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

2 | 3 |

4 | 5 | [![License: GPL v3](https://img.shields.io/badge/License-GPLv3-orange.svg)](https://www.gnu.org/licenses/gpl-3.0) 6 | DOI DOI 7 |
8 | 9 | # Interpretable Gland-Graph Networks using a Neural Aggregator 10 | 11 | IGUANA is a graph neural network built for colon biopsy screening. IGUANA represents a whole-slide image (WSI) as a graph built with nodes on top of glands in the tissue, each node associated with a set of interpretable features. 12 | 13 | For a full description, take a look at our [preprint](https://doi.org/10.1101/2022.10.17.22279804). 14 | 15 | ## Set Up Environment 16 | 17 | ``` 18 | # create base conda environment 19 | conda env create -f environment.yml 20 | 21 | # activate environment 22 | conda activate iguana 23 | 24 | # install PyTorch with pip 25 | pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/cu102/torch_stable.html 26 | 27 | # install PyTorch Geometric and dependencies 28 | pip install torch-scatter -f https://data.pyg.org/whl/torch-1.10.1+cu102.html 29 | pip install torch-sparse -f https://data.pyg.org/whl/torch-1.10.1+cu102.html 30 | pip install torch-geometric 31 | ``` 32 | 33 | ## Repository Structure 34 | 35 | - `doc`: image files used for rendering the README - not necessary for running the code. 36 | - `dataloader`: contains code for loading the data to the model. 37 | - `metrics`: utility scripts and functions for computing metrics/statistics. 38 | - `misc`: miscellaneous scripts and functions. 39 | - `models`: scripts relating to defining the model, the hyperparameters and I/O configuration. 40 | - `run_utils`: main engine and callbacks. 41 | 42 | 43 | ## Inference 44 | 45 | To see the full list of command line arguments for inference and explanation, run `python run_infer.py -h` and `python run_explainer.py -h`, respectively. We have also created two bash scripts to make it easier to run the code with the appropriate arguments. As an example, to run model inference enter: 46 | 47 | ``` 48 | python run_infer.py --gpu= --model_path= --data_dir= --data_info= --stats_dir= 49 | ``` 50 | 51 | You will see above that the `data_info` csv file will need to be incorporated as an argument. This will determine the label and which images to process. By default, the code will process images with values in the fold column equal to 3. If considering a test set, there will be a single 'fold' column named `test_info`. The `fold_nr` and `split_nr` can be added as additional arguments if considering cross validation, which determines the subset of the data from the csv file. 52 | 53 | 54 | ## Interactive Demo 55 | We have made an interactive demo to help visualise the output of our model. Note, this is not optimised for mobile phones and tablets. The demo was built using the TIAToolbox [tile server](https://tia-toolbox.readthedocs.io/en/latest/_autosummary/tiatoolbox.visualization.tileserver.TileServer.html). 56 | 57 | Check out the demo [here](https://iguana.dcs.warwick.ac.uk). 58 | 59 | In the demo, we provide multiple examples of WSI-level results. By default, glands are coloured by their node explanation score, indicating how much they contribute to the slide being predicted as abnormal. Glands can also be coloured by a specific feature using the drop-down menu on the right hand side. 60 | 61 | As you zoom in, smaller objects such as lumen and nuclei will become visible. These are accordingly coloured by their predicted class. For example, epithelial cells are coloured green and lymphocytes red. 62 | 63 | Each histological object can be toggled on/off by clicking the appropriate buton on the right hand side. Also, the colours and the opacity can be altered. 64 | 65 | To see which histological features are contributing to glands being flagged as abnormal, hover over the corresponding node. To view these nodes, toggle the graph on at the bottom-right of the screen. 66 | 67 | ![demo](https://user-images.githubusercontent.com/20071401/201095785-d7ce01c3-9652-425e-b1cd-38dd22672b81.gif) 68 | 69 | ## Sample Data and Weights 70 | We have released a small portion of data to allow researchers to get the code running and see how our graph data is structured. We also include two 'data info' csv files - one as if the data is to be used for cross validation and the other as an external test set. Click [here](https://drive.google.com/drive/folders/14MYdB-Acb93L6IdJfHnlWG7M2pkBn53I?usp=sharing) to download the sample dataset. 71 | 72 | To download the **IGUANA weights** trained on each fold of the UHCW dataset, click [here](https://drive.google.com/drive/folders/1J78dItPqMZcj2BsM4mf69x61wNAuJwKP?usp=share_link). To get the code running, you will also need the stats info used to standardise the input data. This includes the statistics (mean, mean, etc) of the features and the input node degree. Click [here](https://drive.google.com/drive/folders/1vcbf-9YrtoQUpFvf_l73iy5TupVWEdBF?usp=sharing) to download the stats info. 73 | ## License 74 | 75 | Code is under a GPL-3.0 license. See the [LICENSE](https://github.com/TissueImageAnalytics/cerberus/blob/master/LICENSE) file for further details. 76 | 77 | Model weights are licensed under [Attribution-NonCommercial-ShareAlike 4.0 International](http://creativecommons.org/licenses/by-nc-sa/4.0/). Please consider the implications of using the weights under this license. 78 | 79 | ## Cite this repository 80 | 81 | ``` 82 | @article{graham2022screening, 83 | title={Screening of normal endoscopic large bowel biopsies with artificial intelligence: a retrospective study}, 84 | author={Graham, Simon and Minhas, Fayyaz and Bilal, Mohsin and Ali, Mahmoud and Tsang, Yee Wah and Eastwood, Mark and Wahab, Noorul and Jahanifar, Mostafa and Hero, Emily and Dodd, Katherine and others}, 85 | journal={medRxiv}, 86 | year={2022}, 87 | publisher={Cold Spring Harbor Laboratory Press} 88 | } 89 | ``` 90 | -------------------------------------------------------------------------------- /dataloader/graph_loader.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import joblib 3 | import torch 4 | 5 | from torch_geometric.data import Data 6 | 7 | 8 | class FileLoader(torch.utils.data.Dataset): 9 | """Data loader for graph data. The loader will use features defined in features.yml.""" 10 | 11 | def __init__(self, file_list, feat_names, feat_stats, norm, data_clean): 12 | self.file_list = file_list 13 | self.feat_stats = feat_stats 14 | self.feat_names = feat_names 15 | 16 | if feat_stats is not None: 17 | self.mean_stats = np.array([feat_stats["mean"][k] for k in self.feat_names]) 18 | self.median_stats = np.array([feat_stats["median"][k] for k in self.feat_names]) 19 | self.std_stats = np.array([feat_stats["std"][k] for k in self.feat_names]) 20 | self.perc_25_stats = np.array([feat_stats["perc_25"][k] for k in self.feat_names]) 21 | self.perc_75_stats = np.array([feat_stats["perc_75"][k] for k in self.feat_names]) 22 | 23 | # get lower and upper bounds for clipping data (outlier removal) 24 | if data_clean == 'std': 25 | self.local_lower_bounds = self.mean_stats - 3 * self.std_stats 26 | self.local_upper_bounds = self.mean_stats + 3 * self.std_stats 27 | elif data_clean == 'iqr': 28 | iqr = self.perc_75_stats - self.perc_25_stats 29 | self.local_lower_bounds = self.perc_25_stats - 2 * iqr 30 | self.local_upper_bounds = self.perc_75_stats + 2 * iqr 31 | 32 | self.norm = norm 33 | 34 | assert ( 35 | self.norm == "standard" or self.norm == "robust" or self.norm == None 36 | ), "`norm` must be `standard` or `robust`." 37 | self.data_clean = data_clean 38 | assert ( 39 | self.data_clean == "std" or self.data_clean == "iqr" or self.data_clean == None 40 | ), "`data_clean` must be `std` or `iqr`." 41 | return 42 | 43 | def __len__(self): 44 | return len(self.file_list) 45 | 46 | def __getitem__(self, idx): 47 | path = self.file_list[idx] 48 | data = joblib.load(path) 49 | 50 | edge_index = np.array(data["edge_index"]) 51 | feats = data["local_feats"] 52 | 53 | nr_local_feats = len(self.feat_names) 54 | 55 | feats_sub = dict((k, feats[k]) for k in self.feat_names) # get subset of features 56 | feats_sub = np.array(list(feats_sub.values())).astype("float32") # convert to array 57 | feats_sub = np.transpose(feats_sub) # ensure NxF 58 | wsi_name = data["wsi_name"] 59 | obj_id = feats["obj_id"] 60 | 61 | # clean up data - deal with outliers! 62 | if self.data_clean is not None: 63 | # local feats 64 | clipped_feats = [] 65 | for idx in range(nr_local_feats): 66 | feat_single = feats_sub[:, idx] 67 | feat_single[feat_single > self.local_upper_bounds[idx]] = self.local_upper_bounds[idx] 68 | feat_single[feat_single < self.local_lower_bounds[idx]] = self.local_lower_bounds[idx] 69 | clipped_feats.append(feat_single) 70 | feats_sub = np.squeeze(np.dstack(clipped_feats), axis=0) 71 | 72 | # normalise the feature subset 73 | if self.norm == "standard": 74 | feats_sub = ((feats_sub - self.mean_stats) + 1e-8) / (self.std_stats + 1e-8) 75 | elif self.norm == "robust": 76 | feats_sub = ((feats_sub - self.median_stats) + 1e-8) / ((self.perc_75_stats - self.perc_25_stats) + 1e-8) 77 | 78 | label = np.array([data["label"]]) 79 | 80 | x = torch.Tensor(feats_sub).type(torch.float) 81 | edge_index = torch.Tensor(edge_index).type(torch.long) 82 | label = torch.Tensor(label).type(torch.float) 83 | 84 | return Data(x=x, edge_index=edge_index, y=label, obj_id=obj_id, wsi_name=wsi_name) 85 | -------------------------------------------------------------------------------- /dataset.py: -------------------------------------------------------------------------------- 1 | """Dataset info. Modify this with your own data directories for training IGUANA.""" 2 | 3 | import glob 4 | import pandas as pd 5 | import numpy as np 6 | 7 | 8 | class __CoBi(object): 9 | """Defines the Colon Biopsy (CoBi) graph dataset""" 10 | 11 | def __init__(self, fold_nr): 12 | 13 | file_ext = ".dat" 14 | root_dir = '/root/lsf_workspace/proc_slides/cobi/uhcw/graphs' 15 | 16 | self.stats_path = f"{root_dir}/stats" 17 | 18 | self.all_data = glob.glob(f"{root_dir}/data/*{file_ext}") 19 | 20 | # csv file - 1st column indicates the WSI name and each subsequent column gives the fold info 21 | # eg column 2 gives the info for fold1, column 3, gives the info for fold2, etc 22 | # for fold info: 1 denotes training, 2 denotes validation and 3 denotes testing 23 | # if the dataset is an independent test set, use 1 fold info column, with all cells set to 3 24 | fold_info = pd.read_csv(f"{root_dir}/uhcw_info.csv") 25 | 26 | wsi_names = np.array(fold_info.iloc[:, 0]) 27 | fold_info = np.array(fold_info.iloc[:, fold_nr]) 28 | 29 | wsi_train = wsi_names[fold_info==1] 30 | wsi_valid = wsi_names[fold_info==2] 31 | 32 | self.train_list = [] 33 | for wsi_name in wsi_train: 34 | self.train_list.append(f"{root_dir}/data/{wsi_name}{file_ext}") 35 | 36 | self.valid_list = [] 37 | for wsi_name in wsi_valid: 38 | self.valid_list.append(f"{root_dir}/data/{wsi_name}{file_ext}") 39 | 40 | 41 | def get_dataset(name): 42 | """Return a pre-defined dataset object associated with `name`""" 43 | if name.lower() == "cobi_fold1": 44 | return __CoBi(fold_nr=1) 45 | elif name.lower() == "cobi_fold2": 46 | return __CoBi(fold_nr=2) 47 | elif name.lower() == "cobi_fold3": 48 | return __CoBi(fold_nr=3) 49 | else: 50 | assert False, "Unknown dataset `%s`" % name 51 | -------------------------------------------------------------------------------- /doc/iguana.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/TissueImageAnalytics/iguana/c4cd93d0ed67c8d0bfc358f5837d1309963df841/doc/iguana.png -------------------------------------------------------------------------------- /environment.yml: -------------------------------------------------------------------------------- 1 | name: iguana 2 | channels: 3 | - conda-forge 4 | - defaults 5 | dependencies: 6 | - docopt=0.6.2=py37h06a4308_0 7 | - hdf5=1.12.1=nompi_h2386368_104 8 | - joblib=1.1.0=pyhd3eb1b0_0 9 | - matplotlib=3.5.1=py37h06a4308_1 10 | - matplotlib-base=3.5.1=py37ha18d171_1 11 | - numpy=1.21.5=py37h6c91a56_3 12 | - numpy-base=1.21.5=py37ha15fc14_3 13 | - pandas=1.3.5=py37h8c16a72_0 14 | - pillow=9.0.1=py37h22f2fdc_0 15 | - pip=21.2.2=py37h06a4308_0 16 | - progress=1.5=py37h06a4308_0 17 | - python=3.7.12=hb7a2778_100_cpython 18 | - pyyaml=6.0=py37h7f8727e_1 19 | - scikit-image=0.19.2=py37h51133e4_0 20 | - scikit-learn=1.0.2=py37h51133e4_1 21 | - scipy=1.7.3=py37hf2a6cf1_0 22 | - shapely=1.7.1=py37h1728cc4_0 23 | - termcolor=1.1.0=py37h06a4308_1 24 | - tqdm=4.64.0=py37h06a4308_0 25 | - yaml=0.2.5=h7b6447c_0 26 | - pip: 27 | - captum==0.5.0 28 | - imgaug==0.4.0 29 | - opencv-python==4.5.5.64 30 | - opencv-python-headless==4.5.5.64 31 | - seaborn==0.11.2 32 | - tensorboardx==2.5.1 33 | - tiatoolbox==1.3.0 34 | -------------------------------------------------------------------------------- /features.yml: -------------------------------------------------------------------------------- 1 | # below are the selected features used within the framework 2 | # look at the complete list of features obtained after running extract_feats.py in features_all.yml 3 | # you must not swap the ordering of the list! 4 | 5 | features: 6 | - gland-bam 7 | - gland-area 8 | - gland-dist1 9 | - lumen-number 10 | - lumen-gland-ratio 11 | - lumen-bam-mean 12 | - lumen-area-max 13 | - nuclei-gland-epi-count 14 | - nuclei-gland-lym-count 15 | - nuclei-gland-neut-count 16 | - nuclei-gland-eos-count 17 | - nuclei-inter-epi-mean 18 | - nuclei-inter-epi-std 19 | - nuclei-dist-boundary-mean 20 | - nuclei-dist-boundary-std 21 | - nuclei-dist-lumen-mean 22 | - nuclei-dist-lumen-std 23 | - nuclei-area-mean 24 | - nuclei-area-std 25 | - nuclei-focus-lym-prop 26 | - nuclei-focus-plas-prop 27 | - nuclei-focus-neut-prop 28 | - nuclei-focus-eos-prop 29 | - nuclei-focus-conn-prop 30 | - nuclei-inf-density -------------------------------------------------------------------------------- /features_all.yml: -------------------------------------------------------------------------------- 1 | features: 2 | - obj_id 3 | - gland-inflam-prop 4 | - gland-mucous-prop 5 | - gland-debris-prop 6 | - gland-normal-prop 7 | - gland-tumour-prop 8 | - gland-adipose-prop 9 | - gland-stroma-prop 10 | - gland-muscle-prop 11 | - gland-bam 12 | - gland-area 13 | - gland-perimeter 14 | - gland-equiv-diameter 15 | - gland-extent 16 | - gland-convex-area 17 | - gland-solidity 18 | - gland-major-axis-length 19 | - gland-minor-axis-length 20 | - gland-eccentricity 21 | - gland-orientation 22 | - gland-ellipse-centre-x 23 | - gland-ellipse-centre-y 24 | - gland-dist1 25 | - gland-dist2 26 | - gland-dist3 27 | - gland-dist4 28 | - gland-dist5 29 | - gland-dist6 