├── .gitignore ├── Images ├── VAME_Logo-1.png ├── behavior_structure_crop.gif ├── cuttree_vis.jpg └── workflow.png ├── LICENSE ├── README.md ├── VAME.yaml ├── examples ├── convert_maDLC_csv_individual_csv.py ├── demo.py └── video-1.csv ├── reinstall.sh ├── setup.py └── vame ├── __init__.py ├── analysis ├── __init__.py ├── community_analysis.py ├── generative_functions.py ├── gif_creator.py ├── pose_segmentation.py ├── segment_behavior.py ├── tree_hierarchy.py ├── umap_visualization.py └── videowriter.py ├── initialize_project ├── __init__.py └── new.py ├── model ├── __init__.py ├── create_training.py ├── dataloader.py ├── evaluate.py ├── rnn_model.py └── rnn_vae.py └── util ├── __init__.py ├── align_egocentrical.py ├── auxiliary.py ├── csv_to_npy.py └── gif_pose_helper.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | env/ 12 | build/ 13 | develop-eggs/ 14 | dist/ 15 | downloads/ 16 | eggs/ 17 | .eggs/ 18 | #lib/ 19 | lib64/ 20 | parts/ 21 | sdist/ 22 | var/ 23 | wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | 28 | # PyInstaller 29 | # Usually these files are written by a python script from a template 30 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 31 | *.manifest 32 | *.spec 33 | 34 | # Installer logs 35 | pip-log.txt 36 | pip-delete-this-directory.txt 37 | 38 | # Unit test / coverage reports 39 | htmlcov/ 40 | .tox/ 41 | .coverage 42 | .coverage.* 43 | .cache 44 | nosetests.xml 45 | coverage.xml 46 | *.cover 47 | .hypothesis/ 48 | 49 | # Translations 50 | *.mo 51 | *.pot 52 | 53 | # Django stuff: 54 | *.log*.ckpt 55 | snapshot-* 56 | 57 | # Byte-compiled / optimized / DLL files 58 | __pycache__/ 59 | *.py[cod] 60 | local_settings.py 61 | 62 | # Flask stuff: 63 | instance/ 64 | .webassets-cache 65 | 66 | # Scrapy stuff: 67 | .scrapy 68 | 69 | # Sphinx documentation 70 | docs/_build/ 71 | 72 | # PyBuilder 73 | target/ 74 | 75 | # Jupyter Notebook 76 | .ipynb_checkpoints 77 | 78 | # pyenv 79 | .python-version 80 | 81 | # celery beat schedule file 82 | celerybeat-schedule 83 | 84 | # SageMath parsed files 85 | *.sage.py 86 | 87 | # dotenv 88 | .env 89 | 90 | # virtualenv 91 | .venv 92 | venv/ 93 | ENV/ 94 | 95 | # Spyder project settings 96 | .spyderproject 97 | .spyproject 98 | 99 | # Rope project settings 100 | .ropeproject 101 | 102 | # mkdocs documentation 103 | /site 104 | 105 | # mypy 106 | .mypy_cache/ 107 | -------------------------------------------------------------------------------- /Images/VAME_Logo-1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LINCellularNeuroscience/VAME/b97a5180b7c1b78823f9c6ec55853e32f2bd16e5/Images/VAME_Logo-1.png -------------------------------------------------------------------------------- /Images/behavior_structure_crop.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LINCellularNeuroscience/VAME/b97a5180b7c1b78823f9c6ec55853e32f2bd16e5/Images/behavior_structure_crop.gif -------------------------------------------------------------------------------- /Images/cuttree_vis.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LINCellularNeuroscience/VAME/b97a5180b7c1b78823f9c6ec55853e32f2bd16e5/Images/cuttree_vis.jpg -------------------------------------------------------------------------------- /Images/workflow.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/LINCellularNeuroscience/VAME/b97a5180b7c1b78823f9c6ec55853e32f2bd16e5/Images/workflow.png -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ![VAME](https://github.com/LINCellularNeuroscience/VAME/blob/master/Images/VAME_Logo-1.png) 2 | ![workflow](https://github.com/LINCellularNeuroscience/VAME/blob/master/Images/workflow.png) 3 | 4 | # New maintained VAME repository 5 | 6 | This version of VAME is deprecated and no longer maintained, and is made available here as legacy code. VAME is now being maintained at its new home at [https://github.com/EthoML/VAME](https://github.com/EthoML/VAME). There, you will find updated documentation and additional packages. Users can also access a downloadable desktop app for VAME at [https://github.com/EthoML/vame-desktop](https://github.com/EthoML/vame-desktop). 7 | 8 | # VAME in a Nutshell 9 | VAME is a framework to cluster behavioral signals obtained from pose-estimation tools. It is a [PyTorch](https://pytorch.org/) based deep learning framework which leverages the power of recurrent neural networks (RNN) to model sequential data. In order to learn the underlying complex data distribution we use the RNN in a variational autoencoder setting to extract the latent state of the animal in every step of the input time series. 10 | 11 | ![behavior](https://github.com/LINCellularNeuroscience/VAME/blob/master/Images/behavior_structure_crop.gif) 12 | 13 | The workflow of VAME consists of 5 steps and we explain them in detail [here](https://github.com/LINCellularNeuroscience/VAME/wiki/1.-VAME-Workflow). 14 | 15 | ## Installation 16 | To get started we recommend using [Anaconda](https://www.anaconda.com/distribution/) with Python 3.6 or higher. 17 | Here, you can create a [virtual enviroment](https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html) to store all the dependencies necessary for VAME. (you can also use the VAME.yaml file supplied here, byt simply openning the terminal, running `git clone https://github.com/LINCellularNeuroscience/VAME.git`, then type `cd VAME` then run: `conda env create -f VAME.yaml`). 18 | 19 | * Go to the locally cloned VAME directory and run `python setup.py install` in order to install VAME in your active conda environment. 20 | * Install the current stable Pytorch release using the OS-dependent instructions from the [Pytorch website](https://pytorch.org/get-started/locally/). Currently, VAME is tested on PyTorch 1.5. (Note, if you use the conda file we supply, PyTorch is already installed and you don't need to do this step.) 21 | 22 | ## Getting Started 23 | First, you should make sure that you have a GPU powerful enough to train deep learning networks. In our paper, we were using a single Nvidia GTX 1080 Ti GPU to train our network. A hardware guide can be found [here](https://timdettmers.com/2018/12/16/deep-learning-hardware-guide/). Once you have your hardware ready, try VAME following the [workflow guide](https://github.com/LINCellularNeuroscience/VAME/wiki/1.-VAME-Workflow). 24 | 25 | If you want to follow an example first you can download [video-1](https://drive.google.com/file/d/1w6OW9cN_-S30B7rOANvSaR9c3O5KeF0c/view?usp=sharing) here and find the .csv file in our [example](https://github.com/LINCellularNeuroscience/VAME/tree/master/examples) folder. 26 | 27 | ## News 28 | * November 2022: Finally the VAME paper is published! Check it out [on the publisher werbsite](https://www.nature.com/articles/s42003-022-04080-7). In comparison to the preprint version, there is also a practical workflow guide included with many useful instructions on how to use VAME. 29 | * March 2021: We are happy to release VAME 1.0 with a bunch of improvements and new features! These include the community analysis script, a model allowing generation of unseen datapoints, new visualization functions, as well as the much requested function to generate GIF sequences containing UMAP embeddings and trajectories together with the video of the behaving animal. Big thanks also to [@MMathisLab](https://github.com/MMathisLab) for contributing to the OS compatibility and usability of our code. 30 | * November 2020: We uploaded an egocentric alignment [script](https://github.com/LINCellularNeuroscience/VAME/blob/master/examples/align_demo.py) to allow more researcher to use VAME 31 | * October 2020: We updated our manuscript on [Biorxiv](https://www.biorxiv.org/content/10.1101/2020.05.14.095430v2) 32 | * May 2020: Our preprint "Identifying Behavioral Structure from Deep Variational Embeddings of Animal Motion" is out! [Read it on Biorxiv!](https://www.biorxiv.org/content/10.1101/2020.05.14.095430v1) 33 | 34 | ### Authors and Code Contributors 35 | VAME was developed by Kevin Luxem and Pavol Bauer. 36 | 37 | The development of VAME is heavily inspired by [DeepLabCut](https://github.com/DeepLabCut/DeepLabCut/). 38 | As such, the VAME project management codebase has been adapted from the DeepLabCut codebase. 39 | The DeepLabCut 2.0 toolbox is © A. & M.W. Mathis Labs [deeplabcut.org](http:\\deeplabcut.org), released under LGPL v3.0. 40 | The implementation of the VRAE model is partially adapted from the [Timeseries clustering](https://github.com/tejaslodaya/timeseries-clustering-vae) repository developed by [Tejas Lodaya](https://tejaslodaya.com). 41 | 42 | ### References 43 | VAME preprint: [Identifying Behavioral Structure from Deep Variational Embeddings of Animal Motion](https://www.biorxiv.org/content/10.1101/2020.05.14.095430v2)
44 | Kingma & Welling: [Auto-Encoding Variational Bayes](https://arxiv.org/abs/1312.6114)
45 | Pereira & Silveira: [Learning Representations from Healthcare Time Series Data for Unsupervised Anomaly Detection](https://www.joao-pereira.pt/publications/accepted_version_BigComp19.pdf) 46 | 47 | ### License: GPLv3 48 | See the [LICENSE file](../master/LICENSE) for the full statement. 49 | 50 | ### Code Reference (DOI) 51 | [![DOI](https://zenodo.org/badge/254593619.svg)](https://zenodo.org/badge/latestdoi/254593619) 52 | -------------------------------------------------------------------------------- /VAME.yaml: -------------------------------------------------------------------------------- 1 | # VAME.yaml 2 | # 3 | # install: conda env create -f VAME.yaml 4 | # update: conda env update -f VAME.yaml 5 | name: VAME 6 | channels: 7 | - pytorch 8 | - defaults 9 | dependencies: 10 | - python=3.7 11 | - pip 12 | - torchvision 13 | - jupyter 14 | - nb_conda 15 | - pip: 16 | - pytest-shutil 17 | - scipy 18 | - numpy 19 | - matplotlib 20 | - pathlib 21 | - pandas 22 | - ruamel.yaml 23 | - sklearn 24 | - pyyaml 25 | - opencv-python-headless 26 | - h5py 27 | - umap-learn 28 | - networkx 29 | - tqdm 30 | - hmmlearn 31 | -------------------------------------------------------------------------------- /examples/convert_maDLC_csv_individual_csv.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | Created on Mon Feb 7 10:20:44 2022 4 | 5 | @author: Charitha Omprakash, LIN Magdeburg, charitha.omprakash@lin-magdeburg.de 6 | 7 | This file converts a multi-animal DLC CSV to several single animal DLC files. 8 | Those can be used as input to run VAME. 9 | """ 10 | 11 | import pandas, numpy as pd, np 12 | import os 13 | import glob 14 | from pathlib import Path 15 | 16 | def convert_multi_csv_to_individual_csv(csv_files_path): 17 | csvs = sorted(glob.glob(os.path.join(csv_files_path, '*.csv*'))) 18 | 19 | for csv in csvs: 20 | fname = pd.read_csv(csv, header=[0,1,2], index_col=0, skiprows=1) 21 | individuals = fname.columns.get_level_values('individuals').unique() 22 | for ind in individuals: 23 | fname_temp = fname[ind] 24 | fname_temp_path = os.path.splitext(csv)[0] + '_' + ind + '.csv' 25 | fname_temp.to_csv(fname_temp_path, index=True, header=True) 26 | -------------------------------------------------------------------------------- /examples/demo.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import vame 13 | 14 | # These paths have to be set manually 15 | working_directory = '/YOUR/WORKING/DIRECTORY/' 16 | project='Your-VAME-Project' 17 | videos = ['/directory/to/your/video-1','/directory/to/your/video-2','...'] 18 | 19 | # Initialize your project 20 | # Step 1.1: 21 | config = vame.init_new_project(project=project, videos=videos, working_directory=working_directory, videotype='.mp4') 22 | 23 | # After the inital creation of your project you can always access the config.yaml file 24 | # via specifying the path to your project 25 | config = '/YOUR/WORKING/DIRECTORY/Your-VAME-Project-Apr14-2020/config.yaml' 26 | 27 | # As our config.yaml is sometimes still changing a little due to updates, we have here a small function 28 | # to update your config.yaml to the current state. Be aware that this will overwrite your current config.yaml 29 | # and make sure to back up your version if you did parameter changes! 