├── .github └── FUNDING.yml ├── LICENSE ├── README.md ├── football1.mp4 ├── models ├── __init__.py ├── __pycache__ │ ├── __init__.cpython-36.pyc │ ├── common.cpython-36.pyc │ ├── experimental.cpython-36.pyc │ └── yolo.cpython-36.pyc ├── common.py ├── experimental.py └── yolo.py ├── pose-estimate.py ├── requirements.txt └── utils ├── __init__.py ├── __pycache__ ├── __init__.cpython-36.pyc ├── autoanchor.cpython-36.pyc ├── datasets.cpython-36.pyc ├── general.cpython-36.pyc ├── google_utils.cpython-36.pyc ├── loss.cpython-36.pyc ├── metrics.cpython-36.pyc ├── plots.cpython-36.pyc └── torch_utils.cpython-36.pyc ├── activations.py ├── add_nms.py ├── autoanchor.py ├── datasets.py ├── general.py ├── google_utils.py ├── loss.py ├── metrics.py ├── plots.py ├── torch_utils.py └── wandb_logging ├── __init__.py ├── log_dataset.py └── wandb_utils.py /.github/FUNDING.yml: -------------------------------------------------------------------------------- 1 | # These are supported funding model platforms 2 | 3 | github: # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2] 4 | patreon: # Replace with a single Patreon username 5 | open_collective: # Replace with a single Open Collective username 6 | ko_fi: # Replace with a single Ko-fi username 7 | tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel 8 | community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry 9 | liberapay: # Replace with a single Liberapay username 10 | issuehunt: # Replace with a single IssueHunt username 11 | lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry 12 | polar: # Replace with a single Polar username 13 | buy_me_a_coffee: muhammadrizwanm 14 | thanks_dev: # Replace with a single thanks.dev username 15 | custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2'] 16 | -------------------------------------------------------------------------------- /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 | # yolov7-pose-estimation 2 | 3 | ### Steps to Run Code 4 | 5 | - **Google Colab Users**: First, mount the drive: 6 | ```python 7 | from google.colab import drive 8 | drive.mount("/content/drive") 9 | ``` 10 | - **Clone the repository**: 11 | ```bash 12 | git clone https://github.com/RizwanMunawar/yolov7-pose-estimation.git 13 | ``` 14 | - **Go to the cloned folder**: 15 | ```bash 16 | cd yolov7-pose-estimation 17 | ``` 18 | - **Create a virtual environment** (recommended): 19 | ```bash 20 | # Linux 21 | python3 -m venv psestenv 22 | source psestenv/bin/activate 23 | 24 | # Windows 25 | python3 -m venv psestenv 26 | cd psestenv/Scripts 27 | activate 28 | ``` 29 | - **Upgrade pip**: 30 | ```bash 31 | pip install --upgrade pip 32 | ``` 33 | - **Install requirements**: 34 | ```bash 35 | pip install -r requirements.txt 36 | ``` 37 | - **Download YOLOv7 weights** and move to the working directory: 38 | [yolov7-w6-pose.pt](https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7-w6-pose.pt) 39 | 40 | - **Run the code**: 41 | ```bash 42 | python pose-estimate.py 43 | 44 | # Options: 45 | python pose-estimate.py --source "your-video.mp4" --device cpu # For CPU 46 | python pose-estimate.py --source 0 --view-img # For Webcam 47 | python pose-estimate.py --source "rtsp://your-ip" --device 0 --view-img # For LiveStream 48 | ``` 49 | 50 | - Output: The processed video will be saved as **your-file-keypoint.mp4** 51 | 52 | ### RESULTS 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 | 64 | 65 | 66 | 67 |
Football MatchCricket Match FPS & Time ComparisonLive Stream
68 | 69 | ### References 70 | - YOLOv7 Repo: https://github.com/WongKinYiu/yolov7 71 | - Ultralytics: https://github.com/ultralytics/yolov5 72 | 73 | ### 📖 Articles 74 | - [YOLOv7 Training Guide](https://medium.com/augmented-startups/yolov7-training-on-custom-data-b86d23e6623) 75 | - [Computer Vision Roadmap](https://medium.com/augmented-startups/roadmap-for-computer-vision-engineer-45167b94518c) 76 | -------------------------------------------------------------------------------- /football1.mp4: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/football1.mp4 -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- 1 | # init -------------------------------------------------------------------------------- /models/__pycache__/__init__.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/models/__pycache__/__init__.cpython-36.pyc -------------------------------------------------------------------------------- /models/__pycache__/common.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/models/__pycache__/common.cpython-36.pyc -------------------------------------------------------------------------------- /models/__pycache__/experimental.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/models/__pycache__/experimental.cpython-36.pyc -------------------------------------------------------------------------------- /models/__pycache__/yolo.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/models/__pycache__/yolo.cpython-36.pyc -------------------------------------------------------------------------------- /models/experimental.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import random 3 | import torch 4 | import torch.nn as nn 5 | 6 | from models.common import Conv, DWConv 7 | from utils.google_utils import attempt_download 8 | 9 | 10 | class CrossConv(nn.Module): 11 | # Cross Convolution Downsample 12 | def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): 13 | # ch_in, ch_out, kernel, stride, groups, expansion, shortcut 14 | super(CrossConv, self).__init__() 15 | c_ = int(c2 * e) # hidden channels 16 | self.cv1 = Conv(c1, c_, (1, k), (1, s)) 17 | self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) 18 | self.add = shortcut and c1 == c2 19 | 20 | def forward(self, x): 21 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 22 | 23 | 24 | class Sum(nn.Module): 25 | # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 26 | def __init__(self, n, weight=False): # n: number of inputs 27 | super(Sum, self).__init__() 28 | self.weight = weight # apply weights boolean 29 | self.iter = range(n - 1) # iter object 30 | if weight: 31 | self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights 32 | 33 | def forward(self, x): 34 | y = x[0] # no weight 35 | if self.weight: 36 | w = torch.sigmoid(self.w) * 2 37 | for i in self.iter: 38 | y = y + x[i + 1] * w[i] 39 | else: 40 | for i in self.iter: 41 | y = y + x[i + 1] 42 | return y 43 | 44 | 45 | class MixConv2d(nn.Module): 46 | # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 47 | def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): 48 | super(MixConv2d, self).__init__() 49 | groups = len(k) 50 | if equal_ch: # equal c_ per group 51 | i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices 52 | c_ = [(i == g).sum() for g in range(groups)] # intermediate channels 53 | else: # equal weight.numel() per group 54 | b = [c2] + [0] * groups 55 | a = np.eye(groups + 1, groups, k=-1) 56 | a -= np.roll(a, 1, axis=1) 57 | a *= np.array(k) ** 2 58 | a[0] = 1 59 | c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b 60 | 61 | self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) 62 | self.bn = nn.BatchNorm2d(c2) 63 | self.act = nn.LeakyReLU(0.1, inplace=True) 64 | 65 | def forward(self, x): 66 | return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) 67 | 68 | 69 | class Ensemble(nn.ModuleList): 70 | # Ensemble of models 71 | def __init__(self): 72 | super(Ensemble, self).__init__() 73 | 74 | def forward(self, x, augment=False): 75 | y = [] 76 | for module in self: 77 | y.append(module(x, augment)[0]) 78 | # y = torch.stack(y).max(0)[0] # max ensemble 79 | # y = torch.stack(y).mean(0) # mean ensemble 80 | y = torch.cat(y, 1) # nms ensemble 81 | return y, None # inference, train output 82 | 83 | 84 | 85 | 86 | 87 | class ORT_NMS(torch.autograd.Function): 88 | '''ONNX-Runtime NMS operation''' 89 | @staticmethod 90 | def forward(ctx, 91 | boxes, 92 | scores, 93 | max_output_boxes_per_class=torch.tensor([100]), 94 | iou_threshold=torch.tensor([0.45]), 95 | score_threshold=torch.tensor([0.25])): 96 | device = boxes.device 97 | batch = scores.shape[0] 98 | num_det = random.randint(0, 100) 99 | batches = torch.randint(0, batch, (num_det,)).sort()[0].to(device) 100 | idxs = torch.arange(100, 100 + num_det).to(device) 101 | zeros = torch.zeros((num_det,), dtype=torch.int64).to(device) 102 | selected_indices = torch.cat([batches[None], zeros[None], idxs[None]], 0).T.contiguous() 103 | selected_indices = selected_indices.to(torch.int64) 104 | return selected_indices 105 | 106 | @staticmethod 107 | def symbolic(g, boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold): 108 | return g.op("NonMaxSuppression", boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold) 109 | 110 | 111 | class TRT_NMS(torch.autograd.Function): 112 | '''TensorRT NMS operation''' 113 | @staticmethod 114 | def forward( 115 | ctx, 116 | boxes, 117 | scores, 118 | background_class=-1, 119 | box_coding=1, 120 | iou_threshold=0.45, 121 | max_output_boxes=100, 122 | plugin_version="1", 123 | score_activation=0, 124 | score_threshold=0.25, 125 | ): 126 | batch_size, num_boxes, num_classes = scores.shape 127 | num_det = torch.randint(0, max_output_boxes, (batch_size, 1), dtype=torch.int32) 128 | det_boxes = torch.randn(batch_size, max_output_boxes, 4) 129 | det_scores = torch.randn(batch_size, max_output_boxes) 130 | det_classes = torch.randint(0, num_classes, (batch_size, max_output_boxes), dtype=torch.int32) 131 | return num_det, det_boxes, det_scores, det_classes 132 | 133 | @staticmethod 134 | def symbolic(g, 135 | boxes, 136 | scores, 137 | background_class=-1, 138 | box_coding=1, 139 | iou_threshold=0.45, 140 | max_output_boxes=100, 141 | plugin_version="1", 142 | score_activation=0, 143 | score_threshold=0.25): 144 | out = g.op("TRT::EfficientNMS_TRT", 145 | boxes, 146 | scores, 147 | background_class_i=background_class, 148 | box_coding_i=box_coding, 149 | iou_threshold_f=iou_threshold, 150 | max_output_boxes_i=max_output_boxes, 151 | plugin_version_s=plugin_version, 152 | score_activation_i=score_activation, 153 | score_threshold_f=score_threshold, 154 | outputs=4) 155 | nums, boxes, scores, classes = out 156 | return nums, boxes, scores, classes 157 | 158 | 159 | class ONNX_ORT(nn.Module): 160 | '''onnx module with ONNX-Runtime NMS operation.''' 161 | def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=640, device=None): 162 | super().__init__() 163 | self.device = device if device else torch.device("cpu") 164 | self.max_obj = torch.tensor([max_obj]).to(device) 165 | self.iou_threshold = torch.tensor([iou_thres]).to(device) 166 | self.score_threshold = torch.tensor([score_thres]).to(device) 167 | self.max_wh = max_wh # if max_wh != 0 : non-agnostic else : agnostic 168 | self.convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]], 169 | dtype=torch.float32, 170 | device=self.device) 171 | 172 | def forward(self, x): 173 | boxes = x[:, :, :4] 174 | conf = x[:, :, 4:5] 175 | scores = x[:, :, 5:] 176 | scores *= conf 177 | boxes @= self.convert_matrix 178 | max_score, category_id = scores.max(2, keepdim=True) 179 | dis = category_id.float() * self.max_wh 180 | nmsbox = boxes + dis 181 | max_score_tp = max_score.transpose(1, 2).contiguous() 182 | selected_indices = ORT_NMS.apply(nmsbox, max_score_tp, self.max_obj, self.iou_threshold, self.score_threshold) 183 | X, Y = selected_indices[:, 0], selected_indices[:, 2] 184 | selected_boxes = boxes[X, Y, :] 185 | selected_categories = category_id[X, Y, :].float() 186 | selected_scores = max_score[X, Y, :] 187 | X = X.unsqueeze(1).float() 188 | return torch.cat([X, selected_boxes, selected_categories, selected_scores], 1) 189 | 190 | class ONNX_TRT(nn.Module): 191 | '''onnx module with TensorRT NMS operation.''' 192 | def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None ,device=None): 193 | super().__init__() 194 | assert max_wh is None 195 | self.device = device if device else torch.device('cpu') 196 | self.background_class = -1, 197 | self.box_coding = 1, 198 | self.iou_threshold = iou_thres 199 | self.max_obj = max_obj 200 | self.plugin_version = '1' 201 | self.score_activation = 0 202 | self.score_threshold = score_thres 203 | 204 | def forward(self, x): 205 | boxes = x[:, :, :4] 206 | conf = x[:, :, 4:5] 207 | scores = x[:, :, 5:] 208 | scores *= conf 209 | num_det, det_boxes, det_scores, det_classes = TRT_NMS.apply(boxes, scores, self.background_class, self.box_coding, 210 | self.iou_threshold, self.max_obj, 211 | self.plugin_version, self.score_activation, 212 | self.score_threshold) 213 | return num_det, det_boxes, det_scores, det_classes 214 | 215 | 216 | class End2End(nn.Module): 217 | '''export onnx or tensorrt model with NMS operation.''' 218 | def __init__(self, model, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None, device=None): 219 | super().__init__() 220 | device = device if device else torch.device('cpu') 221 | assert isinstance(max_wh,(int)) or max_wh is None 222 | self.model = model.to(device) 223 | self.model.model[-1].end2end = True 224 | self.patch_model = ONNX_TRT if max_wh is None else ONNX_ORT 225 | self.end2end = self.patch_model(max_obj, iou_thres, score_thres, max_wh, device) 226 | self.end2end.eval() 227 | 228 | def forward(self, x): 229 | x = self.model(x) 230 | x = self.end2end(x) 231 | return x 232 | 233 | 234 | 235 | 236 | 237 | def attempt_load(weights, map_location=None): 238 | # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a 239 | model = Ensemble() 240 | for w in weights if isinstance(weights, list) else [weights]: 241 | attempt_download(w) 242 | ckpt = torch.load(w, map_location=map_location) # load 243 | model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model 244 | 245 | # Compatibility updates 246 | for m in model.modules(): 247 | if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: 248 | m.inplace = True # pytorch 1.7.0 compatibility 249 | elif type(m) is nn.Upsample: 250 | m.recompute_scale_factor = None # torch 1.11.0 compatibility 251 | elif type(m) is Conv: 252 | m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility 253 | 254 | if len(model) == 1: 255 | return model[-1] # return model 256 | else: 257 | print('Ensemble created with %s\n' % weights) 258 | for k in ['names', 'stride']: 259 | setattr(model, k, getattr(model[-1], k)) 260 | return model # return ensemble 261 | 262 | 263 | -------------------------------------------------------------------------------- /models/yolo.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import logging 3 | import sys 4 | from copy import deepcopy 5 | 6 | sys.path.append('./') # to run '$ python *.py' files in subdirectories 7 | logger = logging.getLogger(__name__) 8 | import torch 9 | from models.common import * 10 | from models.experimental import * 11 | from utils.autoanchor import check_anchor_order 12 | from utils.general import make_divisible, check_file, set_logging 13 | from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \ 14 | select_device, copy_attr 15 | from utils.loss import SigmoidBin 16 | 17 | try: 18 | import thop # for FLOPS computation 19 | except ImportError: 20 | thop = None 21 | 22 | 23 | class Detect(nn.Module): 24 | stride = None # strides computed during build 25 | export = False # onnx export 26 | end2end = False 27 | include_nms = False 28 | 29 | def __init__(self, nc=80, anchors=(), ch=()): # detection layer 30 | super(Detect, self).