6} | " 183 | "Loss: {training_loss:6.4f} | " 184 | "Total_param: {total_param:6} | " 185 | "Sparse_param: {sparse_param:6} | " 186 | "Mini_Batch_Time: {mini_batch_time:6.4f} | " 187 | "Test Wer: {test_wer:6.4f} | " 188 | "Test Cer: {test_cer:6.4f}".format(**logs[-1])) 189 | 190 | # sche_wer, sche_cer = evaluate(net, test_loader) 191 | scheduler.step(test_wer) 192 | 193 | df = pd.DataFrame(logs) 194 | df.to_csv(WORKPATH + '/log/node{}_{}_{}_m{}_e{}_b{}_{}worker_{}.csv'.format(args.rank - 1, args.mode, 195 | 'an4', args.momentum, 196 | args.epochs, 197 | args.batch_size, 198 | args.world_size - 1, 199 | test_wer)) 200 | # df.to_csv('./log/{}_Node{}_{}.csv'.format(args.file_name, args.dist_rank, datetime.now().strftime("%Y-%m-%d %H:%M:%S")), index_label='index') 201 | print("Finished Training") 202 | 203 | 204 | def evaluate(net, test_loader): 205 | total_cer, total_wer = 0, 0 206 | net.eval() 207 | with torch.no_grad(): 208 | for i, (data) in tqdm(enumerate(test_loader), total=len(test_loader)): 209 | inputs, targets, input_percentages, target_sizes = data 210 | input_sizes = input_percentages.mul_(int(inputs.size(3))).int() 211 | inputs = inputs.to(device) 212 | 213 | # unflatten targets 214 | split_targets = [] 215 | offset = 0 216 | for size in target_sizes: 217 | split_targets.append(targets[offset:offset + size]) 218 | offset += size 219 | 220 | out, output_sizes = net(inputs, input_sizes) 221 | 222 | decoded_output, _ = decoder.decode(out, output_sizes) 223 | target_strings = decoder.convert_to_strings(split_targets) 224 | wer, cer = 0, 0 225 | for x in range(len(target_strings)): 226 | transcript, reference = decoded_output[x][0], target_strings[x][0] 227 | wer += decoder.wer(transcript, reference) / float(len(reference.split())) 228 | cer += decoder.cer(transcript, reference) / float(len(reference)) 229 | total_cer += cer 230 | total_wer += wer 231 | del out 232 | wer = total_wer / len(test_loader.dataset) 233 | cer = total_cer / len(test_loader.dataset) 234 | wer *= 100 235 | cer *= 100 236 | print('Validation Summary Epoch: [{0}]\t' 237 | 'Average WER {wer:.3f}\t' 238 | 'Average CER {cer:.3f}\t'.format(-1, wer=wer, cer=cer)) 239 | return wer, cer 240 | 241 | 242 | if __name__ == "__main__": 243 | pass 244 | # dist.init_process_group(backend='nccl', init_method=args.init_method, rank=args.dist_rank,world_size=args.world_size) 245 | # dist.init_process_group(backend='nccl', init_method=args.init_method, world_size=args.world_size,group_name='mygroup') 246 | # an4(args2,None) 247 | -------------------------------------------------------------------------------- /example/cifar10.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | import socket 4 | import sys 5 | import time 6 | 7 | WORKPATH = os.path.abspath(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) 8 | sys.path.append(WORKPATH) 9 | from torch.optim.lr_scheduler import MultiStepLR 10 | 11 | from core.optim import GradientSGD 12 | from core.utils import constant 13 | from core.utils.GradualWarmupScheduler import GradualWarmupScheduler 14 | from core.utils.serialization import ravel_model_params 15 | from example.main import init_server 16 | 17 | import torchvision 18 | import torchvision.transforms as transforms 19 | from torch.utils.data.distributed import DistributedSampler 20 | 21 | from datetime import datetime 22 | from example.models import * 23 | from sklearn.metrics import classification_report, accuracy_score 24 | import pandas as pd 25 | 26 | 27 | def get_dataset(args, transform_train, transform_test): 28 | """ 29 | :param args: 30 | :param transform: 31 | :return: 32 | """ 33 | if args.dataset == 'MNIST': 34 | trainset = torchvision.datasets.MNIST(root='%s/data' % WORKPATH, train=True, download=True, 35 | transform=transform_train) 36 | testset = torchvision.datasets.MNIST(root='%s/data' % WORKPATH, train=False, download=True, 37 | transform=transform_test) 38 | elif args.dataset == 'cifar10': 39 | trainset = torchvision.datasets.CIFAR10(root='%s/data' % WORKPATH, train=True, download=True, 40 | transform=transform_train) 41 | testset = torchvision.datasets.CIFAR10(root='%s/data' % WORKPATH, train=False, download=True, 42 | transform=transform_test) 43 | 44 | sampler = DistributedSampler(trainset, args.world_size - 1, args.rank - 1) 45 | # sampler = DistributedSampler(trainset, 1, 0) 46 | trainloader = torch.utils.data.DataLoader(trainset, batch_size=args.batch_size, shuffle=False, num_workers=1, 47 | sampler=sampler) 48 | testloader = torch.utils.data.DataLoader(testset, batch_size=args.test_batch_size, shuffle=False, num_workers=1) 49 | return trainloader, testloader 50 | 51 | 52 | def init_net(args): 53 | if args.model == 'AlexNet': 54 | net = AlexNet() 55 | elif args.model == 'ResNet18': 56 | net = ResNet18() 57 | args.test_batch_size = 1000 58 | elif args.model == 'ResNet50': 