30 | vame.update_config(config) 31 | 32 | # Step 1.2: 33 | # Align your behavior videos egocentric and create training dataset: 34 | # pose_ref_index: list of reference coordinate indices for alignment 35 | # Example: 0: snout, 1: forehand_left, 2: forehand_right, 3: hindleft, 4: hindright, 5: tail 36 | vame.egocentric_alignment(config, pose_ref_index=[0,5]) 37 | 38 | # If your experiment is by design egocentrical (e.g. head-fixed experiment on treadmill etc) 39 | # you can use the following to convert your .csv to a .npy array, ready to train vame on it 40 | vame.csv_to_numpy(config) 41 | 42 | # Step 1.3: 43 | # create the training set for the VAME model 44 | vame.create_trainset(config, check_parameter=False) 45 | 46 | # Step 2: 47 | # Train VAME: 48 | vame.train_model(config) 49 | 50 | # Step 3: 51 | # Evaluate model 52 | vame.evaluate_model(config) 53 | 54 | # Step 4: 55 | # Segment motifs/pose 56 | vame.pose_segmentation(config) 57 | 58 | 59 | #------------------------------------------------------------------------------ 60 | #------------------------------------------------------------------------------ 61 | # The following are optional choices to create motif videos, communities/hierarchies of behavior, 62 | # community videos 63 | 64 | # OPTIONIAL: Create motif videos to get insights about the fine grained poses 65 | vame.motif_videos(config, videoType='.mp4') 66 | 67 | # OPTIONAL: Create behavioural hierarchies via community detection 68 | vame.community(config, show_umap=False, cut_tree=2) 69 | 70 | # OPTIONAL: Create community videos to get insights about behavior on a hierarchical scale 71 | vame.community_videos(config) 72 | 73 | # OPTIONAL: Down projection of latent vectors and visualization via UMAP 74 | vame.visualization(config, label=None) #options: label: None, "motif", "community" 75 | 76 | # OPTIONAL: Use the generative model (reconstruction decoder) to sample from 77 | # the learned data distribution, reconstruct random real samples or visualize 78 | # the cluster center for validation 79 | vame.generative_model(config, mode="centers") #options: mode: "sampling", "reconstruction", "centers", "motifs" 80 | 81 | # OPTIONAL: Create a video of an egocentrically aligned mouse + path through 82 | # the community space (similar to our gif on github) to learn more about your representation 83 | # and have something cool to show around ;) 84 | # Note: This function is currently very slow. Once the frames are saved you can create a video 85 | # or gif via e.g. ImageJ or other tools 86 | vame.gif(config, pose_ref_index=[0,5], subtract_background=True, start=None, 87 | length=500, max_lag=30, label='community', file_format='.mp4', crop_size=(300,300)) 88 | -------------------------------------------------------------------------------- /reinstall.sh: -------------------------------------------------------------------------------- 1 | pip uninstall vame 2 | python3 setup.py sdist bdist_wheel 3 | pip install dist/vame-1.0-py3-none-any.whl 4 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | from setuptools import setup, find_packages 2 | 3 | with open("README.md", "r") as fh: 4 | long_description = fh.read() 5 | 6 | setup( 7 | name="vame", 8 | version='1.0', 9 | packages=find_packages(), 10 | entry_points={"console_scripts": "vame = vame:main"}, 11 | author="K. Luxem & P. Bauer", 12 | description="Variational Animal Motion Embedding.", 13 | long_description=long_description, 14 | long_description_content_type="text/markdown", 15 | url="https://https://github.com/LINCellularNeuroscience/VAME/", 16 | setup_requires=[ 17 | "pytest", 18 | ], 19 | install_requires=[ 20 | "pytest-shutil", 21 | "scipy", 22 | "numpy", 23 | "matplotlib", 24 | "pathlib", 25 | "pandas", 26 | "ruamel.yaml", 27 | "sklearn", 28 | "pyyaml", 29 | "opencv-python", 30 | ], 31 | ) 32 | -------------------------------------------------------------------------------- /vame/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | import sys 12 | sys.dont_write_bytecode = True 13 | 14 | from vame.initialize_project import init_new_project 15 | from vame.model import create_trainset 16 | from vame.model import train_model 17 | from vame.model import evaluate_model 18 | from vame.analysis import pose_segmentation 19 | from vame.analysis import motif_videos 20 | from vame.analysis import community 21 | from vame.analysis import community_videos 22 | from vame.analysis import visualization 23 | from vame.analysis import generative_model 24 | from vame.analysis import gif 25 | from vame.util.csv_to_npy import csv_to_numpy 26 | from vame.util.align_egocentrical import egocentric_alignment 27 | from vame.util import auxiliary 28 | from vame.util.auxiliary import update_config 29 | 30 | -------------------------------------------------------------------------------- /vame/analysis/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | import sys 12 | sys.dont_write_bytecode = True 13 | 14 | from vame.analysis.pose_segmentation import pose_segmentation 15 | from vame.analysis.videowriter import motif_videos, community_videos 16 | from vame.analysis.community_analysis import community 17 | from vame.analysis.umap_visualization import visualization 18 | from vame.analysis.generative_functions import generative_model 19 | from vame.analysis.gif_creator import gif 20 | 21 | -------------------------------------------------------------------------------- /vame/analysis/community_analysis.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import umap 14 | import scipy 15 | import pickle 16 | import numpy as np 17 | from pathlib import Path 18 | import matplotlib.pyplot as plt 19 | 20 | from vame.util.auxiliary import read_config 21 | from vame.analysis.tree_hierarchy import graph_to_tree, draw_tree, traverse_tree_cutline 22 | 23 | 24 | def get_adjacency_matrix(labels, n_cluster): 25 | temp_matrix = np.zeros((n_cluster,n_cluster), dtype=np.float64) 26 | adjacency_matrix = np.zeros((n_cluster,n_cluster), dtype=np.float64) 27 | cntMat = np.zeros((n_cluster)) 28 | steps = len(labels) 29 | 30 | for i in range(n_cluster): 31 | for k in range(steps-1): 32 | idx = labels[k] 33 | if idx == i: 34 | idx2 = labels[k+1] 35 | if idx == idx2: 36 | continue 37 | else: 38 | cntMat[idx2] = cntMat[idx2] +1 39 | temp_matrix[i] = cntMat 40 | cntMat = np.zeros((n_cluster)) 41 | 42 | for k in range(steps-1): 43 | idx = labels[k] 44 | idx2 = labels[k+1] 45 | if idx == idx2: 46 | continue 47 | adjacency_matrix[idx,idx2] = 1 48 | adjacency_matrix[idx2,idx] = 1 49 | 50 | transition_matrix = get_transition_matrix(temp_matrix) 51 | 52 | return adjacency_matrix, transition_matrix, temp_matrix 53 | 54 | 55 | def get_transition_matrix(adjacency_matrix, threshold = 0.0): 56 | row_sum=adjacency_matrix.sum(axis=1) 57 | transition_matrix = adjacency_matrix/row_sum[:,np.newaxis] 58 | transition_matrix[transition_matrix <= threshold] = 0 59 | if np.any(np.isnan(transition_matrix)): 60 | transition_matrix=np.nan_to_num(transition_matrix) 61 | return transition_matrix 62 | 63 | 64 | def get_labels(cfg, files, model_name, n_cluster,parameterization): 65 | labels = [] 66 | for file in files: 67 | path_to_file = os.path.join(cfg['project_path'],"results",file,model_name,parameterization+'-'+str(n_cluster),"") 68 | label = np.load(os.path.join(path_to_file,str(n_cluster)+'_km_label_'+file+'.npy')) 69 | labels.append(label) 70 | return labels 71 | 72 | def compute_transition_matrices(files, labels, n_cluster): 73 | transition_matrices = [] 74 | for i, file in enumerate(files): 75 | adj, trans, mat = get_adjacency_matrix(labels[i], n_cluster) 76 | transition_matrices.append(trans) 77 | return transition_matrices 78 | 79 | 80 | def create_community_bag(files, labels, transition_matrices, cut_tree, n_cluster): 81 | # markov chain to tree -> community detection 82 | trees = [] 83 | communities_all = [] 84 | for i, file in enumerate(files): 85 | _, usage = np.unique(labels[i], return_counts=True) 86 | T = graph_to_tree(usage, transition_matrices[i], n_cluster, merge_sel=1) 87 | trees.append(T) 88 | 89 | if cut_tree != None: 90 | community_bag = traverse_tree_cutline(T,cutline=cut_tree) 91 | communities_all.append(community_bag) 92 | draw_tree(T) 93 | else: 94 | draw_tree(T) 95 | plt.pause(0.5) 96 | flag_1 = 'no' 97 | while flag_1 == 'no': 98 | cutline = int(input("Where do you want to cut the Tree? 0/1/2/3/...")) 99 | community_bag = traverse_tree_cutline(T,cutline=cutline) 100 | print(community_bag) 101 | flag_2 = input('\nAre all motifs in the list? (yes/no/restart)') 102 | if flag_2 == 'no': 103 | while flag_2 == 'no': 104 | add = input('Extend list or add in the end? (ext/end)') 105 | if add == "ext": 106 | motif_idx = int(input('Which motif number? ')) 107 | list_idx = int(input('At which position in the list? (pythonic indexing starts at 0) ')) 108 | community_bag[list_idx].append(motif_idx) 109 | if add == "end": 110 | motif_idx = int(input('Which motif number? ')) 111 | community_bag.append([motif_idx]) 112 | print(community_bag) 113 | flag_2 = input('\nAre all motifs in the list? (yes/no/restart)') 114 | if flag_2 == "restart": 115 | continue 116 | if flag_2 == 'yes': 117 | communities_all.append(community_bag) 118 | flag_1 = 'yes' 119 | 120 | return communities_all, trees 121 | 122 | 123 | def get_community_labels(files, labels, communities_all): 124 | # transform parameterized latent vector into communities 125 | community_labels_all = [] 126 | for k, file in enumerate(files): 127 | num_comm = len(communities_all[k]) 128 | 129 | community_labels = np.zeros_like(labels[k]) 130 | for i in range(num_comm): 131 | clust = np.array(communities_all[k][i]) 132 | for j in range(len(clust)): 133 | find_clust = np.where(labels[k] == clust[j])[0] 134 | community_labels[find_clust] = i 135 | 136 | community_labels = np.int64(scipy.signal.medfilt(community_labels, 7)) 137 | community_labels_all.append(community_labels) 138 | 139 | return community_labels_all 140 | 141 | 142 | def umap_embedding(cfg, file, model_name, n_cluster,parameterization): 143 | reducer = umap.UMAP(n_components=2, min_dist=cfg['min_dist'], n_neighbors=cfg['n_neighbors'], 144 | random_state=cfg['random_state']) 145 | 146 | print("UMAP calculation for file %s" %file) 147 | 148 | folder = os.path.join(cfg['project_path'],"results",file,model_name,parameterization+'-'+str(n_cluster),"") 149 | latent_vector = np.load(os.path.join(folder,'latent_vector_'+file+'.npy')) 150 | 151 | num_points = cfg['num_points'] 152 | if num_points > latent_vector.shape[0]: 153 | num_points = latent_vector.shape[0] 154 | print("Embedding %d data points.." %num_points) 155 | 156 | embed = reducer.fit_transform(latent_vector[:num_points,:]) 157 | 158 | return embed 159 | 160 | 161 | def umap_vis(cfg, file, embed, community_labels_all): 162 | num_points = cfg['num_points'] 163 | if num_points > community_labels_all.shape[0]: 164 | num_points = community_labels_all.shape[0] 165 | print("Embedding %d data points.." %num_points) 166 | 167 | num = np.unique(community_labels_all) 168 | 169 | fig = plt.figure(1) 170 | plt.scatter(embed[:,0], embed[:,1], c=community_labels_all[:num_points], cmap='Spectral', s=2, alpha=1) 171 | plt.colorbar(boundaries=np.arange(np.max(num)+2)-0.5).set_ticks(np.arange(np.max(num)+1)) 172 | plt.gca().set_aspect('equal', 'datalim') 173 | plt.grid(False) 174 | 175 | 176 | def community(config, show_umap=False, cut_tree=None): 177 | config_file = Path(config).resolve() 178 | cfg = read_config(config_file) 179 | model_name = cfg['model_name'] 180 | n_cluster = cfg['n_cluster'] 181 | parameterization = cfg['parameterization'] 182 | 183 | files = [] 184 | if cfg['all_data'] == 'No': 185 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 186 | "If you only want to use a specific dataset type filename: \n" 187 | "yes/no/filename ") 188 | else: 189 | all_flag = 'yes' 190 | 191 | if all_flag == 'yes' or all_flag == 'Yes': 192 | for file in cfg['video_sets']: 193 | files.append(file) 194 | 195 | elif all_flag == 'no' or all_flag == 'No': 196 | for file in cfg['video_sets']: 197 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 198 | if use_file == 'yes': 199 | files.append(file) 200 | if use_file == 'no': 201 | continue 202 | else: 203 | files.append(all_flag) 204 | 205 | labels = get_labels(cfg, files, model_name, n_cluster,parameterization) 206 | transition_matrices = compute_transition_matrices(files, labels, n_cluster) 207 | communities_all, trees = create_community_bag(files, labels, transition_matrices, cut_tree, n_cluster) 208 | community_labels_all = get_community_labels(files, labels, communities_all) 209 | 210 | for idx, file in enumerate(files): 211 | path_to_file=os.path.join(cfg['project_path'],"results",file,model_name,parameterization+'-'+str(n_cluster),"") 212 | if not os.path.exists(os.path.join(path_to_file,"community")): 213 | os.mkdir(os.path.join(path_to_file,"community")) 214 | 215 | np.save(os.path.join(path_to_file,"community","transition_matrix_"+file+'.npy'),transition_matrices[idx]) 216 | np.save(os.path.join(path_to_file,"community","community_label_"+file+'.npy'), community_labels_all[idx]) 217 | 218 | with open(os.path.join(path_to_file,"community","hierarchy"+file+".pkl"), "wb") as fp: #Pickling 219 | pickle.dump(communities_all[idx], fp) 220 | 221 | if show_umap == True: 222 | embed = umap_embedding(cfg, file, model_name, n_cluster,parameterization) 223 | umap_vis(cfg, files, embed, community_labels_all[idx]) 224 | 225 | # with open(os.path.join(path_to_file,"community","","hierarchy"+file+".txt"), "rb") as fp: # Unpickling 226 | # b = pickle.load(fp) 227 | 228 | 229 | 230 | 231 | 232 | -------------------------------------------------------------------------------- /vame/analysis/generative_functions.py: -------------------------------------------------------------------------------- 1 | """ 2 | Variational Animal Motion Embedding 1.0-alpha Toolbox 3 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 4 | Leibniz Institute for Neurobiology, Magdeburg, Germany 5 | 6 | https://github.com/LINCellularNeuroscience/VAME 7 | Licensed under GNU General Public License v3.0 8 | """ 9 | 10 | import os 11 | 12 | import torch 13 | import numpy as np 14 | from pathlib import Path 15 | import matplotlib.pyplot as plt 16 | from sklearn.mixture import GaussianMixture 17 | 18 | from vame.util.auxiliary import read_config 19 | from vame.model.rnn_model import RNN_VAE 20 | 21 | 22 | def random_generative_samples_motif(cfg, model, latent_vector,labels,n_cluster): 23 | # Latent sampling and generative model 24 | time_window = cfg['time_window'] 25 | for j in range(n_cluster): 26 | 27 | inds=np.where(labels==j) 28 | motif_latents=latent_vector[inds[0],:] 29 | gm = GaussianMixture(n_components=10).fit(motif_latents) 30 | 31 | # draw sample from GMM 32 | density_sample = gm.sample(10) 33 | 34 | # generate image via model decoder 35 | tensor_sample = torch.from_numpy(density_sample[0]).type('torch.FloatTensor').cuda() 36 | decoder_inputs = tensor_sample.unsqueeze(2).repeat(1, 1, time_window) 37 | decoder_inputs = decoder_inputs.permute(0,2,1) 38 | 39 | image_sample = model.decoder(decoder_inputs, tensor_sample) 40 | recon_sample = image_sample.cpu().detach().numpy() 41 | 42 | 43 | fig, axs = plt.subplots(2,5) 44 | for i in range(5): 45 | axs[0,i].plot(recon_sample[i,...]) 