__init__() 31 | self.nc = nc # number of classes 32 | self.no = nc + 5 # number of outputs per anchor 33 | self.nl = len(anchors) # number of detection layers 34 | self.na = len(anchors[0]) // 2 # number of anchors 35 | self.grid = [torch.zeros(1)] * self.nl # init grid 36 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 37 | self.register_buffer('anchors', a) # shape(nl,na,2) 38 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 39 | self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 40 | 41 | def forward(self, x): 42 | # x = x.copy() # for profiling 43 | z = [] # inference output 44 | self.training |= self.export 45 | for i in range(self.nl): 46 | x[i] = self.m[i](x[i]) # conv 47 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 48 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 49 | 50 | if not self.training: # inference 51 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 52 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 53 | y = x[i].sigmoid() 54 | if not torch.onnx.is_in_onnx_export(): 55 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 56 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 57 | else: 58 | xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 59 | wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 60 | y = torch.cat((xy, wh, y[..., 4:]), -1) 61 | z.append(y.view(bs, -1, self.no)) 62 | 63 | if self.training: 64 | out = x 65 | elif self.end2end: 66 | out = torch.cat(z, 1) 67 | elif self.include_nms: 68 | z = self.convert(z) 69 | out = (z, ) 70 | else: 71 | out = (torch.cat(z, 1), x) 72 | 73 | return out 74 | 75 | @staticmethod 76 | def _make_grid(nx=20, ny=20): 77 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 78 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 79 | 80 | def convert(self, z): 81 | z = torch.cat(z, 1) 82 | box = z[:, :, :4] 83 | conf = z[:, :, 4:5] 84 | score = z[:, :, 5:] 85 | score *= conf 86 | convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]], 87 | dtype=torch.float32, 88 | device=z.device) 89 | box @= convert_matrix 90 | return (box, score) 91 | 92 | 93 | class IDetect(nn.Module): 94 | stride = None # strides computed during build 95 | export = False # onnx export 96 | end2end = False 97 | include_nms = False 98 | 99 | def __init__(self, nc=80, anchors=(), ch=()): # detection layer 100 | super(IDetect, self).__init__() 101 | self.nc = nc # number of classes 102 | self.no = nc + 5 # number of outputs per anchor 103 | self.nl = len(anchors) # number of detection layers 104 | self.na = len(anchors[0]) // 2 # number of anchors 105 | self.grid = [torch.zeros(1)] * self.nl # init grid 106 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 107 | self.register_buffer('anchors', a) # shape(nl,na,2) 108 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 109 | self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 110 | 111 | self.ia = nn.ModuleList(ImplicitA(x) for x in ch) 112 | self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch) 113 | 114 | def forward(self, x): 115 | # x = x.copy() # for profiling 116 | z = [] # inference output 117 | self.training |= self.export 118 | for i in range(self.nl): 119 | x[i] = self.m[i](self.ia[i](x[i])) # conv 120 | x[i] = self.im[i](x[i]) 121 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 122 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 123 | 124 | if not self.training: # inference 125 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 126 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 127 | 128 | y = x[i].sigmoid() 129 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 130 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 131 | z.append(y.view(bs, -1, self.no)) 132 | 133 | return x if self.training else (torch.cat(z, 1), x) 134 | 135 | def fuseforward(self, x): 136 | # x = x.copy() # for profiling 137 | z = [] # inference output 138 | self.training |= self.export 139 | for i in range(self.nl): 140 | x[i] = self.m[i](x[i]) # conv 141 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 142 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 143 | 144 | if not self.training: # inference 145 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 146 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 147 | 148 | y = x[i].sigmoid() 149 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 150 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 151 | z.append(y.view(bs, -1, self.no)) 152 | 153 | if self.training: 154 | out = x 155 | elif self.end2end: 156 | out = torch.cat(z, 1) 157 | elif self.include_nms: 158 | z = self.convert(z) 159 | out = (z, ) 160 | else: 161 | out = (torch.cat(z, 1), x) 162 | 163 | return out 164 | 165 | def fuse(self): 166 | print("IDetect.fuse") 167 | # fuse ImplicitA and Convolution 168 | for i in range(len(self.m)): 169 | c1,c2,_,_ = self.m[i].weight.shape 170 | c1_,c2_, _,_ = self.ia[i].implicit.shape 171 | self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1) 172 | 173 | # fuse ImplicitM and Convolution 174 | for i in range(len(self.m)): 175 | c1,c2, _,_ = self.im[i].implicit.shape 176 | self.m[i].bias *= self.im[i].implicit.reshape(c2) 177 | self.m[i].weight *= self.im[i].implicit.transpose(0,1) 178 | 179 | @staticmethod 180 | def _make_grid(nx=20, ny=20): 181 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 182 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 183 | 184 | def convert(self, z): 185 | z = torch.cat(z, 1) 186 | box = z[:, :, :4] 187 | conf = z[:, :, 4:5] 188 | score = z[:, :, 5:] 189 | score *= conf 190 | convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]], 191 | dtype=torch.float32, 192 | device=z.device) 193 | box @= convert_matrix 194 | return (box, score) 195 | 196 | 197 | class IKeypoint(nn.Module): 198 | stride = None # strides computed during build 199 | export = False # onnx export 200 | 201 | def __init__(self, nc=80, anchors=(), nkpt=17, ch=(), inplace=True, dw_conv_kpt=False): # detection layer 202 | super(IKeypoint, self).__init__() 203 | self.nc = nc # number of classes 204 | self.nkpt = nkpt 205 | self.dw_conv_kpt = dw_conv_kpt 206 | self.no_det=(nc + 5) # number of outputs per anchor for box and class 207 | self.no_kpt = 3*self.nkpt ## number of outputs per anchor for keypoints 208 | self.no = self.no_det+self.no_kpt 209 | self.nl = len(anchors) # number of detection layers 210 | self.na = len(anchors[0]) // 2 # number of anchors 211 | self.grid = [torch.zeros(1)] * self.nl # init grid 212 | self.flip_test = False 213 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 214 | self.register_buffer('anchors', a) # shape(nl,na,2) 215 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 216 | self.m = nn.ModuleList(nn.Conv2d(x, self.no_det * self.na, 1) for x in ch) # output conv 217 | 218 | self.ia = nn.ModuleList(ImplicitA(x) for x in ch) 219 | self.im = nn.ModuleList(ImplicitM(self.no_det * self.na) for _ in ch) 220 | 221 | if self.nkpt is not None: 222 | if self.dw_conv_kpt: #keypoint head is slightly more complex 223 | self.m_kpt = nn.ModuleList( 224 | nn.Sequential(DWConv(x, x, k=3), Conv(x,x), 225 | DWConv(x, x, k=3), Conv(x, x), 226 | DWConv(x, x, k=3), Conv(x,x), 227 | DWConv(x, x, k=3), Conv(x, x), 228 | DWConv(x, x, k=3), Conv(x, x), 229 | DWConv(x, x, k=3), nn.Conv2d(x, self.no_kpt * self.na, 1)) for x in ch) 230 | else: #keypoint head is a single convolution 231 | self.m_kpt = nn.ModuleList(nn.Conv2d(x, self.no_kpt * self.na, 1) for x in ch) 232 | 233 | self.inplace = inplace # use in-place ops (e.g. slice assignment) 234 | 235 | def forward(self, x): 236 | # x = x.copy() # for profiling 237 | z = [] # inference output 238 | self.training |= self.export 239 | for i in range(self.nl): 240 | if self.nkpt is None or self.nkpt==0: 241 | x[i] = self.im[i](self.m[i](self.ia[i](x[i]))) # conv 242 | else : 243 | x[i] = torch.cat((self.im[i](self.m[i](self.ia[i](x[i]))), self.m_kpt[i](x[i])), axis=1) 244 | 245 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 246 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 247 | x_det = x[i][..., :6] 248 | x_kpt = x[i][..., 6:] 249 | 250 | if not self.training: # inference 251 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 252 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 253 | kpt_grid_x = self.grid[i][..., 0:1] 254 | kpt_grid_y = self.grid[i][..., 1:2] 255 | 256 | if self.nkpt == 0: 257 | y = x[i].sigmoid() 258 | else: 259 | y = x_det.sigmoid() 260 | 261 | if self.inplace: 262 | xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 263 | wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2) # wh 264 | if self.nkpt != 0: 265 | x_kpt[..., 0::3] = (x_kpt[..., ::3] * 2. - 0.5 + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy 266 | x_kpt[..., 1::3] = (x_kpt[..., 1::3] * 2. - 0.5 + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy 267 | #x_kpt[..., 0::3] = (x_kpt[..., ::3] + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy 268 | #x_kpt[..., 1::3] = (x_kpt[..., 1::3] + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy 269 | #print('=============') 270 | #print(self.anchor_grid[i].shape) 271 | #print(self.anchor_grid[i][...,0].unsqueeze(4).shape) 272 | #print(x_kpt[..., 0::3].shape) 273 | #x_kpt[..., 0::3] = ((x_kpt[..., 0::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy 274 | #x_kpt[..., 1::3] = ((x_kpt[..., 1::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy 275 | #x_kpt[..., 0::3] = (((x_kpt[..., 0::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy 276 | #x_kpt[..., 1::3] = (((x_kpt[..., 1::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy 277 | x_kpt[..., 2::3] = x_kpt[..., 2::3].sigmoid() 278 | 279 | y = torch.cat((xy, wh, y[..., 4:], x_kpt), dim = -1) 280 | 281 | else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953 282 | xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 283 | wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 284 | if self.nkpt != 0: 285 | y[..., 6:] = (y[..., 6:] * 2. - 0.5 + self.grid[i].repeat((1,1,1,1,self.nkpt))) * self.stride[i] # xy 286 | y = torch.cat((xy, wh, y[..., 4:]), -1) 287 | 288 | z.append(y.view(bs, -1, self.no)) 289 | 290 | return x if self.training else (torch.cat(z, 1), x) 291 | 292 | @staticmethod 293 | def _make_grid(nx=20, ny=20): 294 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 295 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 296 | 297 | 298 | class IAuxDetect(nn.Module): 299 | stride = None # strides computed during build 300 | export = False # onnx export 301 | 302 | def __init__(self, nc=80, anchors=(), ch=()): # detection layer 303 | super(IAuxDetect, self).__init__() 304 | self.nc = nc # number of classes 305 | self.no = nc + 5 # number of outputs per anchor 306 | self.nl = len(anchors) # number of detection layers 307 | self.na = len(anchors[0]) // 2 # number of anchors 308 | self.grid = [torch.zeros(1)] * self.nl # init grid 309 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 310 | self.register_buffer('anchors', a) # shape(nl,na,2) 311 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 312 | self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[:self.nl]) # output conv 313 | self.m2 = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[self.nl:]) # output conv 314 | 315 | self.ia = nn.ModuleList(ImplicitA(x) for x in ch[:self.nl]) 316 | self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch[:self.nl]) 317 | 318 | def forward(self, x): 319 | # x = x.copy() # for profiling 320 | z = [] # inference output 321 | self.training |= self.export 322 | for i in range(self.nl): 323 | x[i] = self.m[i](self.ia[i](x[i])) # conv 324 | x[i] = self.im[i](x[i]) 325 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 326 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 327 | 328 | x[i+self.nl] = self.m2[i](x[i+self.nl]) 329 | x[i+self.nl] = x[i+self.nl].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 330 | 331 | if not self.training: # inference 332 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 333 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 334 | 335 | y = x[i].sigmoid() 336 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 337 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 338 | z.append(y.view(bs, -1, self.no)) 339 | 340 | return x if self.training else (torch.cat(z, 1), x[:self.nl]) 341 | 342 | @staticmethod 343 | def _make_grid(nx=20, ny=20): 344 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 345 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 346 | 347 | 348 | class IBin(nn.Module): 349 | stride = None # strides computed during build 350 | export = False # onnx export 351 | 352 | def __init__(self, nc=80, anchors=(), ch=(), bin_count=21): # detection layer 353 | super(IBin, self).__init__() 354 | self.nc = nc # number of classes 355 | self.bin_count = bin_count 356 | 357 | self.w_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0) 358 | self.h_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0) 359 | # classes, x,y,obj 360 | self.no = nc + 3 + \ 361 | self.w_bin_sigmoid.get_length() + self.h_bin_sigmoid.get_length() # w-bce, h-bce 362 | # + self.x_bin_sigmoid.get_length() + self.y_bin_sigmoid.get_length() 363 | 364 | self.nl = len(anchors) # number of detection layers 365 | self.na = len(anchors[0]) // 2 # number of anchors 366 | self.grid = [torch.zeros(1)] * self.nl # init grid 367 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 368 | self.register_buffer('anchors', a) # shape(nl,na,2) 369 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 370 | self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 371 | 372 | self.ia = nn.ModuleList(ImplicitA(x) for x in ch) 373 | self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch) 374 | 375 | def forward(self, x): 376 | 377 | #self.x_bin_sigmoid.use_fw_regression = True 378 | #self.y_bin_sigmoid.use_fw_regression = True 379 | self.w_bin_sigmoid.use_fw_regression = True 380 | self.h_bin_sigmoid.use_fw_regression = True 381 | 382 | # x = x.copy() # for profiling 383 | z = [] # inference output 384 | self.training |= self.export 385 | for i in range(self.nl): 386 | x[i] = self.m[i](self.ia[i](x[i])) # conv 387 | x[i] = self.im[i](x[i]) 388 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 389 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 390 | 391 | if not self.training: # inference 392 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 393 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 394 | 395 | y = x[i].sigmoid() 396 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy 397 | #y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 398 | 399 | 400 | #px = (self.x_bin_sigmoid.forward(y[..., 0:12]) + self.grid[i][..., 0]) * self.stride[i] 401 | #py = (self.y_bin_sigmoid.forward(y[..., 12:24]) + self.grid[i][..., 1]) * self.stride[i] 402 | 403 | pw = self.w_bin_sigmoid.forward(y[..., 2:24]) * self.anchor_grid[i][..., 0] 404 | ph = self.h_bin_sigmoid.forward(y[..., 24:46]) * self.anchor_grid[i][..., 1] 405 | 406 | #y[..., 0] = px 407 | #y[..., 1] = py 408 | y[..., 2] = pw 409 | y[..., 3] = ph 410 | 411 | y = torch.cat((y[..., 0:4], y[..., 46:]), dim=-1) 412 | 413 | z.append(y.view(bs, -1, y.shape[-1])) 414 | 415 | return x if self.training else (torch.cat(z, 1), x) 416 | 417 | @staticmethod 418 | def _make_grid(nx=20, ny=20): 419 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 420 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 421 | 422 | 423 | class Model(nn.Module): 424 | def __init__(self, cfg='yolor-csp-c.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes 425 | super(Model, self).