├── Dockerfile ├── LICENSE ├── README.md ├── Socail_Distance_Mapping.ipynb ├── data ├── coco.yaml ├── coco128.yaml ├── hyp.finetune.yaml ├── hyp.scratch.yaml ├── images │ ├── bus.jpg │ └── zidane.jpg ├── scripts │ ├── get_coco.sh │ └── get_voc.sh └── voc.yaml ├── detect.py ├── hubconf.py ├── inference └── people1.mp4 ├── requirements.txt ├── test.py ├── train.py ├── tutorial.ipynb └── utils ├── activations.py ├── autoanchor.py ├── datasets.py ├── general.py ├── google_app_engine ├── Dockerfile ├── additional_requirements.txt └── app.yaml ├── google_utils.py ├── loss.py ├── metrics.py ├── plots.py ├── torch_utils.py └── utils.py /Dockerfile: -------------------------------------------------------------------------------- 1 | # Start FROM Nvidia PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch 2 | FROM nvcr.io/nvidia/pytorch:20.10-py3 3 | 4 | # Install dependencies 5 | RUN pip install --upgrade pip 6 | # COPY requirements.txt . 7 | # RUN pip install -r requirements.txt 8 | RUN pip install gsutil 9 | 10 | # Create working directory 11 | RUN mkdir -p /usr/src/app 12 | WORKDIR /usr/src/app 13 | 14 | # Copy contents 15 | COPY . /usr/src/app 16 | 17 | # Copy weights 18 | #RUN python3 -c "from models import *; \ 19 | #attempt_download('weights/yolov5s.pt'); \ 20 | #attempt_download('weights/yolov5m.pt'); \ 21 | #attempt_download('weights/yolov5l.pt')" 22 | 23 | 24 | # --------------------------------------------------- Extras Below --------------------------------------------------- 25 | 26 | # Build and Push 27 | # t=ultralytics/yolov5:latest && sudo docker build -t $t . && sudo docker push $t 28 | # for v in {300..303}; do t=ultralytics/coco:v$v && sudo docker build -t $t . && sudo docker push $t; done 29 | 30 | # Pull and Run 31 | # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all $t 32 | 33 | # Pull and Run with local directory access 34 | # t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all -v "$(pwd)"/coco:/usr/src/coco $t 35 | 36 | # Kill all 37 | # sudo docker kill $(sudo docker ps -q) 38 | 39 | # Kill all image-based 40 | # sudo docker kill $(sudo docker ps -a -q --filter ancestor=ultralytics/yolov5:latest) 41 | 42 | # Bash into running container 43 | # sudo docker container exec -it ba65811811ab bash 44 | 45 | # Bash into stopped container 46 | # sudo docker commit 092b16b25c5b usr/resume && sudo docker run -it --gpus all --ipc=host -v "$(pwd)"/coco:/usr/src/coco --entrypoint=sh usr/resume 47 | 48 | # Send weights to GCP 49 | # python -c "from utils.general import *; strip_optimizer('runs/train/exp0_*/weights/best.pt', 'tmp.pt')" && gsutil cp tmp.pt gs://*.pt 50 | 51 | # Clean up 52 | # docker system prune -a --volumes 53 | -------------------------------------------------------------------------------- /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 | . -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | 2 | 3 |   4 | 5 | ![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg) 6 | 7 | This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. **All code and models are under active development, and are subject to modification or deletion without notice.** Use at your own risk. 8 | 9 | ** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS. EfficientDet data from [google/automl](https://github.com/google/automl) at batch size 8. 10 | 11 | - **August 13, 2020**: [v3.0 release](https://github.com/ultralytics/yolov5/releases/tag/v3.0): nn.Hardswish() activations, data autodownload, native AMP. 12 | - **July 23, 2020**: [v2.0 release](https://github.com/ultralytics/yolov5/releases/tag/v2.0): improved model definition, training and mAP. 13 | - **June 22, 2020**: [PANet](https://arxiv.org/abs/1803.01534) updates: new heads, reduced parameters, improved speed and mAP [364fcfd](https://github.com/ultralytics/yolov5/commit/364fcfd7dba53f46edd4f04c037a039c0a287972). 14 | - **June 19, 2020**: [FP16](https://pytorch.org/docs/stable/nn.html#torch.nn.Module.half) as new default for smaller checkpoints and faster inference [d4c6674](https://github.com/ultralytics/yolov5/commit/d4c6674c98e19df4c40e33a777610a18d1961145). 15 | - **June 9, 2020**: [CSP](https://github.com/WongKinYiu/CrossStagePartialNetworks) updates: improved speed, size, and accuracy (credit to @WongKinYiu for CSP). 16 | - **May 27, 2020**: Public release. YOLOv5 models are SOTA among all known YOLO implementations. 17 | 18 | 19 | ## Pretrained Checkpoints 20 | 21 | | Model | APval | APtest | AP50 | SpeedGPU | FPSGPU || params | FLOPS | 22 | |---------- |------ |------ |------ | -------- | ------| ------ |------ | :------: | 23 | | [YOLOv5s](https://github.com/ultralytics/yolov5/releases) | 37.0 | 37.0 | 56.2 | **2.4ms** | **416** || 7.5M | 13.2B 24 | | [YOLOv5m](https://github.com/ultralytics/yolov5/releases) | 44.3 | 44.3 | 63.2 | 3.4ms | 294 || 21.8M | 39.4B 25 | | [YOLOv5l](https://github.com/ultralytics/yolov5/releases) | 47.7 | 47.7 | 66.5 | 4.4ms | 227 || 47.8M | 88.1B 26 | | [YOLOv5x](https://github.com/ultralytics/yolov5/releases) | **49.2** | **49.2** | **67.7** | 6.9ms | 145 || 89.0M | 166.4B 27 | | | | | | | || | 28 | | [YOLOv5x](https://github.com/ultralytics/yolov5/releases) + TTA|**50.8**| **50.8** | **68.9** | 25.5ms | 39 || 89.0M | 354.3B 29 | | | | | | | || | 30 | | [YOLOv3-SPP](https://github.com/ultralytics/yolov5/releases) | 45.6 | 45.5 | 65.2 | 4.5ms | 222 || 63.0M | 118.0B 31 | 32 | ** APtest denotes COCO [test-dev2017](http://cocodataset.org/#upload) server results, all other AP results denote val2017 accuracy. 33 | ** All AP numbers are for single-model single-scale without ensemble or TTA. **Reproduce mAP** by `python test.py --data coco.yaml --img 640 --conf 0.001 --iou 0.65` 34 | ** SpeedGPU averaged over 5000 COCO val2017 images using a GCP [n1-standard-16](https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types) V100 instance, and includes image preprocessing, FP16 inference, postprocessing and NMS. NMS is 1-2ms/img. **Reproduce speed** by `python test.py --data coco.yaml --img 640 --conf 0.25 --iou 0.45` 35 | ** All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation). 36 | ** Test Time Augmentation ([TTA](https://github.com/ultralytics/yolov5/issues/303)) runs at 3 image sizes. **Reproduce TTA** by `python test.py --data coco.yaml --img 832 --iou 0.65 --augment` 37 | 38 | ## Requirements 39 | 40 | Python 3.8 or later with all [requirements.txt](https://github.com/ultralytics/yolov5/blob/master/requirements.txt) dependencies installed, including `torch>=1.7`. To install run: 41 | ```bash 42 | $ pip install -r requirements.txt 43 | ``` 44 | 45 | 46 | ## Tutorials 47 | 48 | * [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)  🚀 RECOMMENDED 49 | * [Weights & Biases Logging](https://github.com/ultralytics/yolov5/issues/1289)  🌟 NEW 50 | * [Multi-GPU Training](https://github.com/ultralytics/yolov5/issues/475) 51 | * [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36)  ⭐ NEW 52 | * [ONNX and TorchScript Export](https://github.com/ultralytics/yolov5/issues/251) 53 | * [Test-Time Augmentation (TTA)](https://github.com/ultralytics/yolov5/issues/303) 54 | * [Model Ensembling](https://github.com/ultralytics/yolov5/issues/318) 55 | * [Model Pruning/Sparsity](https://github.com/ultralytics/yolov5/issues/304) 56 | * [Hyperparameter Evolution](https://github.com/ultralytics/yolov5/issues/607) 57 | * [Transfer Learning with Frozen Layers](https://github.com/ultralytics/yolov5/issues/1314)  ⭐ NEW 58 | * [TensorRT Deployment](https://github.com/wang-xinyu/tensorrtx) 59 | 60 | 61 | ## Environments 62 | 63 | YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled): 64 | 65 | - **Google Colab Notebook** with free GPU: Open In Colab 66 | - **Kaggle Notebook** with free GPU: [https://www.kaggle.com/ultralytics/yolov5](https://www.kaggle.com/ultralytics/yolov5) 67 | - **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) 68 | - **Docker Image** https://hub.docker.com/r/ultralytics/yolov5. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) ![Docker Pulls](https://img.shields.io/docker/pulls/ultralytics/yolov5?logo=docker) 69 | 70 | 71 | ## Inference 72 | 73 | detect.py runs inference on a variety of sources, downloading models automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases) and saving results to `runs/detect`. 74 | ```bash 75 | $ python detect.py --source 0 # webcam 76 | file.jpg # image 77 | file.mp4 # video 78 | path/ # directory 79 | path/*.jpg # glob 80 | rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa # rtsp stream 81 | rtmp://192.168.1.105/live/test # rtmp stream 82 | http://112.50.243.8/PLTV/88888888/224/3221225900/1.m3u8 # http stream 83 | ``` 84 | 85 | To run inference on example images in `data/images`: 86 | ```bash 87 | $ python detect.py --source data/images --weights yolov5s.pt --conf 0.25 88 | 89 | Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.25, device='', img_size=640, iou_thres=0.45, save_conf=False, save_dir='runs/detect', save_txt=False, source='data/images/', update=False, view_img=False, weights=['yolov5s.pt']) 90 | Using torch 1.7.0+cu101 CUDA:0 (Tesla V100-SXM2-16GB, 16130MB) 91 | 92 | Downloading https://github.com/ultralytics/yolov5/releases/download/v3.1/yolov5s.pt to yolov5s.pt... 100%|██████████████| 14.5M/14.5M [00:00<00:00, 21.3MB/s] 93 | 94 | Fusing layers... 95 | Model Summary: 232 layers, 7459581 parameters, 0 gradients 96 | image 1/2 data/images/bus.jpg: 640x480 4 persons, 1 buss, 1 skateboards, Done. (0.012s) 97 | image 2/2 data/images/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.012s) 98 | Results saved to runs/detect/exp 99 | Done. (0.113s) 100 | ``` 101 | 102 | 103 | ### PyTorch Hub 104 | 105 | To run **batched inference** with YOLOv5 and [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36): 106 | ```python 107 | import torch 108 | from PIL import Image 109 | 110 | # Model 111 | model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True).autoshape() # for PIL/cv2/np inputs and NMS 112 | 113 | # Images 114 | img1 = Image.open('zidane.jpg') 115 | img2 = Image.open('bus.jpg') 116 | imgs = [img1, img2] # batched list of images 117 | 118 | # Inference 119 | prediction = model(imgs, size=640) # includes NMS 120 | ``` 121 | 122 | 123 | ## Training 124 | 125 | Download [COCO](https://github.com/ultralytics/yolov5/blob/master/data/scripts/get_coco.sh) and run command below. Training times for YOLOv5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest `--batch-size` your GPU allows (batch sizes shown for 16 GB devices). 126 | ```bash 127 | $ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64 128 | yolov5m 40 129 | yolov5l 24 130 | yolov5x 16 131 | ``` 132 | 133 | 134 | 135 | ## Citation 136 | 137 | [![DOI](https://zenodo.org/badge/264818686.svg)](https://zenodo.org/badge/latestdoi/264818686) 138 | 139 | 140 | ## About Us 141 | 142 | Ultralytics is a U.S.-based particle physics and AI startup with over 6 years of expertise supporting government, academic and business clients. We offer a wide range of vision AI services, spanning from simple expert advice up to delivery of fully customized, end-to-end production solutions, including: 143 | - **Cloud-based AI** systems operating on **hundreds of HD video streams in realtime.** 144 | - **Edge AI** integrated into custom iOS and Android apps for realtime **30 FPS video inference.** 145 | - **Custom data training**, hyperparameter evolution, and model exportation to any destination. 146 | 147 | For business inquiries and professional support requests please visit us at https://www.ultralytics.com. 148 | 149 | 150 | ## Contact 151 | 152 | **Issues should be raised directly in the repository.** For business inquiries or professional support requests please visit https://www.ultralytics.com or email Glenn Jocher at glenn.jocher@ultralytics.com. 153 | -------------------------------------------------------------------------------- /data/coco.yaml: -------------------------------------------------------------------------------- 1 | # COCO 2017 dataset http://cocodataset.org 2 | # Train command: python train.py --data coco.yaml 3 | # Default dataset location is next to /yolov5: 4 | # /parent_folder 5 | # /coco 6 | # /yolov5 7 | 8 | 9 | # download command/URL (optional) 10 | download: bash data/scripts/get_coco.sh 11 | 12 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 13 | train: ../coco/train2017.txt # 118287 images 14 | val: ../coco/val2017.txt # 5000 images 15 | test: ../coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794 16 | 17 | # number of classes 18 | nc: 80 19 | 20 | # class names 21 | names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 22 | 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 23 | 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 24 | 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 25 | 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 26 | 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 27 | 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 28 | 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 29 | 'hair drier', 'toothbrush'] 30 | 31 | # Print classes 32 | # with open('data/coco.yaml') as f: 33 | # d = yaml.load(f, Loader=yaml.FullLoader) # dict 34 | # for i, x in enumerate(d['names']): 35 | # print(i, x) 36 | -------------------------------------------------------------------------------- /data/coco128.yaml: -------------------------------------------------------------------------------- 1 | # COCO 2017 dataset http://cocodataset.org - first 128 training images 2 | # Train command: python train.py --data coco128.yaml 3 | # Default dataset location is next to /yolov5: 4 | # /parent_folder 5 | # /coco128 6 | # /yolov5 7 | 8 | 9 | # download command/URL (optional) 10 | download: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128.zip 11 | 12 | # train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/] 13 | train: ../coco128/images/train2017/ # 128 images 14 | val: ../coco128/images/train2017/ # 128 images 15 | 16 | # number of classes 17 | nc: 80 18 | 19 | # class names 20 | names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 21 | 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 22 | 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 23 | 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 24 | 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 25 | 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 26 | 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 27 | 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 28 | 'hair drier', 'toothbrush'] 29 | -------------------------------------------------------------------------------- /data/hyp.finetune.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for VOC finetuning 2 | # python train.py --batch 64 --weights yolov5m.pt --data voc.yaml --img 512 --epochs 50 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | # Hyperparameter Evolution Results 7 | # Generations: 306 8 | # P R mAP.5 mAP.5:.95 box obj cls 9 | # Metrics: 0.6 0.936 0.896 0.684 0.0115 0.00805 0.00146 10 | 11 | lr0: 0.0032 12 | lrf: 0.12 13 | momentum: 0.843 14 | weight_decay: 0.00036 15 | warmup_epochs: 2.0 16 | warmup_momentum: 0.5 17 | warmup_bias_lr: 0.05 18 | box: 0.0296 19 | cls: 0.243 20 | cls_pw: 0.631 21 | obj: 0.301 22 | obj_pw: 0.911 23 | iou_t: 0.2 24 | anchor_t: 2.91 25 | # anchors: 3.63 26 | fl_gamma: 0.0 27 | hsv_h: 0.0138 28 | hsv_s: 0.664 29 | hsv_v: 0.464 30 | degrees: 0.373 31 | translate: 0.245 32 | scale: 0.898 33 | shear: 0.602 34 | perspective: 0.0 35 | flipud: 0.00856 36 | fliplr: 0.5 37 | mosaic: 1.0 38 | mixup: 0.243 39 | -------------------------------------------------------------------------------- /data/hyp.scratch.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for COCO training from scratch 2 | # python train.py --batch 40 --cfg yolov5m.yaml --weights '' --data coco.yaml --img 640 --epochs 300 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) 7 | lrf: 0.2 # final OneCycleLR learning rate (lr0 * lrf) 8 | momentum: 0.937 # SGD momentum/Adam beta1 9 | weight_decay: 0.0005 # optimizer weight decay 5e-4 10 | warmup_epochs: 3.0 # warmup epochs (fractions ok) 11 | warmup_momentum: 0.8 # warmup initial momentum 12 | warmup_bias_lr: 0.1 # warmup initial bias lr 13 | box: 0.05 # box loss gain 14 | cls: 0.5 # cls loss gain 15 | cls_pw: 1.0 # cls BCELoss positive_weight 16 | obj: 1.0 # obj loss gain (scale with pixels) 17 | obj_pw: 1.0 # obj BCELoss positive_weight 18 | iou_t: 0.20 # IoU training threshold 19 | anchor_t: 4.0 # anchor-multiple threshold 20 | # anchors: 3 # anchors per output layer (0 to ignore) 21 | fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) 22 | hsv_h: 0.015 # image HSV-Hue augmentation (fraction) 23 | hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) 24 | hsv_v: 0.4 # image HSV-Value augmentation (fraction) 25 | degrees: 0.0 # image rotation (+/- deg) 26 | translate: 0.1 # image translation (+/- fraction) 27 | scale: 0.5 # image scale (+/- gain) 28 | shear: 0.0 # image shear (+/- deg) 29 | perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 30 | flipud: 0.0 # image flip up-down (probability) 31 | fliplr: 0.5 # image flip left-right (probability) 32 | mosaic: 1.0 # image mosaic (probability) 33 | mixup: 0.0 # image mixup (probability) 34 | -------------------------------------------------------------------------------- /data/images/bus.