30 | - lumen-number 31 | - lumen-gland-ratio 32 | - lumen-bam-min 33 | - lumen-bam-max 34 | - lumen-bam-mean 35 | - lumen-bam-std 36 | - lumen-area-min 37 | - lumen-area-max 38 | - lumen-area-mean 39 | - lumen-area-mtd 40 | - lumen-perimeter-min 41 | - lumen-perimeter-max 42 | - lumen-perimeter-mean 43 | - lumen-perimeter-mtd 44 | - lumen-equiv-diameter-min 45 | - lumen-equiv-diameter-max 46 | - lumen-equiv-diameter-mean 47 | - lumen-equiv-diameter-mtd 48 | - lumen-extent-min 49 | - lumen-extent-max 50 | - lumen-extent-mean 51 | - lumen-extent-mtd 52 | - lumen-convex-area-min 53 | - lumen-convex-area-max 54 | - lumen-convex-area-mean 55 | - lumen-convex-area-mtd 56 | - lumen-solidity-min 57 | - lumen-solidity-max 58 | - lumen-solidity-mean 59 | - lumen-solidity-mtd 60 | - lumen-major-axis-length-min 61 | - lumen-major-axis-length-max 62 | - lumen-major-axis-length-mean 63 | - lumen-major-axis-length-mtd 64 | - lumen-minor-axis-length-min 65 | - lumen-minor-axis-length-max 66 | - lumen-minor-axis-length-mean 67 | - lumen-minor-axis-length-mtd 68 | - lumen-eccentricity-min 69 | - lumen-eccentricity-max 70 | - lumen-eccentricity-mean 71 | - lumen-eccentricity-mtd 72 | - lumen-orientation-min 73 | - lumen-orientation-max 74 | - lumen-orientation-mean 75 | - lumen-orientation-mtd 76 | - lumen-ellipse-centre-x-min 77 | - lumen-ellipse-centre-x-max 78 | - lumen-ellipse-centre-x-mean 79 | - lumen-ellipse-centre-x-mtd 80 | - lumen-ellipse-centre-y-min 81 | - lumen-ellipse-centre-y-max 82 | - lumen-ellipse-centre-y-mean 83 | - lumen-ellipse-centre-y-mtd 84 | - nuclei-gland-epi-count 85 | - nuclei-gland-lym-count 86 | - nuclei-gland-plas-count 87 | - nuclei-gland-neut-count 88 | - nuclei-gland-eos-count 89 | - nuclei-gland-nuc-count 90 | - nuclei-inter-epi-min 91 | - nuclei-inter-epi-max 92 | - nuclei-inter-epi-mean 93 | - nuclei-inter-epi-std 94 | - nuclei-dist-boundary-min 95 | - nuclei-dist-boundary-max 96 | - nuclei-dist-boundary-mean 97 | - nuclei-dist-boundary-std 98 | - nuclei-dist-lumen-min 99 | - nuclei-dist-lumen-max 100 | - nuclei-dist-lumen-mean 101 | - nuclei-dist-lumen-std 102 | - nuclei-area-min 103 | - nuclei-area-max 104 | - nuclei-area-mean 105 | - nuclei-area-std 106 | - nuclei-perimeter-min 107 | - nuclei-perimeter-max 108 | - nuclei-perimeter-mean 109 | - nuclei-perimeter-std 110 | - nuclei-equiv-diameter-min 111 | - nuclei-equiv-diameter-max 112 | - nuclei-equiv-diameter-mean 113 | - nuclei-equiv-diameter-std 114 | - nuclei-extent-min 115 | - nuclei-extent-max 116 | - nuclei-extent-mean 117 | - nuclei-extent-std 118 | - nuclei-convex-area-min 119 | - nuclei-convex-area-max 120 | - nuclei-convex-area-mean 121 | - nuclei-convex-area-std 122 | - nuclei-solidity-min 123 | - nuclei-solidity-max 124 | - nuclei-solidity-mean 125 | - nuclei-solidity-std 126 | - nuclei-major-axis-length-min 127 | - nuclei-major-axis-length-max 128 | - nuclei-major-axis-length-mean 129 | - nuclei-major-axis-length-std 130 | - nuclei-minor-axis-length-min 131 | - nuclei-minor-axis-length-max 132 | - nuclei-minor-axis-length-mean 133 | - nuclei-minor-axis-length-std 134 | - nuclei-eccentricity-min 135 | - nuclei-eccentricity-max 136 | - nuclei-eccentricity-mean 137 | - nuclei-eccentricity-std 138 | - nuclei-orientation-min 139 | - nuclei-orientation-max 140 | - nuclei-orientation-mean 141 | - nuclei-orientation-std 142 | - nuclei-ellipse-centre-x-min 143 | - nuclei-ellipse-centre-x-max 144 | - nuclei-ellipse-centre-x-mean 145 | - nuclei-ellipse-centre-x-std 146 | - nuclei-ellipse-centre-y-min 147 | - nuclei-ellipse-centre-y-max 148 | - nuclei-ellipse-centre-y-mean 149 | - nuclei-ellipse-centre-y-std 150 | - colocalisation-neut-neut 151 | - colocalisation-neut-epi 152 | - colocalisation-neut-lym 153 | - colocalisation-neut-plas 154 | - colocalisation-neut-eos 155 | - colocalisation-neut-conn 156 | - colocalisation-epi-neut 157 | - colocalisation-epi-epi 158 | - colocalisation-epi-lym 159 | - colocalisation-epi-plas 160 | - colocalisation-epi-eos 161 | - colocalisation-epi-conn 162 | - colocalisation-lym-neut 163 | - colocalisation-lym-epi 164 | - colocalisation-lym-lym 165 | - colocalisation-lym-plas 166 | - colocalisation-lym-eos 167 | - colocalisation-lym-conn 168 | - colocalisation-plas-neut 169 | - colocalisation-plas-epi 170 | - colocalisation-plas-lym 171 | - colocalisation-plas-plas 172 | - colocalisation-plas-eos 173 | - colocalisation-plas-conn 174 | - colocalisation-eos-neut 175 | - colocalisation-eos-epi 176 | - colocalisation-eos-lym 177 | - colocalisation-eos-plas 178 | - colocalisation-eos-eos 179 | - colocalisation-eos-conn 180 | - colocalisation-conn-neut 181 | - colocalisation-conn-epi 182 | - colocalisation-conn-lym 183 | - colocalisation-conn-plas 184 | - colocalisation-conn-eos 185 | - colocalisation-conn-conn 186 | - nuclei-focus-lym-prop 187 | - nuclei-focus-plas-prop 188 | - nuclei-focus-neut-prop 189 | - nuclei-focus-eos-prop 190 | - nuclei-focus-conn-prop 191 | - nuclei-focus-lym-density 192 | - nuclei-focus-plas-density 193 | - nuclei-focus-neut-density 194 | - nuclei-focus-eos-density 195 | - nuclei-focus-conn-density 196 | - nuclei-inf-density 197 | -------------------------------------------------------------------------------- /metrics/stats_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn.metrics import roc_curve 3 | 4 | 5 | def get_sens_spec_metrics(targets, probs): 6 | """Get the specificity at different sensitivity cut-offs. 7 | 8 | Args: 9 | targets (list): list of true labels 10 | probs (list): list of predicted scores 11 | 12 | Returns: 13 | 14 | """ 15 | binary_labels = np.array(targets.copy()) 16 | specificity_at_95 = [] 17 | specificity_at_97 = [] 18 | specificity_at_98 = [] 19 | specificity_at_99 = [] 20 | specificity_at_100 = [] 21 | 22 | fpr, tpr, _ = roc_curve(binary_labels, probs) 23 | tnr = 1 - fpr 24 | specificity_at_95 = tnr[tpr >= 0.95][0] 25 | specificity_at_97 = tnr[tpr >= 0.97][0] 26 | specificity_at_98 = tnr[tpr >= 0.98][0] 27 | specificity_at_99 = tnr[tpr >= 0.99][0] 28 | specificity_at_100 = tnr[tpr >= 1.0][0] 29 | 30 | return specificity_at_95, specificity_at_97, specificity_at_98, specificity_at_99, specificity_at_100 -------------------------------------------------------------------------------- /misc/bam_utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | Utility functions for computing the BAM metric for a given object 3 | """ 4 | 5 | import numpy as np 6 | import cv2 7 | from skimage import draw 8 | from skimage.measure import regionprops 9 | import matplotlib.pyplot as plt 10 | import scipy 11 | from scipy.interpolate import InterpolatedUnivariateSpline 12 | 13 | 14 | 15 | def get_enclosing_ellipse(cnt): 16 | """Generates a discretised contour for the smallest enclosing ellipse 17 | 18 | Args: 19 | cnt: a 2-column matrix for discritised contour 20 | 21 | Returns: 22 | x: a 2-column matrix for discritised smallest enclosing ellipse 23 | 24 | """ 25 | hull = cv2.convexHull(cnt, clockwise=True).squeeze() # get the convex hull 26 | A, centre, radii, rotation = min_2d_ellipse(hull, 0.01) 27 | 28 | return ellipse_coordinates(A,centre,radii,rotation) 29 | 30 | 31 | def min_2d_ellipse(P, tol): 32 | """Find the minimum covering 2D ellipse which holds all the points 33 | 34 | Based on work by Nima Moshtagh 35 | http://www.mathworks.com/matlabcentral/fileexchange/9542 36 | and also by looking at: 37 | http://cctbx.sourceforge.net/current/python/scitbx.math.minimum_covering_ellipsoid.html 38 | 39 | Args: 40 | P: (d x N) dimnesional matrix containing N points in R^d. 41 | tol: error in the solution with respect to the optimal value. 42 | 43 | Returns: 44 | A: coordinates of the ellipse 45 | centre: centre of ellipse 46 | radii: radii of ellipse 47 | rotation: rotation of ellipse 48 | 49 | """ 50 | (N, d) = np.shape(P) 51 | d = float(d) 52 | 53 | # Q will be our working array 54 | Q = np.vstack([np.copy(P.T), np.ones(N)]) 55 | QT = Q.T 56 | 57 | # initialisations 58 | err = 1.0 + tol 59 | u = (1.0 / N) * np.ones(N) 60 | 61 | # Khachiyan Algorithm 62 | while err > tol: 63 | V = np.dot(Q, np.dot(np.diag(u), QT)) 64 | M = np.diag(np.dot(QT , np.dot(np.linalg.inv(V), Q))) # M the diagonal vector of an NxN matrix 65 | j = np.argmax(M) 66 | maximum = M[j] 67 | step_size = (maximum - d - 1.0) / ((d + 1.0) * (maximum - 1.0)) 68 | new_u = (1.0 - step_size) * u 69 | new_u[j] += step_size 70 | err = np.linalg.norm(new_u - u) 71 | u = new_u 72 | 73 | # centre of the ellipse 74 | centre = np.dot(P.T, u) 75 | 76 | # the A matrix for the ellipse 77 | A = np.linalg.inv( 78 | np.dot(P.T, np.dot(np.diag(u), P)) - 79 | np.array([[a * b for b in centre] for a in centre]) 80 | ) / d 81 | 82 | # Get the values we'd like to return 83 | U, s, rotation = np.linalg.svd(A) 84 | radii = 1.0/np.sqrt(s) 85 | 86 | return A, centre, radii, rotation 87 | 88 | 89 | def ellipse_coordinates(A, centre, radii, rotation, N=20.0): 90 | """Generates the coordinates of the minimum enclosing ellipse 91 | 92 | Args: 93 | A: a 2x2 matrix. 94 | centre: a 2D vector which represents the centre of the ellipsoid. 95 | radii: radii of ellipsoid generated from SVD 96 | rotation: rotation output by SVD 97 | N: the number of grid points for plotting the ellipse; Default: N = 20 98 | 99 | Returns: 100 | coordinates of ellipse 101 | 102 | """ 103 | steps = int(round((2.0 * np.pi + 1/N)*N)) 104 | theta = np.linspace(0.0, 2.0 * np.pi + 1/N, steps) 105 | 106 | # parametric equation of the ellipse 107 | x = radii[0] * np.cos(theta) 108 | y = radii[1] * np.sin(theta) 109 | 110 | # coordinate transform 111 | X = np.matmul(rotation,np.array([x,y])) 112 | X += np.expand_dims(centre.T, -1) 113 | return X.T 114 | 115 | 116 | def ellipse_to_circle(coords): 117 | """Transform ellipse to circle 118 | 119 | Args: 120 | coords: input coordinates of ellipse 121 | 122 | Returns: 123 | circle_coords: coordinates of circle (transformed ellipse) 124 | alpha: orientation in radians 125 | a: major axis length of input ellipse 126 | b: minor axis length of input ellipse 127 | 128 | """ 129 | # get the minimum x and y coordinates 130 | min_x = min(coords[:,1]) 131 | min_y = min(coords[:,0]) 132 | 133 | coords2 = np.round(coords - np.array([min_y, min_x])) 134 | 135 | # generate binary mask from coordinates 136 | shape = (int(max(coords2[:,1]) - min(coords2[:,1])), int(max(coords2[:,0]) - min(coords2[:,0]))) 137 | mask = coords2mask(coords2[:,1], coords2[:,0], shape) 138 | mask = mask.T 139 | # get properties from binary mask 140 | props = regionprops(mask.astype('int')) 141 | orientation = props[0].orientation 142 | a = props[0].major_axis_length 143 | b = props[0].minor_axis_length 144 | centroid = props[0].centroid 145 | 146 | centre = (centroid + np.array([min_y, min_x])) 147 | centre = np.array([centre[1], centre[0]]) 148 | alpha = -orientation 149 | 150 | circle_coords = apply_transform(coords, centre, alpha, a, b) 151 | 152 | return circle_coords, alpha, a, b 153 | 154 | 155 | def apply_transform(coords, centre, alpha, a, b): 156 | """Apply transformation to set of input coordinates 157 | 158 | Args: 159 | coords: input coordinates 160 | centre: centre of input coordinates 161 | alpha: orientation in radians 162 | a: major axis length of input 163 | b: major axis length of input 164 | 165 | Returns: 166 | transformed coordinates 167 | 168 | """ 169 | T1 = [[1, 0, -centre[1]], 170 | [0, 1, -centre[0]], 171 | [0, 0, 1 ]] 172 | 173 | T2 = [[1, 0, centre[1]], 174 | [0, 1, centre[0]], 175 | [0, 0, 1 ]] 176 | 177 | R = [[np.cos(alpha), -np.sin(alpha), 0], 178 | [np.sin(alpha), np.cos(alpha) , 0], 179 | [0 , 0 , 1]] 180 | 181 | S = [[1, 0, 0], 182 | [0, a/b, 0], 183 | [0, 0 , 1]] 184 | 185 | trans_matrix = np.matmul(T2,np.matmul(S,np.matmul(R,T1))) 186 | 187 | coords_ones = np.hstack([coords[:,1:2], coords[:,:1], np.ones([coords.shape[0],1])]) 188 | transformed_coords = np.matmul(trans_matrix, coords_ones[:, [1,0,2]].T).T 189 | return transformed_coords[:, [0,1]] 190 | 191 | 192 | def coords2mask(row_coords, col_coords, shape): 193 | """Convert coordinates to a binary mask 194 | 195 | Args: 196 | row_coords: coordinates in y direction 197 | col_coords: coordinates in x direction 198 | shape: shape of binary mask to output 199 | 200 | Returns: 201 | mask: binary mask 202 | 203 | """ 204 | fill_row_coords, fill_col_coords = draw.polygon(row_coords, col_coords, shape) 205 | mask = np.zeros(shape, dtype=np.bool) 206 | mask[fill_row_coords, fill_col_coords] = True 207 | return mask 208 | 209 | 210 | def best_alignment_metric(circle_coords, trans_coords, show_plots): 211 | """Compute BAM between two sets of coordinates as described in: 212 | 213 | Awan, Ruqayya, et al. "Glandular morphometrics for objective grading 214 | of colorectal adenocarcinoma histology images." Scientific reports 7.1 (2017): 1-12. 