├── assets ├── GT_01.png ├── GT_02.png ├── GT_03.png ├── GT_10.png ├── GT_17.png ├── GT_45.png ├── IDRiD_01.jpg ├── IDRiD_02.jpg ├── IDRiD_03.jpg ├── IDRiD_10.jpg ├── IDRiD_17.jpg └── IDRiD_45.jpg ├── README.md ├── LICENSE └── IDRiDseg.ipynb /assets/GT_01.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_01.png -------------------------------------------------------------------------------- /assets/GT_02.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_02.png -------------------------------------------------------------------------------- /assets/GT_03.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_03.png -------------------------------------------------------------------------------- /assets/GT_10.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_10.png -------------------------------------------------------------------------------- /assets/GT_17.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_17.png -------------------------------------------------------------------------------- /assets/GT_45.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/GT_45.png -------------------------------------------------------------------------------- /assets/IDRiD_01.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_01.jpg -------------------------------------------------------------------------------- /assets/IDRiD_02.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_02.jpg -------------------------------------------------------------------------------- /assets/IDRiD_03.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_03.jpg -------------------------------------------------------------------------------- /assets/IDRiD_10.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_10.jpg -------------------------------------------------------------------------------- /assets/IDRiD_17.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_17.jpg -------------------------------------------------------------------------------- /assets/IDRiD_45.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/HimanshuAgarwal022/IDRiDSegmentation/HEAD/assets/IDRiD_45.jpg -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # IDRiDSegmentation 2 | 3 | Code for Segmentation of retinal lesions associated with diabetic retinopathy as microaneurysms, hemorrhages, hard exudates, and soft exudates. 4 | 5 | It's distributed under the GNU General Public License v3 (or any later 6 | version) - see the accompanying LICENSE file for more details. 7 | 8 | ## Requirements 9 | 10 | You need to have the following python packages installed (may be incomplete): 11 | 12 | - tensorflow 1.X 13 | - opencv 14 | - numpy 15 | - scipy 16 | - sklearn 17 | - matplotlib 18 | - PIL 19 | - pandas 20 | - You need to have downloaded at least the IDRiD training dataset. 21 | 22 | ## Samples 23 | 24 | | Ground Truth(HE) | Predicted | 25 | | :-------------------: | :----------------------: | 26 | | ![](assets/GT_01.png) | ![](assets/IDRiD_01.jpg) | 27 | | ![](assets/GT_02.png) | ![](assets/IDRiD_02.jpg) | 28 | | ![](assets/GT_03.png) | ![](assets/IDRiD_03.jpg) | 29 | | ![](assets/GT_10.png) | ![](assets/IDRiD_10.jpg) | 30 | | ![](assets/GT_17.png) | ![](assets/IDRiD_17.jpg) | 31 | | ![](assets/GT_45.png) | ![](assets/IDRiD_45.jpg) | 32 | *** 33 | 34 | ## Contributors 35 | 36 | - **[Himanshu Agarwal](https://github.com/HimanshuAgarwal022)** 37 | 38 | - **[Jayesh Narayan](https://github.com/jayesh1narayan)** 39 | 40 | *** 41 | 42 | ## Citations 43 | 44 | **[IDRiD Challenge](https://idrid.grand-challenge.org/)** 45 | *** 46 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /IDRiDseg.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "nbformat": 4, 3 | "nbformat_minor": 0, 4 | "metadata": { 5 | "colab": { 6 | "name": "IDRiDseg.ipynb", 7 | "provenance": [], 8 | "collapsed_sections": [ 9 | "AcSma7_RfLnk", 10 | "EeZa_92kuQag", 11 | "_Na2DjTeexJd", 12 | "yrfnYhrjuYYQ", 13 | "BklkTaFzut95", 14 | "cHjVOXKMfXSa", 15 | "EqLtpXRru2kh", 16 | "p2JGpoMrtNx3", 17 | "ByIpNhUk-xQr", 18 | "CM2NY0sRHysI", 19 | "rnNDIjEh_iD3", 20 | "mjewf0kzHheN", 21 | "8hwL8cLfH89I" 22 | ], 23 | "toc_visible": true, 24 | "include_colab_link": true 25 | }, 26 | "kernelspec": { 27 | "name": "python3", 28 | "display_name": "Python 3" 29 | }, 30 | "accelerator": "GPU" 31 | }, 32 | "cells": [ 33 | { 34 | "cell_type": "markdown", 35 | "metadata": { 36 | "id": "view-in-github", 37 | "colab_type": "text" 38 | }, 39 | "source": [ 40 | "\"Open" 41 | ] 42 | }, 43 | { 44 | "cell_type": "markdown", 45 | "metadata": { 46 | "id": "PM9UH-6K18L8" 47 | }, 48 | "source": [ 49 | "# Diabetic Retinopathy Lesion Segmentation.\n", 50 | "---\n", 51 | "\n", 52 | " GNU GENERAL PUBLIC LICENSE\n", 53 | "\n", 54 | " Version 3, 29 June 2007\n", 55 | "\n", 56 | " Copyright (c) [2020] [Himanshu Agarwal]\n", 57 | "\n", 58 | "\n", 59 | "---\n", 60 | "\n" 61 | ] 62 | }, 63 | { 64 | "cell_type": "markdown", 65 | "metadata": { 66 | "id": "AcSma7_RfLnk" 67 | }, 68 | "source": [ 69 | "### load drive" 70 | ] 71 | }, 72 | { 73 | "cell_type": "code", 74 | "metadata": { 75 | "id": "Hf1EOC0CtWWC", 76 | "colab": { 77 | "base_uri": "https://localhost:8080/", 78 | "height": 122 79 | }, 80 | "outputId": "9e20fb59-1936-4b7c-9a3c-3981b8edd276" 81 | }, 82 | "source": [ 83 | "from google.colab import drive\n", 84 | "drive.mount('/content/drive')" 85 | ], 86 | "execution_count": null, 87 | "outputs": [ 88 | { 89 | "output_type": "stream", 90 | "text": [ 91 | "Go to this URL in a browser: https://accounts.google.com/o/oauth2/auth?client_id=947318989803-6bn6qk8qdgf4n4g3pfee6491hc0brc4i.apps.googleusercontent.com&redirect_uri=urn%3aietf%3awg%3aoauth%3a2.0%3aoob&response_type=code&scope=email%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdocs.test%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive.photos.readonly%20https%3a%2f%2fwww.googleapis.com%2fauth%2fpeopleapi.readonly\n", 92 | "\n", 93 | "Enter your authorization code:\n", 94 | "··········\n", 95 | "Mounted at /content/drive\n" 96 | ], 97 | "name": "stdout" 98 | } 99 | ] 100 | }, 101 | { 102 | "cell_type": "markdown", 103 | "metadata": { 104 | "id": "EeZa_92kuQag" 105 | }, 106 | "source": [ 107 | "### convert ground truths to binary" 108 | ] 109 | }, 110 | { 111 | "cell_type": "code", 112 | "metadata": { 113 | "id": "Oi6Fax6vuOoL" 114 | }, 115 | "source": [ 116 | "from PIL import Image\n", 117 | "#from resizeimage import resizeimage\n", 118 | "import os, sys\n", 119 | "\n", 120 | "def cmp(a, b):\n", 121 | " return (a > b) - (a < b) \n", 122 | "\n", 123 | "def resizeImage(infile,file, output_dir, size=(4288,2848)):\n", 124 | " outfile = os.path.splitext(file)[0]\n", 125 | " extension = os.path.splitext(file)[1]\n", 126 | " #print(outfile)\n", 127 | " #print(extension)\n", 128 | " #print(infile)\n", 129 | " #if (cmp(extension, \".jpg\")):\n", 130 | " #print(\"dsd\")\n", 131 | " #return\n", 132 | "\n", 133 | " if infile != outfile:\n", 134 | " try :\n", 135 | " im = Image.open(infile)\n", 136 | " gray = im.convert('L')\n", 137 | " bw = gray.point(lambda x: 0 if x<50 else 255, '1')\n", 138 | " # im = resizeimage.resize_cover(im, [960, 640])\n", 139 | " bw.save(output_dir+outfile[:-3]+extension,\"TIFF\",quality=100)\n", 140 | " #print(\"sucess\")\n", 141 | " #except IOError:\n", 142 | " # print (\"cannot reduce image for \", infile)\n", 143 | " except e:\n", 144 | " print (e)\n", 145 | " \n", 146 | "\n", 147 | "\n", 148 | "output_dir = \"drive/My Drive/data/output/annotations/\"\n", 149 | "annot_dir = \"drive/My Drive/data/annotations/\"\n", 150 | "dir = os.getcwd()\n", 151 | "\n", 152 | "if not os.path.exists(os.path.join(dir,output_dir)):\n", 153 | " os.mkdir(output_dir)\n", 154 | "annot = os.path.join(dir,annot_dir)\n", 155 | "for file in os.listdir(annot):\n", 156 | " #print(file)\n", 157 | " resizeImage(os.path.join(annot,file),file,output_dir)" 158 | ], 159 | "execution_count": null, 160 | "outputs": [] 161 | }, 162 | { 163 | "cell_type": "markdown", 164 | "metadata": { 165 | "id": "_Na2DjTeexJd" 166 | }, 167 | "source": [ 168 | "### CLAHE on input images" 169 | ] 170 | }, 171 | { 172 | "cell_type": "code", 