46 | axs[1,i].plot(recon_sample[i+5,...]) 47 | plt.suptitle('Generated samples for motif '+str(j)) 48 | 49 | def random_generative_samples(cfg, model, latent_vector): 50 | # Latent sampling and generative model 51 | time_window = cfg['time_window'] 52 | gm = GaussianMixture(n_components=10).fit(latent_vector) 53 | 54 | # draw sample from GMM 55 | density_sample = gm.sample(10) 56 | 57 | # generate image via model decoder 58 | tensor_sample = torch.from_numpy(density_sample[0]).type('torch.FloatTensor').cuda() 59 | decoder_inputs = tensor_sample.unsqueeze(2).repeat(1, 1, time_window) 60 | decoder_inputs = decoder_inputs.permute(0,2,1) 61 | 62 | image_sample = model.decoder(decoder_inputs, tensor_sample) 63 | recon_sample = image_sample.cpu().detach().numpy() 64 | 65 | fig, axs = plt.subplots(2,5) 66 | for i in range(5): 67 | axs[0,i].plot(recon_sample[i,...]) 68 | axs[1,i].plot(recon_sample[i+5,...]) 69 | plt.suptitle('Generated samples') 70 | 71 | 72 | def random_reconstruction_samples(cfg, model, latent_vector): 73 | # random samples for reconstruction 74 | time_window = cfg['time_window'] 75 | 76 | rnd = np.random.choice(latent_vector.shape[0], 10) 77 | tensor_sample = torch.from_numpy(latent_vector[rnd]).type('torch.FloatTensor').cuda() 78 | decoder_inputs = tensor_sample.unsqueeze(2).repeat(1, 1, time_window) 79 | decoder_inputs = decoder_inputs.permute(0,2,1) 80 | 81 | image_sample = model.decoder(decoder_inputs, tensor_sample) 82 | recon_sample = image_sample.cpu().detach().numpy() 83 | 84 | fig, axs = plt.subplots(2,5) 85 | for i in range(5): 86 | axs[0,i].plot(recon_sample[i,...]) 87 | axs[1,i].plot(recon_sample[i+5,...]) 88 | plt.suptitle('Reconstructed samples') 89 | 90 | 91 | def visualize_cluster_center(cfg, model, cluster_center): 92 | #Cluster Center 93 | time_window = cfg['time_window'] 94 | animal_centers = cluster_center 95 | 96 | tensor_sample = torch.from_numpy(animal_centers).type('torch.FloatTensor').cuda() 97 | decoder_inputs = tensor_sample.unsqueeze(2).repeat(1, 1, time_window) 98 | decoder_inputs = decoder_inputs.permute(0,2,1) 99 | 100 | image_sample = model.decoder(decoder_inputs, tensor_sample) 101 | recon_sample = image_sample.cpu().detach().numpy() 102 | 103 | num = animal_centers.shape[0] 104 | b = int(np.ceil(num / 5)) 105 | 106 | fig, axs = plt.subplots(5,b) 107 | idx = 0 108 | for k in range(5): 109 | for i in range(b): 110 | axs[k,i].plot(recon_sample[idx,...]) 111 | axs[k,i].set_title("Cluster %d" %idx) 112 | idx +=1 113 | 114 | 115 | def load_model(cfg, model_name): 116 | # load Model 117 | ZDIMS = cfg['zdims'] 118 | FUTURE_DECODER = cfg['prediction_decoder'] 119 | TEMPORAL_WINDOW = cfg['time_window']*2 120 | FUTURE_STEPS = cfg['prediction_steps'] 121 | 122 | NUM_FEATURES = cfg['num_features'] 123 | NUM_FEATURES = NUM_FEATURES - 2 124 | 125 | hidden_size_layer_1 = cfg['hidden_size_layer_1'] 126 | hidden_size_layer_2 = cfg['hidden_size_layer_2'] 127 | hidden_size_rec = cfg['hidden_size_rec'] 128 | hidden_size_pred = cfg['hidden_size_pred'] 129 | dropout_encoder = cfg['dropout_encoder'] 130 | dropout_rec = cfg['dropout_rec'] 131 | dropout_pred = cfg['dropout_pred'] 132 | softplus = cfg['softplus'] 133 | 134 | print('Load model... ') 135 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 136 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 137 | dropout_rec, dropout_pred, softplus).cuda() 138 | 139 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],'model','best_model',model_name+'_'+cfg['Project']+'.pkl'))) 140 | model.eval() 141 | 142 | return model 143 | 144 | 145 | def generative_model(config, mode="sampling"): 146 | config_file = Path(config).resolve() 147 | cfg = read_config(config_file) 148 | model_name = cfg['model_name'] 149 | n_cluster = cfg['n_cluster'] 150 | 151 | files = [] 152 | if cfg['all_data'] == 'No': 153 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 154 | "If you only want to use a specific dataset type filename: \n" 155 | "yes/no/filename ") 156 | else: 157 | all_flag = 'yes' 158 | 159 | if all_flag == 'yes' or all_flag == 'Yes': 160 | for file in cfg['video_sets']: 161 | files.append(file) 162 | 163 | elif all_flag == 'no' or all_flag == 'No': 164 | for file in cfg['video_sets']: 165 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 166 | if use_file == 'yes': 167 | files.append(file) 168 | if use_file == 'no': 169 | continue 170 | else: 171 | files.append(all_flag) 172 | 173 | 174 | model = load_model(cfg, model_name) 175 | 176 | for file in files: 177 | path_to_file=os.path.join(cfg['project_path'],"results",file,model_name,'kmeans-'+str(n_cluster),"") 178 | 179 | if mode == "sampling": 180 | latent_vector = np.load(os.path.join(path_to_file,'latent_vector_'+file+'.npy')) 181 | random_generative_samples(cfg, model, latent_vector) 182 | 183 | if mode == "reconstruction": 184 | latent_vector = np.load(os.path.join(path_to_file,'latent_vector_'+file+'.npy')) 185 | random_reconstruction_samples(cfg, model, latent_vector) 186 | 187 | if mode == "centers": 188 | cluster_center = np.load(os.path.join(path_to_file,'cluster_center_'+file+'.npy')) 189 | visualize_cluster_center(cfg, model, cluster_center) 190 | 191 | if mode == "motifs": 192 | latent_vector = np.load(os.path.join(path_to_file,'latent_vector_'+file+'.npy')) 193 | labels = np.load(os.path.join(path_to_file,"",str(n_cluster)+'_km_label_'+file+'.npy')) 194 | random_generative_samples_motif(cfg, model, latent_vector,labels,n_cluster) 195 | 196 | 197 | 198 | 199 | 200 | 201 | 202 | 203 | 204 | 205 | 206 | -------------------------------------------------------------------------------- /vame/analysis/gif_creator.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import tqdm 14 | import umap 15 | import numpy as np 16 | from pathlib import Path 17 | import matplotlib 18 | from matplotlib import pyplot as plt 19 | from matplotlib.gridspec import GridSpec 20 | from vame.util.auxiliary import read_config 21 | from vame.util.gif_pose_helper import get_animal_frames 22 | 23 | 24 | def create_video(path_to_file, file, embed, clabel, frames, start, length, max_lag, num_points): 25 | # set matplotlib colormap 26 | cmap = matplotlib.cm.gray 27 | cmap_reversed = matplotlib.cm.get_cmap('gray_r') 28 | 29 | # this here generates every frame for your gif. The gif is lastly created by using ImageJ 30 | # the embed variable is my umap embedding, which is for the 2D case a 2xn dimensional vector 31 | fig = plt.figure() 32 | spec = GridSpec(ncols=2, nrows=1, width_ratios=[6, 3]) 33 | ax1 = fig.add_subplot(spec[0]) 34 | ax2 = fig.add_subplot(spec[1]) 35 | ax2.axis('off') 36 | ax2.grid(False) 37 | lag = 0 38 | for i in tqdm.tqdm(range(length)): 39 | if i > max_lag: 40 | lag = i - max_lag 41 | ax1.cla() 42 | ax1.axis('off') 43 | ax1.grid(False) 44 | ax1.scatter(embed[:num_points,0], embed[:num_points,1], c=clabel[:num_points], cmap='Spectral', s=1, alpha=0.4) 45 | ax1.set_aspect('equal', 'datalim') 46 | ax1.plot(embed[start+lag:start+i,0], embed[start+lag:start+i,1],'.b-',alpha=.6, linewidth=2, markersize=4) 47 | ax1.plot(embed[start+i,0], embed[start+i,1], 'gx', markersize=4) 48 | frame = frames[i] 49 | ax2.imshow(frame, cmap=cmap_reversed) 50 | # ax2.set_title("Motif %d,\n Community: %s" % (lbl, motifs[lbl]), fontsize=10) 51 | fig.savefig(os.path.join(path_to_file,"gif_frames",file+'gif_%d.png') %i) 52 | 53 | 54 | def gif(config, pose_ref_index, subtract_background=True, start=None, length=500, 55 | max_lag=30, label='community', file_format='.mp4', crop_size=(300,300)): 56 | 57 | config_file = Path(config).resolve() 58 | cfg = read_config(config_file) 59 | model_name = cfg['model_name'] 60 | n_cluster = cfg['n_cluster'] 61 | param = cfg['parameterization'] 62 | 63 | files = [] 64 | if cfg['all_data'] == 'No': 65 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 66 | "If you only want to use a specific dataset type filename: \n" 67 | "yes/no/filename ") 68 | else: 69 | all_flag = 'yes' 70 | 71 | if all_flag == 'yes' or all_flag == 'Yes': 72 | for file in cfg['video_sets']: 73 | files.append(file) 74 | 75 | elif all_flag == 'no' or all_flag == 'No': 76 | for file in cfg['video_sets']: 77 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 78 | if use_file == 'yes': 79 | files.append(file) 80 | if use_file == 'no': 81 | continue 82 | else: 83 | files.append(all_flag) 84 | 85 | 86 | for file in files: 87 | path_to_file=os.path.join(cfg['project_path'],"results",file,model_name,param+'-'+str(n_cluster),"") 88 | if not os.path.exists(os.path.join(path_to_file,"gif_frames")): 89 | os.mkdir(os.path.join(path_to_file,"gif_frames")) 90 | 91 | embed = np.load(os.path.join(path_to_file,"community","umap_embedding_"+file+'.npy')) 92 | 93 | try: 94 | embed = np.load(os.path.join(path_to_file,"","community","","umap_embedding_"+file+".npy")) 95 | num_points = cfg['num_points'] 96 | if num_points > embed.shape[0]: 97 | num_points = embed.shape[0] 98 | except: 99 | print("Compute embedding for file %s" %file) 100 | reducer = umap.UMAP(n_components=2, min_dist=cfg['min_dist'], n_neighbors=cfg['n_neighbors'], 101 | random_state=cfg['random_state']) 102 | 103 | latent_vector = np.load(os.path.join(path_to_file,"",'latent_vector_'+file+'.npy')) 104 | 105 | num_points = cfg['num_points'] 106 | if num_points > latent_vector.shape[0]: 107 | num_points = latent_vector.shape[0] 108 | print("Embedding %d data points.." %num_points) 109 | 110 | embed = reducer.fit_transform(latent_vector[:num_points,:]) 111 | np.save(os.path.join(path_to_file,"community","umap_embedding_"+file+'.npy'), embed) 112 | 113 | if label == "motif": 114 | umap_label = np.load(os.path.join(path_to_file,str(n_cluster)+"_km_label_"+file+'.npy')) 115 | elif label == "community": 116 | umap_label = np.load(os.path.join(path_to_file,"community","community_label_"+file+'.npy')) 117 | elif label == None: 118 | umap_label = None 119 | 120 | if start == None: 121 | start = np.random.choice(embed[:num_points].shape[0]-length) 122 | else: 123 | start = start 124 | 125 | frames = get_animal_frames(cfg, file, pose_ref_index, start, length, subtract_background, file_format, crop_size) 126 | 127 | create_video(path_to_file, file, embed, umap_label, frames, start, length, max_lag, num_points) 128 | 129 | 130 | 131 | 132 | 133 | 134 | 135 | 136 | 137 | 138 | 139 | 140 | 141 | 142 | 143 | 144 | 145 | 146 | 147 | 148 | 149 | 150 | 151 | 152 | 153 | 154 | 155 | -------------------------------------------------------------------------------- /vame/analysis/pose_segmentation.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import tqdm 14 | import torch 15 | import pickle 16 | import numpy as np 17 | from pathlib import Path 18 | 19 | 20 | from hmmlearn import hmm 21 | from sklearn.cluster import KMeans 22 | 23 | from vame.util.auxiliary import read_config 24 | from vame.model.rnn_model import RNN_VAE 25 | 26 | 27 | def load_model(cfg, model_name, fixed): 28 | use_gpu = torch.cuda.is_available() 29 | if use_gpu: 30 | pass 31 | else: 32 | torch.device("cpu") 33 | 34 | # load Model 35 | ZDIMS = cfg['zdims'] 36 | FUTURE_DECODER = cfg['prediction_decoder'] 37 | TEMPORAL_WINDOW = cfg['time_window']*2 38 | FUTURE_STEPS = cfg['prediction_steps'] 39 | NUM_FEATURES = cfg['num_features'] 40 | if fixed == False: 41 | NUM_FEATURES = NUM_FEATURES - 2 42 | hidden_size_layer_1 = cfg['hidden_size_layer_1'] 43 | hidden_size_layer_2 = cfg['hidden_size_layer_2'] 44 | hidden_size_rec = cfg['hidden_size_rec'] 45 | hidden_size_pred = cfg['hidden_size_pred'] 46 | dropout_encoder = cfg['dropout_encoder'] 47 | dropout_rec = cfg['dropout_rec'] 48 | dropout_pred = cfg['dropout_pred'] 49 | softplus = cfg['softplus'] 50 | 51 | 52 | if use_gpu: 53 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 54 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 55 | dropout_rec, dropout_pred, softplus).cuda() 56 | else: 57 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 58 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 59 | dropout_rec, dropout_pred, softplus).to() 60 | 61 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],'model','best_model',model_name+'_'+cfg['Project']+'.pkl'))) 62 | model.eval() 63 | 64 | return model 65 | 66 | 67 | def embedd_latent_vectors(cfg, files, model, fixed): 68 | project_path = cfg['project_path'] 69 | temp_win = cfg['time_window'] 70 | num_features = cfg['num_features'] 71 | if fixed == False: 72 | num_features = num_features - 2 73 | 74 | use_gpu = torch.cuda.is_available() 75 | if use_gpu: 76 | pass 77 | else: 78 | torch.device("cpu") 79 | 80 | latent_vector_files = [] 81 | 82 | for file in files: 83 | print('Embedding of latent vector for file %s' %file) 84 | data = np.load(os.path.join(project_path,'data',file,file+'-PE-seq-clean.npy')) 85 | latent_vector_list = [] 86 | with torch.no_grad(): 87 | for i in tqdm.tqdm(range(data.shape[1] - temp_win)): 88 | # for i in tqdm.tqdm(range(10000)): 89 | data_sample_np = data[:,i:temp_win+i].T 90 | data_sample_np = np.reshape(data_sample_np, (1, temp_win, num_features)) 91 | if use_gpu: 92 | h_n = model.encoder(torch.from_numpy(data_sample_np).type('torch.FloatTensor').cuda()) 93 | else: 94 | h_n = model.encoder(torch.from_numpy(data_sample_np).type('torch.FloatTensor').to()) 95 | mu, _, _ = model.lmbda(h_n) 96 | latent_vector_list.append(mu.cpu().data.numpy()) 97 | 98 | latent_vector = np.concatenate(latent_vector_list, axis=0) 99 | latent_vector_files.append(latent_vector) 100 | 101 | return latent_vector_files 102 | 103 | 104 | def consecutive(data, stepsize=1): 105 | data = data[:] 106 | return np.split(data, np.where(np.diff(data) != stepsize)[0]+1) 107 | 108 | 109 | def get_motif_usage(label): 110 | motif_usage = np.unique(label, return_counts=True) 111 | cons = consecutive(motif_usage[0]) 112 | if len(cons) != 1: 113 | usage_list = list(motif_usage[1]) 114 | for i in range(len(cons)-1): 115 | a = cons[i+1][0] 116 | b = cons[i][-1] 117 | d = (a-b)-1 118 | for j in range(1,d+1): 119 | index = cons[i][-1]+j 120 | usage_list.insert(index,0) 121 | usage = np.array(usage_list) 122 | motif_usage = usage 123 | else: 124 | motif_usage = motif_usage[1] 125 | 126 | return motif_usage 127 | 128 | 129 | def same_parameterization(cfg, files, latent_vector_files, states, parameterization): 130 | random_state = cfg['random_state_kmeans'] 131 | n_init = cfg['n_init_kmeans'] 132 | 133 | labels = [] 134 | cluster_centers = [] 135 | motif_usages = [] 136 | 137 | latent_vector_cat = np.concatenate(latent_vector_files, axis=0) 138 | 139 | if parameterization == "kmeans": 140 | print("Using kmeans as parameterization!") 