215 | 216 | Args: 217 | circle_coords: circle coordinates 218 | trans_coords: transformed object coordinates 219 | show_plots: show the coordinate plots at each stage of the implementation 220 | 221 | Returns: 222 | d: best alignment distance 223 | R: best index shift 224 | phi: best planar rotation 225 | 226 | """ 227 | # normalise the contour 228 | circle_coords_rs = resample_curve(circle_coords, 300) 229 | trans_coords_rs = resample_curve(trans_coords, 300) 230 | 231 | # convert to complex representation 232 | circle_complex = get_complex(circle_coords_rs) 233 | trans_complex = get_complex(trans_coords_rs) 234 | 235 | d, R, phi = bam_distance(circle_complex, trans_complex) 236 | 237 | if show_plots: 238 | plt.figure(figsize=(5, 5)) 239 | rotated_trans_coords = np.exp(1j*phi)*trans_complex 240 | rotated_and_cycled_trans_coords = np.roll(rotated_trans_coords, R) 241 | 242 | plt.subplot(2,2,1) 243 | plt.plot(trans_coords[:,0],trans_coords[:,1]) # plot cartesian transformed object coordinates 244 | 245 | plt.plot(trans_coords[:,0],trans_coords[:,1],'x', markersize=5) # plot x at ticks 246 | plt.plot(trans_coords[0,0],trans_coords[0,1],'.', markersize=30) # mark the first point 247 | 248 | plt.subplot(2,2,2) 249 | plt.plot(circle_coords[:,0],circle_coords[:,1],'r') # plot cartesian circle coordinates 250 | 251 | plt.plot(circle_coords[:,0],circle_coords[:,1],'rx', markersize=5) # plot x at ticks 252 | plt.plot(circle_coords[0,0],circle_coords[0,1],'r.', markersize=30) # mark the first point 253 | 254 | plt.subplot(2,2,3) 255 | plt.plot(circle_complex.real, circle_complex.imag,'r') # plot complex circle coordinates 256 | 257 | plt.plot(circle_complex.real, circle_complex.imag,'rx', markersize=5) 258 | plt.plot(circle_complex.real[0],circle_complex.imag[0],'r.', markersize=30) 259 | plt.plot(rotated_trans_coords.real, rotated_trans_coords.imag) # plot optimally rotated complex transformed object 260 | plt.plot(rotated_trans_coords.real, rotated_trans_coords.imag,'x', markersize=5) 261 | plt.plot(rotated_trans_coords.real[0], rotated_trans_coords.imag[0],'.', markersize=30) 262 | 263 | plt.subplot(2,2,4) 264 | plt.plot(circle_complex.real, circle_complex.imag,'r') # plot complex circle coordinates 265 | 266 | plt.plot(circle_complex.real, circle_complex.imag,'rx', markersize=5) 267 | plt.plot(circle_complex.real[0], circle_complex.imag[0],'r.', markersize=30) 268 | plt.plot(rotated_and_cycled_trans_coords.real, rotated_and_cycled_trans_coords.imag) # plot optimally rotated and cyclically reordered complex transformed object 269 | plt.plot(rotated_and_cycled_trans_coords.real, rotated_and_cycled_trans_coords.imag,'x', markersize=5) 270 | plt.plot(rotated_and_cycled_trans_coords.real[0],rotated_and_cycled_trans_coords.imag[1],'.', markersize=30) 271 | 272 | plt.show() 273 | 274 | return d, R, phi 275 | 276 | 277 | def bam_distance(u, v): 278 | """Rapidly computes the distance between curves u & v in the plane 279 | 280 | Args: 281 | u: complex vector with shape 1xN 282 | v: complex vector with shape 1xN 283 | 284 | Returns: 285 | d: best alignment distance 286 | R: best index shift 287 | phi: best planar rotation 288 | 289 | """ 290 | sum_u = np.sum(u.real**2 + u.imag**2) 291 | sum_v = np.sum(v.real**2 + v.imag**2) 292 | v_tmp = np.flipud(v) # ! why do this? 293 | Xcorr = np.fft.ifft(np.fft.fft(np.conj(u)) * np.fft.fft(v_tmp)) 294 | 295 | Xcorr2 = abs(Xcorr) 296 | A = np.max(Xcorr2) 297 | idx = np.argmax(Xcorr2) 298 | phi = np.arctan2(Xcorr[idx].imag, Xcorr[idx].real) 299 | R = idx 300 | 301 | summand = sum_u + sum_v - 2*A 302 | if summand < 0: 303 | summand = 0 304 | d = np.sqrt(summand) 305 | 306 | return d, R, phi 307 | 308 | 309 | def resample_curve(coords, N): 310 | """Resample the points on the input curve by interpolation. 311 | 312 | Args: 313 | coords: input coordinates 314 | N: number of sample points 315 | 316 | Returns: 317 | Xn: resampled coordinates 318 | 319 | """ 320 | coords = coords.T 321 | diff = coords[:,1:] - coords[:,:-1] # check same size (measure diff between points) 322 | dist = np.sqrt(diff[0,:]**2 + diff[1,:]**2) # calculate the magnitude between neighbouring coordinates 323 | dist = np.concatenate((np.array([0]),dist), axis=-1) 324 | cum_dist = np.cumsum(dist)/np.sum(dist) # cumulative sum of distances between neighbouring coordinates 325 | 326 | sample_points = np.linspace(1/N,1,N) 327 | interp = np.zeros([2,N]) 328 | for i in range(2): 329 | # generate interpolated points 330 | spline = InterpolatedUnivariateSpline(cum_dist,coords[i,:]) 331 | interp[i,:] = spline(sample_points) 332 | 333 | q = curve_to_q(interp) # include description 334 | qn = ProjectC(q) # include description 335 | Xn = q_to_curve(qn) 336 | 337 | return Xn 338 | 339 | 340 | def get_complex(curve): 341 | """Convert input curve to complex form and translate 342 | it such that it is centred at (0,0) 343 | 344 | Args: 345 | curve: input cartesian coordinates 346 | 347 | Returns: 348 | comp_curve_trans: translated complex curve 349 | 350 | """ 351 | comp_curve = curve[0,:]+1j*curve[1,:] 352 | cf = np.mean(comp_curve) 353 | comp_curve_trans = comp_curve-cf 354 | 355 | return comp_curve_trans 356 | 357 | 358 | def curve_to_q(p): 359 | """Include docstring 360 | 361 | Args: 362 | p: 363 | 364 | """ 365 | #! NEED TO UNDERSTAND WHAT IS GOING ON IN THIS FUNCTION and give appropriate comments 366 | N = p.shape[1] 367 | v = np.zeros([2,N]) 368 | for i in range(2): 369 | v[i,:] = np.gradient(p[i,:], 1/N) 370 | 371 | # unit velocity 372 | L = np.sqrt(np.sqrt(v[0,:]**2 + v[1,:]**2)) # check whether this is the right thing to do 373 | 374 | okPos = L > 1e-5 375 | q = v[:,okPos] / L[okPos] 376 | q[:,~okPos] = np.zeros([2, np.sum(~okPos)]) 377 | 378 | T = q.shape[1] 379 | s = np.linspace(0,1,T) 380 | val = np.trapz(np.sum(q*q, axis=0), s) 381 | return q / np.sqrt(val) 382 | 383 | 384 | def q_to_curve(q): 385 | """Include docstring 386 | 387 | Args: 388 | q: 389 | 390 | """ 391 | #! NEED TO UNDERSTAND WHAT IS GOING ON IN THIS FUNCTION and give appropriate comments 392 | T = q.shape[1] 393 | qnorm = np.sqrt(q[0,:]**2 + q[1,:]**2) 394 | 395 | p = np.zeros([2,T]) 396 | for i in range(2): 397 | p[i,:] = scipy.integrate.cumtrapz(q[i,:]*qnorm, initial=0)/T 398 | return p 399 | 400 | 401 | def ProjectC(q): 402 | """Include description 403 | 404 | Args: 405 | q: 406 | 407 | """ 408 | #! NEED TO UNDERSTAND WHAT IS GOING ON IN THIS FUNCTION and give appropriate comments 409 | T = q.shape[1] 410 | dt = 0.35 # what is this? - provide description 411 | 412 | epsilon = 1e-6 413 | count = 1 414 | res = np.ones([1,2]) 415 | 416 | s = np.linspace(0,1,T) 417 | tmp = np.trapz(np.sum(q*q, axis=0), s) 418 | qnew = q / np.sqrt(tmp) 419 | 420 | while np.linalg.norm(res, ord=2) > epsilon: 421 | if count > 300: 422 | break 423 | 424 | J = np.zeros([2,2]) 425 | for i in range(2): 426 | for j in range(1,2): 427 | J[i,j] = 3 * np.trapz(qnew[i,:]*qnew[j,:], s) 428 | 429 | J = J + J.T 430 | for i in range(2): 431 | J[i,i] = 3 * np.trapz(qnew[i,:]*qnew[i,:], s) 432 | 433 | J = J + np.identity(J.shape[0]) 434 | qnorm = np.sqrt(qnew[0,:]**2 + qnew[1,:]**2) 435 | 436 | G = np.zeros([2]) 437 | for i in range(2): 438 | G[i] = np.trapz(qnew[i,:]*qnorm, s) 439 | res = -G 440 | 441 | if np.linalg.norm(res, ord=2) < epsilon: 442 | break 443 | 444 | x = np.linalg.lstsq(J, res.T, rcond=None)[0] # solve system by least squares 445 | 446 | delG = form_basis_normal_A(qnew) 447 | 448 | tmp = x[0]*delG[0]*dt + x[1]*delG[1]*dt 449 | qnew = qnew + tmp 450 | 451 | count += 1 452 | 453 | tmp = np.trapz(np.sum(qnew*qnew, axis=0), s) 454 | return qnew / np.sqrt(tmp) 455 | 456 | 457 | def form_basis_normal_A(q): 458 | """Include docstring 459 | 460 | Args: 461 | q: 462 | 463 | """ 464 | #! NEED TO UNDERSTAND WHAT IS GOING ON IN THIS FUNCTION and give appropriate comments 465 | T = q.shape[1] 466 | 467 | e = np.identity(2) 468 | Ev = np.zeros([2,T,2]) 469 | for i in range(2): 470 | Ev[:,:,i] = np.tile(np.expand_dims(e[:,i],-1),(1,T)) 471 | 472 | qnorm = np.sqrt(q[0,:]**2 + q[1,:]**2) 473 | 474 | delG = [] 475 | for i in range(2): 476 | tmp1 = np.tile(q[i,:]/qnorm,(2,1)) 477 | tmp2 = np.tile(qnorm,(2,1)) 478 | delG.append(tmp1*q + tmp2*Ev[:,:,i]) 479 | 480 | return delG 481 | -------------------------------------------------------------------------------- /misc/feat_utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | Functions for feature extraction 3 | """ 4 | 5 | import cv2 6 | import numpy as np 7 | import pandas as pd 8 | import uuid 9 | import math 10 | 11 | from shapely.geometry import Polygon 12 | from scipy.spatial import distance, Voronoi, Delaunay 13 | from sklearn.neighbors import KDTree 14 | 15 | from .bam_utils import get_enclosing_ellipse, ellipse_to_circle, apply_transform, best_alignment_metric 16 | 17 | 18 | def get_contour_feats(cnt): 19 | """Get morphological features from input contours.""" 20 | 21 | # calculate some things useful later: 22 | m = cv2.moments(cnt) 23 | 24 | # ** regionprops ** 25 | Area = m["m00"] 26 | if Area > 0: 27 | Perimeter = cv2.arcLength(cnt, True) 28 | # bounding box: x,y,width,height 29 | BoundingBox = cv2.boundingRect(cnt) 30 | # centroid = m10/m00, m01/m00 (x,y) 31 | Centroid = (m["m10"] / m["m00"], m["m01"] / m["m00"]) 32 | 33 | # EquivDiameter: diameter of circle with same area as region 34 | EquivDiameter = np.sqrt(4 * Area / np.pi) 35 | # Extent: ratio of area of region to area of bounding box 36 | Extent = Area / (BoundingBox[2] * BoundingBox[3]) 37 | 38 | # CONVEX HULL stuff 39 | # convex hull vertices 40 | ConvexHull = cv2.convexHull(cnt) 41 | ConvexArea = cv2.contourArea(ConvexHull) 42 | # Solidity := Area/ConvexArea 43 | Solidity = Area / ConvexArea 44 | 45 | # ELLIPSE - determine best-fitting ellipse. 46 | centre, axes, angle = cv2.fitEllipse(cnt) 47 | MAJ = np.argmax(axes) # this is MAJor axis, 1 or 0 48 | MIN = 1 - MAJ # 0 or 1, minor axis 49 | # Note: axes length is 2*radius in that dimension 50 | MajorAxisLength = axes[MAJ] 51 | MinorAxisLength = axes[MIN] 52 | Eccentricity = np.sqrt(1 - (axes[MIN] / axes[MAJ]) ** 2) 53 | Orientation = angle 54 | EllipseCentre = centre # x,y 55 | 56 | else: 57 | Perimeter = 0 58 | EquivDiameter = 0 59 | Extent = 0 60 | ConvexArea = 0 61 | Solidity = 0 62 | MajorAxisLength = 0 63 | MinorAxisLength = 0 64 | Eccentricity = 0 65 | Orientation = 0 66 | EllipseCentre = [0, 0] 67 | 68 | return { 69 | "area": Area, 70 | "perimeter": Perimeter, 71 | "equiv-diameter": EquivDiameter, 72 | "extent": Extent, 73 | "convex-area": ConvexArea, 74 | "solidity": Solidity, 75 | "major-axis-length": MajorAxisLength, 76 | "minor-axis-length": MinorAxisLength, 77 | "eccentricity": Eccentricity, 78 | "orientation": Orientation, 79 | "ellipse-centre-x": EllipseCentre[0], 80 | "ellipse-centre-y": EllipseCentre[1], 81 | } 82 | 83 | 84 | def grab_cnts(input_info, ds_factor): 85 | """Get the contours from the input dictionary and remove excessive coordinates. 86 | 87 | input_info (list): List of input dictionaries 88 | ds_factor (int): Factor for removing coordinates 89 | 90 | """ 91 | output_dict = {} 92 | # first get the input dictionary for a single tissue region 93 | for tissue_idx, input_dict in input_info.items(): 94 | tissue_dict = {} 95 | for inst_id, info in input_dict.items(): 96 | cnt = info["contour"] 97 | cnt = cnt[::ds_factor, :] 98 | tissue_dict[inst_id] = cnt 99 | output_dict[tissue_idx] = tissue_dict 100 | 101 | return output_dict 102 | 103 | 104 | def grab_centroids_type(input_info): 105 | """Get the contours from the input dictionary and remove excessive coordinates. 106 | 107 | input_info (list): List of input dictionaries 108 | 109 | """ 110 | output_dict = {} 111 | # first get the input dictionary for a single tissue region 112 | for tissue_idx, input_dict in input_info.items(): 113 | tissue_dict = {} 114 | for inst_id, info in input_dict.items(): 115 | centroid = info["centroid"] 116 | type = info["type"] 117 | tissue_dict[inst_id] = {"centroid": centroid, "type": type} 118 | output_dict[tissue_idx] = tissue_dict 119 | 120 | return output_dict 121 | 122 | 123 | def convert_to_df(input_dict, return_type=True): 124 | """Convert input dict to dataframe.""" 125 | 126 | cx_list = [] 127 | cy_list = [] 128 | cnt_list = [] 129 | type_list = [] 130 | for info in input_dict.values(): 131 | centroid = info["centroid"] 132 | cx_list.append(centroid[0]) 133 | cy_list.append(centroid[1]) 134 | cnt_list.append(info["contour"]) 135 | if return_type: 136 | type_list.append(info["type"]) 137 | 138 | if return_type: 139 | df = pd.DataFrame(data={"cx": cx_list, "cy": cy_list, "contour": cnt_list, "type": type_list}) 140 | else: 141 | df = pd.DataFrame(data={"cx": cx_list, "cy": cy_list, "contour": cnt_list}) 142 | 143 | return df 144 | 145 | 146 | 147 | def filter_coords_msk(coords, mask, scale=1, mode="contour", label=None): 148 | """Filter input coordinates so that only coordinates within mask remain. 149 | 150 | Args: 151 | coords: input coordinates to filter 152 | mask: labelled tissue mask 153 | scale: processing resolution to mask scale factor 154 | mode: whether to check entire contour or jus the centroid 155 | 156 | """ 157 | 158 | unique_tissue = np.unique(mask).tolist()[1:] 159 | # populate empty dictionary - one per connected component in the tissue mask 160 | output_dict = {} 161 | for idx in unique_tissue: 162 | output_dict[idx] = {} 163 | 164 | # iterate over each object and check to see whether it is within the tissue 165 | for key, value in coords.items(): 166 | # if a label is provided, then only consider the contour if it is equal to the label 167 | if label is not None and value["type"] != label: 168 | continue 169 | contours = value[mode] 170 | if mode == "centroid": 171 | contours = [contours] 172 | in_mask = False 173 | for coord in contours: 174 | coord = coord.astype("float64") 175 | coord *= scale 176 | coord = np.rint(coord).astype("int32") 177 | # make sure coordinate is within the mask 178 | if coord[0] > 0 and coord[1] > 0 and coord[0] < mask.shape[1] and coord[1] < mask.shape[0]: 179 | if np.sum(mask[coord[1], coord[0]]) > 0: 180 | tissue_idx = int(mask[coord[1], coord[0]]) 181 | in_mask = True 182 | 183 | if in_mask: 184 | inst_uuid = uuid.uuid4().hex 185 | # add contour info to corresponding postion in output dictionary 186 | output_dict[tissue_idx][inst_uuid] = value 187 | 188 | return output_dict 189 | 190 | 191 | def filter_coords_msk2(coords, mask1, mask2, scale=1, mode="centroid"): 192 | """Filter input coordinates so that only coordinates within mask remain. 193 | This function returns an additional dictionary to `filter_coords_msk()`, 194 | which only considers objects in a second provided mask. 195 | 196 | Args: 197 | coords: input coordinates to filter 198 | mask1: labelled tissue mask 199 | mask2: binary mask for further sampling 200 | scale: processing resolution to mask scale factor 201 | mode: whether to check entire contour or jus the centroid 202 | 203 | Returns: 204 | output_dict1 (dict): dictionary of objects within mask 205 | output_dict2 (dict): subset of output_dict1 containing only objects that are also in mask2 206 | 207 | """ 208 | 209 | unique_tissue = np.unique(mask1).tolist()[1:] 210 | # populate empty dictionary - one per connected component in the tissue mask 211 | output_dict = {} 212 | output_dict2 = {} 213 | for idx in unique_tissue: 214 | output_dict[idx] = {} 215 | output_dict2[idx] = {} 216 | 217 | # iterate over each object and check to see whether it is within the tissue 218 | for key, value in coords.items(): 219 | 220 | contours = value[mode] 221 | if mode == "centroid": 222 | contours = [contours] 223 | in_mask = False 224 | in_mask2 = False 225 | for coord in contours: 226 | coord = coord.astype("float64") 227 | coord *= scale 228 | coord = np.rint(coord).astype("int32") 229 | # make sure coordinate is within the mask 230 | if coord[0] > 0 and coord[1] > 0 and coord[0] < mask1.shape[1] and coord[1] < mask1.shape[0]: 231 | if np.sum(mask1[coord[1], coord[0]]) > 0: 232 | tissue_idx = int(mask1[coord[1], coord[0]]) 233 | in_mask = True 234 | ###### 235 | if np.sum(mask2[coord[1], coord[0]]) > 0: 236 | # only consider non-epithelilal classes! 237 | if value["type"] != 2: 238 | in_mask2 = True 239 | if in_mask: 240 | inst_uuid = uuid.uuid4().hex 241 | # add contour info to corresponding postion in output dictionary 242 | output_dict[tissue_idx][inst_uuid] = value 243 | if in_mask2: 244 | output_dict2[tissue_idx][inst_uuid] = value 245 | 246 | 247 | return output_dict, output_dict2 248 | 249 | 250 | def filter_coords_cnt(df, contour, mode="contour", return_type=True): 251 | """Filter input coordaintes so that only coordinates within contour remain. 