59 | net = ResNet50() 60 | args.test_batch_size = 1000 61 | elif args.model == 'ResNet101': 62 | net = ResNet101() 63 | args.test_batch_size = 500 64 | if args.cuda: 65 | net = net.cuda() 66 | if args.no_distributed and args.half: 67 | net = net.half() 68 | return net 69 | 70 | 71 | def cifar10(args, optimizer, net): 72 | logs = [] 73 | transform_train = transforms.Compose([ 74 | transforms.RandomCrop(32, padding=4), 75 | transforms.RandomHorizontalFlip(), 76 | transforms.ToTensor(), 77 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), 78 | ]) 79 | 80 | transform_test = transforms.Compose([ 81 | transforms.ToTensor(), 82 | transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), 83 | ]) 84 | trainloader, testloader = get_dataset(args, transform_train, transform_test) 85 | 86 | if args.warmup: 87 | args.lr = args.lr / 10 88 | 89 | # optimizer = DownpourSGD(net.parameters(), lr=args.lr, n_push=args.num_push, n_pull=args.num_pull, model=net) 90 | # scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, cooldown=1, verbose=True, factor=0.25) 91 | scheduler = MultiStepLR(optimizer, milestones=[30, 40], gamma=0.1) 92 | if args.warmup: 93 | scheduler = GradualWarmupScheduler(optimizer, multiplier=10, total_epoch=4, 94 | after_scheduler=scheduler) 95 | compress_ratio = [0.01] * (args.epochs + 10) 96 | compress_ratio[1:4] = [0.25, 0.0625, 0.0625 * 0.25, 0.01] 97 | # train 98 | net.train() 99 | 100 | for epoch in range(1, args.epochs + 1): # loop over the dataset multiple times 101 | # scheduler.step() 102 | 103 | if not args.no_distributed: 104 | optimizer.compress_ratio = compress_ratio[epoch] 105 | print("Training for epoch {}, lr={}".format(epoch, scheduler.optimizer.param_groups[0]['lr'])) 106 | net.train() 107 | # set distributed_sampler.epoch to shuffle data. 108 | trainloader.sampler.set_epoch(epoch) 109 | start = time.time() 110 | for i, data in enumerate(trainloader, 0): 111 | # get the inputs 112 | inputs, labels = data 113 | 114 | if args.cuda: 115 | inputs, labels = inputs.cuda(), labels.cuda() 116 | if args.no_distributed and args.half: 117 | inputs = inputs.half() 118 | 119 | # zero the parameter gradients 120 | optimizer.zero_grad() 121 | # forward + backward + optimize 122 | outputs = net(inputs) 123 | loss = F.cross_entropy(outputs, labels) 124 | loss.backward() 125 | optimizer.step() 126 | _, predicted = torch.max(outputs, 1) 127 | accuracy = accuracy_score(predicted.cpu(), labels.cpu()) 128 | 129 | log_obj = { 130 | 'timestamp': datetime.now(), 131 | 'iteration': i, 132 | 'training_loss': loss.item(), 133 | 'training_accuracy': accuracy, 134 | } 135 | if i % 80 == 0: 136 | print("Timestamp: {timestamp} | " 137 | "Iteration: {iteration:6} | " 138 | "Loss: {training_loss:6.4f} | " 139 | "Accuracy : {training_accuracy:6.4f} | ".format(**log_obj)) 140 | 141 | logs.append(log_obj) 142 | if args.no_distributed or args.rank == 1: 143 | # scheduler.step(logs[-1]['test_loss']) 144 | scheduler.step(epoch) 145 | if True: # print every n mini-batches 146 | end = time.time() 147 | print('minibatch cost :%f, time cost: %f' % ((end - start) / (781 / (args.world_size - 1)), (end - start))) 148 | logs[-1]['test_loss'], logs[-1]['test_accuracy'] = evaluate(net, testloader, args) 149 | print("Timestamp: {timestamp} | " 150 | "Iteration: {iteration:6} | " 151 | "Loss: {training_loss:6.4f} | " 152 | "Accuracy : {training_accuracy:6.4f} | " 153 | "Test Loss: {test_loss:6.4f} | " 154 | "Test Accuracy: {test_accuracy:6.4f}".format(**logs[-1]) 155 | ) 156 | # val_loss, val_accuracy = evaluate(net, testloader, args, verbose=True) 157 | 158 | df = pd.DataFrame(logs) 159 | with open(WORKPATH + '/running.log', 'a+') as f: 160 | running_log = '{},node{}_{}_{}_m{}_e{}_{}.csv'.format(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()), 161 | args.rank - 1, args.mode, 162 | args.model, args.momentum, 163 | epoch, 164 | logs[-1]['test_accuracy']) 165 | f.write(running_log + '\n') 166 | 167 | print(df) 168 | if args.no_distributed: 169 | if args.cuda: 170 | df.to_csv( 171 | WORKPATH + '/log/gpu_{}_{}_m{}_e{}_b{}_{}.csv'.format(args.mode, args.model, args.momentum, args.epochs, 172 | args.batch_size, logs[-1]['test_accuracy']), 173 | index_label='index') 174 | else: 175 | df.to_csv(WORKPATH + '/log/single.csv', index_label='index') 176 | else: 177 | df.to_csv(WORKPATH + '/log/node{}_{}_{}_m{}_e{}_b{}_{}worker_dual_{}.csv'.format(args.rank - 1, args.mode, 178 | args.model, args.momentum, 179 | args.epochs, 180 | args.batch_size, 181 | args.world_size - 1, 182 | logs[-1]['test_accuracy']), 183 | index_label='index') 184 | print('Finished Training') 185 | 186 | 187 | def evaluate(net, testloader, args, verbose=False): 188 | if args.dataset == 'MNIST': 189 | classes = [str(i) for i in range(10)] 190 | else: 191 | classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') 