├── LICENSE ├── README.md ├── datagenerator.py ├── evaluate.py ├── hyperparameters.py ├── layers2D.py ├── layers3D.py ├── losses.py ├── modelmemory.py ├── network.py ├── plotmetrics.ipynb ├── predict.py └── train.py /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Keras U-Net 2 | 3 | ## What is it? 4 | 5 | Keras implementation of a 2D/3D U-Net with the following implementations provided: 6 | * Additive attention -- [Attention U-Net: Learning Where to Look for the Pancreas](https://arxiv.org/abs/1804.03999) 7 | * Inception convolutions w/ dilated convolutions -- [Going Deeper with Convolutions](https://arxiv.org/abs/1409.4842) and [Multi-Scale Context Aggregation by Dilated Convolutions](https://arxiv.org/abs/1511.07122) 8 | * Recurrent convolutions -- [R2U-Net](https://arxiv.org/abs/1802.06955) 9 | * Focal Tversky Loss 10 | * Dice Coefficient Loss 11 | 12 | ## Usage 13 | 14 | ### Dependencies 15 | 16 | This repository depends on the following libraries: 17 | * Tensorflow 18 | * Keras 19 | * Python 3 20 | * Numpy 21 | * Matplotlib 22 | 23 | ### Building your network 24 | 25 | The pre-implemented layers are available in [`layers3D.py`](layers3D.py). Use the layers to build your preferred network configuration in [`network.py`](network.py) 26 | 27 | ##### Example 28 | 29 | ``` 30 | from layers3D import * 31 | from tensorflow.keras.models import Model 32 | 33 | def network(input_img, n_filters=16, dropout=0.5, batchnorm=True): 34 | outputs = inception_block(input_img, n_filters=n_filters, batchnorm=batchnorm, strides=1, recurrent=2) 35 | model = Model(inputs=[input_img], outputs=[outputs]) 36 | return model 37 | ``` 38 | *Refer to [`network.py`](network.py) for a full example* 39 | 40 | ### Data Generator 41 | 42 | Rewrite the `__data_generation()` method in [`datagenerator.py`](datagenerator.py) to supply batches of data during training 43 | 44 | ##### Example 45 | 46 | ``` 47 | def __data_generation(self, list_IDs_temp): 48 | 49 | X = np.empty((self.batch_size, *self.dim, self.n_channels)) 50 | y = np.empty((self.batch_size, *self.dim, self.n_channels)) 51 | 52 | for i, ID in enumerate(list_IDs_temp): 53 | # Write logic for selecting/manipulating X and y here 54 | X[i,] = np.load('path/to/x/ID') 55 | y[i,] = np.load('path/to/y/ID') 56 | 57 | return X, y 58 | ``` 59 | 60 | The `DataGenerator` class in [`train.py`](train.py) takes in `list` arguments containing the ID (filenames) of X and y 61 | 62 | ### Hyperparameters 63 | 64 | Set the appropriate values for the hyper-parameters listed in [`hyperparameters.py`](hyperparameters.py) 65 | 66 | ### Train 67 | 68 | Run [`train.py`](train.py) once all the configuration is done to train your network 69 | 70 | ### Testing 71 | 72 | Run [`evaluate.py`](evaluate.py) or [`predict.py`](predict.py) with the appropriate list_IDs provided to the `DataGenerator` 73 | 74 | -------------------------------------------------------------------------------- /datagenerator.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.utils import Sequence 2 | import numpy as np 3 | 4 | 5 | class DataGenerator(Sequence): 6 | def __init__(self, 7 | list_IDs, 8 | labels=[], 9 | batch_size=1, 10 | dim=(512, 512, 512), 11 | n_channels=1, 12 | n_classes=1, 13 | shuffle=True): 14 | self.dim = dim 15 | self.batch_size = batch_size 16 | self.labels = labels 17 | self.list_IDs = list_IDs 18 | self.n_channels = n_channels 19 | self.n_classes = n_classes 20 | self.shuffle = shuffle 21 | self.on_epoch_end() 22 | 23 | def __len__(self): 24 | 25 | # Counts the number of possible batches that can be made from the total available datasets in list_IDs 26 | # Rule of thumb, num_datasets % batch_size = 0, so every sample is seen 27 | return int(np.floor(len(self.list_IDs) / self.batch_size)) 28 | 29 | def __getitem__(self, index): 30 | 31 | # Gets the indexes of batch_size number of data from list_IDs for one epoch 32 | # If batch_size = 8, 8 files/indexes from list_ID are selected 33 | # Makes sure that on next epoch, the batch does not come from same indexes as the previous batch 34 | # The same batch is not seen again until __len()__ - 1 batches are done 35 | 36 | indexes = self.indexes[index * self.batch_size:(index + 1) * 37 | self.batch_size] 38 | list_IDs_temp = [self.list_IDs[k] for k in indexes] 39 | 40 | X, y = self.