252 | 253 | Args: 254 | df: input dataframe containing coordinates 255 | contour: contours for which to check 256 | mode: whether to check entire contour or jus the centroid 257 | 258 | """ 259 | 260 | assert mode in [ 261 | "contour", 262 | "centroid", 263 | ], "`mode` must either be `contour` or `centroid`." 264 | 265 | output_dict = {} 266 | # iterate over each object and check to see whether it is within the contour (gland in this case) 267 | for idx, row in df.iterrows(): 268 | if mode == "centroid": 269 | cnts = [[row["cx"], row["cy"]]] 270 | else: 271 | cnts = row["contour"] 272 | count = 0 273 | total = 0 274 | for cnt in cnts: 275 | total += 1 276 | cnt = np.rint(cnt).astype("int") 277 | contour = contour.astype("int") 278 | result = cv2.pointPolygonTest(contour, (int(cnt[0]), int(cnt[1])), False) 279 | # check if coordinate lies on or inside the gland contour 280 | if result != -1: 281 | count += 1 282 | # make sure at least 95% of contours are within / on the gland 283 | if count / total > 0.95: 284 | inst_uuid = uuid.uuid4().hex 285 | # add contour info to corresponding postion in output dictionary 286 | if return_type: 287 | output_dict[inst_uuid] = { 288 | "centroid": [row["cx"], row["cy"]], 289 | "contour": row["contour"], 290 | "type": row["type"] 291 | } 292 | else: 293 | output_dict[inst_uuid] = { 294 | "centroid": [row["cx"], row["cy"]], 295 | "contour": row["contour"] 296 | } 297 | 298 | return output_dict 299 | 300 | 301 | def points_list_pairwise_edt(list_a, list_b): 302 | """Get the parwise euclidean distance between two lists of contours. 303 | 304 | Args: 305 | list_a: first list of x,y contour coordinates, with shape Nx2. 306 | list_b: second list of x,y contour coordinates, with shape Nx2. 307 | 308 | Returns: 309 | pairwise euclidean distance. 310 | 311 | """ 312 | pix_x_wrt_cnt_x = np.subtract.outer(list_a[:, 0], list_b[:, 0]) # INNER x CNT 313 | pix_y_wrt_cnt_y = np.subtract.outer(list_a[:, 1], list_b[:, 1]) 314 | pix_x_wrt_cnt_x = pix_x_wrt_cnt_x.flatten() 315 | pix_y_wrt_cnt_y = pix_y_wrt_cnt_y.flatten() 316 | pix_cnt_dst = np.sqrt(pix_x_wrt_cnt_x ** 2 + pix_y_wrt_cnt_y ** 2) # INNER x CNT 317 | pix_cnt_dst = np.reshape(pix_cnt_dst, (list_a.shape[0], list_b.shape[0])) 318 | return pix_cnt_dst 319 | 320 | 321 | def get_centroid(cnt): 322 | """Get the centroid of a set of contour coordinates. 323 | 324 | Args: 325 | cnt: input contour coordinates. 326 | 327 | Returns: 328 | x and y centroid coordinates. 329 | 330 | """ 331 | M = cv2.moments(cnt) 332 | cX = int(M["m10"] / M["m00"]) 333 | cY = int(M["m01"] / M["m00"]) 334 | 335 | return cX, cY 336 | 337 | 338 | def get_boundary_distance(cnts, centroid): 339 | """"Get the euclidean distance of a centroid to the nearest contour boundary. 340 | 341 | Args: 342 | cnts: coordinates of contour 343 | centroid: coordinates of centroid 344 | 345 | """ 346 | min_dst = 100000 347 | for cnt in cnts: 348 | dst = math.sqrt((cnt[0] - centroid[0]) ** 2 + (cnt[1] - centroid[1]) ** 2) 349 | if dst < min_dst: 350 | min_dst = dst 351 | 352 | return min_dst 353 | 354 | 355 | def get_lumen_distance(lumen_info, centroid): 356 | """"Get the euclidean distance of a centroid to the nearest lumen boundary. 357 | 358 | Args: 359 | lumen_info: dict of lumen info 360 | centroid: coordinates of centroid 361 | 362 | """ 363 | 364 | dst_list = [] 365 | for _, info in lumen_info.items(): 366 | cnts = info["contour"] 367 | min_dst = 100000 368 | for cnt in cnts: 369 | dst = math.sqrt((cnt[0] - centroid[0]) ** 2 + (cnt[1] - centroid[1]) ** 2) 370 | if dst < min_dst: 371 | min_dst = dst 372 | dst_list.append(min_dst) 373 | 374 | if len(dst_list) > 0: 375 | return min(dst_list) 376 | else: 377 | return np.nan 378 | 379 | 380 | def get_voronoi_feats(coords): 381 | """Get voronoi diagram features.""" 382 | vor = Voronoi(coords) 383 | regions = vor.regions 384 | vertices = vor.vertices 385 | 386 | area = [] 387 | perim = [] 388 | for region in regions: 389 | if len(region) > 0: 390 | if -1 in region: 391 | region.remove(-1) 392 | if len(region) > 2: 393 | # get the polygon area and perimeter info 394 | region_coords = vertices[region, :].astype('int') 395 | xy_coords = list(zip(region_coords[:, 0], region_coords[:, 1])) 396 | pgon = Polygon(xy_coords) # Assuming the OP's x,y coordinates 397 | area.append(pgon.area) 398 | perim.append(pgon.length) 399 | else: 400 | area.append(0) 401 | perim.append(0) 402 | 403 | return area, perim 404 | 405 | 406 | def find_neighbors(pindex, triang): 407 | """Get the neighbouring vertices of a given vertex from Delaunay triangulation.""" 408 | return triang.vertex_neighbor_vertices[1][triang.vertex_neighbor_vertices[0][pindex]:triang.vertex_neighbor_vertices[0][pindex+1]] 409 | 410 | 411 | def get_delaunay_feats(coords): 412 | """Get Delaunay triangulation features.""" 413 | tess = Delaunay(coords) 414 | simps = tess.simplices 415 | 416 | area = [] 417 | perim = [] 418 | min_dst = [] 419 | max_dst = [] 420 | max_min_dst = [] 421 | for idx, simp in enumerate(simps): 422 | 423 | tri_coords = coords[simp, :] 424 | # just being sure that 3 coordinates exist! 425 | if tri_coords.shape[0] == 3: 426 | # get the edge distance info 427 | dst_list = [] 428 | for idx in range(3): 429 | if idx+1 == 3: 430 | idx2 = 0 431 | else: 432 | idx2 = idx+1 433 | a = tri_coords[idx, :] 434 | b = tri_coords[idx2, :] 435 | dst_list.append(distance.euclidean(a, b)) 436 | min_dst.append(min(dst_list)) 437 | max_dst.append(max(dst_list)) 438 | max_min_dst.append(max(dst_list) / min(dst_list)) 439 | 440 | # get the polygon area and perimeter info 441 | xy_coords = list(zip(tri_coords[:, 0], tri_coords[:, 1])) 442 | pgon = Polygon(xy_coords) # Assuming the OP's x,y coordinates 443 | area.append(pgon.area) 444 | perim.append(pgon.length) 445 | 446 | return area, perim, min_dst, max_dst, max_min_dst 447 | 448 | 449 | def get_k_nearest_from_contour(contour, obj_kdtree, labels, centroids, k=175, nr_samples=None): 450 | """ Get the K nearest nuclei from the contour. 451 | 452 | Args: 453 | contour: input contour 454 | obj_kdtree (sklearn.neighbors.KDTree): KDTree of nuclei centroids 455 | labels: object labels (same index as the tree) 456 | centroids: objects coordinates (same index as the tree) 457 | k: return k nearest objects 458 | 459 | Returns: 460 | output_dst: distances of nearest objects 461 | output_labs: labels of nearest objects 462 | output_cents: coordinates of nearest objects 463 | """ 464 | #! This needs optimisation. Consider geopandas.sindex.SpatialIndex.nearest 465 | if nr_samples < k: 466 | k = nr_samples 467 | 468 | contour = np.array(contour) 469 | dist, inds = obj_kdtree.query(contour, k=k) 470 | 471 | distances = {} 472 | if contour.shape[0] > 1: 473 | unique_inds = np.unique(inds).tolist() 474 | for ind in unique_inds: 475 | dist_subset = dist[inds == ind] 476 | min_dst = np.min(dist_subset) 477 | distances[ind] = min_dst 478 | else: 479 | for index in range(dist.shape[-1]): 480 | distances[inds[0, index]] = dist[0, index] 481 | 482 | distances = {k: v for k, v in sorted(distances.items(), key=lambda item: item[1])} 483 | 484 | if len(list(distances.keys())) < k: 485 | k = len(list(distances.keys())) 486 | 487 | # output dict is {type: distance} 488 | output_labs = [] 489 | output_cents = [] 490 | output_dst = [] 491 | dist_keys = list(distances.keys()) 492 | dist_values = list(distances.values()) 493 | for idx in range(k): 494 | grab_idx = dist_keys[idx] 495 | output_labs.append(labels[grab_idx]) 496 | output_cents.append(centroids[grab_idx]) 497 | output_dst.append(dist_values[idx]) 498 | 499 | return output_dst, output_labs, output_cents 500 | 501 | 502 | def get_nearest_within_radius(centroid, obj_kdtree, labels, r=50, nr_types=6): 503 | """ Get the objects within a fixed radius. 504 | 505 | Args: 506 | contour: input contour 507 | obj_kdtree (sklearn.neighbors.KDTree): KDTree of nuclei centroids 508 | labels: object labels (same index as the tree) 509 | r: return objects within fixed radius 510 | 511 | Returns: 512 | output_dict: frequencies of different nuclei types within radius of r 513 | """ 514 | 515 | centroid = np.array(centroid) 516 | inds = obj_kdtree.query_radius(centroid, r=r, return_distance=False) 517 | inds = np.squeeze(inds).tolist() 518 | 519 | lab_list = [] 520 | for ind in inds: 521 | lab_list.append(labels[ind]) 522 | 523 | # output dict format: {type: frequency} 524 | output_dict = {} 525 | for idx in range(nr_types): 526 | output_dict[idx+1] = lab_list.count(idx+1) 527 | 528 | return output_dict 529 | 530 | 531 | def get_kdtree(input_dict): 532 | """Convert input dictionary of results to KDTree. 533 | 534 | Args: 535 | input_dict: results dictionary. 536 | 537 | Returns: 538 | centroid_kdtree (sklearn.neighbors.KDTree): KD-Tree of object centroids. 539 | 540 | """ 541 | centroid_list = [] 542 | label_list = [] 543 | for key, values in input_dict.items(): 544 | centroid_list.append(values["centroid"]) 545 | label_list.append(values["type"]) 546 | 547 | if len(centroid_list) > 0: 548 | centroid_array = np.array(centroid_list) 549 | centroid_kdtree = KDTree(centroid_array) 550 | else: 551 | centroid_kdtree = None 552 | label_list = None 553 | 554 | return centroid_kdtree, label_list, centroid_list 555 | 556 | 557 | def inter_epi_dst(input_dict, obj_kdtree, labels, lab=2): 558 | """get the stats for distances between epithelial cells. 559 | 560 | Args: 561 | input_dict: 562 | obj_kdtree (sklearn.neighbors.KDTree): KDTree of nuclei centroids in gland 563 | labels: labels of each nucleus 564 | lab: label to consider 565 | 566 | Returns: 567 | mean and std of inter-nuclei distances 568 | 569 | """ 570 | centroid_list = [] 571 | for values in input_dict.values(): 572 | # find nearest object 573 | # only for epi 574 | if values["type"] == lab: 575 | centroid_list.append(values["centroid"]) 576 | 577 | if len(centroid_list) > 1: 578 | dst_list = [] 579 | for centroid in centroid_list: 580 | centroid = np.reshape(centroid, (1, 2)) 581 | dst, _ = obj_kdtree.query(centroid, k=2) 582 | dst_list.append(dst[0,1]) 583 | else: 584 | dst_list = [0] 585 | 586 | return dst_list 587 | 588 | 589 | def get_dst_matrix(list_cnts, sorted=False): 590 | """Get the distance matrix. Measures the distance between all object contours. 591 | 592 | Args: 593 | list_cnts: input list of object contour coordinates. 594 | 595 | 596 | Returns: 597 | dst_matrix: NxN array of matrix of distances between objects. 598 | 599 | """ 600 | nr_objs = len(list_cnts) 601 | dst_matrix = np.zeros([nr_objs, nr_objs]) 602 | 603 | for i in range(nr_objs): 604 | cnt1 = list_cnts[i] 605 | for j in range(nr_objs): 606 | if i != j: 607 | cnt2 = list_cnts[j] 608 | dist = points_list_pairwise_edt(cnt1, cnt2) 609 | # distance between objects is the min dist between 2 contours 610 | dst_matrix[i, j] = np.min(dist) 611 | 612 | if sorted: 613 | dst_matrix = np.sort(dst_matrix, axis=-1) 614 | 615 | return dst_matrix 616 | 617 | 618 | def get_bam(cnt, centroid): 619 | """Get the BAM (best alignment metric).""" 620 | # get enclosing ellipse 621 | ellipse_coords = get_enclosing_ellipse(cnt) 622 | # transform ellipse to circle 623 | circle_coords, alpha, a, b = ellipse_to_circle(ellipse_coords) 624 | # transform original object with same transformation 625 | trans_coords = apply_transform(cnt, centroid, alpha, a, b) 626 | # compute best alignment metric (BAM) 627 | bam_distance, _, _ = best_alignment_metric( 628 | circle_coords, trans_coords, show_plots=False 629 | ) 630 | 631 | return bam_distance 632 | 633 | 634 | def get_tissue_region_info_patch(cnt, mask, relax_pix=50, ds_factor=0.125): 635 | """Get the patch around a given contour by considering the relaxed bounding box.""" 636 | 637 | # contour is at 0.5 mpp 638 | # mask is at 4.0 mpp 639 | # relax_pix operates at mask level 640 | 641 | # first convert contours to correct scale 642 | cnt = cnt * ds_factor 643 | cnt = cnt.astype('int') 644 | x,y,w,h = cv2.boundingRect(cnt) 645 | 646 | x = x - relax_pix 647 | y = y - relax_pix 648 | w = w + 2*relax_pix 649 | h = h + 2*relax_pix 650 | 651 | if x < 0: 652 | x = 0 653 | if y < 0: 654 | y = 0 655 | if x > (mask.shape[1] - relax_pix): 656 | x = mask.shape[1] - relax_pix 657 | if y > (mask.shape[0] - relax_pix): 658 | y = mask.shape[0] - relax_pix 659 | 660 | patch = mask[y: y+h, x: x+w] 661 | 662 | return patch 663 | 664 | 665 | def get_patch_prop(region_info_patch, total_pix, labs): 666 | """Get the proportion of a certain tissue type in a labelled input patch.""" 