├── LICENSE ├── README.md ├── inference ├── images │ └── test.jpg └── output │ └── result.jpg ├── models ├── __init__.py ├── common.py ├── experimental.py ├── onnx_export.py ├── yolo.py ├── yolov3-spp.yaml ├── yolov5l.yaml ├── yolov5m.yaml ├── yolov5s.yaml └── yolov5x.yaml ├── simple_inference.py ├── utils ├── __init__.py ├── torch_utils.py └── utils.py └── weights └── download_weights.sh /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 | # YOLOv5 Pytorch Inference 2 | 3 | ## Inference 4 | 5 | ```bash 6 | $ python simple_inference.py --image inference/images/test.jpg 7 | ``` 8 | 9 | ## Output 10 | 11 | ![](inference/output/result.jpg) 12 | 13 | ## For Training refer 14 | ![ultralytics/yolov5](https://github.com/ultralytics/yolov5) 15 | 16 | ## Licence 17 | [![License: GPL v3](https://img.shields.io/badge/License-GPLv3-blue.svg)](https://www.gnu.org/licenses/gpl-3.0) 18 | 19 | ## Credits 20 | ![YOLOv5 by Ultralytics](https://github.com/ultralytics/yolov5) 21 | -------------------------------------------------------------------------------- /inference/images/test.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Nannigalaxy/yolov5_pytorch_inference/caf9febaeb5266fbcc0cc992c8b73641c3bcf681/inference/images/test.jpg -------------------------------------------------------------------------------- /inference/output/result.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Nannigalaxy/yolov5_pytorch_inference/caf9febaeb5266fbcc0cc992c8b73641c3bcf681/inference/output/result.jpg -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Nannigalaxy/yolov5_pytorch_inference/caf9febaeb5266fbcc0cc992c8b73641c3bcf681/models/__init__.py -------------------------------------------------------------------------------- /models/common.py: -------------------------------------------------------------------------------- 1 | # This file contains modules common to various models 2 | 3 | 4 | from utils.utils import * 5 | 6 | 7 | def DWConv(c1, c2, k=1, s=1, act=True): 8 | # Depthwise convolution 9 | return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act) 10 | 11 | 12 | class Conv(nn.Module): 13 | # Standard convolution 14 | def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups 15 | super(Conv, self).__init__() 16 | self.conv = nn.Conv2d(c1, c2, k, s, k // 2, groups=g, bias=False) 17 | self.bn = nn.BatchNorm2d(c2) 18 | self.act = nn.LeakyReLU(0.1, inplace=True) if act else nn.Identity() 19 | 20 | def forward(self, x): 21 | return self.act(self.bn(self.conv(x))) 22 | 23 | def fuseforward(self, x): 24 | return self.act(self.conv(x)) 25 | 26 | 27 | class Bottleneck(nn.Module): 28 | # Standard bottleneck 29 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 30 | super(Bottleneck, self).__init__() 31 | c_ = int(c2 * e) # hidden channels 32 | self.cv1 = Conv(c1, c_, 1, 1) 33 | self.cv2 = Conv(c_, c2, 3, 1, g=g) 34 | self.add = shortcut and c1 == c2 35 | 36 | def forward(self, x): 37 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 38 | 39 | 40 | class BottleneckCSP(nn.Module): 41 | # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks 42 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 43 | super(BottleneckCSP, self).__init__() 44 | c_ = int(c2 * e) # hidden channels 45 | self.cv1 = Conv(c1, c_, 1, 1) 46 | self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) 47 | self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) 48 | self.cv4 = Conv(c2, c2, 1, 1) 49 | self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) 50 | self.act = nn.LeakyReLU(0.1, inplace=True) 51 | self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) 52 | 53 | def forward(self, x): 54 | y1 = self.cv3(self.m(self.cv1(x))) 55 | y2 = self.cv2(x) 56 | return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) 57 | 58 | 59 | class SPP(nn.Module): 60 | # Spatial pyramid pooling layer used in YOLOv3-SPP 61 | def __init__(self, c1, c2, k=(5, 9, 13)): 62 | super(SPP, self).__init__() 63 | c_ = c1 // 2 # hidden channels 64 | self.cv1 = Conv(c1, c_, 1, 1) 65 | self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) 66 | self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) 67 | 68 | def forward(self, x): 69 | x = self.cv1(x) 70 | return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) 71 | 72 | 73 | class Flatten(nn.Module): 74 | # Use after nn.AdaptiveAvgPool2d(1) to remove last 2 dimensions 75 | def forward(self, x): 76 | return x.view(x.size(0), -1) 77 | 78 | 79 | class Focus(nn.Module): 80 | # Focus wh information into c-space 81 | def __init__(self, c1, c2, k=1): 82 | super(Focus, self).__init__() 83 | self.conv = Conv(c1 * 4, c2, k, 1) 84 | 85 | def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) 86 | return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) 87 | 88 | 89 | class Concat(nn.Module): 90 | # Concatenate a list of tensors along dimension 91 | def __init__(self, dimension=1): 92 | super(Concat, self).__init__() 93 | self.d = dimension 94 | 95 | def forward(self, x): 96 | return torch.cat(x, self.d) 97 | -------------------------------------------------------------------------------- /models/experimental.py: -------------------------------------------------------------------------------- 1 | from models.common import * 2 | 3 | 4 | class Sum(nn.Module): 5 | # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 6 | def __init__(self, n, weight=False): # n: number of inputs 7 | super(Sum, self).__init__() 8 | self.weight = weight # apply weights boolean 9 | self.iter = range(n - 1) # iter object 10 | if weight: 11 | self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights 12 | 13 | def forward(self, x): 14 | y = x[0] # no weight 15 | if self.weight: 16 | w = torch.sigmoid(self.w) * 2 17 | for i in self.iter: 18 | y = y + x[i + 1] * w[i] 19 | else: 20 | for i in self.iter: 21 | y = y + x[i + 1] 22 | return y 23 | 24 | 25 | class GhostConv(nn.Module): 26 | # Ghost Convolution https://github.com/huawei-noah/ghostnet 27 | def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups 28 | super(GhostConv, self).