__data_generation(list_IDs_temp) 41 | 42 | return X, y 43 | 44 | def on_epoch_end(self): 45 | 46 | self.indexes = np.arange(len(self.list_IDs)) 47 | if self.shuffle: 48 | np.random.shuffle(self.indexes) 49 | 50 | def __data_generation(self, list_IDs_temp): 51 | 52 | # Creates an empty placeholder array that will be populated with data that is to be supplied 53 | X = np.empty((self.batch_size, *self.dim, self.n_channels)) 54 | y = np.empty((self.batch_size, *self.dim, self.n_channels)) 55 | 56 | for i, ID in enumerate(list_IDs_temp): 57 | # Write logic for selecting/manipulating X and y here 58 | pass 59 | 60 | return X, y 61 | -------------------------------------------------------------------------------- /evaluate.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.models import load_model 2 | from hyperparameters import * 3 | from datagenerator import DataGenerator 4 | from losses import * 5 | 6 | model = load_model(save_path) 7 | 8 | list_IDs = [] 9 | for filename in os.listdir(test_path): 10 | # Write logic to add filenames of train images to list_IDs which will be processed by DataGenerator 11 | # later on 12 | pass 13 | 14 | evaluate_gen = DataGenerator(list_IDs=list_IDs, 15 | dim=dimensions, 16 | batch_size=batch_size, 17 | shuffle=True) 18 | 19 | # Returns test loss using metric specified in train.py during training 20 | model.evaluate_generator(evaluate_gen, steps=0, verbose=2, workers=20) 21 | -------------------------------------------------------------------------------- /hyperparameters.py: -------------------------------------------------------------------------------- 1 | from losses import * 2 | 3 | ################################################################################ 4 | # Hyperparameters 5 | ################################################################################ 6 | 7 | # Leaky ReLU 8 | alpha = 0.1 9 | 10 | # Input Image Dimensions 11 | # (rows, cols, depth, channels) 12 | input_dimensions = (512, 512, 512, 1) 13 | dimensions = (512, 512, 512) 14 | 15 | # Training parameters 16 | num_initial_filters = 32 17 | batchnorm = True 18 | 19 | # batch_size must be a multiple of num_gpu to ensure GPUs are not starved of data 20 | num_gpu = 8 21 | batch_size = 8 22 | steps_per_epoch = 1 23 | 24 | learning_rate = 0.00001 25 | loss = tversky_loss 26 | metrics = [dice_coef] 27 | epochs = 70000 28 | 29 | # Paths 30 | checkpoint_path = "" 31 | log_path = "" 32 | save_path = "" 33 | train_path = "" 34 | test_path = "" 35 | -------------------------------------------------------------------------------- /layers2D.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.layers import * 2 | import tensorflow.keras.backend as K 3 | from hyperparameters import alpha 4 | K.set_image_data_format('channels_last') 5 | 6 | 7 | def conv2d_block(input_tensor, 8 | n_filters, 9 | kernel_size=3, 10 | batchnorm=True, 11 | strides=1, 12 | dilation_rate=1, 13 | recurrent=1): 14 | 15 | # A wrapper of the Keras Conv2D block to serve as a building block for downsampling layers 16 | # Includes options to use batch normalization, dilation and recurrence 17 | 18 | conv = Conv2D(filters=n_filters, 19 | kernel_size=kernel_size, 20 | strides=strides, 21 | kernel_initializer="he_normal", 22 | padding="same", 23 | dilation_rate=dilation_rate)(input_tensor) 24 | if batchnorm: 25 | conv = BatchNormalization()(conv) 26 | output = LeakyReLU(alpha=alpha)(conv) 27 | 28 | for _ in range(recurrent - 1): 29 | conv = Conv2D(filters=n_filters, 30 | kernel_size=kernel_size, 31 | strides=1, 32 | kernel_initializer="he_normal", 33 | padding="same", 34 | dilation_rate=dilation_rate)(output) 