173 | "metadata": { 174 | "id": "uIIONbM_e4Jz" 175 | }, 176 | "source": [ 177 | "import os, sys\n", 178 | "import numpy as np\n", 179 | "import cv2\n", 180 | "from matplotlib import pyplot as plt\n", 181 | "\n", 182 | "dir = os.getcwd()\n", 183 | "dir_data = os.path.join(dir,\"drive/My Drive/data/output/preimages/\")\n", 184 | "output_dir_data = os.path.join(dir,\"drive/My Drive/data/output/images/\")\n", 185 | "if not os.path.exists(os.path.join(dir,output_dir_data)):\n", 186 | " os.mkdir(output_dir_data)\n", 187 | "\n", 188 | "gridsize = 8\n", 189 | "for file in os.listdir(dir_data):\n", 190 | " bgr = cv2.imread(os.path.join(dir_data,file))\n", 191 | " lab = cv2.cvtColor(bgr, cv2.COLOR_BGR2LAB)\n", 192 | " lab_planes = cv2.split(lab)\n", 193 | " clahe = cv2.createCLAHE(clipLimit=2.0,tileGridSize=(gridsize,gridsize))\n", 194 | " lab_planes[0] = clahe.apply(lab_planes[0])\n", 195 | " lab = cv2.merge(lab_planes)\n", 196 | " bgr = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)\n", 197 | " #if random.uniform(0, 1.0) > 0.75:\n", 198 | " # bgr = cv2.bitwise_not(bgr)\n", 199 | " # plt.imshow(bgr)\n", 200 | " # plt.show()\n", 201 | " cv2.imwrite(os.path.join(output_dir_data,file),bgr)" 202 | ], 203 | "execution_count": null, 204 | "outputs": [] 205 | }, 206 | { 207 | "cell_type": "code", 208 | "metadata": { 209 | "id": "GL-rKiA5hn_T" 210 | }, 211 | "source": [ 212 | "#output_dir_data = os.path.join(dir,\"drive/My Drive/data/output/images/\")\n", 213 | "#for file in os.listdir(output_dir_data):\n", 214 | "# bgr = cv2.imread(os.path.join(output_dir_data,file))\n", 215 | "# rgb = cv2.cvtColor(bgr, cv2.COLOR_BGR2LAB)\n", 216 | "# plt.imshow(rgb)\n", 217 | "# plt.show()\n", 218 | "# bgr = cv2.bitwise_not(bgr)\n", 219 | "# plt.imshow(bgr)\n", 220 | "# plt.show()" 221 | ], 222 | "execution_count": null, 223 | "outputs": [] 224 | }, 225 | { 226 | "cell_type": "markdown", 227 | "metadata": { 228 | "id": "yrfnYhrjuYYQ" 229 | }, 230 | "source": [ 231 | "### extract patches" 232 | ] 233 | }, 234 | { 235 | "cell_type": "code", 236 | "metadata": { 237 | "id": "0BhM3OxTuiWu" 238 | }, 239 | "source": [ 240 | "from PIL import Image\n", 241 | "#from resizeimage import resizeimage\n", 242 | "import os, sys\n", 243 | "import numpy as np\n", 244 | "import cv2\n", 245 | "from matplotlib import pyplot as plt\n", 246 | "\n", 247 | "dir = os.getcwd()\n", 248 | "output_dir_data = \"drive/My Drive/data/output/patches/\"\n", 249 | "output_dir_mask = \"drive/My Drive/data/output/labels/\"\n", 250 | "if not os.path.exists(os.path.join(dir,output_dir_data)):\n", 251 | " os.mkdir(output_dir_data)\n", 252 | "if not os.path.exists(os.path.join(dir,output_dir_mask)):\n", 253 | " os.mkdir(output_dir_mask)\n", 254 | "\n", 255 | "dir_data = os.path.join(dir,\"drive/My Drive/data/output/images/\")\n", 256 | "dir_mask = os.path.join(dir,\"drive/My Drive/data/output/annotations/\")\n", 257 | "\n", 258 | "# im = Image.open(os.path.join(dir_mask,\"IDRiD_06.tif\"))\n", 259 | "# im_crop = im.crop((2000,0,2000+512,0+256))\n", 260 | "# im_crop.show()\n", 261 | "# image_np = np.array(im_crop)\n", 262 | "# print np.sum(image_np)\n", 263 | "\n", 264 | "negative_patches = []\n", 265 | "positive_count = 0\n", 266 | "\n", 267 | "for file in os.listdir(dir_mask):\n", 268 | " outfile = os.path.splitext(file)[0]\n", 269 | " extension = os.path.splitext(file)[1]\n", 270 | " #if (cmp(extension, \".jpg\")):\n", 271 | " # continue\n", 272 | " img = outfile + \".jpg\"\n", 273 | " im = Image.open(os.path.join(dir_mask,file))\n", 274 | " imd = Image.open(os.path.join(dir_data,img))\n", 275 | " # image_np = np.array(im)\n", 276 | " # print np.sum([True, True])\n", 277 | " # im_crop = im.crop((1900,0,1900+512,0+512))\n", 278 | " patch_id = 0\n", 279 | " for i in range(10): #10 6\n", 280 | " \tfor j in range(16): #16 9\n", 281 | " top_y = i*256 #256 512\n", 282 | " if (i==9): #9 5\n", 283 | " top_y = 2336\n", 284 | " top_x = j*256 #256 512\n", 285 | " if (j==15): #15 8\n", 286 | " top_x = 3776\n", 287 | "\n", 288 | " im_crop = im.crop((top_x,top_y,top_x+512,top_y+512))\n", 289 | " imd_crop = imd.crop((top_x,top_y,top_x+512,top_y+512))\n", 290 | " im_crop.save(output_dir_mask+outfile+\"_p\"+str(patch_id)+extension,\"JPEG\",quality=100)\n", 291 | " imd_crop.save(output_dir_data+outfile+\"_p\"+str(patch_id)+extension,\"JPEG\",quality=100)\n", 292 | " if (np.sum(np.array(im_crop)) < 100):\n", 293 | " negative_patches.append(output_dir_mask+outfile+\"_p\"+str(patch_id)+extension)\n", 294 | " else:\n", 295 | " positive_count += 1\n", 296 | "\n", 297 | " patch_id += 1\n", 298 | "\n", 299 | "negative_patches = np.array(negative_patches)\n", 300 | "# np.savetxt(\"negative.csv\", negative_patches, delimiter=\",\", fmt=\"%s\")\n", 301 | "\n", 302 | "negative_count = negative_patches.size\n", 303 | "delete_count = negative_count - 4*positive_count\n", 304 | "np.random.shuffle(negative_patches)\n", 305 | "split_idx = delete_count\n", 306 | "delete_patches = negative_patches[:split_idx]\n", 307 | "\n", 308 | "for idx in range(delete_patches.size):\n", 309 | " os.remove(delete_patches[idx])\n", 310 | " os.remove(os.path.join(output_dir_data,delete_patches[idx][34:]))" 311 | ], 312 | "execution_count": null, 313 | "outputs": [] 314 | }, 315 | { 316 | "cell_type": "markdown", 317 | "metadata": { 318 | "id": "BklkTaFzut95" 319 | }, 320 | "source": [ 321 | "### split in train and test sets" 322 | ] 323 | }, 324 | { 325 | "cell_type": "code", 326 | "metadata": { 327 | "id": "Krh5Z4OHuvBu", 328 | "colab": { 329 | "base_uri": "https://localhost:8080/", 330 | "height": 51 331 | }, 332 | "outputId": "72c55ae1-8c28-446d-eade-1335338daa7e" 333 | }, 334 | "source": [ 335 | "import numpy as np\n", 336 | "import matplotlib.pyplot as plt\n", 337 | "import pandas as pd\n", 338 | "import os\n", 339 | "\n", 340 | "dir = \"drive/My Drive/data/output/patches/\"\n", 341 | "#dir = \"drive/My Drive/data/output/images/\"\n", 342 | "image_paths = os.listdir(dir)\n", 343 | "\n", 344 | "length = len(image_paths)\n", 345 | "img_paths = np.empty(length, dtype=object)\n", 346 | "\n", 347 | "i=0\n", 348 | "for file in image_paths:\n", 349 | "\timg_paths[i] = \"drive/My Drive/data/output/patches/\" + file\n", 350 | "\t#img_paths[i] = \"drive/My Drive/data/output/images/\" + file\n", 351 | "\t#print (img_paths[i])\n", 352 | "\ti+=1\n", 353 | "\n", 354 | "# print (img_paths)\n", 355 | "np.random.shuffle(img_paths)\n", 356 | "#split_idx = int(img_paths.shape[0] * 1)\n", 357 | "split_idx = 11534\n", 358 | "train_paths = img_paths[:split_idx]\n", 359 | "test_paths = img_paths[split_idx:]\n", 360 | "\n", 361 | "train_paths_ = np.copy(train_paths)\n", 362 | "test_paths_ = np.copy(test_paths)\n", 363 | "print(\"train set: \",train_paths.size)\n", 364 | "print(\"test set: \",test_paths.size)\n", 365 | "for i in range(train_paths.size):\n", 366 | "\ttrain_paths_[i] = train_paths[i][35:]\n", 367 | "\t#train_paths_[i] = train_paths[i][34:]\n", 368 | "\ttrain_paths_[i] = \"drive/My Drive/data/output/labels/\" + train_paths_[i]\n", 369 | "\t#train_paths_[i] = \"drive/My Drive/data/output/annotations/\" + train_paths_[i]\n", 370 | "\t#print (train_paths_[i])\n", 371 | "\n", 372 | "#print (\"split\")\n", 373 | "\n", 374 | "for i in range(test_paths.size):\n", 375 | "\ttest_paths_[i] = test_paths[i][35:]\n", 376 | "\t#test_paths_[i] = test_paths[i][34:]\n", 377 | "\ttest_paths_[i] = \"drive/My Drive/data/output/labels/\" + test_paths_[i]\n", 378 | "\t#test_paths_[i] = \"drive/My Drive/data/output/annotations/\" + test_paths_[i]\n", 379 | "\t#print (test_paths_[i])\n", 380 | "\n", 381 | "train_csv = np.stack((train_paths,train_paths_), axis=1)\n", 382 | "test_csv = np.stack((test_paths,test_paths_), axis=1)\n", 383 | "\n", 384 | "np.savetxt(\"train.csv\", train_csv, delimiter=\",\", fmt=\"%s\")\n", 385 | "np.savetxt(\"test.csv\", test_csv, delimiter=\",\", fmt=\"%s\")" 386 | ], 387 | "execution_count": null, 388 | "outputs": [ 389 | { 390 | "output_type": "stream", 391 | "text": [ 392 | "train set: 11534\n", 393 | "test set: 8\n" 394 | ], 395 | "name": "stdout" 396 | } 397 | ] 398 | }, 399 | { 400 | "cell_type": "markdown", 401 | "metadata": { 402 | "id": "cHjVOXKMfXSa" 403 | }, 404 | "source": [ 405 | "### load tensorflow" 406 | ] 407 | }, 408 | { 409 | "cell_type": "code", 410 | "metadata": { 411 | "id": "-YpGHNnLuAt9", 412 | "colab": { 413 | "base_uri": "https://localhost:8080/", 414 | "height": 68 415 | }, 416 | "outputId": "dcd8132f-f5ae-45f0-fcdd-fe89f1910423" 417 | }, 418 | "source": [ 419 | "#import os\n", 420 | "#os.getcwd()\n", 421 | "#os.listdir()\n", 422 | "#os.path.exists('drive/My Drive/')\n", 423 | "#for roots,dirs,files in os.walk('drive/My Drive'): \n", 424 | "# print(roots,dirs,files)\n", 425 | "%tensorflow_version 1.x\n", 426 | "import tensorflow as tf\n", 427 | "device_name = tf.test.gpu_device_name()\n", 428 | "print('TensorFlow Version: {}'.format(tf.