├── .gitignore ├── Calculations ├── SELU_calculations.nb └── SELU_calculations.pdf ├── HTRU2 └── README.md ├── LICENSE ├── Pytorch ├── SelfNormalizingNetworks_CNN_CIFAR10.ipynb ├── SelfNormalizingNetworks_CNN_MNIST.ipynb └── SelfNormalizingNetworks_MLP_MNIST.ipynb ├── README.md ├── SNN-successes └── README.md ├── TF_1_x ├── README.md ├── SelfNormalizingNetworks_CNN_CIFAR10.ipynb ├── SelfNormalizingNetworks_CNN_MNIST.ipynb ├── SelfNormalizingNetworks_MLP_MNIST.ipynb ├── environment.yml ├── getSELUparameters.ipynb └── selu.py ├── TF_2_x ├── CIFAR10-Conv-SELU.py ├── MNIST-Conv-SELU.py ├── MNIST-MLP-SELU.py └── README.md ├── Tox21 └── README.md ├── UCI └── README.md ├── environment.yml └── figure1 ├── README.md ├── create_plots.ipynb ├── run.py └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | */.ipynb_checkpoints/* 2 | */checkpoints/* 3 | */data_set/* 4 | */saved_models/* 5 | */logs/* -------------------------------------------------------------------------------- /Calculations/SELU_calculations.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/bioinf-jku/SNNs/b578499301fcb801f8d4135dbd7cebb246722bfc/Calculations/SELU_calculations.pdf -------------------------------------------------------------------------------- /HTRU2/README.md: -------------------------------------------------------------------------------- 1 | ## HTRU2 data set 2 | - [download](https://archive.ics.uci.edu/ml/machine-learning-databases/00372/HTRU2.zip) 3 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Self-Normalizing Networks 2 | Tutorials and implementations for "Self-normalizing networks"(SNNs) as suggested by Klambauer et al. ([arXiv pre-print](https://arxiv.org/pdf/1706.02515.pdf)). 3 | 4 | ## Versions 5 | - see [environment](environment.yml) file for full list of prerequisites. Tutorial implementations use Tensorflow > 2.0 (Keras) or Pytorch, but versions for Tensorflow 1.x 6 | users based on the deprecated tf.contrib module (with separate [environment](TF_1_x/environment.yml) file) are also available. 7 | 8 | #### Note for Tensorflow >= 1.4 users 9 | Tensorflow >= 1.4 already has the function [tf.nn.selu](https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/nn/selu) and [tf.contrib.nn.alpha_dropout](https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/contrib/nn/alpha_dropout) that implement the SELU activation function and the suggested dropout version. 10 | #### Note for Tensorflow >= 2.0 users 11 | Tensorflow 2.3 already has selu activation function when using high level framework keras, [tf.keras.activations.selu](https://www.tensorflow.org/api_docs/python/tf/keras/activations/selu). 12 | Must be combined with [tf.keras.initializers.LecunNormal](https://www.tensorflow.org/api_docs/python/tf/keras/initializers/LecunNormal), corresponding dropout version is [tf.keras.layers.AlphaDropout](https://www.tensorflow.org/api_docs/python/tf/keras/layers/AlphaDropout). 13 | #### Note for Pytorch users 14 | Pytorch versions >= 0.2 feature [torch.nn.SELU](https://pytorch.org/docs/stable/generated/torch.nn.SELU.html#torch.nn.SELU) and [torch.nn.AlphaDropout](https://pytorch.org/docs/stable/generated/torch.nn.AlphaDropout.html#torch.nn.AlphaDropout), they must be combined with the correct initializer, namely [torch.nn.init.kaiming_normal_](https://pytorch.org/docs/stable/nn.init.html#torch.nn.init.kaiming_normal_) (parameter, mode='fan_in', nonlinearity='linear') 15 | as this is identical to lecun initialisation (mode='fan_in') with a gain of 1 (nonlinearity='linear'). 16 | 17 | 18 | ## Tutorials 19 | 20 | ### Tensorflow 1.x 21 | - Multilayer Perceptron on MNIST ([notebook](TF_1_x/SelfNormalizingNetworks_MLP_MNIST.ipynb)) 22 | - Convolutional Neural Network on MNIST ([notebook](TF_1_x/SelfNormalizingNetworks_CNN_MNIST.ipynb)) 23 | - Convolutional Neural Network on CIFAR10 ([notebook](TF_1_x/SelfNormalizingNetworks_CNN_CIFAR10.ipynb)) 24 | 25 | ### Tensorflow 2.x (Keras) 26 | - Multilayer Perceptron on MNIST ([python script](TF_2_x/MNIST-MLP-SELU.py)) 27 | - Convolutional Neural Network on MNIST ([python script](TF_2_x/MNIST-Conv-SELU.py)) 28 | - Convolutional Neural Network on CIFAR10 ([python script](TF_2_x/CIFAR10-Conv-SELU.py)) 29 | 30 | ### Pytorch 31 | 32 | - Multilayer Perceptron on MNIST ([notebook](Pytorch/SelfNormalizingNetworks_MLP_MNIST.ipynb)) 33 | - Convolutional Neural Network on MNIST ([notebook](Pytorch/SelfNormalizingNetworks_CNN_MNIST.ipynb)) 34 | - Convolutional Neural Network on CIFAR10 ([notebook](Pytorch/SelfNormalizingNetworks_CNN_CIFAR10.ipynb)) 35 | 36 | ## Further material 37 | 38 | ### Design novel SELU functions (Tensorflow 1.x) 39 | - How to obtain the SELU parameters alpha and lambda for arbitrary fixed points ([notebook](TF_1_x/getSELUparameters.ipynb)) 40 | 41 | ### Basic python functions to implement SNNs (Tensorflow 1.x) 42 | are provided as code chunks here: [selu.py](TF_1_x/selu.py) 43 | 44 | ### Notebooks and code to produce Figure 1 (Tensorflow 1.x) 45 | are provided here: [Figure1](figure1/), builds on top of the [biutils](https://github.com/untom/biutils) package. 46 | 47 | ### Calculations and numeric checks of the theorems (Mathematica) 48 | are provided as mathematica notebooks here: 49 | 50 | - [Mathematica notebook](Calculations/SELU_calculations.nb) 51 | - [Mathematica PDF](Calculations/SELU_calculations.pdf) 52 | 53 | ### UCI, Tox21 and HTRU2 data sets 54 | 55 | - [UCI](http://persoal.citius.usc.es/manuel.fernandez.delgado/papers/jmlr/data.tar.gz) 56 | - [Tox21](http://bioinf.jku.at/research/DeepTox/tox21.zip) 57 | - [HTRU2](https://archive.ics.uci.edu/ml/machine-learning-databases/00372/HTRU2.zip) 58 | -------------------------------------------------------------------------------- /SNN-successes/README.md: -------------------------------------------------------------------------------- 1 | # Models and architectures built on Self-Normalizing Networks 2 | 3 | ## GANs 4 | - [THINKING LIKE A MACHINE - GENERATING VISUAL RATIONALES WITH WASSERSTEIN GANS](https://pdfs.semanticscholar.org/dd4c/23a21b1199f34e5003e26d2171d02ba12d45.pdf): Both discriminator and generator trained without batch normalization. 5 | - [Deformable Deep Convolutional Generative Adversarial Network in Microwave Based Hand Gesture Recognition System](https://arxiv.org/abs/1711.01968): 6 | The rate between SELU and SELU+BN proves that SELU itself has the convergence quality of BN. 7 | 8 | ## Convolutional neural networks 9 | - [Effectiveness of Self Normalizing Neural Networks for Text Classification](https://arxiv.org/abs/1905.01338): Applied properties of SNN to CNN to give rise to new arechitecture SCNN (Self normalizaing CNN) which performed better than CNN in text classification. 10 | - [Solving internal covariate shift in deep learning with linked neurons](https://arxiv.org/abs/1712.02609): Show that ultra-deep CNNs without batch normalization can only be trained SELUs (except with the suggested method described by the authors). 11 | - [DCASE 2017 ACOUSTIC SCENE CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK IN TIME SERIES](http://www.cs.tut.fi/sgn/arg/dcase2017/documents/challenge_technical_reports/DCASE2017_Biho_116.pdf): Deep CNN trained without batch normalization. 12 | - [Point-wise Convolutional Neural Network](https://arxiv.org/abs/1712.05245): Training with SELU converges faster than training with ReLU; improved accuracy with SELU. 13 | - [Over the Air Deep Learning Based Radio Signal Classification](https://arxiv.org/abs/1712.04578): Slight performance improvement over ReLU. 14 | - [Convolutional neural networks for structured omics: OmicsCNN and the OmicsConv layer](https://arxiv.org/abs/1710.05918): Deep CNN trained without batch normalization. 15 | - [Searching for Activation Functions](https://arxiv.org/abs/1710.05941): ResNet architectures trained with SELUs probably together with batch normalization. 16 | - [EddyNet: A Deep Neural Network For Pixel-Wise Classification of Oceanic Eddies](https://arxiv.org/abs/1711.03954): Fast CNN training with SELUs. ReLU with BN better at final performance but skip connections not handled appropriately. 17 | - [SMILES2Vec: An Interpretable General-Purpose Deep Neural Network for Predicting Chemical Properties](https://arxiv.org/abs/1712.02034): 20-layer ResNet trained with SELUs. 18 | - [Sentiment Analysis of Tweets in Malayalam Using Long Short-Term Memory Units and Convolutional Neural Nets](https://link.springer.com/chapter/10.1007/978-3-319-71928-3_31) 19 | - [RETUYT in TASS 2017: Sentiment Analysis for Spanish Tweets using SVM and CNN](https://arxiv.org/abs/1710.06393) 20 | 21 | 22 | ## FNNs are finally deep 23 | - [Predicting Adolescent Suicide Attempts with Neural Networks](https://arxiv.org/abs/1711.10057): The use of the SELU activation renders batch normalization 24 | unnecessary. 25 | - [Improving Palliative Care with Deep Learning](https://arxiv.org/abs/1711.06402): An 18-layer neural network with SELUs performed best. 26 | - [An Iterative Closest Points Approach to Neural Generative Models](https://arxiv.org/abs/1711.06562) 27 | - [Retrieval of Surface Ozone from UV-MFRSR Irradiances using Deep Learning](http://uvb.nrel.colostate.edu/UVB/publications/AGU-Retrieval-Surface-Ozone-Deep-Learning.pdf): 6-10 layer networks perform best. 28 | 29 | 30 | ## Reinforcement Learning 31 | - [Automated Cloud Provisioning on AWS using Deep Reinforcement Learning](https://arxiv.org/abs/1709.04305): Deep CNN architecture trained with SELUs. 32 | - [Learning to Run with Actor-Critic Ensemble](https://arxiv.org/abs/1712.08987): Second best method (actor-critic ensemble) at the NIPS2017 "Learning to Run" competition. They have 33 | tried several activation functions and found that the activation function of Scaled Exponential Linear Units (SELU) are superior to ReLU, Leaky ReLU, Tanh and Sigmoid. 34 | 35 | ## Autoencoders 36 | - [Replacement AutoEncoder: A Privacy-Preserving Algorithm for Sensory Data Analysis](https://arxiv.org/abs/1710.06564): Deep autoencoder trained with SELUs. 37 | - [Application of generative autoencoder in de novo molecular design](https://arxiv.org/abs/1711.07839): Faster convergence with SELUs. 38 | 39 | ## Recurrent Neural Networks 40 | - [Sentiment extraction from Consumer-generated noisy short texts](http://sentic.net/sentire2017meisheri.pdf): SNNs used in FC layers. 41 | 42 | 43 | -------------------------------------------------------------------------------- /TF_1_x/README.md: -------------------------------------------------------------------------------- 1 | Tensorflow 1.15 version if you happen to work on Tensorflow versions < 2. 2 | Full list of prerequisites is to be found in [environment](environment.yml) file. -------------------------------------------------------------------------------- /TF_1_x/SelfNormalizingNetworks_CNN_MNIST.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": { 6 | "collapsed": false 7 | }, 8 | "source": [ 9 | "# Tutorial on self-normalizing networks on the MNIST data set: convolutional neural networks\n", 10 | "\n", 11 | "*Author:* Guenter Klambauer, 2017\n", 12 | "\n", 13 | "Derived from: [Aymeric Damien](https://github.com/aymericdamien/TensorFlow-Examples/) " 14 | ] 15 | }, 16 | { 17 | "cell_type": "code", 18 | "execution_count": 1, 19 | "metadata": { 20 | "collapsed": false 21 | }, 22 | "outputs": [ 23 | { 24 | "name": "stdout", 25 | "output_type": "stream", 26 | "text": [ 27 | "WARNING:tensorflow:From :17: read_data_sets (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 28 | "Instructions for updating:\n", 29 | "Please use alternatives such as official/mnist/dataset.py from tensorflow/models.\n", 30 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:260: maybe_download (from tensorflow.contrib.learn.python.learn.datasets.base) is deprecated and will be removed in a future version.\n", 31 | "Instructions for updating:\n", 32 | "Please write your own downloading logic.\n", 33 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:262: extract_images (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 34 | "Instructions for updating:\n", 35 | "Please use tf.data to implement this functionality.