__init__() 29 | c_ = c2 // 2 # hidden channels 30 | self.cv1 = Conv(c1, c_, k, s, g, act) 31 | self.cv2 = Conv(c_, c_, 5, 1, c_, act) 32 | 33 | def forward(self, x): 34 | y = self.cv1(x) 35 | return torch.cat([y, self.cv2(y)], 1) 36 | 37 | 38 | class GhostBottleneck(nn.Module): 39 | # Ghost Bottleneck https://github.com/huawei-noah/ghostnet 40 | def __init__(self, c1, c2, k, s): 41 | super(GhostBottleneck, self).__init__() 42 | c_ = c2 // 2 43 | self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw 44 | DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw 45 | GhostConv(c_, c2, 1, 1, act=False)) # pw-linear 46 | self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), 47 | Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity() 48 | 49 | def forward(self, x): 50 | return self.conv(x) + self.shortcut(x) 51 | 52 | 53 | class ConvPlus(nn.Module): 54 | # Plus-shaped convolution 55 | def __init__(self, c1, c2, k=3, s=1, g=1, bias=True): # ch_in, ch_out, kernel, stride, groups 56 | super(ConvPlus, self).__init__() 57 | self.cv1 = nn.Conv2d(c1, c2, (k, 1), s, (k // 2, 0), groups=g, bias=bias) 58 | self.cv2 = nn.Conv2d(c1, c2, (1, k), s, (0, k // 2), groups=g, bias=bias) 59 | 60 | def forward(self, x): 61 | return self.cv1(x) + self.cv2(x) 62 | 63 | 64 | class MixConv2d(nn.Module): 65 | # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 66 | def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): 67 | super(MixConv2d, self).__init__() 68 | groups = len(k) 69 | if equal_ch: # equal c_ per group 70 | i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices 71 | c_ = [(i == g).sum() for g in range(groups)] # intermediate channels 72 | else: # equal weight.numel() per group 73 | b = [c2] + [0] * groups 74 | a = np.eye(groups + 1, groups, k=-1) 75 | a -= np.roll(a, 1, axis=1) 76 | a *= np.array(k) ** 2 77 | a[0] = 1 78 | c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b 79 | 80 | self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) 81 | self.bn = nn.BatchNorm2d(c2) 82 | self.act = nn.LeakyReLU(0.1, inplace=True) 83 | 84 | def forward(self, x): 85 | return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) 86 | -------------------------------------------------------------------------------- /models/onnx_export.py: -------------------------------------------------------------------------------- 1 | """Exports a pytorch *.pt model to *.onnx format 2 | 3 | Usage: 4 | import torch 5 | $ export PYTHONPATH="$PWD" && python models/onnx_export.py --weights ./weights/yolov5s.pt --img 640 --batch 1 6 | """ 7 | 8 | import argparse 9 | 10 | import onnx 11 | 12 | from models.common import * 13 | 14 | if __name__ == '__main__': 15 | parser = argparse.ArgumentParser() 16 | parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path') 17 | parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') 18 | parser.add_argument('--batch-size', type=int, default=1, help='batch size') 19 | opt = parser.parse_args() 20 | print(opt) 21 | 22 | # Parameters 23 | f = opt.weights.replace('.pt', '.onnx') # onnx filename 24 | img = torch.zeros((opt.batch_size, 3, *opt.img_size)) # image size, (1, 3, 320, 192) iDetection 25 | 26 | # Load pytorch model 27 | google_utils.attempt_download(opt.weights) 28 | model = torch.load(opt.weights, map_location=torch.device('cpu'))['model'].float() 29 | model.eval() 30 | model.fuse() 31 | 32 | # Export to onnx 33 | model.model[-1].export = True # set Detect() layer export=True 34 | _ = model(img) # dry run 35 | torch.onnx.export(model, img, f, verbose=False, opset_version=11, input_names=['images'], 36 | output_names=['output']) # output_names=['classes', 'boxes'] 37 | 38 | # Check onnx model 39 | model = onnx.load(f) # load onnx model 40 | onnx.checker.check_model(model) # check onnx model 41 | print(onnx.helper.printable_graph(model.graph)) # print a human readable representation of the graph 42 | print('Export complete. ONNX model saved to %s\nView with https://github.com/lutzroeder/netron' % f) 43 | -------------------------------------------------------------------------------- /models/yolo.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | import yaml 4 | 5 | from models.experimental import * 6 | 7 | 8 | class Detect(nn.Module): 9 | def __init__(self, nc=80, anchors=()): # detection layer 10 | super(Detect, self).__init__() 11 | self.stride = None # strides computed during build 12 | self.nc = nc # number of classes 13 | self.no = nc + 5 # number of outputs per anchor 14 | self.nl = len(anchors) # number of detection layers 15 | self.na = len(anchors[0]) // 2 # number of anchors 16 | self.grid = [torch.zeros(1)] * self.nl # init grid 17 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 18 | self.register_buffer('anchors', a) # shape(nl,na,2) 19 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 20 | self.export = False # onnx export 21 | 22 | def forward(self, x): 23 | # x = x.copy() # for profiling 24 | z = [] # inference output 25 | self.training |= self.export 26 | for i in range(self.nl): 27 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 28 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 29 | 30 | if not self.training: # inference 31 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 32 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 33 | 34 | y = x[i].sigmoid() 35 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy 36 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 37 | z.append(y.view(bs, -1, self.no)) 38 | 39 | return x if self.training else (torch.cat(z, 1), x) 40 | 41 | @staticmethod 42 | def _make_grid(nx=20, ny=20): 43 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 44 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 45 | 46 | 47 | class Model(nn.Module): 48 | def __init__(self, model_cfg='yolov5s.yaml', ch=3, nc=None): # model, input channels, number of classes 49 | super(Model, self).