__version__))\n", 429 | "if device_name != '/device:GPU:0':\n", 430 | " raise SystemError('GPU device not found')\n", 431 | "print('Found GPU at: {}'.format(device_name))" 432 | ], 433 | "execution_count": null, 434 | "outputs": [ 435 | { 436 | "output_type": "stream", 437 | "text": [ 438 | "TensorFlow 1.x selected.\n", 439 | "TensorFlow Version: 1.15.2\n", 440 | "Found GPU at: /device:GPU:0\n" 441 | ], 442 | "name": "stdout" 443 | } 444 | ] 445 | }, 446 | { 447 | "cell_type": "code", 448 | "metadata": { 449 | "id": "sdVEq0W1we7T" 450 | }, 451 | "source": [ 452 | "from tensorflow.python.client import device_lib\n", 453 | "device_lib.list_local_devices()" 454 | ], 455 | "execution_count": null, 456 | "outputs": [] 457 | }, 458 | { 459 | "cell_type": "code", 460 | "metadata": { 461 | "id": "iJis0dgqxwGB", 462 | "colab": { 463 | "base_uri": "https://localhost:8080/", 464 | "height": 258 465 | }, 466 | "outputId": "2005613f-bf1e-44c7-c8b0-0c61609123f7" 467 | }, 468 | "source": [ 469 | "# memory footprint support libraries/code\n", 470 | "!ln -sf /opt/bin/nvidia-smi /usr/bin/nvidia-smi\n", 471 | "!pip install gputil\n", 472 | "!pip install psutil\n", 473 | "!pip install humanize\n", 474 | "import psutil\n", 475 | "import humanize\n", 476 | "import os\n", 477 | "import GPUtil as GPU\n", 478 | "GPUs = GPU.getGPUs()\n", 479 | "# XXX: only one GPU on Colab and isn’t guaranteed\n", 480 | "gpu = GPUs[0]\n", 481 | "def printm():\n", 482 | " process = psutil.Process(os.getpid())\n", 483 | " print(\"Gen RAM Free: \" + humanize.naturalsize( psutil.virtual_memory().available ), \" | Proc size: \" + humanize.naturalsize( process.memory_info().rss))\n", 484 | " print(\"GPU RAM Free: {0:.0f}MB | Used: {1:.0f}MB | Util {2:3.0f}% | Total {3:.0f}MB\".format(gpu.memoryFree, gpu.memoryUsed, gpu.memoryUtil*100, gpu.memoryTotal))\n", 485 | "printm() " 486 | ], 487 | "execution_count": null, 488 | "outputs": [ 489 | { 490 | "output_type": "stream", 491 | "text": [ 492 | "Collecting gputil\n", 493 | " Downloading https://files.pythonhosted.org/packages/ed/0e/5c61eedde9f6c87713e89d794f01e378cfd9565847d4576fa627d758c554/GPUtil-1.4.0.tar.gz\n", 494 | "Building wheels for collected packages: gputil\n", 495 | " Building wheel for gputil (setup.py) ... \u001b[?25l\u001b[?25hdone\n", 496 | " Created wheel for gputil: filename=GPUtil-1.4.0-cp36-none-any.whl size=7413 sha256=ee2b3d234e82bb1c13f028526f9f680f5971df1f7db91a5bc6c48c72b186221a\n", 497 | " Stored in directory: /root/.cache/pip/wheels/3d/77/07/80562de4bb0786e5ea186911a2c831fdd0018bda69beab71fd\n", 498 | "Successfully built gputil\n", 499 | "Installing collected packages: gputil\n", 500 | "Successfully installed gputil-1.4.0\n", 501 | "Requirement already satisfied: psutil in /usr/local/lib/python3.6/dist-packages (5.4.8)\n", 502 | "Requirement already satisfied: humanize in /usr/local/lib/python3.6/dist-packages (0.5.1)\n", 503 | "Gen RAM Free: 12.4 GB | Proc size: 557.2 MB\n", 504 | "GPU RAM Free: 16015MB | Used: 265MB | Util 2% | Total 16280MB\n" 505 | ], 506 | "name": "stdout" 507 | } 508 | ] 509 | }, 510 | { 511 | "cell_type": "markdown", 512 | "metadata": { 513 | "id": "EqLtpXRru2kh" 514 | }, 515 | "source": [ 516 | "### Train" 517 | ] 518 | }, 519 | { 520 | "cell_type": "code", 521 | "metadata": { 522 | "id": "SIVvZ3JxsepC" 523 | }, 524 | "source": [ 525 | "#!rm 'drive/My Drive/data/models' -rf\n", 526 | "#!rm 'drive/My Drive/data/logs' -rf" 527 | ], 528 | "execution_count": null, 529 | "outputs": [] 530 | }, 531 | { 532 | "cell_type": "code", 533 | "metadata": { 534 | "id": "9jHwMuhTxgBi", 535 | "colab": { 536 | "base_uri": "https://localhost:8080/", 537 | "height": 272 538 | }, 539 | "outputId": "43824013-f102-4ad3-9f4f-c9d73270042e" 540 | }, 541 | "source": [ 542 | "import time\n", 543 | "import os\n", 544 | "import pandas as pd\n", 545 | "import tensorflow as tf\n", 546 | "\n", 547 | "\n", 548 | "def image_augmentation(image, mask):\n", 549 | " \"\"\"Returns (maybe) augmented images\n", 550 | " (1) Random flip (left <--> right)\n", 551 | " (2) Random flip (up <--> down)\n", 552 | " (3) Random brightness\n", 553 | " (4) Random hue\n", 554 | " Args:\n", 555 | " image (3-D Tensor): Image tensor of (H, W, C)\n", 556 | " mask (3-D Tensor): Mask image tensor of (H, W, 1)\n", 557 | " Returns:\n", 558 | " image: Maybe augmented image (same shape as input `image`)\n", 559 | " mask: Maybe augmented mask (same shape as input `mask`)\n", 560 | " \"\"\"\n", 561 | " concat_image = tf.concat([image, mask], axis=-1)\n", 562 | "\n", 563 | " maybe_flipped = tf.image.random_flip_left_right(concat_image)\n", 564 | " maybe_flipped = tf.image.random_flip_up_down(concat_image)\n", 565 | "\n", 566 | " image = maybe_flipped[:, :, :-1]\n", 567 | " mask = maybe_flipped[:, :, -1:]\n", 568 | "\n", 569 | " image = tf.image.random_brightness(image, 0.7)\n", 570 | " image = tf.image.random_hue(image, 0.3)\n", 571 | "\n", 572 | " return image, mask\n", 573 | "\n", 574 | "\n", 575 | "def get_image_mask(queue, augmentation=True):\n", 576 | " \"\"\"Returns `image` and `mask`\n", 577 | " Input pipeline:\n", 578 | " Queue -> CSV -> FileRead -> Decode JPEG\n", 579 | " (1) Queue contains a CSV filename\n", 580 | " (2) Text Reader opens the CSV\n", 581 | " CSV file contains two columns\n", 582 | " [\"path/to/image.jpg\", \"path/to/mask.jpg\"]\n", 583 | " (3) File Reader opens both files\n", 584 | " (4) Decode JPEG to tensors\n", 585 | " Notes:\n", 586 | " height, width = 640, 960\n", 587 | " Returns\n", 588 | " image (3-D Tensor): (640, 960, 3)\n", 589 | " mask (3-D Tensor): (640, 960, 1)\n", 590 | " \"\"\"\n", 591 | " text_reader = tf.TextLineReader(skip_header_lines=1)\n", 592 | " _, csv_content = text_reader.read(queue)\n", 593 | "\n", 594 | " image_path, mask_path = tf.decode_csv(\n", 595 | " csv_content, record_defaults=[[\"\"], [\"\"]])\n", 596 | "\n", 597 | " image_file = tf.read_file(image_path)\n", 598 | " mask_file = tf.read_file(mask_path)\n", 599 | "\n", 600 | " image = tf.image.decode_jpeg(image_file, channels=3)\n", 601 | " image.set_shape([512, 512, 3])\n", 602 | " image = tf.cast(image, tf.float32)\n", 603 | "\n", 604 | " mask = tf.image.decode_jpeg(mask_file, channels=1)\n", 605 | " mask.set_shape([512, 512, 1])\n", 606 | " mask = tf.cast(mask, tf.float32)\n", 607 | " mask = mask / (tf.reduce_max(mask) + 1e-7)\n", 608 | "\n", 609 | " if augmentation:\n", 610 | " image, mask = image_augmentation(image, mask)\n", 611 | "\n", 612 | " return image, mask\n", 613 | "\n", 614 | "\n", 615 | "def conv_conv_pool(input_,\n", 616 | " n_filters,\n", 617 | " training,\n", 618 | " flags,\n", 619 | " name,\n", 620 | " pool=True,\n", 621 | " activation=tf.nn.relu):\n", 622 | " \"\"\"{Conv -> BN -> RELU}x2 -> {Pool, optional}\n", 623 | " Args:\n", 624 | " input_ (4-D Tensor): (batch_size, H, W, C)\n", 625 | " n_filters (list): number of filters [int, int]\n", 626 | " training (1-D Tensor): Boolean Tensor\n", 627 | " name (str): name postfix\n", 628 | " pool (bool): If True, MaxPool2D\n", 629 | " activation: Activaion functions\n", 630 | " Returns:\n", 631 | " net: output of the Convolution operations\n", 632 | " pool (optional): output of the max pooling operations\n", 633 | " \"\"\"\n", 634 | " net = input_\n", 635 | "\n", 636 | " with tf.variable_scope(\"layer{}\".format(name)):\n", 