__init__() 426 | self.traced = False 427 | if isinstance(cfg, dict): 428 | self.yaml = cfg # model dict 429 | else: # is *.yaml 430 | import yaml # for torch hub 431 | self.yaml_file = Path(cfg).name 432 | with open(cfg) as f: 433 | self.yaml = yaml.load(f, Loader=yaml.SafeLoader) # model dict 434 | 435 | # Define model 436 | ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels 437 | if nc and nc != self.yaml['nc']: 438 | logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") 439 | self.yaml['nc'] = nc # override yaml value 440 | if anchors: 441 | logger.info(f'Overriding model.yaml anchors with anchors={anchors}') 442 | self.yaml['anchors'] = round(anchors) # override yaml value 443 | self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist 444 | self.names = [str(i) for i in range(self.yaml['nc'])] # default names 445 | # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))]) 446 | 447 | # Build strides, anchors 448 | m = self.model[-1] # Detect() 449 | if isinstance(m, Detect): 450 | s = 256 # 2x min stride 451 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward 452 | m.anchors /= m.stride.view(-1, 1, 1) 453 | check_anchor_order(m) 454 | self.stride = m.stride 455 | self._initialize_biases() # only run once 456 | # print('Strides: %s' % m.stride.tolist()) 457 | if isinstance(m, IDetect): 458 | s = 256 # 2x min stride 459 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward 460 | m.anchors /= m.stride.view(-1, 1, 1) 461 | check_anchor_order(m) 462 | self.stride = m.stride 463 | self._initialize_biases() # only run once 464 | # print('Strides: %s' % m.stride.tolist()) 465 | if isinstance(m, IAuxDetect): 466 | s = 256 # 2x min stride 467 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))[:4]]) # forward 468 | #print(m.stride) 469 | m.anchors /= m.stride.view(-1, 1, 1) 470 | check_anchor_order(m) 471 | self.stride = m.stride 472 | self._initialize_aux_biases() # only run once 473 | # print('Strides: %s' % m.stride.tolist()) 474 | if isinstance(m, IBin): 475 | s = 256 # 2x min stride 476 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward 477 | m.anchors /= m.stride.view(-1, 1, 1) 478 | check_anchor_order(m) 479 | self.stride = m.stride 480 | self._initialize_biases_bin() # only run once 481 | # print('Strides: %s' % m.stride.tolist()) 482 | if isinstance(m, IKeypoint): 483 | s = 256 # 2x min stride 484 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward 485 | m.anchors /= m.stride.view(-1, 1, 1) 486 | check_anchor_order(m) 487 | self.stride = m.stride 488 | self._initialize_biases_kpt() # only run once 489 | # print('Strides: %s' % m.stride.tolist()) 490 | 491 | # Init weights, biases 492 | initialize_weights(self) 493 | self.info() 494 | logger.info('') 495 | 496 | def forward(self, x, augment=False, profile=False): 497 | if augment: 498 | img_size = x.shape[-2:] # height, width 499 | s = [1, 0.83, 0.67] # scales 500 | f = [None, 3, None] # flips (2-ud, 3-lr) 501 | y = [] # outputs 502 | for si, fi in zip(s, f): 503 | xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) 504 | yi = self.forward_once(xi)[0] # forward 505 | # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save 506 | yi[..., :4] /= si # de-scale 507 | if fi == 2: 508 | yi[..., 1] = img_size[0] - yi[..., 1] # de-flip ud 509 | elif fi == 3: 510 | yi[..., 0] = img_size[1] - yi[..., 0] # de-flip lr 511 | y.append(yi) 512 | return torch.cat(y, 1), None # augmented inference, train 513 | else: 514 | return self.forward_once(x, profile) # single-scale inference, train 515 | 516 | def forward_once(self, x, profile=False): 517 | y, dt = [], [] # outputs 518 | for m in self.model: 519 | if m.f != -1: # if not from previous layer 520 | x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers 521 | 522 | if not hasattr(self, 'traced'): 523 | self.traced=False 524 | 525 | if self.traced: 526 | if isinstance(m, Detect) or isinstance(m, IDetect) or isinstance(m, IAuxDetect) or isinstance(m, IKeypoint): 527 | break 528 | 529 | if profile: 530 | c = isinstance(m, (Detect, IDetect, IAuxDetect, IBin)) 531 | o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS 532 | for _ in range(10): 533 | m(x.copy() if c else x) 534 | t = time_synchronized() 535 | for _ in range(10): 536 | m(x.copy() if c else x) 537 | dt.append((time_synchronized() - t) * 100) 538 | print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type)) 539 | 540 | x = m(x) # run 541 | 542 | y.append(x if m.i in self.save else None) # save output 543 | 544 | if profile: 545 | print('%.1fms total' % sum(dt)) 546 | return x 547 | 548 | def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency 549 | # https://arxiv.org/abs/1708.02002 section 3.3 550 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 551 | m = self.model[-1] # Detect() module 552 | for mi, s in zip(m.m, m.stride): # from 553 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 554 | b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 555 | b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 556 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 557 | 558 | def _initialize_aux_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency 559 | # https://arxiv.org/abs/1708.02002 section 3.3 560 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 561 | m = self.model[-1] # Detect() module 562 | for mi, mi2, s in zip(m.m, m.m2, m.stride): # from 563 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 564 | b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 565 | b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 566 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 567 | b2 = mi2.bias.view(m.na, -1) # conv.bias(255) to (3,85) 568 | b2.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 569 | b2.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 570 | mi2.bias = torch.nn.Parameter(b2.view(-1), requires_grad=True) 571 | 572 | def _initialize_biases_bin(self, cf=None): # initialize biases into Detect(), cf is class frequency 573 | # https://arxiv.org/abs/1708.02002 section 3.3 574 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 575 | m = self.model[-1] # Bin() module 576 | bc = m.bin_count 577 | for mi, s in zip(m.m, m.stride): # from 578 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 579 | old = b[:, (0,1,2,bc+3)].data 580 | obj_idx = 2*bc+4 581 | b[:, :obj_idx].data += math.log(0.6 / (bc + 1 - 0.99)) 582 | b[:, obj_idx].data += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 583 | b[:, (obj_idx+1):].data += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 584 | b[:, (0,1,2,bc+3)].data = old 585 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 586 | 587 | def _initialize_biases_kpt(self, cf=None): # initialize biases into Detect(), cf is class frequency 588 | # https://arxiv.org/abs/1708.02002 section 3.3 589 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 590 | m = self.model[-1] # Detect() module 591 | for mi, s in zip(m.m, m.stride): # from 592 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 593 | b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 594 | b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 595 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 596 | 597 | def _print_biases(self): 598 | m = self.model[-1] # Detect() module 599 | for mi in m.m: # from 600 | b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) 601 | print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean())) 602 | 603 | # def _print_weights(self): 604 | # for m in self.model.modules(): 605 | # if type(m) is Bottleneck: 606 | # print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights 607 | 608 | def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers 609 | print('Fusing layers... ') 610 | for m in self.model.modules(): 611 | if isinstance(m, RepConv): 612 | #print(f" fuse_repvgg_block") 613 | m.fuse_repvgg_block() 614 | elif isinstance(m, RepConv_OREPA): 615 | #print(f" switch_to_deploy") 616 | m.switch_to_deploy() 617 | elif type(m) is Conv and hasattr(m, 'bn'): 618 | m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv 619 | delattr(m, 'bn') # remove batchnorm 620 | m.forward = m.fuseforward # update forward 621 | elif isinstance(m, IDetect): 622 | m.fuse() 623 | m.forward = m.fuseforward 624 | self.info() 625 | return self 626 | 627 | def nms(self, mode=True): # add or remove NMS module 628 | present = type(self.model[-1]) is NMS # last layer is NMS 629 | if mode and not present: 630 | print('Adding NMS... ') 631 | m = NMS() # module 632 | m.f = -1 # from 633 | m.i = self.model[-1].i + 1 # index 634 | self.model.add_module(name='%s' % m.i, module=m) # add 635 | self.eval() 636 | elif not mode and present: 637 | print('Removing NMS... ') 638 | self.model = self.model[:-1] # remove 639 | return self 640 | 641 | def autoshape(self): # add autoShape module 642 | print('Adding autoShape... ') 643 | m = autoShape(self) # wrap model 644 | copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=()) # copy attributes 645 | return m 646 | 647 | def info(self, verbose=False, img_size=640): # print model information 648 | model_info(self, verbose, img_size) 649 | 650 | 651 | def parse_model(d, ch): # model_dict, input_channels(3) 652 | logger.info('\n%3s%18s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments')) 653 | anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] 654 | na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors 655 | no = na * (nc + 5) # number of outputs = anchors * (classes + 5) 656 | 657 | layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out 658 | for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args 659 | m = eval(m) if isinstance(m, str) else m # eval strings 660 | for j, a in enumerate(args): 661 | try: 662 | args[j] = eval(a) if isinstance(a, str) else a # eval strings 663 | except: 664 | pass 665 | 666 | n = max(round(n * gd), 1) if n > 1 else n # depth gain 667 | if m in [nn.Conv2d, Conv, RobustConv, RobustConv2, DWConv, GhostConv, RepConv, RepConv_OREPA, DownC, 668 | SPP, SPPF, SPPCSPC, GhostSPPCSPC, MixConv2d, Focus, Stem, GhostStem, CrossConv, 669 | Bottleneck, BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, 670 | RepBottleneck, RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC, 671 | Res, ResCSPA, ResCSPB, ResCSPC, 672 | RepRes, RepResCSPA, RepResCSPB, RepResCSPC, 673 | ResX, ResXCSPA, ResXCSPB, ResXCSPC, 674 | RepResX, RepResXCSPA, RepResXCSPB, RepResXCSPC, 675 | Ghost, GhostCSPA, GhostCSPB, GhostCSPC, 676 | SwinTransformerBlock, STCSPA, STCSPB, STCSPC, 677 | SwinTransformer2Block, ST2CSPA, ST2CSPB, ST2CSPC]: 678 | c1, c2 = ch[f], args[0] 679 | if c2 != no: # if not output 680 | c2 = make_divisible(c2 * gw, 8) 681 | 682 | args = [c1, c2, *args[1:]] 683 | if m in [DownC, SPPCSPC, GhostSPPCSPC, 684 | BottleneckCSPA, BottleneckCSPB, BottleneckCSPC, 685 | RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC, 686 | ResCSPA, ResCSPB, ResCSPC, 687 | RepResCSPA, RepResCSPB, RepResCSPC, 688 | ResXCSPA, ResXCSPB, ResXCSPC, 689 | RepResXCSPA, RepResXCSPB, RepResXCSPC, 690 | GhostCSPA, GhostCSPB, GhostCSPC, 691 | STCSPA, STCSPB, STCSPC, 692 | ST2CSPA, ST2CSPB, ST2CSPC]: 693 | args.insert(2, n) # number of repeats 694 | n = 1 695 | elif m is nn.BatchNorm2d: 696 | args = [ch[f]] 697 | elif m is Concat: 698 | c2 = sum([ch[x] for x in f]) 699 | elif m is Chuncat: 700 | c2 = sum([ch[x] for x in f]) 701 | elif m is Shortcut: 702 | c2 = ch[f[0]] 703 | elif m is Foldcut: 704 | c2 = ch[f] // 2 705 | elif m in [Detect, IDetect, IAuxDetect, IBin, IKeypoint]: 706 | args.append([ch[x] for x in f]) 707 | if isinstance(args[1], int): # number of anchors 708 | args[1] = [list(range(args[1] * 2))] * len(f) 709 | elif m is ReOrg: 710 | c2 = ch[f] * 4 711 | elif m is Contract: 712 | c2 = ch[f] * args[0] ** 2 713 | elif m is Expand: 714 | c2 = ch[f] // args[0] ** 2 715 | else: 716 | c2 = ch[f] 717 | 718 | m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module 719 | t = str(m)[8:-2].replace('__main__.', '') # module type 720 | np = sum([x.numel() for x in m_.parameters()]) # number params 721 | m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params 722 | logger.info('%3s%18s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print 723 | save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist 724 | layers.append(m_) 725 | if i == 0: 726 | ch = [] 727 | ch.append(c2) 728 | return nn.Sequential(*layers), sorted(save) 729 | 730 | 731 | if __name__ == '__main__': 732 | parser = argparse.ArgumentParser() 733 | parser.add_argument('--cfg', type=str, default='yolor-csp-c.yaml', help='model.yaml') 734 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 735 | parser.add_argument('--profile', action='store_true', help='profile model speed') 736 | opt = parser.parse_args() 737 | opt.cfg = check_file(opt.cfg) # check file 738 | set_logging() 739 | device = select_device(opt.device) 740 | 741 | # Create model 742 | model = Model(opt.cfg).to(device) 743 | model.train() 744 | 745 | if opt.profile: 746 | img = torch.rand(1, 3, 640, 640).to(device) 747 | y = model(img, profile=True) 748 | 749 | # Profile 750 | # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device) 751 | # y = model(img, profile=True) 752 | 753 | # Tensorboard 754 | # from torch.utils.tensorboard import SummaryWriter 755 | # tb_writer = SummaryWriter() 756 | # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/") 757 | # tb_writer.add_graph(model.model, img) # add model to tensorboard 758 | # tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard 759 | -------------------------------------------------------------------------------- /pose-estimate.py: -------------------------------------------------------------------------------- 1 | import cv2 2 | import time 3 | import torch 4 | import argparse 5 | import numpy as np 6 | import matplotlib.pyplot as plt 7 | from torchvision import transforms 8 | from utils.datasets import letterbox 9 | from utils.torch_utils import select_device 10 | from models.experimental import attempt_load 11 | from utils.general import non_max_suppression_kpt,strip_optimizer,xyxy2xywh 12 | from utils.plots import output_to_keypoint, plot_skeleton_kpts,colors,plot_one_box_kpt 13 | 14 | @torch.no_grad() 15 | def run(poseweights="yolov7-w6-pose.pt",source="football1.mp4",device='cpu',view_img=False, 16 | save_conf=False,line_thickness = 3,hide_labels=False, hide_conf=True): 17 | 18 | frame_count = 0 #count no of frames 19 | total_fps = 0 #count total fps 20 | time_list = [] #list to store time 21 | fps_list = [] #list to store fps 22 | 23 | device = select_device(opt.device) #select device 24 | half = device.type != 'cpu' 25 | 26 | model = attempt_load(poseweights, map_location=device) #Load model 27 | _ = model.eval() 28 | names = model.module.names if hasattr(model, 'module') else model.names # get class names 29 | 30 | if source.isnumeric() : 31 | cap = cv2.VideoCapture(int(source)) #pass video to videocapture object 32 | else : 33 | cap = cv2.VideoCapture(source) #pass video to videocapture object 34 | 35 | if (cap.isOpened() == False): #check if videocapture not opened 36 | print('Error while trying to read video. Please check path again') 37 | raise SystemExit() 38 | 39 | else: 40 | frame_width = int(cap.get(3)) #get video frame width 41 | frame_height = int(cap.get(4)) #get video frame height 42 | 43 | 44 | vid_write_image = letterbox(cap.read()[1], (frame_width), stride=64, auto=True)[0] #init videowriter 45 | resize_height, resize_width = vid_write_image.shape[:2] 46 | out_video_name = f"{source.split('/')[-1].split('.')[0]}" 47 | out = cv2.VideoWriter(f"{source}_keypoint.mp4", 48 | cv2.VideoWriter_fourcc(*'mp4v'), 30, 49 | (resize_width, resize_height)) 50 | 51 | while(cap.isOpened): #loop until cap opened or video not complete 52 | 53 | print("Frame {} Processing".format(frame_count+1)) 54 | 55 | ret, frame = cap.read() #get frame and success from video capture 56 | 57 | if ret: #if success is true, means frame exist 58 | orig_image = frame #store frame 59 | image = cv2.cvtColor(orig_image, cv2.COLOR_BGR2RGB) #convert frame to RGB 60 | image = letterbox(image, (frame_width), stride=64, auto=True)[0] 61 | image_ = image.copy() 62 | image = transforms.ToTensor()(image) 63 | image = torch.tensor(np.array([image.numpy()])) 64 | 65 | image = image.to(device) #convert image data to device 66 | image = image.float() #convert image to float precision (cpu) 67 | start_time = time.time() #start time for fps calculation 68 | 69 | with torch.no_grad(): #get predictions 70 | output_data, _ = model(image) 71 | 72 | output_data = non_max_suppression_kpt(output_data, #Apply non max suppression 73 | 0.25, # Conf. Threshold. 74 | 0.65, # IoU Threshold. 75 | nc=model.yaml['nc'], # Number of classes. 76 | nkpt=model.yaml['nkpt'], # Number of keypoints. 77 | kpt_label=True) 78 | 79 | output = output_to_keypoint(output_data) 80 | 81 | im0 = image[0].permute(1, 2, 0) * 255 # Change format [b, c, h, w] to [h, w, c] for displaying the image. 