141 | kmeans = KMeans(init='k-means++', n_clusters=states, random_state=42, n_init=20).fit(latent_vector_cat) 142 | clust_center = kmeans.cluster_centers_ 143 | label = kmeans.predict(latent_vector_cat) 144 | 145 | elif parameterization == "hmm": 146 | if cfg['hmm_trained'] == False: 147 | print("Using a HMM as parameterization!") 148 | hmm_model = hmm.GaussianHMM(n_components=states, covariance_type="full", n_iter=100) 149 | hmm_model.fit(latent_vector_cat) 150 | label = hmm_model.predict(latent_vector_cat) 151 | save_data = os.path.join(cfg['project_path'], "results", "") 152 | with open(save_data+"hmm_trained.pkl", "wb") as file: pickle.dump(hmm_model, file) 153 | else: 154 | print("Using a pretrained HMM as parameterization!") 155 | save_data = os.path.join(cfg['project_path'], "results", "") 156 | with open(save_data+"hmm_trained.pkl", "rb") as file: 157 | hmm_model = pickle.load(file) 158 | label = hmm_model.predict(latent_vector_cat) 159 | 160 | idx = 0 161 | for i, file in enumerate(files): 162 | file_len = latent_vector_files[i].shape[0] 163 | labels.append(label[idx:idx+file_len]) 164 | if parameterization == "kmeans": 165 | cluster_centers.append(clust_center) 166 | 167 | motif_usage = get_motif_usage(label[idx:idx+file_len]) 168 | motif_usages.append(motif_usage) 169 | idx += file_len 170 | 171 | return labels, cluster_centers, motif_usages 172 | 173 | 174 | def individual_parameterization(cfg, files, latent_vector_files, cluster): 175 | random_state = cfg['random_state_kmeans: '] 176 | n_init = cfg['n_init_kmeans'] 177 | 178 | labels = [] 179 | cluster_centers = [] 180 | motif_usages = [] 181 | for i, file in enumerate(files): 182 | print(file) 183 | kmeans = KMeans(init='k-means++', n_clusters=cluster, random_state=random_state, n_init=n_init).fit(latent_vector_files[i]) 184 | clust_center = kmeans.cluster_centers_ 185 | label = kmeans.predict(latent_vector_files[i]) 186 | motif_usage = get_motif_usage(label) 187 | motif_usages.append(motif_usage) 188 | labels.append(label) 189 | cluster_centers.append(clust_center) 190 | 191 | return labels, cluster_centers, motif_usages 192 | 193 | 194 | def pose_segmentation(config): 195 | config_file = Path(config).resolve() 196 | cfg = read_config(config_file) 197 | legacy = cfg['legacy'] 198 | model_name = cfg['model_name'] 199 | n_cluster = cfg['n_cluster'] 200 | fixed = cfg['egocentric_data'] 201 | parameterization = cfg['parameterization'] 202 | 203 | print('Pose segmentation for VAME model: %s \n' %model_name) 204 | 205 | if legacy == True: 206 | from segment_behavior import behavior_segmentation 207 | behavior_segmentation(config, model_name=model_name, cluster_method='kmeans', n_cluster=n_cluster) 208 | 209 | else: 210 | ind_param = cfg['individual_parameterization'] 211 | 212 | for folders in cfg['video_sets']: 213 | if not os.path.exists(os.path.join(cfg['project_path'],"results",folders,model_name,"")): 214 | os.mkdir(os.path.join(cfg['project_path'],"results",folders,model_name,"")) 215 | 216 | files = [] 217 | if cfg['all_data'] == 'No': 218 | all_flag = input("Do you want to qunatify your entire dataset? \n" 219 | "If you only want to use a specific dataset type filename: \n" 220 | "yes/no/filename ") 221 | file = all_flag 222 | 223 | else: 224 | all_flag = 'yes' 225 | 226 | if all_flag == 'yes' or all_flag == 'Yes': 227 | for file in cfg['video_sets']: 228 | files.append(file) 229 | elif all_flag == 'no' or all_flag == 'No': 230 | for file in cfg['video_sets']: 231 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 232 | if use_file == 'yes': 233 | files.append(file) 234 | if use_file == 'no': 235 | continue 236 | else: 237 | files.append(all_flag) 238 | # files.append("mouse-3-1") 239 | # file="mouse-3-1" 240 | 241 | use_gpu = torch.cuda.is_available() 242 | if use_gpu: 243 | print("Using CUDA") 244 | print('GPU active:',torch.cuda.is_available()) 245 | print('GPU used:',torch.cuda.get_device_name(0)) 246 | else: 247 | print("CUDA is not working! Attempting to use the CPU...") 248 | torch.device("cpu") 249 | 250 | if not os.path.exists(os.path.join(cfg['project_path'],"results",file,model_name, parameterization+'-'+str(n_cluster),"")): 251 | new = True 252 | # print("Hello1") 253 | model = load_model(cfg, model_name, fixed) 254 | latent_vectors = embedd_latent_vectors(cfg, files, model, fixed) 255 | 256 | if ind_param == False: 257 | print("For all animals the same parameterization of latent vectors is applied for %d cluster" %n_cluster) 258 | labels, cluster_center, motif_usages = same_parameterization(cfg, files, latent_vectors, n_cluster, parameterization) 259 | else: 260 | print("Individual parameterization of latent vectors for %d cluster" %n_cluster) 261 | labels, cluster_center, motif_usages = individual_parameterization(cfg, files, latent_vectors, n_cluster) 262 | 263 | else: 264 | print('\n' 265 | 'For model %s a latent vector embedding already exists. \n' 266 | 'Parameterization of latent vector with %d k-Means cluster' %(model_name, n_cluster)) 267 | 268 | if os.path.exists(os.path.join(cfg['project_path'],"results",file,model_name, parameterization+'-'+str(n_cluster),"")): 269 | flag = input('WARNING: A parameterization for the chosen cluster size of the model already exists! \n' 270 | 'Do you want to continue? A new parameterization will be computed! (yes/no) ') 271 | else: 272 | flag = 'yes' 273 | 274 | if flag == 'yes': 275 | new = True 276 | latent_vectors = [] 277 | for file in files: 278 | path_to_latent_vector = os.path.join(cfg['project_path'],"results",file,model_name, parameterization+'-'+str(n_cluster),"") 279 | latent_vector = np.load(os.path.join(path_to_latent_vector,'latent_vector_'+file+'.npy')) 280 | latent_vectors.append(latent_vector) 281 | 282 | if ind_param == False: 283 | print("For all animals the same parameterization of latent vectors is applied for %d cluster" %n_cluster) 284 | labels, cluster_center, motif_usages = same_parameterization(cfg, files, latent_vectors, n_cluster, parameterization) 285 | else: 286 | print("Individual parameterization of latent vectors for %d cluster" %n_cluster) 287 | labels, cluster_center, motif_usages = individual_parameterization(cfg, files, latent_vectors, n_cluster) 288 | 289 | else: 290 | print('No new parameterization has been calculated.') 291 | new = False 292 | 293 | # print("Hello2") 294 | if new == True: 295 | for idx, file in enumerate(files): 296 | print(os.path.join(cfg['project_path'],"results",file,"",model_name,parameterization+'-'+str(n_cluster),"")) 297 | if not os.path.exists(os.path.join(cfg['project_path'],"results",file,model_name,parameterization+'-'+str(n_cluster),"")): 298 | try: 299 | os.mkdir(os.path.join(cfg['project_path'],"results",file,"",model_name,parameterization+'-'+str(n_cluster),"")) 300 | except OSError as error: 301 | print(error) 302 | 303 | save_data = os.path.join(cfg['project_path'],"results",file,model_name,parameterization+'-'+str(n_cluster),"") 304 | np.save(os.path.join(save_data,str(n_cluster)+'_km_label_'+file), labels[idx]) 305 | if parameterization=="kmeans": 306 | np.save(os.path.join(save_data,'cluster_center_'+file), cluster_center[idx]) 307 | np.save(os.path.join(save_data,'latent_vector_'+file), latent_vectors[idx]) 308 | np.save(os.path.join(save_data,'motif_usage_'+file), motif_usages[idx]) 309 | 310 | 311 | print("You succesfully extracted motifs with VAME! From here, you can proceed running vame.motif_videos() ") 312 | # "to get an idea of the behavior captured by VAME. This will leave you with short snippets of certain movements." 313 | # "To get the full picture of the spatiotemporal dynamic we recommend applying our community approach afterwards.") 314 | 315 | -------------------------------------------------------------------------------- /vame/analysis/segment_behavior.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import numpy as np 14 | from pathlib import Path 15 | 16 | import torch 17 | import scipy.signal 18 | from sklearn import mixture 19 | from sklearn.cluster import KMeans 20 | 21 | from vame.util.auxiliary import read_config 22 | from vame.model.rnn_vae import RNN_VAE 23 | 24 | 25 | def load_data(PROJECT_PATH, file, data): 26 | X = np.load(os.path.join(PROJECT_PATH,"data",file,"",file+data+'.npy')) 27 | mean = np.load(os.path.join(PROJECT_PATH,"data","train",'seq_mean.npy')) 28 | std = np.load(os.path.join(PROJECT_PATH,"data","train",'seq_std.npy')) 29 | X = (X-mean)/std 30 | return X 31 | 32 | 33 | def kmeans_clustering(context, n_clusters): 34 | kmeans = KMeans(init='k-means++',n_clusters=n_clusters, random_state=42,n_init=15).fit(context) 35 | return kmeans.predict(context) 36 | 37 | 38 | def gmm_clustering(context,n_components): 39 | GMM = mixture.GaussianMixture 40 | gmm = GMM(n_components=n_components,covariance_type='full').fit(context) 41 | return gmm.predict(context) 42 | 43 | 44 | def behavior_segmentation(config, model_name=None, cluster_method='kmeans', n_cluster=[30]): 45 | 46 | config_file = Path(config).resolve() 47 | cfg = read_config(config_file) 48 | 49 | for folders in cfg['video_sets']: 50 | if not os.path.exists(os.path.join(cfg['project_path'],"results",folders,"",model_name)): 51 | os.mkdir(os.path.join(cfg['project_path'],"results",folders,"",model_name)) 52 | 53 | files = [] 54 | if cfg['all_data'] == 'No': 55 | all_flag = input("Do you want to qunatify your entire dataset? \n" 56 | "If you only want to use a specific dataset type filename: \n" 57 | "yes/no/filename ") 58 | else: 59 | all_flag = 'yes' 60 | 61 | if all_flag == 'yes' or all_flag == 'Yes': 62 | for file in cfg['video_sets']: 63 | files.append(file) 64 | elif all_flag == 'no' or all_flag == 'No': 65 | for file in cfg['video_sets']: 66 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 67 | if use_file == 'yes': 68 | files.append(file) 69 | if use_file == 'no': 70 | continue 71 | else: 72 | files.append(all_flag) 73 | 74 | 75 | use_gpu = torch.cuda.is_available() 76 | if use_gpu: 77 | print("Using CUDA") 78 | print('GPU active:',torch.cuda.is_available()) 79 | print('GPU used:',torch.cuda.get_device_name(0)) 80 | else: 81 | print("CUDA is not working! Attempting to use the CPU...") 82 | torch.device("cpu") 83 | 84 | z, z_logger = temporal_quant(cfg, model_name, files, use_gpu) 85 | cluster_latent_space(cfg, files, z, z_logger, cluster_method, n_cluster, model_name) 86 | 87 | 88 | def temporal_quant(cfg, model_name, files, use_gpu): 89 | 90 | SEED = 19 91 | ZDIMS = cfg['zdims'] 92 | FUTURE_DECODER = cfg['prediction_decoder'] 93 | TEMPORAL_WINDOW = cfg['time_window']*2 94 | FUTURE_STEPS = cfg['prediction_steps'] 95 | NUM_FEATURES = cfg['num_features'] 96 | PROJECT_PATH = cfg['project_path'] 97 | hidden_size_layer_1 = cfg['hidden_size_layer_1'] 98 | hidden_size_layer_2 = cfg['hidden_size_layer_2'] 99 | hidden_size_rec = cfg['hidden_size_rec'] 100 | hidden_size_pred = cfg['hidden_size_pred'] 101 | dropout_encoder = cfg['dropout_encoder'] 102 | dropout_rec = cfg['dropout_rec'] 103 | dropout_pred = cfg['dropout_pred'] 104 | temp_win = int(TEMPORAL_WINDOW/2) 105 | 106 | if use_gpu: 107 | torch.cuda.manual_seed(SEED) 108 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 109 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 110 | dropout_rec, dropout_pred).cuda() 111 | else: 112 | torch.cuda.manual_seed(SEED) 113 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 114 