35 | if batchnorm: 36 | conv = BatchNormalization()(conv) 37 | res = LeakyReLU(alpha=alpha)(conv) 38 | output = Add()([output, res]) 39 | 40 | return output 41 | 42 | 43 | def residual_block(input_tensor, 44 | n_filters, 45 | kernel_size=3, 46 | strides=1, 47 | batchnorm=True, 48 | recurrent=1, 49 | dilation_rate=1): 50 | 51 | # A residual block based on the ResNet architecture incorporating use of short-skip connections 52 | # Uses two successive convolution layers by default 53 | 54 | res = conv2d_block(input_tensor, 55 | n_filters=n_filters, 56 | kernel_size=kernel_size, 57 | strides=strides, 58 | batchnorm=batchnorm, 59 | dilation_rate=dilation_rate, 60 | recurrent=recurrent) 61 | res = conv2d_block(res, 62 | n_filters=n_filters, 63 | kernel_size=kernel_size, 64 | strides=1, 65 | batchnorm=batchnorm, 66 | dilation_rate=dilation_rate, 67 | recurrent=recurrent) 68 | 69 | shortcut = conv2d_block(input_tensor, 70 | n_filters=n_filters, 71 | kernel_size=1, 72 | strides=strides, 73 | batchnorm=batchnorm, 74 | dilation_rate=1) 75 | if batchnorm: 76 | shortcut = BatchNormalization()(shortcut) 77 | 78 | output = Add()([shortcut, res]) 79 | return output 80 | 81 | 82 | def inception_block(input_tensor, 83 | n_filters, 84 | kernel_size=3, 85 | strides=1, 86 | batchnorm=True, 87 | recurrent=1, 88 | layers=[]): 89 | 90 | # Inception-style convolutional block similar to InceptionNet 91 | # The first convolution follows the function arguments, while subsequent inception convolutions follow the parameters in 92 | # argument, layers 93 | 94 | # layers is a nested list containing the different secondary inceptions in the format of (kernel_size, dil_rate) 95 | 96 | # E.g => layers=[ [(3,1),(3,1)], [(5,1)], [(3,1),(3,2)] ] 97 | # This will implement 3 sets of secondary convolutions 98 | # Set 1 => 3x3 dil = 1 followed by another 3x3 dil = 1 99 | # Set 2 => 5x5 dil = 1 100 | # Set 3 => 3x3 dil = 1 followed by 3x3 dil = 2 101 | 102 | res = conv2d_block(input_tensor, 103 | n_filters=n_filters, 104 | kernel_size=kernel_size, 105 | strides=strides, 106 | batchnorm=batchnorm, 107 | dilation_rate=1, 108 | recurrent=recurrent) 109 | 110 | temp = [] 111 | for layer in layers: 112 | local_res = res 113 | for conv in layer: 114 | incep_kernel_size = conv[0] 115 | incep_dilation_rate = conv[1] 116 | local_res = conv2d_block(local_res, 117 | n_filters=n_filters, 118 | kernel_size=incep_kernel_size, 119 | strides=1, 120 | batchnorm=batchnorm, 121 | dilation_rate=incep_dilation_rate, 122 | recurrent=recurrent) 123 | temp.append(local_res) 124 | 125 | temp = concatenate(temp) 126 | res = conv2d_block(temp, 127 | n_filters=n_filters, 128 | kernel_size=1, 129 | strides=1, 130 | batchnorm=batchnorm, 131 | dilation_rate=1) 132 | 133 | shortcut = conv2d_block(input_tensor, 134 | n_filters=n_filters, 135 | kernel_size=1, 136 | strides=strides, 137 | batchnorm=batchnorm, 138 | dilation_rate=1) 139 | if batchnorm: 140 | shortcut = BatchNormalization()(shortcut) 141 | 142 | output = Add()([shortcut, res]) 143 | return output 144 | 145 | 146 | def transpose_block(input_tensor, 147 | skip_tensor, 148 | n_filters, 149 | kernel_size=3, 150 | strides=1, 151 | batchnorm=True, 152 | recurrent=1): 153 | 154 | # A wrapper of the Keras Conv2DTranspose block to serve as a building block for upsampling layers 155 | 156 | shape_x = K.int_shape(input_tensor) 157 | shape_xskip = K.int_shape(skip_tensor) 158 | 159 | conv = Conv2DTranspose(filters=n_filters, 160 | kernel_size=kernel_size, 161 | padding='same', 162 | strides=(shape_xskip[1] // shape_x[1], 163 | shape_xskip[2] // shape_x[2]), 164 | kernel_initializer="he_normal")(input_tensor) 165 | conv = LeakyReLU(alpha=alpha)(conv) 166 | 167 | act = conv2d_block(conv, 168 | n_filters=n_filters, 169 | kernel_size=kernel_size, 170 | strides=1, 171 | batchnorm=batchnorm, 172 | dilation_rate=1, 173 | recurrent=recurrent) 174 | output = Concatenate(axis=3)([act, skip_tensor]) 175 | return output 176 | 177 | 178 | def expend_as(tensor, rep): 179 | 180 | # Anonymous lambda function to expand the specified axis by a factor of argument, rep. 181 | # If tensor has shape (512,512,N), lambda will return a tensor of shape (512,512,N*rep), if specified axis=2 182 | 183 | my_repeat = Lambda(lambda x, repnum: K.repeat_elements(x, repnum, axis=3), 184 | arguments={'repnum': rep})(tensor) 185 | return my_repeat 186 | 187 | 188 | def AttnGatingBlock(x, g, inter_shape): 189 | 190 | shape_x = K.int_shape(x) 191 | shape_g = K.int_shape(g) 192 | 193 | # Getting the gating signal to the same number of filters as the inter_shape 194 | phi_g = Conv2D(filters=inter_shape, 195 | kernel_size=1, 196 | strides=1, 197 | padding='same')(g) 198 | 199 | # Getting the x signal to the same shape as the gating signal 200 | theta_x = Conv2D(filters=inter_shape, 201 | kernel_size=3, 202 | strides=(shape_x[1] // shape_g[1], 203 | shape_x[2] // shape_g[2]), 204 | padding='same')(x) 205 | 206 | # Element-wise addition of the gating and x signals 207 | add_xg = add([phi_g, theta_x]) 208 | add_xg = Activation('relu')(add_xg) 209 | 210 | # 1x1x1 convolution 211 | psi = Conv2D(filters=1, kernel_size=1, padding='same')(add_xg) 212 | psi = Activation('sigmoid')(psi) 213 | shape_sigmoid = K.int_shape(psi) 214 | 215 | # Upsampling psi back to the original dimensions of x signal 216 | upsample_sigmoid_xg = UpSampling2D(size=(shape_x[1] // shape_sigmoid[1], 217 | shape_x[2] // 218 | shape_sigmoid[2]))(psi) 219 | 220 | # Expanding the filter axis to the number of filters in the original x signal 221 | upsample_sigmoid_xg = expend_as(upsample_sigmoid_xg, shape_x[3]) 222 | 223 | # Element-wise multiplication of attention coefficients back onto original x signal 224 | attn_coefficients = multiply([upsample_sigmoid_xg, x]) 225 | 226 | # Final 1x1x1 convolution to consolidate attention signal to original x dimensions 227 | output = Conv2D(filters=shape_x[3], 228 | kernel_size=1, 229 | strides=1, 230 | padding='same')(attn_coefficients) 231 | output = BatchNormalization()(output) 232 | return output 233 | 234 | 235 | def GatingSignal(input_tensor, batchnorm=True): 236 | 237 | # 1x1x1 convolution to consolidate gating signal into the required dimensions 238 | # Not required most of the time, unless another ReLU and batch_norm is required on gating signal 239 | 240 | shape = K.int_shape(input_tensor) 241 | conv = Conv2D(filters=shape[3], 242 | kernel_size=1, 243 | strides=1, 244 | padding="same", 245 | kernel_initializer="he_normal")(input_tensor) 246 | if batchnorm: 247 | conv = BatchNormalization()(conv) 248 | output = LeakyReLU(alpha=alpha)(conv) 249 | return output 250 | -------------------------------------------------------------------------------- /layers3D.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.layers import * 2 | import tensorflow.keras.backend as K 3 | from hyperparameters import alpha 4 | K.set_image_data_format('channels_last') 5 | 6 | 7 | def conv3d_block(input_tensor, 8 | n_filters, 9 | kernel_size=3, 10 | batchnorm=True, 11 | strides=1, 12 | dilation_rate=1, 13 | recurrent=1): 14 | 15 | # A wrapper of the Keras Conv3D block to serve as a building block for downsampling layers 16 | # Includes options to use batch normalization, dilation and recurrence 17 | 18 | conv = Conv3D(filters=n_filters, 19 | kernel_size=kernel_size, 20 | strides=strides, 21 | kernel_initializer="he_normal", 22 | padding="same", 23 | dilation_rate=dilation_rate)(input_tensor) 24 | if batchnorm: 25 | conv = BatchNormalization()(conv) 26 | output = LeakyReLU(alpha=alpha)(conv) 27 | 28 | for _ in range(recurrent - 1): 29 | conv = Conv3D(filters=n_filters, 30 | kernel_size=kernel_size, 31 | strides=1, 32 | kernel_initializer="he_normal", 33 | padding="same", 34 | dilation_rate=dilation_rate)(output) 35 | if batchnorm: 36 | conv = BatchNormalization()(conv) 37 | res = LeakyReLU(alpha=alpha)(conv) 38 | output = Add()([output, res]) 39 | 40 | return