667 | output_list = [] 668 | for lab in labs: 669 | lab_tmp = region_info_patch == lab 670 | nr_pix = np.sum(lab_tmp) 671 | 672 | if total_pix == 0: 673 | output_list.append(0) 674 | else: 675 | output_list.append(nr_pix / total_pix) 676 | 677 | return output_list -------------------------------------------------------------------------------- /misc/utils.py: -------------------------------------------------------------------------------- 1 | import glob 2 | import os 3 | import shutil 4 | import joblib 5 | import numpy as np 6 | import cv2 7 | import torch 8 | 9 | import sys 10 | sys.path.append("../") 11 | 12 | 13 | def normalize(mask, dtype=np.uint8): 14 | return (255 * mask / np.amax(mask)).astype(dtype) 15 | 16 | 17 | def get_bounding_box(img): 18 | rows = np.any(img, axis=1) 19 | cols = np.any(img, axis=0) 20 | rmin, rmax = np.where(rows)[0][[0, -1]] 21 | cmin, cmax = np.where(cols)[0][[0, -1]] 22 | # due to python indexing, need to add 1 to max 23 | # else accessing will be 1px in the box, not out 24 | rmax += 1 25 | cmax += 1 26 | return [rmin, rmax, cmin, cmax] 27 | 28 | 29 | def cropping_center(x, crop_shape, batch=False): 30 | orig_shape = x.shape 31 | if not batch: 32 | h0 = int((orig_shape[0] - crop_shape[0]) * 0.5) 33 | w0 = int((orig_shape[1] - crop_shape[1]) * 0.5) 34 | x = x[h0:h0 + crop_shape[0], w0:w0 + crop_shape[1]] 35 | else: 36 | h0 = int((orig_shape[1] - crop_shape[0]) * 0.5) 37 | w0 = int((orig_shape[2] - crop_shape[1]) * 0.5) 38 | x = x[:, h0:h0 + crop_shape[0], w0:w0 + crop_shape[1]] 39 | return x 40 | 41 | 42 | def rm_n_mkdir(dir_path): 43 | if (os.path.isdir(dir_path)): 44 | shutil.rmtree(dir_path) 45 | os.makedirs(dir_path) 46 | 47 | 48 | def get_files(data_dir_list, data_ext): 49 | """Given a list of directories containing data with extention 'data_ext', 50 | generate a list of paths for all files within these directories 51 | 52 | """ 53 | data_files = [] 54 | for sub_dir in data_dir_list: 55 | files_list = glob.glob(sub_dir + '/*' + data_ext) 56 | files_list.sort() # ensure same order 57 | data_files.extend(files_list) 58 | 59 | return data_files 60 | 61 | 62 | def remap_label(pred, by_size=False, ds_factor=None): 63 | """Rename all instance id so that the id is contiguous i.e [0, 1, 2, 3] 64 | not [0, 2, 4, 6]. The ordering of instances (which one comes first) 65 | is preserved unless by_size=True, then the instances will be reordered 66 | so that bigger nucler has smaller ID 67 | 68 | Args: 69 | pred : the 2d array contain instances where each instances is marked 70 | by non-zero integer 71 | by_size : renaming with larger nuclei has smaller id (on-top) 72 | 73 | """ 74 | pred_id = list(np.unique(pred)) 75 | pred_id.remove(0) 76 | if len(pred_id) == 0: 77 | return pred # no label 78 | if by_size: 79 | pred_size = [] 80 | for inst_id in pred_id: 81 | size = (pred == inst_id).sum() 82 | pred_size.append(size) 83 | # sort the id by size in descending order 84 | pair_list = zip(pred_id, pred_size) 85 | pair_list = sorted(pair_list, key=lambda x: x[1], reverse=True) 86 | pred_id, pred_size = zip(*pair_list) 87 | 88 | new_pred = np.zeros(pred.shape, np.int32) 89 | for idx, inst_id in enumerate(pred_id): 90 | new_pred[pred == inst_id] = idx + 1 91 | 92 | return new_pred 93 | 94 | 95 | def get_inst_centroid(inst_map): 96 | inst_centroid_list = [] 97 | inst_id_list = list(np.unique(inst_map)) 98 | for inst_id in inst_id_list[1:]: # avoid 0 i.e background 99 | mask = np.array(inst_map == inst_id, np.uint8) 100 | inst_moment = cv2.moments(mask) 101 | inst_centroid = [(inst_moment["m10"] / inst_moment["m00"]), 102 | (inst_moment["m01"] / inst_moment["m00"])] 103 | inst_centroid_list.append(inst_centroid) 104 | return np.array(inst_centroid_list) 105 | 106 | 107 | def get_local_feat_stats(file_list): 108 | """Calculate mean and standard deviation from features- used for normalisation 109 | of input features before input to GCN. 110 | 111 | Args: 112 | file_list: list of .dat files containing features 113 | 114 | """ 115 | feats_tmp = joblib.load(file_list[0]) 116 | feat_names = list(feats_tmp["local_feats"].keys()) 117 | del feats_tmp 118 | 119 | print("Getting local feature statistics...") 120 | output_dict = {} 121 | mean_dict = {} 122 | median_dict = {} 123 | std_dict = {} 124 | perc_25_dict = {} 125 | perc_75_dict = {} 126 | for feat_name in feat_names: 127 | if feat_name == "obj_id": 128 | mean = 0.0 129 | median = 0.0 130 | std = 0.0 131 | perc_25 = 0.0 132 | perc_75 = 0.0 133 | else: 134 | accumulated_feats_tmp = [] 135 | for filepath in file_list: 136 | feats = joblib.load(filepath)["local_feats"] # hard assumption on .dat file 137 | feats = feats[feat_name].tolist() 138 | accumulated_feats_tmp.extend(np.float32(feats)) 139 | mean = float(np.nanmean(np.array(accumulated_feats_tmp))) 140 | median = float(np.nanmedian(np.array(accumulated_feats_tmp))) 141 | std = float(np.nanstd(np.array(accumulated_feats_tmp))) 142 | perc_25 = float(np.nanpercentile(np.array(accumulated_feats_tmp), q=25)) 143 | perc_75 = float(np.nanpercentile(np.array(accumulated_feats_tmp), q=75)) 144 | 145 | mean_dict[feat_name] = mean 146 | median_dict[feat_name] = median 147 | std_dict[feat_name] = std 148 | perc_25_dict[feat_name] = perc_25 149 | perc_75_dict[feat_name] = perc_75 150 | 151 | output_dict["mean"] = mean_dict 152 | output_dict["median"] = median_dict 153 | output_dict["std"] = std_dict 154 | output_dict["perc_25"] = perc_25_dict 155 | output_dict["perc_75"] = perc_75_dict 156 | 157 | return output_dict 158 | 159 | def get_pna_deg(file_list, feat_names, save_path): 160 | """Compute the maximum in-degree in the training data. Only needed for PNA Conv.""" 161 | 162 | from dataloader.graph_loader import FileLoader 163 | from torch_geometric.utils import degree 164 | 165 | print("Computing maximum node degree for PNA conv...") 166 | 167 | input_dataset = FileLoader(file_list, feat_names, feat_stats=None, norm=None, data_clean=None) 168 | 169 | max_degree = -1 170 | for data in input_dataset: 171 | d = degree(data.edge_index[1], num_nodes=data.num_nodes, dtype=torch.long) 172 | max_degree = max(max_degree, int(d.max())) 173 | 174 | # Compute the in-degree histogram tensor 175 | deg = torch.zeros(max_degree + 1, dtype=torch.long) 176 | for data in input_dataset: 177 | d = degree(data.edge_index[1], num_nodes=data.num_nodes, dtype=torch.long) 178 | deg += torch.bincount(d, minlength=deg.numel()) 179 | 180 | np.save(f"{save_path}/node_deg.npy", deg) 181 | 182 | 183 | def ranking_loss(pred, true): 184 | """Ranking loss. 185 | 186 | Args: 187 | pred: prediction array 188 | true: ground truth array 189 | 190 | """ 191 | loss = 0 192 | c = 0 193 | z = torch.Tensor([0]).to("cuda").type(torch.float32) 194 | for i in range(len(true)-1): 195 | for j in range(i+1, len(true)): 196 | if true[i] != true[j]: 197 | c += 1 198 | dz = pred[i,-1]-pred[j,-1] 199 | dy = true[i]-true[j] 200 | loss += torch.max(z, 1.0-dy*dz) 201 | return loss/c 202 | 203 | 204 | def refine_files(file_list, wsi_info): 205 | """Remove unwanted categories.""" 206 | 207 | wsi_info["Diagnostic Category"].replace( 208 | { 209 | "Normal ": "Normal", 210 | "Abnormal: Non-Neoplastic ": "Abnormal: Non-Neoplastic", 211 | "Abnormal: Neoplastic ": "Abnormal: Neoplastic", 212 | }, 213 | inplace=True, 214 | ) 215 | 216 | refined_list = [] 217 | for filename in file_list: 218 | wsiname = os.path.basename(filename) 219 | wsiname = wsiname[:-4] 220 | 221 | subset = wsi_info.loc[wsi_info["WSI no"] == wsiname] 222 | diagnosis = np.array(subset["Specific Diagnosis"])[0] 223 | if not isinstance(diagnosis, float): 224 | diagnosis = diagnosis.lower() 225 | 226 | category = np.array(subset["Diagnostic Category"])[0] 227 | if category == "Normal": 228 | category = 0 229 | if category == "Abnormal: Non-Neoplastic": 230 | category = 1 231 | if category == "Abnormal: Neoplastic": 232 | category = 2 233 | 234 | # clean up 235 | diagnosis = diagnosis.replace(",", " ") 236 | diagnosis = diagnosis.replace(":", " ") 237 | diagnosis = diagnosis.replace(".", " ") 238 | diagnosis = diagnosis.replace("?", " ") 239 | diagnosis = diagnosis.replace("-", " ") 240 | diagnosis_split = diagnosis.split(" ") 241 | 242 | if "spirochetosis" not in diagnosis_split and "melanosis" not in diagnosis_split and "malanosis" not in diagnosis_split: 243 | refined_list.append(filename) 244 | 245 | return refined_list 246 | 247 | 248 | def get_focus_tissue(wsi_path, tissuetype, results_gland, nr_classes=9, mode="lp", ds_factor=8): 249 | """Get non-glandular area within the issue which is considered for cell quantification. For 250 | biopsies, this is the lamina propria - otherwise, consider the entire non-glandular tissue area! 251 | 252 | Args: 253 | wsi_path: path to the original WSI 254 | tissetype (array): tissue type prediction 255 | results_gland (dict): gland segmentation results 256 | nr_classes (int): Number of classes considered by tissue type prediction 257 | mode (str): if `lp` then consider lamina propria area - otherwise consider entire tissue 258 | ds_factor (int): factor for converting gland segmentation coordinates to appropriate resolution. 259 | 260 | Returns: 261 | out_focus (array): binary map containing tissue region of interest 262 | 263 | """ 264 | from scipy.ndimage import measurements 265 | from skimage.morphology import remove_small_holes 266 | from skimage.morphology.misc import remove_small_objects 267 | 268 | from tiatoolbox.wsicore.wsireader import WSIReader 269 | 270 | wsi_handler = WSIReader.open(wsi_path) 271 | # in XY 272 | wsi_thumb = wsi_handler.slide_thumbnail(resolution=4.0, units="mpp") 273 | wsi_blur = cv2.GaussianBlur( 274 | cv2.cvtColor(wsi_thumb, cv2.COLOR_BGR2GRAY), (3, 3), 0) 275 | 276 | tissuetype = cv2.resize(tissuetype, (wsi_thumb.shape[1], wsi_thumb.shape[0])) 277 | del wsi_thumb 278 | 279 | out_focus = np.zeros([tissuetype.shape[0], tissuetype.shape[1]]) 280 | 281 | if mode == "lp": 282 | # only consider tumour and glandular regions if lamina propria 283 | for i in range(nr_classes): 284 | if i == 1 or i == 2: 285 | tmp = tissuetype == i 286 | out_focus[tmp] = 1 287 | else: 288 | tmp = tissuetype == i 289 | out_focus[tmp] = 0 290 | else: 291 | # consider all tissue 292 | out_focus[out_focus > 0] = 1 293 | 294 | out_focus = remove_small_holes(out_focus.astype("bool"), area_threshold=3900) 295 | out_focus = out_focus.astype("uint8") 296 | 297 | out_focus[out_focus > 0] = 1 298 | 299 | for inst_info in results_gland.values(): 300 | cnt = inst_info["contour"] 301 | cnt = cnt / ds_factor 302 | cnt = np.rint(cnt).astype("int") 303 | cv2.fillPoly(out_focus, pts=[cnt], color=0) 304 | 305 | del results_gland 306 | 307 | out_focus[wsi_blur > 225] = 0 308 | 309 | out_focus_lab = measurements.label(out_focus)[0] 310 | out_focus = remove_small_objects(out_focus_lab.astype("bool"), min_size=2500) 311 | 312 | out_focus[out_focus > 0] = 255 313 | 314 | return out_focus -------------------------------------------------------------------------------- /models/net_desc.py: -------------------------------------------------------------------------------- 1 | from typing import Optional 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | from torch.nn import BatchNorm1d 7 | 8 | from torch_geometric.nn import ( 9 | PNAConv, 10 | GATConv, 11 | global_mean_pool, 12 | ) 13 | from torch_geometric.utils import softmax 14 | from torch_geometric.nn.inits import reset 15 | from torch_scatter import scatter_add 16 | 17 | 18 | def weights_init(m): 19 | if isinstance(m, nn.Linear): 20 | nn.init.kaiming_uniform_(m.weight, mode="fan_out", nonlinearity="relu") 21 | nn.init.constant_(m.bias, 0) 22 | if isinstance(m, nn.BatchNorm1d): 23 | nn.init.constant_(m.weight, 1) 24 | nn.init.constant_(m.bias, 0) 25 | 26 | 27 | class GlobalAtt(torch.nn.Module): 28 | """GlobalAttenion but returning the attention scores.""" 29 | def __init__(self, gate_nn, nn): 30 | super().__init__() 31 | self.gate_nn = gate_nn 32 | self.nn = nn 33 | 34 | self.reset_parameters() 35 | 36 | def reset_parameters(self): 37 | reset(self.gate_nn) 38 | reset(self.nn) 39 | 40 | def forward(self, x: torch.Tensor, batch: Optional[torch.Tensor] = None, 41 | size: Optional[int] = None): 42 | r""" 43 | Args: 44 | x (Tensor): The input node features. 45 | batch (LongTensor, optional): A vector that maps each node to its 46 | respective graph identifier. (default: :obj:`None`) 47 | size (int, optional): The number of graphs in the batch. 48 | (default: :obj:`None`) 49 | """ 50 | if batch is None: 51 | batch = x.new_zeros(x.size(0), dtype=torch.int64) 52 | 53 | x = x.unsqueeze(-1) if x.dim() == 1 else x 54 | size = int(batch.max()) + 1 if size is None else size 55 | 56 | gate = self.gate_nn(x).view(-1, 1) 57 | x = self.nn(x) if self.nn is not None else x 58 | assert gate.dim() == x.dim() and gate.size(0) == x.size(0) 59 | 60 | gate = softmax(gate, batch, num_nodes=size) 61 | out = scatter_add(gate * x, batch, dim=0, dim_size=size) 62 | 63 | return out, gate 64 | 65 | 66 | class PNALayer(nn.Module): 67 | """PNA layer. 68 | 69 | Args: 70 | in_ch: number of input channels. 71 | out_ch: number of output channels. 72 | 73 | """ 74 | 75 | def __init__( 76 | self, 77 | in_ch, 78 | out_ch, 79 | aggregators=['mean', 'min', 'max', 'std'], 80 | scalers = ['identity', 'amplification', 'attenuation'], 81 | deg=None 82 | ): 83 | super().__init__() 84 | 85 | self.gconv = PNAConv(in_ch, out_ch, aggregators=aggregators, scalers=scalers, deg=deg) 86 | 87 | def forward(self, x, edge_index, freeze=False): 88 | if self.training: 89 | with torch.set_grad_enabled(not freeze): 90 | new_feat = self.gconv(x, edge_index) 91 | else: 92 | new_feat = self.gconv(x, edge_index) 93 | 94 | return new_feat 95 | 96 | 97 | class NetDesc(nn.Module): 98 | """Network description.""" 99 | 100 | def __init__( 101 | self, 102 | model_name, 103 | nr_features, 104 | node_degree, 105 | nhid=[12,12,12], 106 | grph_dim=10, 107 | dropout_rate=0.03, 108 | use_edges=0, 109 | label_dim=2, 110 | agg="attention", 111 | return_prob=False, 112 | ): 113 | super().__init__() 114 | self.dropout_rate = dropout_rate 115 | self.use_edges = use_edges 116 | self.agg = agg 117 | self.return_prob = return_prob 118 | 119 | node_degree = torch.Tensor(node_degree) 120 | 121 | if model_name == "graphsage": 122 | self.gconv1 = GraphSageLayer(nhid[0], nhid[1]) 123 | self.gconv2 = GraphSageLayer(nhid[1], nhid[2]) 124 | elif model_name == "gat": 125 | self.gconv1 = GATLayer(nhid[0], nhid[1]) 126 | self.gconv2 = GATLayer(nhid[1], nhid[2]) 127 | elif model_name == "gin": 128 | self.gconv1 = GINLayer(nhid[0], nhid[1], act="relu") 129 | self.gconv2 = GINLayer(nhid[1], nhid[2], act="relu") 130 | elif model_name == "pna": 131 | self.gconv1 = PNALayer(nhid[0], nhid[1], deg=node_degree) 132 | self.gconv2 = PNALayer(nhid[1], nhid[2], deg=node_degree) 133 | elif model_name == "edge": 134 | self.gconv1 = EdgeConvLayer(nhid[0], nhid[1], act="relu") 135 | self.gconv2 = EdgeConvLayer(nhid[1], nhid[2], act="relu") 136 | elif model_name == "linear": 137 | self.gconv1 = nn.Linear(nhid[0], nhid[1]) 138 | self.gconv2 = nn.Linear(nhid[1], nhid[2]) 139 | 140 | self.lin0 = nn.Linear(nr_features, nhid[0]) 141 | self.lin_emb0 = nn.Linear(nhid[0], grph_dim) 142 | self.lin_emb1 = nn.Linear(nhid[1], grph_dim) 143 | self.lin_emb2 = nn.Linear(nhid[2], grph_dim) 144 | 145 | self.lin_merge = nn.Linear(grph_dim*3, grph_dim) 146 | 147 | gate_nn = nn.Sequential(nn.Linear(grph_dim, 1)) 148 | v_nn = nn.Sequential(nn.Linear(grph_dim, grph_dim)) 149 | self.gpool = GlobalAtt(gate_nn, v_nn) 150 | # self.gpool = GlobalAttention(gate_nn, v_nn) 151 | 152 | self.lin_out = nn.Linear(grph_dim, label_dim) 153 | 154 | self.dropout = nn.Dropout(dropout_rate) 155 | 156 | def forward(self, x, edge_index, batch): 157 | edge_attr = None 158 | x0 = self.lin0(x) 159 | x1 = self.gconv1(x0, edge_index) 160 | x2 = self.gconv2(x1, edge_index) 161 | ### 162 | x0 = self.dropout(F.relu(self.lin_emb0(x0))) 163 | x1 = self.dropout(F.relu(self.lin_emb1(x1))) 164 | x2 = self.dropout(F.relu(self.lin_emb2(x2))) 165 | ### 166 | x_combined = F.relu(self.lin_merge(torch.cat((x0, x1, x2), dim=1))) 167 | 168 | # pool over node-level features 169 | if self.agg == 'attention': 170 | att_pool = self.gpool(x_combined, batch) 171 | logits = att_pool[0] 172 | scores = att_pool[1] 173 | scores = None 174 | elif self.agg == 'mean': 175 | scores = self.lin_scores(x_combined) 176 | logits = global_mean_pool(scores, batch) 177 | 178 | output1 = F.softmax(self.lin_out(logits), dim=-1) 179 | output2 = self.lin_out(logits) 180 | 181 | if self.return_prob: 182 | output = output1 183 | else: 184 | output = {"output": output1, "output_log": output2, "node_scores": scores} 185 | 186 | return output 187 | 188 | #### 189 | def create_model(**kwargs): 190 | return NetDesc(**kwargs) 191 | -------------------------------------------------------------------------------- /models/run_desc.