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Akbonline/Social-Distancing-using-YOLOv5/44b63c7593759cfb787e5b17deabc08d7e05bd03/data/images/bus.jpg -------------------------------------------------------------------------------- /data/images/zidane.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Akbonline/Social-Distancing-using-YOLOv5/44b63c7593759cfb787e5b17deabc08d7e05bd03/data/images/zidane.jpg -------------------------------------------------------------------------------- /data/scripts/get_coco.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # COCO 2017 dataset http://cocodataset.org 3 | # Download command: bash data/scripts/get_coco.sh 4 | # Train command: python train.py --data coco.yaml 5 | # Default dataset location is next to /yolov5: 6 | # /parent_folder 7 | # /coco 8 | # /yolov5 9 | 10 | # Download/unzip labels 11 | d='../' # unzip directory 12 | url=https://github.com/ultralytics/yolov5/releases/download/v1.0/ 13 | f='coco2017labels.zip' # 68 MB 14 | echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove 15 | 16 | # Download/unzip images 17 | d='../coco/images' # unzip directory 18 | url=http://images.cocodataset.org/zips/ 19 | f1='train2017.zip' # 19G, 118k images 20 | f2='val2017.zip' # 1G, 5k images 21 | f3='test2017.zip' # 7G, 41k images (optional) 22 | for f in $f1 $f2; do 23 | echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove 24 | done 25 | -------------------------------------------------------------------------------- /data/scripts/get_voc.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC/ 3 | # Download command: bash data/scripts/get_voc.sh 4 | # Train command: python train.py --data voc.yaml 5 | # Default dataset location is next to /yolov5: 6 | # /parent_folder 7 | # /VOC 8 | # /yolov5 9 | 10 | start=$(date +%s) 11 | mkdir -p ../tmp 12 | cd ../tmp/ 13 | 14 | # Download/unzip images and labels 15 | d='.' # unzip directory 16 | url=https://github.com/ultralytics/yolov5/releases/download/v1.0/ 17 | f1=VOCtrainval_06-Nov-2007.zip # 446MB, 5012 images 18 | f2=VOCtest_06-Nov-2007.zip # 438MB, 4953 images 19 | f3=VOCtrainval_11-May-2012.zip # 1.95GB, 17126 images 20 | for f in $f1 $f2 $f3; do 21 | echo 'Downloading' $url$f ' ...' && curl -L $url$f -o $f && unzip -q $f -d $d && rm $f # download, unzip, remove 22 | done 23 | 24 | end=$(date +%s) 25 | runtime=$((end - start)) 26 | echo "Completed in" $runtime "seconds" 27 | 28 | echo "Splitting dataset..." 29 | python3 - "$@" <train.txt 89 | cat 2007_train.txt 2007_val.txt 2007_test.txt 2012_train.txt 2012_val.txt >train.all.txt 90 | 91 | python3 - "$@" <= 1 79 | p, s, im0 = path[i], '%g: ' % i, im0s[i].copy() 80 | else: 81 | p, s, im0 = path, '', im0s 82 | 83 | save_path = str(Path(out) / Path(p).name) 84 | s += '%gx%g ' % img.shape[2:] # print string 85 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] #  normalization gain whwh 86 | if det is not None and len(det): 87 | # Rescale boxes from img_size to im0 size 88 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 89 | 90 | # Print results 91 | for c in det[:, -1].unique(): 92 | n = (det[:, -1] == c).sum() # detections per class 93 | s += '%g %ss, ' % (n, names[int(c)]) # add to string 94 | 95 | # Write results 96 | for *xyxy, conf, cls in det: 97 | if save_txt: # Write to file 98 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 99 | with open(save_path[:save_path.rfind('.')] + '.txt', 'a') as file: 100 | file.write(('%g ' * 5 + '\n') % (cls, *xywh)) # label format 101 | 102 | if save_img or view_img: # Add bbox to image 103 | label = '%s %.2f' % (names[int(cls)], conf) 104 | if label is not None: 105 | if (label.split())[0] == 'person': 106 | people_coords.append(xyxy) 107 | # plot_one_box(xyxy, im0, line_thickness=3) 108 | plot_dots_on_people(xyxy, im0) 109 | 110 | # Plot lines connecting people 111 | distancing(people_coords, im0, dist_thres_lim=(200,250)) 112 | 113 | # Print time (inference + NMS) 114 | print('%sDone. (%.3fs)' % (s, t2 - t1)) 115 | 116 | # Stream results 117 | if view_img: 118 | cv2.imshow(p, im0) 119 | if cv2.waitKey(1) == ord('q'): # q to quit 120 | raise StopIteration 121 | 122 | # Save results (image with detections) 123 | if save_img: 124 | if dataset.mode == 'images': 125 | cv2.imwrite(save_path, im0) 126 | else: 127 | if vid_path != save_path: # new video 128 | vid_path = save_path 129 | if isinstance(vid_writer, cv2.VideoWriter): 130 | vid_writer.release() # release previous video writer 131 | 132 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 133 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 134 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 135 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*opt.fourcc), fps, (w, h)) 136 | vid_writer.write(im0) 137 | 138 | if save_txt or save_img: 139 | print('Results saved to %s' % os.getcwd() + os.sep + out) 140 | if platform == 'darwin': # MacOS 141 | os.system('open ' + save_path) 142 | 143 | print('Done. (%.3fs)' % (time.time() - t0)) 144 | 145 | 146 | if __name__ == '__main__': 147 | parser = argparse.ArgumentParser() 148 | parser.add_argument('--weights', type=str, default='weights/yolov5s.pt', help='model.pt path') 149 | parser.add_argument('--source', type=str, default='inference/images', help='source') # file/folder, 0 for webcam 150 | parser.add_argument('--output', type=str, default='inference/output', help='output folder') # output folder 151 | parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)') 152 | parser.add_argument('--conf-thres', type=float, default=0.4, help='object confidence threshold') 153 | parser.add_argument('--iou-thres', type=float, default=0.5, help='IOU threshold for NMS') 154 | parser.add_argument('--fourcc', type=str, default='mp4v', help='output video codec (verify ffmpeg support)') 155 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 156 | parser.add_argument('--view-img', action='store_true', help='display results') 157 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 158 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class') 159 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 160 | parser.add_argument('--augment', action='store_true', help='augmented inference') 161 | opt = parser.parse_args() 162 | opt.img_size = check_img_size(opt.img_size) 163 | print(opt) 164 | 165 | with torch.no_grad(): 166 | detect() 167 | -------------------------------------------------------------------------------- /hubconf.py: -------------------------------------------------------------------------------- 1 | """File for accessing YOLOv5 via PyTorch Hub https://pytorch.org/hub/ 2 | 3 | Usage: 4 | import torch 5 | model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True, channels=3, classes=80) 6 | """ 7 | 8 | dependencies = ['torch', 'yaml'] 9 | 10 | import os 11 | 12 | import torch 13 | 14 | from models.yolo import Model 15 | from utils import google_utils 16 | 17 | 18 | def create(name, pretrained, channels, classes): 19 | """Creates a specified YOLOv5 model 20 | 21 | Arguments: 22 | name (str): name of model, i.e. 'yolov5s' 23 | pretrained (bool): load pretrained weights into the model 24 | channels (int): number of input channels 25 | classes (int): number of model classes 26 | 27 | Returns: 28 | pytorch model 29 | """ 30 | config = os.path.join(os.path.dirname(__file__), 'models', '%s.yaml' % name) # model.yaml path 31 | model = Model(config, channels, classes) 32 | if pretrained: 33 | ckpt = '%s.pt' % name # checkpoint filename 34 | google_utils.attempt_download(ckpt) # download if not found locally 35 | state_dict = torch.load(ckpt, map_location=torch.device('cpu'))['model'].float().state_dict() # to FP32 36 | state_dict = {k: v for k, v in state_dict.items() if model.state_dict()[k].shape == v.shape} # filter 37 | model.load_state_dict(state_dict, strict=False) # load 38 | return model 39 | 40 | 41 | def yolov5s(pretrained=False, channels=3, classes=80): 42 | """YOLOv5-small model from https://github.com/ultralytics/yolov5 43 | 44 | Arguments: 45 | pretrained (bool): load pretrained weights into the model, default=False 46 | channels (int): number of input channels, default=3 47 | classes (int): number of model classes, default=80 48 | 49 | Returns: 50 | pytorch model 51 | """ 52 | return create('yolov5s', pretrained, channels, classes) 53 | 54 | 55 | def yolov5m(pretrained=False, channels=3, classes=80): 56 | """YOLOv5-medium model from https://github.com/ultralytics/yolov5 57 | 58 | Arguments: 59 | pretrained (bool): load pretrained weights into the model, default=False 60 | channels (int): number of input channels, default=3 61 | classes (int): number of model classes, default=80 62 | 63 | Returns: 64 | pytorch model 65 | """ 66 | return create('yolov5m', pretrained, channels, classes) 67 | 68 | 69 | def yolov5l(pretrained=False, channels=3, classes=80): 70 | """YOLOv5-large model from https://github.com/ultralytics/yolov5 71 | 72 | Arguments: 73 | pretrained (bool): load pretrained weights into the model, default=False 74 | channels (int): number of input channels, default=3 75 | classes (int): number of model classes, default=80 76 | 77 | Returns: 78 | pytorch model 79 | """ 80 | return create('yolov5l', pretrained, channels, classes) 81 | 82 | 83 | def yolov5x(pretrained=False, channels=3, classes=80): 84 | """YOLOv5-xlarge model from https://github.com/ultralytics/yolov5 85 | 86 | Arguments: 87 | pretrained (bool): load pretrained weights into the model, default=False 88 | channels (int): number of input channels, default=3 89 | classes (int): number of model classes, default=80 90 | 91 | Returns: 92 | pytorch model 93 | """ 94 | return create('yolov5x', pretrained, channels, classes) 95 | -------------------------------------------------------------------------------- /inference/people1.mp4: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Akbonline/Social-Distancing-using-YOLOv5/44b63c7593759cfb787e5b17deabc08d7e05bd03/inference/people1.mp4 -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # pip install -r requirements.txt 2 | 3 | # base ---------------------------------------- 4 | Cython 5 | matplotlib>=3.2.2 6 | numpy>=1.18.5 7 | opencv-python>=4.1.2 8 | Pillow 9 | PyYAML>=5.3 10 | scipy>=1.4.1 11 | tensorboard>=2.2 12 | torch>=1.6.0 13 | torchvision>=0.7.0 14 | tqdm>=4.41.0 15 | 16 | # logging ------------------------------------- 17 | # wandb 18 | 19 | # plotting ------------------------------------ 20 | seaborn 21 | pandas 22 | 23 | # export -------------------------------------- 24 | # coremltools==4.0 25 | # onnx>=1.8.0 26 | # scikit-learn==0.19.2 # for coreml quantization 27 | 28 | # extras -------------------------------------- 29 | # thop # FLOPS computation 30 | # pycocotools>=2.0 # COCO mAP 31 | -------------------------------------------------------------------------------- /test.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import json 3 | import os 4 | from pathlib import Path 5 | from threading import Thread 6 | 7 | import numpy as np 8 | import torch 9 | import yaml 10 | from tqdm import tqdm 11 | 12 | from models.experimental import attempt_load 13 | from utils.datasets import create_dataloader 14 | from utils.general import coco80_to_coco91_class, check_dataset, check_file, check_img_size, box_iou, \ 15 | non_max_suppression, scale_coords, xyxy2xywh, xywh2xyxy, set_logging, increment_path 16 | from utils.loss import compute_loss 17 | from utils.metrics import ap_per_class, ConfusionMatrix 18 | from utils.plots import plot_images, output_to_target, plot_study_txt 19 | from utils.torch_utils import select_device, time_synchronized 20 | 21 | 22 | def test(data, 23 | weights=None, 24 | batch_size=32, 25 | imgsz=640, 26 | conf_thres=0.001, 27 | iou_thres=0.6, # for NMS 28 | save_json=False, 29 | single_cls=False, 30 | augment=False, 31 | verbose=False, 32 | model=None, 33 | dataloader=None, 34 | save_dir=Path(''), # for saving images 35 | save_txt=False, # for auto-labelling 36 | save_hybrid=False, # for hybrid auto-labelling 37 | save_conf=False, # save auto-label confidences 38 | plots=True, 39 | log_imgs=0): # number of logged images 40 | 41 | # Initialize/load model and set device 42 | training = model is not None 43 | if training: # called by train.py 44 | device = next(model.parameters()).device # get model device 45 | 46 | else: # called directly 47 | set_logging() 48 | device = select_device(opt.device, batch_size=batch_size) 49 | 50 | # Directories 51 | save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run 52 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 53 | 54 | # Load model 55 | model = attempt_load(weights, map_location=device) # load FP32 model 56 | imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size 57 | 58 | # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 59 | # if device.type != 'cpu' and torch.cuda.device_count() > 1: 60 | # model = nn.DataParallel(model) 61 | 62 | # Half 63 | half = device.type != 'cpu' # half precision only supported on CUDA 64 | if half: 65 | model.half() 66 | 67 | # Configure 68 | model.eval() 69 | is_coco = data.endswith('coco.yaml') # is COCO dataset 70 | with open(data) as f: 71 | data = yaml.load(f, Loader=yaml.FullLoader) # model dict 72 | check_dataset(data) # check 73 | nc = 1 if single_cls else int(data['nc']) # number of classes 74 | iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for mAP@0.5:0.95 75 | niou = iouv.numel() 76 | 77 | # Logging 78 | log_imgs, wandb = min(log_imgs, 100), None # ceil 79 | try: 80 | import wandb # Weights & Biases 81 | except ImportError: 82 | log_imgs = 0 83 | 84 | # Dataloader 85 | if not training: 86 | img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img 87 | _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once 88 | path = data['test'] if opt.task == 'test' else data['val'] # path to val/test images 89 | dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True)[0] 90 | 91 | seen = 0 92 | confusion_matrix = ConfusionMatrix(nc=nc) 93 | names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} 94 | coco91class = coco80_to_coco91_class() 95 | s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP@.5', 'mAP@.5:.95') 96 | p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. 97 | loss = torch.zeros(3, device=device) 98 | jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] 99 | for batch_i, (img, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): 100 | img = img.to(device, non_blocking=True) 101 | img = img.half() if half else img.float() # uint8 to fp16/32 102 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 103 | targets = targets.to(device) 104 | nb, _, height, width = img.shape # batch size, channels, height, width 105 | 106 | with torch.no_grad(): 107 | # Run model 108 | t = time_synchronized() 109 | inf_out, train_out = model(img, augment=augment) # inference and training outputs 110 | t0 += time_synchronized() - t 111 | 112 | # Compute loss 113 | if training: 114 | loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # box, obj, cls 115 | 116 | # Run NMS 117 | targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels 118 | lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling 119 | t = time_synchronized() 120 | output = non_max_suppression(inf_out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb) 121 | t1 += time_synchronized() - t 122 | 123 | # Statistics per image 124 | for si, pred in enumerate(output): 125 | labels = targets[targets[:, 0] == si, 1:] 126 | nl = len(labels) 127 | tcls = labels[:, 0].tolist() if nl else [] # target class 128 | path = Path(paths[si]) 129 | seen += 1 130 | 131 | if len(pred) == 0: 132 | if nl: 133 | stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) 134 | continue 135 | 136 | # Predictions 137 | predn = pred.clone() 138 | scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred 139 | 140 | # Append to text file 141 | if save_txt: 142 | gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0]] # normalization gain whwh 143 | for *xyxy, conf, cls in predn.tolist(): 144 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 145 | line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format 146 | with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: 147 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 148 | 149 | # W&B logging 150 | if plots and len(wandb_images) < log_imgs: 151 | box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, 152 | "class_id": int(cls), 153 | "box_caption": "%s %.3f" % (names[cls], conf), 154 | "scores": {"class_score": conf}, 155 | "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] 156 | boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space 157 | wandb_images.append(wandb.Image(img[si], boxes=boxes, caption=path.name)) 158 | 159 | # Append to pycocotools JSON dictionary 160 | if save_json: 161 | # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... 