192 | net.eval() 193 | total = 0 194 | correct = 0 195 | test_loss = 0 196 | with torch.no_grad(): 197 | for data in testloader: 198 | images, labels = data 199 | if args.cuda: 200 | images, labels = images.cuda(), labels.cuda() 201 | if args.no_distributed and args.half: 202 | images = images.half() 203 | outputs = net(images) 204 | _, predicted = torch.max(outputs, 1) 205 | test_loss += F.cross_entropy(outputs, labels).item() 206 | total += labels.size(0) 207 | correct += (predicted == labels).sum() 208 | 209 | fake_test_accuracy = accuracy_score(predicted.cpu(), labels.cpu()) 210 | test_accuracy = correct.item() / total 211 | print('%f,%f,%f|%f,%s' % ( 212 | test_accuracy, correct.item(), total, fake_test_accuracy, str((predicted == labels).sum()))) 213 | if verbose: 214 | print('Loss: {:.3f}'.format(test_loss)) 215 | print('Accuracy: {:.3f}'.format(test_accuracy)) 216 | print(classification_report(predicted, labels, target_names=classes)) 217 | 218 | return test_loss, test_accuracy 219 | 220 | 221 | if __name__ == "__main__": 222 | parser = argparse.ArgumentParser(description='Distbelief training example') 223 | 224 | parser.add_argument('--test-batch-size', type=int, default=20000, metavar='N', 225 | help='input batch size for testing (default: 10000)') 226 | parser.add_argument('--epochs', type=int, default=50, metavar='N', help='number of epochs to train (default: 20)') 227 | parser.add_argument('--batch-size', type=int, default=64, metavar='N', 228 | help='input batch size for training (default: 64)') 229 | parser.add_argument('--lr', type=float, default=0.1, metavar='LR', help='learning rate (default: 0.1)') 230 | parser.add_argument('--momentum', type=float, default=0.9, metavar='momentum', help='momentum (default: 0.9') 231 | parser.add_argument('--wd', '--weight-decay', default=5e-4, type=float, 232 | metavar='W', help='weight decay (default: 5e-4)', 233 | dest='weight_decay') 234 | parser.add_argument('--cuda', action='store_true', default=True, help='use CUDA for training') 235 | parser.add_argument('--warmup', action='store_true', default=False, help='use warmup or not') 236 | parser.add_argument('--log-interval', type=int, default=10, metavar='N', help='how often to evaluate and print out') 237 | parser.add_argument('--no-distributed', action='store_true', default=False, 238 | help='whether to use DownpourSGD or normal SGD') 239 | parser.add_argument('--rank', type=int, default=0, metavar='N', 240 | help='rank of current process (0 is server, 1+ is training node)') 241 | parser.add_argument('--world-size', type=int, default=3, metavar='N', help='size of the world') 242 | # parser.add_argument('--server', action='store_true', default=False, help='server node?') 243 | parser.add_argument('--dataset', type=str, default='cifar10', help='which dataset to train on') 244 | parser.add_argument('--master', type=str, default='localhost', help='ip address of the master (server) node') 245 | parser.add_argument('--port', type=str, default='29500', help='port on master node to communicate with') 246 | parser.add_argument('--mode', type=str, default='gradient_sgd', help='gradient_sgd, dgc, Aji or asgd') 247 | parser.add_argument('--model', type=str, default='ResNet18', help='AlexNet, ResNet18, ResNet50') 248 | parser.add_argument('--network-interface', type=str, default='enp3s0', 249 | help='By default, Gloo backends will try to find the right network interface to use. ' 250 | 'If the automatically detected interface is not correct, you can override it ') 251 | args = parser.parse_args() 252 | if args.network_interface: 253 | os.environ['GLOO_SOCKET_IFNAME'] = args.network_interface 254 | if args.cuda: 255 | if socket.gethostname() == 'yan-pc': 256 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 1) 257 | elif 'gn' in socket.gethostname(): 258 | print('init in th') 259 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 4) 260 | else: 261 | os.environ['CUDA_VISIBLE_DEVICES'] = '%d' % (args.rank % 2) 262 | # os.environ['CUDA_VISIBLE_DEVICES'] = '1' 263 | print('Using device%s, device count:%d' % (os.environ['CUDA_VISIBLE_DEVICES'], torch.cuda.device_count())) 264 | 265 | net = None 266 | 267 | if args.dataset == 'cifar10': 268 | args.warmup = False 269 | net = init_net(args) 270 | print('MODEL:%s, momentum:%f' % (args.model, args.momentum)) 271 | assert net is not None 272 | constant.MODEL_SIZE = ravel_model_params(net).numel() 273 | if args.rank == 0 and not args.no_distributed: 274 | if args.cuda is False: 275 | print('server init in cpu') 276 | init_server(args, net) 277 | else: 278 | optimizer = GradientSGD(net.parameters(), lr=args.lr, model=net, momentum=args.momentum, 279 | weight_decay=args.weight_decay, 280 | args=args) 281 | cifar10(args, optimizer, net) 282 | -------------------------------------------------------------------------------- /example/graph.py: -------------------------------------------------------------------------------- 1 | """ 2 | plots accuracy (test and train) vs. time 3 | """ 4 | import matplotlib as mpl 5 | mpl.use('TkAgg') 6 | 7 | import matplotlib.pyplot as plt 8 | import pandas as pd 9 | 10 | colors = ['blue', 'green', 'red', 'orange', 'magenta'] 11 | files_to_read = ['log/single.csv', 'log/gpu.csv', 'log/node1.csv', 'log/node2.csv', 'log/node3.csv'] 12 | log_dataframes = list(map(pd.read_csv, files_to_read)) 13 | 14 | for df in log_dataframes: 15 | df['timestamp'] = pd.to_datetime(df['timestamp']) 16 | df['timestamp'] -= df['timestamp'].min() 17 | 18 | 19 | def plot_train(df, label, color): 20 | plt.plot(df['timestamp'].dt.seconds / 3600.0, 21 | df['training_accuracy'].rolling(50).mean(), 22 | label=label, 23 | color=color) 24 | 25 | def plot_test(df, label, color): 26 | plt.plot(df.dropna()['timestamp'].dt.seconds / 3600.0, 27 | df.dropna()['test_accuracy'].rolling(5).mean(), 28 | label=label, 29 | color=color) 30 | 31 | 32 | fig1 = plt.figure(figsize=(20, 10)) 33 | 34 | for color, filename, df in zip(colors, files_to_read, log_dataframes): 35 | plot_train(df, filename, color) 36 | 37 | plt.ylabel('Training Accuracy') 38 | plt.xlabel('Time (hours)') 39 | plt.legend() 40 | plt.title("Training Accuracy vs. Time (50 iteration rolling average, freq: 3, lr: 0.1)") 41 | plt.savefig('train_time.png') 42 | 43 | fig = plt.figure(figsize=(20, 10)) 44 | 45 | for color, filename, df in zip(colors, files_to_read, log_dataframes): 46 | plot_test(df, filename, color) 47 | 48 | plt.ylabel('Test Accuracy') 49 | plt.xlabel('Time (hours)') 50 | plt.legend() 51 | plt.title("Test Accuracy vs. Time (5 iteration rolling average, freq: 3, lr: 0.1)") 52 | plt.savefig('test_time.png') 53 | -------------------------------------------------------------------------------- /example/main.py: -------------------------------------------------------------------------------- 1 | import os 2 | import sys 3 | 4 | WORKPATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) 5 | print(WORKPATH) 6 | sys.path.append(WORKPATH) 7 | 8 | from core.utils.serialization import ravel_model_params 9 | 10 | from core.utils import constant 11 | import torch.distributed as dist 12 | from core.server import GradientServer 13 | 14 | 15 | def init_server(args, net): 16 | print('init server') 17 | dist.init_process_group('gloo', init_method='file://%s/sharedfile' % WORKPATH, group_name='mygroup', 18 | world_size=args.world_size, rank=args.rank) 19 | 20 | if args.cuda: 21 | model = net.cuda() 22 | else: 23 | model = net 24 | size_list = [i.data.numel() for i in net.parameters()] 25 | threads_num = dist.get_world_size() - 1 26 | threads = [] 27 | global_model = ravel_model_params(model) 28 | constant.MODEL_SIZE = global_model.numel() 29 | synced_model = global_model.clone() 30 | for i in range(1, threads_num + 1): 31 | th = GradientServer(model=model, rank=i, worker_num=args.world_size, global_model=global_model, 32 | synced_model=synced_model, size_list=size_list, args=args) 33 | threads.append(th) 34 | th.start() 35 | for t in threads: 36 | t.join() 37 | -------------------------------------------------------------------------------- /example/models.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | 4 | class LeNet(nn.Module): 5 | def __init__(self): 6 | super(LeNet, self).__init__() 7 | self.conv1 = nn.Conv2d(3, 6, kernel_size=5) 8 | self.conv2 = nn.Conv2d(6, 16, kernel_size=5) 9 | self.conv2_drop = nn.Dropout2d() 10 | self.fc1 = nn.Linear(16 * 5 * 5, 120) 11 | self.fc2 = nn.Linear(120, 84) 12 | self.fc3 = nn.Linear(84, 10) 13 | 14 | def forward(self, x): 15 | x = F.relu(F.max_pool2d(self.conv1(x), 2)) 16 | x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) 17 | x = x.view(-1, 16 * 5 * 5) 18 | x = F.relu(self.fc1(x)) 19 | x = F.dropout(x, training=self.training) 20 | x = F.relu(self.fc2(x)) 21 | x = self.fc3(x) 22 | return x 23 | 24 | 25 | class ConvBN(nn.Module): 26 | def __init__(self, c_in, c_out, bn_weight_init=1.0): 27 | super().__init__() 28 | self.conv = nn.Conv2d(c_in, c_out, kernel_size=3, stride=1, padding=1, bias=False) 29 | self.bn = nn.BatchNorm2d(c_out) 30 | self.bn.weight.data.fill_(bn_weight_init) 31 | self.relu = nn.ReLU(True) 32 | 33 | def forward(self, inputs): 34 | return self.relu(self.bn(self.conv(inputs))) 35 | 36 | 37 | class ResBlk(nn.Module): 38 | def __init__(self, c_in, c_out, pool, res=False): 39 | super().__init__() 40 | self.conv_bn = ConvBN(c_in, c_out) 41 | self.pool = pool 42 | self.res = res 43 | if self.res: 44 | self.res1 = ConvBN(c_out, c_out) 45 | self.res2 = ConvBN(c_out, c_out) 46 | 47 | def forward(self, inputs): 48 | h = self.pool(self.conv_bn(inputs)) 49 | if self.res: 50 | h = h + self.res2(self.res1(h)) 51 | return h 52 | 53 | 54 | class DavidNet(nn.Module): 55 | def __init__(self, c=64, weight=0.125): 56 | super().