82 | im0 = im0.cpu().numpy().astype(np.uint8) 83 | 84 | im0 = cv2.cvtColor(im0, cv2.COLOR_RGB2BGR) #reshape image format to (BGR) 85 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 86 | 87 | for i, pose in enumerate(output_data): # detections per image 88 | 89 | if len(output_data): #check if no pose 90 | for c in pose[:, 5].unique(): # Print results 91 | n = (pose[:, 5] == c).sum() # detections per class 92 | print("No of Objects in Current Frame : {}".format(n)) 93 | 94 | for det_index, (*xyxy, conf, cls) in enumerate(reversed(pose[:,:6])): #loop over poses for drawing on frame 95 | c = int(cls) # integer class 96 | kpts = pose[det_index, 6:] 97 | label = None if opt.hide_labels else (names[c] if opt.hide_conf else f'{names[c]} {conf:.2f}') 98 | plot_one_box_kpt(xyxy, im0, label=label, color=colors(c, True), 99 | line_thickness=opt.line_thickness,kpt_label=True, kpts=kpts, steps=3, 100 | orig_shape=im0.shape[:2]) 101 | 102 | 103 | end_time = time.time() #Calculatio for FPS 104 | fps = 1 / (end_time - start_time) 105 | total_fps += fps 106 | frame_count += 1 107 | 108 | fps_list.append(total_fps) #append FPS in list 109 | time_list.append(end_time - start_time) #append time in list 110 | 111 | # Stream results 112 | if view_img: 113 | cv2.imshow("YOLOv7 Pose Estimation Demo", im0) 114 | cv2.waitKey(1) # 1 millisecond 115 | 116 | out.write(im0) #writing the video frame 117 | 118 | else: 119 | break 120 | 121 | cap.release() 122 | # cv2.destroyAllWindows() 123 | avg_fps = total_fps / frame_count 124 | print(f"Average FPS: {avg_fps:.3f}") 125 | 126 | #plot the comparision graph 127 | plot_fps_time_comparision(time_list=time_list,fps_list=fps_list) 128 | 129 | 130 | def parse_opt(): 131 | parser = argparse.ArgumentParser() 132 | parser.add_argument('--poseweights', nargs='+', type=str, default='yolov7-w6-pose.pt', help='model path(s)') 133 | parser.add_argument('--source', type=str, default='football1.mp4', help='video/0 for webcam') #video source 134 | parser.add_argument('--device', type=str, default='cpu', help='cpu/0,1,2,3(gpu)') #device arugments 135 | parser.add_argument('--view-img', action='store_true', help='display results') #display results 136 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') #save confidence in txt writing 137 | parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') #box linethickness 138 | parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') #box hidelabel 139 | parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') #boxhideconf 140 | opt = parser.parse_args() 141 | return opt 142 | 143 | #function for plot fps and time comparision graph 144 | def plot_fps_time_comparision(time_list,fps_list): 145 | plt.figure() 146 | plt.xlabel('Time (s)') 147 | plt.ylabel('FPS') 148 | plt.title('FPS and Time Comparision Graph') 149 | plt.plot(time_list, fps_list,'b',label="FPS & Time") 150 | plt.savefig("FPS_and_Time_Comparision_pose_estimate.png") 151 | 152 | 153 | #main function 154 | def main(opt): 155 | run(**vars(opt)) 156 | 157 | if __name__ == "__main__": 158 | opt = parse_opt() 159 | strip_optimizer(opt.device,opt.poseweights) 160 | main(opt) 161 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # Usage: pip install -r requirements.txt 2 | 3 | # Base ---------------------------------------- 4 | matplotlib>=3.2.2 5 | numpy>=1.18.5 6 | opencv-python>=4.1.1 7 | Pillow>=7.1.2 8 | PyYAML>=5.3.1 9 | requests>=2.23.0 10 | scipy>=1.4.1 11 | torch>=1.7.0,!=1.12.0 12 | torchvision>=0.8.1,!=0.13.0 13 | tqdm>=4.41.0 14 | protobuf<4.21.3 15 | 16 | # Logging ------------------------------------- 17 | tensorboard>=2.4.1 18 | # wandb 19 | 20 | # Plotting ------------------------------------ 21 | pandas>=1.1.4 22 | seaborn>=0.11.0 23 | 24 | 25 | # Extras -------------------------------------- 26 | ipython # interactive notebook 27 | psutil # system utilization 28 | thop # FLOPs computation 29 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- 1 | # init -------------------------------------------------------------------------------- /utils/__pycache__/__init__.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/__init__.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/autoanchor.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/autoanchor.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/datasets.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/datasets.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/general.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/general.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/google_utils.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/google_utils.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/loss.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/loss.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/metrics.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/metrics.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/plots.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/plots.cpython-36.pyc -------------------------------------------------------------------------------- /utils/__pycache__/torch_utils.cpython-36.pyc: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RizwanMunawar/yolov7-pose-estimation/6c81c3ade8d80e401cbad0bdb65f4682949b974f/utils/__pycache__/torch_utils.cpython-36.pyc -------------------------------------------------------------------------------- /utils/activations.py: -------------------------------------------------------------------------------- 1 | # Activation functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | # SiLU https://arxiv.org/pdf/1606.08415.pdf ---------------------------------------------------------------------------- 9 | class SiLU(nn.Module): # export-friendly version of nn.SiLU() 10 | @staticmethod 11 | def forward(x): 12 | return x * torch.sigmoid(x) 13 | 14 | 15 | class Hardswish(nn.Module): # export-friendly version of nn.Hardswish() 16 | @staticmethod 17 | def forward(x): 18 | # return x * F.hardsigmoid(x) # for torchscript and CoreML 19 | return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX 20 | 21 | 22 | class MemoryEfficientSwish(nn.Module): 23 | class F(torch.autograd.Function): 24 | @staticmethod 25 | def forward(ctx, x): 26 | ctx.save_for_backward(x) 27 | return x * torch.sigmoid(x) 28 | 29 | @staticmethod 30 | def backward(ctx, grad_output): 31 | x = ctx.saved_tensors[0] 32 | sx = torch.sigmoid(x) 33 | return grad_output * (sx * (1 + x * (1 - sx))) 34 | 35 | def forward(self, x): 36 | return self.F.apply(x) 37 | 38 | 39 | # Mish https://github.com/digantamisra98/Mish -------------------------------------------------------------------------- 40 | class Mish(nn.Module): 41 | @staticmethod 42 | def forward(x): 43 | return x * F.softplus(x).tanh() 44 | 45 | 46 | class MemoryEfficientMish(nn.Module): 47 | class F(torch.autograd.Function): 48 | @staticmethod 49 | def forward(ctx, x): 50 | ctx.save_for_backward(x) 51 | return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) 52 | 53 | @staticmethod 54 | def backward(ctx, grad_output): 55 | x = ctx.saved_tensors[0] 56 | sx = torch.sigmoid(x) 57 | fx = F.softplus(x).tanh() 58 | return grad_output * (fx + x * sx * (1 - fx * fx)) 59 | 60 | def forward(self, x): 61 | return self.F.apply(x) 62 | 63 | 64 | # FReLU https://arxiv.org/abs/2007.11824 ------------------------------------------------------------------------------- 65 | class FReLU(nn.Module): 66 | def __init__(self, c1, k=3): # ch_in, kernel 67 | super().__init__() 68 | self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) 69 | self.bn = nn.BatchNorm2d(c1) 70 | 71 | def forward(self, x): 72 | return torch.max(x, self.bn(self.conv(x))) 73 | -------------------------------------------------------------------------------- /utils/add_nms.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import onnx 3 | from onnx import shape_inference 4 | try: 5 | import onnx_graphsurgeon as gs 6 | except Exception as e: 7 | print('Import onnx_graphsurgeon failure: %s' % e) 8 | 9 | import logging 10 | 11 | LOGGER = logging.getLogger(__name__) 12 | 13 | class RegisterNMS(object): 14 | def __init__( 15 | self, 16 | onnx_model_path: str, 17 | precision: str = "fp32", 18 | ): 19 | 20 | self.graph = gs.import_onnx(onnx.load(onnx_model_path)) 21 | assert self.graph 22 | LOGGER.info("ONNX graph created successfully") 23 | # Fold constants via ONNX-GS that PyTorch2ONNX may have missed 24 | self.graph.fold_constants() 25 | self.precision = precision 26 | self.batch_size = 1 27 | def infer(self): 28 | """ 29 | Sanitize the graph by cleaning any unconnected nodes, do a topological resort, 30 | and fold constant inputs values. When possible, run shape inference on the 31 | ONNX graph to determine tensor shapes. 32 | """ 33 | for _ in range(3): 34 | count_before = len(self.graph.nodes) 35 | 36 | self.graph.cleanup().toposort() 37 | try: 38 | for node in self.graph.nodes: 39 | for o in node.outputs: 40 | o.shape = None 41 | model = gs.export_onnx(self.graph) 42 | model = shape_inference.infer_shapes(model) 43 | self.graph = gs.import_onnx(model) 44 | except Exception as e: 45 | LOGGER.info(f"Shape inference could not be performed at this time:\n{e}") 46 | try: 47 | self.graph.fold_constants(fold_shapes=True) 48 | except TypeError as e: 49 | LOGGER.error( 50 | "This version of ONNX GraphSurgeon does not support folding shapes, " 51 | f"please upgrade your onnx_graphsurgeon module. Error:\n{e}" 52 | ) 53 | raise 54 | 55 | count_after = len(self.graph.nodes) 56 | if count_before == count_after: 57 | # No new folding occurred in this iteration, so we can stop for now. 58 | break 59 | 60 | def save(self, output_path): 61 | """ 62 | Save the ONNX model to the given location. 63 | Args: 64 | output_path: Path pointing to the location where to write 65 | out the updated ONNX model. 66 | """ 67 | self.graph.cleanup().toposort() 68 | model = gs.export_onnx(self.graph) 69 | onnx.save(model, output_path) 70 | LOGGER.info(f"Saved ONNX model to {output_path}") 71 | 72 | def register_nms( 73 | self, 74 | *, 75 | score_thresh: float = 0.25, 76 | nms_thresh: float = 0.45, 77 | detections_per_img: int = 100, 78 | ): 79 | """ 80 | Register the ``EfficientNMS_TRT`` plugin node. 81 | NMS expects these shapes for its input tensors: 82 | - box_net: [batch_size, number_boxes, 4] 83 | - class_net: [batch_size, number_boxes, number_labels] 84 | Args: 85 | score_thresh (float): The scalar threshold for score (low scoring boxes are removed). 86 | nms_thresh (float): The scalar threshold for IOU (new boxes that have high IOU 87 | overlap with previously selected boxes are removed). 88 | detections_per_img (int): Number of best detections to keep after NMS. 89 | """ 90 | 91 | self.infer() 92 | # Find the concat node at the end of the network 93 | op_inputs = self.graph.outputs 94 | op = "EfficientNMS_TRT" 95 | attrs = { 96 | "plugin_version": "1", 97 | "background_class": -1, # no background class 98 | "max_output_boxes": detections_per_img, 99 | "score_threshold": score_thresh, 100 | "iou_threshold": nms_thresh, 101 | "score_activation": False, 102 | "box_coding": 0, 103 | } 104 | 105 | if self.precision == "fp32": 106 | dtype_output = np.float32 107 | elif self.precision == "fp16": 108 | dtype_output = np.float16 109 | else: 110 | raise NotImplementedError(f"Currently not supports precision: {self.precision}") 111 | 112 | # NMS Outputs 113 | output_num_detections = gs.Variable( 114 | name="num_detections", 115 | dtype=np.int32, 116 | shape=[self.batch_size, 1], 117 | ) # A scalar indicating the number of valid detections per batch image. 118 | output_boxes = gs.Variable( 119 | name="detection_boxes", 120 | dtype=dtype_output, 121 | shape=[self.batch_size, detections_per_img, 4], 122 | ) 123 | output_scores = gs.Variable( 124 | name="detection_scores", 125 | dtype=dtype_output, 126 | shape=[self.batch_size, detections_per_img], 127 | ) 128 | output_labels = gs.Variable( 129 | name="detection_classes", 130 | dtype=np.int32, 131 | shape=[self.batch_size, detections_per_img], 132 | ) 133 | 134 | op_outputs = [output_num_detections, output_boxes, output_scores, output_labels] 135 | 136 | # Create the NMS Plugin node with the selected inputs. The outputs of the node will also 137 | # become the final outputs of the graph. 138 | self.graph.layer(op=op, name="batched_nms", inputs=op_inputs, outputs=op_outputs, attrs=attrs) 139 | LOGGER.info(f"Created NMS plugin '{op}' with attributes: {attrs}") 140 | 141 | self.graph.outputs = op_outputs 142 | 143 | self.infer() 144 | 145 | def save(self, output_path): 146 | """ 147 | Save the ONNX model to the given location. 148 | Args: 149 | output_path: Path pointing to the location where to write 150 | out the updated ONNX model. 151 | """ 152 | self.graph.cleanup().toposort() 153 | model = gs.export_onnx(self.graph) 154 | onnx.save(model, output_path) 155 | LOGGER.info(f"Saved ONNX model to {output_path}") 156 | -------------------------------------------------------------------------------- /utils/autoanchor.py: -------------------------------------------------------------------------------- 1 | # Auto-anchor utils 2 | 3 | import numpy as np 4 | import torch 5 | import yaml 6 | from scipy.cluster.vq import kmeans 7 | from tqdm import tqdm 8 | 9 | from utils.general import colorstr 10 | 11 | 12 | def check_anchor_order(m): 13 | # Check anchor order against stride order for YOLO Detect() module m, and correct if necessary 14 | a = m.anchor_grid.prod(-1).view(-1) # anchor area 15 | da = a[-1] - a[0] # delta a 16 | ds = m.stride[-1] - m.stride[0] # delta s 17 | if da.sign() != ds.sign(): # same order 18 | print('Reversing anchor order') 19 | m.anchors[:] = m.anchors.flip(0) 20 | m.anchor_grid[:] = m.anchor_grid.flip(0) 21 | 22 | 23 | def check_anchors(dataset, model, thr=4.0, imgsz=640): 24 | # Check anchor fit to data, recompute if necessary 25 | prefix = colorstr('autoanchor: ') 26 | print(f'\n{prefix}Analyzing anchors... ', end='') 27 | m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() 28 | shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) 29 | scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale 30 | wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh 31 | 32 | def metric(k): # compute metric 33 | r = wh[:, None] / k[None] 34 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 35 | best = x.max(1)[0] # best_x 36 | aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold 37 | bpr = (best > 1. / thr).float().mean() # best possible recall 38 | return bpr, aat 39 | 40 | anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors 41 | bpr, aat = metric(anchors) 42 | print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='') 43 | if bpr < 0.98: # threshold to recompute 44 | print('. Attempting to improve anchors, please wait...') 45 | na = m.anchor_grid.numel() // 2 # number of anchors 46 | try: 47 | anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) 48 | except Exception as e: 49 | print(f'{prefix}ERROR: {e}') 50 | new_bpr = metric(anchors)[0] 51 | if new_bpr > bpr: # replace anchors 52 | anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) 53 | m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference 54 | m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss 55 | check_anchor_order(m) 56 | print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.') 57 | else: 58 | print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.') 59 | print('') # newline 60 | 61 | 62 | def kmean_anchors(path='./data/coco.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): 63 | """ Creates kmeans-evolved anchors from training dataset 64 | 65 | Arguments: 66 | path: path to dataset *.yaml, or a loaded dataset 67 | n: number of anchors 68 | img_size: image size used for training 69 | thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 70 | gen: generations to evolve anchors using genetic algorithm 71 | verbose: print all results 72 | 73 | Return: 74 | k: kmeans evolved anchors 75 | 76 | Usage: 77 | from utils.autoanchor import *; _ = kmean_anchors() 78 | """ 79 | thr = 1. / thr 80 | prefix = colorstr('autoanchor: ') 81 | 82 | def metric(k, wh): # compute metrics 83 | r = wh[:, None] / k[None] 84 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 85 | # x = wh_iou(wh, torch.tensor(k)) # iou metric 86 | return x, x.max(1)[0] # x, best_x 87 | 88 | def anchor_fitness(k): # mutation fitness 89 | _, best = metric(torch.tensor(k, dtype=torch.float32), wh) 90 | return (best * (best > thr).float()).mean() # fitness 91 | 92 | def print_results(k): 93 | k = k[np.argsort(k.prod(1))] # sort small to large 94 | x, best = metric(k, wh0) 95 | bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr 96 | print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr') 97 | print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' 98 | f'past_thr={x[x > thr].mean():.3f}-mean: ', end='') 99 | for i, x in enumerate(k): 100 | print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg 101 | return k 102 | 103 | if isinstance(path, str): # *.yaml file 104 | with open(path) as f: 105 | data_dict = yaml.load(f, Loader=yaml.SafeLoader) # model dict 106 | from utils.datasets import LoadImagesAndLabels 107 | dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) 108 | else: 109 | dataset = path # dataset 110 | 111 | # Get label wh 112 | shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) 113 | wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh 114 | 115 | # Filter 116 | i = (wh0 < 3.0).any(1).sum() 117 | if i: 118 | print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.') 