__init__() 50 | if type(model_cfg) is dict: 51 | self.md = model_cfg # model dict 52 | else: # is *.yaml 53 | with open(model_cfg) as f: 54 | self.md = yaml.load(f, Loader=yaml.FullLoader) # model dict 55 | 56 | # Define model 57 | if nc: 58 | self.md['nc'] = nc # override yaml value 59 | self.model, self.save = parse_model(self.md, ch=[ch]) # model, savelist, ch_out 60 | # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))]) 61 | 62 | # Build strides, anchors 63 | m = self.model[-1] # Detect() 64 | m.stride = torch.tensor([64 / x.shape[-2] for x in self.forward(torch.zeros(1, ch, 64, 64))]) # forward 65 | m.anchors /= m.stride.view(-1, 1, 1) 66 | self.stride = m.stride 67 | 68 | # Init weights, biases 69 | torch_utils.initialize_weights(self) 70 | self._initialize_biases() # only run once 71 | torch_utils.model_info(self) 72 | print('') 73 | 74 | def forward(self, x, augment=False, profile=False): 75 | if augment: 76 | img_size = x.shape[-2:] # height, width 77 | s = [0.83, 0.67] # scales 78 | y = [] 79 | for i, xi in enumerate((x, 80 | torch_utils.scale_img(x.flip(3), s[0]), # flip-lr and scale 81 | torch_utils.scale_img(x, s[1]), # scale 82 | )): 83 | # cv2.imwrite('img%g.jpg' % i, 255 * xi[0].numpy().transpose((1, 2, 0))[:, :, ::-1]) 84 | y.append(self.forward_once(xi)[0]) 85 | 86 | y[1][..., :4] /= s[0] # scale 87 | y[1][..., 0] = img_size[1] - y[1][..., 0] # flip lr 88 | y[2][..., :4] /= s[1] # scale 89 | return torch.cat(y, 1), None # augmented inference, train 90 | else: 91 | return self.forward_once(x, profile) # single-scale inference, train 92 | 93 | def forward_once(self, x, profile=False): 94 | y, dt = [], [] # outputs 95 | for m in self.model: 96 | if m.f != -1: # if not from previous layer 97 | x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers 98 | 99 | if profile: 100 | import thop 101 | o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # FLOPS 102 | t = torch_utils.time_synchronized() 103 | for _ in range(10): 104 | _ = m(x) 105 | dt.append((torch_utils.time_synchronized() - t) * 100) 106 | print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type)) 107 | 108 | x = m(x) # run 109 | y.append(x if m.i in self.save else None) # save output 110 | 111 | if profile: 112 | print('%.1fms total' % sum(dt)) 113 | return x 114 | 115 | def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency 116 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 117 | m = self.model[-1] # Detect() module 118 | for f, s in zip(m.f, m.stride): #  from 119 | mi = self.model[f % m.i] 120 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 121 | b[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 122 | b[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 123 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 124 | 125 | def _print_biases(self): 126 | m = self.model[-1] # Detect() module 127 | for f in sorted([x % m.i for x in m.f]): #  from 128 | b = self.model[f].bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) 129 | print(('%g Conv2d.bias:' + '%10.3g' * 6) % (f, *b[:5].mean(1).tolist(), b[5:].mean())) 130 | 131 | # def _print_weights(self): 132 | # for m in self.model.modules(): 133 | # if type(m) is Bottleneck: 134 | # print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights 135 | 136 | def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers 137 | print('Fusing layers...') 138 | for m in self.model.modules(): 139 | if type(m) is Conv: 140 | m.conv = torch_utils.fuse_conv_and_bn(m.conv, m.bn) # update conv 141 | m.bn = None # remove batchnorm 142 | m.forward = m.fuseforward # update forward 143 | torch_utils.model_info(self) 144 | 145 | 146 | def parse_model(md, ch): # model_dict, input_channels(3) 147 | print('\n%3s%15s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments')) 148 | anchors, nc, gd, gw = md['anchors'], md['nc'], md['depth_multiple'], md['width_multiple'] 149 | na = (len(anchors[0]) // 2) # number of anchors 150 | no = na * (nc + 5) # number of outputs = anchors * (classes + 5) 151 | 152 | layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out 153 | for i, (f, n, m, args) in enumerate(md['backbone'] + md['head']): # from, number, module, args 154 | m = eval(m) if isinstance(m, str) else m # eval strings 155 | for j, a in enumerate(args): 156 | try: 157 | args[j] = eval(a) if isinstance(a, str) else a # eval strings 158 | except: 159 | pass 160 | 161 | n = max(round(n * gd), 1) if n > 1 else n # depth gain 162 | if m in [nn.Conv2d, Conv, Bottleneck, SPP, DWConv, MixConv2d, Focus, ConvPlus, BottleneckCSP]: 163 | c1, c2 = ch[f], args[0] 164 | 165 | # Normal 166 | # if i > 0 and args[0] != no: # channel expansion factor 167 | # ex = 1.75 # exponential (default 2.0) 168 | # e = math.log(c2 / ch[1]) / math.log(2) 169 | # c2 = int(ch[1] * ex ** e) 170 | # if m != Focus: 171 | c2 = make_divisible(c2 * gw, 8) if c2 != no else c2 172 | 173 | # Experimental 174 | # if i > 0 and args[0] != no: # channel expansion factor 175 | # ex = 1 + gw # exponential (default 2.0) 176 | # ch1 = 32 # ch[1] 177 | # e = math.log(c2 / ch1) / math.log(2) # level 1-n 178 | # c2 = int(ch1 * ex ** e) 179 | # if m != Focus: 180 | # c2 = make_divisible(c2, 8) if c2 != no else c2 181 | 182 | args = [c1, c2, *args[1:]] 183 | if m is BottleneckCSP: 184 | args.insert(2, n) 185 | n = 1 186 | elif m is nn.BatchNorm2d: 187 | args = [ch[f]] 188 | elif m is Concat: 189 | c2 = sum([ch[-1 if x == -1 else x + 1] for x in f]) 190 | elif m is Detect: 191 | f = f or list(reversed([(-1 if j == i else j - 1) for j, x in enumerate(ch) if x == no])) 192 | else: 193 | c2 = ch[f] 194 | 195 | m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module 196 | t = str(m)[8:-2].replace('__main__.', '') # module type 197 | np = sum([x.numel() for x in m_.parameters()]) # number params 198 | m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params 199 | print('%3s%15s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print 200 | save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist 201 | layers.append(m_) 202 | ch.append(c2) 203 | return nn.Sequential(*layers), sorted(save) 204 | 205 | 206 | if __name__ == '__main__': 207 | parser = argparse.ArgumentParser() 208 | parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml') 209 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 210 | opt = parser.parse_args() 211 | opt.cfg = glob.glob('./**/' + opt.cfg, recursive=True)[0] # find file 212 | device = torch_utils.select_device(opt.device) 213 | 214 | # Create model 215 | model = Model(opt.cfg).to(device) 216 | model.train() 217 | 218 | # Profile 219 | # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device) 220 | # y = model(img, profile=True) 221 | # print([y[0].shape] + [x.shape for x in y[1]]) 222 | 223 | # ONNX export 224 | # model.model[-1].export = True 225 | # torch.onnx.export(model, img, f.replace('.yaml', '.onnx'), verbose=True, opset_version=11) 226 | 227 | # Tensorboard 228 | # from torch.utils.tensorboard import SummaryWriter 229 | # tb_writer = SummaryWriter() 230 | # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/") 231 | # tb_writer.add_graph(model.model, img) # add model to tensorboard 232 | # tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard 233 | -------------------------------------------------------------------------------- /models/yolov3-spp.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # darknet53 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Conv, [32, 3, 1]], # 0 16 | [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 17 | [-1, 1, Bottleneck, [64]], 18 | [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 19 | [-1, 2, Bottleneck, [128]], 20 | [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 21 | [-1, 8, Bottleneck, [256]], 22 | [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 23 | [-1, 8, Bottleneck, [512]], 24 | [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 25 | [-1, 4, Bottleneck, [1024]], # 10 26 | ] 27 | 28 | # yolov3-spp head 29 | # na = len(anchors[0]) 30 | head: 31 | [[-1, 1, Bottleneck, [1024, False]], # 11 32 | [-1, 1, SPP, [512, [5, 9, 13]]], 33 | [-1, 1, Conv, [1024, 3, 1]], 34 | [-1, 1, Conv, [512, 1, 1]], 35 | [-1, 1, Conv, [1024, 3, 1]], 36 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 16 (P5/32-large) 37 | 38 | [-3, 1, Conv, [256, 1, 1]], 39 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 40 | [[-1, 8], 1, Concat, [1]], # cat backbone P4 41 | [-1, 1, Bottleneck, [512, False]], 42 | [-1, 1, Bottleneck, [512, False]], 43 | [-1, 1, Conv, [256, 1, 1]], 44 | [-1, 1, Conv, [512, 3, 1]], 45 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 24 (P4/16-medium) 46 | 47 | [-3, 1, Conv, [128, 1, 1]], 48 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 49 | [[-1, 6], 1, Concat, [1]], # cat backbone P3 50 | [-1, 1, Bottleneck, [256, False]], 51 | [-1, 2, Bottleneck, [256, False]], 52 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 30 (P3/8-small) 53 | 54 | [[], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 55 | ] 56 | -------------------------------------------------------------------------------- /models/yolov5l.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.0 # model depth multiple 4 | width_multiple: 1.0 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # yolov5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 1-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 2-P2/4 17 | [-1, 3, Bottleneck, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 4-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 6-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 6, BottleneckCSP, [1024]], # 10 25 | ] 26 | 27 | # yolov5 head 28 | head: 29 | [[-1, 3, BottleneckCSP, [1024, False]], # 11 30 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 12 (P5/32-large) 31 | 32 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 33 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 34 | [-1, 1, Conv, [512, 1, 1]], 35 | [-1, 3, BottleneckCSP, [512, False]], 36 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 17 (P4/16-medium) 37 | 38 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 39 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 40 | [-1, 1, Conv, [256, 1, 1]], 41 | [-1, 3, BottleneckCSP, [256, False]], 42 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 22 (P3/8-small) 43 | 44 | [[], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 45 | ] 46 | -------------------------------------------------------------------------------- /models/yolov5m.