__init__() 57 | pool = nn.MaxPool2d(2) 58 | self.init_conv_bn = ConvBN(3, c) 59 | self.blk1 = ResBlk(c, c * 2, pool, res=True) 60 | # self.blk1 = ResBlk(c, c*2, pool) 61 | self.blk2 = ResBlk(c * 2, c * 4, pool) 62 | self.blk3 = ResBlk(c * 4, c * 8, pool, res=True) 63 | # self.blk3 = ResBlk(c*4, c*8, pool) 64 | self.pool = nn.AdaptiveMaxPool2d((1, 1)) 65 | self.linear = nn.Linear(c * 8, 10, bias=False) 66 | 67 | self.weight = weight 68 | self.loss = nn.CrossEntropyLoss(reduction='sum') 69 | 70 | def to_dev(self, device): 71 | self.to(device) 72 | self.blk1.to(device) 73 | self.blk2.to(device) 74 | self.blk3.to(device) 75 | return self 76 | 77 | def forward(self, x): 78 | h = self.pool(self.blk3(self.blk2(self.blk1(self.init_conv_bn(x))))) 79 | h = h.view(h.size(0), h.size(1)) 80 | h = self.linear(h) * self.weight 81 | # loss = self.loss(h, y) 82 | # correct = (h.max(dim = 1)[1] == y).sum() 83 | return h 84 | 85 | # class AlexNet(nn.Module): 86 | # def __init__(self, num_classes=10): 87 | # super(AlexNet, self).__init__() 88 | # self.features = nn.Sequential( 89 | # nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1), 90 | # nn.ReLU(inplace=True), 91 | # nn.MaxPool2d(kernel_size=2), 92 | # nn.Conv2d(64, 192, kernel_size=3, padding=1), 93 | # nn.ReLU(inplace=True), 94 | # nn.MaxPool2d(kernel_size=2), 95 | # nn.Conv2d(192, 384, kernel_size=3, padding=1), 96 | # nn.ReLU(inplace=True), 97 | # nn.Conv2d(384, 256, kernel_size=3, padding=1), 98 | # nn.ReLU(inplace=True), 99 | # nn.Conv2d(256, 256, kernel_size=3, padding=1), 100 | # nn.ReLU(inplace=True), 101 | # nn.MaxPool2d(kernel_size=2), 102 | # ) 103 | # self.classifier = nn.Sequential( 104 | # nn.Dropout(), 105 | # nn.Linear(256 * 2 * 2, 4096), 106 | # nn.ReLU(inplace=True), 107 | # nn.Dropout(), 108 | # nn.Linear(4096, 4096), 109 | # nn.ReLU(inplace=True), 110 | # nn.Linear(4096, num_classes), 111 | # ) 112 | # 113 | # def forward(self, x): 114 | # x = self.features(x) 115 | # x = x.view(x.size(0), 256 * 2 * 2) 116 | # x = self.classifier(x) 117 | # return x 118 | class AlexNet(nn.Module): 119 | def __init__(self, num_classes=10): 120 | super(AlexNet, self).__init__() 121 | self.features = nn.Sequential( 122 | nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=5), 123 | nn.ReLU(inplace=True), 124 | nn.MaxPool2d(kernel_size=2, stride=2), 125 | nn.Conv2d(64, 192, kernel_size=5, padding=2), 126 | nn.ReLU(inplace=True), 127 | nn.MaxPool2d(kernel_size=2, stride=2), 128 | nn.Conv2d(192, 384, kernel_size=3, padding=1), 129 | nn.ReLU(inplace=True), 130 | nn.Conv2d(384, 256, kernel_size=3, padding=1), 131 | nn.ReLU(inplace=True), 132 | nn.Conv2d(256, 256, kernel_size=3, padding=1), 133 | nn.ReLU(inplace=True), 134 | nn.MaxPool2d(kernel_size=2, stride=2), 135 | ) 136 | self.classifier = nn.Linear(256, num_classes) 137 | 138 | def forward(self, x): 139 | x = self.features(x) 140 | x = x.view(x.size(0), -1) 141 | x = self.classifier(x) 142 | return x 143 | 144 | import torch 145 | import torch.nn as nn 146 | import torch.nn.functional as F 147 | 148 | 149 | class BasicBlock(nn.Module): 150 | expansion = 1 151 | 152 | def __init__(self, in_planes, planes, stride=1): 153 | super(BasicBlock, self).__init__() 154 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 155 | self.bn1 = nn.BatchNorm2d(planes) 156 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) 157 | self.bn2 = nn.BatchNorm2d(planes) 158 | 159 | self.shortcut = nn.Sequential() 160 | if stride != 1 or in_planes != self.expansion * planes: 161 | self.shortcut = nn.Sequential( 162 | nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False), 163 | nn.BatchNorm2d(self.expansion * planes) 164 | ) 165 | 166 | def forward(self, x): 167 | out = F.relu(self.bn1(self.conv1(x))) 168 | out = self.bn2(self.conv2(out)) 169 | out += self.shortcut(x) 170 | out = F.relu(out) 171 | return out 172 | 173 | 174 | class Bottleneck(nn.Module): 175 | expansion = 4 176 | 177 | def __init__(self, in_planes, planes, stride=1): 178 | super(Bottleneck, self).__init__() 179 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) 180 | self.bn1 = nn.BatchNorm2d(planes) 181 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 182 | self.bn2 = nn.BatchNorm2d(planes) 183 | self.conv3 = nn.Conv2d(planes, self.expansion * planes, kernel_size=1, bias=False) 184 | self.bn3 = nn.BatchNorm2d(self.expansion * planes) 185 | 186 | self.shortcut = nn.Sequential() 187 | if stride != 1 or in_planes != self.expansion * planes: 188 | self.shortcut = nn.Sequential( 189 | nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False), 190 | nn.BatchNorm2d(self.expansion * planes) 191 | ) 192 | 193 | def forward(self, x): 194 | out = F.relu(self.bn1(self.conv1(x))) 195 | out = F.relu(self.bn2(self.conv2(out))) 196 | out = self.bn3(self.conv3(out)) 197 | out += self.shortcut(x) 198 | out = F.relu(out) 199 | return out 200 | 201 | 202 | class ResNet(nn.Module): 203 | def __init__(self, block, num_blocks, num_classes=10): 204 | super(ResNet, self).