119 | wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels 120 | # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 121 | 122 | # Kmeans calculation 123 | print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...') 124 | s = wh.std(0) # sigmas for whitening 125 | k, dist = kmeans(wh / s, n, iter=30) # points, mean distance 126 | assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}') 127 | k *= s 128 | wh = torch.tensor(wh, dtype=torch.float32) # filtered 129 | wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered 130 | k = print_results(k) 131 | 132 | # Plot 133 | # k, d = [None] * 20, [None] * 20 134 | # for i in tqdm(range(1, 21)): 135 | # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance 136 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) 137 | # ax = ax.ravel() 138 | # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') 139 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh 140 | # ax[0].hist(wh[wh[:, 0]<100, 0],400) 141 | # ax[1].hist(wh[wh[:, 1]<100, 1],400) 142 | # fig.savefig('wh.png', dpi=200) 143 | 144 | # Evolve 145 | npr = np.random 146 | f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma 147 | pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar 148 | for _ in pbar: 149 | v = np.ones(sh) 150 | while (v == 1).all(): # mutate until a change occurs (prevent duplicates) 151 | v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) 152 | kg = (k.copy() * v).clip(min=2.0) 153 | fg = anchor_fitness(kg) 154 | if fg > f: 155 | f, k = fg, kg.copy() 156 | pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' 157 | if verbose: 158 | print_results(k) 159 | 160 | return print_results(k) 161 | -------------------------------------------------------------------------------- /utils/google_utils.py: -------------------------------------------------------------------------------- 1 | # Google utils: https://cloud.google.com/storage/docs/reference/libraries 2 | 3 | import os 4 | import platform 5 | import subprocess 6 | import time 7 | from pathlib import Path 8 | 9 | import requests 10 | import torch 11 | 12 | 13 | def gsutil_getsize(url=''): 14 | # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du 15 | s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') 16 | return eval(s.split(' ')[0]) if len(s) else 0 # bytes 17 | 18 | 19 | def attempt_download(file, repo='WongKinYiu/yolov7'): 20 | # Attempt file download if does not exist 21 | file = Path(str(file).strip().replace("'", '').lower()) 22 | 23 | if not file.exists(): 24 | try: 25 | response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json() # github api 26 | assets = [x['name'] for x in response['assets']] # release assets 27 | tag = response['tag_name'] # i.e. 'v1.0' 28 | except: # fallback plan 29 | assets = ['yolov7.pt'] 30 | tag = subprocess.check_output('git tag', shell=True).decode().split()[-1] 31 | 32 | name = file.name 33 | if name in assets: 34 | msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/' 35 | redundant = False # second download option 36 | try: # GitHub 37 | url = f'https://github.com/{repo}/releases/download/{tag}/{name}' 38 | print(f'Downloading {url} to {file}...') 39 | torch.hub.download_url_to_file(url, file) 40 | assert file.exists() and file.stat().st_size > 1E6 # check 41 | except Exception as e: # GCP 42 | print(f'Download error: {e}') 43 | assert redundant, 'No secondary mirror' 44 | url = f'https://storage.googleapis.com/{repo}/ckpt/{name}' 45 | print(f'Downloading {url} to {file}...') 46 | os.system(f'curl -L {url} -o {file}') # torch.hub.download_url_to_file(url, weights) 47 | finally: 48 | if not file.exists() or file.stat().st_size < 1E6: # check 49 | file.unlink(missing_ok=True) # remove partial downloads 50 | print(f'ERROR: Download failure: {msg}') 51 | print('') 52 | return 53 | 54 | 55 | def gdrive_download(id='', file='tmp.zip'): 56 | # Downloads a file from Google Drive. from yolov7.utils.google_utils import *; gdrive_download() 57 | t = time.time() 58 | file = Path(file) 59 | cookie = Path('cookie') # gdrive cookie 60 | print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') 61 | file.unlink(missing_ok=True) # remove existing file 62 | cookie.unlink(missing_ok=True) # remove existing cookie 63 | 64 | # Attempt file download 65 | out = "NUL" if platform.system() == "Windows" else "/dev/null" 66 | os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') 67 | if os.path.exists('cookie'): # large file 68 | s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' 69 | else: # small file 70 | s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' 71 | r = os.system(s) # execute, capture return 72 | cookie.unlink(missing_ok=True) # remove existing cookie 73 | 74 | # Error check 75 | if r != 0: 76 | file.unlink(missing_ok=True) # remove partial 77 | print('Download error ') # raise Exception('Download error') 78 | return r 79 | 80 | # Unzip if archive 81 | if file.suffix == '.zip': 82 | print('unzipping... ', end='') 83 | os.system(f'unzip -q {file}') # unzip 84 | file.unlink() # remove zip to free space 85 | 86 | print(f'Done ({time.time() - t:.1f}s)') 87 | return r 88 | 89 | 90 | def get_token(cookie="./cookie"): 91 | with open(cookie) as f: 92 | for line in f: 93 | if "download" in line: 94 | return line.split()[-1] 95 | return "" 96 | 97 | # def upload_blob(bucket_name, source_file_name, destination_blob_name): 98 | # # Uploads a file to a bucket 99 | # # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python 100 | # 101 | # storage_client = storage.Client() 102 | # bucket = storage_client.get_bucket(bucket_name) 103 | # blob = bucket.blob(destination_blob_name) 104 | # 105 | # blob.upload_from_filename(source_file_name) 106 | # 107 | # print('File {} uploaded to {}.'.format( 108 | # source_file_name, 109 | # destination_blob_name)) 110 | # 111 | # 112 | # def download_blob(bucket_name, source_blob_name, destination_file_name): 113 | # # Uploads a blob from a bucket 114 | # storage_client = storage.Client() 115 | # bucket = storage_client.get_bucket(bucket_name) 116 | # blob = bucket.blob(source_blob_name) 117 | # 118 | # blob.download_to_filename(destination_file_name) 119 | # 120 | # print('Blob {} downloaded to {}.'.format( 121 | # source_blob_name, 122 | # destination_file_name)) 123 | -------------------------------------------------------------------------------- /utils/metrics.py: -------------------------------------------------------------------------------- 1 | # Model validation metrics 2 | 3 | from pathlib import Path 4 | 5 | import matplotlib.pyplot as plt 6 | import numpy as np 7 | import torch 8 | 9 | from . import general 10 | 11 | 12 | def fitness(x): 13 | # Model fitness as a weighted combination of metrics 14 | w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] 15 | return (x[:, :4] * w).sum(1) 16 | 17 | 18 | def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=()): 19 | """ Compute the average precision, given the recall and precision curves. 20 | Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. 21 | # Arguments 22 | tp: True positives (nparray, nx1 or nx10). 23 | conf: Objectness value from 0-1 (nparray). 24 | pred_cls: Predicted object classes (nparray). 25 | target_cls: True object classes (nparray). 26 | plot: Plot precision-recall curve at mAP@0.5 27 | save_dir: Plot save directory 28 | # Returns 29 | The average precision as computed in py-faster-rcnn. 30 | """ 31 | 32 | # Sort by objectness 33 | i = np.argsort(-conf) 34 | tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] 35 | 36 | # Find unique classes 37 | unique_classes = np.unique(target_cls) 38 | nc = unique_classes.shape[0] # number of classes, number of detections 39 | 40 | # Create Precision-Recall curve and compute AP for each class 41 | px, py = np.linspace(0, 1, 1000), [] # for plotting 42 | ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000)) 43 | for ci, c in enumerate(unique_classes): 44 | i = pred_cls == c 45 | n_l = (target_cls == c).sum() # number of labels 46 | n_p = i.sum() # number of predictions 47 | 48 | if n_p == 0 or n_l == 0: 49 | continue 50 | else: 51 | # Accumulate FPs and TPs 52 | fpc = (1 - tp[i]).cumsum(0) 53 | tpc = tp[i].cumsum(0) 54 | 55 | # Recall 56 | recall = tpc / (n_l + 1e-16) # recall curve 57 | r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases 58 | 59 | # Precision 60 | precision = tpc / (tpc + fpc) # precision curve 61 | p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score 62 | 63 | # AP from recall-precision curve 64 | for j in range(tp.shape[1]): 65 | ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) 66 | if plot and j == 0: 67 | py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 68 | 69 | # Compute F1 (harmonic mean of precision and recall) 70 | f1 = 2 * p * r / (p + r + 1e-16) 71 | if plot: 72 | plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names) 73 | plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1') 74 | plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision') 75 | plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall') 76 | 77 | i = f1.mean(0).argmax() # max F1 index 78 | return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32') 79 | 80 | 81 | def compute_ap(recall, precision): 82 | """ Compute the average precision, given the recall and precision curves 83 | # Arguments 84 | recall: The recall curve (list) 85 | precision: The precision curve (list) 86 | # Returns 87 | Average precision, precision curve, recall curve 88 | """ 89 | 90 | # Append sentinel values to beginning and end 91 | mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01])) 92 | mpre = np.concatenate(([1.], precision, [0.])) 93 | 94 | # Compute the precision envelope 95 | mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) 96 | 97 | # Integrate area under curve 98 | method = 'interp' # methods: 'continuous', 'interp' 99 | if method == 'interp': 100 | x = np.linspace(0, 1, 101) # 101-point interp (COCO) 101 | ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate 102 | else: # 'continuous' 103 | i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes 104 | ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve 105 | 106 | return ap, mpre, mrec 107 | 108 | 109 | class ConfusionMatrix: 110 | # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix 111 | def __init__(self, nc, conf=0.25, iou_thres=0.45): 112 | self.matrix = np.zeros((nc + 1, nc + 1)) 113 | self.nc = nc # number of classes 114 | self.conf = conf 115 | self.iou_thres = iou_thres 116 | 117 | def process_batch(self, detections, labels): 118 | """ 119 | Return intersection-over-union (Jaccard index) of boxes. 120 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 121 | Arguments: 122 | detections (Array[N, 6]), x1, y1, x2, y2, conf, class 123 | labels (Array[M, 5]), class, x1, y1, x2, y2 124 | Returns: 125 | None, updates confusion matrix accordingly 126 | """ 127 | detections = detections[detections[:, 4] > self.conf] 128 | gt_classes = labels[:, 0].int() 129 | detection_classes = detections[:, 5].int() 130 | iou = general.box_iou(labels[:, 1:], detections[:, :4]) 131 | 132 | x = torch.where(iou > self.iou_thres) 133 | if x[0].shape[0]: 134 | matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() 135 | if x[0].shape[0] > 1: 136 | matches = matches[matches[:, 2].argsort()[::-1]] 137 | matches = matches[np.unique(matches[:, 1], return_index=True)[1]] 138 | matches = matches[matches[:, 2].argsort()[::-1]] 139 | matches = matches[np.unique(matches[:, 0], return_index=True)[1]] 140 | else: 141 | matches = np.zeros((0, 3)) 142 | 143 | n = matches.shape[0] > 0 144 | m0, m1, _ = matches.transpose().astype(np.int16) 145 | for i, gc in enumerate(gt_classes): 146 | j = m0 == i 147 | if n and sum(j) == 1: 148 | self.matrix[gc, detection_classes[m1[j]]] += 1 # correct 149 | else: 150 | self.matrix[self.nc, gc] += 1 # background FP 151 | 152 | if n: 153 | for i, dc in enumerate(detection_classes): 154 | if not any(m1 == i): 155 | self.matrix[dc, self.nc] += 1 # background FN 156 | 157 | def matrix(self): 158 | return self.matrix 159 | 160 | def plot(self, save_dir='', names=()): 161 | try: 162 | import seaborn as sn 163 | 164 | array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize 165 | array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) 166 | 167 | fig = plt.figure(figsize=(12, 9), tight_layout=True) 168 | sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size 169 | labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels 170 | sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True, 171 | xticklabels=names + ['background FP'] if labels else "auto", 172 | yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) 173 | fig.axes[0].set_xlabel('True') 174 | fig.axes[0].set_ylabel('Predicted') 175 | fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) 176 | except Exception as e: 177 | pass 178 | 179 | def print(self): 180 | for i in range(self.nc + 1): 181 | print(' '.join(map(str, self.matrix[i]))) 182 | 183 | 184 | # Plots ---------------------------------------------------------------------------------------------------------------- 185 | 186 | def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()): 187 | # Precision-recall curve 188 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 189 | py = np.stack(py, axis=1) 190 | 191 | if 0 < len(names) < 21: # display per-class legend if < 21 classes 192 | for i, y in enumerate(py.T): 193 | ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision) 194 | else: 195 | ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) 196 | 197 | ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) 198 | ax.set_xlabel('Recall') 199 | ax.set_ylabel('Precision') 200 | ax.set_xlim(0, 1) 201 | ax.set_ylim(0, 1) 202 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 203 | fig.savefig(Path(save_dir), dpi=250) 204 | 205 | 206 | def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'): 207 | # Metric-confidence curve 208 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 209 | 210 | if 0 < len(names) < 21: # display per-class legend if < 21 classes 211 | for i, y in enumerate(py): 212 | ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric) 213 | else: 214 | ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric) 215 | 216 | y = py.mean(0) 217 | ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}') 218 | ax.set_xlabel(xlabel) 219 | ax.set_ylabel(ylabel) 220 | ax.set_xlim(0, 1) 221 | ax.set_ylim(0, 1) 222 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 223 | fig.savefig(Path(save_dir), dpi=250) 224 | -------------------------------------------------------------------------------- /utils/plots.py: -------------------------------------------------------------------------------- 1 | # Plotting utils 2 | 3 | import glob 4 | import math 5 | import os 6 | import random 7 | from copy import copy 8 | from pathlib import Path 9 | 10 | import cv2 11 | import matplotlib 12 | import matplotlib.pyplot as plt 13 | import numpy as np 14 | import pandas as pd 15 | import seaborn as sns 16 | import torch 17 | import yaml 18 | from PIL import Image, ImageDraw, ImageFont 19 | from scipy.signal import butter, filtfilt 20 | 21 | from utils.general import xywh2xyxy, xyxy2xywh 22 | from utils.metrics import fitness 23 | 24 | # Settings 25 | matplotlib.rc('font', **{'size': 11}) 26 | matplotlib.use('Agg') # for writing to files only 27 | 28 | class Colors: 29 | # Ultralytics color palette https://ultralytics.com/ 30 | def __init__(self): 31 | self.palette = [self.hex2rgb(c) for c in matplotlib.colors.TABLEAU_COLORS.values()] 32 | self.n = len(self.palette) 33 | 34 | def __call__(self, i, bgr=False): 35 | c = self.palette[int(i) % self.n] 36 | return (c[2], c[1], c[0]) if bgr else c 37 | 38 | @staticmethod 39 | def hex2rgb(h): # rgb order (PIL) 40 | return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) 41 | 42 | 43 | def plot_one_box_kpt(x, im, color=None, label=None, line_thickness=3, kpt_label=False, kpts=None, steps=2, orig_shape=None): 44 | # Plots one bounding box on image 'im' using OpenCV 45 | assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to plot_on_box() input image.' 