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 115 | dropout_rec, dropout_pred).to() 116 | 117 | if cfg['snapshot'] == 'yes': 118 | if use_gpu: 119 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model","snapshots",model_name+'_'+cfg['Project']+'_epoch_'+cfg['snapshot_epoch']+'.pkl'))) 120 | else: 121 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model","snapshots",model_name+'_'+cfg['Project']+'_epoch_'+cfg['snapshot_epoch']+'.pkl'),map_location=torch.device('cpu'))) 122 | else: 123 | if use_gpu: 124 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model",model_name+'_'+cfg['Project']+'.pkl'))) 125 | else: 126 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model",model_name+'_'+cfg['Project']+'.pkl'),map_location=torch.device('cpu'))) 127 | 128 | model.eval() 129 | 130 | z_list = [] 131 | z_logger = [] 132 | logger = 0 133 | for file in files: 134 | print("Computing latent space for %s " %file) 135 | z_logger.append(logger) 136 | 137 | 138 | data=cfg['load_data'] 139 | X = load_data(PROJECT_PATH, file, data) 140 | 141 | if X.shape[0] > X.shape[1]: 142 | X=X.T 143 | 144 | num_frames = len(X[0,:]) - temp_win 145 | window_start = int(temp_win/2) 146 | idx = int(temp_win/2) 147 | x_decoded = [] 148 | 149 | with torch.no_grad(): 150 | for i in range(num_frames): 151 | if idx >= num_frames: 152 | break 153 | data = X[:,idx-window_start:idx+window_start] 154 | data = np.reshape(data, (1,temp_win,NUM_FEATURES)) 155 | if use_gpu: 156 | dataTorch = torch.from_numpy(data).type(torch.FloatTensor).cuda() 157 | else: 158 | dataTorch = torch.from_numpy(data).type(torch.FloatTensor).to() 159 | h_n = model.encoder(dataTorch) 160 | latent, _, _ = model.lmbda(h_n) 161 | z = latent.cpu().data.numpy() 162 | x_decoded.append(z) 163 | idx += 1 164 | 165 | z_temp = np.concatenate(x_decoded,axis=0) 166 | logger_temp = len(z_temp) 167 | logger += logger_temp 168 | z_list.append(z_temp) 169 | 170 | z_array= np.concatenate(z_list) 171 | z_logger.append(len(z_array)) 172 | 173 | return z_array, z_logger 174 | 175 | 176 | def cluster_latent_space(cfg, files, z_data, z_logger, cluster_method, n_cluster, model_name): 177 | 178 | for cluster in n_cluster: 179 | if cluster_method == 'kmeans': 180 | print('Behavior segmentation via k-Means for %d cluster.' %cluster) 181 | data_labels = kmeans_clustering(z_data, n_clusters=cluster) 182 | data_labels = np.int64(scipy.signal.medfilt(data_labels, cfg['median_filter'])) 183 | 184 | if cluster_method == 'GMM': 185 | print('Behavior segmentation via GMM.') 186 | data_labels = gmm_clustering(z_data, n_components=cluster) 187 | data_labels = np.int64(scipy.signal.medfilt(data_labels, cfg['median_filter'])) 188 | 189 | for idx, file in enumerate(files): 190 | print("Segmentation for file %s..." %file ) 191 | if not os.path.exists(os.path.join(cfg['project_path'],"results",file,"",model_name,"",cluster_method+'-'+str(cluster))): 192 | os.mkdir(os.path.join(cfg['project_path'],"results",file,"",model_name,"",cluster_method+'-'+str(cluster))) 193 | 194 | save_data = os.path.join(cfg['project_path'],"results",file,"",model_name,"") 195 | z_latent = z_data[z_logger[idx]:z_logger[idx+1],:] 196 | labels = data_labels[z_logger[idx]:z_logger[idx+1]] 197 | 198 | 199 | if cluster_method == 'kmeans': 200 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+str(cluster)+'_km_label_'+file, labels) 201 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+'latent_vector_'+file, z_latent) 202 | 203 | if cluster_method == 'GMM': 204 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+str(cluster)+'_gmm_label_'+file, labels) 205 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+'latent_vector_'+file, z_latent) 206 | 207 | if cluster_method == 'all': 208 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+str(cluster)+'_km_label_'+file, labels) 209 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+str(cluster)+'_gmm_label_'+file, labels) 210 | np.save(save_data+cluster_method+'-'+str(cluster)+'/'+'latent_vector_'+file, z_latent) 211 | 212 | 213 | -------------------------------------------------------------------------------- /vame/analysis/tree_hierarchy.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | 13 | import numpy as np 14 | import networkx as nx 15 | import random 16 | from matplotlib import pyplot as plt 17 | 18 | def hierarchy_pos(G, root=None, width=.5, vert_gap = 0.2, vert_loc = 0, xcenter = 0.5): 19 | 20 | ''' 21 | From Joel's answer at https://stackoverflow.com/a/29597209/2966723. 22 | ''' 23 | if not nx.is_tree(G): 24 | raise TypeError('cannot use hierarchy_pos on a graph that is not a tree') 25 | 26 | if root is None: 27 | if isinstance(G, nx.DiGraph): 28 | root = next(iter(nx.topological_sort(G))) #allows back compatibility with nx version 1.11 29 | else: 30 | root = random.choice(list(G.nodes)) 31 | 32 | def _hierarchy_pos(G, root, width=1., vert_gap = 0.2, vert_loc = 0, xcenter = 0.5, pos = None, parent = None): 33 | if pos is None: 34 | pos = {root:(xcenter,vert_loc)} 35 | else: 36 | pos[root] = (xcenter, vert_loc) 37 | children = list(G.neighbors(root)) 38 | if not isinstance(G, nx.DiGraph) and parent is not None: 39 | children.remove(parent) 40 | if len(children)!=0: 41 | dx = width/len(children) 42 | nextx = xcenter - width/2 - dx/2 43 | for child in children: 44 | nextx += dx 45 | pos = _hierarchy_pos(G,child, width = dx, vert_gap = vert_gap, 46 | vert_loc = vert_loc-vert_gap, xcenter=nextx, 47 | pos=pos, parent = root) 48 | return pos 49 | 50 | 51 | return _hierarchy_pos(G, root, width, vert_gap, vert_loc, xcenter) 52 | 53 | 54 | def merge_func(transition_matrix, n_cluster, motif_norm, merge_sel): 55 | 56 | if merge_sel == 0: 57 | # merge nodes with highest transition probability 58 | cost = np.max(transition_matrix) 59 | merge_nodes = np.where(cost == transition_matrix) 60 | 61 | if merge_sel == 1: 62 | 63 | cost_temp = 100 64 | for i in range(n_cluster): 65 | for j in range(n_cluster): 66 | try: 67 | cost = (motif_norm[i] + motif_norm[j]) / np.abs(transition_matrix[i,j] + transition_matrix[j,i] ) 68 | except ZeroDivisionError: 69 | print("Error: Transition probabilities between motif "+str(i)+" and motif "+str(j)+ " are zero.") 70 | if cost <= cost_temp: 71 | cost_temp = cost 72 | merge_nodes = (np.array([i]), np.array([j])) 73 | 74 | return merge_nodes 75 | 76 | 77 | def graph_to_tree(motif_usage, transition_matrix, n_cluster, merge_sel=1): 78 | 79 | if merge_sel == 1: 80 | # motif_usage_temp = np.load(path_to_file+'/behavior_quantification/motif_usage.npy') 81 | motif_usage_temp = motif_usage 82 | motif_usage_temp_colsum = motif_usage_temp.sum(axis=0) 83 | motif_norm = motif_usage_temp/motif_usage_temp_colsum 84 | motif_norm_temp = motif_norm.copy() 85 | else: 86 | motif_norm_temp = None 87 | 88 | merging_nodes = [] 89 | hierarchy_nodes = [] 90 | trans_mat_temp = transition_matrix.copy() 91 | is_leaf = np.ones((n_cluster), dtype='int') 92 | node_label = [] 93 | leaf_idx = [] 94 | 95 | if np.any(transition_matrix.sum(axis=1) == 0): 96 | temp = np.where(transition_matrix.sum(axis=1)==0) 97 | reduction = len(temp) + 1 98 | else: 99 | reduction = 1 100 | 101 | for i in range(n_cluster-reduction): 102 | 103 | # max_tr = np.max(trans_mat_temp) #merge function 104 | # nodes = np.where(max_tr == trans_mat_temp) 105 | nodes = merge_func(trans_mat_temp, n_cluster, motif_norm_temp, merge_sel) 106 | 107 | if np.size(nodes) >= 2: 108 | nodes = np.array([nodes[0][0], nodes[1][0]]) 109 | 110 | if is_leaf[nodes[0]] == 1: 111 | is_leaf[nodes[0]] = 0 112 | node_label.append('leaf_left_'+str(i)) 113 | leaf_idx.append(1) 114 | 115 | elif is_leaf[nodes[0]] == 0: 116 | node_label.append('h_'+str(i)+'_'+str(nodes[0])) 117 | leaf_idx.append(0) 118 | 119 | if is_leaf[nodes[1]] == 1: 120 | is_leaf[nodes[1]] = 0 121 | node_label.append('leaf_right_'+str(i)) 122 | hierarchy_nodes.append('h_'+str(i)+'_'+str(nodes[1])) 123 | leaf_idx.append(1) 124 | 125 | elif is_leaf[nodes[1]] == 0: 126 | node_label.append('h_'+str(i)+'_'+str(nodes[1])) 127 | hierarchy_nodes.append('h_'+str(i)+'_'+str(nodes[1])) 128 | leaf_idx.append(0) 129 | 130 | merging_nodes.append(nodes) 131 | 132 | node1_trans_x = trans_mat_temp[nodes[0],:] 133 | node2_trans_x = trans_mat_temp[nodes[1],:] 134 | 135 | node1_trans_y = trans_mat_temp[:,nodes[0]] 136 | node2_trans_y = trans_mat_temp[:,nodes[1]] 137 | 138 | new_node_trans_x = node1_trans_x + node2_trans_x 139 | new_node_trans_y = node1_trans_y + node2_trans_y 140 | 141 | trans_mat_temp[nodes[1],:] = new_node_trans_x 142 | trans_mat_temp[:,nodes[1]] = new_node_trans_y 143 | 144 | trans_mat_temp[nodes[0],:] = 0 145 | trans_mat_temp[:,nodes[0]] = 0 146 | 147 | trans_mat_temp[nodes[1],nodes[1]] = 0 148 | 149 | if merge_sel == 1: 150 | motif_norm_1 = motif_norm_temp[nodes[0]] 151 | motif_norm_2 = motif_norm_temp[nodes[1]] 152 | 153 | new_motif = motif_norm_1 + motif_norm_2 154 | 155 | motif_norm_temp[nodes[0]] = 0 156 | motif_norm_temp[nodes[1]] = 0 157 | 158 | motif_norm_temp[nodes[1]] = new_motif 159 | 160 | merge = np.array(merging_nodes) 161 | # merge = np.concatenate((merge),axis=1).T 162 | 163 | T = nx.Graph() 164 | 165 | T.add_node('Root') 166 | node_dict = {} 167 | 168 | if leaf_idx[-1] == 0: 169 | temp_node = 'h_'+str(merge[-1,1])+'_'+str(28) 170 | T.add_edge(temp_node, 'Root') 171 | node_dict[merge[-1,1]] = temp_node 172 | 173 | if leaf_idx[-1] == 1: 174 | T.add_edge(merge[-1,1], 'Root') 175 | 176 | if leaf_idx[-2] == 0: 177 | temp_node = 'h_'+str(merge[-1,0])+'_'+str(28) 178 | T.add_edge(temp_node, 'Root') 179 | node_dict[merge[-1,0]] = temp_node 180 | 181 | if leaf_idx[-2] == 1: 182 | T.add_edge(merge[-1,0], 'Root') 183 | 184 | idx = len(leaf_idx)-3 185 | 186 | if np.any(transition_matrix.sum(axis=1) == 0): 187 | temp = np.where(transition_matrix.sum(axis=1)==0) 188 | reduction = len(temp) + 2 189 | else: 190 | reduction = 2 191 | 192 | for i in range(n_cluster-reduction)[::-1]: 193 | 194 | if leaf_idx[idx-1] == 1: 195 | if merge[i,1] in node_dict: 196 | T.add_edge(merge[i,0], node_dict[merge[i,1]]) 197 | else: 198 | T.add_edge(merge[i,0], temp_node) 199 | 200 | if leaf_idx[idx] == 1: 201 | if merge[i,1] in node_dict: 202 | T.add_edge(merge[i,1], node_dict[merge[i,1]]) 203 | else: 204 | T.add_edge(merge[i,1], temp_node) 205 | 206 | if leaf_idx[idx] == 0: 207 | new_node = 'h_'+str(merge[i,1])+'_'+str(i) 208 | if merge[i,1] in node_dict: 209 | T.add_edge(node_dict[merge[i,1]], new_node) 210 | else: 211 | T.add_edge(temp_node, new_node) 212 | # node_dict[merge[i,1]] = new_node 213 | 214 | if leaf_idx[idx-1] == 1: 215 | temp_node = new_node 216 | node_dict[merge[i,1]] = new_node 217 | else: 218 | new_node_2 = 'h_'+str(merge[i,0])+'_'+str(i) 219 | # temp_node = 'h_'+str(merge[i,0])+'_'+str(i) 220 | T.add_edge(node_dict[merge[i,1]], new_node_2) 221 | # node_dict[merge[i,0]] = temp_node 222 | node_dict[merge[i,1]] = new_node 223 | node_dict[merge[i,0]] = new_node_2 224 | # temp_node = new_node 225 | 226 | elif leaf_idx[idx-1] == 0: 227 | new_node = 'h_'+str(merge[i,0])+'_'+str(i) 228 | if merge[i,1] in node_dict: 229 | T.add_edge(node_dict[merge[i,1]], new_node) 230 | else: 231 | T.add_edge(temp_node, new_node) 232 | node_dict[merge[i,0]] = new_node 233 | 234 | if leaf_idx[idx] == 1: 235 | temp_node = new_node 236 | else: 237 | new_node = 'h_'+str(merge[i,1])+'_'+str(i) 238 | T.add_edge(temp_node, new_node) 239 | node_dict[merge[i,1]] = new_node 240 | temp_node = new_node 241 | 242 | idx -= 2 243 | 244 | return T 245 | 246 | 247 | def draw_tree(T): 248 | # pos = nx.drawing.layout.fruchterman_reingold_layout(T) 249 | pos = hierarchy_pos(T,'Root',width=.5, vert_gap = 0.1, vert_loc = 0, xcenter = 50) 250 | fig = plt.figure(2) 251 | nx.draw_networkx(T, pos) 252 | figManager = plt.get_current_fig_manager() 253 | figManager.window.showMaximized() 254 | 255 | 256 | def traverse_tree(T, root_node=None): 257 | if root_node == None: 258 | node=['Root'] 259 | else: 260 | node=[root_node] 261 | traverse_list = [] 262 | traverse_preorder = '{' 263 | 264 | def _traverse_tree(T, node, traverse_preorder): 265 | traverse_preorder += str(node[0]) 266 | traverse_list.append(node[0]) 267 | children = list(T.neighbors(node[0])) 268 | 269 | if len(children) == 3: 270 | # print(children) 271 | for child in children: 272 | if child in traverse_list: 273 | # print(child) 274 | children.remove(child) 275 | 276 | if len(children) > 1: 277 | traverse_preorder += '{' 278 | traverse_preorder_temp = _traverse_tree(T, [children[0]], '') 279 | traverse_preorder += traverse_preorder_temp 280 | 281 | traverse_preorder += '}{' 282 | 283 | traverse_preorder_temp = _traverse_tree(T, [children[1]], '') 284 | traverse_preorder += traverse_preorder_temp 285 | traverse_preorder += '}' 286 | 287 | return traverse_preorder 288 | 289 | traverse_preorder = _traverse_tree(T, node, traverse_preorder) 290 | traverse_preorder += '}' 291 | 292 | return traverse_preorder 293 | 294 | 295 | 296 | def _traverse_tree(T, node, traverse_preorder,traverse_list): 297 | traverse_preorder += str(node[0]) 298 | traverse_list.append(node[0]) 299 | children = list(T.neighbors(node[0])) 300 | 301 | if len(children) == 3: 302 | # print(children) 303 | for child in children: 304 | if child in traverse_list: 305 | # print(child) 306 | children.remove(child) 307 | 308 | if len(children) > 1: 309 | traverse_preorder += '{' 310 | traverse_preorder_temp = _traverse_tree(T, [children[0]], '',traverse_list) 311 | traverse_preorder += traverse_preorder_temp 312 | 313 | traverse_preorder += '}{' 314 | 315 | traverse_preorder_temp = _traverse_tree(T, [children[1]], '',traverse_list) 316 | traverse_preorder += traverse_preorder_temp 317 | traverse_preorder += '}' 318 | 319 | return traverse_preorder 320 | 321 | def traverse_tree(T, root_node=None): 322 | if root_node == None: 323 | node=['Root'] 324 | else: 325 | node=[root_node] 326 | traverse_list = [] 327 | traverse_preorder = '{' 328 | traverse_preorder = _traverse_tree(T, node, traverse_preorder,traverse_list) 329 | traverse_preorder += '}' 330 | 331 | return traverse_preorder 332 | 333 | 334 | def _traverse_tree_cutline(T, node, traverse_list, cutline, level, community_bag, community_list=None): 335 | cmap = plt.get_cmap("tab10") 336 | traverse_list.append(node[0]) 337 | if community_list is not None and type(node[0]) is not str: 338 | community_list.append(node[0]) 339 | children = list(T.neighbors(node[0])) 340 | 341 | if len(children) == 3: 342 | # print(children) 343 | for child in children: 344 | if child in traverse_list: 345 | # print(child) 346 | children.remove(child) 347 | 348 | if len(children) > 1: 349 | if nx.shortest_path_length(T,'Root',node[0])==cutline: 350 | #create new list 351 | traverse_list1 = [] 352 | traverse_list2 = [] 353 | community_bag = _traverse_tree_cutline(T, [children[0]], traverse_list, cutline, level+1, community_bag, traverse_list1) 354 | community_bag = _traverse_tree_cutline(T, [children[1]], traverse_list, cutline, level+1, community_bag, traverse_list2) 355 | joined_list=traverse_list1+traverse_list2 356 | community_bag.append(joined_list) 357 | if type(node[0]) is not str: #append itself 358 | community_bag.append([node[0]]) 359 | else: 360 | community_bag = _traverse_tree_cutline(T, [children[0]], traverse_list, cutline, level+1, community_bag, community_list) 361 | community_bag = _traverse_tree_cutline(T, [children[1]], traverse_list, cutline, level+1, community_bag, community_list) 362 | 363 | return community_bag 364 | 365 | 366 | def traverse_tree_cutline(T, root_node=None,cutline=2): 367 | if root_node == None: 368 | node=['Root'] 369 | else: 370 | node=[root_node] 371 | traverse_list = [] 372 | color_map = [] 373 | community_bag=[] 374 | level = 0 375 | community_bag = _traverse_tree_cutline(T, node, traverse_list,cutline, level, color_map,community_bag) 376 | 377 | return community_bag 378 | -------------------------------------------------------------------------------- /vame/analysis/umap_visualization.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import umap 14 | import numpy as np 15 | from pathlib import Path 16 | import matplotlib.pyplot as plt 17 | from matplotlib.gridspec import GridSpec 18 | from mpl_toolkits.mplot3d import Axes3D 19 | 20 | from vame.util.auxiliary import read_config 21 | 22 | 23 | def umap_vis(file, embed, num_points): 24 | fig = plt.figure(1) 25 | plt.scatter(embed[:num_points,0], embed[:num_points,1], s=2, alpha=.5) 26 | plt.gca().set_aspect('equal', 'datalim') 27 | plt.grid(False) 28 | 29 | 30 | def umap_label_vis(file, embed, label, n_cluster, num_points): 31 | fig = plt.figure(1) 32 | plt.scatter(embed[:num_points,0], embed[:num_points,1], c=label[:num_points], cmap='Spectral', s=2, alpha=.7) 33 | plt.colorbar(boundaries=np.arange(n_cluster+1)-0.5).set_ticks(np.arange(n_cluster)) 34 | plt.gca().set_aspect('equal', 'datalim') 35 | plt.grid(False) 36 | 37 | 38 | def umap_vis_comm(file, embed, community_label, num_points): 39 | num = np.unique(community_label).shape[0] 40 | fig = plt.figure(1) 41 | plt.scatter(embed[:num_points,0], embed[:num_points,1], c=community_label[:num_points], cmap='Spectral', s=2, alpha=.7) 42 | plt.colorbar(boundaries=np.arange(num+1)-0.5).set_ticks(np.arange(num)) 43 | plt.gca().set_aspect('equal', 'datalim') 44 | plt.grid(False) 45 | 46 | 47 | def visualization(config, label=None): 48 | config_file = Path(config).resolve() 49 | cfg = read_config(config_file) 50 | model_name = cfg['model_name'] 51 | n_cluster = cfg['n_cluster'] 52 | param = cfg['parameterization'] 53 | 54 | files = [] 55 | if cfg['all_data'] == 'No': 56 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 57 | "If you only want to use a specific dataset type filename: \n" 58 | "yes/no/filename ") 59 | else: 60 | all_flag = 'yes' 61 | 62 | if all_flag == 'yes' or all_flag == 'Yes': 63 | for file in cfg['video_sets']: 64 | files.append(file) 65 | 66 | elif all_flag == 'no' or all_flag == 'No': 67 | for file in cfg['video_sets']: 68 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 69 | if use_file == 'yes': 70 | files.append(file) 71 | if use_file == 'no': 72 | continue 73 | else: 74 | files.append(all_flag) 75 | 76 | for idx, file in enumerate(files): 77 | path_to_file=os.path.join(cfg['project_path'],"results",file,"",model_name,"",param+'-'+str(n_cluster)) 78 | 79 | try: 80 | embed = np.load(os.path.join(path_to_file,"","community","","umap_embedding_"+file+".npy")) 81 | num_points = cfg['num_points'] 82 | if num_points > embed.shape[0]: 83 | num_points = embed.shape[0] 84 | except: 85 | if not os.path.exists(os.path.join(path_to_file,"community")): 86 | os.mkdir(os.path.join(path_to_file,"community")) 87 | print("Compute embedding for file %s" %file) 88 | reducer = umap.UMAP(n_components=2, min_dist=cfg['min_dist'], n_neighbors=cfg['n_neighbors'], 89 | random_state=cfg['random_state']) 90 | 91 | latent_vector = np.load(os.path.join(path_to_file,"",'latent_vector_'+file+'.npy')) 92 | 93 | num_points = cfg['num_points'] 94 | if num_points > latent_vector.shape[0]: 95 | num_points = latent_vector.shape[0] 96 | print("Embedding %d data points.." %num_points) 97 | 98 | embed = reducer.fit_transform(latent_vector[:num_points,:]) 99 | np.save(os.path.join(path_to_file,"community","umap_embedding_"+file+'.npy'), embed) 100 | 101 | print("Visualizing %d data points.. " %num_points) 102 | if label == None: 103 | umap_vis(file, embed, num_points) 104 | 105 | if label == 'motif': 106 | motif_label = np.load(os.path.join(path_to_file,"",str(n_cluster)+'_km_label_'+file+'.npy')) 107 | umap_label_vis(file, embed, motif_label, n_cluster, num_points) 108 | 109 | if label == "community": 110 | community_label = np.load(os.path.join(path_to_file,"","community","","community_label_"+file+".npy")) 111 | umap_vis_comm(file, embed, community_label, num_points) 112 | 113 | 114 | 115 | 116 | 117 | 118 | 119 | 120 | 121 | 122 | 123 | -------------------------------------------------------------------------------- /vame/analysis/videowriter.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | from pathlib import Path 14 | import numpy as np 15 | import cv2 as cv 16 | import tqdm 17 | 18 | from vame.util.auxiliary import read_config 19 | 20 | 21 | def get_cluster_vid(cfg, path_to_file, file, n_cluster, videoType, flag): 22 | if flag == "motif": 23 | print("Motif videos getting created for "+file+" ...") 24 | labels = np.load(os.path.join(path_to_file,str(n_cluster)+'_km_label_'+file+'.npy')) 25 | if flag == "community": 26 | print("Community videos getting created for "+file+" ...") 27 | labels = np.load(os.path.join(path_to_file,"community",'community_label_'+file+'.npy')) 28 | capture = cv.VideoCapture(os.path.join(cfg['project_path'],"videos",file+videoType)) 29 | 30 | if capture.isOpened(): 31 | width = capture.get(cv.CAP_PROP_FRAME_WIDTH) 32 | height = capture.get(cv.CAP_PROP_FRAME_HEIGHT) 33 | # print('width, height:', width, height) 34 | 35 | fps = 25#capture.get(cv.CAP_PROP_FPS) 36 | # print('fps:', fps) 37 | 38 | cluster_start = cfg['time_window'] / 2 39 | for cluster in range(n_cluster): 40 | print('Cluster: %d' %(cluster)) 41 | cluster_lbl = np.where(labels == cluster) 42 | cluster_lbl = cluster_lbl[0] 43 | 44 | if flag == "motif": 45 | output = os.path.join(path_to_file,"cluster_videos",file+'-motif_%d.avi' %cluster) 46 | if flag == "community": 47 | output = os.path.join(path_to_file,"community_videos",file+'-community_%d.avi' %cluster) 48 | 49 | video = cv.VideoWriter(output, cv.VideoWriter_fourcc('M','J','P','G'), fps, (int(width), int(height))) 50 | 51 | if len(cluster_lbl) < cfg['length_of_motif_video']: 52 | vid_length = len(cluster_lbl) 53 | else: 54 | vid_length = cfg['length_of_motif_video'] 55 | 56 | for num in tqdm.tqdm(range(vid_length)): 57 | idx = cluster_lbl[num] 58 | capture.set(1,idx+cluster_start) 59 | ret, frame = capture.read() 60 | video.write(frame) 61 | 62 | video.release() 63 | capture.release() 64 | 65 | 66 | def motif_videos(config, videoType='.mp4'): 67 | config_file = Path(config).resolve() 68 | cfg = read_config(config_file) 69 | model_name = cfg['model_name'] 70 | n_cluster = cfg['n_cluster'] 71 | param = cfg['parameterization'] 72 | flag = 'motif' 73 | 74 | files = [] 75 | if cfg['all_data'] == 'No': 76 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 77 | "If you only want to use a specific dataset type filename: \n" 78 | "yes/no/filename ") 79 | else: 80 | all_flag = 'yes' 81 | 82 | if all_flag == 'yes' or all_flag == 'Yes': 83 | for file in cfg['video_sets']: 84 | files.append(file) 85 | 86 | elif all_flag == 'no' or all_flag == 'No': 87 | for file in cfg['video_sets']: 88 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 89 | if use_file == 'yes': 90 | files.append(file) 91 | if use_file == 'no': 92 | continue 93 | else: 94 | files.append(all_flag) 95 | 96 | print("Cluster size is: %d " %n_cluster) 97 | for file in files: 98 | path_to_file=os.path.join(cfg['project_path'],"results",file,model_name,param+'-'+str(n_cluster),"") 99 | if not os.path.exists(os.path.join(path_to_file,"cluster_videos")): 100 | os.mkdir(os.path.join(path_to_file,"cluster_videos")) 101 | 102 | get_cluster_vid(cfg, path_to_file, file, n_cluster, videoType, flag) 103 | 104 | print("All videos have been created!") 105 | 106 | 107 | def community_videos(config, videoType='.mp4'): 108 | config_file = Path(config).resolve() 109 | cfg = read_config(config_file) 110 | model_name = cfg['model_name'] 111 | n_cluster = cfg['n_cluster'] 112 | param = cfg['parameterization'] 113 | flag = 'community' 114 | 115 | files = [] 116 | if cfg['all_data'] == 'No': 117 | all_flag = input("Do you want to write motif videos for your entire dataset? \n" 118 | "If you only want to use a specific dataset type filename: \n" 119 | "yes/no/filename ") 120 | else: 121 | all_flag = 'yes' 122 | 123 | if all_flag == 'yes' or all_flag == 'Yes': 124 | for file in cfg['video_sets']: 125 | files.append(file) 126 | 127 | elif all_flag == 'no' or all_flag == 'No': 128 | for file in cfg['video_sets']: 129 | use_file = input("Do you want to quantify " + file + "? yes/no: ") 130 | if use_file == 'yes': 131 | files.append(file) 132 | if use_file == 'no': 133 | continue 134 | else: 135 | files.append(all_flag) 136 | 137 | print("Cluster size is: %d " %n_cluster) 138 | for file in files: 139 | path_to_file=os.path.join(cfg['project_path'],"results",file,model_name,param+'-'+str(n_cluster),"") 140 | if not os.path.exists(os.path.join(path_to_file,"community_videos")): 141 | os.mkdir(os.path.join(path_to_file,"community_videos")) 142 | 143 | get_cluster_vid(cfg, path_to_file, file, n_cluster, videoType, flag) 144 | 145 | print("All videos have been created!") 146 | -------------------------------------------------------------------------------- /vame/initialize_project/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | import sys 12 | sys.dont_write_bytecode = True 13 | 14 | from vame.initialize_project.new import init_new_project 15 | -------------------------------------------------------------------------------- /vame/initialize_project/new.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | 11 | The following code is adapted from: 12 | 13 | DeepLabCut2.0 Toolbox (deeplabcut.org) 14 | © A. & M. Mathis Labs 15 | https://github.com/AlexEMG/DeepLabCut 16 | Please see AUTHORS for contributors. 17 | https://github.com/AlexEMG/DeepLabCut/blob/master/AUTHORS 18 | Licensed under GNU Lesser General Public License v3.0 19 | """ 20 | 21 | import os 22 | from pathlib import Path 23 | import shutil 24 | 25 | def init_new_project(project, videos, working_directory=None, videotype='.mp4'): 26 | from datetime import datetime as dt 27 | from vame.util import auxiliary 28 | date = dt.today() 29 | month = date.strftime("%B") 30 | day = date.day 31 | year = date.year 32 | d = str(month[0:3]+str(day)) 33 | date = dt.today().strftime('%Y-%m-%d') 34 | 35 | if working_directory == None: 36 | working_directory = '.' 37 | 38 | wd = Path(working_directory).resolve() 39 | project_name = '{pn}-{date}'.format(pn=project, date=d+'-'+str(year)) 40 | 41 | project_path = wd / project_name 42 | 43 | 44 | if project_path.exists(): 45 | print('Project "{}" already exists!'.format(project_path)) 46 | return 47 | 48 | video_path = project_path / 'videos' 49 | data_path = project_path / 'data' 50 | results_path = project_path / 'results' 51 | model_path = project_path / 'model' 52 | 53 | for p in [video_path, data_path, results_path, model_path]: 54 | p.mkdir(parents=True) 55 | print('Created "{}"'.format(p)) 56 | 57 | vids = [] 58 | for i in videos: 59 | #Check if it is a folder 60 | if os.path.isdir(i): 61 | vids_in_dir = [os.path.join(i,vp) for vp in os.listdir(i) if videotype in vp] 62 | vids = vids + vids_in_dir 63 | if len(vids_in_dir)==0: 64 | print("No videos found in",i) 65 | print("Perhaps change the videotype, which is currently set to:", videotype) 66 | else: 67 | videos = vids 68 | print(len(vids_in_dir)," videos from the directory" ,i, "were added to the project.") 