\n", 36 | "Extracting /tmp/data/train-images-idx3-ubyte.gz\n", 37 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:267: extract_labels (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 38 | "Instructions for updating:\n", 39 | "Please use tf.data to implement this functionality.\n", 40 | "Extracting /tmp/data/train-labels-idx1-ubyte.gz\n", 41 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:110: dense_to_one_hot (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 42 | "Instructions for updating:\n", 43 | "Please use tf.one_hot on tensors.\n", 44 | "Extracting /tmp/data/t10k-images-idx3-ubyte.gz\n", 45 | "Extracting /tmp/data/t10k-labels-idx1-ubyte.gz\n", 46 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:290: DataSet.__init__ (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 47 | "Instructions for updating:\n", 48 | "Please use alternatives such as official/mnist/dataset.py from tensorflow/models.\n", 49 | "1.15.0\n" 50 | ] 51 | } 52 | ], 53 | "source": [ 54 | "import tensorflow as tf\n", 55 | "import numpy as np\n", 56 | "\n", 57 | "import numbers\n", 58 | "from tensorflow.python.framework import ops\n", 59 | "from tensorflow.python.framework import tensor_shape\n", 60 | "from tensorflow.python.framework import tensor_util\n", 61 | "from tensorflow.python.ops import math_ops\n", 62 | "from tensorflow.python.ops import random_ops\n", 63 | "from tensorflow.python.ops import array_ops\n", 64 | "from tensorflow.python.layers import utils\n", 65 | "\n", 66 | "from sklearn.preprocessing import StandardScaler\n", 67 | "\n", 68 | "# Import MNIST data\n", 69 | "from tensorflow.examples.tutorials.mnist import input_data\n", 70 | "mnist = input_data.read_data_sets(\"/tmp/data/\", one_hot=True)\n", 71 | "\n", 72 | "print(tf.__version__)" 73 | ] 74 | }, 75 | { 76 | "cell_type": "markdown", 77 | "metadata": {}, 78 | "source": [ 79 | "### (1) Definition of scaled exponential linear units (SELUs)" 80 | ] 81 | }, 82 | { 83 | "cell_type": "code", 84 | "execution_count": 2, 85 | "metadata": { 86 | "collapsed": true 87 | }, 88 | "outputs": [], 89 | "source": [ 90 | "def selu(x):\n", 91 | " with ops.name_scope('elu') as scope:\n", 92 | " alpha = 1.6732632423543772848170429916717\n", 93 | " scale = 1.0507009873554804934193349852946\n", 94 | " return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x))" 95 | ] 96 | }, 97 | { 98 | "cell_type": "markdown", 99 | "metadata": {}, 100 | "source": [ 101 | "### (2) Definition of dropout variant for SNNs\n" 102 | ] 103 | }, 104 | { 105 | "cell_type": "code", 106 | "execution_count": 3, 107 | "metadata": { 108 | "collapsed": true 109 | }, 110 | "outputs": [], 111 | "source": [ 112 | "def dropout_selu(x, rate, alpha= -1.7580993408473766, fixedPointMean=0.0, fixedPointVar=1.0, \n", 113 | " noise_shape=None, seed=None, name=None, training=False):\n", 114 | " \"\"\"Dropout to a value with rescaling.\"\"\"\n", 115 | "\n", 116 | " def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name):\n", 117 | " keep_prob = 1.0 - rate\n", 118 | " x = ops.convert_to_tensor(x, name=\"x\")\n", 119 | " if isinstance(keep_prob, numbers.Real) and not 0 < keep_prob <= 1:\n", 120 | " raise ValueError(\"keep_prob must be a scalar tensor or a float in the \"\n", 121 | " \"range (0, 1], got %g\" % keep_prob)\n", 122 | " keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name=\"keep_prob\")\n", 123 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 124 | "\n", 125 | " alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name=\"alpha\")\n", 126 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 127 | "\n", 128 | " if tensor_util.constant_value(keep_prob) == 1:\n", 129 | " return x\n", 130 | "\n", 131 | " noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x)\n", 132 | " random_tensor = keep_prob\n", 133 | " random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype)\n", 134 | " binary_tensor = math_ops.floor(random_tensor)\n", 135 | " ret = x * binary_tensor + alpha * (1-binary_tensor)\n", 136 | "\n", 137 | " a = tf.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * tf.pow(alpha-fixedPointMean,2) + fixedPointVar)))\n", 138 | "\n", 139 | " b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha)\n", 140 | " ret = a * ret + b\n", 141 | " ret.set_shape(x.get_shape())\n", 142 | " return ret\n", 143 | "\n", 144 | " with ops.name_scope(name, \"dropout\", [x]) as name:\n", 145 | " return utils.smart_cond(training,\n", 146 | " lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name),\n", 147 | " lambda: array_ops.identity(x))\n" 148 | ] 149 | }, 150 | { 151 | "cell_type": "markdown", 152 | "metadata": {}, 153 | "source": [ 154 | "### (3) Scale input to zero mean and unit variance" 155 | ] 156 | }, 157 | { 158 | "cell_type": "code", 159 | "execution_count": 4, 160 | "metadata": { 161 | "collapsed": false 162 | }, 163 | "outputs": [], 164 | "source": [ 165 | "scaler = StandardScaler().fit(mnist.train.images)" 166 | ] 167 | }, 168 | { 169 | "cell_type": "code", 170 | "execution_count": 5, 171 | "metadata": { 172 | "collapsed": true 173 | }, 174 | "outputs": [], 175 | "source": [ 176 | "# Parameters\n", 177 | "learning_rate = 0.025\n", 178 | "training_iters = 50\n", 179 | "batch_size = 128\n", 180 | "display_step = 1\n", 181 | "\n", 182 | "# Network Parameters\n", 183 | "n_input = 784 # MNIST data input (img shape: 28*28)\n", 184 | "n_classes = 10 # MNIST total classes (0-9 digits)\n", 185 | "keep_prob_ReLU = 0.5 # Dropout, probability to keep units\n", 186 | "dropout_prob_SNN = 0.05 # Dropout, probability to dropout units\n", 187 | "\n", 188 | "# tf Graph input\n", 189 | "x = tf.placeholder(tf.float32, [None, n_input])\n", 190 | "y = tf.placeholder(tf.float32, [None, n_classes])\n", 191 | "keep_prob = tf.placeholder(tf.float32) #dropout (keep probability for ReLU)\n", 192 | "dropout_prob = tf.placeholder(tf.float32) #dropout (dropout probability for SNN)\n", 193 | "is_training = tf.placeholder(tf.bool)" 194 | ] 195 | }, 196 | { 197 | "cell_type": "code", 198 | "execution_count": 6, 199 | "metadata": { 200 | "collapsed": true 201 | }, 202 | "outputs": [], 203 | "source": [ 204 | "# Create some wrappers for simplicity\n", 205 | "def conv2d(x, W, b, strides=1):\n", 206 | " # Conv2D wrapper, with bias and relu activation\n", 207 | " x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')\n", 208 | " x = tf.nn.bias_add(x, b)\n", 209 | " return tf.nn.relu(x)\n", 210 | "\n", 211 | "def conv2d_SNN(x, W, b, strides=1):\n", 212 | " # Conv2D wrapper, with bias and relu activation\n", 213 | " x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')\n", 214 | " x = tf.nn.bias_add(x, b)\n", 215 | " return selu(x)\n", 216 | "\n", 217 | "def maxpool2d(x, k=2):\n", 218 | " # MaxPool2D wrapper\n", 219 | " return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],\n", 220 | " padding='SAME')" 221 | ] 222 | }, 223 | { 224 | "cell_type": "code", 225 | "execution_count": 7, 226 | "metadata": { 227 | "collapsed": true 228 | }, 229 | "outputs": [], 230 | "source": [ 231 | "# Create model\n", 232 | "def conv_net_ReLU(x, weights, biases, keep_prob):\n", 233 | " # Reshape input picture\n", 234 | " x = tf.reshape(x, shape=[-1, 28, 28, 1])\n", 235 | "\n", 236 | " # Convolution Layer\n", 237 | " conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n", 238 | " # Max Pooling (down-sampling)\n", 239 | " conv1 = maxpool2d(conv1, k=2)\n", 240 | "\n", 241 | " # Convolution Layer\n", 242 | " conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n", 243 | " # Max Pooling (down-sampling)\n", 244 | " conv2 = maxpool2d(conv2, k=2)\n", 245 | "\n", 246 | " # Fully connected layer\n", 247 | " # Reshape conv2 output to fit fully connected layer input\n", 248 | " fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])\n", 249 | " fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n", 250 | " fc1 = tf.nn.relu(fc1)\n", 251 | " \n", 252 | " # Apply Dropout\n", 253 | " fc1 = tf.nn.dropout(fc1, keep_prob)\n", 254 | "\n", 255 | " # Output, class prediction\n", 256 | " out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n", 257 | " return out" 258 | ] 259 | }, 260 | { 261 | "cell_type": "code", 262 | "execution_count": 8, 263 | "metadata": { 264 | "collapsed": true 265 | }, 266 | "outputs": [], 267 | "source": [ 268 | "# Create model\n", 269 | "def conv_net_SNN(x, weights, biases, dropout_prob, is_training):\n", 270 | " # Reshape input picture\n", 271 | " x = tf.reshape(x, shape=[-1, 28, 28, 1])\n", 272 | "\n", 273 | " # Convolution Layer\n", 274 | " conv1 = conv2d_SNN(x, weights['wc1'], biases['bc1'],)\n", 275 | " # Max Pooling (down-sampling)\n", 276 | " conv1 = maxpool2d(conv1, k=2)\n", 277 | "\n", 278 | " # Convolution Layer\n", 279 | " conv2 = conv2d_SNN(conv1, weights['wc2'], biases['bc2'])\n", 280 | " # Max Pooling (down-sampling)\n", 281 | " conv2 = maxpool2d(conv2, k=2)\n", 282 | "\n", 283 | " # Fully connected layer\n", 284 | " # Reshape conv2 output to fit fully connected layer input\n", 285 | " fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])\n", 286 | " fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n", 287 | " fc1 = selu(fc1)\n", 288 | " \n", 289 | " # Apply Dropout\n", 290 | " fc1 = dropout_selu(fc1, dropout_prob,training=is_training)\n", 291 | "\n", 292 | " # Output, class prediction\n", 293 | " out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n", 294 | " return out" 295 | ] 296 | }, 297 | { 298 | "cell_type": "code", 299 | "execution_count": 9, 300 | "metadata": { 301 | "collapsed": true 302 | }, 303 | "outputs": [], 304 | "source": [ 305 | "# RELU: Store layers weight & bias\n", 306 | "## Improved with MSRA initialization\n", 307 | "\n", 308 | "weights = {\n", 309 | " # 5x5 conv, 1 input, 32 outputs\n", 310 | " 'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32],stddev=np.sqrt(2/25)) ),\n", 311 | " # 5x5 conv, 32 inputs, 64 outputs\n", 312 | " 'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64],stddev=np.sqrt(2/(25*32)))),\n", 313 | " # fully connected, 7*7*64 inputs, 1024 outputs\n", 314 | " 'wd1': tf.Variable(tf.random_normal([7*7*64, 1024],stddev=np.sqrt(2/(7*7*64)))),\n", 315 | " # 1024 inputs, 10 outputs (class prediction)\n", 316 | " 'out': tf.Variable(tf.random_normal([1024, n_classes],stddev=np.sqrt(2/(1024))))\n", 317 | "}\n", 318 | "\n", 319 | "biases = {\n", 320 | " 'bc1': tf.Variable(tf.random_normal([32],stddev=0)),\n", 321 | " 'bc2': tf.Variable(tf.random_normal([64],stddev=0)),\n", 322 | " 'bd1': tf.Variable(tf.random_normal([1024],stddev=0)),\n", 323 | " 'out': tf.Variable(tf.random_normal([n_classes],stddev=0))\n", 324 | "}" 325 | ] 326 | }, 327 | { 328 | "cell_type": "markdown", 329 | "metadata": {}, 330 | "source": [ 331 | "### (4) Initialization with STDDEV of sqrt(1/n)" 332 | ] 333 | }, 334 | { 335 | "cell_type": "code", 336 | "execution_count": 10, 337 | "metadata": { 338 | "collapsed": false 339 | }, 340 | "outputs": [], 341 | "source": [ 342 | "# SNN: Store layers weight & bias\n", 343 | "weights2 = {\n", 344 | " # 5x5 conv, 1 input, 32 outputs\n", 345 | " 'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32],stddev=np.sqrt(1/25)) ),\n", 346 | " # 5x5 conv, 32 inputs, 64 outputs\n", 347 | " 'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64],stddev=np.sqrt(1/(25*32)))),\n", 348 | " # fully connected, 7*7*64 inputs, 1024 outputs\n", 349 | " 'wd1': tf.Variable(tf.random_normal([7*7*64, 1024],stddev=np.sqrt(1/(7*7*64)))),\n", 350 | " # 1024 inputs, 10 outputs (class prediction)\n", 351 | " 'out': tf.Variable(tf.random_normal([1024, n_classes],stddev=np.sqrt(1/(1024))))\n", 352 | "}\n", 353 | "\n", 354 | "biases2 = {\n", 355 | " 'bc1': tf.Variable(tf.random_normal([32],stddev=0)),\n", 356 | " 'bc2': tf.Variable(tf.random_normal([64],stddev=0)),\n", 357 | " 'bd1': tf.Variable(tf.random_normal([1024],stddev=0)),\n", 358 | " 'out': tf.Variable(tf.random_normal([n_classes],stddev=0))\n", 359 | "}\n" 360 | ] 361 | }, 362 | { 363 | "cell_type": "code", 364 | "execution_count": 11, 365 | "metadata": { 366 | "collapsed": false 367 | }, 368 | "outputs": [ 369 | { 370 | "name": "stdout", 371 | "output_type": "stream", 372 | "text": [ 373 | "WARNING:tensorflow:From :23: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.