637 | " for i, F in enumerate(n_filters):\n", 638 | " net = tf.layers.conv2d(\n", 639 | " net,\n", 640 | " F, (3, 3),\n", 641 | " activation=None,\n", 642 | " padding='same',\n", 643 | " kernel_regularizer=tf.contrib.layers.l2_regularizer(flags.reg),\n", 644 | " name=\"conv_{}\".format(i + 1))\n", 645 | " net = tf.layers.batch_normalization(\n", 646 | " net, training=training, name=\"bn_{}\".format(i + 1))\n", 647 | " net = activation(net, name=\"relu{}_{}\".format(name, i + 1))\n", 648 | "\n", 649 | " if pool is False:\n", 650 | " return net\n", 651 | "\n", 652 | " pool = tf.layers.max_pooling2d(\n", 653 | " net, (2, 2), strides=(2, 2), name=\"pool_{}\".format(name))\n", 654 | "\n", 655 | " return net, pool\n", 656 | "\n", 657 | "\n", 658 | "def upconv_concat(inputA, input_B, n_filter, flags, name):\n", 659 | " \"\"\"Upsample `inputA` and concat with `input_B`\n", 660 | " Args:\n", 661 | " input_A (4-D Tensor): (N, H, W, C)\n", 662 | " input_B (4-D Tensor): (N, 2*H, 2*H, C2)\n", 663 | " name (str): name of the concat operation\n", 664 | " Returns:\n", 665 | " output (4-D Tensor): (N, 2*H, 2*W, C + C2)\n", 666 | " \"\"\"\n", 667 | " up_conv = upconv_2D(inputA, n_filter, flags, name)\n", 668 | "\n", 669 | " return tf.concat(\n", 670 | " [up_conv, input_B], axis=-1, name=\"concat_{}\".format(name))\n", 671 | "\n", 672 | "\n", 673 | "def upconv_2D(tensor, n_filter, flags, name):\n", 674 | " \"\"\"Up Convolution `tensor` by 2 times\n", 675 | " Args:\n", 676 | " tensor (4-D Tensor): (N, H, W, C)\n", 677 | " n_filter (int): Filter Size\n", 678 | " name (str): name of upsampling operations\n", 679 | " Returns:\n", 680 | " output (4-D Tensor): (N, 2 * H, 2 * W, C)\n", 681 | " \"\"\"\n", 682 | "\n", 683 | " return tf.layers.conv2d_transpose(\n", 684 | " tensor,\n", 685 | " filters=n_filter,\n", 686 | " kernel_size=2,\n", 687 | " strides=2,\n", 688 | " kernel_regularizer=tf.contrib.layers.l2_regularizer(flags.reg),\n", 689 | " name=\"upsample_{}\".format(name))\n", 690 | "\n", 691 | "\n", 692 | "def make_unet(X, training, flags=None):\n", 693 | " \"\"\"Build a U-Net architecture\n", 694 | " Args:\n", 695 | " X (4-D Tensor): (N, H, W, C)\n", 696 | " training (1-D Tensor): Boolean Tensor is required for batchnormalization layers\n", 697 | " Returns:\n", 698 | " output (4-D Tensor): (N, H, W, C)\n", 699 | " Same shape as the `input` tensor\n", 700 | " Notes:\n", 701 | " U-Net: Convolutional Networks for Biomedical Image Segmentation\n", 702 | " https://arxiv.org/abs/1505.04597\n", 703 | " \"\"\"\n", 704 | " net = X / 127.5 - 1\n", 705 | " conv1, pool1 = conv_conv_pool(net, [16, 16], training, flags, name=1)\n", 706 | " conv2, pool2 = conv_conv_pool(pool1, [32, 32], training, flags, name=2)\n", 707 | " conv3, pool3 = conv_conv_pool(pool2, [64, 64], training, flags, name=3)\n", 708 | " conv4, pool4 = conv_conv_pool(pool3, [128, 128], training, flags, name=4)\n", 709 | " conv5, pool5 = conv_conv_pool(pool4, [256, 256], training, flags, name=5)\n", 710 | " conv6, pool6 = conv_conv_pool(pool5, [512, 512], training, flags, name=6)\n", 711 | " conv7 = conv_conv_pool(\n", 712 | " pool6, [1024, 1024], training, flags, name=7, pool=False)\n", 713 | "\n", 714 | " up8 = upconv_concat(conv7, conv6, 512, flags, name=8)\n", 715 | " conv8 = conv_conv_pool(up8, [512, 512], training, flags, name=8, pool=False)\n", 716 | "\n", 717 | " up9 = upconv_concat(conv8, conv5, 256, flags, name=9)\n", 718 | " conv9 = conv_conv_pool(up9, [256, 256], training, flags, name=9, pool=False)\n", 719 | "\n", 720 | " up10 = upconv_concat(conv9, conv4, 128, flags, name=10)\n", 721 | " conv10 = conv_conv_pool(up10, [128, 128], training, flags, name=10, pool=False)\n", 722 | "\n", 723 | " up11 = upconv_concat(conv10, conv3, 64, flags, name=11)\n", 724 | " conv11 = conv_conv_pool(up11, [64, 64], training, flags, name=11, pool=False)\n", 725 | "\n", 726 | " up12 = upconv_concat(conv11, conv2, 32, flags, name=12)\n", 727 | " conv12 = conv_conv_pool(up12, [32, 32], training, flags, name=12, pool=False)\n", 728 | "\n", 729 | " up13 = upconv_concat(conv12, conv1, 16, flags, name=13)\n", 730 | " conv13 = conv_conv_pool(up13, [16, 16], training, flags, name=13, pool=False)\n", 731 | "\n", 732 | " # return tf.layers.conv2d(\n", 733 | " # conv13,\n", 734 | " # 1, (1, 1),\n", 735 | " # name='final',\n", 736 | " # activation=tf.nn.sigmoid,\n", 737 | " # padding='same')\n", 738 | " return tf.layers.conv2d(conv13,1, (1, 1),name='final',activation=None,padding='same')\n", 739 | "\n", 740 | "def BCE_(y_pred, y_true):\n", 741 | " # weight ratio = 9:1\n", 742 | " # 9-1=8\n", 743 | " class_weights = tf.constant([8],dtype=tf.float32)\n", 744 | " tensor_one = tf.constant([1],dtype=tf.float32)\n", 745 | "\n", 746 | " pred_flat = tf.reshape(y_pred, [-1, 1])\n", 747 | " true_flat = tf.reshape(y_true, [-1, 1])\n", 748 | "\n", 749 | " weight_map = tf.multiply(true_flat, class_weights)\n", 750 | " weight_map = tf.add(weight_map, tensor_one)\n", 751 | "\n", 752 | " loss_map = tf.nn.sigmoid_cross_entropy_with_logits(logits=pred_flat, labels=true_flat)\n", 753 | " loss_map = tf.multiply(loss_map, weight_map)\n", 754 | " loss = tf.reduce_mean(loss_map)\n", 755 | " return loss\n", 756 | "\n", 757 | "def IOU_(y_pred, y_true):\n", 758 | " \"\"\"Returns a (approx) IOU score\n", 759 | " intesection = y_pred.flatten() * y_true.flatten()\n", 760 | " Then, IOU = 2 * intersection / (y_pred.sum() + y_true.sum() + 1e-7) + 1e-7\n", 761 | " Args:\n", 762 | " y_pred (4-D array): (N, H, W, 1)\n", 763 | " y_true (4-D array): (N, H, W, 1)\n", 764 | " Returns:\n", 765 | " float: IOU score\n", 766 | " \"\"\"\n", 767 | " H, W, _ = y_pred.get_shape().as_list()[1:]\n", 768 | " threshold = 0.7\n", 769 | " pred_flat = tf.reshape(y_pred, [-1, H * W])\n", 770 | " true_flat = tf.reshape(y_true, [-1, H * W])\n", 771 | " pred = tf.cast(pred_flat > threshold, dtype=tf.float32)\n", 772 | " true = tf.cast(true_flat > threshold, dtype=tf.float32)\n", 773 | " intersection = tf.reduce_sum(pred * true, axis=1) + 1e-7\n", 774 | " denominator = tf.reduce_sum(pred, axis=1) + tf.reduce_sum(true, axis=1) + 1e-7\n", 775 | "\n", 776 | " return tf.reduce_mean(intersection / denominator)\n", 777 | "\n", 778 | "\n", 779 | "def make_train_op(y_pred, y_true):\n", 780 | " \"\"\"Returns a training operation\n", 781 | " Args:\n", 782 | " y_pred (4-D Tensor): (N, H, W, 1)\n", 783 | " y_true (4-D Tensor): (N, H, W, 1)\n", 784 | " Returns:\n", 785 | " train_op: minimize operation\n", 786 | " \"\"\"\n", 787 | " # loss = -IOU_(y_pred, y_true)\n", 788 | " loss = BCE_(y_pred, y_true)\n", 789 | "\n", 790 | " global_step = tf.train.get_or_create_global_step()\n", 791 | "\n", 792 | " # optim = tf.train.AdamOptimizer()\n", 793 | " optim = tf.train.AdamOptimizer(1e-4)\n", 794 | " return optim.minimize(loss, global_step=global_step)\n", 795 | "class flags:\n", 796 | " epochs = 100\n", 797 | " batch_size = 8\n", 798 | " logdir = \"drive/My Drive/data/logs/\"\n", 799 | " reg = 0.1\n", 800 | " ckdir = \"drive/My Drive/data/models/\"\n", 801 | "\n", 802 | "\n", 803 | "'''def read_flags():\n", 804 | " \"\"\"Returns flags\"\"\"\n", 805 | "\n", 806 | " import argparse\n", 807 | "\n", 808 | " parser = argparse.ArgumentParser(\n", 809 | " formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n", 810 | " parser.add_argument(\n", 811 | " \"--epochs\", default=1, type=int, help=\"Number of epochs\")\n", 812 | "\n", 813 | " parser.add_argument(\"--batch-size\", default=8, type=int, help=\"Batch size\")\n", 814 | "\n", 815 | " parser.add_argument(\n", 816 | " \"--logdir\", default=\"logdir\", help=\"Tensorboard log directory\")\n", 817 | "\n", 818 | " parser.add_argument(\n", 819 | " \"--reg\", type=float, default=0.1, help=\"L2 Regularizer Term\")\n", 820 | "\n", 821 | " parser.add_argument(\n", 822 | " \"--ckdir\", default=\"models\", help=\"Checkpoint directory\")\n", 823 | "\n", 824 | " flags = parser.parse_args()\n", 825 | " return flags'''\n", 826 | "\n", 827 | "\n", 828 | "def main():\n", 829 | " train = pd.read_csv(\"./train.csv\")\n", 830 | " n_train = train.shape[0]\n", 831 | "\n", 832 | " test = pd.read_csv(\"./test.csv\")\n", 833 | " n_test = test.shape[0]\n", 834 | "\n", 835 | " current_time = time.strftime(\"%m/%d/%H/%M/%S\")\n", 836 | " train_logdir = os.path.join(flags.logdir, \"train\", current_time)\n", 837 | " test_logdir = os.path.join(flags.logdir, \"test\", current_time)\n", 838 | "\n", 839 | " tf.reset_default_graph()\n", 840 | " X = tf.placeholder(tf.float32, shape=[None, 512, 512, 3], name=\"X\")\n", 841 | " y = tf.placeholder(tf.float32, shape=[None, 512, 512, 1], name=\"y\")\n", 842 | " mode = tf.placeholder(tf.bool, name=\"mode\")\n", 843 | "\n", 844 | " pred = make_unet(X, mode, flags)\n", 845 | "\n", 846 | " tf.add_to_collection(\"inputs\", X)\n", 847 | " tf.add_to_collection(\"inputs\", mode)\n", 848 | " tf.add_to_collection(\"outputs\", pred)\n", 849 | "\n", 850 | " tf.summary.histogram(\"Predicted Mask\", pred)\n", 851 | " tf.summary.image(\"Predicted Mask\", pred)\n", 852 | "\n", 853 | " update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\n", 854 | "\n", 855 | " with tf.control_dependencies(update_ops):\n", 856 | " train_op = make_train_op(pred, y)\n", 857 | "\n", 858 | " IOU_op = IOU_(pred, y)\n", 859 | " IOU_op = tf.Print(IOU_op, [IOU_op])\n", 860 | " tf.summary.scalar(\"IOU\", IOU_op)\n", 861 | "\n", 862 | " train_csv = tf.train.string_input_producer(['train.csv'])\n", 863 | " test_csv = tf.train.string_input_producer(['test.csv'])\n", 864 | " train_image, train_mask = get_image_mask(train_csv)\n", 865 | " test_image, test_mask = get_image_mask(test_csv, augmentation=False)\n", 866 | "\n", 867 | " X_batch_op, y_batch_op = tf.train.shuffle_batch(\n", 868 | " [train_image, train_mask],\n", 869 | " batch_size=flags.batch_size,\n", 870 | " capacity=flags.batch_size * 5,\n", 871 | " min_after_dequeue=flags.batch_size * 2,\n", 872 | " allow_smaller_final_batch=True)\n", 873 | "\n", 874 | " X_test_op, y_test_op = tf.train.batch(\n", 875 | " [test_image, test_mask],\n", 876 | " batch_size=flags.batch_size,\n", 877 | " capacity=flags.batch_size * 2,\n", 878 | " allow_smaller_final_batch=True)\n", 879 | "\n", 880 | " summary_op = tf.summary.merge_all()\n", 881 | "\n", 882 | " with tf.Session() as sess:\n", 883 | " train_summary_writer = tf.summary.FileWriter(train_logdir, sess.graph)\n", 884 | " test_summary_writer = tf.summary.FileWriter(test_logdir)\n", 885 | "\n", 886 | " init = tf.global_variables_initializer()\n", 887 | " sess.run(init)\n", 888 | "\n", 889 | " saver = tf.train.Saver()\n", 890 | " if os.path.exists(flags.ckdir):\n", 891 | " latest_check_point = tf.train.latest_checkpoint(flags.ckdir)\n", 892 | " saver.restore(sess, latest_check_point)\n", 893 | " print('model restored!')\n", 894 | "\n", 895 | " else:\n", 896 | " #try:\n", 897 | " # os.rmdir(flags.ckdir)\n", 898 | " #except IOError:\n", 899 | " # pass\n", 900 | " os.mkdir(flags.ckdir)\n", 901 | "\n", 902 | " try:\n", 903 | " global_step = tf.train.get_global_step(sess.graph)\n", 904 | "\n", 905 | " coord = tf.train.Coordinator()\n", 906 | " threads = tf.train.start_queue_runners(coord=coord)\n", 907 | " start = time.time()\n", 908 | " for epoch in range(90,flags.epochs):\n", 909 | " print('%d epochs in %fs' % (epoch, 62539+time.time()-start))\n", 910 | " for step in range(0, n_train, flags.batch_size):\n", 911 | " #print(\"%d train steps in %fs\" %(step, time.time()-start))\n", 912 | " X_batch, y_batch = sess.run([X_batch_op, y_batch_op])\n", 913 | "\n", 914 | " _, step_iou, step_summary, global_step_value = sess.run(\n", 915 | " [train_op, IOU_op, summary_op, global_step],\n", 916 | " feed_dict={X: X_batch,\n", 917 | " y: y_batch,\n", 918 | " mode: True})\n", 919 | "\n", 920 | " train_summary_writer.add_summary(step_summary,\n", 921 | " global_step_value)\n", 922 | "\n", 923 | " total_iou = 0\n", 924 | " for step in range(0, n_test, flags.batch_size):\n", 925 | " #print(\"%d test steps in %fs\" %(step, time.time()-start))\n", 926 | " X_test, y_test = sess.run([X_test_op, y_test_op])\n", 927 | " step_iou, step_summary = sess.run(\n", 928 | " [IOU_op, summary_op],\n", 929 | " feed_dict={X: X_test,\n", 930 | " y: y_test,\n", 931 | " mode: False})\n", 932 | "\n", 933 | " total_iou += step_iou * X_test.shape[0]\n", 934 | "\n", 935 | " test_summary_writer.add_summary(step_summary,(epoch + 1) * (step + 1))\n", 936 | "\n", 937 | " #saver.save(sess, \"{}/model.ckpt\".format(flags.ckdir))\n", 938 | " #print('the %d epoch in %fs, saved successfully' % (epoch, time.time()-start))\n", 939 | "\n", 940 | " finally:\n", 941 | " coord.request_stop()\n", 942 | " coord.join(threads)\n", 943 | " saver.save(sess, \"{}/model.ckpt\".format(flags.ckdir))\n", 944 | " print('model saved successfully')\n", 945 | "\n", 946 | "\n", 947 | "\n", 948 | "if __name__ == '__main__':\n", 949 | " #flags = read_flags()\n", 950 | " main()" 951 | ], 952 | "execution_count": null, 953 | "outputs": [ 954 | { 955 | "output_type": "stream", 956 | "text": [ 957 | "INFO:tensorflow:Summary name Predicted Mask is illegal; using Predicted_Mask instead.\n", 958 | "INFO:tensorflow:Summary name Predicted Mask is illegal; using Predicted_Mask instead.\n", 959 | "INFO:tensorflow:Restoring parameters from drive/My Drive/data/models/model.ckpt\n", 960 | "model restored!\n", 961 | "90 epochs in 62539.000005s\n", 962 | "91 epochs in 63012.843219s\n", 963 | "92 epochs in 63480.010369s\n", 964 | "93 epochs in 63947.371985s\n", 965 | "94 epochs in 64414.004312s\n", 966 | "95 epochs in 64883.230457s\n", 967 | "96 epochs in 65350.314995s\n", 968 | "97 epochs in 65820.104939s\n", 969 | "98 epochs in 66288.175114s\n", 970 | "99 epochs in 66757.795235s\n", 971 | "model saved successfully\n" 972 | ], 973 | "name": "stdout" 974 | } 975 | ] 976 | }, 977 | { 978 | "cell_type": "code", 979 | "metadata": { 980 | "id": "TQOF55DOjN_B" 981 | }, 982 | "source": [ 983 | "%load_ext tensorboard\n", 984 | "%tensorboard --logdir 'drive/My Drive/data/logs'\n", 985 | "#%reload_ext tensorboard" 986 | ], 987 | "execution_count": null, 988 | "outputs": [] 989 | }, 990 | { 991 | "cell_type": "markdown", 992 | "metadata": { 993 | "id": "p2JGpoMrtNx3" 994 | }, 995 | "source": [ 996 | "### Test on a single image" 997 | ] 998 | }, 999 | { 1000 | "cell_type": "code", 1001 | "metadata": { 1002 | "id": "AXJmE8wYtUJ2", 1003 | "colab": { 1004 | "base_uri": "https://localhost:8080/", 1005 | "height": 139 1006 | }, 1007 | "outputId": "b50675a1-3018-42ed-a51c-9d2e17fd8f4a" 1008 | }, 1009 | "source": [ 1010 | "import tensorflow as tf\n", 1011 | "import cv2\n", 1012 | "import pandas as pd\n", 1013 | "import matplotlib.pyplot as plt\n", 1014 | "import numpy as np\n", 1015 | "from scipy.ndimage.measurements import label\n", 1016 | "from PIL import Image\n", 1017 | "from scipy.special import expit\n", 1018 | "\n", 1019 | "saver = tf.train.import_meta_graph(\"drive/My Drive/data/models/model.ckpt.meta\")\n", 1020 | "sess = tf.InteractiveSession()\n", 1021 | "saver.restore(sess, \"drive/My Drive/data/models/model.ckpt\")\n", 1022 | "X, mode = tf.get_collection(\"inputs\")[:2]\n", 1023 | "pred = tf.get_collection(\"outputs\")[0]\n", 1024 | "\n", 1025 | "def read_image(image_path, gray=False):\n", 1026 | " \"\"\"Returns an image array\n", 1027 | " Args:\n", 1028 | " image_path (str): Path to image.jpg\n", 1029 | " Returns:\n", 1030 | " 3-D array: RGB numpy image array\n", 1031 | " \"\"\"\n", 1032 | " if gray:\n", 1033 | " return cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)\n", 1034 | " \n", 1035 | " image = cv2.imread(image_path) \n", 1036 | " return cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n", 1037 | "\n", 1038 | "def pipeline(image, threshold=0.5, image_WH=(512, 512)):\n", 1039 | " image = np.copy(image)\n", 1040 | " H, W, C = image.shape\n", 1041 | " \n", 1042 | " if (W, H) != image_WH:\n", 1043 | " image = cv2.resize(image, image_WH)\n", 1044 | " \n", 1045 | " mask_pred = sess.run(pred, feed_dict={X: np.expand_dims(image, 0),\n", 1046 | " mode: False})\n", 1047 | " \n", 1048 | " mask_pred = np.squeeze(mask_pred)\n", 1049 | " mask_pred = expit(mask_pred)\n", 1050 | " # mask_pred = mask_pred > threshold\n", 1051 | " return