69 | else: 70 | if os.path.isfile(i): 71 | vids = vids + [i] 72 | videos = vids 73 | 74 | 75 | videos = [Path(vp) for vp in videos] 76 | video_names = [] 77 | dirs_data = [data_path/Path(i.stem) for i in videos] 78 | for p in dirs_data: 79 | """ 80 | Creates directory under data 81 | """ 82 | p.mkdir(parents = True, exist_ok = True) 83 | video_names.append(p.stem) 84 | 85 | dirs_results = [results_path/Path(i.stem) for i in videos] 86 | for p in dirs_results: 87 | """ 88 | Creates directory under results 89 | """ 90 | p.mkdir(parents = True, exist_ok = True) 91 | 92 | destinations = [video_path.joinpath(vp.name) for vp in videos] 93 | 94 | os.mkdir(str(project_path)+'/'+'videos/pose_estimation/') 95 | os.mkdir(str(project_path)+'/model/pretrained_model') 96 | 97 | print("Copying the videos \n") 98 | for src, dst in zip(videos, destinations): 99 | shutil.copy(os.fspath(src),os.fspath(dst)) 100 | 101 | cfg_file,ruamelFile = auxiliary.create_config_template() 102 | cfg_file 103 | 104 | cfg_file['Project']=str(project) 105 | cfg_file['project_path']=str(project_path)+'/' 106 | cfg_file['test_fraction']=0.1 107 | cfg_file['video_sets']=video_names 108 | cfg_file['all_data']='yes' 109 | cfg_file['load_data']='-PE-seq-clean' 110 | cfg_file['anneal_function']='linear' 111 | cfg_file['batch_size']=256 112 | cfg_file['max_epochs']=500 113 | cfg_file['transition_function']='GRU' 114 | cfg_file['beta']=1 115 | cfg_file['zdims']=30 116 | cfg_file['learning_rate']=5e-4 117 | cfg_file['time_window']=30 118 | cfg_file['prediction_decoder']=1 119 | cfg_file['prediction_steps']=15 120 | cfg_file['model_convergence']=50 121 | cfg_file['model_snapshot']=50 122 | cfg_file['num_features']=12 123 | cfg_file['savgol_filter']=True 124 | cfg_file['savgol_length']=5 125 | cfg_file['savgol_order']=2 126 | cfg_file['hidden_size_layer_1']=256 127 | cfg_file['hidden_size_layer_2']=256 128 | cfg_file['dropout_encoder']=0 129 | cfg_file['hidden_size_rec']=256 130 | cfg_file['dropout_rec']=0 131 | cfg_file['hidden_size_pred']=256 132 | cfg_file['dropout_pred']=0 133 | cfg_file['kl_start']=2 134 | cfg_file['annealtime']=4 135 | cfg_file['mse_reconstruction_reduction']='sum' 136 | cfg_file['mse_prediction_reduction']='sum' 137 | cfg_file['kmeans_loss']=cfg_file['zdims'] 138 | cfg_file['kmeans_lambda']=0.1 139 | cfg_file['scheduler']=1 140 | cfg_file['length_of_motif_video'] = 1000 141 | cfg_file['noise'] = False 142 | cfg_file['scheduler_step_size'] = 100 143 | cfg_file['legacy'] = False 144 | cfg_file['individual_parameterization'] = False 145 | cfg_file['random_state_kmeans'] = 42 146 | cfg_file['n_init_kmeans'] = 15 147 | cfg_file['model_name']='VAME' 148 | cfg_file['n_cluster'] = 15 149 | cfg_file['pretrained_weights'] = False 150 | cfg_file['pretrained_model']='None' 151 | cfg_file['min_dist'] = 0.1 152 | cfg_file['n_neighbors'] = 200 153 | cfg_file['random_state'] = 42 154 | cfg_file['num_points'] = 30000 155 | cfg_file['scheduler_gamma'] = 0.2 156 | cfg_file['softplus'] = False 157 | cfg_file['pose_confidence'] = 0.99 158 | cfg_file['iqr_factor'] = 4 159 | cfg_file['robust'] = True 160 | cfg_file['beta_norm'] = False 161 | cfg_file['n_layers'] = 1 162 | cfg_file['axis'] = None 163 | cfg_file['parameterization'] = 'hmm' 164 | cfg_file['hmm_trained'] = False 165 | 166 | projconfigfile=os.path.join(str(project_path),'config.yaml') 167 | # Write dictionary to yaml config file 168 | auxiliary.write_config(projconfigfile,cfg_file) 169 | 170 | print('A VAME project has been created. \n') 171 | print('Now its time to prepare your data for VAME. ' 172 | 'The first step is to move your pose .csv file (e.g. DeepLabCut .csv) into the ' 173 | '//YOUR//VAME//PROJECT//videos//pose_estimation folder. From here you can call ' 174 | 'either the function vame.egocentric_alignment() or if your data is by design egocentric ' 175 | 'call vame.csv_to_numpy(). This will prepare the data in .csv into the right format to start ' 176 | 'working with VAME.') 177 | 178 | return projconfigfile 179 | -------------------------------------------------------------------------------- /vame/model/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import sys 13 | sys.dont_write_bytecode = True 14 | 15 | from vame.model.create_training import create_trainset 16 | from vame.model.dataloader import SEQUENCE_DATASET 17 | from vame.model.rnn_vae import train_model 18 | from vame.model.evaluate import evaluate_model 19 | 20 | -------------------------------------------------------------------------------- /vame/model/create_training.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import numpy as np 14 | import pandas as pd 15 | from pathlib import Path 16 | import scipy.signal 17 | from scipy.stats import iqr 18 | import matplotlib.pyplot as plt 19 | 20 | from vame.util.auxiliary import read_config 21 | 22 | 23 | #Helper function to return indexes of nans 24 | def nan_helper(y): 25 | return np.isnan(y), lambda z: z.nonzero()[0] 26 | 27 | #Interpolates all nan values of given array 28 | def interpol(arr): 29 | y = np.transpose(arr) 30 | nans, x = nan_helper(y) 31 | y[nans]= np.interp(x(nans), x(~nans), y[~nans]) 32 | arr = np.transpose(y) 33 | return arr 34 | 35 | def plot_check_parameter(cfg, iqr_val, num_frames, X_true, X_med, anchor_1, anchor_2): 36 | plot_X_orig = np.concatenate(X_true, axis=0).T 37 | plot_X_med = X_med.copy() 38 | iqr_cutoff = cfg['iqr_factor']*iqr_val 39 | 40 | plt.figure() 41 | plt.plot(plot_X_orig.T) 42 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 43 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 44 | plt.title("Full Signal z-scored") 45 | plt.legend() 46 | 47 | if num_frames > 1000: 48 | rnd = np.random.choice(num_frames) 49 | 50 | plt.figure() 51 | plt.plot(plot_X_med[:,rnd:rnd+1000].T) 52 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 53 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 54 | plt.title("Filtered signal z-scored") 55 | plt.legend() 56 | 57 | plt.figure() 58 | plt.plot(plot_X_orig[:,rnd:rnd+1000].T) 59 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 60 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 61 | plt.title("Original signal z-scored") 62 | plt.legend() 63 | 64 | plt.figure() 65 | plt.plot(plot_X_orig[:,rnd:rnd+1000].T, 'g', alpha=0.5) 66 | plt.plot(plot_X_med[:,rnd:rnd+1000].T, '--m', alpha=0.6) 67 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 68 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 69 | plt.title("Overlayed z-scored") 70 | plt.legend() 71 | 72 | # plot_X_orig = np.delete(plot_X_orig.T, anchor_1, 1) 73 | # plot_X_orig = np.delete(plot_X_orig, anchor_2, 1) 74 | # mse = (np.square(plot_X_orig[rnd:rnd+1000, :] - plot_X_med[:,rnd:rnd+1000].T)).mean(axis=0) 75 | 76 | 77 | else: 78 | plt.figure() 79 | plt.plot(plot_X_med.T) 80 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 81 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 82 | plt.title("Filtered signal z-scored") 83 | plt.legend() 84 | 85 | plt.figure() 86 | plt.plot(plot_X_orig.T) 87 | plt.axhline(y=iqr_cutoff, color='r', linestyle='--', label="IQR cutoff") 88 | plt.axhline(y=-iqr_cutoff, color='r', linestyle='--') 89 | plt.title("Original signal z-scored") 90 | plt.legend() 91 | 92 | print("Please run the function with check_parameter=False if you are happy with the results") 93 | 94 | def traindata_aligned(cfg, files, testfraction, num_features, savgol_filter, check_parameter): 95 | 96 | X_train = [] 97 | pos = [] 98 | pos_temp = 0 99 | pos.append(0) 100 | 101 | if check_parameter == True: 102 | X_true = [] 103 | files = [files[0]] 104 | 105 | for file in files: 106 | print("z-scoring of file %s" %file) 107 | path_to_file = os.path.join(cfg['project_path'],"data", file, file+'-PE-seq.npy') 108 | data = np.load(path_to_file) 109 | 110 | X_mean = np.mean(data,axis=None) 111 | X_std = np.std(data, axis=None) 112 | X_z = (data.T - X_mean) / X_std 113 | 114 | # Introducing artificial error spikes 115 | # rang = [1.5, 2, 2.5, 3, 3.5, 3, 3, 2.5, 2, 1.5] 116 | # for i in range(num_frames): 117 | # if i % 300 == 0: 118 | # rnd = np.random.choice(12,2) 119 | # for j in range(10): 120 | # X_z[i+j, rnd[0]] = X_z[i+j, rnd[0]] * rang[j] 121 | # X_z[i+j, rnd[1]] = X_z[i+j, rnd[1]] * rang[j] 122 | 123 | if check_parameter == True: 124 | X_z_copy = X_z.copy() 125 | X_true.append(X_z_copy) 126 | 127 | if cfg['robust'] == True: 128 | iqr_val = iqr(X_z) 129 | print("IQR value: %.2f, IQR cutoff: %.2f" %(iqr_val, cfg['iqr_factor']*iqr_val)) 130 | for i in range(X_z.shape[0]): 131 | for marker in range(X_z.shape[1]): 132 | if X_z[i,marker] > cfg['iqr_factor']*iqr_val: 133 | X_z[i,marker] = np.nan 134 | 135 | elif X_z[i,marker] < -cfg['iqr_factor']*iqr_val: 136 | X_z[i,marker] = np.nan 137 | 138 | X_z = interpol(X_z) 139 | 140 | X_len = len(data.T) 141 | pos_temp += X_len 142 | pos.append(pos_temp) 143 | X_train.append(X_z) 144 | 145 | X = np.concatenate(X_train, axis=0) 146 | # X_std = np.std(X) 147 | 148 | detect_anchors = np.std(X.T, axis=1) 149 | sort_anchors = np.sort(detect_anchors) 150 | if sort_anchors[0] == sort_anchors[1]: 151 | anchors = np.where(detect_anchors == sort_anchors[0])[0] 152 | anchor_1_temp = anchors[0] 153 | anchor_2_temp = anchors[1] 154 | 155 | else: 156 | anchor_1_temp = int(np.where(detect_anchors == sort_anchors[0])[0]) 157 | anchor_2_temp = int(np.where(detect_anchors == sort_anchors[1])[0]) 158 | 159 | if anchor_1_temp > anchor_2_temp: 160 | anchor_1 = anchor_1_temp 161 | anchor_2 = anchor_2_temp 162 | 163 | else: 164 | anchor_1 = anchor_2_temp 165 | anchor_2 = anchor_1_temp 166 | 167 | X = np.delete(X, anchor_1, 1) 168 | X = np.delete(X, anchor_2, 1) 169 | 170 | X = X.T 171 | 172 | if savgol_filter: 173 | X_med = scipy.signal.savgol_filter(X, cfg['savgol_length'], cfg['savgol_order']) 174 | else: 175 | X_med = X 176 | 177 | num_frames = len(X_med.T) 178 | test = int(num_frames*testfraction) 179 | 180 | z_test =X_med[:,:test] 181 | z_train = X_med[:,test:] 182 | 183 | if check_parameter == True: 184 | plot_check_parameter(cfg, iqr_val, num_frames, X_true, X_med, anchor_1, anchor_2) 185 | 186 | else: 187 | #save numpy arrays the the test/train info: 188 | np.save(os.path.join(cfg['project_path'],"data", "train",'train_seq.npy'), z_train) 189 | np.save(os.path.join(cfg['project_path'],"data", "train", 'test_seq.npy'), z_test) 190 | 191 | for i, file in enumerate(files): 192 | np.save(os.path.join(cfg['project_path'],"data", file, file+'-PE-seq-clean.npy'), X_med[:,pos[i]:pos[i+1]]) 193 | 194 | print('Lenght of train data: %d' %len(z_train.T)) 195 | print('Lenght of test data: %d' %len(z_test.T)) 196 | 197 | 198 | def traindata_fixed(cfg, files, testfraction, num_features, savgol_filter, check_parameter): 199 | X_train = [] 200 | pos = [] 201 | pos_temp = 0 202 | pos.append(0) 203 | 204 | if check_parameter == True: 205 | X_true = [] 206 | rnd_file = np.random.choice(len(files)) 207 | files = [files[0]] 208 | 209 | for file in files: 210 | print("z-scoring of file %s" %file) 211 | path_to_file = os.path.join(cfg['project_path'],"data", file, file+'-PE-seq.npy') 212 | data = np.load(path_to_file) 213 | X_mean = np.mean(data,axis=None) 214 | X_std = np.std(data, axis=None) 215 | X_z = (data.T - X_mean) / X_std 216 | 217 | if check_parameter == True: 218 | X_z_copy = X_z.copy() 219 | X_true.append(X_z_copy) 220 | 221 | if cfg['robust'] == True: 222 | iqr_val = iqr(X_z) 223 | print("IQR value: %.2f, IQR cutoff: %.2f" %(iqr_val, cfg['iqr_factor']*iqr_val)) 224 | for i in range(X_z.shape[0]): 225 | for marker in range(X_z.shape[1]): 226 | if X_z[i,marker] > cfg['iqr_factor']*iqr_val: 227 | X_z[i,marker] = np.nan 228 | 229 | elif X_z[i,marker] < -cfg['iqr_factor']*iqr_val: 230 | X_z[i,marker] = np.nan 231 | 232 | X_z[i,:] = interpol(X_z[i,:]) 233 | 234 | X_len = len(data.T) 235 | pos_temp += X_len 236 | pos.append(pos_temp) 237 | X_train.append(X_z) 238 | 239 | X = np.concatenate(X_train, axis=0).T 240 | 241 | if savgol_filter: 242 | X_med = scipy.signal.savgol_filter(X, cfg['savgol_length'], cfg['savgol_order']) 243 | else: 244 | X_med = X 245 | 246 | num_frames = len(X_med.T) 247 | test = int(num_frames*testfraction) 248 | 249 | z_test =X_med[:,:test] 250 | z_train = X_med[:,test:] 251 | 252 | if check_parameter == True: 253 | plot_check_parameter(cfg, iqr_val, num_frames, X_true, X_med) 254 | 255 | else: 256 | #save numpy arrays the the test/train info: 257 | np.save(os.path.join(cfg['project_path'],"data", "train",'train_seq.npy'), z_train) 258 | np.save(os.path.join(cfg['project_path'],"data", "train", 'test_seq.npy'), z_test) 259 | 260 | for i, file in enumerate(files): 261 | np.save(os.path.join(cfg['project_path'],"data", file, file+'-PE-seq-clean.npy'), X_med[:,pos[i]:pos[i+1]]) 262 | 263 | print('Lenght of train data: %d' %len(z_train.T)) 264 | print('Lenght of test data: %d' %len(z_test.T)) 265 | 266 | 267 | def create_trainset(config, check_parameter=False): 268 | config_file = Path(config).resolve() 269 | cfg = read_config(config_file) 270 | legacy = cfg['legacy'] 271 | fixed = cfg['egocentric_data'] 272 | 273 | if not os.path.exists(os.path.join(cfg['project_path'],'data','train',"")): 274 | os.mkdir(os.path.join(cfg['project_path'],'data','train',"")) 275 | 276 | files = [] 277 | if cfg['all_data'] == 'No': 278 | for file in cfg['video_sets']: 279 | use_file = input("Do you want to train on " + file + "? yes/no: ") 280 | if use_file == 'yes': 281 | files.append(file) 282 | if use_file == 'no': 283 | continue 284 | else: 285 | for file in cfg['video_sets']: 286 | files.append(file) 287 | 288 | print("Creating training dataset...") 