162 | image_id = int(path.stem) if path.stem.isnumeric() else path.stem 163 | box = xyxy2xywh(predn[:, :4]) # xywh 164 | box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner 165 | for p, b in zip(pred.tolist(), box.tolist()): 166 | jdict.append({'image_id': image_id, 167 | 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 168 | 'bbox': [round(x, 3) for x in b], 169 | 'score': round(p[4], 5)}) 170 | 171 | # Assign all predictions as incorrect 172 | correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) 173 | if nl: 174 | detected = [] # target indices 175 | tcls_tensor = labels[:, 0] 176 | 177 | # target boxes 178 | tbox = xywh2xyxy(labels[:, 1:5]) 179 | scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels 180 | if plots: 181 | confusion_matrix.process_batch(pred, torch.cat((labels[:, 0:1], tbox), 1)) 182 | 183 | # Per target class 184 | for cls in torch.unique(tcls_tensor): 185 | ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # prediction indices 186 | pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # target indices 187 | 188 | # Search for detections 189 | if pi.shape[0]: 190 | # Prediction to target ious 191 | ious, i = box_iou(predn[pi, :4], tbox[ti]).max(1) # best ious, indices 192 | 193 | # Append detections 194 | detected_set = set() 195 | for j in (ious > iouv[0]).nonzero(as_tuple=False): 196 | d = ti[i[j]] # detected target 197 | if d.item() not in detected_set: 198 | detected_set.add(d.item()) 199 | detected.append(d) 200 | correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn 201 | if len(detected) == nl: # all targets already located in image 202 | break 203 | 204 | # Append statistics (correct, conf, pcls, tcls) 205 | stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) 206 | 207 | # Plot images 208 | if plots and batch_i < 3: 209 | f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels 210 | Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start() 211 | f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions 212 | Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names), daemon=True).start() 213 | 214 | # Compute statistics 215 | stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy 216 | if len(stats) and stats[0].any(): 217 | p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) 218 | p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1) # [P, R, AP@0.5, AP@0.5:0.95] 219 | mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() 220 | nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class 221 | else: 222 | nt = torch.zeros(1) 223 | 224 | # Print results 225 | pf = '%20s' + '%12.3g' * 6 # print format 226 | print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) 227 | 228 | # Print results per class 229 | if verbose and nc > 1 and len(stats): 230 | for i, c in enumerate(ap_class): 231 | print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) 232 | 233 | # Print speeds 234 | t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple 235 | if not training: 236 | print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) 237 | 238 | # Plots 239 | if plots: 240 | confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) 241 | if wandb and wandb.run: 242 | wandb.log({"Images": wandb_images}) 243 | wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg'))]}) 244 | 245 | # Save JSON 246 | if save_json and len(jdict): 247 | w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights 248 | anno_json = '../coco/annotations/instances_val2017.json' # annotations json 249 | pred_json = str(save_dir / f"{w}_predictions.json") # predictions json 250 | print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) 251 | with open(pred_json, 'w') as f: 252 | json.dump(jdict, f) 253 | 254 | try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb 255 | from pycocotools.coco import COCO 256 | from pycocotools.cocoeval import COCOeval 257 | 258 | anno = COCO(anno_json) # init annotations api 259 | pred = anno.loadRes(pred_json) # init predictions api 260 | eval = COCOeval(anno, pred, 'bbox') 261 | if is_coco: 262 | eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] # image IDs to evaluate 263 | eval.evaluate() 264 | eval.accumulate() 265 | eval.summarize() 266 | map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) 267 | except Exception as e: 268 | print(f'pycocotools unable to run: {e}') 269 | 270 | # Return results 271 | if not training: 272 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 273 | print(f"Results saved to {save_dir}{s}") 274 | model.float() # for training 275 | maps = np.zeros(nc) + map 276 | for i, c in enumerate(ap_class): 277 | maps[c] = ap[i] 278 | return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t 279 | 280 | 281 | if __name__ == '__main__': 282 | parser = argparse.ArgumentParser(prog='test.py') 283 | parser.add_argument('--weights', nargs='+', type=str, default='yolov5s.pt', help='model.pt path(s)') 284 | parser.add_argument('--data', type=str, default='data/coco128.yaml', help='*.data path') 285 | parser.add_argument('--batch-size', type=int, default=32, help='size of each image batch') 286 | parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)') 287 | parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold') 288 | parser.add_argument('--iou-thres', type=float, default=0.6, help='IOU threshold for NMS') 289 | parser.add_argument('--task', default='val', help="'val', 'test', 'study'") 290 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 291 | parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset') 292 | parser.add_argument('--augment', action='store_true', help='augmented inference') 293 | parser.add_argument('--verbose', action='store_true', help='report mAP by class') 294 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 295 | parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt') 296 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 297 | parser.add_argument('--save-json', action='store_true', help='save a cocoapi-compatible JSON results file') 298 | parser.add_argument('--project', default='runs/test', help='save to project/name') 299 | parser.add_argument('--name', default='exp', help='save to project/name') 300 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 301 | opt = parser.parse_args() 302 | opt.save_json |= opt.data.endswith('coco.yaml') 303 | opt.data = check_file(opt.data) # check file 304 | print(opt) 305 | 306 | if opt.task in ['val', 'test']: # run normally 307 | test(opt.data, 308 | opt.weights, 309 | opt.batch_size, 310 | opt.img_size, 311 | opt.conf_thres, 312 | opt.iou_thres, 313 | opt.save_json, 314 | opt.single_cls, 315 | opt.augment, 316 | opt.verbose, 317 | save_txt=opt.save_txt | opt.save_hybrid, 318 | save_hybrid=opt.save_hybrid, 319 | save_conf=opt.save_conf, 320 | ) 321 | 322 | elif opt.task == 'study': # run over a range of settings and save/plot 323 | for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 324 | f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem) # filename to save to 325 | x = list(range(320, 800, 64)) # x axis 326 | y = [] # y axis 327 | for i in x: # img-size 328 | print('\nRunning %s point %s...' % (f, i)) 329 | r, _, t = test(opt.data, weights, opt.batch_size, i, opt.conf_thres, opt.iou_thres, opt.save_json, 330 | plots=False) 331 | y.append(r + t) # results and times 332 | np.savetxt(f, y, fmt='%10.4g') # save 333 | os.system('zip -r study.zip study_*.txt') 334 | plot_study_txt(f, x) # plot 335 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import logging 3 | import os 4 | import random 5 | import time 6 | from pathlib import Path 7 | from threading import Thread 8 | from warnings import warn 9 | 10 | import math 11 | import numpy as np 12 | import torch.distributed as dist 13 | import torch.nn as nn 14 | import torch.nn.functional as F 15 | import torch.optim as optim 16 | import torch.optim.lr_scheduler as lr_scheduler 17 | import torch.utils.data 18 | import yaml 19 | from torch.cuda import amp 20 | from torch.nn.parallel import DistributedDataParallel as DDP 21 | from torch.utils.tensorboard import SummaryWriter 22 | from tqdm import tqdm 23 | 24 | import test # import test.py to get mAP after each epoch 25 | from models.experimental import attempt_load 26 | from models.yolo import Model 27 | from utils.autoanchor import check_anchors 28 | from utils.datasets import create_dataloader 29 | from utils.general import labels_to_class_weights, increment_path, labels_to_image_weights, init_seeds, \ 30 | fitness, strip_optimizer, get_latest_run, check_dataset, check_file, check_git_status, check_img_size, \ 31 | print_mutation, set_logging 32 | from utils.google_utils import attempt_download 33 | from utils.loss import compute_loss 34 | from utils.plots import plot_images, plot_labels, plot_results, plot_evolution 35 | from utils.torch_utils import ModelEMA, select_device, intersect_dicts, torch_distributed_zero_first 36 | 37 | logger = logging.getLogger(__name__) 38 | 39 | try: 40 | import wandb 41 | except ImportError: 42 | wandb = None 43 | logger.info("Install Weights & Biases for experiment logging via 'pip install wandb' (recommended)") 44 | 45 | 46 | def train(hyp, opt, device, tb_writer=None, wandb=None): 47 | logger.info(f'Hyperparameters {hyp}') 48 | save_dir, epochs, batch_size, total_batch_size, weights, rank = \ 49 | Path(opt.save_dir), opt.epochs, opt.batch_size, opt.total_batch_size, opt.weights, opt.global_rank 50 | 51 | # Directories 52 | wdir = save_dir / 'weights' 53 | wdir.mkdir(parents=True, exist_ok=True) # make dir 54 | last = wdir / 'last.pt' 55 | best = wdir / 'best.pt' 56 | results_file = save_dir / 'results.txt' 57 | 58 | # Save run settings 59 | with open(save_dir / 'hyp.yaml', 'w') as f: 60 | yaml.dump(hyp, f, sort_keys=False) 61 | with open(save_dir / 'opt.yaml', 'w') as f: 62 | yaml.dump(vars(opt), f, sort_keys=False) 63 | 64 | # Configure 65 | plots = not opt.evolve # create plots 66 | cuda = device.type != 'cpu' 67 | init_seeds(2 + rank) 68 | with open(opt.data) as f: 69 | data_dict = yaml.load(f, Loader=yaml.FullLoader) # data dict 70 | with torch_distributed_zero_first(rank): 71 | check_dataset(data_dict) # check 72 | train_path = data_dict['train'] 73 | test_path = data_dict['val'] 74 | nc, names = (1, ['item']) if opt.single_cls else (int(data_dict['nc']), data_dict['names']) # number classes, names 75 | assert len(names) == nc, '%g names found for nc=%g dataset in %s' % (len(names), nc, opt.data) # check 76 | 77 | # Model 78 | pretrained = weights.endswith('.pt') 79 | if pretrained: 80 | with torch_distributed_zero_first(rank): 81 | attempt_download(weights) # download if not found locally 82 | ckpt = torch.load(weights, map_location=device) # load checkpoint 83 | if hyp.get('anchors'): 84 | ckpt['model'].yaml['anchors'] = round(hyp['anchors']) # force autoanchor 85 | model = Model(opt.cfg or ckpt['model'].yaml, ch=3, nc=nc).to(device) # create 86 | exclude = ['anchor'] if opt.cfg or hyp.get('anchors') else [] # exclude keys 87 | state_dict = ckpt['model'].float().state_dict() # to FP32 88 | state_dict = intersect_dicts(state_dict, model.state_dict(), exclude=exclude) # intersect 89 | model.load_state_dict(state_dict, strict=False) # load 90 | logger.info('Transferred %g/%g items from %s' % (len(state_dict), len(model.state_dict()), weights)) # report 91 | else: 92 | model = Model(opt.cfg, ch=3, nc=nc).to(device) # create 93 | 94 | # Freeze 95 | freeze = [] # parameter names to freeze (full or partial) 96 | for k, v in model.named_parameters(): 97 | v.requires_grad = True # train all layers 98 | if any(x in k for x in freeze): 99 | print('freezing %s' % k) 100 | v.requires_grad = False 101 | 102 | # Optimizer 103 | nbs = 64 # nominal batch size 104 | accumulate = max(round(nbs / total_batch_size), 1) # accumulate loss before optimizing 105 | hyp['weight_decay'] *= total_batch_size * accumulate / nbs # scale weight_decay 106 | 107 | pg0, pg1, pg2 = [], [], [] # optimizer parameter groups 108 | for k, v in model.named_modules(): 109 | if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): 110 | pg2.append(v.bias) # biases 111 | if isinstance(v, nn.BatchNorm2d): 112 | pg0.append(v.weight) # no decay 113 | elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): 114 | pg1.append(v.weight) # apply decay 115 | 116 | if opt.adam: 117 | optimizer = optim.Adam(pg0, lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum 118 | else: 119 | optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) 120 | 121 | optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay 122 | optimizer.add_param_group({'params': pg2}) # add pg2 (biases) 123 | logger.info('Optimizer groups: %g .bias, %g conv.weight, %g other' % (len(pg2), len(pg1), len(pg0))) 124 | del pg0, pg1, pg2 125 | 126 | # Scheduler https://arxiv.org/pdf/1812.01187.pdf 127 | # https://pytorch.org/docs/stable/_modules/torch/optim/lr_scheduler.html#OneCycleLR 128 | lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - hyp['lrf']) + hyp['lrf'] # cosine 129 | scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) 130 | # plot_lr_scheduler(optimizer, scheduler, epochs) 131 | 132 | # Logging 133 | if wandb and wandb.run is None: 134 | opt.hyp = hyp # add hyperparameters 135 | wandb_run = wandb.init(config=opt, resume="allow", 136 | project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, 137 | name=save_dir.stem, 138 | id=ckpt.get('wandb_id') if 'ckpt' in locals() else None) 139 | loggers = {'wandb': wandb} # loggers dict 140 | 141 | # Resume 142 | start_epoch, best_fitness = 0, 0.0 143 | if pretrained: 144 | # Optimizer 145 | if ckpt['optimizer'] is not None: 146 | optimizer.load_state_dict(ckpt['optimizer']) 147 | best_fitness = ckpt['best_fitness'] 148 | 149 | # Results 150 | if ckpt.get('training_results') is not None: 151 | with open(results_file, 'w') as file: 152 | file.write(ckpt['training_results']) # write results.txt 153 | 154 | # Epochs 155 | start_epoch = ckpt['epoch'] + 1 156 | if opt.resume: 157 | assert start_epoch > 0, '%s training to %g epochs is finished, nothing to resume.' % (weights, epochs) 158 | if epochs < start_epoch: 159 | logger.info('%s has been trained for %g epochs. Fine-tuning for %g additional epochs.' % 160 | (weights, ckpt['epoch'], epochs)) 161 | epochs += ckpt['epoch'] # finetune additional epochs 162 | 163 | del ckpt, state_dict 164 | 165 | # Image sizes 166 | gs = int(max(model.stride)) # grid size (max stride) 167 | imgsz, imgsz_test = [check_img_size(x, gs) for x in opt.img_size] # verify imgsz are gs-multiples 168 | 169 | # DP mode 170 | if cuda and rank == -1 and torch.cuda.device_count() > 1: 171 | model = torch.nn.DataParallel(model) 172 | 173 | # SyncBatchNorm 174 | if opt.sync_bn and cuda and rank != -1: 175 | model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) 176 | logger.info('Using SyncBatchNorm()') 177 | 178 | # EMA 179 | ema = ModelEMA(model) if rank in [-1, 0] else None 180 | 181 | # DDP mode 182 | if cuda and rank != -1: 183 | model = DDP(model, device_ids=[opt.local_rank], output_device=opt.local_rank) 184 | 185 | # Trainloader 186 | dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt, 187 | hyp=hyp, augment=True, cache=opt.cache_images, rect=opt.rect, rank=rank, 188 | world_size=opt.world_size, workers=opt.workers, 189 | image_weights=opt.image_weights) 190 | mlc = np.concatenate(dataset.labels, 0)[:, 0].max() # max label class 191 | nb = len(dataloader) # number of batches 192 | assert mlc < nc, 'Label class %g exceeds nc=%g in %s. Possible class labels are 0-%g' % (mlc, nc, opt.data, nc - 1) 193 | 194 | # Process 0 195 | if rank in [-1, 0]: 196 | ema.updates = start_epoch * nb // accumulate # set EMA updates 197 | testloader = create_dataloader(test_path, imgsz_test, total_batch_size, gs, opt, # testloader 198 | hyp=hyp, cache=opt.cache_images and not opt.notest, rect=True, 199 | rank=-1, world_size=opt.world_size, workers=opt.workers, pad=0.5)[0] 200 | 201 | if not opt.resume: 202 | labels = np.concatenate(dataset.labels, 0) 203 | c = torch.tensor(labels[:, 0]) # classes 204 | # cf = torch.bincount(c.long(), minlength=nc) + 1. # frequency 205 | # model._initialize_biases(cf.to(device)) 206 | if plots: 207 | Thread(target=plot_labels, args=(labels, save_dir, loggers), daemon=True).start() 208 | if tb_writer: 209 | tb_writer.add_histogram('classes', c, 0) 210 | 211 | # Anchors 212 | if not opt.noautoanchor: 213 | check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) 214 | 215 | # Model parameters 216 | hyp['cls'] *= nc / 80. # scale coco-tuned hyp['cls'] to current dataset 217 | model.nc = nc # attach number of classes to model 218 | model.hyp = hyp # attach hyperparameters to model 219 | model.gr = 1.0 # iou loss ratio (obj_loss = 1.0 or iou) 220 | model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights 221 | model.names = names 222 | 223 | # Start training 224 | t0 = time.time() 225 | nw = max(round(hyp['warmup_epochs'] * nb), 1000) # number of warmup iterations, max(3 epochs, 1k iterations) 226 | # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training 227 | maps = np.zeros(nc) # mAP per class 228 | results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) 229 | scheduler.last_epoch = start_epoch - 1 # do not move 230 | scaler = amp.GradScaler(enabled=cuda) 231 | logger.info('Image sizes %g train, %g test\n' 232 | 'Using %g dataloader workers\nLogging results to %s\n' 233 | 'Starting training for %g epochs...' % (imgsz, imgsz_test, dataloader.num_workers, save_dir, epochs)) 234 | for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ 235 | model.train() 236 | 237 | # Update image weights (optional) 238 | if opt.image_weights: 239 | # Generate indices 240 | if rank in [-1, 0]: 241 | cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights 242 | iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights 243 | dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx 244 | # Broadcast if DDP 245 | if rank != -1: 246 | indices = (torch.tensor(dataset.indices) if rank == 0 else torch.zeros(dataset.n)).int() 247 | dist.broadcast(indices, 0) 248 | if rank != 0: 249 | dataset.indices = indices.cpu().numpy() 250 | 251 | # Update mosaic border 252 | # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) 253 | # dataset.mosaic_border = [b - imgsz, -b] # height, width borders 254 | 255 | mloss = torch.zeros(4, device=device) # mean losses 256 | if rank != -1: 257 | dataloader.sampler.set_epoch(epoch) 258 | pbar = enumerate(dataloader) 259 | logger.info(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'total', 'targets', 'img_size')) 260 | if rank in [-1, 0]: 261 | pbar = tqdm(pbar, total=nb) # progress bar 262 | optimizer.zero_grad() 263 | for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- 264 | ni = i + nb * epoch # number integrated batches (since train start) 265 | imgs = imgs.to(device, non_blocking=True).float() / 255.0 # uint8 to float32, 0-255 to 0.0-1.0 266 | 267 | # Warmup 268 | if ni <= nw: 269 | xi = [0, nw] # x interp 270 | # model.