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.67 # model depth multiple 4 | width_multiple: 0.75 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # yolov5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 1-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 2-P2/4 17 | [-1, 3, Bottleneck, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 4-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 6-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 6, BottleneckCSP, [1024]], # 10 25 | ] 26 | 27 | # yolov5 head 28 | head: 29 | [[-1, 3, BottleneckCSP, [1024, False]], # 11 30 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 12 (P5/32-large) 31 | 32 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 33 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 34 | [-1, 1, Conv, [512, 1, 1]], 35 | [-1, 3, BottleneckCSP, [512, False]], 36 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 17 (P4/16-medium) 37 | 38 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 39 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 40 | [-1, 1, Conv, [256, 1, 1]], 41 | [-1, 3, BottleneckCSP, [256, False]], 42 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 22 (P3/8-small) 43 | 44 | [[], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 45 | ] 46 | -------------------------------------------------------------------------------- /models/yolov5s.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # yolov5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 1-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 2-P2/4 17 | [-1, 3, Bottleneck, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 4-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 6-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 6, BottleneckCSP, [1024]], # 10 25 | ] 26 | 27 | # yolov5 head 28 | head: 29 | [[-1, 3, BottleneckCSP, [1024, False]], # 11 30 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 12 (P5/32-large) 31 | 32 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 33 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 34 | [-1, 1, Conv, [512, 1, 1]], 35 | [-1, 3, BottleneckCSP, [512, False]], 36 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 17 (P4/16-medium) 37 | 38 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 39 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 40 | [-1, 1, Conv, [256, 1, 1]], 41 | [-1, 3, BottleneckCSP, [256, False]], 42 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 22 (P3/8-small) 43 | 44 | [[], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 45 | ] 46 | -------------------------------------------------------------------------------- /models/yolov5x.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 1.33 # model depth multiple 4 | width_multiple: 1.25 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # yolov5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 1-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 2-P2/4 17 | [-1, 3, Bottleneck, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 4-P3/8 19 | [-1, 9, BottleneckCSP, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 6-P4/16 21 | [-1, 9, BottleneckCSP, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 8-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 6, BottleneckCSP, [1024]], # 10 25 | ] 26 | 27 | # yolov5 head 28 | head: 29 | [[-1, 3, BottleneckCSP, [1024, False]], # 11 30 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 12 (P5/32-large) 31 | 32 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 33 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 34 | [-1, 1, Conv, [512, 1, 1]], 35 | [-1, 3, BottleneckCSP, [512, False]], 36 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 17 (P4/16-medium) 37 | 38 | [-2, 1, nn.Upsample, [None, 2, 'nearest']], 39 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 40 | [-1, 1, Conv, [256, 1, 1]], 41 | [-1, 3, BottleneckCSP, [256, False]], 42 | [-1, 1, nn.Conv2d, [na * (nc + 5), 1, 1]], # 22 (P3/8-small) 43 | 44 | [[], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 45 | ] 46 | -------------------------------------------------------------------------------- /simple_inference.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import cv2 3 | import numpy as np 4 | from utils.utils import non_max_suppression, attempt_download 5 | import argparse 6 | 7 | 8 | parser = argparse.ArgumentParser() 9 | parser.add_argument('--weights', type=str, default='weights/yolov5x.pt', help='model.pt path') 10 | parser.add_argument('--image', type=str, default='inference/images/test.jpg', help='Input image') 11 | parser.add_argument('--output_dir', type=str, default='inference/output/', help='output directory') 12 | parser.add_argument('--thres', type=float, default=0.4, help='object confidence threshold') 13 | opt = parser.parse_args() 14 | 15 | 16 | ''' 17 | Class Labels 18 | Num : 80 19 | ''' 20 | 21 | classnames = ['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 | 32 | label = {} 33 | for i, name in enumerate(classnames): 34 | label[i]=name 35 | 36 | 37 | 38 | # load pre-trained model 39 | weights = opt.weights 40 | attempt_download(weights) 41 | 42 | # try: 43 | model = torch.load(weights)['model'].float() 44 | model.eval() 45 | # except: 46 | # print('[ERROR] check the model') 47 | 48 | 49 | def image_loader(img,imsize): 50 | ''' 51 | processes input image for inference 52 | ''' 53 | h, w = img.shape[:2] 54 | img = cv2.resize(img,(imsize,imsize)) 55 | img = img[:, :, ::-1].transpose(2, 0, 1) 56 | img = np.ascontiguousarray(img) 57 | img = torch.from_numpy(img) 58 | img = img.float() 59 | img /= 255.0 60 | img = img.unsqueeze(0) 61 | return img, h, w 62 | 63 | 64 | def get_pred(img): 65 | ''' 66 | returns prediction in numpy array 67 | ''' 68 | imsize = 640 69 | img, h, w = image_loader(img,imsize) 70 | pred = model(img)[0] 71 | pred = non_max_suppression(pred, conf_thres=opt.thres, fast=True) # conf_thres is confidence thresold 72 | if pred[0] is not None: 73 | gain = imsize / max(h,w) 74 | pad = (imsize - w * gain) / 2, (imsize - h * gain) / 2 # wh padding 75 | pred = pred[0] 76 | 77 | pred[:, [0, 2]] -= pad[0] # x padding 78 | pred[:, [1, 3]] -= pad[1] # y padding 79 | pred[:, :4] /= gain 80 | pred[:, 0].clamp_(0, w) # x1 81 | pred[:, 1].clamp_(0, h) # y1 82 | pred[:, 2].clamp_(0, w) # x2 83 | pred[:, 3].clamp_(0, h) # y2 84 | 85 | pred = pred.detach().numpy() 86 | 87 | return pred 88 | 89 | 90 | path = opt.image 91 | 92 | image = cv2.imread(path) 93 | 94 | if image is not None: 95 | prediction = get_pred(image) 96 | 97 | if prediction is not None: 98 | for pred in prediction: 99 | 100 | x1 = int(pred[0]) 101 | y1 = int(pred[1]) 102 | x2 = int(pred[2]) 103 | y2 = int(pred[3]) 104 | 105 | start = (x1,y1) 106 | end = (x2,y2) 107 | 108 | pred_data = f'{label[pred[-1]]} {str(pred[-2]*100)[:5]}%' 109 | print(pred_data) 110 | color = (0,255,0) 111 | image = cv2.rectangle(image, start, end, color) 112 | image = cv2.putText(image, pred_data, (x1,y1+25), cv2.FONT_HERSHEY_SIMPLEX, 1, color, 2, cv2.LINE_AA) 113 | cv2.imwrite(opt.output_dir+'result.jpg', image) 114 | 115 | else: 116 | print('[ERROR] check input image') 117 | 118 | 119 | 120 | 121 | 122 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Nannigalaxy/yolov5_pytorch_inference/caf9febaeb5266fbcc0cc992c8b73641c3bcf681/utils/__init__.py -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | import math 2 | import os 3 | import time 4 | from copy import deepcopy 5 | 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | import torchvision.models as models 11 | 12 | 13 | def init_seeds(seed=0): 14 | torch.manual_seed(seed) 15 | 16 | # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html 17 | if seed == 0: # slower, more reproducible 18 | cudnn.deterministic = True 19 | cudnn.benchmark = False 20 | else: # faster, less reproducible 21 | cudnn.deterministic = False 22 | cudnn.benchmark = True 23 | 24 | 25 | def select_device(device='', apex=False, batch_size=None): 26 | # device = 'cpu' or '0' or '0,1,2,3' 27 | cpu_request = device.lower() == 'cpu' 28 | if device and not cpu_request: # if device requested other than 'cpu' 29 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 30 | assert torch.cuda.is_available(), 'CUDA unavailable, invalid device %s requested' % device # check availablity 31 | 32 | cuda = False if cpu_request else torch.cuda.is_available() 33 | if cuda: 34 | c = 1024 ** 2 # bytes to MB 35 | ng = torch.cuda.device_count() 36 | if ng > 1 and batch_size: # check that batch_size is compatible with device_count 37 | assert batch_size % ng == 0, 'batch-size %g not multiple of GPU count %g' % (batch_size, ng) 38 | x = [torch.cuda.get_device_properties(i) for i in range(ng)] 39 | s = 'Using CUDA ' + ('Apex ' if apex else '') # apex for mixed precision https://github.com/NVIDIA/apex 40 | for i in range(0, ng): 41 | if i == 1: 42 | s = ' ' * len(s) 43 | print("%sdevice%g _CudaDeviceProperties(name='%s', total_memory=%dMB)" % 44 | (s, i, x[i].name, x[i].total_memory / c)) 45 | else: 46 | print('Using CPU') 47 | 48 | print('') # skip a line 49 | return torch.device('cuda:0' if cuda else 'cpu') 50 | 51 | 52 | def time_synchronized(): 53 | torch.cuda.synchronize() if torch.cuda.is_available() else None 54 | return time.time() 55 | 56 | 57 | def initialize_weights(model): 58 | for m in model.modules(): 59 | t = type(m) 60 | if t is nn.Conv2d: 61 | pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 62 | elif t is nn.BatchNorm2d: 63 | m.eps = 1e-4 64 | m.momentum = 0.03 65 | elif t in [nn.LeakyReLU, nn.ReLU, nn.ReLU6]: 66 | m.inplace = True 67 | 68 | 69 | def find_modules(model, mclass=nn.Conv2d): 70 | # finds layer indices matching module class 'mclass' 71 | return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] 72 | 73 | 74 | def fuse_conv_and_bn(conv, bn): 75 | # https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 76 | with torch.no_grad(): 77 | # init 78 | fusedconv = torch.nn.Conv2d(conv.in_channels, 79 | conv.out_channels, 80 | kernel_size=conv.kernel_size, 81 | stride=conv.stride, 82 | padding=conv.padding, 83 | bias=True) 84 | 85 | # prepare filters 86 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 87 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 88 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size())) 89 | 90 | # prepare spatial bias 91 | if conv.bias is not None: 92 | b_conv = conv.bias 93 | else: 94 | b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) 95 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 96 | fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) 97 | 98 | return fusedconv 99 | 100 | 101 | def model_info(model, verbose=False): 102 | # Plots a line-by-line description of a PyTorch model 103 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 104 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 105 | if verbose: 106 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 107 | for i, (name, p) in enumerate(model.named_parameters()): 108 | name = name.replace('module_list.', '') 109 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 110 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 111 | 112 | try: # FLOPS 113 | from thop import profile 114 | macs, _ = profile(model, inputs=(torch.zeros(1, 3, 480, 640),), verbose=False) 115 | fs = ', %.1f GFLOPS' % (macs / 1E9 * 2) 116 | except: 117 | fs = '' 118 | 119 | print('Model Summary: %g layers, %g parameters, %g gradients%s' % (len(list(model.parameters())), n_p, n_g, fs)) 120 | 121 | 122 | def load_classifier(name='resnet101', n=2): 123 | # Loads a pretrained model reshaped to n-class output 124 | model = models.