output 41 | 42 | 43 | def residual_block(input_tensor, 44 | n_filters, 45 | kernel_size=3, 46 | strides=1, 47 | batchnorm=True, 48 | recurrent=1, 49 | dilation_rate=1): 50 | 51 | # A residual block based on the ResNet architecture incorporating use of short-skip connections 52 | # Uses two successive convolution layers by default 53 | 54 | res = conv3d_block(input_tensor, 55 | n_filters=n_filters, 56 | kernel_size=kernel_size, 57 | strides=strides, 58 | batchnorm=batchnorm, 59 | dilation_rate=dilation_rate, 60 | recurrent=recurrent) 61 | res = conv3d_block(res, 62 | n_filters=n_filters, 63 | kernel_size=kernel_size, 64 | strides=1, 65 | batchnorm=batchnorm, 66 | dilation_rate=dilation_rate, 67 | recurrent=recurrent) 68 | 69 | shortcut = conv3d_block(input_tensor, 70 | n_filters=n_filters, 71 | kernel_size=1, 72 | strides=strides, 73 | batchnorm=batchnorm, 74 | dilation_rate=1) 75 | if batchnorm: 76 | shortcut = BatchNormalization()(shortcut) 77 | 78 | output = Add()([shortcut, res]) 79 | return output 80 | 81 | 82 | def inception_block(input_tensor, 83 | n_filters, 84 | kernel_size=3, 85 | strides=1, 86 | batchnorm=True, 87 | recurrent=1, 88 | layers=[]): 89 | 90 | # Inception-style convolutional block similar to InceptionNet 91 | # The first convolution follows the function arguments, while subsequent inception convolutions follow the parameters in 92 | # argument, layers 93 | 94 | # layers is a nested list containing the different secondary inceptions in the format of (kernel_size, dil_rate) 95 | 96 | # E.g => layers=[ [(3,1),(3,1)], [(5,1)], [(3,1),(3,2)] ] 97 | # This will implement 3 sets of secondary convolutions 98 | # Set 1 => 3x3 dil = 1 followed by another 3x3 dil = 1 99 | # Set 2 => 5x5 dil = 1 100 | # Set 3 => 3x3 dil = 1 followed by 3x3 dil = 2 101 | 102 | res = conv3d_block(input_tensor, 103 | n_filters=n_filters, 104 | kernel_size=kernel_size, 105 | strides=strides, 106 | batchnorm=batchnorm, 107 | dilation_rate=1, 108 | recurrent=recurrent) 109 | 110 | temp = [] 111 | for layer in layers: 112 | local_res = res 113 | for conv in layer: 114 | incep_kernel_size = conv[0] 115 | incep_dilation_rate = conv[1] 116 | local_res = conv3d_block(local_res, 117 | n_filters=n_filters, 118 | kernel_size=incep_kernel_size, 119 | strides=1, 120 | batchnorm=batchnorm, 121 | dilation_rate=incep_dilation_rate, 122 | recurrent=recurrent) 123 | temp.append(local_res) 124 | 125 | temp = concatenate(temp) 126 | res = conv3d_block(temp, 127 | n_filters=n_filters, 128 | kernel_size=1, 129 | strides=1, 130 | batchnorm=batchnorm, 131 | dilation_rate=1) 132 | 133 | shortcut = conv3d_block(input_tensor, 134 | n_filters=n_filters, 135 | kernel_size=1, 136 | strides=strides, 137 | batchnorm=batchnorm, 138 | dilation_rate=1) 139 | if batchnorm: 140 | shortcut = BatchNormalization()(shortcut) 141 | 142 | output = Add()([shortcut, res]) 143 | return output 144 | 145 | 146 | def transpose_block(input_tensor, 147 | skip_tensor, 148 | n_filters, 149 | kernel_size=3, 150 | strides=1, 151 | batchnorm=True, 152 | recurrent=1): 153 | 154 | # A wrapper of the Keras Conv3DTranspose block to serve as a building block for upsampling layers 155 | 156 | shape_x = K.int_shape(input_tensor) 157 | shape_xskip = K.int_shape(skip_tensor) 158 | 159 | conv = Conv3DTranspose(filters=n_filters, 160 | kernel_size=kernel_size, 161 | padding='same', 162 | strides=(shape_xskip[1] // shape_x[1], 163 | shape_xskip[2] // shape_x[2], 164 | shape_xskip[3] // shape_x[3]), 165 | kernel_initializer="he_normal")(input_tensor) 166 | conv = LeakyReLU(alpha=alpha)(conv) 167 | 168 | act = conv3d_block(conv, 169 | n_filters=n_filters, 170 | kernel_size=kernel_size, 171 | strides=1, 172 | batchnorm=batchnorm, 173 | dilation_rate=1, 174 | recurrent=recurrent) 175 | output = Concatenate(axis=4)([act, skip_tensor]) 176 | return output 177 | 178 | 179 | def expend_as(tensor, rep): 180 | 181 | # Anonymous lambda function to expand the specified axis by a factor of