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) 271 | accumulate = max(1, np.interp(ni, xi, [1, nbs / total_batch_size]).round()) 272 | for j, x in enumerate(optimizer.param_groups): 273 | # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 274 | x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)]) 275 | if 'momentum' in x: 276 | x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) 277 | 278 | # Multi-scale 279 | if opt.multi_scale: 280 | sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size 281 | sf = sz / max(imgs.shape[2:]) # scale factor 282 | if sf != 1: 283 | ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple) 284 | imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) 285 | 286 | # Forward 287 | with amp.autocast(enabled=cuda): 288 | pred = model(imgs) # forward 289 | loss, loss_items = compute_loss(pred, targets.to(device), model) # loss scaled by batch_size 290 | if rank != -1: 291 | loss *= opt.world_size # gradient averaged between devices in DDP mode 292 | 293 | # Backward 294 | scaler.scale(loss).backward() 295 | 296 | # Optimize 297 | if ni % accumulate == 0: 298 | scaler.step(optimizer) # optimizer.step 299 | scaler.update() 300 | optimizer.zero_grad() 301 | if ema: 302 | ema.update(model) 303 | 304 | # Print 305 | if rank in [-1, 0]: 306 | mloss = (mloss * i + loss_items) / (i + 1) # update mean losses 307 | mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) 308 | s = ('%10s' * 2 + '%10.4g' * 6) % ( 309 | '%g/%g' % (epoch, epochs - 1), mem, *mloss, targets.shape[0], imgs.shape[-1]) 310 | pbar.set_description(s) 311 | 312 | # Plot 313 | if plots and ni < 3: 314 | f = save_dir / f'train_batch{ni}.jpg' # filename 315 | Thread(target=plot_images, args=(imgs, targets, paths, f), daemon=True).start() 316 | # if tb_writer: 317 | # tb_writer.add_image(f, result, dataformats='HWC', global_step=epoch) 318 | # tb_writer.add_graph(model, imgs) # add model to tensorboard 319 | elif plots and ni == 3 and wandb: 320 | wandb.log({"Mosaics": [wandb.Image(str(x), caption=x.name) for x in save_dir.glob('train*.jpg')]}) 321 | 322 | # end batch ------------------------------------------------------------------------------------------------ 323 | # end epoch ---------------------------------------------------------------------------------------------------- 324 | 325 | # Scheduler 326 | lr = [x['lr'] for x in optimizer.param_groups] # for tensorboard 327 | scheduler.step() 328 | 329 | # DDP process 0 or single-GPU 330 | if rank in [-1, 0]: 331 | # mAP 332 | if ema: 333 | ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'gr', 'names', 'stride']) 334 | final_epoch = epoch + 1 == epochs 335 | if not opt.notest or final_epoch: # Calculate mAP 336 | results, maps, times = test.test(opt.data, 337 | batch_size=total_batch_size, 338 | imgsz=imgsz_test, 339 | model=ema.ema, 340 | single_cls=opt.single_cls, 341 | dataloader=testloader, 342 | save_dir=save_dir, 343 | plots=plots and final_epoch, 344 | log_imgs=opt.log_imgs if wandb else 0) 345 | 346 | # Write 347 | with open(results_file, 'a') as f: 348 | f.write(s + '%10.4g' * 7 % results + '\n') # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) 349 | if len(opt.name) and opt.bucket: 350 | os.system('gsutil cp %s gs://%s/results/results%s.txt' % (results_file, opt.bucket, opt.name)) 351 | 352 | # Log 353 | tags = ['train/box_loss', 'train/obj_loss', 'train/cls_loss', # train loss 354 | 'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 355 | 'val/box_loss', 'val/obj_loss', 'val/cls_loss', # val loss 356 | 'x/lr0', 'x/lr1', 'x/lr2'] # params 357 | for x, tag in zip(list(mloss[:-1]) + list(results) + lr, tags): 358 | if tb_writer: 359 | tb_writer.add_scalar(tag, x, epoch) # tensorboard 360 | if wandb: 361 | wandb.log({tag: x}) # W&B 362 | 363 | # Update best mAP 364 | fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95] 365 | if fi > best_fitness: 366 | best_fitness = fi 367 | 368 | # Save model 369 | save = (not opt.nosave) or (final_epoch and not opt.evolve) 370 | if save: 371 | with open(results_file, 'r') as f: # create checkpoint 372 | ckpt = {'epoch': epoch, 373 | 'best_fitness': best_fitness, 374 | 'training_results': f.read(), 375 | 'model': ema.ema, 376 | 'optimizer': None if final_epoch else optimizer.state_dict(), 377 | 'wandb_id': wandb_run.id if wandb else None} 378 | 379 | # Save last, best and delete 380 | torch.save(ckpt, last) 381 | if best_fitness == fi: 382 | torch.save(ckpt, best) 383 | del ckpt 384 | # end epoch ---------------------------------------------------------------------------------------------------- 385 | # end training 386 | 387 | if rank in [-1, 0]: 388 | # Strip optimizers 389 | for f in [last, best]: 390 | if f.exists(): # is *.pt 391 | strip_optimizer(f) # strip optimizer 392 | os.system('gsutil cp %s gs://%s/weights' % (f, opt.bucket)) if opt.bucket else None # upload 393 | 394 | # Plots 395 | if plots: 396 | plot_results(save_dir=save_dir) # save as results.png 397 | if wandb: 398 | files = ['results.png', 'precision_recall_curve.png', 'confusion_matrix.png'] 399 | wandb.log({"Results": [wandb.Image(str(save_dir / f), caption=f) for f in files 400 | if (save_dir / f).exists()]}) 401 | logger.info('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600)) 402 | 403 | # Test best.pt 404 | if opt.data.endswith('coco.yaml') and nc == 80: # if COCO 405 | results, _, _ = test.test(opt.data, 406 | batch_size=total_batch_size, 407 | imgsz=imgsz_test, 408 | model=attempt_load(best if best.exists() else last, device).half(), 409 | single_cls=opt.single_cls, 410 | dataloader=testloader, 411 | save_dir=save_dir, 412 | save_json=True, # use pycocotools 413 | plots=False) 414 | 415 | else: 416 | dist.destroy_process_group() 417 | 418 | wandb.run.finish() if wandb and wandb.run else None 419 | torch.cuda.empty_cache() 420 | return results 421 | 422 | 423 | if __name__ == '__main__': 424 | parser = argparse.ArgumentParser() 425 | parser.add_argument('--weights', type=str, default='yolov5s.pt', help='initial weights path') 426 | parser.add_argument('--cfg', type=str, default='', help='model.yaml path') 427 | parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') 428 | parser.add_argument('--hyp', type=str, default='data/hyp.scratch.yaml', help='hyperparameters path') 429 | parser.add_argument('--epochs', type=int, default=300) 430 | parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs') 431 | parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='[train, test] image sizes') 432 | parser.add_argument('--rect', action='store_true', help='rectangular training') 433 | parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training') 434 | parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') 435 | parser.add_argument('--notest', action='store_true', help='only test final epoch') 436 | parser.add_argument('--noautoanchor', action='store_true', help='disable autoanchor check') 437 | parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters') 438 | parser.add_argument('--bucket', type=str, default='', help='gsutil bucket') 439 | parser.add_argument('--cache-images', action='store_true', help='cache images for faster training') 440 | parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training') 441 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 442 | parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%') 443 | parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') 444 | parser.add_argument('--adam', action='store_true', help='use torch.optim.Adam() optimizer') 445 | parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode') 446 | parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify') 447 | parser.add_argument('--log-imgs', type=int, default=16, help='number of images for W&B logging, max 100') 448 | parser.add_argument('--workers', type=int, default=8, help='maximum number of dataloader workers') 449 | parser.add_argument('--project', default='runs/train', help='save to project/name') 450 | parser.add_argument('--name', default='exp', help='save to project/name') 451 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 452 | opt = parser.parse_args() 453 | 454 | # Set DDP variables 455 | opt.total_batch_size = opt.batch_size 456 | opt.world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1 457 | opt.global_rank = int(os.environ['RANK']) if 'RANK' in os.environ else -1 458 | set_logging(opt.global_rank) 459 | if opt.global_rank in [-1, 0]: 460 | check_git_status() 461 | 462 | # Resume 463 | if opt.resume: # resume an interrupted run 464 | ckpt = opt.resume if isinstance(opt.resume, str) else get_latest_run() # specified or most recent path 465 | assert os.path.isfile(ckpt), 'ERROR: --resume checkpoint does not exist' 466 | with open(Path(ckpt).parent.parent / 'opt.yaml') as f: 467 | opt = argparse.Namespace(**yaml.load(f, Loader=yaml.FullLoader)) # replace 468 | opt.cfg, opt.weights, opt.resume = '', ckpt, True 469 | logger.info('Resuming training from %s' % ckpt) 470 | else: 471 | # opt.hyp = opt.hyp or ('hyp.finetune.yaml' if opt.weights else 'hyp.scratch.yaml') 472 | opt.data, opt.cfg, opt.hyp = check_file(opt.data), check_file(opt.cfg), check_file(opt.hyp) # check files 473 | assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified' 474 | opt.img_size.extend([opt.img_size[-1]] * (2 - len(opt.img_size))) # extend to 2 sizes (train, test) 475 | opt.name = 'evolve' if opt.evolve else opt.name 476 | opt.save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok | opt.evolve) # increment run 477 | 478 | # DDP mode 479 | device = select_device(opt.device, batch_size=opt.batch_size) 480 | if opt.local_rank != -1: 481 | assert torch.cuda.device_count() > opt.local_rank 482 | torch.cuda.set_device(opt.local_rank) 483 | device = torch.device('cuda', opt.local_rank) 484 | dist.init_process_group(backend='nccl', init_method='env://') # distributed backend 485 | assert opt.batch_size % opt.world_size == 0, '--batch-size must be multiple of CUDA device count' 486 | opt.batch_size = opt.total_batch_size // opt.world_size 487 | 488 | # Hyperparameters 489 | with open(opt.hyp) as f: 490 | hyp = yaml.load(f, Loader=yaml.FullLoader) # load hyps 491 | if 'box' not in hyp: 492 | warn('Compatibility: %s missing "box" which was renamed from "giou" in %s' % 493 | (opt.hyp, 'https://github.com/ultralytics/yolov5/pull/1120')) 494 | hyp['box'] = hyp.pop('giou') 495 | 496 | # Train 497 | logger.info(opt) 498 | if not opt.evolve: 499 | tb_writer = None # init loggers 500 | if opt.global_rank in [-1, 0]: 501 | logger.info(f'Start Tensorboard with "tensorboard --logdir {opt.project}", view at http://localhost:6006/') 502 | tb_writer = SummaryWriter(opt.save_dir) # Tensorboard 503 | train(hyp, opt, device, tb_writer, wandb) 504 | 505 | # Evolve hyperparameters (optional) 506 | else: 507 | # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit) 508 | meta = {'lr0': (1, 1e-5, 1e-1), # initial learning rate (SGD=1E-2, Adam=1E-3) 509 | 'lrf': (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf) 510 | 'momentum': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1 511 | 'weight_decay': (1, 0.0, 0.001), # optimizer weight decay 512 | 'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok) 513 | 'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum 514 | 'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr 515 | 'box': (1, 0.02, 0.2), # box loss gain 516 | 'cls': (1, 0.2, 4.0), # cls loss gain 517 | 'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight 518 | 'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels) 519 | 'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight 520 | 'iou_t': (0, 0.1, 0.7), # IoU training threshold 521 | 'anchor_t': (1, 2.0, 8.0), # anchor-multiple threshold 522 | 'anchors': (2, 2.0, 10.0), # anchors per output grid (0 to ignore) 523 | 'fl_gamma': (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma=1.5) 524 | 'hsv_h': (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction) 525 | 'hsv_s': (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction) 526 | 'hsv_v': (1, 0.0, 0.9), # image HSV-Value augmentation (fraction) 527 | 'degrees': (1, 0.0, 45.0), # image rotation (+/- deg) 528 | 'translate': (1, 0.0, 0.9), # image translation (+/- fraction) 529 | 'scale': (1, 0.0, 0.9), # image scale (+/- gain) 530 | 'shear': (1, 0.0, 10.0), # image shear (+/- deg) 531 | 'perspective': (0, 0.0, 0.001), # image perspective (+/- fraction), range 0-0.001 532 | 'flipud': (1, 0.0, 1.0), # image flip up-down (probability) 533 | 'fliplr': (0, 0.0, 1.0), # image flip left-right (probability) 534 | 'mosaic': (1, 0.0, 1.0), # image mixup (probability) 535 | 'mixup': (1, 0.0, 1.0)} # image mixup (probability) 536 | 537 | assert opt.local_rank == -1, 'DDP mode not implemented for --evolve' 538 | opt.notest, opt.nosave = True, True # only test/save final epoch 539 | # ei = [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indices 540 | yaml_file = Path(opt.save_dir) / 'hyp_evolved.yaml' # save best result here 541 | if opt.bucket: 542 | os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt if exists 543 | 544 | for _ in range(300): # generations to evolve 545 | if Path('evolve.txt').exists(): # if evolve.txt exists: select best hyps and mutate 546 | # Select parent(s) 547 | parent = 'single' # parent selection method: 'single' or 'weighted' 548 | x = np.loadtxt('evolve.txt', ndmin=2) 549 | n = min(5, len(x)) # number of previous results to consider 550 | x = x[np.argsort(-fitness(x))][:n] # top n mutations 551 | w = fitness(x) - fitness(x).min() # weights 552 | if parent == 'single' or len(x) == 1: 553 | # x = x[random.randint(0, n - 1)] # random selection 554 | x = x[random.choices(range(n), weights=w)[0]] # weighted selection 555 | elif parent == 'weighted': 556 | x = (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination 557 | 558 | # Mutate 559 | mp, s = 0.8, 0.2 # mutation probability, sigma 560 | npr = np.random 561 | npr.seed(int(time.time())) 562 | g = np.array([x[0] for x in meta.values()]) # gains 0-1 563 | ng = len(meta) 564 | v = np.ones(ng) 565 | while all(v == 1): # mutate until a change occurs (prevent duplicates) 566 | v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0) 567 | for i, k in enumerate(hyp.keys()): # plt.hist(v.ravel(), 300) 568 | hyp[k] = float(x[i + 7] * v[i]) # mutate 569 | 570 | # Constrain to limits 571 | for k, v in meta.items(): 572 | hyp[k] = max(hyp[k], v[1]) # lower