mask_pred\n", 1052 | "\n", 1053 | "image_path = \"drive/My Drive/data/output/images/IDRiD_17.jpg\"\n", 1054 | "image = read_image(image_path)\n", 1055 | "predicted_image = np.zeros((2848, 4288), dtype=float)\n", 1056 | "\n", 1057 | "for i in range(10): #10 6\n", 1058 | " for j in range(16): #16 9\n", 1059 | " top_y = i*256 #256 512\n", 1060 | " if (i==9): #9 5\n", 1061 | " top_y = 2336\n", 1062 | " top_x = j*256 #256 512\n", 1063 | " if (j==15): #15 8\n", 1064 | " top_x = 3776\n", 1065 | "\n", 1066 | " image_crop = image[top_y:top_y+512, top_x:top_x+512]\n", 1067 | " predicted_crop = pipeline(image_crop)\n", 1068 | " predicted_image[top_y:top_y+512, top_x:top_x+512] = np.maximum(predicted_image[top_y:top_y+512, top_x:top_x+512], predicted_crop)\n", 1069 | "\n", 1070 | "threshold = 0.7\n", 1071 | "predicted_image = predicted_image > threshold\n", 1072 | "(unique, counts) = np.unique(predicted_image.astype('uint8')*255, return_counts=True)\n", 1073 | "frequencies = np.asarray((unique, counts)).T\n", 1074 | "print(frequencies)\n", 1075 | "\n", 1076 | "predicted_save = Image.fromarray((predicted_image.astype('uint8'))*255)\n", 1077 | "predicted_save.save(\"test_predicted.jpg\", \"JPEG\")" 1078 | ], 1079 | "execution_count": null, 1080 | "outputs": [ 1081 | { 1082 | "output_type": "stream", 1083 | "text": [ 1084 | "WARNING:tensorflow:From /tensorflow-1.15.2/python3.6/tensorflow_core/python/training/queue_runner_impl.py:391: QueueRunner.__init__ (from tensorflow.python.training.queue_runner_impl) is deprecated and will be removed in a future version.\n", 1085 | "Instructions for updating:\n", 1086 | "To construct input pipelines, use the `tf.data` module.\n", 1087 | "INFO:tensorflow:Restoring parameters from drive/My Drive/data/models/model.ckpt\n", 1088 | "[[ 0 10224722]\n", 1089 | " [ 255 1987502]]\n" 1090 | ], 1091 | "name": "stdout" 1092 | } 1093 | ] 1094 | }, 1095 | { 1096 | "cell_type": "markdown", 1097 | "metadata": { 1098 | "id": "ByIpNhUk-xQr" 1099 | }, 1100 | "source": [ 1101 | "### Generate probability maps for the dataset using the trained model" 1102 | ] 1103 | }, 1104 | { 1105 | "cell_type": "code", 1106 | "metadata": { 1107 | "id": "1zed1ERU_FyF", 1108 | "colab": { 1109 | "base_uri": "https://localhost:8080/", 1110 | "height": 88 1111 | }, 1112 | "outputId": "67f45da6-6b60-459b-c072-9d364ff20b58" 1113 | }, 1114 | "source": [ 1115 | "import tensorflow as tf\n", 1116 | "import cv2\n", 1117 | "import pandas as pd\n", 1118 | "import matplotlib.pyplot as plt\n", 1119 | "import numpy as np\n", 1120 | "from scipy.ndimage.measurements import label\n", 1121 | "from PIL import Image\n", 1122 | "from scipy.special import expit\n", 1123 | "import os, sys\n", 1124 | "\n", 1125 | "saver = tf.train.import_meta_graph(\"drive/My Drive/data/models/model.ckpt.meta\")\n", 1126 | "sess = tf.InteractiveSession()\n", 1127 | "saver.restore(sess, \"drive/My Drive/data/models/model.ckpt\")\n", 1128 | "X, mode = tf.get_collection(\"inputs\")[:2]\n", 1129 | "pred = tf.get_collection(\"outputs\")[0]\n", 1130 | "\n", 1131 | "def read_image(image_path, gray=False):\n", 1132 | " \"\"\"Returns an image array\n", 1133 | " Args:\n", 1134 | " image_path (str): Path to image.jpg\n", 1135 | " Returns:\n", 1136 | " 3-D array: RGB numpy image array\n", 1137 | " \"\"\"\n", 1138 | " if gray:\n", 1139 | " return cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)\n", 1140 | " \n", 1141 | " image = cv2.imread(image_path) \n", 1142 | " return cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n", 1143 | "\n", 1144 | "def pipeline(image, image_WH=(512, 512)):\n", 1145 | " image = np.copy(image)\n", 1146 | " H, W, C = image.shape\n", 1147 | " \n", 1148 | " if (W, H) != image_WH:\n", 1149 | " image = cv2.resize(image, image_WH)\n", 1150 | " \n", 1151 | " mask_pred = sess.run(pred, feed_dict={X: np.expand_dims(image, 0),\n", 1152 | " mode: False})\n", 1153 | " \n", 1154 | " mask_pred = np.squeeze(mask_pred)\n", 1155 | " mask_pred = expit(mask_pred)\n", 1156 | " # mask_pred = mask_pred > threshold\n", 1157 | " return mask_pred\n", 1158 | "\n", 1159 | "output_dir = \"drive/My Drive/data/output/prob/\"\n", 1160 | "dir = os.getcwd()\n", 1161 | "\n", 1162 | "if not os.path.exists(os.path.join(dir,output_dir)):\n", 1163 | " os.mkdir(output_dir)\n", 1164 | "\n", 1165 | "for image_path in os.listdir(os.path.join(dir,\"drive/My Drive/data/output/images/\")):\n", 1166 | " image = read_image(\"drive/My Drive/data/output/images/\"+image_path)\n", 1167 | " predicted_image = np.zeros((2848, 4288), dtype=float)\n", 1168 | "\n", 1169 | " for i in range(10): #10 6\n", 1170 | " for j in range(16): #16 9\n", 1171 | " top_y = i*256 #256 512\n", 1172 | " if (i==9): #9 5\n", 1173 | " top_y = 2336\n", 1174 | " top_x = j*256 #256 512\n", 1175 | " if (j==15): #15 8\n", 1176 | " top_x = 3776\n", 1177 | "\n", 1178 | " image_crop = image[top_y:top_y+512, top_x:top_x+512]\n", 1179 | " predicted_crop = pipeline(image_crop)\n", 1180 | " predicted_image[top_y:top_y+512, top_x:top_x+512] = np.maximum(predicted_image[top_y:top_y+512, top_x:top_x+512], predicted_crop)\n", 1181 | "\n", 1182 | " # threshold = 0.5\n", 1183 | " # predicted_image = predicted_image > threshold\n", 1184 | " #(unique, counts) = np.unique((predicted_image*255).astype('uint8'), return_counts=True)\n", 1185 | " #frequencies = np.asarray((unique, counts)).T\n", 1186 | " #print(frequencies)\n", 1187 | " predicted_save = Image.fromarray((predicted_image*255).astype('uint8'))\n", 1188 | " predicted_save.save(output_dir+image_path, \"JPEG\", quality=100)" 1189 | ], 1190 | "execution_count": null, 1191 | "outputs": [ 1192 | { 1193 | "output_type": "stream", 1194 | "text": [ 1195 | "INFO:tensorflow:Restoring parameters from drive/My Drive/data/models/model.ckpt\n" 1196 | ], 1197 | "name": "stdout" 1198 | }, 1199 | { 1200 | "output_type": "stream", 1201 | "text": [ 1202 | "/tensorflow-1.15.2/python3.6/tensorflow_core/python/client/session.py:1750: UserWarning: An interactive session is already active. This can cause out-of-memory errors in some cases. You must explicitly call `InteractiveSession.close()` to release resources held by the other session(s).\n", 1203 | " warnings.warn('An interactive session is already active. This can '\n" 1204 | ], 1205 | "name": "stderr" 1206 | } 1207 | ] 1208 | }, 1209 | { 1210 | "cell_type": "markdown", 1211 | "metadata": { 1212 | "id": "CM2NY0sRHysI" 1213 | }, 1214 | "source": [ 1215 | "### plot FROC curve" 1216 | ] 1217 | }, 1218 | { 1219 | "cell_type": "code", 1220 | "metadata": { 1221 | "id": "Aq8hygaTHxv-", 1222 | "colab": { 1223 | "base_uri": "https://localhost:8080/", 1224 | "height": 986 1225 | }, 1226 | "outputId": "fc83bffd-4f6e-4a40-9828-ebd1dbb384b0" 1227 | }, 1228 | "source": [ 1229 | "from PIL import Image\n", 1230 | "import numpy as np\n", 1231 | "import os, sys\n", 1232 | "import matplotlib\n", 1233 | "matplotlib.use('Agg')\n", 1234 | "import matplotlib.pyplot as plt\n", 1235 | "\n", 1236 | "training_size=81\n", 1237 | "gt_dir = \"drive/My Drive/data/output/annotations/\"\n", 1238 | "prob_dir = \"drive/My Drive/data/output/prob/\"\n", 1239 | "true_p=0\n", 1240 | "actual_p=0\n", 1241 | "pred_p=0\n", 1242 | "false_p=0\n", 1243 | "\n", 1244 | "thresh_list = [0, 0.00001, 0.0001, 0.001, 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99, 0.999, 0.9999, 0.99999, 1]\n", 1245 | "\n", 1246 | "dir = os.getcwd()\n", 1247 | "thresh_size = len(thresh_list)\n", 1248 | "sn = np.empty(thresh_size, dtype=float)\n", 1249 | "fppi = np.empty(thresh_size, dtype=float)\n", 1250 | "thresh_array = np.array(thresh_list)\n", 1251 | "\n", 1252 | "for th in range(thresh_size):\n", 1253 | "\tthreshold = thresh_array[th]\n", 1254 | "\tprint (threshold)\n", 1255 | "\ttrue_p=0\n", 1256 | "\tactual_p=0\n", 1257 | "\tpred_p=0\n", 1258 | "\tfalse_p=0\n", 1259 | "\n", 1260 | "\tfor image_path in os.listdir(os.path.join(dir,gt_dir)):\n", 1261 | "\t\t# print image_path\n", 1262 | "\t\tim_gt = Image.open(gt_dir+image_path)\n", 1263 | "\t\timg = os.path.splitext(image_path)[0] + \".jpg\"\n", 1264 | "\t\tim_prob = Image.open(prob_dir+img)\n", 