46 | tl = line_thickness or round(0.002 * (im.shape[0] + im.shape[1]) / 2) + 1 # line/font thickness 47 | color = color or [random.randint(0, 255) for _ in range(3)] 48 | c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) 49 | cv2.rectangle(im, c1, c2, (255,0,0), thickness=tl*1//3, lineType=cv2.LINE_AA) 50 | if label: 51 | if len(label.split(' ')) > 1: 52 | label = label.split(' ')[-1] 53 | tf = max(tl - 1, 1) # font thickness 54 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 6, thickness=tf)[0] 55 | c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 56 | cv2.rectangle(im, c1, c2, color, -1, cv2.LINE_AA) # filled 57 | cv2.putText(im, label, (c1[0], c1[1] - 2), 0, tl / 6, [225, 255, 255], thickness=tf//2, lineType=cv2.LINE_AA) 58 | if kpt_label: 59 | plot_skeleton_kpts(im, kpts, steps, orig_shape=orig_shape) 60 | 61 | colors = Colors() 62 | 63 | def color_list(): 64 | def hex2rgb(h): 65 | return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) 66 | 67 | return [hex2rgb(h) for h in matplotlib.colors.TABLEAU_COLORS.values()] # or BASE_ (8), CSS4_ (148), XKCD_ (949) 68 | 69 | 70 | def hist2d(x, y, n=100): 71 | # 2d histogram used in labels.png and evolve.png 72 | xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) 73 | hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) 74 | xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) 75 | yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) 76 | return np.log(hist[xidx, yidx]) 77 | 78 | 79 | def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): 80 | # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy 81 | def butter_lowpass(cutoff, fs, order): 82 | nyq = 0.5 * fs 83 | normal_cutoff = cutoff / nyq 84 | return butter(order, normal_cutoff, btype='low', analog=False) 85 | 86 | b, a = butter_lowpass(cutoff, fs, order=order) 87 | return filtfilt(b, a, data) # forward-backward filter 88 | 89 | 90 | def plot_one_box(x, img, color=None, label=None, line_thickness=1): 91 | # Plots one bounding box on image img 92 | tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 2 # line/font thickness 93 | color = color or [random.randint(0, 255) for _ in range(3)] 94 | c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) 95 | cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA) 96 | if label: 97 | tf = max(tl - 1, 1) # font thickness 98 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 99 | c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 100 | cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled 101 | cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) 102 | 103 | 104 | def plot_one_box_PIL(box, img, color=None, label=None, line_thickness=None): 105 | img = Image.fromarray(img) 106 | draw = ImageDraw.Draw(img) 107 | line_thickness = line_thickness or max(int(min(img.size) / 200), 2) 108 | draw.rectangle(box, width=line_thickness, outline=tuple(color)) # plot 109 | if label: 110 | fontsize = max(round(max(img.size) / 40), 12) 111 | font = ImageFont.truetype("Arial.ttf", fontsize) 112 | txt_width, txt_height = font.getsize(label) 113 | draw.rectangle([box[0], box[1] - txt_height + 4, box[0] + txt_width, box[1]], fill=tuple(color)) 114 | draw.text((box[0], box[1] - txt_height + 1), label, fill=(255, 255, 255), font=font) 115 | return np.asarray(img) 116 | 117 | 118 | def plot_wh_methods(): # from utils.plots import *; plot_wh_methods() 119 | # Compares the two methods for width-height anchor multiplication 120 | # https://github.com/ultralytics/yolov3/issues/168 121 | x = np.arange(-4.0, 4.0, .1) 122 | ya = np.exp(x) 123 | yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2 124 | 125 | fig = plt.figure(figsize=(6, 3), tight_layout=True) 126 | plt.plot(x, ya, '.-', label='YOLOv3') 127 | plt.plot(x, yb ** 2, '.-', label='YOLOR ^2') 128 | plt.plot(x, yb ** 1.6, '.-', label='YOLOR ^1.6') 129 | plt.xlim(left=-4, right=4) 130 | plt.ylim(bottom=0, top=6) 131 | plt.xlabel('input') 132 | plt.ylabel('output') 133 | plt.grid() 134 | plt.legend() 135 | fig.savefig('comparison.png', dpi=200) 136 | 137 | 138 | def output_to_target(output): 139 | # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] 140 | targets = [] 141 | for i, o in enumerate(output): 142 | for *box, conf, cls in o.cpu().numpy(): 143 | targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) 144 | return np.array(targets) 145 | 146 | 147 | def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): 148 | # Plot image grid with labels 149 | 150 | if isinstance(images, torch.Tensor): 151 | images = images.cpu().float().numpy() 152 | if isinstance(targets, torch.Tensor): 153 | targets = targets.cpu().numpy() 154 | 155 | # un-normalise 156 | if np.max(images[0]) <= 1: 157 | images *= 255 158 | 159 | tl = 3 # line thickness 160 | tf = max(tl - 1, 1) # font thickness 161 | bs, _, h, w = images.shape # batch size, _, height, width 162 | bs = min(bs, max_subplots) # limit plot images 163 | ns = np.ceil(bs ** 0.5) # number of subplots (square) 164 | 165 | # Check if we should resize 166 | scale_factor = max_size / max(h, w) 167 | if scale_factor < 1: 168 | h = math.ceil(scale_factor * h) 169 | w = math.ceil(scale_factor * w) 170 | 171 | colors = color_list() # list of colors 172 | mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init 173 | for i, img in enumerate(images): 174 | if i == max_subplots: # if last batch has fewer images than we expect 175 | break 176 | 177 | block_x = int(w * (i // ns)) 178 | block_y = int(h * (i % ns)) 179 | 180 | img = img.transpose(1, 2, 0) 181 | if scale_factor < 1: 182 | img = cv2.resize(img, (w, h)) 183 | 184 | mosaic[block_y:block_y + h, block_x:block_x + w, :] = img 185 | if len(targets) > 0: 186 | image_targets = targets[targets[:, 0] == i] 187 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 188 | classes = image_targets[:, 1].astype('int') 189 | labels = image_targets.shape[1] == 6 # labels if no conf column 190 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 191 | 192 | if boxes.shape[1]: 193 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 194 | boxes[[0, 2]] *= w # scale to pixels 195 | boxes[[1, 3]] *= h 196 | elif scale_factor < 1: # absolute coords need scale if image scales 197 | boxes *= scale_factor 198 | boxes[[0, 2]] += block_x 199 | boxes[[1, 3]] += block_y 200 | for j, box in enumerate(boxes.T): 201 | cls = int(classes[j]) 202 | color = colors[cls % len(colors)] 203 | cls = names[cls] if names else cls 204 | if labels or conf[j] > 0.25: # 0.25 conf thresh 205 | label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j]) 206 | plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl) 207 | 208 | # Draw image filename labels 209 | if paths: 210 | label = Path(paths[i]).name[:40] # trim to 40 char 211 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 212 | cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf, 213 | lineType=cv2.LINE_AA) 214 | 215 | # Image border 216 | cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3) 217 | 218 | if fname: 219 | r = min(1280. / max(h, w) / ns, 1.0) # ratio to limit image size 220 | mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA) 221 | # cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB)) # cv2 save 222 | Image.fromarray(mosaic).save(fname) # PIL save 223 | return mosaic 224 | 225 | 226 | def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): 227 | # Plot LR simulating training for full epochs 228 | optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals 229 | y = [] 230 | for _ in range(epochs): 231 | scheduler.step() 232 | y.append(optimizer.param_groups[0]['lr']) 233 | plt.plot(y, '.-', label='LR') 234 | plt.xlabel('epoch') 235 | plt.ylabel('LR') 236 | plt.grid() 237 | plt.xlim(0, epochs) 238 | plt.ylim(0) 239 | plt.savefig(Path(save_dir) / 'LR.png', dpi=200) 240 | plt.close() 241 | 242 | 243 | def plot_test_txt(): # from utils.plots import *; plot_test() 244 | # Plot test.txt histograms 245 | x = np.loadtxt('test.txt', dtype=np.float32) 246 | box = xyxy2xywh(x[:, :4]) 247 | cx, cy = box[:, 0], box[:, 1] 248 | 249 | fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) 250 | ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) 251 | ax.set_aspect('equal') 252 | plt.savefig('hist2d.png', dpi=300) 253 | 254 | fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) 255 | ax[0].hist(cx, bins=600) 256 | ax[1].hist(cy, bins=600) 257 | plt.savefig('hist1d.png', dpi=200) 258 | 259 | 260 | def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() 261 | # Plot targets.txt histograms 262 | x = np.loadtxt('targets.txt', dtype=np.float32).T 263 | s = ['x targets', 'y targets', 'width targets', 'height targets'] 264 | fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) 265 | ax = ax.ravel() 266 | for i in range(4): 267 | ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std())) 268 | ax[i].legend() 269 | ax[i].set_title(s[i]) 270 | plt.savefig('targets.jpg', dpi=200) 271 | 272 | 273 | def plot_study_txt(path='', x=None): # from utils.plots import *; plot_study_txt() 274 | # Plot study.txt generated by test.py 275 | fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True) 276 | # ax = ax.ravel() 277 | 278 | fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) 279 | # for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolor-p6', 'yolor-w6', 'yolor-e6', 'yolor-d6']]: 280 | for f in sorted(Path(path).glob('study*.txt')): 281 | y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T 282 | x = np.arange(y.shape[1]) if x is None else np.array(x) 283 | s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)'] 284 | # for i in range(7): 285 | # ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) 286 | # ax[i].set_title(s[i]) 287 | 288 | j = y[3].argmax() + 1 289 | ax2.plot(y[6, 1:j], y[3, 1:j] * 1E2, '.-', linewidth=2, markersize=8, 290 | label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) 291 | 292 | ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], 293 | 'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet') 294 | 295 | ax2.grid(alpha=0.2) 296 | ax2.set_yticks(np.arange(20, 60, 5)) 297 | ax2.set_xlim(0, 57) 298 | ax2.set_ylim(30, 55) 299 | ax2.set_xlabel('GPU Speed (ms/img)') 300 | ax2.set_ylabel('COCO AP val') 301 | ax2.legend(loc='lower right') 302 | plt.savefig(str(Path(path).name) + '.png', dpi=300) 303 | 304 | 305 | def plot_labels(labels, names=(), save_dir=Path(''), loggers=None): 306 | # plot dataset labels 307 | print('Plotting labels... ') 308 | c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes 309 | nc = int(c.max() + 1) # number of classes 310 | colors = color_list() 311 | x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) 312 | 313 | # seaborn correlogram 314 | sns.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9)) 315 | plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) 316 | plt.close() 317 | 318 | # matplotlib labels 319 | matplotlib.use('svg') # faster 320 | ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() 321 | ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) 322 | ax[0].set_ylabel('instances') 323 | if 0 < len(names) < 30: 324 | ax[0].set_xticks(range(len(names))) 325 | ax[0].set_xticklabels(names, rotation=90, fontsize=10) 326 | else: 327 | ax[0].set_xlabel('classes') 328 | sns.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9) 329 | sns.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9) 330 | 331 | # rectangles 332 | labels[:, 1:3] = 0.5 # center 333 | labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 334 | img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) 335 | for cls, *box in labels[:1000]: 336 | ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10]) # plot 337 | ax[1].imshow(img) 338 | ax[1].axis('off') 339 | 340 | for a in [0, 1, 2, 3]: 341 | for s in ['top', 'right', 'left', 'bottom']: 342 | ax[a].spines[s].set_visible(False) 343 | 344 | plt.savefig(save_dir / 'labels.jpg', dpi=200) 345 | matplotlib.use('Agg') 346 | plt.close() 347 | 348 | # loggers 349 | for k, v in loggers.items() or {}: 350 | if k == 'wandb' and v: 351 | v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}, commit=False) 352 | 353 | 354 | def plot_evolution(yaml_file='data/hyp.finetune.yaml'): # from utils.plots import *; plot_evolution() 355 | # Plot hyperparameter evolution results in evolve.txt 356 | with open(yaml_file) as f: 357 | hyp = yaml.load(f, Loader=yaml.SafeLoader) 358 | x = np.loadtxt('evolve.txt', ndmin=2) 359 | f = fitness(x) 360 | # weights = (f - f.min()) ** 2 # for weighted results 361 | plt.figure(figsize=(10, 12), tight_layout=True) 362 | matplotlib.rc('font', **{'size': 8}) 363 | for i, (k, v) in enumerate(hyp.items()): 364 | y = x[:, i + 7] 365 | # mu = (y * weights).sum() / weights.sum() # best weighted result 366 | mu = y[f.argmax()] # best single result 367 | plt.subplot(6, 5, i + 1) 368 | plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none') 369 | plt.plot(mu, f.max(), 'k+', markersize=15) 370 | plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters 371 | if i % 5 != 0: 372 | plt.yticks([]) 373 | print('%15s: %.3g' % (k, mu)) 374 | plt.savefig('evolve.png', dpi=200) 375 | print('\nPlot saved as evolve.png') 376 | 377 | 378 | def profile_idetection(start=0, stop=0, labels=(), save_dir=''): 379 | # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection() 380 | ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel() 381 | s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS'] 382 | files = list(Path(save_dir).glob('frames*.txt')) 383 | for fi, f in enumerate(files): 384 | try: 385 | results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows 386 | n = results.shape[1] # number of rows 387 | x = np.arange(start, min(stop, n) if stop else n) 388 | results = results[:, x] 389 | t = (results[0] - results[0].min()) # set t0=0s 390 | results[0] = x 391 | for i, a in enumerate(ax): 392 | if i < len(results): 393 | label = labels[fi] if len(labels) else f.stem.replace('frames_', '') 394 | a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5) 395 | a.set_title(s[i]) 396 | a.set_xlabel('time (s)') 397 | # if fi == len(files) - 1: 398 | # a.set_ylim(bottom=0) 399 | for side in ['top', 'right']: 400 | a.spines[side].set_visible(False) 401 | else: 402 | a.remove() 403 | except Exception as e: 404 | print('Warning: Plotting error for %s; %s' % (f, e)) 405 | 406 | ax[1].legend() 407 | plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200) 408 | 409 | 410 | def plot_results_overlay(start=0, stop=0): # from utils.plots import *; plot_results_overlay() 411 | # Plot training 'results*.txt', overlaying train and val losses 412 | s = ['train', 'train', 'train', 'Precision', 'mAP@0.5', 'val', 'val', 'val', 'Recall', 'mAP@0.5:0.95'] # legends 413 | t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles 414 | for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): 415 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 416 | n = results.shape[1] # number of rows 417 | x = range(start, min(stop, n) if stop else n) 418 | fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True) 419 | ax = ax.ravel() 420 | for i in range(5): 421 | for j in [i, i + 5]: 422 | y = results[j, x] 423 | ax[i].plot(x, y, marker='.', label=s[j]) 424 | # y_smooth = butter_lowpass_filtfilt(y) 425 | # ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j]) 426 | 427 | ax[i].set_title(t[i]) 428 | ax[i].legend() 429 | ax[i].set_ylabel(f) if i == 0 else None # add filename 430 | fig.savefig(f.replace('.txt', '.png'), dpi=200) 431 | 432 | 433 | def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''): 434 | # Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp') 435 | fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) 436 | ax = ax.ravel() 437 | s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall', 438 | 'val Box', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95'] 439 | if bucket: 440 | # files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id] 441 | files = ['results%g.txt' % x for x in id] 442 | c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id) 443 | os.system(c) 444 | else: 445 | files = list(Path(save_dir).glob('results*.txt')) 446 | assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir) 447 | for fi, f in enumerate(files): 448 | try: 449 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 450 | n = results.shape[1] # number of rows 451 | x = range(start, min(stop, n) if stop else n) 452 | for i in range(10): 453 | y = results[i, x] 454 | if i in [0, 1, 2, 5, 6, 7]: 455 | y[y == 0] = np.nan # don't show zero loss values 456 | # y /= y[0] # normalize 457 | label = labels[fi] if len(labels) else f.stem 458 | ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8) 459 | ax[i].set_title(s[i]) 460 | # if i in [5, 6, 7]: # share train and val loss y axes 461 | # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) 462 | except Exception as e: 463 | print('Warning: Plotting error for %s; %s' % (f, e)) 464 | 465 | ax[1].legend() 466 | fig.savefig(Path(save_dir) / 'results.png', dpi=200) 467 | 468 | 469 | def output_to_keypoint(output): 470 | # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] 471 | targets = [] 472 | for i, o in enumerate(output): 473 | kpts = o[:,6:] 474 | o = o[:,:6] 475 | for index, (*box, conf, cls) in enumerate(o.cpu().numpy()): 476 | targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf, *list(kpts.cpu().numpy()[index])]) 477 | return np.array(targets) 478 | 479 | 480 | def plot_skeleton_kpts(im, kpts, steps, orig_shape=None): 481 | #Plot the skeleton and keypointsfor coco datatset 482 | palette = np.array([[255, 128, 0], [255, 153, 51], [255, 178, 102], 483 | [230, 230, 0], [255, 153, 255], [153, 204, 255], 484 | [255, 102, 255], [255, 51, 255], [102, 178, 255], 485 | [51, 153, 255], [255, 153, 153], [255, 102, 102], 486 | [255, 51, 51], [153, 255, 153], [102, 255, 102], 487 | [51, 255, 51], [0, 255, 0], [0, 0, 255], [255, 0, 0], 488 | [255, 255, 255]]) 489 | 490 | skeleton = [[16, 14], [14, 12], [17, 15], [15, 13], [12, 13], [6, 12], 491 | [7, 13], [6, 7], [6, 8], [7, 9], [8, 10], [9, 11], [2, 3], 492 | [1, 2], [1, 3], [2, 4], [3, 5], [4, 6], [5, 7]] 493 | 494 | pose_limb_color = palette[[9, 9, 9, 9, 7, 7, 7, 0, 0, 0, 0, 0, 16, 16, 16, 16, 16, 16, 16]] 495 | pose_kpt_color = palette[[16, 16, 16, 16, 16, 0, 0, 0, 0, 0, 0, 9, 9, 9, 9, 9, 9]] 496 | radius = 5 497 | num_kpts = len(kpts) // steps 498 | 499 | for kid in range(num_kpts): 500 | r, g, b = pose_kpt_color[kid] 501 | x_coord, y_coord = kpts[steps * kid], kpts[steps * kid + 1] 502 | if not (x_coord % 640 == 0 or y_coord % 640 == 0): 503 | if steps == 3: 504 | conf = kpts[steps * kid + 2] 505 | if conf < 0.5: 506 | continue 507 | cv2.circle(im, (int(x_coord), int(y_coord)), radius, (int(r), int(g), int(b)), -1) 508 | 509 | for sk_id, sk in enumerate(skeleton): 510 | r, g, b = pose_limb_color[sk_id] 511 | pos1 = (int(kpts[(sk[0]-1)*steps]), int(kpts[(sk[0]-1)*steps+1])) 512 | pos2 = (int(kpts[(sk[1]-1)*steps]), int(kpts[(sk[1]-1)*steps+1])) 513 | if steps == 3: 514 | conf1 = kpts[(sk[0]-1)*steps+2] 515 | conf2 = kpts[(sk[1]-1)*steps+2] 516 | if conf1<0.5 or conf2<0.5: 517 | continue 518 | if pos1[0]%640 == 0 or pos1[1]%640==0 or pos1[0]<0 or pos1[1]<0: 519 | continue 520 | if pos2[0] % 640 == 0 or pos2[1] % 640 == 0 or pos2[0]<0 or pos2[1]<0: 521 | continue 522 | cv2.line(im, pos1, pos2, (int(r), int(g), int(b)), thickness=2) 523 | -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | # YOLOR PyTorch utils 2 | 3 | import datetime 4 | import logging 5 | import math 6 | import os 7 | import platform 8 | import subprocess 9 | import time 10 | from contextlib import contextmanager 11 | from copy import deepcopy 12 | from pathlib import Path 13 | 14 | import torch 15 | import torch.backends.cudnn as cudnn 16 | import torch.nn as nn 17 | import torch.nn.functional as F 18 | import torchvision 19 | 20 | try: 21 | import thop # for FLOPS computation 22 | except ImportError: 23 | thop = None 24 | logger = logging.getLogger(__name__) 25 | 26 | 27 | @contextmanager 28 | def torch_distributed_zero_first(local_rank: int): 29 | """ 30 | Decorator to make all processes in distributed training wait for each local_master to do something. 31 | """ 32 | if local_rank not in [-1, 0]: 33 | torch.distributed.barrier() 34 | yield 35 | if local_rank == 0: 36 | torch.distributed.barrier() 37 | 38 | 39 | def init_torch_seeds(seed=0): 40 | # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html 41 | torch.manual_seed(seed) 42 | if seed == 0: # slower, more reproducible 43 | cudnn.benchmark, cudnn.deterministic = False, True 44 | else: # faster, less reproducible 45 | cudnn.benchmark, cudnn.deterministic = True, False 46 | 47 | 48 | def date_modified(path=__file__): 49 | # return human-readable file modification date, i.e. '2021-3-26' 50 | t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime) 51 | return f'{t.year}-{t.month}-{t.day}' 52 | 53 | 54 | def git_describe(path=Path(__file__).parent): # path must be a directory 55 | # return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe 56 | s = f'git -C {path} describe --tags --long --always' 57 | try: 58 | return subprocess.check_output(s, shell=True, stderr=subprocess.STDOUT).decode()[:-1] 59 | except subprocess.CalledProcessError as e: 60 | return '' # not a git repository 61 | 62 | 63 | def select_device(device='', batch_size=None): 64 | # device = 'cpu' or '0' or '0,1,2,3' 65 | s = f'YOLOR 🚀 {git_describe() or date_modified()} torch {torch.__version__} ' # string 66 | cpu = device.lower() == 'cpu' 67 | if cpu: 68 | os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False 69 | elif device: # non-cpu device requested 70 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 71 | assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested' # check availability 72 | 73 | cuda = not cpu and torch.cuda.is_available() 74 | if cuda: 75 | n = torch.cuda.device_count() 76 | if n > 1 and batch_size: # check that batch_size is compatible with device_count 77 | assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}' 78 | space = ' ' * len(s) 79 | for i, d in enumerate(device.split(',') if device else range(n)): 80 | p = torch.cuda.get_device_properties(i) 81 | s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n" # bytes to MB 82 | else: 83 | s += 'CPU\n' 84 | 85 | logger.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) # emoji-safe 86 | return torch.device('cuda:0' if cuda else 'cpu') 87 | 88 | 89 | def time_synchronized(): 90 | # pytorch-accurate time 91 | if torch.cuda.is_available(): 92 | torch.cuda.synchronize() 93 | return time.time() 94 | 95 | 96 | def profile(x, ops, n=100, device=None): 97 | # profile a pytorch module or list of modules. Example usage: 98 | # x = torch.randn(16, 3, 640, 640) # input 99 | # m1 = lambda x: x * torch.sigmoid(x) 100 | # m2 = nn.SiLU() 101 | # profile(x, [m1, m2], n=100) # profile speed over 100 iterations 102 | 103 | device = device or torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') 104 | x = x.to(device) 105 | x.requires_grad = True 106 | print(torch.__version__, device.type, torch.cuda.get_device_properties(0) if device.type == 'cuda' else '') 107 | print(f"\n{'Params':>12s}{'GFLOPS':>12s}{'forward (ms)':>16s}{'backward (ms)':>16s}{'input':>24s}{'output':>24s}") 108 | for m in ops if isinstance(ops, list) else [ops]: 109 | m = m.to(device) if hasattr(m, 'to') else m # device 110 | m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m # type 111 | dtf, dtb, t = 0., 0., [0., 0., 0.] # dt forward, backward 112 | try: 113 | flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPS 114 | except: 115 | flops = 0 116 | 117 | for _ in range(n): 118 | t[0] = time_synchronized() 119 | y = m(x) 120 | t[1] = time_synchronized() 121 | try: 122 | _ = y.sum().backward() 123 | t[2] = time_synchronized() 124 | except: # no backward method 125 | t[2] = float('nan') 126 | dtf += (t[1] - t[0]) * 1000 / n # ms per op forward 127 | dtb += (t[2] - t[1]) * 1000 / n # ms per op backward 128 | 129 | s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' 130 | s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list' 131 | p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0 # parameters 132 | print(f'{p:12}{flops:12.4g}{dtf:16.4g}{dtb:16.4g}{str(s_in):>24s}{str(s_out):>24s}') 133 | 134 | 135 | def is_parallel(model): 136 | return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) 137 | 138 | 139 | def intersect_dicts(da, db, exclude=()): 140 | # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values 141 | return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape} 142 | 143 | 144 | def initialize_weights(model): 145 | for m in model.modules(): 146 | t = type(m) 147 | if t is nn.Conv2d: 148 | pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 149 | elif t is nn.BatchNorm2d: 150 | m.eps = 1e-3 151 | m.momentum = 0.03 152 | elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]: 153 | m.inplace = True 154 | 155 | 156 | def find_modules(model, mclass=nn.Conv2d): 157 | # Finds layer indices matching module class 'mclass' 158 | return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] 159 | 160 | 161 | def sparsity(model): 162 | # Return global model sparsity 163 | a, b = 0., 0. 164 | for p in model.parameters(): 165 | a += p.numel() 166 | b += (p == 0).sum() 167 | return b / a 168 | 169 | 170 | def prune(model, amount=0.3): 171 | # Prune model to requested global sparsity 172 | import torch.nn.utils.prune as prune 173 | print('Pruning model... ', end='') 174 | for name, m in model.named_modules(): 175 | if isinstance(m, nn.Conv2d): 176 | prune.l1_unstructured(m, name='weight', amount=amount) # prune 177 | prune.remove(m, 'weight') # make permanent 178 | print(' %.3g global sparsity' % sparsity(model)) 179 | 180 | 181 | def fuse_conv_and_bn(conv, bn): 182 | # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 183 | fusedconv = nn.Conv2d(conv.in_channels, 184 | conv.out_channels, 185 | kernel_size=conv.kernel_size, 186 | stride=conv.stride, 187 | padding=conv.padding, 188 | groups=conv.groups, 189 | bias=True).requires_grad_(False).to(conv.weight.device) 190 | 191 | # prepare filters 192 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 193 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 194 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape)) 195 | 196 | # prepare spatial bias 197 | b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias 198 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 199 | fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) 200 | 201 | return fusedconv 202 | 203 | 204 | def model_info(model, verbose=False, img_size=640): 205 | # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] 206 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 207 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 208 | if verbose: 209 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 210 | for i, (name, p) in enumerate(model.named_parameters()): 211 | name = name.replace('module_list.', '') 212 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 213 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 214 | 215 | try: # FLOPS 216 | from thop import profile 217 | stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32 218 | img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input 219 | flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPS 220 | img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float 221 | fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPS 222 | except (ImportError, Exception): 223 | fs = '' 224 | 225 | logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") 226 | 227 | 228 | def load_classifier(name='resnet101', n=2): 229 | # Loads a pretrained model reshaped to n-class output 230 | model = torchvision.models.__dict__[name](pretrained=True) 231 | 232 | # ResNet model properties 233 | # input_size = [3, 224, 224] 234 | # input_space = 'RGB' 235 | # input_range = [0, 1] 236 | # mean = [0.485, 0.456, 0.406] 237 | # std = [0.229, 0.224, 0.225] 238 | 239 | # Reshape output to n classes 240 | filters = model.fc.weight.shape[1] 241 | model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True) 242 | model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True) 243 | model.fc.out_features = n 244 | return model 245 | 246 | 247 | def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416) 248 | # scales img(bs,3,y,x) by ratio constrained to gs-multiple 249 | if ratio == 1.0: 250 | return img 251 | else: 252 | h, w = img.shape[2:] 253 | s = (int(h * ratio), int(w * ratio)) # new size 254 | img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize 255 | if not same_shape: # pad/crop img 256 | h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] 257 | return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean 258 | 259 | 260 | def copy_attr(a, b, include=(), exclude=()): 261 | # Copy attributes from b to a, options to only include [...] and to exclude [...] 262 | for k, v in b.__dict__.items(): 263 | if (len(include) and k not in include) or k.startswith('_') or k in exclude: 264 | continue 265 | else: 266 | setattr(a, k, v) 267 | 268 | 269 | class ModelEMA: 270 | """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models 271 | Keep a moving average of everything in the model state_dict (parameters and buffers). 272 | This is intended to allow functionality like 273 | https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage 274 | A smoothed version of the weights is necessary for some training schemes to perform well. 275 | This class is sensitive where it is initialized in the sequence of model init, 276 | GPU assignment and distributed training wrappers. 277 | """ 278 | 279 | def __init__(self, model, decay=0.9999, updates=0): 280 | # Create EMA 281 | self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA 282 | # if next(model.parameters()).device.type != 'cpu': 283 | # self.ema.half() # FP16 EMA 284 | self.updates = updates # number of EMA updates 285 | self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) 286 | for p in self.ema.parameters(): 287 | p.requires_grad_(False) 288 | 289 | def update(self, model): 290 | # Update EMA parameters 291 | with torch.no_grad(): 292 | self.updates += 1 293 | d = self.decay(self.updates) 294 | 295 | msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict 296 | for k, v in self.ema.state_dict().items(): 297 | if v.dtype.is_floating_point: 298 | v *= d 299 | v += (1. - d) * msd[k].detach() 300 | 301 | def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): 302 | # Update EMA attributes 303 | copy_attr(self.ema, model, include, exclude) 304 | 305 | 306 | class BatchNormXd(torch.nn.modules.batchnorm._BatchNorm): 307 | def _check_input_dim(self, input): 308 | # The only difference between BatchNorm1d, BatchNorm2d, BatchNorm3d, etc 309 | # is this method that is overwritten by the sub-class 310 | # This original goal of this method was for tensor sanity checks 311 | # If you're ok bypassing those sanity checks (eg. if you trust your inference 312 | # to provide the right dimensional inputs), then you can just use this method 313 | # for easy conversion from SyncBatchNorm 314 | # (unfortunately, SyncBatchNorm does not store the original class - if it did 315 | # we could return the one that was originally created) 316 | return 317 | 318 | def revert_sync_batchnorm(module): 319 | # this is very similar to the function that it is trying to revert: 320 | # https://github.com/pytorch/pytorch/blob/c8b3686a3e4ba63dc59e5dcfe5db3430df256833/torch/nn/modules/batchnorm.py#L679 321 | module_output = module 322 | if isinstance(module, torch.nn.modules.batchnorm.SyncBatchNorm): 323 | new_cls = BatchNormXd 324 | module_output = BatchNormXd(module.num_features, 325 | module.eps, module.momentum, 326 | module.affine, 327 | module.track_running_stats) 328 | if module.affine: 329 | with torch.no_grad(): 330 | module_output.weight = module.weight 331 | module_output.bias = module.bias 332 | module_output.running_mean = module.running_mean 333 | module_output.running_var = module.running_var 334 | module_output.num_batches_tracked = module.num_batches_tracked 335 | if hasattr(module, "qconfig"): 336 | module_output.qconfig = module.qconfig 337 | for name, child in module.named_children(): 338 | module_output.add_module(name, revert_sync_batchnorm(child)) 339 | del module 340 | return module_output 341 | 342 | 343 | class TracedModel(nn.Module): 344 | 345 | def __init__(self, model=None, device=None, img_size=(640,640)): 346 | super(TracedModel, self).