__init__() 205 | self.in_planes = 64 206 | 207 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) 208 | self.bn1 = nn.BatchNorm2d(64) 209 | self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) 210 | self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) 211 | self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) 212 | self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) 213 | self.linear = nn.Linear(512 * block.expansion, num_classes) 214 | 215 | def _make_layer(self, block, planes, num_blocks, stride): 216 | strides = [stride] + [1] * (num_blocks - 1) 217 | layers = [] 218 | for stride in strides: 219 | layers.append(block(self.in_planes, planes, stride)) 220 | self.in_planes = planes * block.expansion 221 | return nn.Sequential(*layers) 222 | 223 | def forward(self, x): 224 | out = F.relu(self.bn1(self.conv1(x))) 225 | out = self.layer1(out) 226 | out = self.layer2(out) 227 | out = self.layer3(out) 228 | out = self.layer4(out) 229 | out = F.avg_pool2d(out, 4) 230 | out = out.view(out.size(0), -1) 231 | out = self.linear(out) 232 | return out 233 | 234 | 235 | def ResNet18(): 236 | return ResNet(BasicBlock, [2, 2, 2, 2]) 237 | 238 | 239 | def ResNet34(): 240 | return ResNet(BasicBlock, [3, 4, 6, 3]) 241 | 242 | 243 | def ResNet50(): 244 | return ResNet(Bottleneck, [3, 4, 6, 3]) 245 | 246 | 247 | def ResNet101(): 248 | return ResNet(Bottleneck, [3, 4, 23, 3]) 249 | 250 | 251 | def ResNet152(): 252 | return ResNet(Bottleneck, [3, 8, 36, 3]) 253 | 254 | 255 | def test(): 256 | net = ResNet18() 257 | y = net(torch.randn(1, 3, 32, 32)) 258 | print(y.size()) 259 | -------------------------------------------------------------------------------- /example/pssh_script.py: -------------------------------------------------------------------------------- 1 | # from __future__ import print_function 2 | import argparse 3 | import threading 4 | 5 | from pssh.clients import ParallelSSHClient 6 | from pssh.utils import enable_host_logger 7 | 8 | enable_host_logger() 9 | 10 | stdout = [] 11 | 12 | 13 | def Print(host, host_out): 14 | for line in host_out.stdout: 15 | try: 16 | stdout.append(line) 17 | except Exception as e: 18 | print(e) 19 | 20 | 21 | if __name__ == '__main__': 22 | parser = argparse.ArgumentParser(description='Distbelief training example') 23 | parser.add_argument('--where', type=str, default='522', help='522 or th') 24 | args = parser.parse_args() 25 | # where = '522' 26 | print(args) 27 | if args.where == '522': 28 | threads = [] 29 | server = ['192.168.3.100'] 30 | worker = ['192.168.3.101', '192.168.3.102', '192.168.3.103', '192.168.3.104'] 31 | process_per_worker = 1 32 | 33 | # server 34 | host_args = ['--rank %d' % 0] 35 | client = ParallelSSHClient(server, timeout=10000, proxy_host='172.18.233.41', proxy_user='yan', 36 | proxy_port=10000, ) 37 | command = '/home/yan/anaconda3/envs/torch1.3/bin/python /share/DGS/example/cifar10.py --network-interface enp3s0 --world-size ' + str( 38 | len(worker) * process_per_worker + 1) + ' %s' 39 | output = client.run_command(command, host_args=host_args, use_pty=True, timeout=10000) 40 | for host, host_out in output.items(): 41 | t = threading.Thread(target=Print, args=(host, host_out)) 42 | t.start() 43 | threads.append(t) 44 | # worker 45 | for i in range(process_per_worker): 46 | host_args = ['--rank %d' % (j * process_per_worker - i) for j in range(1, len(worker) + 1)] 47 | print(host_args) 48 | client = ParallelSSHClient(worker, timeout=10000, proxy_host='172.18.233.41', proxy_user='yan', 49 | proxy_port=10000, ) 50 | command = '/home/yan/anaconda3/envs/torch1.3/bin/python /share/DGS/example/cifar10.py --network-interface enp3s0 --mode aji --world-size ' + str( 51 | len(worker) * process_per_worker + 1) + ' %s' 52 | output = client.run_command(command, host_args=host_args, use_pty=True, timeout=10000) 53 | for host, host_out in output.items(): 54 | t = threading.Thread(target=Print, args=(host, host_out)) 55 | t.start() 56 | threads.append(t) 57 | for thread in threads: 58 | thread.join() 59 | exit(12580) 60 | elif args.where == 'th': 61 | pass 62 | elif args.where == 'v100': 63 | threads = [] 64 | server = ['gpu2'] 65 | worker = ['gpu9', 'gpu15', 'gpu18', 'gpu26'] 66 | process_per_worker = 4 67 | # server 68 | host_args = ['--rank %d' % 0] 69 | client = ParallelSSHClient(server, timeout=10000) 70 | command = '/GPUFS/app_GPU/application/anaconda3/5.3.1/envs/pytorch-py36/bin/python /GPUFS/sysu_wgwu_8/GradientServer/DGS/example/Imagenet_dist.py -data /GPUFS/sysu_wgwu_8/ImageNet --world-size ' + str( 71 | len(worker) * process_per_worker + 1) + ' %s' 72 | output = client.run_command(command, host_args=host_args, use_pty=True, timeout=10000) 73 | for host, host_out in output.items(): 74 | t = threading.Thread(target=Print, args=(host, host_out)) 75 | t.start() 76 | threads.append(t) 77 | # worker 78 | for i in range(1, process_per_worker + 1): 79 | host_args = ['--rank %d' % j for j in range(i, i + len(worker) * process_per_worker, process_per_worker)] 80 | print(host_args) 81 | client = ParallelSSHClient(worker, timeout=10000) 82 | command = '/GPUFS/app_GPU/application/anaconda3/5.3.1/envs/pytorch-py36/bin/python /GPUFS/sysu_wgwu_8/GradientServer/DGS/example/Imagenet_dist.py -data /GPUFS/sysu_wgwu_8/ImageNet --mode aji --epoch 90 --momentum 0.45 --world-size ' + str( 83 | len(worker) * process_per_worker + 1) + ' %s' 84 | output = client.run_command(command, host_args=host_args, use_pty=True, timeout=10000) 85 | for host, host_out in output.items(): 86 | t = threading.Thread(target=Print, args=(host, host_out)) 87 | t.start() 88 | threads.append(t) 89 | for thread in threads: 90 | thread.join() 91 | exit(12580) 92 | -------------------------------------------------------------------------------- /example/tinyimagenet.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | from datetime import datetime 4 | 5 | import torch 6 | import torch.nn as nn 7 | import torch.nn.parallel 8 | import torch.optim 9 | import torch.utils.data 10 | import torch.utils.data.distributed 11 | import torchvision.datasets as datasets 12 | import torchvision.models as models 13 | import torchvision.transforms as transforms 14 | 15 | parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') 16 | # parser.add_argument('data', metavar='DIR', 17 | # help='path to dataset') 18 | parser.add_argument('--gpu', default=None, type=int, 19 | help='GPU id to use.') 20 | 21 | 22 | def adjust_learning_rate(optimizer, epoch): 23 | """Sets the learning rate to the initial LR decayed by 10 every 30 epochs""" 24 | lr = 0.01 * (0.1 ** (epoch // 30)) 25 | for param_group in optimizer.param_groups: 26 | param_group['lr'] = lr 27 | 28 | 29 | def accuracy(output, target, topk=(1,)): 30 | """Computes the accuracy over the k top predictions for the specified values of k""" 31 | with torch.no_grad(): 32 | maxk = max(topk) 33 | batch_size = target.size(0) 34 | 35 | _, pred = output.topk(maxk, 1, True, True) 36 | pred = pred.t() 37 | correct = pred.eq(target.view(1, -1).expand_as(pred)) 38 | 39 | res = [] 40 | for k in topk: 41 | correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) 42 | res.append(correct_k.mul_(100.0 / batch_size)) 43 | return res 44 | 45 | 46 | def train(train_loader, model, criterion, optimizer, epoch): 47 | # switch to train mode 48 | model.train() 49 | losses = list() 50 | top1 = list() 51 | top5 = list() 52 | 53 | for i, (input, target) in enumerate(train_loader): 54 | 55 | # compute output 56 | input, target = input.cuda(), target.cuda() 57 | output = model(input) 58 | loss = criterion(output, target) 59 | 60 | losses.append(loss.item()) 61 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 62 | top1.append(acc1[0].item()) 63 | top5.append(acc5[0].item()) 64 | 65 | optimizer.zero_grad() 66 | loss.backward() 67 | optimizer.step() 68 | 69 | if i % 1 == 0: 70 | print(str(datetime.now()) + ' Epoch: [{0}][{1}/{2}]\t' 71 | 'Loss {loss:.4f} ({loss_avg:.4f})\t' 72 | 'Acc@1 {top1:.3f} ({top1_avg:.3f})\t' 73 | 'Acc@5 {top5:.3f} ({top5_avg:.3f})'.format( 74 | epoch, i, len(train_loader), loss=loss.item(), loss_avg=sum(losses) / float(len(losses)), 75 | top1=acc1[0].item(), top1_avg=sum(top1) / float(len(top1)), 76 | top5=acc5[0].item(), top5_avg=sum(top5) / float(len(top5)))) 77 | 78 | 79 | def validate(test_loader, model, criterion): 80 | # switch to train mode 81 | model.eval() 82 | losses = list() 83 | top1 = list() 84 | top5 = list() 85 | 86 | for i, (input, target) in enumerate(test_loader): 87 | 88 | # compute output 89 | input, target = input.cuda(), target.cuda() 90 | output = model(input) 91 | loss = criterion(output, target) 92 | 93 | losses.append(loss.item()) 94 | acc1, acc5 = accuracy(output, target, topk=(1, 5)) 95 | top1.append(acc1[0].item()) 96 | top5.append(acc5[0].item()) 97 | 98 | if i % 10 == 0: 99 | print('[{0}/{1}]\t' 100 | 'Loss {loss:.4f} ({loss_avg:.4f})\t' 101 | 'Acc@1 {top1:.3f} ({top1_avg:.3f})\t' 102 | 'Acc@5 {top5:.3f} ({top5_avg:.3f})'.format( 103 | i, len(test_loader), loss=loss.item(), loss_avg=sum(losses) / float(len(losses)), 104 | top1=acc1[0].item(), top1_avg=sum(top1) / float(len(top1)), 105 | top5=acc5[0].item(), top5_avg=sum(top5) / float(len(top5)))) 106 | 107 | print(' * Acc@1 {top1_avg:.3f} Acc@5 {top5_avg:.3f}' 108 | .format(top1_avg=sum(top1) / float(len(top1)), top5_avg=sum(top5) / float(len(top5)))) 109 | 110 | 111 | def main(): 112 | args = parser.parse_args() 113 | args.gpu = 0 114 | # args.data = '/share/DGS/data/tiny-imagenet-200/' 115 | args.data = '/home/yan/data/' 116 | traindir = os.path.join(args.data, 'train') # /train/ を指定されたパスに追加 117 | testdir = os.path.join(args.data, 'val') 118 | normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], 119 | std=[0.229, 0.224, 0.225]) # 正規化定数 120 | 121 | train_dataset = datasets.ImageFolder( 122 | traindir, 123 | transforms.Compose([ 124 | transforms.RandomResizedCrop(224), # 画像をサイズ224に切り出しもしくはリサイズ 125 | transforms.RandomHorizontalFlip(), # ランダムに画像をフリップ（水増し） 126 | transforms.ToTensor(), 127 | normalize, 128 | ])) 129 | 130 | train_loader = torch.utils.data.DataLoader( 131 | train_dataset, batch_size=256, shuffle=True, 132 | num_workers=4, pin_memory=True) 133 | 134 | test_loader = torch.utils.data.DataLoader( 135 | datasets.ImageFolder(testdir, transforms.Compose([ 136 | transforms.Resize(256), 137 | transforms.CenterCrop(224), 138 | transforms.ToTensor(), 139 | normalize, 140 | ])), 141 | batch_size=64, shuffle=False, 142 | num_workers=4, pin_memory=True) 143 | 144 | model = models.resnet18(num_classes=1000) 145 | # model = EfficientNet.from_name('efficientnet-b0') 146 | criterion = nn.CrossEntropyLoss().cuda(args.gpu) 147 | optimizer = torch.optim.SGD(model.parameters(), 0.01, 148 | momentum=0.9, 149 | weight_decay=1e-4) 150 | 151 | if args.gpu is not None: 152 | model = model.cuda(args.gpu) 153 | 154 | for epoch in range(0, 100): 155 | adjust_learning_rate(optimizer, epoch) 156 | 157 | # train for one epoch 158 | train(train_loader, model, criterion, optimizer, epoch) 159 | 160 | # evaluate on validation set 161 | acc1 = validate(test_loader, model, criterion) 162 | 163 | 164 | main() 165 | -------------------------------------------------------------------------------- /notebook/Untitled.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": 6, 6 | "metadata": {}, 7 | "outputs": [ 8 | { 9 | "data": { 10 | "text/html": [ 11 | "" 12 | ], 13 | "text/vnd.plotly.v1+html": [ 14 | "" 15 | ] 16 | }, 17 | "metadata": {}, 18 | "output_type": "display_data" 19 | }, 20 | { 21 | "data": { 22 | "application/vnd.plotly.v1+json": { 23 | "data": [ 24 | { 25 | "type": "scatter", 26 | "uid": "a27bdfe2-f472-11e8-84d1-7d4a4274f24f", 27 | "x": [ 28 | 1, 29 | 2, 30 | 3, 31 | 4 32 | ], 33 | "y": [ 34 | 4, 35 | 3, 36 | 2, 37 | 1 38 | ] 39 | } 40 | ], 41 | "layout": { 42 | "title": "hello world" 43 | } 44 | }, 45 | "text/html": [ 46 | "
" 47 | ], 48 | "text/vnd.plotly.v1+html": [ 49 | "
" 50 | ] 51 | }, 52 | "metadata": {}, 53 | "output_type": "display_data" 54 | } 55 | ], 56 | "source": [ 57 | "import plotly\n", 58 | "import plotly.graph_objs as go\n", 59 | "\n", 60 | "plotly.offline.init_notebook_mode(connected=True)\n", 61 | "\n", 62 | "plotly.offline.iplot({\n", 63 | " \"data\": [go.Scatter(x=[1, 2, 3, 4], y=[4, 3, 2, 1])],\n", 64 | " \"layout\": go.Layout(title=\"hello world\")\n", 65 | "})" 66 | ] 67 | }, 68 | { 69 | "cell_type": "code", 70 | "execution_count": null, 71 | "metadata": {}, 72 | "outputs": [], 73 | "source": [] 74 | } 75 | ], 76 | "metadata": { 77 | "kernelspec": { 78 | "display_name": "Python 3", 79 | "language": "python", 80 | "name": "python3" 81 | }, 82 | "language_info": { 83 | "codemirror_mode": { 84 | "name": "ipython", 85 | "version": 3 86 | }, 87 | "file_extension": ".py", 88 | "mimetype": "text/x-python", 89 | "name": "python", 90 | "nbconvert_exporter": "python", 91 | "pygments_lexer": "ipython3", 92 | "version": "3.7.0" 93 | } 94 | }, 95 | "nbformat": 4, 96 | "nbformat_minor": 2 97 | } 98 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | bcrypt==3.1.7 2 | certifi==2019.9.11 3 | cffi==1.13.0 4 | cryptography==3.2 5 | future==0.18.2 6 | gevent==1.4.0 7 | greenlet==0.4.15 8 | joblib==0.13.2 9 | numpy==1.17.2 10 | olefile==0.46 11 | pandas==0.25.2 12 | parallel-ssh==1.9.1 13 | paramiko==2.6.0 14 | Pillow==7.1.0 15 | pycparser==2.19 16 | PyNaCl==1.3.0 17 | python-dateutil==2.8.0 18 | pytz==2019.3 19 | scikit-learn==0.21.3 20 | scipy==1.3.1 21 | six==1.12.0 22 | ssh2-python==0.18.0.post1 23 | torch==1.3.0 24 | torchvision==0.4.1a0+d94043a 25 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import setuptools 2 | 3 | with open("README.md", "r") as fh: 4 | long_description = fh.read() 5 | 6 | setuptools.setup( 7 | name="pytorch-DGS", 8 | version="0.1.0", 9 | author="Jesse Cai", 10 | author_email="jcjessecai@gmail.com", 11 | description="Distributed training for pytorch", 12 | long_description=long_description, 13 | long_description_content_type="text/markdown", 14 | url="https://github.com/ucla-labx/DGS", 15 | packages=setuptools.find_packages(), 16 | classifiers=( 17 | "Programming Language :: Python :: 3", 18 | "License :: OSI Approved :: MIT License", 19 | "Operating System :: OS Independent", 20 | ), 21 | ) 22 | --------------------------------------------------------------------------------