1265 | "\t\tarr_gt = np.array(im_gt)\n", 1266 | "\t\t#(unique, counts) = np.unique(arr_gt, return_counts=True)\n", 1267 | "\t\t#frequencies = np.asarray((unique, counts)).T\n", 1268 | "\t\t#print(\"arr_gt: \",frequencies)\n", 1269 | "\t\tarr_prob = (np.array(im_prob)).astype(float)/255\n", 1270 | "\t\t#(unique, counts) = np.unique(arr_prob, return_counts=True)\n", 1271 | "\t\t#frequencies = np.asarray((unique, counts)).T\n", 1272 | "\t\t#print(\"arr_prob: \",frequencies)\n", 1273 | "\t\tarr_pred = (arr_prob > threshold).astype('uint8')\n", 1274 | "\t\t#(unique, counts) = np.unique(arr_pred, return_counts=True)\n", 1275 | "\t\t#frequencies = np.asarray((unique, counts)).T\n", 1276 | "\t\t#print(\"arr_pred: \",frequencies)\n", 1277 | "\t\ttp = np.sum(np.logical_and(arr_gt, arr_pred))\n", 1278 | "\t\tap = np.sum(arr_gt)\n", 1279 | "\t\tpp = np.sum(arr_pred)\n", 1280 | "\t\ttrue_p += tp\n", 1281 | "\t\tactual_p += ap\n", 1282 | "\t\tpred_p += pp\n", 1283 | "\t\tfalse_p += (pp-tp)\n", 1284 | "\n", 1285 | "\tsn[th] = float(true_p)/float(actual_p)\n", 1286 | "\tprint (\"sn: \", sn[th])\n", 1287 | "\tfppi[th] = float(false_p)/float(training_size)\n", 1288 | "\tprint (\"fppi: \", fppi[th])\n", 1289 | "\n", 1290 | "plt.plot(fppi, sn)\n", 1291 | "plt.ylabel('SN')\n", 1292 | "plt.xlabel('FPs per image')\n", 1293 | "plt.savefig('drive/My Drive/data/froc.png')" 1294 | ], 1295 | "execution_count": null, 1296 | "outputs": [ 1297 | { 1298 | "output_type": "stream", 1299 | "text": [ 1300 | "0.0\n", 1301 | "sn: 0.9997918539315748\n", 1302 | "fppi: 433847.8024691358\n", 1303 | "1e-05\n", 1304 | "sn: 0.9997918539315748\n", 1305 | "fppi: 433847.8024691358\n", 1306 | "0.0001\n", 1307 | "sn: 0.9997918539315748\n", 1308 | "fppi: 433847.8024691358\n", 1309 | "0.001\n", 1310 | "sn: 0.9997918539315748\n", 1311 | "fppi: 433847.8024691358\n", 1312 | "0.01\n", 1313 | "sn: 0.9994911652226554\n", 1314 | "fppi: 298734.4567901235\n", 1315 | "0.1\n", 1316 | "sn: 0.9943516045316544\n", 1317 | "fppi: 163824.34567901236\n", 1318 | "0.2\n", 1319 | "sn: 0.9918498585006056\n", 1320 | "fppi: 133057.64197530865\n", 1321 | "0.3\n", 1322 | "sn: 0.9889898217071691\n", 1323 | "fppi: 113391.08641975309\n", 1324 | "0.4\n", 1325 | "sn: 0.985346916103857\n", 1326 | "fppi: 97763.20987654322\n", 1327 | "0.5\n", 1328 | "sn: 0.9808762180787789\n", 1329 | "fppi: 85151.83950617284\n", 1330 | "0.6\n", 1331 | "sn: 0.9739881305827571\n", 1332 | "fppi: 72394.50617283951\n", 1333 | "0.7\n", 1334 | "sn: 0.9626368323303055\n", 1335 | "fppi: 59131.567901234564\n", 1336 | "0.8\n", 1337 | "sn: 0.9414782104509493\n", 1338 | "fppi: 44431.61728395062\n", 1339 | "0.9\n", 1340 | "sn: 0.892350447299412\n", 1341 | "fppi: 26864.46913580247\n", 1342 | "0.99\n", 1343 | "sn: 0.5611847614316616\n", 1344 | "fppi: 2681.543209876543\n", 1345 | "0.999\n", 1346 | "sn: 0.001640410645737987\n", 1347 | "fppi: 13.061728395061728\n", 1348 | "0.9999\n", 1349 | "sn: 0.001640410645737987\n", 1350 | "fppi: 13.061728395061728\n", 1351 | "0.99999\n", 1352 | "sn: 0.001640410645737987\n", 1353 | "fppi: 13.061728395061728\n", 1354 | "1.0\n", 1355 | "sn: 0.0\n", 1356 | "fppi: 0.0\n" 1357 | ], 1358 | "name": "stdout" 1359 | } 1360 | ] 1361 | }, 1362 | { 1363 | "cell_type": "markdown", 1364 | "metadata": { 1365 | "id": "rnNDIjEh_iD3" 1366 | }, 1367 | "source": [ 1368 | "### Generate segmented output masks from the probability maps" 1369 | ] 1370 | }, 1371 | { 1372 | "cell_type": "code", 1373 | "metadata": { 1374 | "id": "37-KEnaJ_lxg", 1375 | "colab": { 1376 | "base_uri": "https://localhost:8080/", 1377 | "height": 88 1378 | }, 1379 | "outputId": "4cb1dd97-c2dd-4144-f29e-5576fc42da3a" 1380 | }, 1381 | "source": [ 1382 | "import tensorflow as tf\n", 1383 | "import cv2\n", 1384 | "import pandas as pd\n", 1385 | "import matplotlib.pyplot as plt\n", 1386 | "import numpy as np\n", 1387 | "from scipy.ndimage.measurements import label\n", 1388 | "from PIL import Image\n", 1389 | "from scipy.special import expit\n", 1390 | "import os, sys\n", 1391 | "\n", 1392 | "saver = tf.train.import_meta_graph(\"drive/My Drive/data/models/model.ckpt.meta\")\n", 1393 | "sess = tf.InteractiveSession()\n", 1394 | "saver.restore(sess, \"drive/My Drive/data/models/model.ckpt\")\n", 1395 | "X, mode = tf.get_collection(\"inputs\")[:2]\n", 1396 | "pred = tf.get_collection(\"outputs\")[0]\n", 1397 | "\n", 1398 | "def read_image(image_path, gray=False):\n", 1399 | " \"\"\"Returns an image array\n", 1400 | " Args:\n", 1401 | " image_path (str): Path to image.jpg\n", 1402 | " Returns:\n", 1403 | " 3-D array: RGB numpy image array\n", 1404 | " \"\"\"\n", 1405 | " if gray:\n", 1406 | " return cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)\n", 1407 | " \n", 1408 | " image = cv2.imread(image_path) \n", 1409 | " return cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n", 1410 | "\n", 1411 | "def pipeline(image, image_WH=(512, 512)):\n", 1412 | " image = np.copy(image)\n", 1413 | " H, W, C = image.shape\n", 1414 | " \n", 1415 | " if (W, H) != image_WH:\n", 1416 | " image = cv2.resize(image, image_WH)\n", 1417 | " \n", 1418 | " mask_pred = sess.run(pred, feed_dict={X: np.expand_dims(image, 0),\n", 1419 | " mode: False})\n", 1420 | " \n", 1421 | " mask_pred = np.squeeze(mask_pred)\n", 1422 | " mask_pred = expit(mask_pred)\n", 1423 | " # mask_pred = mask_pred > threshold\n", 1424 | " return mask_pred\n", 1425 | "\n", 1426 | "output_dir = \"drive/My Drive/data/output/predicted/\"\n", 1427 | "dir = os.getcwd()\n", 1428 | "\n", 1429 | "if not os.path.exists(os.path.join(dir,output_dir)):\n", 1430 | " os.mkdir(output_dir)\n", 1431 | "\n", 1432 | "for image_path in os.listdir(os.path.join(dir,\"drive/My Drive/data/output/prob/\")):\n", 1433 | " im_prob = Image.open(\"drive/My Drive/data/output/prob/\"+image_path)\n", 1434 | " arr_prob = (np.array(im_prob)).astype(float)/255\n", 1435 | " threshold = 0.8\n", 1436 | " arr_pred = (arr_prob > threshold).astype('uint8')\n", 1437 | " # image = read_image(\"test_data/\"+image_path)\n", 1438 | " # predicted_image = np.zeros((2848, 4288), dtype=float)\n", 1439 | "\n", 1440 | " # for i in range(10):\n", 1441 | " # for j in range(16):\n", 1442 | " # top_y = i*256\n", 1443 | " # if (i==9):\n", 1444 | " # top_y = 2336\n", 1445 | " # top_x = j*256\n", 1446 | " # if (j==15):\n", 1447 | " # top_x = 3776\n", 1448 | "\n", 1449 | " # image_crop = image[top_y:top_y+512, top_x:top_x+512]\n", 1450 | " # predicted_crop = pipeline(image_crop)\n", 1451 | " # predicted_image[top_y:top_y+512, top_x:top_x+512] = np.maximum(predicted_image[top_y:top_y+512, top_x:top_x+512], predicted_crop)\n", 1452 | " #(unique, counts) = np.unique(arr_pred*255, return_counts=True)\n", 1453 | " #frequencies = np.asarray((unique, counts)).T\n", 1454 | " #print(frequencies)\n", 1455 | " predicted_save = Image.fromarray(arr_pred*255)\n", 1456 | " predicted_save.save(output_dir+image_path, \"JPEG\", quality=100)" 1457 | ], 1458 | "execution_count": null, 1459 | "outputs": [ 1460 | { 1461 | "output_type": "stream", 1462 | "text": [ 1463 | "INFO:tensorflow:Restoring parameters from drive/My Drive/data/models/model.ckpt\n" 1464 | ], 1465 | "name": "stdout" 1466 | }, 1467 | { 1468 | "output_type": "stream", 1469 | "text": [ 1470 | "/tensorflow-1.15.2/python3.6/tensorflow_core/python/client/session.py:1750: UserWarning: An interactive session is already active. This can cause out-of-memory errors in some cases. You must explicitly call `InteractiveSession.close()` to release resources held by the other session(s).