limit 573 | hyp[k] = min(hyp[k], v[2]) # upper limit 574 | hyp[k] = round(hyp[k], 5) # significant digits 575 | 576 | # Train mutation 577 | results = train(hyp.copy(), opt, device, wandb=wandb) 578 | 579 | # Write mutation results 580 | print_mutation(hyp.copy(), results, yaml_file, opt.bucket) 581 | 582 | # Plot results 583 | plot_evolution(yaml_file) 584 | print(f'Hyperparameter evolution complete. Best results saved as: {yaml_file}\n' 585 | f'Command to train a new model with these hyperparameters: $ python train.py --hyp {yaml_file}') 586 | -------------------------------------------------------------------------------- /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 | # Swish https://arxiv.org/pdf/1905.02244.pdf --------------------------------------------------------------------------- 9 | class Swish(nn.Module): # 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) 69 | self.bn = nn.BatchNorm2d(c1) 70 | 71 | def forward(self, x): 72 | return torch.max(x, self.bn(self.conv(x))) 73 | -------------------------------------------------------------------------------- /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 | 10 | def check_anchor_order(m): 11 | # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary 12 | a = m.anchor_grid.prod(-1).view(-1) # anchor area 13 | da = a[-1] - a[0] # delta a 14 | ds = m.stride[-1] - m.stride[0] # delta s 15 | if da.sign() != ds.sign(): # same order 16 | print('Reversing anchor order') 17 | m.anchors[:] = m.anchors.flip(0) 18 | m.anchor_grid[:] = m.anchor_grid.flip(0) 19 | 20 | 21 | def check_anchors(dataset, model, thr=4.0, imgsz=640): 22 | # Check anchor fit to data, recompute if necessary 23 | print('\nAnalyzing anchors... ', end='') 24 | m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() 25 | shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) 26 | scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale 27 | wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh 28 | 29 | def metric(k): # compute metric 30 | r = wh[:, None] / k[None] 31 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 32 | best = x.max(1)[0] # best_x 33 | aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold 34 | bpr = (best > 1. / thr).float().mean() # best possible recall 35 | return bpr, aat 36 | 37 | bpr, aat = metric(m.anchor_grid.clone().cpu().view(-1, 2)) 38 | print('anchors/target = %.2f, Best Possible Recall (BPR) = %.4f' % (aat, bpr), end='') 39 | if bpr < 0.98: # threshold to recompute 40 | print('. Attempting to improve anchors, please wait...') 41 | na = m.anchor_grid.numel() // 2 # number of anchors 42 | new_anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) 43 | new_bpr = metric(new_anchors.reshape(-1, 2))[0] 44 | if new_bpr > bpr: # replace anchors 45 | new_anchors = torch.tensor(new_anchors, device=m.anchors.device).type_as(m.anchors) 46 | m.anchor_grid[:] = new_anchors.clone().view_as(m.anchor_grid) # for inference 47 | m.anchors[:] = new_anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss 48 | check_anchor_order(m) 49 | print('New anchors saved to model. Update model *.yaml to use these anchors in the future.') 50 | else: 51 | print('Original anchors better than new anchors. Proceeding with original anchors.') 52 | print('') # newline 53 | 54 | 55 | def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): 56 | """ Creates kmeans-evolved anchors from training dataset 57 | 58 | Arguments: 59 | path: path to dataset *.yaml, or a loaded dataset 60 | n: number of anchors 61 | img_size: image size used for training 62 | thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 63 | gen: generations to evolve anchors using genetic algorithm 64 | verbose: print all results 65 | 66 | Return: 67 | k: kmeans evolved anchors 68 | 69 | Usage: 70 | from utils.autoanchor import *; _ = kmean_anchors() 71 | """ 72 | thr = 1. / thr 73 | 74 | def metric(k, wh): # compute metrics 75 | r = wh[:, None] / k[None] 76 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 77 | # x = wh_iou(wh, torch.tensor(k)) # iou metric 78 | return x, x.max(1)[0] # x, best_x 79 | 80 | def anchor_fitness(k): # mutation fitness 81 | _, best = metric(torch.tensor(k, dtype=torch.float32), wh) 82 | return (best * (best > thr).float()).mean() # fitness 83 | 84 | def print_results(k): 85 | k = k[np.argsort(k.prod(1))] # sort small to large 86 | x, best = metric(k, wh0) 87 | bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr 88 | print('thr=%.2f: %.4f best possible recall, %.2f anchors past thr' % (thr, bpr, aat)) 89 | print('n=%g, img_size=%s, metric_all=%.3f/%.3f-mean/best, past_thr=%.3f-mean: ' % 90 | (n, img_size, x.mean(), best.mean(), x[x > thr].mean()), end='') 91 | for i, x in enumerate(k): 92 | print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg 93 | return k 94 | 95 | if isinstance(path, str): # *.yaml file 96 | with open(path) as f: 97 | data_dict = yaml.load(f, Loader=yaml.FullLoader) # model dict 98 | from utils.datasets import LoadImagesAndLabels 99 | dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) 100 | else: 101 | dataset = path # dataset 102 | 103 | # Get label wh 104 | shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) 105 | wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh 106 | 107 | # Filter 108 | i = (wh0 < 3.0).any(1).sum() 109 | if i: 110 | print('WARNING: Extremely small objects found. ' 111 | '%g of %g labels are < 3 pixels in width or height.' % (i, len(wh0))) 112 | wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels 113 | 114 | # Kmeans calculation 115 | print('Running kmeans for %g anchors on %g points...' % (n, len(wh))) 116 | s = wh.std(0) # sigmas for whitening 117 | k, dist = kmeans(wh / s, n, iter=30) # points, mean distance 118 | k *= s 119 | wh = torch.tensor(wh, dtype=torch.float32) # filtered 120 | wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered 121 | k = print_results(k) 122 | 123 | # Plot 124 | # k, d = [None] * 20, [None] * 20 125 | # for i in tqdm(range(1, 21)): 126 | # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance 127 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) 128 | # ax = ax.ravel() 129 | # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') 130 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh 131 | # ax[0].hist(wh[wh[:, 0]<100, 0],400) 132 | # ax[1].hist(wh[wh[:, 1]<100, 1],400) 133 | # fig.savefig('wh.png', dpi=200) 134 | 135 | # Evolve 136 | npr = np.random 137 | f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma 138 | pbar = tqdm(range(gen), desc='Evolving anchors with Genetic Algorithm') # progress bar 139 | for _ in pbar: 140 | v = np.ones(sh) 141 | while (v == 1).all(): # mutate until a change occurs (prevent duplicates) 142 | v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) 143 | kg = (k.copy() * v).clip(min=2.0) 144 | fg = anchor_fitness(kg) 145 | if fg > f: 146 | f, k = fg, kg.copy() 147 | pbar.desc = 'Evolving anchors with Genetic Algorithm: fitness = %.4f' % f 148 | if verbose: 149 | print_results(k) 150 | 151 | return print_results(k) 152 | -------------------------------------------------------------------------------- /utils/general.py: -------------------------------------------------------------------------------- 1 | # General utils 2 | 3 | import glob 4 | import logging 5 | import os 6 | import platform 7 | import random 8 | import re 9 | import subprocess 10 | import time 11 | from pathlib import Path 12 | 13 | import cv2 14 | import math 15 | import numpy as np 16 | import torch 17 | import torchvision 18 | import yaml 19 | 20 | from utils.google_utils import gsutil_getsize 21 | from utils.metrics import fitness 22 | from utils.torch_utils import init_torch_seeds 23 | 24 | # Settings 25 | torch.set_printoptions(linewidth=320, precision=5, profile='long') 26 | np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 27 | cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) 28 | 29 | 30 | def set_logging(rank=-1): 31 | logging.basicConfig( 32 | format="%(message)s", 33 | level=logging.INFO if rank in [-1, 0] else logging.WARN) 34 | 35 | 36 | def init_seeds(seed=0): 37 | random.seed(seed) 38 | np.random.seed(seed) 39 | init_torch_seeds(seed) 40 | 41 | 42 | def get_latest_run(search_dir='.'): 43 | # Return path to most recent 'last.pt' in /runs (i.e. to --resume from) 44 | last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True) 45 | return max(last_list, key=os.path.getctime) if last_list else '' 46 | 47 | 48 | def check_git_status(): 49 | # Suggest 'git pull' if repo is out of date 50 | if platform.system() in ['Linux', 'Darwin'] and not os.path.isfile('/.dockerenv'): 51 | s = subprocess.check_output('if [ -d .git ]; then git fetch && git status -uno; fi', shell=True).decode('utf-8') 52 | if 'Your branch is behind' in s: 53 | print(s[s.find('Your branch is behind'):s.find('\n\n')] + '\n') 54 | 55 | 56 | def check_img_size(img_size, s=32): 57 | # Verify img_size is a multiple of stride s 58 | new_size = make_divisible(img_size, int(s)) # ceil gs-multiple 59 | if new_size != img_size: 60 | print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size)) 61 | return new_size 62 | 63 | 64 | def check_file(file): 65 | # Search for file if not found 66 | if os.path.isfile(file) or file == '': 67 | return file 68 | else: 69 | files = glob.glob('./**/' + file, recursive=True) # find file 70 | assert len(files), 'File Not Found: %s' % file # assert file was found 71 | assert len(files) == 1, "Multiple files match '%s', specify exact path: %s" % (file, files) # assert unique 72 | return files[0] # return file 73 | 74 | 75 | def check_dataset(dict): 76 | # Download dataset if not found locally 77 | val, s = dict.get('val'), dict.get('download') 78 | if val and len(val): 79 | val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path 80 | if not all(x.exists() for x in val): 81 | print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()]) 82 | if s and len(s): # download script 83 | print('Downloading %s ...' % s) 84 | if s.startswith('http') and s.endswith('.zip'): # URL 85 | f = Path(s).name # filename 86 | torch.hub.download_url_to_file(s, f) 87 | r = os.system('unzip -q %s -d ../ && rm %s' % (f, f)) # unzip 88 | else: # bash script 89 | r = os.system(s) 90 | print('Dataset autodownload %s\n' % ('success' if r == 0 else 'failure')) # analyze return value 91 | else: 92 | raise Exception('Dataset not found.') 93 | 94 | 95 | def make_divisible(x, divisor): 96 | # Returns x evenly divisible by divisor 97 | return math.ceil(x / divisor) * divisor 98 | 99 | 100 | def labels_to_class_weights(labels, nc=80): 101 | # Get class weights (inverse frequency) from training labels 102 | if labels[0] is None: # no labels loaded 103 | return torch.Tensor() 104 | 105 | labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO 106 | classes = labels[:, 0].astype(np.int) # labels = [class xywh] 107 | weights = np.bincount(classes, minlength=nc) # occurrences per class 108 | 109 | # Prepend gridpoint count (for uCE training) 110 | # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image 111 | # weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start 112 | 113 | weights[weights == 0] = 1 # replace empty bins with 1 114 | weights = 1 / weights # number of targets per class 115 | weights /= weights.sum() # normalize 116 | return torch.from_numpy(weights) 117 | 118 | 119 | def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)): 120 | # Produces image weights based on class_weights and image contents 121 | class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels]) 122 | image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1) 123 | # index = random.choices(range(n), weights=image_weights, k=1) # weight image sample 124 | return image_weights 125 | 126 | 127 | def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) 128 | # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ 129 | # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n') 130 | # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n') 131 | # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco 132 | # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet 133 | x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 134 | 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 135 | 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] 136 | return x 137 | 138 | 139 | def xyxy2xywh(x): 140 | # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right 141 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 142 | y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center 143 | y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center 144 | y[:, 2] = x[:, 2] - x[:, 0] # width 145 | y[:, 3] = x[:, 3] - x[:, 1] # height 146 | return y 147 | 148 | 149 | def xywh2xyxy(x): 150 | # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right 151 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 152 | y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x 153 | y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y 154 | y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x 155 | y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y 156 | return y 157 | 158 | 159 | def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None): 160 | # Rescale coords (xyxy) from img1_shape to img0_shape 161 | if ratio_pad is None: # calculate from img0_shape 162 | gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new 163 | pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding 164 | else: 165 | gain = ratio_pad[0][0] 166 | pad = ratio_pad[1] 167 | 168 | coords[:, [0, 2]] -= pad[0] # x padding 169 | coords[:, [1, 3]] -= pad[1] # y padding 170 | coords[:, :4] /= gain 171 | clip_coords(coords, img0_shape) 172 | return coords 173 | 174 | 175 | def clip_coords(boxes, img_shape): 176 | # Clip bounding xyxy bounding boxes to image shape (height, width) 177 | boxes[:, 0].clamp_(0, img_shape[1]) # x1 178 | boxes[:, 1].clamp_(0, img_shape[0]) # y1 179 | boxes[:, 2].clamp_(0, img_shape[1]) # x2 180 | boxes[:, 3].clamp_(0, img_shape[0]) # y2 181 | 182 | 183 | def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-9): 184 | # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4 185 | box2 = box2.T 186 | 187 | # Get the coordinates of bounding boxes 188 | if x1y1x2y2: # x1, y1, x2, y2 = box1 189 | b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3] 190 | b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3] 191 | else: # transform from xywh to xyxy 192 | b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2 193 | b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2 194 | b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2 195 | b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2 196 | 197 | # Intersection area 198 | inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ 199 | (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) 200 | 201 | # Union Area 202 | w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps 203 | w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps 204 | union = w1 * h1 + w2 * h2 - inter + eps 205 | 206 | iou = inter / union 207 | if GIoU or DIoU or CIoU: 208 | cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width 209 | ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height 210 | if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1 211 | c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared 212 | rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + 213 | (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared 214 | if DIoU: 215 | return iou - rho2 / c2 # DIoU 216 | elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47 217 | v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2) 218 | with torch.no_grad(): 219 | alpha = v / ((1 + eps) - iou + v) 220 | return iou - (rho2 / c2 + v * alpha) # CIoU 221 | else: # GIoU https://arxiv.org/pdf/1902.09630.pdf 222 | c_area = cw * ch + eps # convex area 223 | return iou - (c_area - union) / c_area # GIoU 224 | else: 225 | return iou # IoU 226 | 227 | 228 | def box_iou(box1, box2): 229 | # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py 230 | """ 231 | Return intersection-over-union (Jaccard index) of boxes. 