\n", 374 | "Instructions for updating:\n", 375 | "Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.\n", 376 | "WARNING:tensorflow:From :5: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n", 377 | "Instructions for updating:\n", 378 | "Use tf.where in 2.0, which has the same broadcast rule as np.where\n", 379 | "WARNING:tensorflow:From :12: scalar (from tensorflow.python.framework.tensor_shape) is deprecated and will be removed in a future version.\n", 380 | "Instructions for updating:\n", 381 | "Use tf.TensorShape([]).\n", 382 | "WARNING:tensorflow:From :6: softmax_cross_entropy_with_logits (from tensorflow.python.ops.nn_ops) is deprecated and will be removed in a future version.\n", 383 | "Instructions for updating:\n", 384 | "\n", 385 | "Future major versions of TensorFlow will allow gradients to flow\n", 386 | "into the labels input on backprop by default.\n", 387 | "\n", 388 | "See `tf.nn.softmax_cross_entropy_with_logits_v2`.\n", 389 | "\n" 390 | ] 391 | } 392 | ], 393 | "source": [ 394 | "# Construct model\n", 395 | "pred_ReLU = conv_net_ReLU(x, weights, biases, keep_prob)\n", 396 | "pred_SNN = conv_net_SNN(x, weights2, biases2, dropout_prob,is_training)\n", 397 | "\n", 398 | "# Define loss and optimizer\n", 399 | "cost_ReLU = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred_ReLU, labels=y))\n", 400 | "cost_SNN = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred_SNN, labels=y))\n", 401 | "\n", 402 | "optimizer_ReLU = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost_ReLU)\n", 403 | "optimizer_SNN = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost_SNN)\n", 404 | "\n", 405 | "# Evaluate ReLU model\n", 406 | "correct_pred_ReLU = tf.equal(tf.argmax(pred_ReLU, 1), tf.argmax(y, 1))\n", 407 | "accuracy_ReLU = tf.reduce_mean(tf.cast(correct_pred_ReLU, tf.float32))\n", 408 | "\n", 409 | "# Evaluate SNN model\n", 410 | "correct_pred_SNN = tf.equal(tf.argmax(pred_SNN, 1), tf.argmax(y, 1))\n", 411 | "accuracy_SNN = tf.reduce_mean(tf.cast(correct_pred_SNN, tf.float32))\n", 412 | "\n", 413 | "\n", 414 | "# Initializing the variables\n", 415 | "init = tf.global_variables_initializer()" 416 | ] 417 | }, 418 | { 419 | "cell_type": "code", 420 | "execution_count": 12, 421 | "metadata": { 422 | "collapsed": true 423 | }, 424 | "outputs": [], 425 | "source": [ 426 | "training_loss_protocol_ReLU = []\n", 427 | "training_loss_protocol_SNN = []" 428 | ] 429 | }, 430 | { 431 | "cell_type": "code", 432 | "execution_count": 13, 433 | "metadata": { 434 | "collapsed": false 435 | }, 436 | "outputs": [ 437 | { 438 | "name": "stdout", 439 | "output_type": "stream", 440 | "text": [ 441 | "RELU: Nbr of updates: 1, Minibatch Loss= 2.529600, Training Accuracy= 0.14062\n", 442 | "SNN: Nbr of updates: 1, Minibatch Loss= 2.043923, Training Accuracy= 0.51562\n", 443 | "RELU: Nbr of updates: 2, Minibatch Loss= 2.308014, Training Accuracy= 0.26562\n", 444 | "SNN: Nbr of updates: 2, Minibatch Loss= 1.727961, Training Accuracy= 0.51562\n", 445 | "RELU: Nbr of updates: 3, Minibatch Loss= 2.068156, Training Accuracy= 0.38281\n", 446 | "SNN: Nbr of updates: 3, Minibatch Loss= 1.357726, Training Accuracy= 0.64844\n", 447 | "RELU: Nbr of updates: 4, Minibatch Loss= 1.670266, Training Accuracy= 0.62500\n", 448 | "SNN: Nbr of updates: 4, Minibatch Loss= 0.805312, Training Accuracy= 0.77344\n", 449 | "RELU: Nbr of updates: 5, Minibatch Loss= 1.599183, Training Accuracy= 0.67969\n", 450 | "SNN: Nbr of updates: 5, Minibatch Loss= 0.634932, Training Accuracy= 0.85156\n", 451 | "RELU: Nbr of updates: 6, Minibatch Loss= 1.444664, Training Accuracy= 0.71094\n", 452 | "SNN: Nbr of updates: 6, Minibatch Loss= 0.472679, Training Accuracy= 0.85156\n", 453 | "RELU: Nbr of updates: 7, Minibatch Loss= 1.424880, Training Accuracy= 0.75000\n", 454 | "SNN: Nbr of updates: 7, Minibatch Loss= 0.443895, Training Accuracy= 0.86719\n", 455 | "RELU: Nbr of updates: 8, Minibatch Loss= 1.238384, Training Accuracy= 0.70312\n", 456 | "SNN: Nbr of updates: 8, Minibatch Loss= 0.433366, Training Accuracy= 0.87500\n", 457 | "RELU: Nbr of updates: 9, Minibatch Loss= 1.223281, Training Accuracy= 0.74219\n", 458 | "SNN: Nbr of updates: 9, Minibatch Loss= 0.446125, Training Accuracy= 0.91406\n", 459 | "RELU: Nbr of updates: 10, Minibatch Loss= 1.126184, Training Accuracy= 0.82812\n", 460 | "SNN: Nbr of updates: 10, Minibatch Loss= 0.353353, Training Accuracy= 0.92188\n", 461 | "RELU: Nbr of updates: 11, Minibatch Loss= 0.945392, Training Accuracy= 0.83594\n", 462 | "SNN: Nbr of updates: 11, Minibatch Loss= 0.323721, Training Accuracy= 0.91406\n", 463 | "RELU: Nbr of updates: 12, Minibatch Loss= 1.061059, Training Accuracy= 0.72656\n", 464 | "SNN: Nbr of updates: 12, Minibatch Loss= 0.319183, Training Accuracy= 0.91406\n", 465 | "RELU: Nbr of updates: 13, Minibatch Loss= 1.053779, Training Accuracy= 0.71094\n", 466 | "SNN: Nbr of updates: 13, Minibatch Loss= 0.341918, Training Accuracy= 0.90625\n", 467 | "RELU: Nbr of updates: 14, Minibatch Loss= 0.861647, Training Accuracy= 0.82031\n", 468 | "SNN: Nbr of updates: 14, Minibatch Loss= 0.247850, Training Accuracy= 0.94531\n", 469 | "RELU: Nbr of updates: 15, Minibatch Loss= 0.859886, Training Accuracy= 0.85938\n", 470 | "SNN: Nbr of updates: 15, Minibatch Loss= 0.300610, Training Accuracy= 0.91406\n", 471 | "RELU: Nbr of updates: 16, Minibatch Loss= 0.803375, Training Accuracy= 0.82031\n", 472 | "SNN: Nbr of updates: 16, Minibatch Loss= 0.233539, Training Accuracy= 0.96094\n", 473 | "RELU: Nbr of updates: 17, Minibatch Loss= 0.770378, Training Accuracy= 0.85156\n", 474 | "SNN: Nbr of updates: 17, Minibatch Loss= 0.292139, Training Accuracy= 0.90625\n", 475 | "RELU: Nbr of updates: 18, Minibatch Loss= 0.703464, Training Accuracy= 0.81250\n", 476 | "SNN: Nbr of updates: 18, Minibatch Loss= 0.214945, Training Accuracy= 0.96094\n", 477 | "RELU: Nbr of updates: 19, Minibatch Loss= 0.667796, Training Accuracy= 0.90625\n", 478 | "SNN: Nbr of updates: 19, Minibatch Loss= 0.267386, Training Accuracy= 0.92969\n", 479 | "RELU: Nbr of updates: 20, Minibatch Loss= 0.739539, Training Accuracy= 0.82812\n", 480 | "SNN: Nbr of updates: 20, Minibatch Loss= 0.301252, Training Accuracy= 0.92188\n", 481 | "RELU: Nbr of updates: 21, Minibatch Loss= 0.712775, Training Accuracy= 0.82031\n", 482 | "SNN: Nbr of updates: 21, Minibatch Loss= 0.229995, Training Accuracy= 0.94531\n", 483 | "RELU: Nbr of updates: 22, Minibatch Loss= 0.668107, Training Accuracy= 0.83594\n", 484 | "SNN: Nbr of updates: 22, Minibatch Loss= 0.302777, Training Accuracy= 0.92188\n", 485 | "RELU: Nbr of updates: 23, Minibatch Loss= 0.662650, Training Accuracy= 0.80469\n", 486 | "SNN: Nbr of updates: 23, Minibatch Loss= 0.277247, Training Accuracy= 0.91406\n", 487 | "RELU: Nbr of updates: 24, Minibatch Loss= 0.751235, Training Accuracy= 0.82031\n", 488 | "SNN: Nbr of updates: 24, Minibatch Loss= 0.312389, Training Accuracy= 0.89844\n", 489 | "RELU: Nbr of updates: 25, Minibatch Loss= 0.610983, Training Accuracy= 0.85938\n", 490 | "SNN: Nbr of updates: 25, Minibatch Loss= 0.230509, Training Accuracy= 0.93750\n", 491 | "RELU: Nbr of updates: 26, Minibatch Loss= 0.598711, Training Accuracy= 0.85156\n", 492 | "SNN: Nbr of updates: 26, Minibatch Loss= 0.257897, Training Accuracy= 0.92969\n", 493 | "RELU: Nbr of updates: 27, Minibatch Loss= 0.649527, Training Accuracy= 0.82031\n", 494 | "SNN: Nbr of updates: 27, Minibatch Loss= 0.225033, Training Accuracy= 0.92969\n", 495 | "RELU: Nbr of updates: 28, Minibatch Loss= 0.513970, Training Accuracy= 0.92188\n", 496 | "SNN: Nbr of updates: 28, Minibatch Loss= 0.170374, Training Accuracy= 0.94531\n", 497 | "RELU: Nbr of updates: 29, Minibatch Loss= 0.548257, Training Accuracy= 0.83594\n", 498 | "SNN: Nbr of updates: 29, Minibatch Loss= 0.211176, Training Accuracy= 0.92188\n", 499 | "RELU: Nbr of updates: 30, Minibatch Loss= 0.541867, Training Accuracy= 0.88281\n", 500 | "SNN: Nbr of updates: 30, Minibatch Loss= 0.221660, Training Accuracy= 0.96094\n", 501 | "RELU: Nbr of updates: 31, Minibatch Loss= 0.485644, Training Accuracy= 0.89844\n", 502 | "SNN: Nbr of updates: 31, Minibatch Loss= 0.212724, Training Accuracy= 0.92188\n", 503 | "RELU: Nbr of updates: 32, Minibatch Loss= 0.450281, Training Accuracy= 0.85938\n", 504 | "SNN: Nbr of updates: 32, Minibatch Loss= 0.169182, Training Accuracy= 0.96094\n", 505 | "RELU: Nbr of updates: 33, Minibatch Loss= 0.633430, Training Accuracy= 0.75000\n", 506 | "SNN: Nbr of updates: 33, Minibatch Loss= 0.212616, Training Accuracy= 0.96875\n", 507 | "RELU: Nbr of updates: 34, Minibatch Loss= 0.477125, Training Accuracy= 0.89062\n", 508 | "SNN: Nbr of updates: 34, Minibatch Loss= 0.202256, Training Accuracy= 0.96094\n", 509 | "RELU: Nbr of updates: 35, Minibatch Loss= 0.549391, Training Accuracy= 0.82812\n", 510 | "SNN: Nbr of updates: 35, Minibatch Loss= 0.310254, Training Accuracy= 0.89844\n", 511 | "RELU: Nbr of updates: 36, Minibatch Loss= 0.368259, Training Accuracy= 0.90625\n", 512 | "SNN: Nbr of updates: 36, Minibatch Loss= 0.153444, Training Accuracy= 0.93750\n", 513 | "RELU: Nbr of updates: 37, Minibatch Loss= 0.492477, Training Accuracy= 0.86719\n", 514 | "SNN: Nbr of updates: 37, Minibatch Loss= 0.211344, Training Accuracy= 0.93750\n", 515 | "RELU: Nbr of updates: 38, Minibatch Loss= 0.556713, Training Accuracy= 0.87500\n", 516 | "SNN: Nbr of updates: 38, Minibatch Loss= 0.203756, Training Accuracy= 0.95312\n", 517 | "RELU: Nbr of updates: 39, Minibatch Loss= 0.442008, Training Accuracy= 0.88281\n", 518 | "SNN: Nbr of updates: 39, Minibatch Loss= 0.194691, Training Accuracy= 0.94531\n", 519 | "RELU: Nbr of updates: 40, Minibatch Loss= 0.427592, Training Accuracy= 0.90625\n", 520 | "SNN: Nbr of updates: 40, Minibatch Loss= 0.226508, Training Accuracy= 0.91406\n", 521 | "RELU: Nbr of updates: 41, Minibatch Loss= 0.598306, Training Accuracy= 0.82812\n", 522 | "SNN: Nbr of updates: 41, Minibatch Loss= 0.207904, Training Accuracy= 0.93750\n", 523 | "RELU: Nbr of updates: 42, Minibatch Loss= 0.551276, Training Accuracy= 0.85938\n", 524 | "SNN: Nbr of updates: 42, Minibatch Loss= 0.198079, Training Accuracy= 0.94531\n", 525 | "RELU: Nbr of updates: 43, Minibatch Loss= 0.404157, Training Accuracy= 0.89844\n", 526 | "SNN: Nbr of updates: 43, Minibatch Loss= 0.127414, Training Accuracy= 0.98438\n", 527 | "RELU: Nbr of updates: 44, Minibatch Loss= 0.368151, Training Accuracy= 0.93750\n", 528 | "SNN: Nbr of updates: 44, Minibatch Loss= 0.153246, Training Accuracy= 0.96875\n", 529 | "RELU: Nbr of updates: 45, Minibatch Loss= 0.461047, Training Accuracy= 0.88281\n", 530 | "SNN: Nbr of updates: 45, Minibatch Loss= 0.242575, Training Accuracy= 0.92969\n", 531 | "RELU: Nbr of updates: 46, Minibatch Loss= 0.456105, Training Accuracy= 0.89844\n", 532 | "SNN: Nbr of updates: 46, Minibatch Loss= 0.231932, Training Accuracy= 0.93750\n", 533 | "RELU: Nbr of updates: 47, Minibatch Loss= 0.414762, Training Accuracy= 0.89062\n", 534 | "SNN: Nbr of updates: 47, Minibatch Loss= 0.234552, Training Accuracy= 0.95312\n", 535 | "RELU: Nbr of updates: 48, Minibatch Loss= 0.371672, Training Accuracy= 0.90625\n", 536 | "SNN: Nbr of updates: 48, Minibatch Loss= 0.172536, Training Accuracy= 0.96094\n", 537 | "RELU: Nbr of updates: 49, Minibatch Loss= 0.396432, Training Accuracy= 0.91406\n", 538 | "SNN: Nbr of updates: 49, Minibatch Loss= 0.162846, Training Accuracy= 0.95312\n", 539 | "RELU: Nbr of updates: 50, Minibatch Loss= 0.281247, Training Accuracy= 0.96875\n", 540 | "SNN: Nbr of updates: 50, Minibatch Loss= 0.125570, Training Accuracy= 0.97656\n", 541 | "Optimization Finished!