\n", 1471 | " warnings.warn('An interactive session is already active. This can '\n" 1472 | ], 1473 | "name": "stderr" 1474 | } 1475 | ] 1476 | }, 1477 | { 1478 | "cell_type": "markdown", 1479 | "metadata": { 1480 | "id": "mjewf0kzHheN" 1481 | }, 1482 | "source": [ 1483 | "### Calculate sensitivity and precison values for individual images" 1484 | ] 1485 | }, 1486 | { 1487 | "cell_type": "code", 1488 | "metadata": { 1489 | "id": "UfPnyrHXHm4D" 1490 | }, 1491 | "source": [ 1492 | "from PIL import Image\n", 1493 | "import numpy as np\n", 1494 | "import os, sys\n", 1495 | "\n", 1496 | "training_size=80\n", 1497 | "gt_dir = \"drive/My Drive/data/output/annotations/\"\n", 1498 | "pred_dir = \"drive/My Drive/data/output/predicted/\"\n", 1499 | "sn = np.empty(training_size, dtype=float)\n", 1500 | "ppv = np.empty(training_size, dtype=float)\n", 1501 | "sp = np.empty(training_size, dtype=float)\n", 1502 | "image_paths = np.empty(training_size, dtype=object)\n", 1503 | "\n", 1504 | "dir = os.getcwd()\n", 1505 | "i=0\n", 1506 | "for image_path in os.listdir(os.path.join(dir,gt_dir)):\n", 1507 | "\timage_paths[i] = image_path\n", 1508 | "\tim_gt = Image.open(gt_dir+image_path)\n", 1509 | "\timg = os.path.splitext(image_path)[0] + \".jpg\"\n", 1510 | "\tim_pred = Image.open(pred_dir+img)\n", 1511 | "\tarr_gt = np.array(im_gt)\n", 1512 | "\tarr_pred = np.array(im_pred)\n", 1513 | "\tarr_pred = arr_pred > 0\n", 1514 | "\t#(unique, counts) = np.unique(arr_gt, return_counts=True)\n", 1515 | "\t#frequencies = np.asarray((unique, counts)).T\n", 1516 | "\t#print(\"arr_gt: \",frequencies)\n", 1517 | "\t#(unique, counts) = np.unique(arr_pred, return_counts=True)\n", 1518 | "\t#frequencies = np.asarray((unique, counts)).T\n", 1519 | "\t#print(\"arr_pred: \",frequencies)\n", 1520 | "\ttrue_p = np.sum(np.logical_and(arr_gt, arr_pred))\n", 1521 | "\tactual_p = np.sum(arr_gt)\n", 1522 | "\tpred_p = np.sum(arr_pred)\n", 1523 | "\t\n", 1524 | "\tfalse_p = pred_p - true_p\n", 1525 | "\tactual_n = 4288*2848 - actual_p\n", 1526 | "\ttrue_n = actual_n - false_p\n", 1527 | "\t#print (\"True pos: \", true_p)\n", 1528 | "\t#print (\"Actual pos: \", actual_p)\n", 1529 | "\t#print (\"Pred pos: \", pred_p)\n", 1530 | "\tif actual_p == 0:\n", 1531 | "\t\tsn[i] = 1\n", 1532 | "\telse:\n", 1533 | "\t\tsn[i] = float(true_p)/float(actual_p)\n", 1534 | "\tif pred_p == 0:\n", 1535 | "\t\tppv[i] = 1\n", 1536 | "\telse:\n", 1537 | "\t\tppv[i] = float(true_p)/float(pred_p)\n", 1538 | "\t#print (i)\n", 1539 | "\tif actual_n == 0:\n", 1540 | "\t\tsp[i] = 1\n", 1541 | "\telse:\n", 1542 | "\t\tsp[i] = float(true_n)/float(actual_n)\n", 1543 | "\ti+=1\n", 1544 | "\n", 1545 | "sn_csv = np.stack((image_paths,sn), axis=1)\n", 1546 | "ppv_csv = np.stack((image_paths,ppv), axis=1)\n", 1547 | "sp_csv = np.stack((image_paths,sp), axis=1)\n", 1548 | "\n", 1549 | "np.savetxt(\"drive/My Drive/data/sn.csv\", sn_csv, delimiter=\",\", fmt=\"%s\")\n", 1550 | "np.savetxt(\"drive/My Drive/data/ppv.csv\", ppv_csv, delimiter=\",\", fmt=\"%s\")\n", 1551 | "np.savetxt(\"drive/My Drive/data/sp.csv\", sp_csv, delimiter=\",\", fmt=\"%s\")" 1552 | ], 1553 | "execution_count": null, 1554 | "outputs": [] 1555 | }, 1556 | { 1557 | "cell_type": "markdown", 1558 | "metadata": { 1559 | "id": "8hwL8cLfH89I" 1560 | }, 1561 | "source": [ 1562 | "### compute average statistics" 1563 | ] 1564 | }, 1565 | { 1566 | "cell_type": "code", 1567 | "metadata": { 1568 | "id": "GJaPdfF89P7x", 1569 | "colab": { 1570 | "base_uri": "https://localhost:8080/", 1571 | "height": 119 1572 | }, 1573 | "outputId": "3cca101e-3813-4268-caea-a2efd1bdca76" 1574 | }, 1575 | "source": [ 1576 | "from PIL import Image\n", 1577 | "import numpy as np\n", 1578 | "import os, sys\n", 1579 | "import matplotlib\n", 1580 | "matplotlib.use('Agg')\n", 1581 | "import matplotlib.pyplot as plt\n", 1582 | "#np.set_printoptions(threshold=sys.maxsize)\n", 1583 | "\n", 1584 | "gt_dir = \"drive/My Drive/data/output/annotations/\"\n", 1585 | "pred_dir = \"drive/My Drive/data/output/predicted/\"\n", 1586 | "true_p=0\n", 1587 | "actual_p=0\n", 1588 | "pred_p=0\n", 1589 | "false_p=0\n", 1590 | "false_n=0\n", 1591 | "actual_n=0\n", 1592 | "true_n=0\n", 1593 | "pred_n=0\n", 1594 | "\n", 1595 | "dir = os.getcwd()\n", 1596 | "\n", 1597 | "for image_path in os.listdir(os.path.join(dir,gt_dir)):\n", 1598 | " im_gt = Image.open(gt_dir+image_path)\n", 1599 | " img = os.path.splitext(image_path)[0] + \".jpg\"\n", 1600 | " im_pred = Image.open(pred_dir+img)\n", 1601 | " arr_gt = np.array(im_gt)\n", 1602 | " arr_pred = np.array(im_pred)\n", 1603 | " arr_pred = arr_pred > 0\n", 1604 | " #(unique, counts) = np.unique(arr_pred, return_counts=True)\n", 1605 | " #frequencies = np.asarray((unique, counts)).T\n", 1606 | " #print(\"arr_pred: \",frequencies)\n", 1607 | "\n", 1608 | " tp = np.sum(np.logical_and(arr_gt, arr_pred))\n", 1609 | " #print(\"tp: \",tp)\n", 1610 | " ap = np.sum(arr_gt)\n", 1611 | " #print(\"ap: \",ap)\n", 1612 | " pp = np.sum(arr_pred)\n", 1613 | "\n", 1614 | " fp = pp - tp\n", 1615 | " an = 4288*2848 - ap\n", 1616 | " pn = 4288*2848 - pp\n", 1617 | " tn = an - fp\n", 1618 | " fn = pn - tn\n", 1619 | "\n", 1620 | " true_p += tp\n", 1621 | " actual_p += ap\n", 1622 | " pred_p += pp\n", 1623 | " pred_n += pn\n", 1624 | " false_p += fp\n", 1625 | " actual_n += an\n", 1626 | " true_n += tn\n", 1627 | " false_n +=fn\n", 1628 | "\n", 1629 | "\n", 1630 | "sn = float(true_p)/float(actual_p)\n", 1631 | "ppv = float(true_p)/float(pred_p)\n", 1632 | "sp = float(true_n)/float(actual_n)\n", 1633 | "npv = float(true_n)/float(pred_n)\n", 1634 | "acc = float(true_p + true_n)/float(actual_p + actual_n)\n", 1635 | "f1 = float(2*true_p)/float((2*true_p)+false_p+false_n)\n", 1636 | "\n", 1637 | "print (\"Sensitivity/Recall/True Positive Rate(TPR)(TP/P): \", sn)\n", 1638 | "print (\"Precision/Positive Predictive Value(PPV)(TP/TP+FP): \", ppv)\n", 1639 | "print (\"Specificity/Selectivity/True Negative Rate(TNR)(TN/N): \", sp)\n", 1640 | "print (\"Negative Predictive Value(NPV)(TN/TN+FN): \", npv)\n", 1641 | "print (\"Accuracy(TP+TN/P+N)\",acc)\n", 1642 | "print (\"F1 Score(2TP/2TP+FP+FN)\",f1)" 1643 | ], 1644 | "execution_count": null, 1645 | "outputs": [ 1646 | { 1647 | "output_type": "stream", 1648 | "text": [ 1649 | "Sensitivity/Recall/True Positive Rate(TPR)(TP/P): 0.9425759438650518\n", 1650 | "Precision/Positive Predictive Value(PPV)(TP/TP+FP): 0.7188802057557396\n", 1651 | "Specificity/Selectivity/True Negative Rate(TNR)(TN/N): 0.9961816109225247\n", 1652 | "Negative Predictive Value(NPV)(TN/TN+FN): 0.9994032052285645\n", 1653 | "Accuracy(TP+TN/P+N) 0.9956319893084176\n", 1654 | "F1 Score(2TP/2TP+FP+FN) 0.8156690606860414\n" 1655 | ], 1656 | "name": "stdout" 1657 | } 1658 | ] 1659 | }, 1660 | { 1661 | "cell_type": "markdown", 1662 | "metadata": { 1663 | "id": "r9LjYv56IQ7Q" 1664 | }, 1665 | "source": [ 1666 | "### precision score" 1667 | ] 1668 | }, 1669 | { 1670 | "cell_type": "code", 1671 | "metadata": { 1672 | "id": "Ott8N-YVISa9", 1673 | "colab": { 1674 | "base_uri": "https://localhost:8080/", 1675 | "height": 34 1676 | }, 1677 | "outputId": "6ae66a0c-f6bc-439a-8aea-3faa125b7c53" 1678 | }, 1679 | "source": [ 1680 | "from PIL import Image\n", 1681 | "import numpy as np\n", 1682 | "import os, sys\n", 1683 | "from sklearn.metrics import average_precision_score\n", 1684 | "\n", 1685 | "training_size=80\n", 1686 | "gt_dir = \"drive/My Drive/data/output/annotations/\"\n", 1687 | "prob_dir = \"drive/My Drive/data/output/prob/\"\n", 1688 | "\n", 1689 | "dir = os.getcwd()\n", 1690 | "\n", 1691 | "i=0\n", 1692 | "sum_pav=0\n", 1693 | "for image_path in os.listdir(os.path.join(dir,gt_dir)):\n", 1694 | "\t# print image_path\n", 1695 | "\tim_gt = Image.open(gt_dir+image_path)\n", 1696 | "\timg = os.path.splitext(image_path)[0] + \".jpg\"\n", 1697 | "\tim_prob = Image.open(prob_dir+img)\n", 1698 | "\tarr_gt = (np.array(im_gt)).astype(bool)\n", 1699 | "\tarr_prob = (np.array(im_prob)).astype(float)/255\n", 1700 | "\t#(unique, counts) = np.unique(arr_gt, return_counts=True)\n", 1701 | "\t#frequencies = np.asarray((unique, counts)).T\n", 1702 | "\t#print(\"arr_prob: \",frequencies)\n", 1703 | "\tpav = average_precision_score(arr_gt.reshape((-1)),arr_prob.reshape((-1)))\n", 1704 | "\tsum_pav = sum_pav+pav\n", 1705 | "\ti = i+1\n", 1706 | "\n", 1707 | "mpav = sum_pav/i\n", 1708 | "print(mpav)" 1709 | ], 1710 | "execution_count": null, 1711 | "outputs": [ 1712 | { 1713 | "output_type": "stream", 1714 | "text": [ 1715 | "0.7908225819055167\n" 1716 | ], 1717 | "name": "stdout" 1718 | } 1719 | ] 1720 | } 1721 | ] 1722 | } --------------------------------------------------------------------------------