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn.functional as F 4 | from sklearn.metrics import roc_auc_score 5 | 6 | import sys 7 | sys.path.append("..") 8 | from metrics.stats_utils import get_sens_spec_metrics 9 | from misc.utils import ranking_loss 10 | 11 | 12 | def train_step(batch_data, run_info, device="cuda"): 13 | run_info, state_info = run_info 14 | 15 | weight = torch.tensor([1, 1]).to(device).type(torch.float32) 16 | 17 | loss_func_dict = { 18 | "ce": torch.nn.CrossEntropyLoss(weight), 19 | "ranking": ranking_loss 20 | } 21 | # use 'ema' to add for EMA calculation, must be scalar! 22 | result_dict = {"EMA": {}} 23 | track_value = lambda name, value: result_dict["EMA"].update({name: value}) 24 | 25 | #### 26 | batch = batch_data.batch 27 | edge_index = batch_data.edge_index 28 | feats = batch_data.x 29 | label = batch_data.y 30 | 31 | #! below is specific to CoBi! 32 | # data is 3-class -> convert to 2 class (normal vs abnormal) 33 | label_orig = label.clone() # make copy of original 3 class label for evaluation 34 | label[label > 1] = 1 35 | 36 | batch = batch.to(device).type(torch.int64) 37 | edge_index = edge_index.to(device).type(torch.long) 38 | feats = feats.to(device).type(torch.float32) 39 | label = torch.squeeze(label.to(device).type(torch.int64)) 40 | 41 | #### 42 | model = run_info["net"]["desc"] 43 | optimizer = run_info["net"]["optimizer"] 44 | #### 45 | model.train() 46 | model.zero_grad() # not rnn so don't accumulate 47 | out_dict = model(feats, edge_index, batch) 48 | out = out_dict["output_log"] 49 | prob = out 50 | 51 | #### 52 | loss = 0 53 | loss_opts = run_info["net"]["extra_info"]["loss"] 54 | 55 | for loss_name, loss_weight in loss_opts.items(): 56 | loss_func = loss_func_dict[loss_name] 57 | loss_args = [out, label] 58 | term_loss = loss_func(*loss_args) 59 | track_value("loss_%s" % loss_name, term_loss.cpu().item()) 60 | loss += loss_weight * term_loss 61 | loss = torch.unsqueeze(loss, 0) 62 | 63 | track_value("overall_loss", loss.cpu().item()) 64 | 65 | # * gradient update 66 | loss.backward() 67 | optimizer.step() 68 | #### 69 | 70 | # * Its up to user to define the protocol to process the raw output per step! 71 | result_dict["raw"] = {"true": label, "true_orig": label_orig, "prob": prob} 72 | 73 | return result_dict 74 | 75 | 76 | def valid_step(batch_data, run_info, device="cuda"): 77 | run_info, state_info = run_info 78 | #### 79 | batch = batch_data.batch 80 | edge_index = batch_data.edge_index 81 | feats = batch_data.x 82 | label = batch_data.y 83 | 84 | # data is 3-class -> convert to 2 class (normal vs abnormal) 85 | label_orig = label.clone() # make copy of original 3 class label for evaluation 86 | label[label > 1] = 1 87 | 88 | batch = batch.to(device).type(torch.int64) 89 | edge_index = edge_index.to(device).type(torch.long) 90 | feats = feats.to(device).type(torch.float32) 91 | label = label.to(device).type(torch.int64) 92 | 93 | #### 94 | model = run_info["net"]["desc"] 95 | model.eval() # inference mode 96 | 97 | # -------------------------------------------------------------- 98 | with torch.no_grad(): # dont compute gradient 99 | out_dict = model(feats, edge_index, batch) 100 | prob = out_dict["output"] 101 | 102 | # * Its up to user to define the protocol to process the raw output per step! 103 | result_dict = {"raw": {"true": label, "true_orig": label_orig, "prob": prob}} 104 | 105 | return result_dict 106 | 107 | 108 | def infer_step(batch_data, model, device="cuda"): 109 | #### 110 | batch = batch_data.batch 111 | edge_index = batch_data.edge_index 112 | feats = batch_data.x 113 | label = batch_data.y 114 | wsi_name = batch_data.wsi_name 115 | 116 | # data is 3-class -> convert to 2 class (normal vs abnormal) 117 | label_orig = label.clone() # make copy of original 3 class label for evaluation 118 | label[label > 1] = 1 119 | 120 | batch = batch.to(device).type(torch.int64) 121 | edge_index = edge_index.to(device).type(torch.long) 122 | feats = feats.to(device).type(torch.float32) 123 | label = label.to(device).type(torch.int64) 124 | 125 | #### 126 | model.eval() # inference mode 127 | 128 | # -------------------------------------------------------------- 129 | with torch.no_grad(): # dont compute gradient 130 | out_dict = model(feats, edge_index, batch) 131 | prob = out_dict["output"] 132 | node_scores = out_dict["node_scores"] 133 | 134 | # * Its up to user to define the protocol to process the raw output per step! 135 | result_dict = {"true": label, "true_orig": label_orig, "prob": prob, "node_scores": node_scores, "wsi_name": wsi_name} 136 | 137 | return result_dict 138 | 139 | 140 | def proc_valid_step_output(raw_data): 141 | track_dict = {"scalar": {}, "image": {}} 142 | 143 | def track_value(name, value, vtype): 144 | return track_dict[vtype].update({name: value}) 145 | 146 | prob = raw_data["prob"] 147 | true = raw_data["true"] 148 | true_orig = np.array(raw_data["true_orig"]) 149 | num_examples = len(raw_data["true"]) 150 | 151 | prob_list = [] 152 | true_list = [] 153 | for idx in range(num_examples): 154 | graph_prob = prob[idx][1] 155 | graph_true = true[idx] 156 | prob_list.append(graph_prob.cpu()) 157 | true_list.append(graph_true.cpu()) 158 | 159 | prob = np.array(prob_list) 160 | pred = prob.copy() 161 | pred[pred >= 0.5] = 1 162 | pred[pred < 0.5] = 0 163 | true = np.array(true_list) 164 | 165 | auc = roc_auc_score(true, prob) 166 | 167 | spec_95, spec_97, spec_98, spec_99, spec_100 = get_sens_spec_metrics(true, prob) 168 | 169 | track_value("AUC-ROC", auc, "scalar") 170 | track_value("Specifity_at_95_Sensitivity", spec_95, "scalar") 171 | track_value("Specifity_at_97_Sensitivity", spec_97, "scalar") 172 | track_value("Specifity_at_98_Sensitivity", spec_98, "scalar") 173 | track_value("Specifity_at_99_Sensitivity", spec_99, "scalar") 174 | 175 | return track_dict 176 | 177 | -------------------------------------------------------------------------------- /run_infer.py: -------------------------------------------------------------------------------- 1 | """run_infer.py 2 | 3 | Process slides with IGUANA. 4 | 5 | Usage: 6 | run_infer.py [--gpu=] [--model_path=] [--model_name=] [--data_dir=] \ 7 | [--data_info=] [--stats_dir=] [--output_dir=] [--batch_size=] \ 8 | [--fold_nr=] [--split_nr=] [--num_workers=] 9 | run_infer.py (-h | --help) 10 | run_infer.py --version 11 | 12 | Options: 13 | -h --help Show this string. 14 | --version Show version. 15 | --gpu= GPU list. [default: 0] 16 | --model_path= Path to saved checkpoint. 17 | --model_name= Type of graph convolution used. [default: pna] 18 | --data_dir= Path to where graph data is stored. 19 | --data_info= Path to where data information csv file is stored 20 | --stats_dir= Location of feaure stats directory for input standardisation. 21 | --output_dir= Path where output will be saved. [default: output/] 22 | --batch_size= Batch size. [default: 1] 23 | --fold_nr= Fold number considered during cross validation. Don't change if considering independent test set. [default: 1] 24 | --split_nr= Only consider slides in the data info csv according to this selected number. [default: 3] 25 | --num_workers= Number of workers. [default: 8] 26 | 27 | """ 28 | 29 | import os 30 | import yaml 31 | from docopt import docopt 32 | import tqdm 33 | import numpy as np 34 | import pandas as pd 35 | from importlib import import_module 36 | import glob 37 | from sklearn.metrics import roc_auc_score, precision_recall_curve, auc 38 | 39 | import torch 40 | from torch_geometric.data import DataLoader 41 | 42 | from dataloader.graph_loader import FileLoader 43 | from metrics.stats_utils import get_sens_spec_metrics 44 | from misc.utils import rm_n_mkdir 45 | 46 | import warnings 47 | warnings.filterwarnings('ignore') 48 | 49 | 50 | def get_labels_scores(wsi_names, scores, gt, binarize=True): 51 | """Align the scores and labels.""" 52 | labels_output = [] 53 | scores_output = [] 54 | for idx, wsi_name in enumerate(wsi_names): 55 | score = scores[idx] 56 | gt_subset = gt[gt["wsi_name"] == wsi_name] 57 | lab = list(gt_subset["label"]) 58 | if len(lab) > 0: 59 | lab = int(lab[0]) 60 | if binarize: 61 | if lab > 0: 62 | lab = 1 63 | labels_output.append(lab) 64 | scores_output.append(score) 65 | return labels_output, scores_output 66 | 67 | 68 | class InferBase(object): 69 | def __init__(self, **kwargs): 70 | self.run_step = None 71 | for variable, value in kwargs.items(): 72 | self.__setattr__(variable, value) 73 | self.__load_model() 74 | return 75 | 76 | def __load_model(self): 77 | """Create the model, load the checkpoint and define 78 | associated run steps to process each data batch 79 | 80 | """ 81 | model_desc = import_module('models.net_desc') 82 | model_creator = getattr(model_desc, 'create_model') 83 | 84 | # TODO: deal with parsing multi level model desc 85 | net = model_creator( 86 | model_name=self.model_name, 87 | nr_features=len(self.feat_names), 88 | node_degree=self.node_degree).to('cuda') 89 | saved_state_dict = torch.load(self.model_path) 90 | net.load_state_dict(saved_state_dict['desc'], strict=True) 91 | 92 | run_desc = import_module('models.run_desc') 93 | self.run_step = lambda input_batch: getattr( 94 | run_desc, 'infer_step')(input_batch, net) 95 | return 96 | 97 | 98 | class Infer(InferBase): 99 | def __run_model(self, file_list): 100 | 101 | print('Loading feature statistics...') 102 | with open(f"{self.stats_path}/stats_dict.yml") as fptr: 103 | stats_dict = yaml.full_load(fptr) 104 | 105 | input_dataset = FileLoader( 106 | file_list, self.feat_names, feat_stats=stats_dict, norm="standard", data_clean="std" 107 | ) 108 | 109 | dataloader = DataLoader(input_dataset, 110 | num_workers=self.nr_procs, 111 | batch_size=self.batch_size, 112 | shuffle=False, 113 | drop_last=False 114 | ) 115 | 116 | pbar = tqdm.tqdm(desc='Processsing', leave=True, 117 | total=int(len(dataloader)), 118 | ncols=80, ascii=True, position=0) 119 | 120 | pred_all = [] 121 | prob_all = [] 122 | true_all = [] 123 | wsi_name_all = [] 124 | for _, batch_data in enumerate(dataloader): 125 | 126 | batch_output = self.run_step(batch_data) 127 | pred_list = [] 128 | prob_list = [] 129 | true_list = [] 130 | wsi_name_list = [] 131 | 132 | prob = batch_output['prob'] 133 | true = batch_output['true'] 134 | wsi_name = batch_output['wsi_name'][0] 135 | num_examples = len(batch_output['true']) 136 | 137 | for idx in range(num_examples): 138 | pred_tmp = torch.argmax(prob[idx]) 139 | prob_tmp = prob[idx][1] 140 | true_tmp = true[idx] 141 | pred_list.append(pred_tmp.cpu()) 142 | prob_list.append(prob_tmp.cpu()) 143 | true_list.append(true_tmp.cpu()) 144 | wsi_name_list.append(wsi_name) 145 | 146 | pred_all.extend(pred_list) 147 | prob_all.extend(prob_list) 148 | true_all.extend(true_list) 149 | wsi_name_all.extend(wsi_name_list) 150 | 151 | pbar.update() 152 | pbar.close() 153 | return np.array(pred_all), np.array(prob_all), np.array(true_all), np.array(wsi_name_all) 154 | 155 | 156 | def __get_stats(self, prob, true): 157 | # AUC-ROC 158 | auc_roc = roc_auc_score(true, prob) 159 | # AUC-PR 160 | pr, re, _ = precision_recall_curve(true, prob) 161 | auc_pr = auc(re, pr) 162 | # specificity @ given sensitivity 163 | spec_95, spec_97, spec_98, spec_99, spec_100 = get_sens_spec_metrics(true, prob) 164 | 165 | print('='*50) 166 | print("AUC-ROC:", auc_roc) 167 | print("AUC-PR:", auc_pr) 168 | print("Specifity_at_97_Sensitivity:", spec_97) 169 | print("Specifity_at_98_Sensitivity:", spec_98) 170 | print("Specifity_at_99_Sensitivity:", spec_99) 171 | 172 | def process_files(self): 173 | 174 | # select the slides according to selected fold_nr and split_nr 175 | # independent test set should have split_nr all equal to 3 176 | data_info = pd.read_csv(self.data_info) 177 | file_list = [] 178 | for row in data_info.iterrows(): 179 | wsi_name = row[1].iloc[0] 180 | if row[1].iloc[self.fold_nr] == self.split_nr: 181 | file_list.append(f"{self.data_path}/{wsi_name}.dat") 182 | file_list.sort() # to always ensure same input ordering 183 | 184 | print('Number of WSI graphs:', len(file_list)) 185 | print('-'*50) 186 | 187 | pred, prob, true, wsi_names = self.__run_model(file_list) 188 | 189 | # save results to a single csv file 190 | df = pd.DataFrame(data = {'wsi_name': wsi_names, 'score': prob, "pred": pred, 'label': true}) 191 | df.to_csv(f"{self.output_path}/results.csv") 192 | 193 | # get stats 194 | true, prob = get_labels_scores(wsi_names, prob, data_info) 195 | self.__get_stats(prob, true) 196 | 197 | 198 | #------------------------------------------------------------------------------------------------------- 199 | 200 | if __name__ == '__main__': 201 | args = docopt(__doc__, version='IGUANA Inference v1.0') 202 | print(args) 203 | 204 | os.environ['CUDA_VISIBLE_DEVICES'] = args['--gpu'] 205 | 206 | # get the subset of features to be input to the GNN 207 | with open("features.yml") as fptr: 208 | feat_names = list(yaml.full_load(fptr).values())[0] 209 | 210 | # load node degree 211 | stats_path = args["--stats_dir"] 212 | if args["--model_name"] == "pna": 213 | node_degree = np.load(f"{stats_path}/node_deg.npy") 214 | else: 215 | node_degree = None 216 | 217 | if not os.path.exists(args["--output_dir"]): 218 | rm_n_mkdir(args["--output_dir"]) 219 | 220 | #TODO Batch size must be set at 1 at the moment - fix this! 221 | args = { 222 | "model_name": args["--model_name"], 223 | "model_path": args["--model_path"], 224 | "stats_path": stats_path, 225 | "node_degree": node_degree, 226 | "data_path": args["--data_dir"], 227 | "data_info": args["--data_info"], 228 | "feat_names": feat_names, 229 | "batch_size": int(args["--batch_size"]), 230 | "nr_procs": int(args["--num_workers"]), 231 | "output_path": args["--output_dir"], 232 | "fold_nr": int(args["--fold_nr"]), 233 | "split_nr": int(args["--split_nr"]), 234 | } 235 | 236 | infer = Infer(**args) 237 | infer.process_files() 238 | 239 | -------------------------------------------------------------------------------- /run_infer.sh: -------------------------------------------------------------------------------- 1 | python run_infer.py \ 2 | --gpu=0 \ 3 | --model_path="/root/lsf_workspace/iguana_data/weights/iguana_fold1.tar" \ 4 | --model_name="pna" \ 5 | --data_dir="/root/lsf_workspace/proc_slides/cobi/colchester/graphs/data" \ 6 | --data_info="/root/lsf_workspace/proc_slides/cobi/colchester/graphs/colchester_info.csv" \ 7 | --stats_dir="/root/lsf_workspace/proc_slides/cobi/uhcw/graphs/stats" \ 8 | --output_dir="output_test/" \ -------------------------------------------------------------------------------- /run_utils/callbacks/base.py: -------------------------------------------------------------------------------- 1 | 2 | import cv2 3 | import matplotlib.pyplot as plt 4 | import numpy as np 5 | import torch 6 | from scipy.stats import mode as major_value 7 | from sklearn.metrics import confusion_matrix 8 | 9 | 10 | #### 11 | class BaseCallbacks(object): 12 | def __init__(self): 13 | self.engine_trigger = False 14 | 15 | def reset(self): 16 | pass 17 | 18 | def run(self, state, event): 19 | pass 20 | 21 | #### 22 | class TrackLr(BaseCallbacks): 23 | """ 24 | Add learning rate to tracking 25 | """ 26 | def __init__(self, per_n_epoch=1, per_n_step=None): 27 | super().