\n", 542 | "\n", 543 | "ReLU: Testing Accuracy: 0.8339844\n", 544 | "SNN: Testing Accuracy: 0.9140625\n" 545 | ] 546 | } 547 | ], 548 | "source": [ 549 | "# Launch the graph\n", 550 | "gpu_options = tf.GPUOptions(allow_growth=True)\n", 551 | "with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:\n", 552 | " sess.run(init)\n", 553 | " step = 0\n", 554 | " # Keep training until reach max iterations\n", 555 | " while step < training_iters:\n", 556 | " batch_x, batch_y = mnist.train.next_batch(batch_size)\n", 557 | " batch_x_norm = scaler.transform(batch_x)\n", 558 | " # Run optimization op (backprop)\n", 559 | " sess.run(optimizer_ReLU, feed_dict={x: batch_x, y: batch_y,\n", 560 | " keep_prob: keep_prob_ReLU})\n", 561 | " sess.run(optimizer_SNN, feed_dict={x: batch_x_norm, y: batch_y,\n", 562 | " dropout_prob: dropout_prob_SNN,is_training:True})\n", 563 | " \n", 564 | " \n", 565 | " if step % display_step == 0:\n", 566 | " #batch_x, batch_y = mnist.test.next_batch(batch_size)\n", 567 | " #batch_x_norm = scaler.transform(batch_x)\n", 568 | " # Calculate batch loss and accuracy\n", 569 | " loss_ReLU, acc_ReLU = sess.run([cost_ReLU, accuracy_ReLU], feed_dict={x: batch_x,\n", 570 | " y: batch_y,\n", 571 | " keep_prob: 1.0})\n", 572 | " training_loss_protocol_ReLU.append(loss_ReLU)\n", 573 | " \n", 574 | " loss_SNN, acc_SNN = sess.run([cost_SNN, accuracy_SNN], feed_dict={x: batch_x_norm,\n", 575 | " y: batch_y,\n", 576 | " dropout_prob: 0.0, is_training:False})\n", 577 | " training_loss_protocol_SNN.append(loss_SNN)\n", 578 | " \n", 579 | " print( \"RELU: Nbr of updates: \" + str(step+1) + \", Minibatch Loss= \" + \\\n", 580 | " \"{:.6f}\".format(loss_ReLU) + \", Training Accuracy= \" + \\\n", 581 | " \"{:.5f}\".format(acc_ReLU))\n", 582 | " \n", 583 | " print( \"SNN: Nbr of updates: \" + str(step+1) + \", Minibatch Loss= \" + \\\n", 584 | " \"{:.6f}\".format(loss_SNN) + \", Training Accuracy= \" + \\\n", 585 | " \"{:.5f}\".format(acc_SNN))\n", 586 | " step += 1\n", 587 | " print(\"Optimization Finished!\\n\")\n", 588 | "\n", 589 | " # Calculate accuracy for 256 mnist test images\n", 590 | " print(\"ReLU: Testing Accuracy:\", sess.run(accuracy_ReLU, feed_dict={x: mnist.test.images[:512],\n", 591 | " y: mnist.test.labels[:512],\n", 592 | " keep_prob: 1.0}))\n", 593 | " print(\"SNN: Testing Accuracy:\", sess.run(accuracy_SNN, feed_dict={x: scaler.transform(mnist.test.images[:512]),\n", 594 | " y: mnist.test.labels[:512],\n", 595 | " dropout_prob: 0.0, is_training:False}))" 596 | ] 597 | }, 598 | { 599 | "cell_type": "code", 600 | "execution_count": 14, 601 | "metadata": { 602 | "collapsed": false, 603 | "scrolled": true, 604 | "pycharm": { 605 | "name": "#%%\n" 606 | } 607 | }, 608 | "outputs": [ 609 | { 610 | "data": { 611 | "text/plain": "
", 612 | "image/png": "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\n" 613 | }, 614 | "metadata": { 615 | "needs_background": "light" 616 | }, 617 | "output_type": "display_data" 618 | } 619 | ], 620 | "source": [ 621 | "import matplotlib.pyplot as plt\n", 622 | "\n", 623 | "fig, ax = plt.subplots()\n", 624 | "ax.plot( training_loss_protocol_ReLU, label='Loss ReLU-CNN')\n", 625 | "ax.plot( training_loss_protocol_SNN, label='Loss SNN')\n", 626 | "ax.set_yscale('log') # log scale\n", 627 | "ax.set_xlabel('iterations/updates')\n", 628 | "ax.set_ylabel('training loss')\n", 629 | "fig.tight_layout()\n", 630 | "ax.legend()\n", 631 | "plt.show()" 632 | ] 633 | } 634 | ], 635 | "metadata": { 636 | "kernelspec": { 637 | "name": "python3", 638 | "language": "python", 639 | "display_name": "Python 3" 640 | }, 641 | "language_info": { 642 | "codemirror_mode": { 643 | "name": "ipython", 644 | "version": 3 645 | }, 646 | "file_extension": ".py", 647 | "mimetype": "text/x-python", 648 | "name": "python", 649 | "nbconvert_exporter": "python", 650 | "pygments_lexer": "ipython3", 651 | "version": "3.5.2" 652 | } 653 | }, 654 | "nbformat": 4, 655 | "nbformat_minor": 0 656 | } -------------------------------------------------------------------------------- /TF_1_x/SelfNormalizingNetworks_MLP_MNIST.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": {}, 6 | "source": [ 7 | "# Tutorial on self-normalizing networks on the MNIST data set: multi-layer perceptrons\n", 8 | "\n", 9 | "*Author:* Guenter Klambauer, 2017\n", 10 | "\n", 11 | "Derived from: [Aymeric Damien](https://github.com/aymericdamien/TensorFlow-Examples/) \n" 12 | ] 13 | }, 14 | { 15 | "cell_type": "code", 16 | "execution_count": 1, 17 | "metadata": { 18 | "collapsed": false 19 | }, 20 | "outputs": [ 21 | { 22 | "name": "stdout", 23 | "output_type": "stream", 24 | "text": [ 25 | "WARNING:tensorflow:From :16: read_data_sets (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 26 | "Instructions for updating:\n", 27 | "Please use alternatives such as official/mnist/dataset.py from tensorflow/models.\n", 28 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:260: maybe_download (from tensorflow.contrib.learn.python.learn.datasets.base) is deprecated and will be removed in a future version.\n", 29 | "Instructions for updating:\n", 30 | "Please write your own downloading logic.\n", 31 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:262: extract_images (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 32 | "Instructions for updating:\n", 33 | "Please use tf.data to implement this functionality.\n", 34 | "Extracting /tmp/data/train-images-idx3-ubyte.gz\n", 35 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:267: extract_labels (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 36 | "Instructions for updating:\n", 37 | "Please use tf.data to implement this functionality.\n", 38 | "Extracting /tmp/data/train-labels-idx1-ubyte.gz\n", 39 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:110: dense_to_one_hot (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 40 | "Instructions for updating:\n", 41 | "Please use tf.one_hot on tensors.\n", 42 | "Extracting /tmp/data/t10k-images-idx3-ubyte.gz\n", 43 | "Extracting /tmp/data/t10k-labels-idx1-ubyte.gz\n", 44 | "WARNING:tensorflow:From /home/kai/miniconda3/envs/snn_tf1_compat/lib/python3.7/site-packages/tensorflow_core/contrib/learn/python/learn/datasets/mnist.py:290: DataSet.__init__ (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.\n", 45 | "Instructions for updating:\n", 46 | "Please use alternatives such as official/mnist/dataset.py from tensorflow/models.\n", 47 | "1.15.0\n" 48 | ] 49 | } 50 | ], 51 | "source": [ 52 | "import tensorflow as tf\n", 53 | "import numpy as np\n", 54 | "from sklearn.preprocessing import StandardScaler\n", 55 | "\n", 56 | "import numbers\n", 57 | "from tensorflow.python.framework import ops\n", 58 | "from tensorflow.python.framework import tensor_shape\n", 59 | "from tensorflow.python.framework import tensor_util\n", 60 | "from tensorflow.python.ops import math_ops\n", 61 | "from tensorflow.python.ops import random_ops\n", 62 | "from tensorflow.python.ops import array_ops\n", 63 | "from tensorflow.python.layers import utils\n", 64 | "\n", 65 | "# Import MINST data\n", 66 | "from tensorflow.examples.tutorials.mnist import input_data\n", 67 | "mnist = input_data.read_data_sets(\"/tmp/data/\", one_hot=True)\n", 68 | "\n", 69 | "print(tf.__version__)" 70 | ] 71 | }, 72 | { 73 | "cell_type": "code", 74 | "execution_count": 2, 75 | "metadata": { 76 | "collapsed": true 77 | }, 78 | "outputs": [], 79 | "source": [ 80 | "# Parameters\n", 81 | "learning_rate = 0.05\n", 82 | "training_epochs = 15\n", 83 | "batch_size = 100\n", 84 | "display_step = 1\n", 85 | "\n", 86 | "# Network Parameters\n", 87 | "n_hidden_1 = 784 # 1st layer number of features\n", 88 | "n_hidden_2 = 784 # 2nd layer number of features\n", 89 | "n_input = 784 # MNIST data input (img shape: 28*28)\n", 90 | "n_classes = 10 # MNIST total classes (0-9 digits)\n", 91 | "\n", 92 | "# tf Graph input\n", 93 | "x = tf.placeholder(\"float\", [None, n_input])\n", 94 | "y = tf.placeholder(\"float\", [None, n_classes])\n", 95 | "dropoutRate = tf.placeholder(tf.float32)\n", 96 | "is_training= tf.placeholder(tf.bool)" 97 | ] 98 | }, 99 | { 100 | "cell_type": "markdown", 101 | "metadata": {}, 102 | "source": [ 103 | "### (1) Definition of scaled exponential linear units (SELUs)" 104 | ] 105 | }, 106 | { 107 | "cell_type": "code", 108 | "execution_count": 3, 109 | "metadata": { 110 | "collapsed": true 111 | }, 112 | "outputs": [], 113 | "source": [ 114 | "def selu(x):\n", 115 | " with ops.name_scope('elu') as scope:\n", 116 | " alpha = 1.6732632423543772848170429916717\n", 117 | " scale = 1.0507009873554804934193349852946\n", 118 | " return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x))" 119 | ] 120 | }, 121 | { 122 | "cell_type": "markdown", 123 | "metadata": {}, 124 | "source": [ 125 | "### (2) Definition of dropout variant for SNNs" 126 | ] 127 | }, 128 | { 129 | "cell_type": "code", 130 | "execution_count": 4, 131 | "metadata": { 132 | "collapsed": true 133 | }, 134 | "outputs": [], 135 | "source": [ 136 | "def dropout_selu(x, rate, alpha= -1.7580993408473766, fixedPointMean=0.0, fixedPointVar=1.0, \n", 137 | " noise_shape=None, seed=None, name=None, training=False):\n", 138 | " \"\"\"Dropout to a value with rescaling.\"\"\"\n", 139 | "\n", 140 | " def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name):\n", 141 | " keep_prob = 1.0 - rate\n", 142 | " x = ops.convert_to_tensor(x, name=\"x\")\n", 143 | " if isinstance(keep_prob, numbers.Real) and not 0 < keep_prob <= 1:\n", 144 | " raise ValueError(\"keep_prob must be a scalar tensor or a float in the \"\n", 145 | " \"range (0, 1], got %g\" % keep_prob)\n", 146 | " keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name=\"keep_prob\")\n", 147 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 148 | "\n", 149 | " alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name=\"alpha\")\n", 150 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 151 | "\n", 152 | " if tensor_util.constant_value(keep_prob) == 1:\n", 153 | " return x\n", 154 | "\n", 155 | " noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x)\n", 156 | " random_tensor = keep_prob\n", 157 | " random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype)\n", 158 | " binary_tensor = math_ops.floor(random_tensor)\n", 159 | " ret = x * binary_tensor + alpha * (1-binary_tensor)\n", 160 | "\n", 161 | " a = tf.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * tf.pow(alpha-fixedPointMean,2) + fixedPointVar)))\n", 162 | "\n", 163 | " b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha)\n", 164 | " ret = a * ret + b\n", 165 | " ret.set_shape(x.get_shape())\n", 166 | " return ret\n", 167 | "\n", 168 | " with ops.name_scope(name, \"dropout\", [x]) as name:\n", 169 | " return utils.smart_cond(training,\n", 170 | " lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name),\n", 171 | " lambda: array_ops.identity(x))\n" 172 | ] 173 | }, 174 | { 175 | "cell_type": "markdown", 176 | "metadata": {}, 177 | "source": [ 178 | "### (3) Input data scaled to zero mean and unit variance" 179 | ] 180 | }, 181 | { 182 | "cell_type": "code", 183 | "execution_count": 5, 184 | "metadata": { 185 | "collapsed": true 186 | }, 187 | "outputs": [], 188 | "source": [ 189 | "# (1) Scale input to zero mean and unit variance\n", 190 | "scaler = StandardScaler().fit(mnist.train.images)" 191 | ] 192 | }, 193 | { 194 | "cell_type": "code", 195 | "execution_count": 6, 196 | "metadata": { 197 | "collapsed": true 198 | }, 199 | "outputs": [], 200 | "source": [ 201 | "# Tensorboard\n", 202 | "logs_path = '~/tmp'" 203 | ] 204 | }, 205 | { 206 | "cell_type": "code", 207 | "execution_count": 7, 208 | "metadata": { 209 | "collapsed": true 210 | }, 211 | "outputs": [], 212 | "source": [ 213 | "# Create model\n", 214 | "def multilayer_perceptron(x, weights, biases, rate, is_training):\n", 215 | " # Hidden layer with SELU activation\n", 216 | " layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])\n", 217 | " #netI_1 = layer_1\n", 218 | " layer_1 = selu(layer_1)\n", 219 | " layer_1 = dropout_selu(layer_1,rate, training=is_training)\n", 220 | " \n", 221 | " # Hidden layer with SELU activation\n", 222 | " layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])\n", 223 | " #netI_2 = layer_2\n", 224 | " layer_2 = selu(layer_2)\n", 225 | " layer_2 = dropout_selu(layer_2,rate, training=is_training)\n", 226 | "\n", 227 | " # Output layer with linear activation\n", 228 | " out_layer = tf.matmul(layer_2, weights['out']) + biases['out']\n", 229 | " return out_layer" 230 | ] 231 | }, 232 | { 233 | "cell_type": "markdown", 234 | "metadata": {}, 235 | "source": [ 236 | "### (4) Initialization with STDDEV of sqrt(1/n)" 237 | ] 238 | }, 239 | { 240 | "cell_type": "code", 241 | "execution_count": 8, 242 | "metadata": { 243 | "collapsed": false 244 | }, 245 | "outputs": [ 246 | { 247 | "name": "stdout", 248 | "output_type": "stream", 249 | "text": [ 250 | "WARNING:tensorflow:From :5: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n", 251 | "Instructions for updating:\n", 252 | "Use tf.where in 2.0, which has the same broadcast rule as np.where\n", 253 | "WARNING:tensorflow:From :12: scalar (from tensorflow.python.framework.tensor_shape) is deprecated and will be removed in a future version.