__init__() 347 | 348 | print(" Convert model to Traced-model... ") 349 | self.stride = model.stride 350 | self.names = model.names 351 | self.model = model 352 | 353 | self.model = revert_sync_batchnorm(self.model) 354 | self.model.to('cpu') 355 | self.model.eval() 356 | 357 | self.detect_layer = self.model.model[-1] 358 | self.model.traced = True 359 | 360 | rand_example = torch.rand(1, 3, img_size, img_size) 361 | 362 | traced_script_module = torch.jit.trace(self.model, rand_example, strict=False) 363 | #traced_script_module = torch.jit.script(self.model) 364 | traced_script_module.save("traced_model.pt") 365 | print(" traced_script_module saved! ") 366 | self.model = traced_script_module 367 | self.model.to(device) 368 | self.detect_layer.to(device) 369 | print(" model is traced! \n") 370 | 371 | def forward(self, x, augment=False, profile=False): 372 | out = self.model(x) 373 | out = self.detect_layer(out) 374 | return out -------------------------------------------------------------------------------- /utils/wandb_logging/__init__.py: -------------------------------------------------------------------------------- 1 | # init -------------------------------------------------------------------------------- /utils/wandb_logging/log_dataset.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | import yaml 4 | 5 | from wandb_utils import WandbLogger 6 | 7 | WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' 8 | 9 | 10 | def create_dataset_artifact(opt): 11 | with open(opt.data) as f: 12 | data = yaml.load(f, Loader=yaml.SafeLoader) # data dict 13 | logger = WandbLogger(opt, '', None, data, job_type='Dataset Creation') 14 | 15 | 16 | if __name__ == '__main__': 17 | parser = argparse.ArgumentParser() 18 | parser.add_argument('--data', type=str, default='data/coco.yaml', help='data.yaml path') 19 | parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') 20 | parser.add_argument('--project', type=str, default='YOLOR', help='name of W&B Project') 21 | opt = parser.parse_args() 22 | opt.resume = False # Explicitly disallow resume check for dataset upload job 23 | 24 | create_dataset_artifact(opt) 25 | -------------------------------------------------------------------------------- /utils/wandb_logging/wandb_utils.py: -------------------------------------------------------------------------------- 1 | import json 2 | import sys 3 | from pathlib import Path 4 | 5 | import torch 6 | import yaml 7 | from tqdm import tqdm 8 | 9 | sys.path.append(str(Path(__file__).parent.parent.parent)) # add utils/ to path 10 | from utils.datasets import LoadImagesAndLabels 11 | from utils.datasets import img2label_paths 12 | from utils.general import colorstr, xywh2xyxy, check_dataset 13 | 14 | try: 15 | import wandb 16 | from wandb import init, finish 17 | except ImportError: 18 | wandb = None 19 | 20 | WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' 21 | 22 | 23 | def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX): 24 | return from_string[len(prefix):] 25 | 26 | 27 | def check_wandb_config_file(data_config_file): 28 | wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1)) # updated data.yaml path 29 | if Path(wandb_config).is_file(): 30 | return wandb_config 31 | return data_config_file 32 | 33 | 34 | def get_run_info(run_path): 35 | run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX)) 36 | run_id = run_path.stem 37 | project = run_path.parent.stem 38 | model_artifact_name = 'run_' + run_id + '_model' 39 | return run_id, project, model_artifact_name 40 | 41 | 42 | def check_wandb_resume(opt): 43 | process_wandb_config_ddp_mode(opt) if opt.global_rank not in [-1, 0] else None 44 | if isinstance(opt.resume, str): 45 | if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): 46 | if opt.global_rank not in [-1, 0]: # For resuming DDP runs 47 | run_id, project, model_artifact_name = get_run_info(opt.resume) 48 | api = wandb.Api() 49 | artifact = api.artifact(project + '/' + model_artifact_name + ':latest') 50 | modeldir = artifact.download() 51 | opt.weights = str(Path(modeldir) / "last.pt") 52 | return True 53 | return None 54 | 55 | 56 | def process_wandb_config_ddp_mode(opt): 57 | with open(opt.data) as f: 58 | data_dict = yaml.load(f, Loader=yaml.SafeLoader) # data dict 59 | train_dir, val_dir = None, None 60 | if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX): 61 | api = wandb.Api() 62 | train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias) 63 | train_dir = train_artifact.download() 64 | train_path = Path(train_dir) / 'data/images/' 65 | data_dict['train'] = str(train_path) 66 | 67 | if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX): 68 | api = wandb.Api() 69 | val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias) 70 | val_dir = val_artifact.download() 71 | val_path = Path(val_dir) / 'data/images/' 72 | data_dict['val'] = str(val_path) 73 | if train_dir or val_dir: 74 | ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml') 75 | with open(ddp_data_path, 'w') as f: 76 | yaml.dump(data_dict, f) 77 | opt.data = ddp_data_path 78 | 79 | 80 | class WandbLogger(): 81 | def __init__(self, opt, name, run_id, data_dict, job_type='Training'): 82 | # Pre-training routine -- 83 | self.job_type = job_type 84 | self.wandb, self.wandb_run, self.data_dict = wandb, None if not wandb else wandb.run, data_dict 85 | # It's more elegant to stick to 1 wandb.init call, but useful config data is overwritten in the WandbLogger's wandb.init call 86 | if isinstance(opt.resume, str): # checks resume from artifact 87 | if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): 88 | run_id, project, model_artifact_name = get_run_info(opt.resume) 89 | model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name 90 | assert wandb, 'install wandb to resume wandb runs' 91 | # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config 92 | self.wandb_run = wandb.init(id=run_id, project=project, resume='allow') 93 | opt.resume = model_artifact_name 94 | elif self.wandb: 95 | self.wandb_run = wandb.init(config=opt, 96 | resume="allow", 97 | project='YOLOR' if opt.project == 'runs/train' else Path(opt.project).stem, 98 | name=name, 99 | job_type=job_type, 100 | id=run_id) if not wandb.run else wandb.run 101 | if self.wandb_run: 102 | if self.job_type == 'Training': 103 | if not opt.resume: 104 | wandb_data_dict = self.check_and_upload_dataset(opt) if opt.upload_dataset else data_dict 105 | # Info useful for resuming from artifacts 106 | self.wandb_run.config.opt = vars(opt) 107 | self.wandb_run.config.data_dict = wandb_data_dict 108 | self.data_dict = self.setup_training(opt, data_dict) 109 | if self.job_type == 'Dataset Creation': 110 | self.data_dict = self.check_and_upload_dataset(opt) 111 | else: 112 | prefix = colorstr('wandb: ') 113 | print(f"{prefix}Install Weights & Biases for YOLOR logging with 'pip install wandb' (recommended)") 114 | 115 | def check_and_upload_dataset(self, opt): 116 | assert wandb, 'Install wandb to upload dataset' 117 | check_dataset(self.data_dict) 118 | config_path = self.log_dataset_artifact(opt.data, 119 | opt.single_cls, 120 | 'YOLOR' if opt.project == 'runs/train' else Path(opt.project).stem) 121 | print("Created dataset config file ", config_path) 122 | with open(config_path) as f: 123 | wandb_data_dict = yaml.load(f, Loader=yaml.SafeLoader) 124 | return wandb_data_dict 125 | 126 | def setup_training(self, opt, data_dict): 127 | self.log_dict, self.current_epoch, self.log_imgs = {}, 0, 16 # Logging Constants 128 | self.bbox_interval = opt.bbox_interval 129 | if isinstance(opt.resume, str): 130 | modeldir, _ = self.download_model_artifact(opt) 131 | if modeldir: 132 | self.weights = Path(modeldir) / "last.pt" 133 | config = self.wandb_run.config 134 | opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp = str( 135 | self.weights), config.save_period, config.total_batch_size, config.bbox_interval, config.epochs, \ 136 | config.opt['hyp'] 137 | data_dict = dict(self.wandb_run.config.data_dict) # eliminates the need for config file to resume 138 | if 'val_artifact' not in self.__dict__: # If --upload_dataset is set, use the existing artifact, don't download 139 | self.train_artifact_path, self.train_artifact = self.download_dataset_artifact(data_dict.get('train'), 140 | opt.artifact_alias) 141 | self.val_artifact_path, self.val_artifact = self.download_dataset_artifact(data_dict.get('val'), 142 | opt.artifact_alias) 143 | self.result_artifact, self.result_table, self.val_table, self.weights = None, None, None, None 144 | if self.train_artifact_path is not None: 145 | train_path = Path(self.train_artifact_path) / 'data/images/' 146 | data_dict['train'] = str(train_path) 147 | if self.val_artifact_path is not None: 148 | val_path = Path(self.val_artifact_path) / 'data/images/' 149 | data_dict['val'] = str(val_path) 150 | self.val_table = self.val_artifact.get("val") 151 | self.map_val_table_path() 152 | if self.val_artifact is not None: 153 | self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") 154 | self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"]) 155 | if opt.bbox_interval == -1: 156 | self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1 157 | return data_dict 158 | 159 | def download_dataset_artifact(self, path, alias): 160 | if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX): 161 | dataset_artifact = wandb.use_artifact(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias) 162 | assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'" 163 | datadir = dataset_artifact.download() 164 | return datadir, dataset_artifact 165 | return None, None 166 | 167 | def download_model_artifact(self, opt): 168 | if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): 169 | model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest") 170 | assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist' 171 | modeldir = model_artifact.download() 172 | epochs_trained = model_artifact.metadata.get('epochs_trained') 173 | total_epochs = model_artifact.metadata.get('total_epochs') 174 | assert epochs_trained < total_epochs, 'training to %g epochs is finished, nothing to resume.' % ( 175 | total_epochs) 176 | return modeldir, model_artifact 177 | return None, None 178 | 179 | def log_model(self, path, opt, epoch, fitness_score, best_model=False): 180 | model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', type='model', metadata={ 181 | 'original_url': str(path), 182 | 'epochs_trained': epoch + 1, 183 | 'save period': opt.save_period, 184 | 'project': opt.project, 185 | 'total_epochs': opt.epochs, 186 | 'fitness_score': fitness_score 187 | }) 188 | model_artifact.add_file(str(path / 'last.pt'), name='last.pt') 189 | wandb.log_artifact(model_artifact, 190 | aliases=['latest', 'epoch ' + str(self.current_epoch), 'best' if best_model else '']) 191 | print("Saving model artifact on epoch ", epoch + 1) 192 | 193 | def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False): 194 | with open(data_file) as f: 195 | data = yaml.load(f, Loader=yaml.SafeLoader) # data dict 196 | nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names']) 197 | names = {k: v for k, v in enumerate(names)} # to index dictionary 198 | self.train_artifact = self.create_dataset_table(LoadImagesAndLabels( 199 | data['train']), names, name='train') if data.get('train') else None 200 | self.val_artifact = self.create_dataset_table(LoadImagesAndLabels( 201 | data['val']), names, name='val') if data.get('val') else None 202 | if data.get('train'): 203 | data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train') 204 | if data.get('val'): 205 | data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val') 206 | path = data_file if overwrite_config else '_wandb.'.join(data_file.rsplit('.', 1)) # updated data.yaml path 207 | data.pop('download', None) 208 | with open(path, 'w') as f: 209 | yaml.dump(data, f) 210 | 211 | if self.job_type == 'Training': # builds correct artifact pipeline graph 212 | self.wandb_run.use_artifact(self.val_artifact) 213 | self.wandb_run.use_artifact(self.train_artifact) 214 | self.val_artifact.wait() 215 | self.val_table = self.val_artifact.get('val') 216 | self.map_val_table_path() 217 | else: 218 | self.wandb_run.log_artifact(self.train_artifact) 219 | self.wandb_run.log_artifact(self.val_artifact) 220 | return path 221 | 222 | def map_val_table_path(self): 223 | self.val_table_map = {} 224 | print("Mapping dataset") 225 | for i, data in enumerate(tqdm(self.val_table.data)): 226 | self.val_table_map[data[3]] = data[0] 227 | 228 | def create_dataset_table(self, dataset, class_to_id, name='dataset'): 229 | # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging 230 | artifact = wandb.Artifact(name=name, type="dataset") 231 | img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None 232 | img_files = tqdm(dataset.img_files) if not img_files else img_files 233 | for img_file in img_files: 234 | if Path(img_file).is_dir(): 235 | artifact.add_dir(img_file, name='data/images') 236 | labels_path = 'labels'.join(dataset.path.rsplit('images', 1)) 237 | artifact.add_dir(labels_path, name='data/labels') 238 | else: 239 | artifact.add_file(img_file, name='data/images/' + Path(img_file).name) 240 | label_file = Path(img2label_paths([img_file])[0]) 241 | artifact.add_file(str(label_file), 242 | name='data/labels/' + label_file.name) if label_file.exists() else None 243 | table = wandb.Table(columns=["id", "train_image", "Classes", "name"]) 244 | class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()]) 245 | for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)): 246 | height, width = shapes[0] 247 | labels[:, 2:] = (xywh2xyxy(labels[:, 2:].view(-1, 4))) * torch.Tensor([width, height, width, height]) 248 | box_data, img_classes = [], {} 249 | for cls, *xyxy in labels[:, 1:].tolist(): 250 | cls = int(cls) 251 | box_data.append({"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, 252 | "class_id": cls, 253 | "box_caption": "%s" % (class_to_id[cls]), 254 | "scores": {"acc": 1}, 255 | "domain": "pixel"}) 256 | img_classes[cls] = class_to_id[cls] 257 | boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space 258 | table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), json.dumps(img_classes), 259 | Path(paths).name) 260 | artifact.add(table, name) 261 | return artifact 262 | 263 | def log_training_progress(self, predn, path, names): 264 | if self.val_table and self.result_table: 265 | class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()]) 266 | box_data = [] 267 | total_conf = 0 268 | for *xyxy, conf, cls in predn.tolist(): 269 | if conf >= 0.25: 270 | box_data.append( 271 | {"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, 272 | "class_id": int(cls), 273 | "box_caption": "%s %.3f" % (names[cls], conf), 274 | "scores": {"class_score": conf}, 275 | "domain": "pixel"}) 276 | total_conf = total_conf + conf 277 | boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space 278 | id = self.val_table_map[Path(path).name] 279 | self.result_table.add_data(self.current_epoch, 280 | id, 281 | wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set), 282 | total_conf / max(1, len(box_data)) 283 | ) 284 | 285 | def log(self, log_dict): 286 | if self.wandb_run: 287 | for key, value in log_dict.items(): 288 | self.log_dict[key] = value 289 | 290 | def end_epoch(self, best_result=False): 291 | if self.wandb_run: 292 | wandb.log(self.log_dict) 293 | self.log_dict = {} 294 | if self.result_artifact: 295 | train_results = wandb.JoinedTable(self.val_table, self.result_table, "id") 296 | self.result_artifact.add(train_results, 'result') 297 | wandb.log_artifact(self.result_artifact, aliases=['latest', 'epoch ' + str(self.current_epoch), 298 | ('best' if best_result else '')]) 299 | self.result_table = wandb.Table(["epoch", "id", "prediction", "avg_confidence"]) 300 | self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") 301 | 302 | def finish_run(self): 303 | if self.wandb_run: 304 | if self.log_dict: 305 | wandb.log(self.log_dict) 306 | wandb.run.finish() 307 | --------------------------------------------------------------------------------