argument, rep. 182 | # If tensor has shape (512,512,N), lambda will return a tensor of shape (512,512,N*rep), if specified axis=2 183 | 184 | my_repeat = Lambda(lambda x, repnum: K.repeat_elements(x, repnum, axis=4), 185 | arguments={'repnum': rep})(tensor) 186 | return my_repeat 187 | 188 | 189 | def AttnGatingBlock(x, g, inter_shape): 190 | 191 | shape_x = K.int_shape(x) 192 | shape_g = K.int_shape(g) 193 | 194 | # Getting the gating signal to the same number of filters as the inter_shape 195 | phi_g = Conv3D(filters=inter_shape, 196 | kernel_size=1, 197 | strides=1, 198 | padding='same')(g) 199 | 200 | # Getting the x signal to the same shape as the gating signal 201 | theta_x = Conv3D(filters=inter_shape, 202 | kernel_size=3, 203 | strides=(shape_x[1] // shape_g[1], 204 | shape_x[2] // shape_g[2], 205 | shape_x[3] // shape_g[3]), 206 | padding='same')(x) 207 | 208 | # Element-wise addition of the gating and x signals 209 | add_xg = add([phi_g, theta_x]) 210 | add_xg = Activation('relu')(add_xg) 211 | 212 | # 1x1x1 convolution 213 | psi = Conv3D(filters=1, kernel_size=1, padding='same')(add_xg) 214 | psi = Activation('sigmoid')(psi) 215 | shape_sigmoid = K.int_shape(psi) 216 | 217 | # Upsampling psi back to the original dimensions of x signal 218 | upsample_sigmoid_xg = UpSampling3D( 219 | size=(shape_x[1] // shape_sigmoid[1], shape_x[2] // shape_sigmoid[2], 220 | shape_x[3] // shape_sigmoid[3]))(psi) 221 | 222 | # Expanding the filter axis to the number of filters in the original x signal 223 | upsample_sigmoid_xg = expend_as(upsample_sigmoid_xg, shape_x[4]) 224 | 225 | # Element-wise multiplication of attention coefficients back onto original x signal 226 | attn_coefficients = multiply([upsample_sigmoid_xg, x]) 227 | 228 | # Final 1x1x1 convolution to consolidate attention signal to original x dimensions 229 | output = Conv3D(filters=shape_x[4], 230 | kernel_size=1, 231 | strides=1, 232 | padding='same')(attn_coefficients) 233 | output = BatchNormalization()(output) 234 | return output 235 | 236 | 237 | def GatingSignal(input_tensor, batchnorm=True): 238 | 239 | # 1x1x1 convolution to consolidate gating signal into the required dimensions 240 | # Not required most of the time, unless another ReLU and batch_norm is required on gating signal 241 | 242 | shape = K.int_shape(input_tensor) 243 | conv = Conv3D(filters=shape[4], 244 | kernel_size=1, 245 | strides=1, 246 | padding="same", 247 | kernel_initializer="he_normal")(input_tensor) 248 | if batchnorm: 249 | conv = BatchNormalization()(conv) 250 | output = LeakyReLU(alpha=alpha)(conv) 251 | return output 252 | -------------------------------------------------------------------------------- /losses.py: -------------------------------------------------------------------------------- 1 | import tensorflow.keras.backend as K 2 | 3 | 4 | def dice_coef(y_true, y_pred, smooth=1): 5 | y_true_f = K.flatten(y_true) 6 | y_pred_f = K.flatten(y_pred) 7 | intersection = K.sum(y_true_f * y_pred_f) 8 | return (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + 9 | smooth) 10 | 11 | 12 | def dice_coef_loss(y_true, y_pred): 13 | return 1 - dice_coef(y_true, y_pred) 14 | 15 | 16 | def tversky(y_true, y_pred, smooth=1, alpha=0.7): 17 | y_true_pos = K.flatten(y_true) 18 | y_pred_pos = K.flatten(y_pred) 19 | true_pos = K.sum(y_true_pos * y_pred_pos) 20 | false_neg = K.sum(y_true_pos * (1 - y_pred_pos)) 21 | false_pos = K.sum((1 - y_true_pos) * y_pred_pos) 22 | return (true_pos + smooth) / (true_pos + alpha * false_neg + 23 | (1 - alpha) * false_pos + smooth) 24 | 25 | 26 | def tversky_loss(y_true, y_pred): 27 | return 1 - tversky(y_true, y_pred) 28 | 29 | 30 | def focal_tversky_loss(y_true, y_pred, gamma=0.75): 31 | tv = tversky(y_true, y_pred) 32 | return K.pow((1 - tv), gamma) 33 | -------------------------------------------------------------------------------- /modelmemory.py: -------------------------------------------------------------------------------- 