\n", 254 | "Instructions for updating:\n", 255 | "Use tf.TensorShape([]).\n", 256 | "WARNING:tensorflow:From :16: softmax_cross_entropy_with_logits (from tensorflow.python.ops.nn_ops) is deprecated and will be removed in a future version.\n", 257 | "Instructions for updating:\n", 258 | "\n", 259 | "Future major versions of TensorFlow will allow gradients to flow\n", 260 | "into the labels input on backprop by default.\n", 261 | "\n", 262 | "See `tf.nn.softmax_cross_entropy_with_logits_v2`.\n", 263 | "\n" 264 | ] 265 | } 266 | ], 267 | "source": [ 268 | "# Store layers weight & bias\n", 269 | "weights = {\n", 270 | " 'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1],stddev=np.sqrt(1/n_input))),\n", 271 | " 'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2],stddev=np.sqrt(1/n_hidden_1))),\n", 272 | " 'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes],stddev=np.sqrt(1/n_hidden_2)))\n", 273 | "}\n", 274 | "biases = {\n", 275 | " 'b1': tf.Variable(tf.random_normal([n_hidden_1],stddev=0)),\n", 276 | " 'b2': tf.Variable(tf.random_normal([n_hidden_2],stddev=0)),\n", 277 | " 'out': tf.Variable(tf.random_normal([n_classes],stddev=0))\n", 278 | "}\n", 279 | "# Construct model\n", 280 | "pred = multilayer_perceptron(x, weights, biases, rate=dropoutRate, is_training=is_training)\n", 281 | "\n", 282 | "# Define loss and optimizer\n", 283 | "cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))\n", 284 | "optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)\n", 285 | "\n", 286 | " # Test model\n", 287 | "correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))\n", 288 | "# Calculate accuracy\n", 289 | "accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))\n", 290 | " \n", 291 | "# Initializing the variables\n", 292 | "init = tf.global_variables_initializer()" 293 | ] 294 | }, 295 | { 296 | "cell_type": "code", 297 | "execution_count": 9, 298 | "metadata": { 299 | "collapsed": false 300 | }, 301 | "outputs": [], 302 | "source": [ 303 | "# Create a histogramm for weights\n", 304 | "tf.summary.histogram(\"weights2\", weights['h2'])\n", 305 | "tf.summary.histogram(\"weights1\", weights['h1'])\n", 306 | "\n", 307 | "# Create a summary to monitor cost tensor\n", 308 | "tf.summary.scalar(\"loss\", cost)\n", 309 | "# Create a summary to monitor accuracy tensor\n", 310 | "tf.summary.scalar(\"accuracy\", accuracy)\n", 311 | "# Merge all summaries into a single op\n", 312 | "merged_summary_op = tf.summary.merge_all()" 313 | ] 314 | }, 315 | { 316 | "cell_type": "code", 317 | "execution_count": 10, 318 | "metadata": { 319 | "collapsed": false, 320 | "pycharm": { 321 | "name": "#%%\n" 322 | } 323 | }, 324 | "outputs": [ 325 | { 326 | "name": "stdout", 327 | "output_type": "stream", 328 | "text": [ 329 | "Epoch: 0001 cost= 0.378099435\n", 330 | "Train-Accuracy: 0.94 Train-Loss: 0.17783636\n", 331 | "Validation-Accuracy: 0.9375 Val-Loss: 0.19006419 \n", 332 | "\n", 333 | "Epoch: 0002 cost= 0.260677475\n", 334 | "Train-Accuracy: 0.98 Train-Loss: 0.17112085\n", 335 | "Validation-Accuracy: 0.953125 Val-Loss: 0.18273504 \n", 336 | "\n", 337 | "Epoch: 0003 cost= 0.209526656\n", 338 | "Train-Accuracy: 0.99 Train-Loss: 0.038006987\n", 339 | "Validation-Accuracy: 0.9433594 Val-Loss: 0.18926144 \n", 340 | "\n", 341 | "Epoch: 0004 cost= 0.172529011\n", 342 | "Train-Accuracy: 0.99 Train-Loss: 0.04645816\n", 343 | "Validation-Accuracy: 0.9550781 Val-Loss: 0.18212736 \n", 344 | "\n", 345 | "Epoch: 0005 cost= 0.143096214\n", 346 | "Train-Accuracy: 1.0 Train-Loss: 0.023179742\n", 347 | "Validation-Accuracy: 0.98046875 Val-Loss: 0.08168894 \n", 348 | "\n", 349 | "Epoch: 0006 cost= 0.120568850\n", 350 | "Train-Accuracy: 0.99 Train-Loss: 0.042775374\n", 351 | "Validation-Accuracy: 0.96484375 Val-Loss: 0.15286936 \n", 352 | "\n", 353 | "Epoch: 0007 cost= 0.109314255\n", 354 | "Train-Accuracy: 0.99 Train-Loss: 0.023606239\n", 355 | "Validation-Accuracy: 0.96875 Val-Loss: 0.13182782 \n", 356 | "\n", 357 | "Epoch: 0008 cost= 0.095551941\n", 358 | "Train-Accuracy: 1.0 Train-Loss: 0.016547382\n", 359 | "Validation-Accuracy: 0.9511719 Val-Loss: 0.15446913 \n", 360 | "\n", 361 | "Epoch: 0009 cost= 0.086288710\n", 362 | "Train-Accuracy: 1.0 Train-Loss: 0.007214434\n", 363 | "Validation-Accuracy: 0.9746094 Val-Loss: 0.19779019 \n", 364 | "\n", 365 | "Epoch: 0010 cost= 0.076114546\n", 366 | "Train-Accuracy: 1.0 Train-Loss: 0.006380536\n", 367 | "Validation-Accuracy: 0.96484375 Val-Loss: 0.123645395 \n", 368 | "\n", 369 | "Epoch: 0011 cost= 0.069839616\n", 370 | "Train-Accuracy: 1.0 Train-Loss: 0.016955461\n", 371 | "Validation-Accuracy: 0.9824219 Val-Loss: 0.07528831 \n", 372 | "\n", 373 | "Epoch: 0012 cost= 0.063571214\n", 374 | "Train-Accuracy: 1.0 Train-Loss: 0.0137194265\n", 375 | "Validation-Accuracy: 0.9765625 Val-Loss: 0.119060434 \n", 376 | "\n", 377 | "Epoch: 0013 cost= 0.058336736\n", 378 | "Train-Accuracy: 0.99 Train-Loss: 0.014294322\n", 379 | "Validation-Accuracy: 0.97265625 Val-Loss: 0.12580527 \n", 380 | "\n", 381 | "Epoch: 0014 cost= 0.053450938\n", 382 | "Train-Accuracy: 1.0 Train-Loss: 0.019194987\n", 383 | "Validation-Accuracy: 0.97265625 Val-Loss: 0.06765857 \n", 384 | "\n", 385 | "Epoch: 0015 cost= 0.048051286\n", 386 | "Train-Accuracy: 1.0 Train-Loss: 0.008367514\n", 387 | "Validation-Accuracy: 0.9589844 Val-Loss: 0.17628105 \n", 388 | "\n" 389 | ] 390 | } 391 | ], 392 | "source": [ 393 | "# Launch the graph\n", 394 | "gpu_options = tf.GPUOptions(allow_growth=True)\n", 395 | "with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:\n", 396 | " sess.run(init)\n", 397 | "\n", 398 | " summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())\n", 399 | "\n", 400 | " # Training cycle\n", 401 | " for epoch in range(training_epochs):\n", 402 | " avg_cost = 0.\n", 403 | " total_batch = int(mnist.train.num_examples/batch_size)\n", 404 | " # Loop over all batches\n", 405 | " for i in range(total_batch):\n", 406 | " batch_x, batch_y = mnist.train.next_batch(batch_size)\n", 407 | " batch_x = scaler.transform(batch_x)\n", 408 | " # Run optimization op (backprop) and cost op (to get loss value)\n", 409 | " _, c = sess.run([optimizer, cost], feed_dict={x: batch_x,\n", 410 | " y: batch_y, dropoutRate: 0.05, is_training:True})\n", 411 | "\n", 412 | " # Compute average loss\n", 413 | " avg_cost += c / total_batch\n", 414 | " # Display logs per epoch step\n", 415 | " if epoch % display_step == 0:\n", 416 | " print (\"Epoch:\", '%04d' % (epoch+1), \"cost=\",\"{:.9f}\".format(avg_cost))\n", 417 | " \n", 418 | " accTrain, costTrain = sess.run([accuracy, cost],\n", 419 | " feed_dict={x: batch_x, y: batch_y, \n", 420 | " dropoutRate: 0.0, is_training:False})\n", 421 | " \n", 422 | " print(\"Train-Accuracy:\", accTrain,\"Train-Loss:\", costTrain)\n", 423 | "\n", 424 | " batch_x_test, batch_y_test = mnist.test.next_batch(512)\n", 425 | " batch_x_test = scaler.transform(batch_x_test)\n", 426 | "\n", 427 | " accTest, costVal = sess.run([accuracy, cost], feed_dict={x: batch_x_test, y: batch_y_test, \n", 428 | " dropoutRate: 0.0, is_training:False})\n", 429 | "\n", 430 | " print(\"Validation-Accuracy:\", accTest,\"Val-Loss:\", costVal,\"\\n\")" 431 | ] 432 | } 433 | ], 434 | "metadata": { 435 | "kernelspec": { 436 | "name": "python3", 437 | "language": "python", 438 | "display_name": "Python 3" 439 | }, 440 | "language_info": { 441 | "codemirror_mode": { 442 | "name": "ipython", 443 | "version": 3 444 | }, 445 | "file_extension": ".py", 446 | "mimetype": "text/x-python", 447 | "name": "python", 448 | "nbconvert_exporter": "python", 449 | "pygments_lexer": "ipython3", 450 | "version": "3.5.2" 451 | } 452 | }, 453 | "nbformat": 4, 454 | "nbformat_minor": 0 455 | } -------------------------------------------------------------------------------- /TF_1_x/environment.yml: -------------------------------------------------------------------------------- 1 | name: snn_tf1_compat 2 | channels: 3 | - conda-forge 4 | - defaults 5 | dependencies: 6 | - python=3.7 7 | - pip=21.0.1 8 | - numpy=1.20.0 9 | - scipy=1.6.0 10 | - sympy=1.7.1 11 | - pandas=1.2.2 12 | - scikit-learn=0.24.1 13 | - matplotlib=3.3.4 14 | - jupyter=1.0.0 15 | # CPU mode only, comment if using gpu mode 16 | # - tensorflow=1.15.0 17 | # GPU mode, take take that your gpu might need a different cudatoolkit! 18 | - cudatoolkit=10.0 19 | - tensorflow-gpu=1.15.0 -------------------------------------------------------------------------------- /TF_1_x/getSELUparameters.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": {}, 6 | "source": [ 7 | "# Obtain the SELU parameters for arbitrary fixed points\n", 8 | "\n", 9 | "*Author:* Guenter Klambauer, 2017\n" 10 | ] 11 | }, 12 | { 13 | "cell_type": "code", 14 | "execution_count": 1, 15 | "metadata": {}, 16 | "outputs": [], 17 | "source": [ 18 | "import numpy as np\n", 19 | "from scipy.special import erf,erfc\n", 20 | "from sympy import Symbol, solve, nsolve" 21 | ] 22 | }, 23 | { 24 | "cell_type": "markdown", 25 | "metadata": {}, 26 | "source": [ 27 | "### Function to obtain the parameters for the SELU with arbitrary fixed point (mean variance)" 28 | ] 29 | }, 30 | { 31 | "cell_type": "code", 32 | "execution_count": 2, 33 | "metadata": {}, 34 | "outputs": [], 35 | "source": [ 36 | "def getSeluParameters(fixedpointMean=0,fixedpointVar=1):\n", 37 | " \"\"\" Finding the parameters of the SELU activation function. The function returns alpha and lambda for the desired fixed point. \"\"\"\n", 38 | " \n", 39 | " import sympy\n", 40 | " from sympy import Symbol, solve, nsolve\n", 41 | "\n", 42 | " aa = Symbol('aa')\n", 43 | " ll = Symbol('ll')\n", 44 | " nu = fixedpointMean \n", 45 | " tau = fixedpointVar \n", 46 | "\n", 47 | " mean = 0.5*ll*(nu + np.exp(-nu**2/(2*tau))*np.sqrt(2/np.pi)*np.sqrt(tau) + \\\n", 48 | " nu*erf(nu/(np.sqrt(2*tau))) - aa*erfc(nu/(np.sqrt(2*tau))) + \\\n", 49 | " np.exp(nu+tau/2)*aa*erfc((nu+tau)/(np.sqrt(2*tau))))\n", 50 | "\n", 51 | " var = 0.5*ll**2*(np.exp(-nu**2/(2*tau))*np.sqrt(2/np.pi*tau)*nu + (nu**2+tau)* \\\n", 52 | " (1+erf(nu/(np.sqrt(2*tau)))) + aa**2 *erfc(nu/(np.sqrt(2*tau))) \\\n", 53 | " - aa**2 * 2 *np.exp(nu+tau/2)*erfc((nu+tau)/(np.sqrt(2*tau)))+ \\\n", 54 | " aa**2*np.exp(2*(nu+tau))*erfc((nu+2*tau)/(np.sqrt(2*tau))) ) - mean**2\n", 55 | "\n", 56 | " eq1 = mean - nu\n", 57 | " eq2 = var - tau\n", 58 | "\n", 59 | " res = nsolve( (eq2, eq1), (aa,ll), (1.67,1.05))\n", 60 | " return float(res[0]),float(res[1])\n" 61 | ] 62 | }, 63 | { 64 | "cell_type": "code", 65 | "execution_count": 3, 66 | "metadata": {}, 67 | "outputs": [ 68 | { 69 | "data": { 70 | "text/plain": [ 71 | "(1.6732632423543778, 1.0507009873554805)" 72 | ] 73 | }, 74 | "execution_count": 3, 75 | "metadata": {}, 76 | "output_type": "execute_result" 77 | } 78 | ], 79 | "source": [ 80 | "### To recover the parameters of the SELU with mean zero and unit variance\n", 81 | "getSeluParameters(0,1)" 82 | ] 83 | }, 84 | { 85 | "cell_type": "code", 86 | "execution_count": 4, 87 | "metadata": {}, 88 | "outputs": [ 89 | { 90 | "data": { 91 | "text/plain": [ 92 | "(1.9769021954242014, 1.073851239616046)" 93 | ] 94 | }, 95 | "execution_count": 4, 96 | "metadata": {}, 97 | "output_type": "execute_result" 98 | } 99 | ], 100 | "source": [ 101 | "### To obtain new parameters for mean zero and variance 2\n", 102 | "myFixedPointMean = -0.1\n", 103 | "myFixedPointVar = 2.0\n", 104 | "myAlpha, myLambda = getSeluParameters(myFixedPointMean,myFixedPointVar)\n", 105 | "getSeluParameters(myFixedPointMean,myFixedPointVar)" 106 | ] 107 | }, 108 | { 109 | "cell_type": "markdown", 110 | "metadata": {}, 111 | "source": [ 112 | "### Adjust the SELU function and Dropout to your new parameters" 113 | ] 114 | }, 115 | { 116 | "cell_type": "code", 117 | "execution_count": 5, 118 | "metadata": {}, 119 | "outputs": [], 120 | "source": [ 121 | "def selu(x):\n", 122 | " with ops.name_scope('elu') as scope:\n", 123 | " alpha = myAlpha\n", 124 | " scale = myLambda\n", 125 | " return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x))" 126 | ] 127 | }, 128 | { 129 | "cell_type": "code", 130 | "execution_count": 6, 131 | "metadata": {}, 132 | "outputs": [], 133 | "source": [ 134 | "def dropout_selu(x, rate, alpha= -myAlpha*myLambda, fixedPointMean=myFixedPointMean, fixedPointVar=myFixedPointVar, \n", 135 | " noise_shape=None, seed=None, name=None, training=False):\n", 136 | " \"\"\"Dropout to a value with rescaling.