232 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 233 | Arguments: 234 | box1 (Tensor[N, 4]) 235 | box2 (Tensor[M, 4]) 236 | Returns: 237 | iou (Tensor[N, M]): the NxM matrix containing the pairwise 238 | IoU values for every element in boxes1 and boxes2 239 | """ 240 | 241 | def box_area(box): 242 | # box = 4xn 243 | return (box[2] - box[0]) * (box[3] - box[1]) 244 | 245 | area1 = box_area(box1.T) 246 | area2 = box_area(box2.T) 247 | 248 | # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) 249 | inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2) 250 | return inter / (area1[:, None] + area2 - inter) # iou = inter / (area1 + area2 - inter) 251 | 252 | 253 | def wh_iou(wh1, wh2): 254 | # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2 255 | wh1 = wh1[:, None] # [N,1,2] 256 | wh2 = wh2[None] # [1,M,2] 257 | inter = torch.min(wh1, wh2).prod(2) # [N,M] 258 | return inter / (wh1.prod(2) + wh2.prod(2) - inter) # iou = inter / (area1 + area2 - inter) 259 | 260 | 261 | def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, classes=None, agnostic=False, labels=()): 262 | """Performs Non-Maximum Suppression (NMS) on inference results 263 | 264 | Returns: 265 | detections with shape: nx6 (x1, y1, x2, y2, conf, cls) 266 | """ 267 | 268 | nc = prediction.shape[2] - 5 # number of classes 269 | xc = prediction[..., 4] > conf_thres # candidates 270 | 271 | # Settings 272 | min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height 273 | max_det = 300 # maximum number of detections per image 274 | time_limit = 10.0 # seconds to quit after 275 | redundant = True # require redundant detections 276 | multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img) 277 | merge = False # use merge-NMS 278 | 279 | t = time.time() 280 | output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0] 281 | for xi, x in enumerate(prediction): # image index, image inference 282 | # Apply constraints 283 | # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height 284 | x = x[xc[xi]] # confidence 285 | 286 | # Cat apriori labels if autolabelling 287 | if labels and len(labels[xi]): 288 | l = labels[xi] 289 | v = torch.zeros((len(l), nc + 5), device=x.device) 290 | v[:, :4] = l[:, 1:5] # box 291 | v[:, 4] = 1.0 # conf 292 | v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls 293 | x = torch.cat((x, v), 0) 294 | 295 | # If none remain process next image 296 | if not x.shape[0]: 297 | continue 298 | 299 | # Compute conf 300 | x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf 301 | 302 | # Box (center x, center y, width, height) to (x1, y1, x2, y2) 303 | box = xywh2xyxy(x[:, :4]) 304 | 305 | # Detections matrix nx6 (xyxy, conf, cls) 306 | if multi_label: 307 | i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T 308 | x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) 309 | else: # best class only 310 | conf, j = x[:, 5:].max(1, keepdim=True) 311 | x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] 312 | 313 | # Filter by class 314 | if classes: 315 | x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] 316 | 317 | # Apply finite constraint 318 | # if not torch.isfinite(x).all(): 319 | # x = x[torch.isfinite(x).all(1)] 320 | 321 | # If none remain process next image 322 | n = x.shape[0] # number of boxes 323 | if not n: 324 | continue 325 | 326 | # Sort by confidence 327 | # x = x[x[:, 4].argsort(descending=True)] 328 | 329 | # Batched NMS 330 | c = x[:, 5:6] * (0 if agnostic else max_wh) # classes 331 | boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores 332 | i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS 333 | if i.shape[0] > max_det: # limit detections 334 | i = i[:max_det] 335 | if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) 336 | # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) 337 | iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix 338 | weights = iou * scores[None] # box weights 339 | x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes 340 | if redundant: 341 | i = i[iou.sum(1) > 1] # require redundancy 342 | 343 | output[xi] = x[i] 344 | if (time.time() - t) > time_limit: 345 | break # time limit exceeded 346 | 347 | return output 348 | 349 | 350 | def strip_optimizer(f='weights/best.pt', s=''): # from utils.general import *; strip_optimizer() 351 | # Strip optimizer from 'f' to finalize training, optionally save as 's' 352 | x = torch.load(f, map_location=torch.device('cpu')) 353 | x['optimizer'] = None 354 | x['training_results'] = None 355 | x['epoch'] = -1 356 | x['model'].half() # to FP16 357 | for p in x['model'].parameters(): 358 | p.requires_grad = False 359 | torch.save(x, s or f) 360 | mb = os.path.getsize(s or f) / 1E6 # filesize 361 | print('Optimizer stripped from %s,%s %.1fMB' % (f, (' saved as %s,' % s) if s else '', mb)) 362 | 363 | 364 | def print_mutation(hyp, results, yaml_file='hyp_evolved.yaml', bucket=''): 365 | # Print mutation results to evolve.txt (for use with train.py --evolve) 366 | a = '%10s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys 367 | b = '%10.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values 368 | c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) 369 | print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c)) 370 | 371 | if bucket: 372 | url = 'gs://%s/evolve.txt' % bucket 373 | if gsutil_getsize(url) > (os.path.getsize('evolve.txt') if os.path.exists('evolve.txt') else 0): 374 | os.system('gsutil cp %s .' % url) # download evolve.txt if larger than local 375 | 376 | with open('evolve.txt', 'a') as f: # append result 377 | f.write(c + b + '\n') 378 | x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows 379 | x = x[np.argsort(-fitness(x))] # sort 380 | np.savetxt('evolve.txt', x, '%10.3g') # save sort by fitness 381 | 382 | # Save yaml 383 | for i, k in enumerate(hyp.keys()): 384 | hyp[k] = float(x[0, i + 7]) 385 | with open(yaml_file, 'w') as f: 386 | results = tuple(x[0, :7]) 387 | c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) 388 | f.write('# Hyperparameter Evolution Results\n# Generations: %g\n# Metrics: ' % len(x) + c + '\n\n') 389 | yaml.dump(hyp, f, sort_keys=False) 390 | 391 | if bucket: 392 | os.system('gsutil cp evolve.txt %s gs://%s' % (yaml_file, bucket)) # upload 393 | 394 | 395 | def apply_classifier(x, model, img, im0): 396 | # applies a second stage classifier to yolo outputs 397 | im0 = [im0] if isinstance(im0, np.ndarray) else im0 398 | for i, d in enumerate(x): # per image 399 | if d is not None and len(d): 400 | d = d.clone() 401 | 402 | # Reshape and pad cutouts 403 | b = xyxy2xywh(d[:, :4]) # boxes 404 | b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square 405 | b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad 406 | d[:, :4] = xywh2xyxy(b).long() 407 | 408 | # Rescale boxes from img_size to im0 size 409 | scale_coords(img.shape[2:], d[:, :4], im0[i].shape) 410 | 411 | # Classes 412 | pred_cls1 = d[:, 5].long() 413 | ims = [] 414 | for j, a in enumerate(d): # per item 415 | cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])] 416 | im = cv2.resize(cutout, (224, 224)) # BGR 417 | # cv2.imwrite('test%i.jpg' % j, cutout) 418 | 419 | im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 420 | im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32 421 | im /= 255.0 # 0 - 255 to 0.0 - 1.0 422 | ims.append(im) 423 | 424 | pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction 425 | x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections 426 | 427 | return x 428 | 429 | 430 | def increment_path(path, exist_ok=True, sep=''): 431 | # Increment path, i.e. runs/exp --> runs/exp{sep}0, runs/exp{sep}1 etc. 432 | path = Path(path) # os-agnostic 433 | if (path.exists() and exist_ok) or (not path.exists()): 434 | return str(path) 435 | else: 436 | dirs = glob.glob(f"{path}{sep}*") # similar paths 437 | matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs] 438 | i = [int(m.groups()[0]) for m in matches if m] # indices 439 | n = max(i) + 1 if i else 2 # increment number 440 | return f"{path}{sep}{n}" # update path 441 | -------------------------------------------------------------------------------- /utils/google_app_engine/Dockerfile: -------------------------------------------------------------------------------- 1 | FROM gcr.io/google-appengine/python 2 | 3 | # Create a virtualenv for dependencies. This isolates these packages from 4 | # system-level packages. 5 | # Use -p python3 or -p python3.7 to select python version. Default is version 2. 6 | RUN virtualenv /env -p python3 7 | 8 | # Setting these environment variables are the same as running 9 | # source /env/bin/activate. 10 | ENV VIRTUAL_ENV /env 11 | ENV PATH /env/bin:$PATH 12 | 13 | RUN apt-get update && apt-get install -y python-opencv 14 | 15 | # Copy the application's requirements.txt and run pip to install all 16 | # dependencies into the virtualenv. 17 | ADD requirements.txt /app/requirements.txt 18 | RUN pip install -r /app/requirements.txt 19 | 20 | # Add the application source code. 21 | ADD . /app 22 | 23 | # Run a WSGI server to serve the application. gunicorn must be declared as 24 | # a dependency in requirements.txt. 25 | CMD gunicorn -b :$PORT main:app 26 | -------------------------------------------------------------------------------- /utils/google_app_engine/additional_requirements.txt: -------------------------------------------------------------------------------- 1 | # add these requirements in your app on top of the existing ones 2 | pip==18.1 3 | Flask==1.0.2 4 | gunicorn==19.9.0 5 | -------------------------------------------------------------------------------- /utils/google_app_engine/app.yaml: -------------------------------------------------------------------------------- 1 | runtime: custom 2 | env: flex 3 | 4 | service: yolov5app 5 | 6 | liveness_check: 7 | initial_delay_sec: 600 8 | 9 | manual_scaling: 10 | instances: 1 11 | resources: 12 | cpu: 1 13 | memory_gb: 4 14 | disk_size_gb: 20 -------------------------------------------------------------------------------- /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 torch 10 | 11 | 12 | def gsutil_getsize(url=''): 13 | # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du 14 | s = subprocess.check_output('gsutil du %s' % url, shell=True).decode('utf-8') 15 | return eval(s.split(' ')[0]) if len(s) else 0 # bytes 16 | 17 | 18 | def attempt_download(weights): 19 | # Attempt to download pretrained weights if not found locally 20 | weights = str(weights).strip().replace("'", '') 21 | file = Path(weights).name.lower() 22 | 23 | msg = weights + ' missing, try downloading from https://github.com/ultralytics/yolov5/releases/' 24 | models = ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt'] # available models 25 | redundant = False # offer second download option 26 | 27 | if file in models and not os.path.isfile(weights): 28 | # Google Drive 29 | # d = {'yolov5s.pt': '1R5T6rIyy3lLwgFXNms8whc-387H0tMQO', 30 | # 'yolov5m.pt': '1vobuEExpWQVpXExsJ2w-Mbf3HJjWkQJr', 31 | # 'yolov5l.pt': '1hrlqD1Wdei7UT4OgT785BEk1JwnSvNEV', 32 | # 'yolov5x.pt': '1mM8aZJlWTxOg7BZJvNUMrTnA2AbeCVzS'} 33 | # r = gdrive_download(id=d[file], name=weights) if file in d else 1 34 | # if r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6: # check 35 | # return 36 | 37 | try: # GitHub 38 | url = 'https://github.com/ultralytics/yolov5/releases/download/v3.1/' + file 39 | print('Downloading %s to %s...' % (url, weights)) 40 | torch.hub.download_url_to_file(url, weights) 41 | assert os.path.exists(weights) and os.path.getsize(weights) > 1E6 # check 42 | except Exception as e: # GCP 43 | print('Download error: %s' % e) 44 | assert redundant, 'No secondary mirror' 45 | url = 'https://storage.googleapis.com/ultralytics/yolov5/ckpt/' + file 46 | print('Downloading %s to %s...' % (url, weights)) 47 | r = os.system('curl -L %s -o %s' % (url, weights)) # torch.hub.download_url_to_file(url, weights) 48 | finally: 49 | if not (os.path.exists(weights) and os.path.getsize(weights) > 1E6): # check 50 | os.remove(weights) if os.path.exists(weights) else None # remove partial downloads 51 | print('ERROR: Download failure: %s' % msg) 52 | print('') 53 | return 54 | 55 | 56 | def gdrive_download(id='1n_oKgR81BJtqk75b00eAjdv03qVCQn2f', name='coco128.zip'): 57 | # Downloads a file from Google Drive. from utils.google_utils import *; gdrive_download() 58 | t = time.time() 59 | 60 | print('Downloading https://drive.google.com/uc?export=download&id=%s as %s... ' % (id, name), end='') 61 | os.remove(name) if os.path.exists(name) else None # remove existing 62 | os.remove('cookie') if os.path.exists('cookie') else None 63 | 64 | # Attempt file download 65 | out = "NUL" if platform.system() == "Windows" else "/dev/null" 66 | os.system('curl -c ./cookie -s -L "drive.google.com/uc?export=download&id=%s" > %s ' % (id, out)) 67 | if os.path.exists('cookie'): # large file 68 | s = 'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm=%s&id=%s" -o %s' % (get_token(), id, name) 69 | else: # small file 70 | s = 'curl -s -L -o %s "drive.google.com/uc?export=download&id=%s"' % (name, id) 71 | r = os.system(s) # execute, capture return 72 | os.remove('cookie') if os.path.exists('cookie') else None 73 | 74 | # Error check 75 | if r != 0: 76 | os.remove(name) if os.path.exists(name) else None # remove partial 77 | print('Download error ') # raise Exception('Download error') 78 | return r 79 | 80 | # Unzip if archive 81 | if name.endswith('.zip'): 82 | print('unzipping... ', end='') 83 | os.system('unzip -q %s' % name) # unzip 84 | os.remove(name) # remove zip to free space 85 | 86 | print('Done (%.1fs)' % (time.time() - t)) 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/loss.py: -------------------------------------------------------------------------------- 1 | # Loss functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | 6 | from utils.general import bbox_iou 7 | from utils.torch_utils import is_parallel 8 | 9 | 10 | def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 11 | # return positive, negative label smoothing BCE targets 12 | return 1.0 - 0.5 * eps, 0.5 * eps 13 | 14 | 15 | class BCEBlurWithLogitsLoss(nn.Module): 16 | # BCEwithLogitLoss() with reduced missing label effects. 17 | def __init__(self, alpha=0.05): 18 | super(BCEBlurWithLogitsLoss, self).__init__() 19 | self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() 20 | self.alpha = alpha 21 | 22 | def forward(self, pred, true): 23 | loss = self.loss_fcn(pred, true) 24 | pred = torch.sigmoid(pred) # prob from logits 25 | dx = pred - true # reduce only missing label effects 26 | # dx = (pred - true).abs() # reduce missing label and false label effects 27 | alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) 28 | loss *= alpha_factor 29 | return loss.mean() 30 | 31 | 32 | class FocalLoss(nn.Module): 33 | # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 34 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 35 | super(FocalLoss, self).__init__() 36 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 37 | self.gamma = gamma 38 | self.alpha = alpha 39 | self.reduction = loss_fcn.reduction 40 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 41 | 42 | def forward(self, pred, true): 43 | loss = self.loss_fcn(pred, true) 44 | # p_t = torch.exp(-loss) 45 | # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability 46 | 47 | # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py 48 | pred_prob = torch.sigmoid(pred) # prob from logits 49 | p_t = true * pred_prob + (1 - true) * (1 - pred_prob) 50 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 51 | modulating_factor = (1.0 - p_t) ** self.gamma 52 | loss *= alpha_factor * modulating_factor 53 | 54 | if self.reduction == 'mean': 55 | return loss.mean() 56 | elif self.reduction == 'sum': 57 | return loss.sum() 58 | else: # 'none' 59 | return loss 60 | 61 | 62 | class QFocalLoss(nn.Module): 63 | # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 64 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 65 | super(QFocalLoss, self).__init__() 66 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 67 | self.gamma = gamma 68 | self.alpha = alpha 69 | self.reduction = loss_fcn.reduction 70 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 71 | 72 | def forward(self, pred, true): 73 | loss = self.loss_fcn(pred, true) 74 | 75 | pred_prob = torch.sigmoid(pred) # prob from logits 76 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 77 | modulating_factor = torch.abs(true - pred_prob) ** self.gamma 78 | loss *= alpha_factor * modulating_factor 79 | 80 | if self.reduction == 'mean': 81 | return loss.mean() 82 | elif self.reduction == 'sum': 83 | return loss.sum() 84 | else: # 'none' 85 | return loss 86 | 87 | 88 | def compute_loss(p, targets, model): # predictions, targets, model 89 | device = targets.device 90 | lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device) 91 | tcls, tbox, indices, anchors = build_targets(p, targets, model) # targets 92 | h = model.hyp # hyperparameters 93 | 94 | # Define criteria 95 | BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([h['cls_pw']])).to(device) 96 | BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([h['obj_pw']])).to(device) 97 | 98 | # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 99 | cp, cn = smooth_BCE(eps=0.0) 100 | 101 | # Focal loss 102 | g = h['fl_gamma'] # focal loss gamma 103 | if g > 0: 104 | BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) 105 | 106 | # Losses 107 | nt = 0 # number of targets 108 | no = len(p) # number of outputs 109 | balance = [4.0, 1.0, 0.4] if no == 3 else [4.0, 1.0, 0.4, 0.1] # P3-5 or P3-6 110 | for i, pi in enumerate(p): # layer index, layer predictions 111 | b, a, gj, gi = indices[i] # image, anchor, gridy, gridx 112 | tobj = torch.zeros_like(pi[..., 0], device=device) # target obj 113 | 114 | n = b.shape[0] # number of targets 115 | if n: 116 | nt += n # cumulative targets 117 | ps = pi[b, a, gj, gi] # prediction subset corresponding to targets 118 | 119 | # Regression 120 | pxy = ps[:, :2].sigmoid() * 2. - 0.5 121 | pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i] 122 | pbox = torch.cat((pxy, pwh), 1).to(device) # predicted box 123 | iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target) 124 | lbox += (1.0 - iou).mean() # iou loss 125 | 126 | # Objectness 127 | tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio 128 | 129 | # Classification 130 | if model.nc > 1: # cls loss (only if multiple classes) 131 | t = torch.full_like(ps[:, 5:], cn, device=device) # targets 132 | t[range(n), tcls[i]] = cp 133 | lcls += BCEcls(ps[:, 5:], t) # BCE 134 | 135 | # Append targets to text file 136 | # with open('targets.txt', 'a') as file: 137 | # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] 138 | 139 | lobj += BCEobj(pi[..., 4], tobj) * balance[i] # obj loss 140 | 141 | s = 3 / no # output count scaling 142 | lbox *= h['box'] * s 143 | lobj *= h['obj'] * s * (1.4 if no == 4 else 1.) 