__init__() 28 | self.per_n_epoch = per_n_epoch 29 | self.per_n_step = per_n_step 30 | 31 | def run(self, state, event): 32 | # logging learning rate, decouple into another callback? 33 | run_info = state.run_info 34 | for net_name, net_info in run_info.items(): 35 | lr = net_info['optimizer'].param_groups[0]['lr'] 36 | state.tracked_step_output['scalar']['lr-%s' % net_name] = lr 37 | return 38 | 39 | #### 40 | class ScheduleLr(BaseCallbacks): 41 | """ 42 | Trigger all scheduler 43 | """ 44 | def __init__(self): 45 | super().__init__() 46 | 47 | def run(self, state, event): 48 | # logging learning rate, decouple into another callback? 49 | run_info = state.run_info 50 | for net_name, net_info in run_info.items(): 51 | net_info['lr_scheduler'].step() 52 | return 53 | 54 | #### 55 | class TriggerEngine(BaseCallbacks): 56 | def __init__(self, triggered_engine_name, nr_epoch=1): 57 | self.engine_trigger = True 58 | self.triggered_engine_name = triggered_engine_name 59 | self.triggered_engine = None 60 | self.nr_epoch = nr_epoch 61 | 62 | def run(self, state, event): 63 | self.triggered_engine.run(chained=True, 64 | nr_epoch=self.nr_epoch, 65 | shared_state=state) 66 | return 67 | #### 68 | class CheckpointSaver(BaseCallbacks): 69 | """ 70 | Must declare save dir first in the shared global state of the 71 | attached engine 72 | """ 73 | def run(self, state, event): 74 | if not state.logging: 75 | return 76 | for net_name, net_info in state.run_info.items(): 77 | net_checkpoint = {} 78 | for key, value in net_info.items(): 79 | if key != 'extra_info': 80 | net_checkpoint[key] = value.state_dict() 81 | torch.save(net_checkpoint, '%s/%s_epoch=%d.tar' % 82 | (state.log_dir, net_name, state.curr_epoch)) 83 | return 84 | 85 | #### 86 | class AccumulateRawOutput(BaseCallbacks): 87 | def run(self, state, event): 88 | step_output = state.step_output['raw'] 89 | accumulated_output = state.epoch_accumulated_output 90 | 91 | for key, step_value in step_output.items(): 92 | if key in accumulated_output: 93 | accumulated_output[key].extend(list(step_value)) 94 | else: 95 | accumulated_output[key] = list(step_value) 96 | return 97 | #### 98 | class ScalarMovingAverage(BaseCallbacks): 99 | """ 100 | Calculate the running average for all scalar output of 101 | each runstep of the attached RunEngine 102 | """ 103 | 104 | def __init__(self, alpha=0.95): 105 | super().__init__() 106 | self.alpha = alpha 107 | self.tracking_dict = {} 108 | 109 | def run(self, state, event): 110 | # TODO: protocol for dynamic key retrieval for EMA 111 | step_output = state.step_output['EMA'] 112 | 113 | for key, current_value in step_output.items(): 114 | if key in self.tracking_dict: 115 | old_ema_value = self.tracking_dict[key] 116 | # calculate the exponential moving average 117 | new_ema_value = old_ema_value * self.alpha + (1.0 - self.alpha) * current_value 118 | self.tracking_dict[key] = new_ema_value 119 | else: # init for variable which appear for the first time 120 | new_ema_value = current_value 121 | self.tracking_dict[key] = new_ema_value 122 | 123 | state.tracked_step_output['scalar'] = self.tracking_dict 124 | return 125 | 126 | #### 127 | class ProcessAccumulatedRawOutput(BaseCallbacks): 128 | def __init__(self, proc_func, per_n_epoch=1): 129 | # TODO: allow dynamically attach specific procesing for `type` 130 | super().__init__() 131 | self.per_n_epoch = per_n_epoch 132 | self.proc_func = proc_func 133 | 134 | def run(self, state, event): 135 | current_epoch = state.curr_epoch 136 | # if current_epoch % self.per_n_epoch != 0: return 137 | raw_data = state.epoch_accumulated_output 138 | track_dict = self.proc_func(raw_data) 139 | # update global shared states 140 | state.tracked_step_output = track_dict 141 | return 142 | 143 | #### 144 | class VisualizeOutput(BaseCallbacks): 145 | def __init__(self, proc_func, per_n_epoch=1): 146 | super().__init__() 147 | # TODO: option to dump viz per epoch or per n step 148 | self.per_n_epoch = per_n_epoch 149 | self.proc_func = proc_func 150 | 151 | def run(self, state, event): 152 | current_epoch = state.curr_epoch 153 | raw_output = state.step_output['raw'] 154 | viz_image = self.proc_func(raw_output) 155 | state.tracked_step_output['image']['output'] = viz_image 156 | return 157 | -------------------------------------------------------------------------------- /run_utils/callbacks/logging.py: -------------------------------------------------------------------------------- 1 | 2 | import json 3 | import random 4 | 5 | import matplotlib.pyplot as plt 6 | import numpy as np 7 | from matplotlib.lines import Line2D 8 | from termcolor import colored 9 | 10 | from .base import BaseCallbacks 11 | from .serialize import fig2data, serialize 12 | 13 | # TODO: logging for all printed info on the terminal 14 | 15 | 16 | #### 17 | class LoggingGradient(BaseCallbacks): 18 | """ 19 | Will log per each training step 20 | """ 21 | def _pyplot_grad_flow(self, named_parameters): 22 | """ 23 | Plots the gradients flowing through different layers in the net during training. 24 | "_pyplot_grad_flow(self.model.named_parameters())" to visualize the gradient flow 25 | 26 | ! Very slow if triggered per steps because of CPU <=> GPU 27 | """ 28 | ave_grads = [] 29 | max_grads = [] 30 | layers = [] 31 | for n, p in named_parameters: 32 | if(p.requires_grad) and ("bias" not in n): 33 | layers.append(n) 34 | ave_grads.append(p.grad.abs().mean().cpu().item()) 35 | max_grads.append(p.grad.abs().max().cpu().item()) 36 | fig = plt.figure(figsize=(10, 10)) 37 | plt.bar(np.arange(len(max_grads)), max_grads, 38 | alpha=0.1, lw=1, color="c") 39 | plt.bar(np.arange(len(max_grads)), ave_grads, 40 | alpha=0.1, lw=1, color="b") 41 | plt.hlines(0, 0, len(ave_grads)+1, lw=2, color="k") 42 | plt.xticks(range(0, len(ave_grads), 1), layers, rotation="vertical") 43 | plt.xlim(left=0, right=len(ave_grads)) 44 | # zoom in on the lower gradient regions 45 | plt.xlabel("Layers") 46 | plt.ylabel("average gradient") 47 | plt.title("Gradient flow") 48 | plt.grid(True) 49 | plt.legend([Line2D([0], [0], color="c", lw=4), 50 | Line2D([0], [0], color="b", lw=4), 51 | Line2D([0], [0], color="k", lw=4)], 52 | ['max-gradient', 'mean-gradient', 'zero-gradient']) 53 | fig = np.transpose(fig2data(fig), axes=[2, 0, 1]) # HWC => CHW 54 | plt.close() 55 | return fig 56 | 57 | def run(self, state, event): 58 | 59 | if random.random() > 0.05: return 60 | curr_step = state.curr_global_step 61 | 62 | # logging the grad of all trainable parameters 63 | tfwriter = state.log_info['tfwriter'] 64 | run_info = state.run_info 65 | for net_name, net_info in run_info.items(): 66 | netdesc = net_info['desc'].module 67 | for param_name, param in netdesc.named_parameters(): 68 | param_grad = param.grad 69 | # TODO: sync test None or epislon for pytorch 1.4 vs 1.5 70 | if param_grad is None: continue 71 | tfwriter.add_histogram( 72 | "%s_grad/%s" % (net_name, param_name), 73 | param_grad.detach().cpu().numpy().flatten(), 74 | global_step=curr_step) # ditribute into 10 bins (np default) 75 | tfwriter.add_histogram( 76 | "%s_para/%s" % (net_name, param_name), 77 | param.detach().cpu().numpy().flatten(), 78 | global_step=curr_step) # ditribute into 10 bins (np default) 79 | return 80 | 81 | 82 | #### 83 | class LoggingEpochOutput(BaseCallbacks): 84 | """ 85 | Must declare save dir first in the shared global state of the 86 | attached engine 87 | """ 88 | def __init__(self, per_n_epoch=1): 89 | super().__init__() 90 | self.per_n_epoch = per_n_epoch 91 | 92 | def run(self, state, event): 93 | 94 | # only logging every n epochs also 95 | if state.curr_epoch % self.per_n_epoch != 0: 96 | return 97 | 98 | # TODO: rename to differentiate global vs local epoch 99 | if state.global_state is not None: 100 | current_epoch = str(state.global_state.curr_epoch) 101 | else: 102 | current_epoch = str(state.curr_epoch) 103 | 104 | output = state.tracked_step_output 105 | 106 | def get_serializable_values(output_format): 107 | log_dict = {} 108 | # get type and variable that is serializable 109 | # to console or other logging format (json, tensorboard) 110 | for variable_type, variable_dict in output.items(): 111 | for value_name, value in variable_dict.items(): 112 | value_name = '%s-%s' % (state.attached_engine_name, 113 | value_name) 114 | new_format = serialize(value, variable_type, output_format) 115 | if new_format is not None: 116 | log_dict[value_name] = new_format 117 | return log_dict 118 | 119 | # * Serialize to Console 120 | # align the console print output 121 | formatted_values = get_serializable_values('console') 122 | max_length = len(max(formatted_values.keys(), key=len)) 123 | for value_name, value_text in formatted_values.items(): 124 | value_name = colored(value_name.ljust(max_length), 'green') 125 | print('------%s : %s' % (value_name, value_text)) 126 | 127 | # TODO: [CRITICAL] fix passing this between engine 128 | # if not state.logging: return 129 | 130 | # * Serialize to JSON file 131 | stat_dict = get_serializable_values('json') 132 | # json stat log file, update and overwrite 133 | with open(state.log_info['json_file']) as json_file: 134 | json_data = json.load(json_file) 135 | 136 | if current_epoch in json_data: 137 | old_stat_dict = json_data[current_epoch] 138 | stat_dict.update(old_stat_dict) 139 | current_epoch_dict = {current_epoch: stat_dict} 140 | json_data.update(current_epoch_dict) 141 | 142 | # TODO: may corrupt 143 | with open(state.log_info['json_file'], 'w') as json_file: 144 | json.dump(json_data, json_file) 145 | 146 | # * Serialize to Tensorboard 147 | tfwriter = state.log_info['tfwriter'] 148 | formatted_values = get_serializable_values('tensorboard') 149 | # ! may need to flush to force update 150 | for value_name, value in formatted_values.items(): 151 | # TODO: dynamically call this 152 | if value[0] == 'scalar': 153 | tfwriter.add_scalar(value_name, value[1], current_epoch) 154 | elif value[0] == 'image': 155 | tfwriter.add_image(value_name, value[1], current_epoch, 156 | dataformats='HWC') 157 | # tfwriter.flush() 158 | 159 | return 160 | -------------------------------------------------------------------------------- /run_utils/callbacks/serialize.py: -------------------------------------------------------------------------------- 1 | 2 | import cv2 3 | import matplotlib 4 | import numpy as np 5 | from matplotlib import pyplot as plt 6 | 7 | # * syn where to set this 8 | # must use 'Agg' to plot out onto image 9 | matplotlib.use('Agg') 10 | 11 | #### 12 | def fig2data(fig, dpi=180): 13 | """ 14 | Convert a Matplotlib figure to a 4D numpy array with RGBA channels and return it 15 | Args: 16 | fig: a matplotlib figure 17 | 18 | Return: a numpy 3D array of RGBA values 19 | """ 20 | buf = io.BytesIO() 21 | fig.savefig(buf, format="png", dpi=dpi) 22 | buf.seek(0) 23 | img_arr = np.frombuffer(buf.getvalue(), dtype=np.uint8) 24 | buf.close() 25 | img = cv2.imdecode(img_arr, 1) 26 | img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 27 | return img 28 | #### 29 | 30 | 31 | #### 32 | class _Scalar(object): 33 | @staticmethod 34 | def to_console(value): 35 | return '%0.5f' % value 36 | 37 | @staticmethod 38 | def to_json(value): 39 | return value 40 | 41 | @staticmethod 42 | def to_tensorboard(value): 43 | return 'scalar', value 44 | 45 | #### 46 | class _ConfusionMatrix(object): 47 | @staticmethod 48 | def to_console(value): 49 | value = pd.DataFrame(value) 50 | value.index.name = 'True' 51 | value.columns.name = 'Pred' 52 | formatted_value = value.to_string() 53 | return '\n' + formatted_value 54 | 55 | @staticmethod 56 | def to_json(value): 57 | value = pd.DataFrame(value) 58 | value.index.name = 'True' 59 | value.columns.name = 'Pred' 60 | value = value.unstack().rename('value').reset_index() 61 | value = pd.Series({'conf_mat': value}) 62 | formatted_value = value.to_json(orient='records') 63 | return formatted_value 64 | 65 | @staticmethod 66 | def to_tensorboard(value): 67 | def plot_confusion_matrix(cm, 68 | target_names, 69 | title='Confusion matrix', 70 | cmap=None, 71 | normalize=False): 72 | """ 73 | given a sklearn confusion matrix (cm), make a nice plot 74 | 75 | Arguments 76 | --------- 77 | cm: confusion matrix from sklearn.metrics.confusion_matrix 78 | 79 | target_names: given classification classes such as [0, 1, 2] 80 | the class names, for example: ['high', 'medium', 'low'] 81 | 82 | title: the text to display at the top of the matrix 83 | 84 | cmap: the gradient of the values displayed from matplotlib.pyplot.cm 85 | see http://matplotlib.org/examples/color/colormaps_reference.html 86 | plt.get_cmap('jet') or plt.cm.Blues 87 | 88 | normalize: If False, plot the raw numbers 89 | If True, plot the proportions 90 | 91 | Usage 92 | ----- 93 | plot_confusion_matrix(cm = cm, # confusion matrix created by 94 | # sklearn.metrics.confusion_matrix 95 | normalize = True, # show proportions 96 | target_names = y_labels_vals, # list of names of the classes 97 | title = best_estimator_name) # title of graph 98 | 99 | Citiation 100 | --------- 101 | http://scikit-learn.org/stable/auto_examples/model_selection/plot_confusion_matrix.html 102 | 103 | """ 104 | import matplotlib.pyplot as plt 105 | import numpy as np 106 | import itertools 107 | 108 | accuracy = np.trace(cm) / np.sum(cm).astype('float') 109 | misclass = 1 - accuracy 110 | 111 | if cmap is None: 112 | cmap = plt.get_cmap('Blues') 113 | 114 | plt.figure(figsize=(8, 6)) 115 | plt.imshow(cm, interpolation='nearest', cmap=cmap) 116 | plt.title(title) 117 | plt.colorbar() 118 | 119 | if target_names is not None: 120 | tick_marks = np.arange(len(target_names)) 121 | plt.xticks(tick_marks, target_names, rotation=45) 122 | plt.yticks(tick_marks, target_names) 123 | 124 | if normalize: 125 | cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis] 126 | 127 | 128 | thresh = cm.max() / 1.5 if normalize else cm.max() / 2 129 | for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])): 130 | if normalize: 131 | plt.text(j, i, "{:0.4f}".format(cm[i, j]), 132 | horizontalalignment="center", 133 | color="white" if cm[i, j] > thresh else "black") 134 | else: 135 | plt.text(j, i, "{:,}".format(cm[i, j]), 136 | horizontalalignment="center", 137 | color="white" if cm[i, j] > thresh else "black") 138 | 139 | 140 | plt.tight_layout() 141 | plt.ylabel('True label') 142 | plt.xlabel('Predicted label\naccuracy={:0.4f}; misclass={:0.4f}'.format(accuracy, misclass)) 143 | 144 | plot_confusion_matrix(value, ['0', '1']) 145 | img = fig2data(plt.gcf()) 146 | plt.close() 147 | return 'image', img 148 | 149 | #### 150 | class _Image(object): 151 | @staticmethod 152 | def to_console(value): 153 | # TODO: add warn for not possible or sthg here 154 | return None 155 | 156 | @staticmethod 157 | def to_json(value): 158 | # TODO: add warn for not possible or sthg here 159 | return None 160 | 161 | @staticmethod 162 | def to_tensorboard(value): 163 | # TODO: add method 164 | return 'image', value 165 | 166 | 167 | __converter_dict = { 168 | 'scalar': _Scalar, 169 | 'conf_mat': _ConfusionMatrix, 170 | 'image': _Image 171 | } 172 | 173 | 174 | #### 175 | def serialize(value, input_format, output_format): 176 | converter = __converter_dict[input_format] 177 | if output_format == 'console': 178 | return converter.to_console(value) 179 | elif output_format == 'json': 180 | return converter.to_json(value) 181 | elif output_format == 'tensorboard': 182 | return converter.to_tensorboard(value) 183 | else: 184 | assert False, 'Unknown format' 185 | -------------------------------------------------------------------------------- /run_utils/engine.py: -------------------------------------------------------------------------------- 1 | import tqdm 2 | from enum import Enum 3 | 4 | 5 | #### 6 | class Events(Enum): 7 | EPOCH_STARTED = "epoch_started" 8 | EPOCH_COMPLETED = "epoch_completed" 9 | STEP_STARTED = "step_started" 10 | STEP_COMPLETED = "step_completed" 11 | STARTED = "started" 12 | COMPLETED = "completed" 13 | EXCEPTION_RAISED = "exception_raised" 14 | 15 | 16 | #### 17 | class State(object): 18 | """ 19 | An object that is used to pass internal and 20 | user-defined state between event handlers 21 | """ 22 | 23 | def __init__(self): 24 | # settings propagated from config 25 | self.logging = None 26 | self.log_dir = None 27 | self.log_info = None 28 | 29 | # internal variable 30 | self.curr_epoch_step = 0 # current step in epoch 31 | self.curr_global_step = 0 # current global step 32 | self.curr_epoch = 0 # current global epoch 33 | 34 | # TODO: [LOW] better document this 35 | # for outputing value that will be tracked per step 36 | # "scalar" will always be printed out and added to the tensorboard 37 | # "images" will need dedicated function to process and added to the tensorboard 38 | 39 | # ! naming should match with types supported for serialize 40 | # TODO: Need way to dynamically adding new types 41 | self.tracked_step_output = { 42 | "scalar": {}, # type : {variable_name : variablee_value} 43 | "image": {}, 44 | } 45 | # TODO: find way to known which method bind/interact with which value 46 | 47 | self.epoch_accumulated_output = {} # all output of the current epoch 48 | 49 | # TODO: soft reset for pertain variable for N epochs 50 | self.run_accumulated_output = [] # of run until reseted 51 | 52 | # holder for output returned after current runstep 53 | # * depend on the type of training i.e GAN, the updated accumulated may be different 54 | self.step_output = None 55 | 56 | self.global_state = None 57 | return 58 | 59 | def reset_variable(self): 60 | # type : {variable_name : variable_value} 61 | self.tracked_step_output = {k: {} for k in self.tracked_step_output.keys()} 62 | 63 | # TODO: [CRITICAL] refactor this 64 | if self.curr_epoch % self.pertain_n_epoch_output == 0: 65 | self.run_accumulated_output = [] 66 | 67 | self.epoch_accumulated_output = {} 68 | 69 | # * depend on the type of training i.e GAN, the updated accumulated may be different 70 | self.step_output = None # holder for output returned after current runstep 71 | return 72 | 73 | 74 | #### 75 | class RunEngine(object): 76 | """ 77 | TODO: Include docstring 78 | """ 79 | 80 | def __init__( 81 | self, 82 | engine_name=None, 83 | dataloader=None, 84 | run_step=None, 85 | run_info=None, 86 | log_info=None, # TODO: refactor this with trainer.py 87 | ): 88 | 89 | # * auto set all input as object variables 90 | self.engine_name = engine_name 91 | self.run_step = run_step 92 | self.dataloader = dataloader 93 | 94 | # * global variable/object holder shared between all event handler 95 | self.state = State() 96 | # * check if correctly referenced, not new copies 97 | self.state.attached_engine_name = engine_name # TODO: redundant? 98 | self.state.run_info = run_info 99 | self.state.log_info = log_info 100 | self.state.batch_size = dataloader.batch_size 101 | 102 | # TODO: [CRITICAL] match all the mechanism outline with opt 103 | self.state.pertain_n_epoch_output = 1 if engine_name == "valid" else 1 104 | 105 | self.event_handler_dict = {event: [] for event in Events} 106 | 107 | # TODO: think about this more 108 | # to share global state across a chain of RunEngine such as 109 | # from the engine for training to engine for validation 110 | 111 | # 112 | self.terminate = False 113 | return 114 | 115 | def __reset_state(self): 116 | # TODO: think about this more, looks too redundant 117 | new_state = State() 118 | new_state.attached_engine_name = self.state.attached_engine_name 119 | new_state.run_info = self.state.run_info 120 | new_state.log_info = self.state.log_info 121 | self.state = new_state 122 | return 123 | 124 | def __trigger_events(self, event): 125 | for callback in self.event_handler_dict[event]: 126 | callback.run(self.state, event) 127 | # TODO: exception and throwing error with name or sthg to allow trace back 128 | return 129 | 130 | # TODO: variable to indicate output dependency between handler ! 131 | def add_event_handler(self, event_name, handler): 132 | self.event_handler_dict[event_name].append(handler) 133 | 134 | # ! Put into trainer.py ? 135 | def run(self, nr_epoch=1, shared_state=None, chained=False): 136 | 137 | # TODO: refactor this 138 | if chained: 139 | self.state.curr_epoch = 0 140 | self.state.global_state = shared_state 141 | 142 | while self.state.curr_epoch < nr_epoch: 143 | self.state.reset_variable() # * reset all EMA holder per epoch 144 | 145 | if not chained: 146 | print("----------------EPOCH %d" % (self.state.curr_epoch + 1)) 147 | 148 | self.__trigger_events(Events.EPOCH_STARTED) 149 | 150 | pbar_format = ( 151 | "Processing: |{bar}| " 152 | "{n_fmt}/{total_fmt}[{elapsed}<{remaining},{rate_fmt}]" 153 | ) 154 | if self.engine_name == "train": 155 | pbar_format += ( 156 | "Batch = {postfix[1][Batch]:0.5f}|" "EMA = {postfix[1][EMA]:0.5f}" 157 | ) 158 | # * changing print char may break the bar so avoid it 159 | pbar = tqdm.tqdm( 160 | total=len(self.dataloader), 161 | leave=True, 162 | initial=0, 163 | bar_format=pbar_format, 164 | ascii=True, 165 | postfix=["", dict(Batch=float("NaN"), EMA=float("NaN"))], 166 | ) 167 | else: 168 | pbar = tqdm.tqdm( 169 | total=len(self.dataloader), 170 | leave=True, 171 | bar_format=pbar_format, 172 | ascii=True, 173 | ) 174 | 175 | for data_batch in self.dataloader: 176 | data = data_batch 177 | 178 | self.__trigger_events(Events.STEP_STARTED) 179 | 180 | step_run_info = [ 181 | self.state.run_info, 182 | { 183 | "epoch": self.state.curr_epoch, 184 | "step": self.state.curr_global_step, 185 | }, 186 | ] 187 | step_output = self.run_step(data_batch, step_run_info) 188 | self.state.step_output = step_output 189 | 190 | self.__trigger_events(Events.STEP_COMPLETED) 191 | self.state.curr_global_step += 1 192 | self.state.curr_epoch_step += 1 193 | 194 | if self.engine_name == "train": 195 | pbar.postfix[1]["Batch"] = step_output["EMA"]["overall_loss"] 196 | pbar.postfix[1]["EMA"] = self.state.tracked_step_output["scalar"][ 197 | "overall_loss" 198 | ] 199 | pbar.update() 200 | pbar.close() # to flush out the bar before doing end of epoch reporting 201 | self.state.curr_epoch += 1 202 | self.__trigger_events(Events.EPOCH_COMPLETED) 203 | 204 | # TODO: [CRITICAL] align the protocol 205 | self.state.run_accumulated_output.append( 206 | self.state.epoch_accumulated_output 207 | ) 208 | 209 | return 210 | 211 | -------------------------------------------------------------------------------- /run_utils/utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import random 3 | import shutil 4 | from collections import OrderedDict 5 | 6 | import numpy as np 7 | import torch 8 | import torch.nn as nn 9 | from imgaug import imgaug as ia 10 | from termcolor import colored 11 | from torch.autograd import Variable 12 | 13 | 14 | #### 15 | def check_manual_seed(seed): 16 | """ 17 | If manual seed is not specified, choose a 18 | random one and communicate it to the user. 19 | 20 | Args: 21 | seed: seed to check 22 | """ 23 | 24 | seed = seed or random.randint(1, 10000) 25 | random.seed(seed) 26 | np.random.seed(seed) 27 | torch.manual_seed(seed) 28 | torch.cuda.manual_seed(seed) 29 | # ia.random.seed(seed) 30 | 31 | print('Using manual seed: {seed}'.format(seed=seed)) 32 | return 33 | 34 | 35 | #### 36 | def check_log_dir(log_dir): 37 | """ 38 | Check if log directory exists 39 | 40 | Args: 41 | log_dir: path to logs 42 | """ 43 | 44 | if os.path.isdir(log_dir): 45 | colored_word = colored('WARNING', color='red', attrs=['bold', 'blink']) 46 | print('%s: %s exist!' % 47 | (colored_word, colored(log_dir, attrs=['underline']))) 48 | while (True): 49 | print('Select Action: d (delete) / q (quit)', end='') 50 | key = input() 51 | if key == 'd': 52 | shutil.rmtree(log_dir) 53 | break 54 | elif key == 'q': 55 | exit() 56 | else: 57 | color_word = colored('ERR', color='red') 58 | print('---[%s] Unrecognize Characters!' % colored_word) 59 | return 60 | 61 | def get_model_summary(model, input_size, batch_size=-1, device=torch.device('cpu'), dtypes=None): 62 | """ 63 | Reusable utility layers such as pool or upsample will also get printed, but their printed values will 64 | be corresponding to the last call 65 | """ 66 | 67 | if dtypes == None: 68 | dtypes = [torch.FloatTensor]*len(input_size) 69 | 70 | summary_str = '' 71 | 72 | def register_hook(module): 73 | def hook(module, input, output): 74 | class_name = str(module.__class__).split(".")[-1].split("'")[0] 75 | module_idx = len(summary) 76 | 77 | m_key = module.name if module.name != '' else "%s" % class_name 78 | 79 | summary[m_key] = OrderedDict() 80 | summary[m_key]["input_shape"] = list(input[0].size()) 81 | summary[m_key]["input_shape"][0] = batch_size 82 | if isinstance(output, (list, tuple)): 83 | summary[m_key]["output_shape"] = [ 84 | [-1] + list(o.size())[1:] for o in output 85 | ] 86 | elif isinstance(output, dict): 87 | summary[m_key]["output_shape"] = [ 88 | [-1] + list(o.size())[1:] for o in output.values() 89 | ] 90 | elif isinstance(output, torch.Tensor): 91 | summary[m_key]["output_shape"] = list(output.size()) 92 | summary[m_key]["output_shape"][0] = batch_size 93 | 94 | params = 0 95 | if hasattr(module, "weight") and hasattr(module.weight, "size"): 96 | params += torch.prod(torch.LongTensor(list(module.weight.size()))) 97 | summary[m_key]["trainable"] = module.weight.requires_grad 98 | if hasattr(module, "bias") and hasattr(module.bias, "size"): 99 | params += torch.prod(torch.LongTensor(list(module.bias.size()))) 100 | summary[m_key]["nb_params"] = params 101 | 102 | if len(list(module.children())) == 0: 103 | hooks.append(module.register_forward_hook(hook)) 104 | 105 | # multiple inputs to the network 106 | if isinstance(input_size, tuple): 107 | input_size = [input_size] 108 | 109 | # batch_size of 2 for batchnorm 110 | x = [torch.rand(2, *in_size).type(dtype).to(device=device) 111 | for in_size, dtype in zip(input_size, dtypes)] 112 | 113 | # create properties 114 | summary = OrderedDict() 115 | hooks = [] 116 | 117 | # create layer name according to hierachy names 118 | for name, module in model.named_modules(): 119 | module.name = name 120 | 121 | # register hook 122 | model.apply(register_hook) 123 | 124 | # make a forward pass 125 | model(*x) 126 | 127 | # aligning name to the left 128 | max_name_length = len(max(summary.keys(), key=len)) 129 | summary = [(k.ljust(max_name_length), v) for k, v in summary.items()] 130 | summary = OrderedDict(summary) 131 | 132 | # remove these hooks 133 | for h in hooks: 134 | h.remove() 135 | 136 | header_line = "{} {:>25} {:>15}".format("Layer Name".center(max_name_length), "Output Shape", "Param #") 137 | summary_str += "".join('-' for _ in range(len(header_line))) + "\n" 138 | summary_str += header_line + "\n" 139 | summary_str += "".join('=' for _ in range(len(header_line))) + "\n" 140 | total_params = 0 141 | total_output = 0 142 | trainable_params = 0 143 | for layer in summary: 144 | # input_shape, output_shape, trainable, nb_params 145 | line_new = "{:>20} {:>25} {:>15}".format( 146 | layer, 147 | str(summary[layer]["output_shape"]), 148 | "{0:,}".format(summary[layer]["nb_params"]), 149 | ) 150 | total_params += summary[layer]["nb_params"] 151 | 152 | total_output += np.prod(summary[layer]["output_shape"]) 153 | if "trainable" in summary[layer]: 154 | if summary[layer]["trainable"] == True: 155 | trainable_params += summary[layer]["nb_params"] 156 | summary_str += line_new + "\n" 157 | 158 | # assume 4 bytes/number (float on cuda). 159 | total_input_size = abs(np.prod(sum(input_size, ())) * batch_size * 4. / (1024 ** 2.)) 160 | total_output_size = abs(2. * total_output * 4. / (1024 ** 2.)) # x2 for gradients 161 | total_params_size = abs(total_params * 4. / (1024 ** 2.)) 162 | total_size = total_params_size + total_output_size + total_input_size 163 | 164 | summary_str += "".join('=' for _ in range(len(header_line))) + "\n" 165 | summary_str += "Total params: {0:,}".format(total_params) + "\n" 166 | summary_str += "Trainable params: {0:,}".format(trainable_params) + "\n" 167 | summary_str += "Non-trainable params: {0:,}".format(total_params - 168 | trainable_params) + "\n" 169 | summary_str += "".join('-' for _ in range(len(header_line))) + "\n" 170 | summary_str += "Input size (MB): %0.2f" % total_input_size + "\n" 171 | summary_str += "Forward/backward pass size (MB): %0.2f" % total_output_size + "\n" 172 | summary_str += "Params size (MB): %0.2f" % total_params_size + "\n" 173 | summary_str += "Estimated Total Size (MB): %0.2f" % total_size + "\n" 174 | summary_str += "".join('-' for _ in range(len(header_line))) + "\n" 175 | return summary_str 176 | --------------------------------------------------------------------------------