1 | import tensorflow.keras.backend as K 2 | import numpy as np 3 | 4 | 5 | def memory_usage(batch_size, model): 6 | shapes_mem_count = 0 7 | for l in model.layers: 8 | single_layer_mem = 1 9 | for s in l.output_shape: 10 | if s is None: 11 | continue 12 | single_layer_mem *= s 13 | shapes_mem_count += single_layer_mem 14 | 15 | trainable_count = np.sum( 16 | [K.count_params(w) for w in model.trainable_weights]) 17 | non_trainable_count = np.sum( 18 | [K.count_params(w) for w in model.non_trainable_weights]) 19 | 20 | number_size = 4.0 21 | if K.floatx() == 'float16': 22 | number_size = 2.0 23 | if K.floatx() == 'float64': 24 | number_size = 8.0 25 | 26 | total_memory = number_size * (batch_size * shapes_mem_count + 27 | trainable_count + non_trainable_count) 28 | gbytes = np.round(total_memory / (1024.0**3), 3) 29 | return gbytes 30 | -------------------------------------------------------------------------------- /network.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.models import Model 2 | from tensorflow.keras.layers import Conv3D 3 | from layers3D import * 4 | 5 | # Use the functions provided in layers3D to build the network 6 | 7 | 8 | def network(input_img, n_filters=16, batchnorm=True): 9 | 10 | # contracting path 11 | 12 | c0 = inception_block(input_img, 13 | n_filters=n_filters, 14 | batchnorm=batchnorm, 15 | strides=1, 16 | recurrent=2, 17 | layers=[[(3, 1), (3, 1)], [(3, 2)]]) # 512x512x512 18 | 19 | c1 = inception_block(c0, 20 | n_filters=n_filters * 2, 21 | batchnorm=batchnorm, 22 | strides=2, 23 | recurrent=2, 24 | layers=[[(3, 1), (3, 1)], [(3, 2)]]) # 256x256x256 25 | 26 | c2 = inception_block(c1, 27 | n_filters=n_filters * 4, 28 | batchnorm=batchnorm, 29 | strides=2, 30 | recurrent=2, 31 | layers=[[(3, 1), (3, 1)], [(3, 2)]]) # 128x128x128 32 | 33 | c3 = inception_block(c2, 34 | n_filters=n_filters * 8, 35 | batchnorm=batchnorm, 36 | strides=2, 37 | recurrent=2, 38 | layers=[[(3, 1), (3, 1)], [(3, 2)]]) # 64x64x64 39 | 40 | # bridge 41 | 42 | b0 = inception_block(c3, 43 | n_filters=n_filters * 16, 44 | batchnorm=batchnorm, 45 | strides=2, 46 | recurrent=2, 47 | layers=[[(3, 1), (3, 1)], [(3, 2)]]) # 32x32x32 48 | 49 | # expansive path 50 | 51 | attn0 = AttnGatingBlock(c3, b0, n_filters * 16) 52 | u0 = transpose_block(b0, 53 | attn0, 54 | n_filters=n_filters * 8, 55 | batchnorm=batchnorm, 56 | recurrent=2) # 64x64x64 57 | 58 | attn1 = AttnGatingBlock(c2, u0, n_filters * 8) 59 | u1 = transpose_block(u0, 60 | attn1, 61 | n_filters=n_filters * 4, 62 | batchnorm=batchnorm, 63 | recurrent=2) # 128x128x128 64 | 65 | attn2 = AttnGatingBlock(c1, u1, n_filters * 4) 66 | u2 = transpose_block(u1, 67 | attn2, 68 | n_filters=n_filters * 2, 69 | batchnorm=batchnorm, 70 | recurrent=2) # 256x256x256 71 | 72 | u3 = transpose_block(u2, 73 | c0, 74 | n_filters=n_filters, 75 | batchnorm=batchnorm, 76 | recurrent=2) # 512x512x512 77 | 78 | outputs = Conv3D(filters=1, kernel_size=1, strides=1, 79 | activation='sigmoid')(u3) 80 | model = Model(inputs=[input_img], outputs=[outputs]) 81 | return model 82 | -------------------------------------------------------------------------------- /plotmetrics.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "code", 5 | "execution_count": null, 6 | "metadata": {}, 7 | "outputs": [], 8 | "source": [ 9 | "# yapf: disable\n", 10 | "import matplotlib.pyplot as plt\n", 11 | "import numpy as np\n", 12 | "%matplotlib inline\n", 13 | "\n", 14 | "epoch, metric, loss = np.loadtxt('', delimiter=',', unpack=True, skiprows=1)\n", 15 | "fig = plt.figure(figsize=(20, 10))\n", 16 | "ax = fig.add_axes([1, 1, 1, 1])\n", 17 | "#ax.set_yscale('log')\n", 18 | "ax.plot(epoch, metric)\n", 19 | "plt.show()" 20 | ] 21 | }, 22 | { 23 | "cell_type": "code", 24 | "execution_count": null, 25 | "metadata": {}, 26 | "outputs": [], 27 | "source": [] 28 | } 29 | ], 30 | "metadata": { 31 | "kernelspec": { 32 | "display_name": "Python 3", 33 | "language": "python", 34 | "name": "python3" 