\"\"\"\n", 137 | "\n", 138 | " def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name):\n", 139 | " keep_prob = 1.0 - rate\n", 140 | " x = ops.convert_to_tensor(x, name=\"x\")\n", 141 | " if isinstance(keep_prob, numbers.Real) and not 0 < keep_prob <= 1:\n", 142 | " raise ValueError(\"keep_prob must be a scalar tensor or a float in the \"\n", 143 | " \"range (0, 1], got %g\" % keep_prob)\n", 144 | " keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name=\"keep_prob\")\n", 145 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 146 | "\n", 147 | " alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name=\"alpha\")\n", 148 | " keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())\n", 149 | "\n", 150 | " if tensor_util.constant_value(keep_prob) == 1:\n", 151 | " return x\n", 152 | "\n", 153 | " noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x)\n", 154 | " random_tensor = keep_prob\n", 155 | " random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype)\n", 156 | " binary_tensor = math_ops.floor(random_tensor)\n", 157 | " ret = x * binary_tensor + alpha * (1-binary_tensor)\n", 158 | "\n", 159 | "\n", 160 | " a = tf.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * tf.pow(alpha-fixedPointMean,2) + fixedPointVar)))\n", 161 | " \n", 162 | " b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha)\n", 163 | " ret = a * ret + b\n", 164 | " ret.set_shape(x.get_shape())\n", 165 | " return ret\n", 166 | "\n", 167 | " with ops.name_scope(name, \"dropout\", [x]) as name:\n", 168 | " return utils.smart_cond(training,\n", 169 | " lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name),\n", 170 | " lambda: array_ops.identity(x))\n" 171 | ] 172 | }, 173 | { 174 | "cell_type": "code", 175 | "execution_count": 7, 176 | "metadata": { 177 | "scrolled": true 178 | }, 179 | "outputs": [ 180 | { 181 | "name": "stdout", 182 | "output_type": "stream", 183 | "text": [ 184 | "WARNING:tensorflow:From :5: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\n", 185 | "Instructions for updating:\n", 186 | "Use tf.where in 2.0, which has the same broadcast rule as np.where\n", 187 | "WARNING:tensorflow:From :12: scalar (from tensorflow.python.framework.tensor_shape) is deprecated and will be removed in a future version.\n", 188 | "Instructions for updating:\n", 189 | "Use tf.TensorShape([]).\n", 190 | "mean/var should be at: -0.1 / 2.0\n", 191 | "Input data mean/var: -0.114250198007 / 1.969062447548\n", 192 | "After selu: -0.114625535905 / 1.961883544922\n", 193 | "After dropout mean/var -0.127010405064 / 1.969492197037\n" 194 | ] 195 | } 196 | ], 197 | "source": [ 198 | "import tensorflow as tf\n", 199 | "import numpy as np\n", 200 | "\n", 201 | "import numbers\n", 202 | "from tensorflow.python.framework import ops\n", 203 | "from tensorflow.python.framework import tensor_shape\n", 204 | "from tensorflow.python.framework import tensor_util\n", 205 | "from tensorflow.python.ops import math_ops\n", 206 | "from tensorflow.python.ops import random_ops\n", 207 | "from tensorflow.python.ops import array_ops\n", 208 | "from tensorflow.python.layers import utils\n", 209 | "\n", 210 | "\n", 211 | "x = tf.Variable(tf.random.normal([10000],mean=myFixedPointMean, stddev=np.sqrt(myFixedPointVar)))\n", 212 | "w = selu(x)\n", 213 | "y = dropout_selu(w,0.2,training=True)\n", 214 | "init = tf.global_variables_initializer()\n", 215 | " \n", 216 | "gpu_options = tf.GPUOptions(allow_growth=True)\n", 217 | "with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:\n", 218 | " sess.run(init)\n", 219 | " z,zz, zzz = sess.run([x, w, y]) \n", 220 | " #print(z)\n", 221 | " #print(zz)\n", 222 | " print(\"mean/var should be at:\", myFixedPointMean, \"/\", myFixedPointVar)\n", 223 | " print(\"Input data mean/var: \", \"{:.12f}\".format(np.mean(z)), \"/\", \"{:.12f}\".format(np.var(z))) \n", 224 | " print(\"After selu: \", \"{:.12f}\".format(np.mean(zz)), \"/\", \"{:.12f}\".format(np.var(zz)))\n", 225 | " print(\"After dropout mean/var\", \"{:.12f}\".format(np.mean(zzz)), \"/\", \"{:.12f}\".format(np.var(zzz)))" 226 | ] 227 | }, 228 | { 229 | "cell_type": "markdown", 230 | "metadata": {}, 231 | "source": [ 232 | "### For completeness: These are the correct expressions for mean zero and unit variance" 233 | ] 234 | }, 235 | { 236 | "cell_type": "code", 237 | "execution_count": 8, 238 | "metadata": {}, 239 | "outputs": [], 240 | "source": [ 241 | "myAlpha = -np.sqrt(2/np.pi) / (np.exp(0.5) * erfc(1/np.sqrt(2))-1 ) \n", 242 | "myLambda = (1-np.sqrt(np.exp(1))*erfc(1/np.sqrt(2))) * \\\n", 243 | " np.sqrt( 2*np.pi/ (2 + np.pi -2*np.sqrt(np.exp(1))*(2+np.pi)*erfc(1/np.sqrt(2)) + \\\n", 244 | " np.exp(1)*np.pi*erfc(1/np.sqrt(2))**2 + 2*np.exp(2)*erfc(np.sqrt(2))))" 245 | ] 246 | }, 247 | { 248 | "cell_type": "code", 249 | "execution_count": 9, 250 | "metadata": {}, 251 | "outputs": [ 252 | { 253 | "name": "stdout", 254 | "output_type": "stream", 255 | "text": [ 256 | "Alpha parameter of the SELU: 1.6732632423543778\n", 257 | "Lambda parameter of the SELU: 1.0507009873554807\n" 258 | ] 259 | } 260 | ], 261 | "source": [ 262 | "print(\"Alpha parameter of the SELU: \", myAlpha)\n", 263 | "print(\"Lambda parameter of the SELU: \", myLambda)" 264 | ] 265 | } 266 | ], 267 | "metadata": { 268 | "kernelspec": { 269 | "display_name": "Python 3", 270 | "language": "python", 271 | "name": "python3" 272 | }, 273 | "language_info": { 274 | "codemirror_mode": { 275 | "name": "ipython", 276 | "version": 3 277 | }, 278 | "file_extension": ".py", 279 | "mimetype": "text/x-python", 280 | "name": "python", 281 | "nbconvert_exporter": "python", 282 | "pygments_lexer": "ipython3", 283 | "version": "3.7.9" 284 | } 285 | }, 286 | "nbformat": 4, 287 | "nbformat_minor": 1 288 | } 289 | -------------------------------------------------------------------------------- /TF_1_x/selu.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | ''' 3 | Tensorflow Implementation of the Scaled ELU function and Dropout 4 | ''' 5 | 6 | from __future__ import absolute_import, division, print_function 7 | import numbers 8 | from tensorflow.contrib import layers 9 | from tensorflow.python.framework import ops 10 | from tensorflow.python.framework import tensor_shape 11 | from tensorflow.python.framework import tensor_util 12 | from tensorflow.python.ops import math_ops 13 | from tensorflow.python.ops import random_ops 14 | from tensorflow.python.ops import array_ops 15 | from tensorflow.python.layers import utils 16 | import tensorflow as tf 17 | 18 | # (1) scale inputs to zero mean and unit variance 19 | 20 | 21 | # (2) use SELUs 22 | def selu(x): 23 | with ops.name_scope('elu') as scope: 24 | alpha = 1.6732632423543772848170429916717 25 | scale = 1.0507009873554804934193349852946 26 | return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x)) 27 | 28 | 29 | # (3) initialize weights with stddev sqrt(1/n) 30 | # e.g. use: 31 | initializer = layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN') 32 | 33 | 34 | # (4) use this dropout 35 | def dropout_selu(x, rate, alpha= -1.7580993408473766, fixedPointMean=0.0, fixedPointVar=1.0, 36 | noise_shape=None, seed=None, name=None, training=False): 37 | """Dropout to a value with rescaling.""" 38 | 39 | def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name): 40 | keep_prob = 1.0 - rate 41 | x = ops.convert_to_tensor(x, name="x") 42 | if isinstance(keep_prob, numbers.Real) and not 0 < keep_prob <= 1: 43 | raise ValueError("keep_prob must be a scalar tensor or a float in the " 44 | "range (0, 1], got %g" % keep_prob) 45 | keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name="keep_prob") 46 | keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar()) 47 | 48 | alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name="alpha") 49 | alpha.get_shape().assert_is_compatible_with(tensor_shape.scalar()) 50 | 51 | if tensor_util.constant_value(keep_prob) == 1: 52 | return x 53 | 54 | noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x) 55 | random_tensor = keep_prob 56 | random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype) 57 | binary_tensor = math_ops.floor(random_tensor) 58 | ret = x * binary_tensor + alpha * (1-binary_tensor) 59 | 60 | a = math_ops.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * math_ops.pow(alpha-fixedPointMean,2) + fixedPointVar))) 61 | 62 | b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha) 63 | ret = a * ret + b 64 | ret.set_shape(x.get_shape()) 65 | return ret 66 | 67 | with ops.name_scope(name, "dropout", [x]) as name: 68 | return utils.smart_cond(training, 69 | lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name), 70 | lambda: array_ops.identity(x)) 71 | -------------------------------------------------------------------------------- /TF_2_x/CIFAR10-Conv-SELU.py: -------------------------------------------------------------------------------- 1 | # Adapted KERAS tutorial 2 | 3 | import tensorflow as tf 4 | import tensorflow.keras as keras 5 | from tensorflow.keras.datasets import cifar10 6 | from tensorflow.keras.preprocessing.image import ImageDataGenerator 7 | from tensorflow.keras.models import Sequential 8 | from tensorflow.keras.layers import Dense, AlphaDropout, Activation, Flatten 9 | from tensorflow.keras.layers import Conv2D, MaxPooling2D 10 | 11 | 12 | import os 13 | import pickle 14 | import numpy as np 15 | 16 | batch_size = 32 17 | num_classes = 10 18 | epochs = 2 19 | data_augmentation = True 20 | num_predictions = 20 21 | save_dir = os.path.join(os.getcwd(), 'saved_models') 22 | model_name = 'keras_cifar10_trained_model.h5' 23 | 24 | # list devices so you can check whether your gpu is available 25 | print(tf.config.list_physical_devices()) 26 | 27 | # The data, shuffled and split between train and test sets: 28 | (x_train, y_train), (x_test, y_test) = cifar10.load_data() 29 | print('x_train shape:', x_train.shape) 30 | print(x_train.shape[0], 'train samples') 31 | print(x_test.shape[0], 'test samples') 32 | 33 | # Convert class vectors to binary class matrices. 34 | y_train = keras.utils.to_categorical(y_train, num_classes) 35 | y_test = keras.utils.to_categorical(y_test, num_classes) 36 | 37 | model = Sequential() 38 | 39 | model.add(Conv2D(32, (3, 3), padding='same', 40 | input_shape=x_train.shape[1:],kernel_initializer='lecun_normal',bias_initializer='zeros')) 41 | model.add(Activation('selu')) 42 | model.add(Conv2D(32, (3, 3),kernel_initializer='lecun_normal',bias_initializer='zeros')) 43 | model.add(Activation('selu')) 44 | model.add(MaxPooling2D(pool_size=(2, 2))) 45 | model.add(AlphaDropout(0.1)) 46 | 47 | model.add(Conv2D(64, (3, 3), padding='same',kernel_initializer='lecun_normal',bias_initializer='zeros')) 48 | model.add(Activation('selu')) 49 | model.add(Conv2D(64, (3, 3),kernel_initializer='lecun_normal',bias_initializer='zeros')) 50 | model.add(Activation('selu')) 51 | model.add(MaxPooling2D(pool_size=(2, 2))) 52 | model.add(AlphaDropout(0.1)) 53 | 54 | model.add(Flatten()) 55 | model.add(Dense(512,kernel_initializer='lecun_normal',bias_initializer='zeros')) 56 | model.add(Activation('selu')) 57 | model.add(AlphaDropout(0.2)) 58 | model.add(Dense(num_classes,kernel_initializer='lecun_normal',bias_initializer='zeros')) 59 | model.add(Activation('softmax')) 60 | 61 | # initiate RMSprop optimizer 62 | opt = keras.optimizers.RMSprop(lr=0.0001, decay=1e-6) 63 | 64 | # Let's train the model using RMSprop 65 | model.compile(loss='categorical_crossentropy', 66 | optimizer=opt, 67 | metrics=['accuracy']) 68 | 69 | x_train = x_train.astype('float32') 70 | x_test = x_test.astype('float32') 71 | x_train /= 255 72 | x_test /= 255 73 | 74 | if not data_augmentation: 75 | print('Not using data augmentation.') 