144 | lcls *= h['cls'] * s 145 | bs = tobj.shape[0] # batch size 146 | 147 | loss = lbox + lobj + lcls 148 | return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach() 149 | 150 | 151 | def build_targets(p, targets, model): 152 | # Build targets for compute_loss(), input targets(image,class,x,y,w,h) 153 | det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module 154 | na, nt = det.na, targets.shape[0] # number of anchors, targets 155 | tcls, tbox, indices, anch = [], [], [], [] 156 | gain = torch.ones(7, device=targets.device) # normalized to gridspace gain 157 | ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) 158 | targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices 159 | 160 | g = 0.5 # bias 161 | off = torch.tensor([[0, 0], 162 | [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m 163 | # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm 164 | ], device=targets.device).float() * g # offsets 165 | 166 | for i in range(det.nl): 167 | anchors = det.anchors[i] 168 | gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain 169 | 170 | # Match targets to anchors 171 | t = targets * gain 172 | if nt: 173 | # Matches 174 | r = t[:, :, 4:6] / anchors[:, None] # wh ratio 175 | j = torch.max(r, 1. / r).max(2)[0] < model.hyp['anchor_t'] # compare 176 | # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) 177 | t = t[j] # filter 178 | 179 | # Offsets 180 | gxy = t[:, 2:4] # grid xy 181 | gxi = gain[[2, 3]] - gxy # inverse 182 | j, k = ((gxy % 1. < g) & (gxy > 1.)).T 183 | l, m = ((gxi % 1. < g) & (gxi > 1.)).T 184 | j = torch.stack((torch.ones_like(j), j, k, l, m)) 185 | t = t.repeat((5, 1, 1))[j] 186 | offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] 187 | else: 188 | t = targets[0] 189 | offsets = 0 190 | 191 | # Define 192 | b, c = t[:, :2].long().T # image, class 193 | gxy = t[:, 2:4] # grid xy 194 | gwh = t[:, 4:6] # grid wh 195 | gij = (gxy - offsets).long() 196 | gi, gj = gij.T # grid xy indices 197 | 198 | # Append 199 | a = t[:, 6].long() # anchor indices 200 | indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices 201 | tbox.append(torch.cat((gxy - gij, gwh), 1)) # box 202 | anch.append(anchors[a]) # anchors 203 | tcls.append(c) # class 204 | 205 | return tcls, tbox, indices, anch 206 | -------------------------------------------------------------------------------- /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='precision-recall_curve.png', 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 | 39 | # Create Precision-Recall curve and compute AP for each class 40 | px, py = np.linspace(0, 1, 1000), [] # for plotting 41 | pr_score = 0.1 # score to evaluate P and R https://github.com/ultralytics/yolov3/issues/898 42 | s = [unique_classes.shape[0], tp.shape[1]] # number class, number iou thresholds (i.e. 10 for mAP0.5...0.95) 43 | ap, p, r = np.zeros(s), np.zeros(s), np.zeros(s) 44 | for ci, c in enumerate(unique_classes): 45 | i = pred_cls == c 46 | n_l = (target_cls == c).sum() # number of labels 47 | n_p = i.sum() # number of predictions 48 | 49 | if n_p == 0 or n_l == 0: 50 | continue 51 | else: 52 | # Accumulate FPs and TPs 53 | fpc = (1 - tp[i]).cumsum(0) 54 | tpc = tp[i].cumsum(0) 55 | 56 | # Recall 57 | recall = tpc / (n_l + 1e-16) # recall curve 58 | r[ci] = np.interp(-pr_score, -conf[i], recall[:, 0]) # r at pr_score, negative x, xp because xp decreases 59 | 60 | # Precision 61 | precision = tpc / (tpc + fpc) # precision curve 62 | p[ci] = np.interp(-pr_score, -conf[i], precision[:, 0]) # p at pr_score 63 | 64 | # AP from recall-precision curve 65 | for j in range(tp.shape[1]): 66 | ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) 67 | if plot and (j == 0): 68 | py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 69 | 70 | # Compute F1 score (harmonic mean of precision and recall) 71 | f1 = 2 * p * r / (p + r + 1e-16) 72 | 73 | if plot: 74 | plot_pr_curve(px, py, ap, save_dir, names) 75 | 76 | return p, r, ap, f1, unique_classes.astype('int32') 77 | 78 | 79 | def compute_ap(recall, precision): 80 | """ Compute the average precision, given the recall and precision curves 81 | # Arguments 82 | recall: The recall curve (list) 83 | precision: The precision curve (list) 84 | # Returns 85 | Average precision, precision curve, recall curve 86 | """ 87 | 88 | # Append sentinel values to beginning and end 89 | mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01])) 90 | mpre = np.concatenate(([1.], precision, [0.])) 91 | 92 | # Compute the precision envelope 93 | mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) 94 | 95 | # Integrate area under curve 96 | method = 'interp' # methods: 'continuous', 'interp' 97 | if method == 'interp': 98 | x = np.linspace(0, 1, 101) # 101-point interp (COCO) 99 | ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate 100 | else: # 'continuous' 101 | i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes 102 | ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve 103 | 104 | return ap, mpre, mrec 105 | 106 | 107 | class ConfusionMatrix: 108 | # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix 109 | def __init__(self, nc, conf=0.25, iou_thres=0.45): 110 | self.matrix = np.zeros((nc + 1, nc + 1)) 111 | self.nc = nc # number of classes 112 | self.conf = conf 113 | self.iou_thres = iou_thres 114 | 115 | def process_batch(self, detections, labels): 116 | """ 117 | Return intersection-over-union (Jaccard index) of boxes. 118 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 119 | Arguments: 120 | detections (Array[N, 6]), x1, y1, x2, y2, conf, class 121 | labels (Array[M, 5]), class, x1, y1, x2, y2 122 | Returns: 123 | None, updates confusion matrix accordingly 124 | """ 125 | detections = detections[detections[:, 4] > self.conf] 126 | gt_classes = labels[:, 0].int() 127 | detection_classes = detections[:, 5].int() 128 | iou = general.box_iou(labels[:, 1:], detections[:, :4]) 129 | 130 | x = torch.where(iou > self.iou_thres) 131 | if x[0].shape[0]: 132 | matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() 133 | if x[0].shape[0] > 1: 134 | matches = matches[matches[:, 2].argsort()[::-1]] 135 | matches = matches[np.unique(matches[:, 1], return_index=True)[1]] 136 | matches = matches[matches[:, 2].argsort()[::-1]] 137 | matches = matches[np.unique(matches[:, 0], return_index=True)[1]] 138 | else: 139 | matches = np.zeros((0, 3)) 140 | 141 | n = matches.shape[0] > 0 142 | m0, m1, _ = matches.transpose().astype(np.int16) 143 | for i, gc in enumerate(gt_classes): 144 | j = m0 == i 145 | if n and sum(j) == 1: 146 | self.matrix[gc, detection_classes[m1[j]]] += 1 # correct 147 | else: 148 | self.matrix[gc, self.nc] += 1 # background FP 149 | 150 | if n: 151 | for i, dc in enumerate(detection_classes): 152 | if not any(m1 == i): 153 | self.matrix[self.nc, dc] += 1 # background FN 154 | 155 | def matrix(self): 156 | return self.matrix 157 | 158 | def plot(self, save_dir='', names=()): 159 | try: 160 | import seaborn as sn 161 | 162 | array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize 163 | array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) 164 | 165 | fig = plt.figure(figsize=(12, 9), tight_layout=True) 166 | sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size 167 | labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels 168 | sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True, 169 | xticklabels=names + ['background FN'] if labels else "auto", 170 | yticklabels=names + ['background FP'] if labels else "auto").set_facecolor((1, 1, 1)) 171 | fig.axes[0].set_xlabel('True') 172 | fig.axes[0].set_ylabel('Predicted') 173 | fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) 174 | except Exception as e: 175 | pass 176 | 177 | def print(self): 178 | for i in range(self.nc + 1): 179 | print(' '.join(map(str, self.matrix[i]))) 180 | 181 | 182 | # Plots ---------------------------------------------------------------------------------------------------------------- 183 | 184 | def plot_pr_curve(px, py, ap, save_dir='.', names=()): 185 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 186 | py = np.stack(py, axis=1) 187 | 188 | if 0 < len(names) < 21: # show mAP in legend if < 10 classes 189 | for i, y in enumerate(py.T): 190 | ax.plot(px, y, linewidth=1, label=f'{names[i]} %.3f' % ap[i, 0]) # plot(recall, precision) 191 | else: 192 | ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) 193 | 194 | ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) 195 | ax.set_xlabel('Recall') 196 | ax.set_ylabel('Precision') 197 | ax.set_xlim(0, 1) 198 | ax.set_ylim(0, 1) 199 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 200 | fig.savefig(Path(save_dir) / 'precision_recall_curve.png', dpi=250) 201 | -------------------------------------------------------------------------------- /utils/plots.py: -------------------------------------------------------------------------------- 1 | # Plotting utils 2 | 3 | import glob 4 | import os 5 | import random 6 | from copy import copy 7 | from pathlib import Path 8 | 9 | import cv2 10 | import math 11 | import matplotlib 12 | import matplotlib.pyplot as plt 13 | import numpy as np 14 | import torch 15 | import yaml 16 | from PIL import Image, ImageDraw 17 | from scipy.signal import butter, filtfilt 18 | 19 | from utils.general import xywh2xyxy, xyxy2xywh 20 | from utils.metrics import fitness 21 | 22 | # Settings 23 | matplotlib.rc('font', **{'size': 11}) 24 | matplotlib.use('Agg') # for writing to files only 25 | 26 | 27 | def color_list(): 28 | # Return first 10 plt colors as (r,g,b) https://stackoverflow.com/questions/51350872/python-from-color-name-to-rgb 29 | def hex2rgb(h): 30 | return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) 31 | 32 | return [hex2rgb(h) for h in plt.rcParams['axes.prop_cycle'].by_key()['color']] 33 | 34 | 35 | def hist2d(x, y, n=100): 36 | # 2d histogram used in labels.png and evolve.png 37 | xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) 38 | hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) 39 | xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) 40 | yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) 41 | return np.log(hist[xidx, yidx]) 42 | 43 | 44 | def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): 45 | # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy 46 | def butter_lowpass(cutoff, fs, order): 47 | nyq = 0.5 * fs 48 | normal_cutoff = cutoff / nyq 49 | return butter(order, normal_cutoff, btype='low', analog=False) 50 | 51 | b, a = butter_lowpass(cutoff, fs, order=order) 52 | return filtfilt(b, a, data) # forward-backward filter 53 | 54 | 55 | def plot_one_box(x, img, color=None, label=None, line_thickness=None): 56 | # Plots one bounding box on image img 57 | tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness 58 | color = color or [random.randint(0, 255) for _ in range(3)] 59 | c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) 60 | cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA) 61 | if label: 62 | tf = max(tl - 1, 1) # font thickness 63 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 64 | c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 65 | cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled 66 | cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) 67 | 68 | 69 | def plot_wh_methods(): # from utils.plots import *; plot_wh_methods() 70 | # Compares the two methods for width-height anchor multiplication 71 | # https://github.com/ultralytics/yolov3/issues/168 72 | x = np.arange(-4.0, 4.0, .1) 73 | ya = np.exp(x) 74 | yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2 75 | 76 | fig = plt.figure(figsize=(6, 3), tight_layout=True) 77 | plt.plot(x, ya, '.-', label='YOLOv3') 78 | plt.plot(x, yb ** 2, '.-', label='YOLOv5 ^2') 79 | plt.plot(x, yb ** 1.6, '.-', label='YOLOv5 ^1.6') 80 | plt.xlim(left=-4, right=4) 81 | plt.ylim(bottom=0, top=6) 82 | plt.xlabel('input') 83 | plt.ylabel('output') 84 | plt.grid() 85 | plt.legend() 86 | fig.savefig('comparison.png', dpi=200) 87 | 88 | 89 | def output_to_target(output): 90 | # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] 91 | targets = [] 92 | for i, o in enumerate(output): 93 | for *box, conf, cls in o.cpu().numpy(): 94 | targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) 95 | return np.array(targets) 96 | 97 | 98 | def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): 99 | # Plot image grid with labels 100 | 101 | if isinstance(images, torch.Tensor): 102 | images = images.cpu().float().numpy() 103 | if isinstance(targets, torch.Tensor): 104 | targets = targets.cpu().numpy() 105 | 106 | # un-normalise 107 | if np.max(images[0]) <= 1: 108 | images *= 255 109 | 110 | tl = 3 # line thickness 111 | tf = max(tl - 1, 1) # font thickness 112 | bs, _, h, w = images.shape # batch size, _, height, width 113 | bs = min(bs, max_subplots) # limit plot images 114 | ns = np.ceil(bs ** 0.5) # number of subplots (square) 115 | 116 | # Check if we should resize 117 | scale_factor = max_size / max(h, w) 118 | if scale_factor < 1: 119 | h = math.ceil(scale_factor * h) 120 | w = math.ceil(scale_factor * w) 121 | 122 | colors = color_list() # list of colors 123 | mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init 124 | for i, img in enumerate(images): 125 | if i == max_subplots: # if last batch has fewer images than we expect 126 | break 127 | 128 | block_x = int(w * (i // ns)) 129 | block_y = int(h * (i % ns)) 130 | 131 | img = img.transpose(1, 2, 0) 132 | if scale_factor < 1: 133 | img = cv2.resize(img, (w, h)) 134 | 135 | mosaic[block_y:block_y + h, block_x:block_x + w, :] = img 136 | if len(targets) > 0: 137 | image_targets = targets[targets[:, 0] == i] 138 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 139 | classes = image_targets[:, 1].astype('int') 140 | labels = image_targets.shape[1] == 6 # labels if no conf column 141 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 142 | 143 | if boxes.shape[1]: 144 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 145 | boxes[[0, 2]] *= w # scale to pixels 146 | boxes[[1, 3]] *= h 147 | elif scale_factor < 1: # absolute coords need scale if image scales 148 | boxes *= scale_factor 149 | boxes[[0, 2]] += block_x 150 | boxes[[1, 3]] += block_y 151 | for j, box in enumerate(boxes.T): 152 | cls = int(classes[j]) 153 | color = colors[cls % len(colors)] 154 | cls = names[cls] if names else cls 155 | if labels or conf[j] > 0.25: # 0.25 conf thresh 156 | label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j]) 157 | plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl) 158 | 159 | # Draw image filename labels 160 | if paths: 161 | label = Path(paths[i]).name[:40] # trim to 40 char 162 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 163 | cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf, 164 | lineType=cv2.LINE_AA) 165 | 166 | # Image border 167 | cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3) 168 | 169 | if fname: 170 | r = min(1280. / max(h, w) / ns, 1.0) # ratio to limit image size 171 | mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA) 172 | # cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB)) # cv2 save 173 | Image.fromarray(mosaic).save(fname) # PIL save 174 | return mosaic 175 | 176 | 177 | def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): 178 | # Plot LR simulating training for full epochs 179 | optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals 180 | y = [] 181 | for _ in range(epochs): 182 | scheduler.step() 183 | y.append(optimizer.param_groups[0]['lr']) 184 | plt.plot(y, '.