__dict__[name](pretrained=True) 125 | 126 | # Display model properties 127 | input_size = [3, 224, 224] 128 | input_space = 'RGB' 129 | input_range = [0, 1] 130 | mean = [0.485, 0.456, 0.406] 131 | std = [0.229, 0.224, 0.225] 132 | for x in [input_size, input_space, input_range, mean, std]: 133 | print(x + ' =', eval(x)) 134 | 135 | # Reshape output to n classes 136 | filters = model.fc.weight.shape[1] 137 | model.fc.bias = torch.nn.Parameter(torch.zeros(n), requires_grad=True) 138 | model.fc.weight = torch.nn.Parameter(torch.zeros(n, filters), requires_grad=True) 139 | model.fc.out_features = n 140 | return model 141 | 142 | 143 | def scale_img(img, ratio=1.0, same_shape=False): # img(16,3,256,416), r=ratio 144 | # scales img(bs,3,y,x) by ratio 145 | h, w = img.shape[2:] 146 | s = (int(h * ratio), int(w * ratio)) # new size 147 | img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize 148 | if not same_shape: # pad/crop img 149 | gs = 32 # (pixels) grid size 150 | h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] 151 | return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean 152 | 153 | 154 | class ModelEMA: 155 | """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models 156 | Keep a moving average of everything in the model state_dict (parameters and buffers). 157 | This is intended to allow functionality like 158 | https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage 159 | A smoothed version of the weights is necessary for some training schemes to perform well. 160 | E.g. Google's hyper-params for training MNASNet, MobileNet-V3, EfficientNet, etc that use 161 | RMSprop with a short 2.4-3 epoch decay period and slow LR decay rate of .96-.99 requires EMA 162 | smoothing of weights to match results. Pay attention to the decay constant you are using 163 | relative to your update count per epoch. 164 | To keep EMA from using GPU resources, set device='cpu'. This will save a bit of memory but 165 | disable validation of the EMA weights. Validation will have to be done manually in a separate 166 | process, or after the training stops converging. 167 | This class is sensitive where it is initialized in the sequence of model init, 168 | GPU assignment and distributed training wrappers. 169 | I've tested with the sequence in my own train.py for torch.DataParallel, apex.DDP, and single-GPU. 170 | """ 171 | 172 | def __init__(self, model, decay=0.9999, device=''): 173 | # make a copy of the model for accumulating moving average of weights 174 | self.ema = deepcopy(model) 175 | self.ema.eval() 176 | self.updates = 0 # number of EMA updates 177 | self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) 178 | self.device = device # perform ema on different device from model if set 179 | if device: 180 | self.ema.to(device=device) 181 | for p in self.ema.parameters(): 182 | p.requires_grad_(False) 183 | 184 | def update(self, model): 185 | self.updates += 1 186 | d = self.decay(self.updates) 187 | with torch.no_grad(): 188 | if type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel): 189 | msd, esd = model.module.state_dict(), self.ema.module.state_dict() 190 | else: 191 | msd, esd = model.state_dict(), self.ema.state_dict() 192 | 193 | for k, v in esd.items(): 194 | if v.dtype.is_floating_point: 195 | v *= d 196 | v += (1. - d) * msd[k].detach() 197 | 198 | def update_attr(self, model): 199 | # Assign attributes (which may change during training) 200 | for k in model.__dict__.keys(): 201 | if not k.startswith('_'): 202 | setattr(self.ema, k, getattr(model, k)) 203 | -------------------------------------------------------------------------------- /utils/utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import random 3 | import subprocess 4 | import time 5 | from pathlib import Path 6 | 7 | import numpy as np 8 | import torch 9 | import torchvision 10 | import torch.nn as nn 11 | 12 | 13 | 14 | def xywh2xyxy(x): 15 | # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right 16 | y = torch.zeros_like(x) if isinstance(x, torch.Tensor) else np.zeros_like(x) 17 | y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x 18 | y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y 19 | y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x 20 | y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y 21 | return y 22 | 23 | 24 | 25 | def box_iou(box1, box2): 26 | # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py 27 | """ 28 | Return intersection-over-union (Jaccard index) of boxes. 29 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 30 | Arguments: 31 | box1 (Tensor[N, 4]) 32 | box2 (Tensor[M, 4]) 33 | Returns: 34 | iou (Tensor[N, M]): the NxM matrix containing the pairwise 35 | IoU values for every element in boxes1 and boxes2 36 | """ 37 | 38 | def box_area(box): 39 | # box = 4xn 40 | return (box[2] - box[0]) * (box[3] - box[1]) 41 | 42 | area1 = box_area(box1.t()) 43 | area2 = box_area(box2.t()) 44 | 45 | # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) 46 | inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2) 47 | return inter / (area1[:, None] + area2 - inter) # iou = inter / (area1 + area2 - inter) 48 | 49 | 50 | 51 | 52 | 53 | def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, fast=False, classes=None, agnostic=False): 54 | """Performs Non-Maximum Suppression (NMS) on inference results 55 | 56 | Returns: 57 | detections with shape: nx6 (x1, y1, x2, y2, conf, cls) 58 | """ 59 | if prediction.dtype is torch.float16: 60 | prediction = prediction.float() # to FP32 61 | 62 | nc = prediction[0].shape[1] - 5 # number of classes 63 | xc = prediction[..., 4] > conf_thres # candidates 64 | 65 | # Settings 66 | min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height 67 | max_det = 300 # maximum number of detections per image 68 | time_limit = 10.0 # seconds to quit after 69 | redundant = True # require redundant detections 70 | fast |= conf_thres > 0.001 # fast mode 71 | multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img) 72 | if fast: 73 | merge = False 74 | else: 75 | merge = True # merge for best mAP (adds 0.5ms/img) 76 | 77 | t = time.time() 78 | output = [None] * prediction.shape[0] 79 | for xi, x in enumerate(prediction): # image index, image inference 80 | # Apply constraints 81 | # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height 82 | x = x[xc[xi]] # confidence 83 | 84 | # If none remain process next image 85 | if not x.shape[0]: 86 | continue 87 | 88 | # Compute conf 89 | x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf 90 | 91 | # Box (center x, center y, width, height) to (x1, y1, x2, y2) 92 | box = xywh2xyxy(x[:, :4]) 93 | 94 | # Detections matrix nx6 (xyxy, conf, cls) 95 | if multi_label: 96 | i, j = (x[:, 5:] > conf_thres).nonzero().t() 97 | x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) 98 | else: # best class only 99 | conf, j = x[:, 5:].max(1, keepdim=True) 100 | x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] 101 | 102 | # Filter by class 103 | if classes: 104 | x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] 105 | 106 | # Apply finite constraint 107 | # if not torch.isfinite(x).all(): 108 | # x = x[torch.isfinite(x).all(1)] 109 | 110 | # If none remain process next image 111 | n = x.shape[0] # number of boxes 112 | if not n: 113 | continue 114 | 115 | # Sort by confidence 116 | # x = x[x[:, 4].argsort(descending=True)] 117 | 118 | # Batched NMS 119 | c = x[:, 5:6] * (0 if agnostic else max_wh) # classes 120 | boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores 121 | i = torchvision.ops.boxes.nms(boxes, scores, iou_thres) 122 | if i.shape[0] > max_det: # limit detections 123 | i = i[:max_det] 124 | if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) 125 | try: # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) 126 | iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix 127 | weights = iou * scores[None] # box weights 128 | x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes 129 | if redundant: 130 | i = i[iou.sum(1) > 1] # require redundancy 131 | except: # possible CUDA error https://github.com/ultralytics/yolov3/issues/1139 132 | print(x, i, x.shape, i.shape) 133 | pass 134 | 135 | output[xi] = x[i] 136 | if (time.time() - t) > time_limit: 137 | break # time limit exceeded 138 | 139 | return output 140 | 141 | 142 | def attempt_download(weights): 143 | # Attempt to download pretrained weights if not found locally 144 | weights = weights.strip() 145 | msg = weights + ' missing, try downloading from https://drive.google.com/drive/folders/1Drs_Aiu7xx6S-ix95f9kNsA6ueKRpN2J' 146 | 147 | r = 1 148 | if len(weights) > 0 and not os.path.isfile(weights): 149 | d = {'yolov3-spp.pt': '1mM67oNw4fZoIOL1c8M3hHmj66d8e-ni_', # yolov3-spp.yaml 150 | 'yolov5s.pt': '1R5T6rIyy3lLwgFXNms8whc-387H0tMQO', # yolov5s.yaml 151 | 'yolov5m.pt': '1vobuEExpWQVpXExsJ2w-Mbf3HJjWkQJr', # yolov5m.yaml 152 | 'yolov5l.pt': '1hrlqD1Wdei7UT4OgT785BEk1JwnSvNEV', # yolov5l.yaml 153 | 'yolov5x.pt': '1mM8aZJlWTxOg7BZJvNUMrTnA2AbeCVzS', # yolov5x.yaml 154 | } 155 | 156 | file = Path(weights).name 157 | if file in d: 158 | r = gdrive_download(id=d[file], name=weights) 159 | 160 | if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6): # weights exist and > 1MB 161 | os.remove(weights) if os.path.exists(weights) else None # remove partial downloads 162 | s = "curl -L -o %s 'https://storage.googleapis.com/ultralytics/yolov5/ckpt/%s'" % (weights, file) 163 | r = os.system(s) # execute, capture return values 164 | 165 | # Error check 166 | if not (r == 0 and os.path.exists(weights) and os.path.getsize(weights) > 1E6): # weights exist and > 1MB 167 | os.remove(weights) if os.path.exists(weights) else None # remove partial downloads 168 | raise Exception(msg) 169 | 170 | 171 | def gdrive_download(id='1HaXkef9z6y5l4vUnCYgdmEAj61c6bfWO', name='coco.zip'): 172 | # https://gist.github.com/tanaikech/f0f2d122e05bf5f971611258c22c110f 173 | # Downloads a file from Google Drive, accepting presented query 174 | # from utils.google_utils import *; gdrive_download() 175 | t = time.time() 176 | 177 | print('Downloading https://drive.google.com/uc?export=download&id=%s as %s... ' % (id, name), end='') 178 | os.remove(name) if os.path.exists(name) else None # remove existing 179 | os.remove('cookie') if os.path.exists('cookie') else None 180 | 181 | # Attempt file download 182 | os.system("curl -c ./cookie -s -L \"https://drive.google.com/uc?export=download&id=%s\" > /dev/null" % id) 183 | if os.path.exists('cookie'): # large file 184 | s = "curl -Lb ./cookie \"https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=%s\" -o %s" % ( 185 | id, name) 186 | else: # small file 187 | s = "curl -s -L -o %s 'https://drive.google.com/uc?export=download&id=%s'" % (name, id) 188 | r = os.system(s) # execute, capture return values 189 | os.remove('cookie') if os.path.exists('cookie') else None 190 | 191 | # Error check 192 | if r != 0: 193 | os.remove(name) if os.path.exists(name) else None # remove partial 194 | print('Download error ') # raise Exception('Download error') 195 | return r 196 | 197 | # Unzip if archive 198 | if name.endswith('.zip'): 199 | print('unzipping... ', end='') 200 | os.system('unzip -q %s' % name) # unzip 201 | os.remove(name) # remove zip to free space 202 | 203 | print('Done (%.1fs)' % (time.time() - t)) 204 | return r -------------------------------------------------------------------------------- /weights/download_weights.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Download common models 3 | 4 | python3 -c "from utils.google_utils import *; 5 | attempt_download('weights/yolov5s.pt'); 6 | attempt_download('weights/yolov5m.pt'); 7 | attempt_download('weights/yolov5l.pt'); 8 | attempt_download('weights/yolov5x.pt')" 9 | --------------------------------------------------------------------------------