76 | model.fit(x_train, y_train, 77 | batch_size=batch_size, 78 | epochs=epochs, 79 | validation_data=(x_test, y_test), 80 | shuffle=True) 81 | else: 82 | print('Using real-time data augmentation.') 83 | # This will do preprocessing and realtime data augmentation: 84 | datagen = ImageDataGenerator( 85 | featurewise_center=False, # set input mean to 0 over the dataset 86 | samplewise_center=False, # set each sample mean to 0 87 | featurewise_std_normalization=False, # divide inputs by std of the dataset 88 | samplewise_std_normalization=False, # divide each input by its std 89 | zca_whitening=False, # apply ZCA whitening 90 | rotation_range=0, # randomly rotate images in the range (degrees, 0 to 180) 91 | width_shift_range=0.1, # randomly shift images horizontally (fraction of total width) 92 | height_shift_range=0.1, # randomly shift images vertically (fraction of total height) 93 | horizontal_flip=True, # randomly flip images 94 | vertical_flip=False) # randomly flip images 95 | 96 | # Compute quantities required for feature-wise normalization 97 | # (std, mean, and principal components if ZCA whitening is applied). 98 | datagen.fit(x_train) 99 | 100 | # Fit the model on the batches generated by datagen.flow(). 101 | model.fit(datagen.flow(x_train, y_train, 102 | batch_size=batch_size), 103 | steps_per_epoch=x_train.shape[0] // batch_size, 104 | epochs=epochs, 105 | validation_data=(x_test, y_test)) 106 | 107 | # Save model and weights 108 | if not os.path.isdir(save_dir): 109 | os.makedirs(save_dir) 110 | model_path = os.path.join(save_dir, model_name) 111 | model.save(model_path) 112 | print('Saved trained model at %s ' % model_path) 113 | 114 | # Load label names to use in prediction results 115 | label_list_path = 'datasets/cifar-10-batches-py/batches.meta' 116 | 117 | 118 | keras_dir = os.path.expanduser(os.path.join('~', '.keras')) 119 | datadir_base = os.path.expanduser(keras_dir) 120 | if not os.access(datadir_base, os.W_OK): 121 | datadir_base = os.path.join('/tmp', '.keras') 122 | label_list_path = os.path.join(datadir_base, label_list_path) 123 | 124 | with open(label_list_path, mode='rb') as f: 125 | labels = pickle.load(f) 126 | 127 | # Evaluate model with test data set and share sample prediction results 128 | evaluation = model.evaluate(datagen.flow(x_test, y_test, 129 | batch_size=batch_size), 130 | steps=x_test.shape[0] // batch_size) 131 | 132 | print('Model Accuracy = %.5f' % (evaluation[1])) 133 | 134 | f = open('CIFAR10_SELU_results.txt', 'a') 135 | f.write(' Test accuracy:' + str(evaluation[1]) + '\n') 136 | f.close() 137 | 138 | 139 | predict_gen = model.predict(datagen.flow(x_test, y_test, 140 | batch_size=batch_size), 141 | steps=x_test.shape[0] // batch_size) 142 | 143 | for predict_index, predicted_y in enumerate(predict_gen): 144 | actual_label = labels['label_names'][np.argmax(y_test[predict_index])] 145 | predicted_label = labels['label_names'][np.argmax(predicted_y)] 146 | print('Actual Label = %s vs. Predicted Label = %s' % (actual_label, 147 | predicted_label)) 148 | if predict_index == num_predictions: 149 | break 150 | -------------------------------------------------------------------------------- /TF_2_x/MNIST-Conv-SELU.py: -------------------------------------------------------------------------------- 1 | # Adapted KERAS tutorial 2 | 3 | import tensorflow as tf 4 | import tensorflow.keras as keras 5 | from tensorflow.keras.datasets import mnist 6 | from tensorflow.keras.models import Sequential 7 | from tensorflow.keras.layers import Dense, AlphaDropout, Flatten 8 | from tensorflow.keras.layers import Conv2D, MaxPooling2D 9 | from tensorflow.keras import backend as K 10 | 11 | 12 | batch_size = 128 13 | num_classes = 10 14 | epochs = 5 15 | 16 | # input image dimensions 17 | img_rows, img_cols = 28, 28 18 | 19 | # list devices so you can check whether your gpu is available 20 | print(tf.config.list_physical_devices()) 21 | 22 | # the data, shuffled and split between train and test sets 23 | (x_train, y_train), (x_test, y_test) = mnist.load_data() 24 | 25 | if K.image_data_format() == 'channels_first': 26 | x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols) 27 | x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols) 28 | input_shape = (1, img_rows, img_cols) 29 | else: 30 | x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) 31 | x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) 32 | input_shape = (img_rows, img_cols, 1) 33 | 34 | x_train = x_train.astype('float32') 35 | x_test = x_test.astype('float32') 36 | x_train /= 255 37 | #x_train = (x_train - np.mean(x_train))/np.std(x_train) 38 | 39 | x_test /= 255 40 | #x_test = (x_test - np.mean(x_train))/np.std(x_train) 41 | 42 | # create validation file 43 | x_val = x_train[:10000] 44 | x_train = x_train[10000:] 45 | y_val = y_train[:10000] 46 | y_train = y_train[10000:] 47 | 48 | print('x_train shape:', x_train.shape) 49 | print(x_train.shape[0], 'train samples') 50 | print(x_val.shape[0], 'val samples') 51 | print(x_test.shape[0], 'test samples') 52 | 53 | # convert class vectors to binary class matrices 54 | y_train = keras.utils.to_categorical(y_train, num_classes) 55 | y_val = keras.utils.to_categorical(y_val, num_classes) 56 | y_test = keras.utils.to_categorical(y_test, num_classes) 57 | 58 | model = Sequential() 59 | model.add(Conv2D(32, (3, 3),activation='selu',kernel_initializer='lecun_normal',bias_initializer='zeros')) 60 | model.add(Conv2D(64, (3, 3), activation='selu',kernel_initializer='lecun_normal',bias_initializer='zeros')) 61 | model.add(MaxPooling2D(pool_size=(2, 2))) 62 | model.add(AlphaDropout(0.05)) 63 | model.add(Flatten()) 64 | model.add(Dense(512, activation='selu',kernel_initializer='lecun_normal',bias_initializer='zeros')) 65 | model.add(AlphaDropout(0.05)) 66 | model.add(Dense(num_classes, activation='softmax',kernel_initializer='lecun_normal',bias_initializer='zeros')) 67 | 68 | model.compile(loss=keras.losses.categorical_crossentropy, 69 | optimizer=keras.optimizers.Adam(learning_rate=1e-3), 70 | metrics=['accuracy']) 71 | 72 | model.fit(x_train, y_train, 73 | batch_size=batch_size, 74 | epochs=epochs, 75 | verbose=1, 76 | validation_data=(x_val, y_val)) 77 | 78 | score = model.evaluate(x_test, y_test, verbose=0) 79 | print('Test loss:', score[0]) 80 | print('Test accuracy:', score[1]) 81 | 82 | 83 | -------------------------------------------------------------------------------- /TF_2_x/MNIST-MLP-SELU.py: -------------------------------------------------------------------------------- 1 | # Adapted KERAS tutorial 2 | 3 | import tensorflow as tf 4 | import tensorflow.keras as keras 5 | from tensorflow.keras.datasets import mnist 6 | from tensorflow.keras.models import Sequential 7 | from tensorflow.keras.layers import Dense, AlphaDropout, Flatten 8 | from tensorflow.keras.layers import Conv2D, MaxPooling2D 9 | from tensorflow.keras import backend as K 10 | import numpy as np 11 | 12 | 13 | batch_size = 128 14 | num_classes = 10 15 | epochs = 20 16 | 17 | # input image dimensions 18 | img_rows, img_cols = 28, 28 19 | 20 | # list devices so you can check whether your gpu is available 21 | print(tf.config.list_physical_devices()) 22 | 23 | # the data, shuffled and split between train and test sets 24 | (x_train, y_train), (x_test, y_test) = mnist.load_data() 25 | 26 | if K.image_data_format() == 'channels_first': 27 | x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols) 28 | x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols) 29 | input_shape = (1, img_rows, img_cols) 30 | else: 31 | x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) 32 | x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) 33 | input_shape = (img_rows, img_cols, 1) 34 | 35 | x_train = x_train.astype('float32') 36 | x_test = x_test.astype('float32') 37 | x_train /= 255 38 | #x_train = (x_train - np.mean(x_train))/np.std(x_train) 39 | 40 | x_test /= 255 41 | #x_test = (x_test - np.mean(x_train))/np.std(x_train) 42 | 43 | # create validation file 44 | x_val = x_train[:10000] 45 | x_train = x_train[10000:] 46 | y_val = y_train[:10000] 47 | y_train = y_train[10000:] 48 | 49 | print('x_train shape:', x_train.shape) 50 | print(x_train.shape[0], 'train samples') 51 | print(x_val.shape[0], 'val samples') 52 | print(x_test.shape[0], 'test samples') 53 | 54 | # convert class vectors to binary class matrices 55 | y_train = keras.utils.to_categorical(y_train, num_classes) 56 | y_val = keras.utils.to_categorical(y_val, num_classes) 57 | y_test = keras.utils.to_categorical(y_test, num_classes) 58 | 59 | model = Sequential() 60 | model.add(Flatten()) 61 | model.add(Dense(512, activation='selu',kernel_initializer='lecun_normal',bias_initializer='zeros')) 62 | model.add(AlphaDropout(0.05)) 63 | model.add(Dense(256, activation='selu',kernel_initializer='lecun_normal',bias_initializer='zeros')) 64 | model.add(AlphaDropout(0.05)) 65 | model.add(Dense(num_classes, activation='softmax',kernel_initializer='lecun_normal',bias_initializer='zeros')) 66 | 67 | model.compile(loss=keras.losses.categorical_crossentropy, 68 | optimizer=keras.optimizers.Adam(learning_rate=0.001), 69 | metrics=['accuracy']) 70 | 71 | model.fit(x_train, y_train, 72 | batch_size=batch_size, 73 | epochs=epochs, 74 | verbose=1, 75 | validation_data=(x_val, y_val)) 76 | 77 | score = model.evaluate(x_test, y_test, verbose=0) 78 | print('Test loss:', score[0]) 79 | print('Test accuracy:', score[1]) 80 | 81 | -------------------------------------------------------------------------------- /TF_2_x/README.md: -------------------------------------------------------------------------------- 1 | Both Keras examples are based on dependencies described in [this](../environment.yml) environment file. -------------------------------------------------------------------------------- /Tox21/README.md: -------------------------------------------------------------------------------- 1 | ## Tox21 data set 2 | - [download](http://bioinf.jku.at/research/DeepTox/tox21.zip) 3 | 4 | -------------------------------------------------------------------------------- /UCI/README.md: -------------------------------------------------------------------------------- 1 | ## UCI data set 2 | - [download from original source](http://persoal.citius.usc.es/manuel.fernandez.delgado/papers/jmlr/data.tar.gz) 3 | - [download our processed version of the data set](http://www.bioinf.jku.at/people/klambauer/data_py.zip) 4 | -------------------------------------------------------------------------------- /environment.yml: -------------------------------------------------------------------------------- 1 | name: snn 2 | channels: 3 | - pytorch 4 | - conda-forge 5 | - defaults 6 | dependencies: 7 | - python=3.7 8 | - pip=21.0.1 9 | - numpy=1.20.0 10 | - scipy=1.6.0 11 | - sympy=1.7.1 12 | - pandas=1.2.2 13 | - scikit-learn=0.24.1 14 | - matplotlib=3.3.4 15 | - jupyter=1.0.0 16 | - tqdm=4.56.0 17 | - tensorflow=2.3.0 18 | # GPU compatibility, check your nvidia drivers to find what toolkit is supported by your GPU 19 | - cudatoolkit=10.1 20 | # Pytorch dependencies 21 | - pytorch=1.7.1 22 | - torchvision=0.8.2 23 | -------------------------------------------------------------------------------- /figure1/README.md: -------------------------------------------------------------------------------- 1 | # Reproducing Figure 1 2 | 3 | This contains the code necessary to reproduce Figure 1 from the SNN paper. Note that the code uses the [biutils](https://github.com/untom/biutils) package to load the MNIST/CIFAR10 datasets. 4 | You can use 5 | 6 | `pip install git+https://github.com/untom/biutils.git` 7 | 8 | to easily install the package. 