-', label='LR') 185 | plt.xlabel('epoch') 186 | plt.ylabel('LR') 187 | plt.grid() 188 | plt.xlim(0, epochs) 189 | plt.ylim(0) 190 | plt.savefig(Path(save_dir) / 'LR.png', dpi=200) 191 | 192 | 193 | def plot_test_txt(): # from utils.plots import *; plot_test() 194 | # Plot test.txt histograms 195 | x = np.loadtxt('test.txt', dtype=np.float32) 196 | box = xyxy2xywh(x[:, :4]) 197 | cx, cy = box[:, 0], box[:, 1] 198 | 199 | fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) 200 | ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) 201 | ax.set_aspect('equal') 202 | plt.savefig('hist2d.png', dpi=300) 203 | 204 | fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) 205 | ax[0].hist(cx, bins=600) 206 | ax[1].hist(cy, bins=600) 207 | plt.savefig('hist1d.png', dpi=200) 208 | 209 | 210 | def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() 211 | # Plot targets.txt histograms 212 | x = np.loadtxt('targets.txt', dtype=np.float32).T 213 | s = ['x targets', 'y targets', 'width targets', 'height targets'] 214 | fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) 215 | ax = ax.ravel() 216 | for i in range(4): 217 | ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std())) 218 | ax[i].legend() 219 | ax[i].set_title(s[i]) 220 | plt.savefig('targets.jpg', dpi=200) 221 | 222 | 223 | def plot_study_txt(path='', x=None): # from utils.plots import *; plot_study_txt() 224 | # Plot study.txt generated by test.py 225 | fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True) 226 | ax = ax.ravel() 227 | 228 | fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) 229 | for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolov5s', 'yolov5m', 'yolov5l', 'yolov5x']]: 230 | y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T 231 | x = np.arange(y.shape[1]) if x is None else np.array(x) 232 | s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)'] 233 | for i in range(7): 234 | ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) 235 | ax[i].set_title(s[i]) 236 | 237 | j = y[3].argmax() + 1 238 | ax2.plot(y[6, :j], y[3, :j] * 1E2, '.-', linewidth=2, markersize=8, 239 | label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) 240 | 241 | ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], 242 | 'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet') 243 | 244 | ax2.grid() 245 | ax2.set_xlim(0, 30) 246 | ax2.set_ylim(28, 50) 247 | ax2.set_yticks(np.arange(30, 55, 5)) 248 | ax2.set_xlabel('GPU Speed (ms/img)') 249 | ax2.set_ylabel('COCO AP val') 250 | ax2.legend(loc='lower right') 251 | plt.savefig('test_study.png', dpi=300) 252 | 253 | 254 | def plot_labels(labels, save_dir=Path(''), loggers=None): 255 | # plot dataset labels 256 | c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes 257 | nc = int(c.max() + 1) # number of classes 258 | colors = color_list() 259 | 260 | # seaborn correlogram 261 | try: 262 | import seaborn as sns 263 | import pandas as pd 264 | x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) 265 | sns.pairplot(x, corner=True, diag_kind='hist', kind='scatter', markers='o', 266 | plot_kws=dict(s=3, edgecolor=None, linewidth=1, alpha=0.02), 267 | diag_kws=dict(bins=50)) 268 | plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) 269 | plt.close() 270 | except Exception as e: 271 | pass 272 | 273 | # matplotlib labels 274 | matplotlib.use('svg') # faster 275 | ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() 276 | ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) 277 | ax[0].set_xlabel('classes') 278 | ax[2].scatter(b[0], b[1], c=hist2d(b[0], b[1], 90), cmap='jet') 279 | ax[2].set_xlabel('x') 280 | ax[2].set_ylabel('y') 281 | ax[3].scatter(b[2], b[3], c=hist2d(b[2], b[3], 90), cmap='jet') 282 | ax[3].set_xlabel('width') 283 | ax[3].set_ylabel('height') 284 | 285 | # rectangles 286 | labels[:, 1:3] = 0.5 # center 287 | labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 288 | img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) 289 | for cls, *box in labels[:1000]: 290 | ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10]) # plot 291 | ax[1].imshow(img) 292 | ax[1].axis('off') 293 | 294 | for a in [0, 1, 2, 3]: 295 | for s in ['top', 'right', 'left', 'bottom']: 296 | ax[a].spines[s].set_visible(False) 297 | 298 | plt.savefig(save_dir / 'labels.jpg', dpi=200) 299 | matplotlib.use('Agg') 300 | plt.close() 301 | 302 | # loggers 303 | for k, v in loggers.items() or {}: 304 | if k == 'wandb' and v: 305 | v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}) 306 | 307 | 308 | def plot_evolution(yaml_file='data/hyp.finetune.yaml'): # from utils.plots import *; plot_evolution() 309 | # Plot hyperparameter evolution results in evolve.txt 310 | with open(yaml_file) as f: 311 | hyp = yaml.load(f, Loader=yaml.FullLoader) 312 | x = np.loadtxt('evolve.txt', ndmin=2) 313 | f = fitness(x) 314 | # weights = (f - f.min()) ** 2 # for weighted results 315 | plt.figure(figsize=(10, 12), tight_layout=True) 316 | matplotlib.rc('font', **{'size': 8}) 317 | for i, (k, v) in enumerate(hyp.items()): 318 | y = x[:, i + 7] 319 | # mu = (y * weights).sum() / weights.sum() # best weighted result 320 | mu = y[f.argmax()] # best single result 321 | plt.subplot(6, 5, i + 1) 322 | plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none') 323 | plt.plot(mu, f.max(), 'k+', markersize=15) 324 | plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters 325 | if i % 5 != 0: 326 | plt.yticks([]) 327 | print('%15s: %.3g' % (k, mu)) 328 | plt.savefig('evolve.png', dpi=200) 329 | print('\nPlot saved as evolve.png') 330 | 331 | 332 | def plot_results_overlay(start=0, stop=0): # from utils.plots import *; plot_results_overlay() 333 | # Plot training 'results*.txt', overlaying train and val losses 334 | s = ['train', 'train', 'train', 'Precision', 'mAP@0.5', 'val', 'val', 'val', 'Recall', 'mAP@0.5:0.95'] # legends 335 | t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles 336 | for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): 337 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 338 | n = results.shape[1] # number of rows 339 | x = range(start, min(stop, n) if stop else n) 340 | fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True) 341 | ax = ax.ravel() 342 | for i in range(5): 343 | for j in [i, i + 5]: 344 | y = results[j, x] 345 | ax[i].plot(x, y, marker='.', label=s[j]) 346 | # y_smooth = butter_lowpass_filtfilt(y) 347 | # ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j]) 348 | 349 | ax[i].set_title(t[i]) 350 | ax[i].legend() 351 | ax[i].set_ylabel(f) if i == 0 else None # add filename 352 | fig.savefig(f.replace('.txt', '.png'), dpi=200) 353 | 354 | 355 | def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''): 356 | # Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp') 357 | fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) 358 | ax = ax.ravel() 359 | s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall', 360 | 'val Box', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95'] 361 | if bucket: 362 | # files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id] 363 | files = ['results%g.txt' % x for x in id] 364 | c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id) 365 | os.system(c) 366 | else: 367 | files = list(Path(save_dir).glob('results*.txt')) 368 | assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir) 369 | for fi, f in enumerate(files): 370 | try: 371 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 372 | n = results.shape[1] # number of rows 373 | x = range(start, min(stop, n) if stop else n) 374 | for i in range(10): 375 | y = results[i, x] 376 | if i in [0, 1, 2, 5, 6, 7]: 377 | y[y == 0] = np.nan # don't show zero loss values 378 | # y /= y[0] # normalize 379 | label = labels[fi] if len(labels) else f.stem 380 | ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8) 381 | ax[i].set_title(s[i]) 382 | # if i in [5, 6, 7]: # share train and val loss y axes 383 | # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) 384 | except Exception as e: 385 | print('Warning: Plotting error for %s; %s' % (f, e)) 386 | 387 | ax[1].legend() 388 | fig.savefig(Path(save_dir) / 'results.png', dpi=200) 389 | -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | # PyTorch utils 2 | 3 | import logging 4 | import os 5 | import time 6 | from contextlib import contextmanager 7 | from copy import deepcopy 8 | 9 | import math 10 | import torch 11 | import torch.backends.cudnn as cudnn 12 | import torch.nn as nn 13 | import torch.nn.functional as F 14 | import torchvision 15 | 16 | logger = logging.getLogger(__name__) 17 | 18 | 19 | @contextmanager 20 | def torch_distributed_zero_first(local_rank: int): 21 | """ 22 | Decorator to make all processes in distributed training wait for each local_master to do something. 23 | """ 24 | if local_rank not in [-1, 0]: 25 | torch.distributed.barrier() 26 | yield 27 | if local_rank == 0: 28 | torch.distributed.barrier() 29 | 30 | 31 | def init_torch_seeds(seed=0): 32 | # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html 33 | torch.manual_seed(seed) 34 | if seed == 0: # slower, more reproducible 35 | cudnn.deterministic = True 36 | cudnn.benchmark = False 37 | else: # faster, less reproducible 38 | cudnn.deterministic = False 39 | cudnn.benchmark = True 40 | 41 | 42 | def select_device(device='', batch_size=None): 43 | # device = 'cpu' or '0' or '0,1,2,3' 44 | cpu_request = device.lower() == 'cpu' 45 | if device and not cpu_request: # if device requested other than 'cpu' 46 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 47 | assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device # check availablity 48 | 49 | cuda = False if cpu_request else torch.cuda.is_available() 50 | if cuda: 51 | c = 1024 ** 2 # bytes to MB 52 | ng = torch.cuda.device_count() 53 | if ng > 1 and batch_size: # check that batch_size is compatible with device_count 54 | assert batch_size % ng == 0, 'batch-size %g not multiple of GPU count %g' % (batch_size, ng) 55 | x = [torch.cuda.get_device_properties(i) for i in range(ng)] 56 | s = f'Using torch {torch.__version__} ' 57 | for i in range(0, ng): 58 | if i == 1: 59 | s = ' ' * len(s) 60 | logger.info("%sCUDA:%g (%s, %dMB)" % (s, i, x[i].name, x[i].total_memory / c)) 61 | else: 62 | logger.info(f'Using torch {torch.__version__} CPU') 63 | 64 | logger.info('') # skip a line 65 | return torch.device('cuda:0' if cuda else 'cpu') 66 | 67 | 68 | def time_synchronized(): 69 | torch.cuda.synchronize() if torch.cuda.is_available() else None 70 | return time.time() 71 | 72 | 73 | def is_parallel(model): 74 | return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) 75 | 76 | 77 | def intersect_dicts(da, db, exclude=()): 78 | # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values 79 | 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} 80 | 81 | 82 | def initialize_weights(model): 83 | for m in model.modules(): 84 | t = type(m) 85 | if t is nn.Conv2d: 86 | pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 87 | elif t is nn.BatchNorm2d: 88 | m.eps = 1e-3 89 | m.momentum = 0.03 90 | elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]: 91 | m.inplace = True 92 | 93 | 94 | def find_modules(model, mclass=nn.Conv2d): 95 | # Finds layer indices matching module class 'mclass' 96 | return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] 97 | 98 | 99 | def sparsity(model): 100 | # Return global model sparsity 101 | a, b = 0., 0. 102 | for p in model.parameters(): 103 | a += p.numel() 104 | b += (p == 0).sum() 105 | return b / a 106 | 107 | 108 | def prune(model, amount=0.3): 109 | # Prune model to requested global sparsity 110 | import torch.nn.utils.prune as prune 111 | print('Pruning model... ', end='') 112 | for name, m in model.named_modules(): 113 | if isinstance(m, nn.Conv2d): 114 | prune.l1_unstructured(m, name='weight', amount=amount) # prune 115 | prune.remove(m, 'weight') # make permanent 116 | print(' %.3g global sparsity' % sparsity(model)) 117 | 118 | 119 | def fuse_conv_and_bn(conv, bn): 120 | # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 121 | fusedconv = nn.Conv2d(conv.in_channels, 122 | conv.out_channels, 123 | kernel_size=conv.kernel_size, 124 | stride=conv.stride, 125 | padding=conv.padding, 126 | groups=conv.groups, 127 | bias=True).requires_grad_(False).to(conv.weight.device) 128 | 129 | # prepare filters 130 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 131 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 132 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size())) 133 | 134 | # prepare spatial bias 135 | b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias 136 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 137 | fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) 138 | 139 | return fusedconv 140 | 141 | 142 | def model_info(model, verbose=False, img_size=640): 143 | # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] 144 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 145 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 146 | if verbose: 147 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 148 | for i, (name, p) in enumerate(model.named_parameters()): 149 | name = name.replace('module_list.', '') 150 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 151 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 152 | 153 | try: # FLOPS 154 | from thop import profile 155 | stride = int(model.stride.max()) if hasattr(model, 'stride') else 32 156 | img = torch.zeros((1, 3, stride, stride), device=next(model.parameters()).device) # input 157 | flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride FLOPS 158 | img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float 159 | fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 FLOPS 160 | except (ImportError, Exception): 161 | fs = '' 162 | 163 | logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") 164 | 165 | 166 | def load_classifier(name='resnet101', n=2): 167 | # Loads a pretrained model reshaped to n-class output 168 | model = torchvision.models.__dict__[name](pretrained=True) 169 | 170 | # ResNet model properties 171 | # input_size = [3, 224, 224] 172 | # input_space = 'RGB' 173 | # input_range = [0, 1] 174 | # mean = [0.485, 0.456, 0.406] 175 | # std = [0.229, 0.224, 0.225] 176 | 177 | # Reshape output to n classes 178 | filters = model.fc.weight.shape[1] 179 | model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True) 180 | model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True) 181 | model.fc.out_features = n 182 | return model 183 | 184 | 185 | def scale_img(img, ratio=1.0, same_shape=False): # img(16,3,256,416), r=ratio 186 | # scales img(bs,3,y,x) by ratio 187 | if ratio == 1.0: 188 | return img 189 | else: 190 | h, w = img.shape[2:] 191 | s = (int(h * ratio), int(w * ratio)) # new size 192 | img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize 193 | if not same_shape: # pad/crop img 194 | gs = 32 # (pixels) grid size 195 | h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] 196 | return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean 197 | 198 | 199 | def copy_attr(a, b, include=(), exclude=()): 200 | # Copy attributes from b to a, options to only include [...] and to exclude [...] 201 | for k, v in b.__dict__.items(): 202 | if (len(include) and k not in include) or k.startswith('_') or k in exclude: 203 | continue 204 | else: 205 | setattr(a, k, v) 206 | 207 | 208 | class ModelEMA: 209 | """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models 210 | Keep a moving average of everything in the model state_dict (parameters and buffers). 211 | This is intended to allow functionality like 212 | https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage 213 | A smoothed version of the weights is necessary for some training schemes to perform well. 214 | This class is sensitive where it is initialized in the sequence of model init, 215 | GPU assignment and distributed training wrappers. 216 | """ 217 | 218 | def __init__(self, model, decay=0.9999, updates=0): 219 | # Create EMA 220 | self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA 221 | # if next(model.parameters()).device.type != 'cpu': 222 | # self.ema.half() # FP16 EMA 223 | self.updates = updates # number of EMA updates 224 | self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) 225 | for p in self.ema.parameters(): 226 | p.requires_grad_(False) 227 | 228 | def update(self, model): 229 | # Update EMA parameters 230 | with torch.no_grad(): 231 | self.updates += 1 232 | d = self.decay(self.updates) 233 | 234 | msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict 235 | for k, v in self.ema.state_dict().items(): 236 | if v.dtype.is_floating_point: 237 | v *= d 238 | v += (1. - d) * msd[k].detach() 239 | 240 | def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): 241 | # Update EMA attributes 242 | copy_attr(self.ema, model, include, exclude) 243 | --------------------------------------------------------------------------------