35 | }, 36 | "language_info": { 37 | "codemirror_mode": { 38 | "name": "ipython", 39 | "version": 3 40 | }, 41 | "file_extension": ".py", 42 | "mimetype": "text/x-python", 43 | "name": "python", 44 | "nbconvert_exporter": "python", 45 | "pygments_lexer": "ipython3", 46 | "version": "3.7.7" 47 | } 48 | }, 49 | "nbformat": 4, 50 | "nbformat_minor": 4 51 | } 52 | -------------------------------------------------------------------------------- /predict.py: -------------------------------------------------------------------------------- 1 | from tensorflow.keras.models import load_model 2 | from hyperparameters import * 3 | from datagenerator import DataGenerator 4 | import numpy as np 5 | from losses import * 6 | 7 | model = load_model(save_path) 8 | 9 | list_IDs = [] 10 | for filename in os.listdir(test_path): 11 | # Write logic to add filenames of train images to list_IDs which will be processed by DataGenerator 12 | # later on 13 | pass 14 | 15 | evaluate_gen = DataGenerator(list_IDs=list_IDs, 16 | dim=dimensions, 17 | batch_size=batch_size, 18 | shuffle=True) 19 | 20 | # Returns Numpy arrays of predictions 21 | model.predict_generator(evaluate_gen, steps=0, verbose=2, workers=20) 22 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import os 2 | import tensorflow as tf 3 | from tensorflow.keras.layers import Input 4 | from tensorflow.keras.utils import multi_gpu_model 5 | from tensorflow.keras.optimizers import Adam 6 | from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau, CSVLogger, TerminateOnNaN 7 | import tensorflow.keras.backend as K 8 | from hyperparameters import * 9 | from losses import * 10 | from network import network 11 | from datagenerator import DataGenerator 12 | from modelmemory import memory_usage 13 | 14 | ################################################################################ 15 | # Compiling 16 | ################################################################################ 17 | 18 | input_img = Input(input_dimensions) 19 | with tf.device('/cpu:0'): 20 | model = network(input_img, 21 | n_filters=num_initial_filters, 22 | batchnorm=batchnorm) 23 | 24 | if num_gpu > 1: 25 | parallel_model = multi_gpu_model(model, gpus=num_gpu, cpu_merge=False) 26 | parallel_model.compile(optimizer=Adam(lr=learning_rate), 27 | loss=loss, 28 | metrics=metrics) 29 | else: 30 | model.compile(optimizer=Adam(lr=learning_rate), loss=loss, metrics=metrics) 31 | 32 | callbacks = [ 33 | EarlyStopping(monitor='', patience=400, verbose=1), 34 | ReduceLROnPlateau(factor=0.1, 35 | monitor='', 36 | patience=50, 37 | min_lr=0.00001, 38 | verbose=1, 39 | mode='max'), 40 | ModelCheckpoint(checkpoint_path, 41 | monitor='', 42 | mode='max', 43 | verbose=0, 44 | save_best_only=True), 45 | CSVLogger(log_path, separator=',', append=True), 46 | TerminateOnNaN() 47 | ] 48 | 49 | # Prints a rough estimation of the GPU memory per GPU required to store the model 50 | print('Memory Footprint/GPU: ' + str(memory_usage(1, model)) + 'GB') 51 | 52 | ################################################################################ 53 | # Training 54 | ################################################################################ 55 | 56 | list_IDs = [] 57 | for filename in os.listdir(train_path): 58 | # Write logic to add filenames of train images to list_IDs which will be processed by DataGenerator 59 | # later on 60 | pass 61 | 62 | train_gen = DataGenerator(list_IDs=list_IDs, 63 | dim=dimensions, 64 | batch_size=batch_size, 65 | shuffle=True) 66 | 67 | if num_gpu > 1: 68 | parallel_model.fit_generator(train_gen, 69 | steps_per_epoch=steps_per_epoch, 70 | epochs=epochs, 71 | verbose=2, 72 | callbacks=callbacks, 73 | workers=20) 74 | else: 75 | model.fit_generator(train_gen, 76 | steps_per_epoch=steps_per_epoch, 77 | epochs=epochs, 78 | verbose=2, 79 | callbacks=callbacks, 80 | workers=20) 81 | 82 | model.save(save_path) 83 | --------------------------------------------------------------------------------