9 | 10 | The data for the plot was created by running 11 | 12 | ./run.py -g 0 -d 08 -a selu -l 1e-5 -e 2000 --dataset mnist 13 | ./run.py -g 1 -d 16 -a selu -l 1e-5 -e 2000 --dataset mnist 14 | ./run.py -g 2 -d 32 -a selu -l 1e-5 -e 2000 --dataset mnist 15 | ./run.py -g 3 -d 08 -a relu --batchnorm -l 1e-5 -e 2000 --dataset mnist 16 | ./run.py -g 0 -d 16 -a relu --batchnorm -l 1e-5 -e 2000 --dataset mnist 17 | ./run.py -g 1 -d 32 -a relu --batchnorm -l 1e-5 -e 2000 --dataset mnist 18 | 19 | ./run.py -g 0 -d 08 -a selu -l 1e-5 -e 2000 --dataset cifar10 20 | ./run.py -g 1 -d 16 -a selu -l 1e-5 -e 2000 --dataset cifar10 21 | ./run.py -g 2 -d 32 -a selu -l 1e-5 -e 2000 --dataset cifar10 22 | ./run.py -g 3 -d 08 -a relu --batchnorm -l 1e-5 -e 2000 --dataset cifar10 23 | ./run.py -g 0 -d 16 -a relu --batchnorm -l 1e-5 -e 2000 --dataset cifar10 24 | ./run.py -g 1 -d 32 -a relu --batchnorm -l 1e-5 -e 2000 --dataset cifar10 25 | 26 | The plots where then created using `create_plots.ipynb`. 27 | -------------------------------------------------------------------------------- /figure1/run.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | 3 | import os 4 | 5 | from sklearn.preprocessing import StandardScaler 6 | 7 | import biutils # used to load the dataset 8 | import utils 9 | 10 | 11 | def model(dataset, n_layers, n_hidden, activation, dropout_rate, use_batchnorm): 12 | 13 | x_tr, y_tr, x_va, y_va = biutils.load_dataset(dataset) 14 | s = StandardScaler() 15 | s.fit(x_tr) 16 | x_tr = s.transform(x_tr) 17 | x_va = s.transform(x_va) 18 | 19 | if n_hidden == -1: # use as many hidden# as there are input features 20 | n_hidden = x_tr.shape[1] 21 | 22 | if activation == 'relu': 23 | act_fn = tf.nn.relu 24 | init_scale = 2.0 25 | elif activation == 'tanh': 26 | act_fn = tf.nn.tanh 27 | init_scale = 1.0 28 | elif activation == 'selu': 29 | act_fn = utils.selu 30 | init_scale = 1.0 31 | else: 32 | assert False, "Unknown activation" 33 | 34 | tf.reset_default_graph() 35 | x = tf.placeholder(np.float32, [None, x_tr.shape[1]], name="x") 36 | y = tf.placeholder(np.float32, [None, y_tr.shape[1]], name="y") 37 | is_training = tf.placeholder_with_default(tf.constant(False, tf.bool), shape=[], name='is_training') 38 | 39 | h = x 40 | if dropout_rate > 0.0: 41 | h = tf.layers.dropout(h, 0.2, training=is_training) 42 | 43 | for i in range(n_layers): 44 | s = np.sqrt(init_scale/h.get_shape().as_list()[1]) 45 | init = tf.random_normal_initializer(stddev=s) 46 | h = tf.layers.dense(h, n_hidden, activation=act_fn, name='layer%d' % i, kernel_initializer=init) 47 | if use_batchnorm: 48 | h = tf.layers.batch_normalization(h, training=is_training) 49 | if dropout_rate > 0.0: 50 | h = tf.layers.dropout(h, dropout_rate, training=is_training) 51 | 52 | with tf.variable_scope('output_layer') as scope: 53 | o = tf.layers.dense(h, y_tr.shape[1], activation=None, name=scope) 54 | scope.reuse_variables() 55 | 56 | return (x_tr, y_tr, x_va, y_va), (x, y, is_training), o 57 | 58 | 59 | def run(n_layers, n_hidden, n_epochs, learning_rate, dataset, activation, logdir_base='/tmp', 60 | batch_size=64, dropout_rate=0.0, use_batchnorm=False): 61 | 62 | ld = '%s%s_d%02d_h%d_l%1.0e_%s' % (activation, 63 | 'bn' if use_batchnorm else '', 64 | n_layers, n_hidden, learning_rate, 65 | utils.get_timestamp()) 66 | logdir = os.path.join(logdir_base, dataset, ld) 67 | print(logdir) 68 | 69 | dataset, variables, logits = model(dataset, n_layers, n_hidden, activation, dropout_rate, use_batchnorm) 70 | x_tr, y_tr, x_va, y_va = dataset 71 | x, y, is_training = variables 72 | 73 | prob_op = tf.nn.softmax(logits) 74 | loss_op = tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=logits) 75 | 76 | optimizer = tf.train.GradientDescentOptimizer(learning_rate) 77 | variables_to_train = tf.trainable_variables() 78 | grads = optimizer.compute_gradients(loss_op, variables_to_train) 79 | global_step = tf.train.get_global_step() 80 | train_op = optimizer.apply_gradients(grads, global_step=global_step) 81 | 82 | loss_val = tf.Variable(0.0, trainable=False, dtype=np.float32) 83 | acc_op, acc_upd = tf.metrics.accuracy(tf.argmax(y, 1), tf.argmax(prob_op, 1), name='accuracy') 84 | acc_tr_op = tf.summary.scalar('acc_tr', acc_op) 85 | acc_va_op = tf.summary.scalar('acc_va', acc_op) 86 | loss_tr_op = tf.summary.scalar('loss_tr', loss_val / x_tr.shape[0]) 87 | loss_va_op = tf.summary.scalar('loss_va', loss_val / x_va.shape[0]) 88 | metric_vars = [i for i in tf.local_variables() if i.name.split('/')[0] == 'accuracy'] 89 | reset_op = [tf.variables_initializer(metric_vars), loss_val.assign(0.0)] 90 | loss_upd = loss_val.assign_add(tf.reduce_sum(loss_op)) 91 | smry_tr = tf.summary.merge([acc_tr_op, loss_tr_op]) 92 | 93 | smry_va = tf.summary.merge([acc_va_op, loss_va_op]) 94 | config = tf.ConfigProto(intra_op_parallelism_threads=2, 95 | use_per_session_threads=True, 96 | gpu_options = tf.GPUOptions(allow_growth=True) 97 | ) 98 | with tf.Session(config=config) as sess: 99 | log = tf.summary.FileWriter(logdir, sess.graph) 100 | saver = tf.train.Saver(max_to_keep=100) 101 | sess.run(tf.global_variables_initializer()) 102 | sess.run(tf.local_variables_initializer()) 103 | fd_tr = {is_training: True} 104 | 105 | for cur_epoch in range(n_epochs): 106 | # get stats over whole training set 107 | for fd in utils.generate_minibatches(batch_size, [x, y], [x_tr, y_tr], feed_dict=fd_tr, shuffle=False): 108 | sess.run([acc_upd, loss_upd], feed_dict=fd) 109 | log.add_summary(sess.run(smry_tr, feed_dict=fd), cur_epoch) 110 | sess.run(reset_op) 111 | 112 | # training 113 | for fd in utils.generate_minibatches(batch_size, [x, y], [x_tr, y_tr], feed_dict=fd_tr): 114 | sess.run([train_op], feed_dict=fd) 115 | 116 | # validation 117 | for fd in utils.generate_minibatches(batch_size, [x, y], [x_va, y_va], shuffle=False): 118 | sess.run([acc_upd, loss_upd], feed_dict=fd) 119 | smry, acc = sess.run([smry_va, acc_op]) 120 | log.add_summary(smry, cur_epoch) 121 | sess.run(reset_op) 122 | print("%3d: %3.3f" % (cur_epoch, acc), flush=True) 123 | 124 | if cur_epoch % 250 == 0 and cur_epoch > 0: 125 | saver.save(sess, os.path.join(logdir, 'model'), global_step=cur_epoch) 126 | 127 | 128 | from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter 129 | parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter) 130 | parser.add_argument("-n", "--nhidden", type=int, help='hidden units (-1: use input size)', default=-1) 131 | parser.add_argument("-d", "--depth", type=int, help='number of hidden layers', default=3) 132 | parser.add_argument("-a", "--activation", choices=['relu', 'selu', 'tanh'], default='relu') 133 | parser.add_argument("-b", "--batchsize", type=int, help='batch size', default=128) 134 | parser.add_argument("-e", "--epochs", type=int, help='number of training epochs', default=30) 135 | parser.add_argument("-l", "--learningrate", type=float, help='learning rate', default=1e-5) 136 | parser.add_argument("-g", "--gpuid", type=str, help='GPU to use (leave blank for CPU only)', default="") 137 | parser.add_argument("--batchnorm", help='use batchnorm', action="store_true") 138 | parser.add_argument("--dropout", type=float, help='hidden dropout rate (implies input-dropout of 0.2)', default=0.0) 139 | parser.add_argument("--dataset", type=str, help='name of dataset', default='mnist_bgimg') 140 | parser.add_argument("--logdir", type=str, help='directory for TF logs and summaries', default="logs") 141 | 142 | # by parsing the arguments already, we can bail out now instead of waiting 143 | # for TF to load, in case the arguments aren't ok 144 | args = parser.parse_args() 145 | os.environ['CUDA_VISIBLE_DEVICES'] = args.gpuid 146 | 147 | import numpy as np 148 | import tensorflow as tf 149 | 150 | logdir_base = os.getcwd() 151 | run(args.depth, args.nhidden, args.epochs, args.learningrate, args.dataset, 152 | args.activation, args.logdir, args.batchsize, args.dropout, args.batchnorm) 153 | -------------------------------------------------------------------------------- /figure1/utils.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | ''' 3 | Tensorflow Implementation of the Scaled ELU function and Dropout 4 | ''' 5 | 6 | import numbers 7 | 8 | import numpy as np 9 | import tensorflow as tf 10 | from tensorflow.contrib import layers 11 | from tensorflow.python.framework import ops 12 | from tensorflow.python.framework import tensor_shape 13 | from tensorflow.python.framework import tensor_util 14 | from tensorflow.python.layers import utils 15 | from tensorflow.python.ops import array_ops 16 | from tensorflow.python.ops import math_ops 17 | from tensorflow.python.ops import random_ops 18 | 19 | 20 | # (1) scale inputs to zero mean and unit variance 21 | 22 | 23 | # (2) use SELUs 24 | def selu(x): 25 | with ops.name_scope('elu') as scope: 26 | alpha = 1.6732632423543772848170429916717 27 | scale = 1.0507009873554804934193349852946 28 | return scale*tf.where(x>=0.0, x, alpha*tf.nn.elu(x)) 29 | 30 | 31 | # (3) initialize weights with stddev sqrt(1/n) 32 | # e.g. use: 33 | initializer = layers.variance_scaling_initializer(factor=1.0, mode='FAN_IN') 34 | 35 | 36 | # (4) use this dropout 37 | def dropout_selu(x, rate, alpha= -1.7580993408473766, fixedPointMean=0.0, fixedPointVar=1.0, 38 | noise_shape=None, seed=None, name=None, training=False): 39 | """Dropout to a value with rescaling.""" 40 | 41 | def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name): 42 | keep_prob = 1.0 - rate 43 | x = ops.convert_to_tensor(x, name="x") 44 | if isinstance(keep_prob, numbers.Real) and not 0 < keep_prob <= 1: 45 | raise ValueError("keep_prob must be a scalar tensor or a float in the " 46 | "range (0, 1], got %g" % keep_prob) 47 | keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name="keep_prob") 48 | keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar()) 49 | 50 | alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name="alpha") 51 | keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar()) 52 | 53 | if tensor_util.constant_value(keep_prob) == 1: 54 | return x 55 | 56 | noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x) 57 | random_tensor = keep_prob 58 | random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype) 59 | binary_tensor = math_ops.floor(random_tensor) 60 | ret = x * binary_tensor + alpha * (1-binary_tensor) 61 | 62 | a = math_ops.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * math_ops.pow(alpha-fixedPointMean,2) + fixedPointVar))) 63 | 64 | b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha) 65 | ret = a * ret + b 66 | ret.set_shape(x.get_shape()) 67 | return ret 68 | 69 | with ops.name_scope(name, "dropout", [x]) as name: 70 | return utils.smart_cond(training, 71 | lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name), 72 | lambda: array_ops.identity(x)) 73 | 74 | 75 | def get_timestamp(fmt='%y%m%d_%H%M'): 76 | '''Returns a string that contains the current date and time. 77 | 78 | Suggested formats: 79 | short_format=%y%m%d_%H%M (default) 80 | long format=%Y%m%d_%H%M%S 81 | ''' 82 | import datetime 83 | now = datetime.datetime.now() 84 | return datetime.datetime.strftime(now, fmt) 85 | 86 | 87 | def generate_slices(n, slice_size, allow_smaller_final_batch=True): 88 | """Generates slices of given slice_size up to n""" 89 | start, end = 0, 0 90 | for pack_num in range(int(n / slice_size)): 91 | end = start + slice_size 92 | yield slice(start, end, None) 93 | start = end 94 | # last slice might not be a full batch 95 | if allow_smaller_final_batch: 96 | if end < n: 97 | yield slice(end, n, None) 98 | 99 | 100 | def generate_minibatches(batch_size, ph_list, data_list, n_epochs=1, 101 | allow_smaller_final_batch=False, shuffle=True, 102 | feed_dict=None): 103 | cnt_epochs = 0 104 | assert len(ph_list) == len(data_list), "Passed different number of data and placeholders" 105 | assert len(data_list) >= 0, "Passed empty lists" 106 | 107 | n_samples = data_list[0].shape[0] 108 | n_items = len(data_list) 109 | 110 | while True: 111 | if shuffle: 112 | idx = np.arange(n_samples) 113 | np.random.shuffle(idx) 114 | for i in range(n_items): 115 | data_list[i] = data_list[i][idx] 116 | 117 | if feed_dict is None: 118 | feed_dict = {} 119 | 120 | for s in generate_slices(n_samples, batch_size, allow_smaller_final_batch): 121 | for i in range(n_items): 122 | ph = ph_list[i] 123 | d = data_list[i][s] 124 | feed_dict[ph] = d 125 | yield feed_dict 126 | cnt_epochs += 1 127 | if n_epochs is not None and cnt_epochs >= n_epochs: 128 | break 129 | --------------------------------------------------------------------------------