289 | if cfg['robust'] == True: 290 | print("Using robust setting to eliminate outliers! IQR factor: %d" %cfg['iqr_factor']) 291 | 292 | if fixed == False: 293 | print("Creating trainset from the vame.egocentrical_alignment() output ") 294 | traindata_aligned(cfg, files, cfg['test_fraction'], cfg['num_features'], cfg['savgol_filter'], check_parameter) 295 | else: 296 | print("Creating trainset from the vame.csv_to_numpy() output ") 297 | traindata_fixed(cfg, files, cfg['test_fraction'], cfg['num_features'], cfg['savgol_filter'], check_parameter) 298 | 299 | if check_parameter == False: 300 | print("A training and test set has been created. Next step: vame.train_model()") 301 | -------------------------------------------------------------------------------- /vame/model/dataloader.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import torch 13 | from torch.utils.data.dataset import Dataset 14 | import numpy as np 15 | import os 16 | 17 | 18 | class SEQUENCE_DATASET(Dataset): 19 | def __init__(self,path_to_file,data,train,temporal_window): 20 | self.temporal_window = temporal_window 21 | self.X = np.load(path_to_file+data) 22 | if self.X.shape[0] > self.X.shape[1]: 23 | self.X=self.X.T 24 | 25 | self.data_points = len(self.X[0,:]) 26 | 27 | if train and not os.path.exists(os.path.join(path_to_file,'seq_mean.npy')): 28 | print("Compute mean and std for temporal dataset.") 29 | self.mean = np.mean(self.X) 30 | self.std = np.std(self.X) 31 | np.save(path_to_file+'seq_mean.npy', self.mean) 32 | np.save(path_to_file+'seq_std.npy', self.std) 33 | else: 34 | self.mean = np.load(path_to_file+'seq_mean.npy') 35 | self.std = np.load(path_to_file+'seq_std.npy') 36 | 37 | if train: 38 | print('Initialize train data. Datapoints %d' %self.data_points) 39 | else: 40 | print('Initialize test data. Datapoints %d' %self.data_points) 41 | 42 | def __len__(self): 43 | return self.data_points 44 | 45 | def __getitem__(self, index): 46 | temp_window = self.temporal_window 47 | 48 | nf = self.data_points 49 | start = np.random.choice(nf-temp_window) 50 | end = start+temp_window 51 | 52 | sequence = self.X[:,start:end] 53 | 54 | sequence = (sequence-self.mean)/self.std 55 | 56 | return torch.from_numpy(sequence) 57 | 58 | 59 | 60 | 61 | 62 | -------------------------------------------------------------------------------- /vame/model/evaluate.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import torch 14 | import numpy as np 15 | from pathlib import Path 16 | from matplotlib import pyplot as plt 17 | import torch.utils.data as Data 18 | 19 | from vame.util.auxiliary import read_config 20 | from vame.model.rnn_vae import RNN_VAE 21 | from vame.model.dataloader import SEQUENCE_DATASET 22 | 23 | use_gpu = torch.cuda.is_available() 24 | if use_gpu: 25 | pass 26 | else: 27 | torch.device("cpu") 28 | 29 | 30 | def plot_reconstruction(filepath, test_loader, seq_len_half, model, model_name, 31 | FUTURE_DECODER, FUTURE_STEPS, suffix=None): 32 | #x = test_loader.__iter__().next() 33 | dataiter = iter(test_loader) 34 | x = next(dataiter) 35 | x = x.permute(0,2,1) 36 | if use_gpu: 37 | data = x[:,:seq_len_half,:].type('torch.FloatTensor').cuda() 38 | data_fut = x[:,seq_len_half:seq_len_half+FUTURE_STEPS,:].type('torch.FloatTensor').cuda() 39 | else: 40 | data = x[:,:seq_len_half,:].type('torch.FloatTensor').to() 41 | data_fut = x[:,seq_len_half:seq_len_half+FUTURE_STEPS,:].type('torch.FloatTensor').to() 42 | if FUTURE_DECODER: 43 | x_tilde, future, latent, mu, logvar = model(data) 44 | 45 | fut_orig = data_fut.cpu() 46 | fut_orig = fut_orig.data.numpy() 47 | fut = future.cpu() 48 | fut = fut.detach().numpy() 49 | 50 | else: 51 | x_tilde, latent, mu, logvar = model(data) 52 | 53 | data_orig = data.cpu() 54 | data_orig = data_orig.data.numpy() 55 | data_tilde = x_tilde.cpu() 56 | data_tilde = data_tilde.detach().numpy() 57 | 58 | if FUTURE_DECODER: 59 | fig, axs = plt.subplots(2, 5) 60 | fig.suptitle('Reconstruction [top] and future prediction [bottom] of input sequence') 61 | for i in range(5): 62 | axs[0,i].plot(data_orig[i,...], color='k', label='Sequence Data') 63 | axs[0,i].plot(data_tilde[i,...], color='r', linestyle='dashed', label='Sequence Reconstruction') 64 | 65 | axs[1,i].plot(fut_orig[i,...], color='k') 66 | axs[1,i].plot(fut[i,...], color='r', linestyle='dashed') 67 | axs[0,0].set(xlabel='time steps', ylabel='reconstruction') 68 | axs[1,0].set(xlabel='time steps', ylabel='predction') 69 | fig.savefig(os.path.join(filepath,"evaluate",'Future_Reconstruction.png')) 70 | 71 | else: 72 | fig, ax1 = plt.subplots(1, 5) 73 | for i in range(5): 74 | fig.suptitle('Reconstruction of input sequence') 75 | ax1[i].plot(data_orig[i,...], color='k', label='Sequence Data') 76 | ax1[i].plot(data_tilde[i,...], color='r', linestyle='dashed', label='Sequence Reconstruction') 77 | fig.set_tight_layout(True) 78 | if not suffix: 79 | fig.savefig(os.path.join(filepath,'evaluate','Reconstruction_'+model_name+'.png'), bbox_inches='tight') 80 | elif suffix: 81 | fig.savefig(os.path.join(filepath,'evaluate','Reconstruction_'+model_name+'_'+suffix+'.png'), bbox_inches='tight') 82 | 83 | 84 | def plot_loss(cfg, filepath, model_name): 85 | basepath = os.path.join(cfg['project_path'],"model","model_losses") 86 | train_loss = np.load(os.path.join(basepath,'train_losses_'+model_name+'.npy')) 87 | test_loss = np.load(os.path.join(basepath,'test_losses_'+model_name+'.npy')) 88 | mse_loss_train = np.load(os.path.join(basepath,'mse_train_losses_'+model_name+'.npy')) 89 | mse_loss_test = np.load(os.path.join(basepath,'mse_test_losses_'+model_name+'.npy')) 90 | # km_loss = np.load(os.path.join(basepath,'kmeans_losses_'+model_name+'.npy'), allow_pickle=True) 91 | km_losses = np.load(os.path.join(basepath,'kmeans_losses_'+model_name+'.npy')) 92 | kl_loss = np.load(os.path.join(basepath,'kl_losses_'+model_name+'.npy')) 93 | fut_loss = np.load(os.path.join(basepath,'fut_losses_'+model_name+'.npy')) 94 | 95 | # km_losses = [] 96 | # for i in range(len(km_loss)): 97 | # km = km_loss[i].cpu().detach().numpy() 98 | # km_losses.append(km) 99 | 100 | fig, (ax1) = plt.subplots(1, 1) 101 | fig.suptitle('Losses of our Model') 102 | ax1.set(xlabel='Epochs', ylabel='loss [log-scale]') 103 | ax1.set_yscale("log") 104 | ax1.plot(train_loss, label='Train-Loss') 105 | ax1.plot(test_loss, label='Test-Loss') 106 | ax1.plot(mse_loss_train, label='MSE-Train-Loss') 107 | ax1.plot(mse_loss_test, label='MSE-Test-Loss') 108 | ax1.plot(km_losses, label='KMeans-Loss') 109 | ax1.plot(kl_loss, label='KL-Loss') 110 | ax1.plot(fut_loss, label='Prediction-Loss') 111 | ax1.legend() 112 | #fig.savefig(filepath+'evaluate/'+'MSE-and-KL-Loss'+model_name+'.png') 113 | fig.savefig(os.path.join(filepath,"evaluate",'MSE-and-KL-Loss'+model_name+'.png')) 114 | 115 | 116 | def eval_temporal(cfg, use_gpu, model_name, fixed, snapshot=None, suffix=None): 117 | SEED = 19 118 | ZDIMS = cfg['zdims'] 119 | FUTURE_DECODER = cfg['prediction_decoder'] 120 | TEMPORAL_WINDOW = cfg['time_window']*2 121 | FUTURE_STEPS = cfg['prediction_steps'] 122 | NUM_FEATURES = cfg['num_features'] 123 | if fixed == False: 124 | NUM_FEATURES = NUM_FEATURES - 2 125 | TEST_BATCH_SIZE = 64 126 | PROJECT_PATH = cfg['project_path'] 127 | hidden_size_layer_1 = cfg['hidden_size_layer_1'] 128 | hidden_size_layer_2 = cfg['hidden_size_layer_2'] 129 | hidden_size_rec = cfg['hidden_size_rec'] 130 | hidden_size_pred = cfg['hidden_size_pred'] 131 | dropout_encoder = cfg['dropout_encoder'] 132 | dropout_rec = cfg['dropout_rec'] 133 | dropout_pred = cfg['dropout_pred'] 134 | softplus = cfg['softplus'] 135 | 136 | filepath = os.path.join(cfg['project_path'],"model") 137 | 138 | 139 | seq_len_half = int(TEMPORAL_WINDOW/2) 140 | if use_gpu: 141 | torch.cuda.manual_seed(SEED) 142 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 143 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 144 | dropout_rec, dropout_pred, softplus).cuda() 145 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model",model_name+'_'+cfg['Project']+'.pkl'))) 146 | else: 147 | model = RNN_VAE(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 148 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 149 | dropout_rec, dropout_pred, softplus).to() 150 | if not snapshot: 151 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],"model","best_model",model_name+'_'+cfg['Project']+'.pkl'), map_location=torch.device('cpu'))) 152 | elif snapshot: 153 | model.load_state_dict(torch.load(snapshot), map_location=torch.device('cpu')) 154 | model.eval() #toggle evaluation mode 155 | 156 | testset = SEQUENCE_DATASET(os.path.join(cfg['project_path'],"data", "train",""), data='test_seq.npy', train=False, temporal_window=TEMPORAL_WINDOW) 157 | test_loader = Data.DataLoader(testset, batch_size=TEST_BATCH_SIZE, shuffle=True, drop_last=True) 158 | 159 | if not snapshot: 160 | plot_reconstruction(filepath, test_loader, seq_len_half, model, model_name, FUTURE_DECODER, FUTURE_STEPS)#, suffix=suffix 161 | elif snapshot: 162 | plot_reconstruction(filepath, test_loader, seq_len_half, model, model_name, FUTURE_DECODER, FUTURE_STEPS, suffix=suffix)#, 163 | if use_gpu: 164 | plot_loss(cfg, filepath, model_name) 165 | else: 166 | plot_loss(cfg, filepath, model_name) 167 | # pass #note, loading of losses needs to be adapted for CPU use #TODO 168 | 169 | 170 | def evaluate_model(config, use_snapshots=False): 171 | """ 172 | Evaluation of testset. 173 | 174 | Parameters 175 | ---------- 176 | config : str 177 | Path to config file. 178 | model_name : str 179 | name of model (same as in config.yaml) 180 | use_snapshots : bool 181 | Whether to plot for all snapshots or only the best model. 182 | """ 183 | config_file = Path(config).resolve() 184 | cfg = read_config(config_file) 185 | #legacy = cfg['legacy'] 186 | model_name = cfg['model_name'] 187 | fixed = cfg['egocentric_data'] 188 | 189 | if not os.path.exists(os.path.join(cfg['project_path'],"model","evaluate")): 190 | os.mkdir(os.path.join(cfg['project_path'],"model","evaluate")) 191 | 192 | use_gpu = torch.cuda.is_available() 193 | if use_gpu: 194 | print("Using CUDA") 195 | print('GPU active:',torch.cuda.is_available()) 196 | print('GPU used:',torch.cuda.get_device_name(0)) 197 | else: 198 | torch.device("cpu") 199 | print("CUDA is not working, or a GPU is not found; using CPU!") 200 | 201 | print("\n\nEvaluation of %s model. \n" %model_name) 202 | if not use_snapshots: 203 | eval_temporal(cfg, use_gpu, model_name, fixed)#suffix=suffix 204 | elif use_snapshots: 205 | snapshots=os.listdir(os.path.join(cfg['project_path'],'model','best_model','snapshots')) 206 | for snap in snapshots: 207 | fullpath = os.path.join(cfg['project_path'],"model","best_model","snapshots",snap) 208 | epoch=snap.split('_')[-1] 209 | eval_temporal(cfg, use_gpu, model_name, fixed, snapshot=fullpath, suffix='snapshot'+str(epoch)) 210 | #eval_temporal(cfg, use_gpu, model_name, legacy=legacy, suffix='bestModel') 211 | 212 | print("You can find the results of the evaluation in '/Your-VAME-Project-Apr30-2020/model/evaluate/' \n" 213 | "OPTIONS:\n" 214 | "- vame.pose_segmentation() to identify behavioral motifs.\n" 215 | "- re-run the model for further fine tuning. Check again with vame.evaluate_model()") 216 | -------------------------------------------------------------------------------- /vame/model/rnn_model.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | Variational Animal Motion Embedding 0.1 Toolbox 4 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 5 | Leibniz Institute for Neurobiology, Magdeburg, Germany 6 | 7 | https://github.com/LINCellularNeuroscience/VAME 8 | Licensed under GNU General Public License v3.0 9 | 10 | The Model is partially adapted from the Timeseries Clustering repository developed by Tejas Lodaya: 11 | https://github.com/tejaslodaya/timeseries-clustering-vae/blob/master/vrae/vrae.py 12 | """ 13 | 14 | 15 | import torch 16 | from torch import nn 17 | from torch.autograd import Variable 18 | 19 | 20 | # NEW MODEL WITH SMALL ALTERATIONS 21 | """ MODEL """ 22 | 23 | class Encoder(nn.Module): 24 | def __init__(self, NUM_FEATURES, hidden_size_layer_1, hidden_size_layer_2, dropout_encoder): 25 | super(Encoder, self).__init__() 26 | 27 | self.input_size = NUM_FEATURES 28 | self.hidden_size = hidden_size_layer_1 29 | self.hidden_size_2 = hidden_size_layer_2 30 | self.n_layers = 2 31 | self.dropout = dropout_encoder 32 | self.bidirectional = True 33 | 34 | self.encoder_rnn = nn.GRU(input_size=self.input_size, hidden_size=self.hidden_size, num_layers=self.n_layers, 35 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=self.bidirectional)#UNRELEASED! 36 | 37 | 38 | self.hidden_factor = (2 if self.bidirectional else 1) * self.n_layers 39 | 40 | def forward(self, inputs): 41 | outputs_1, hidden_1 = self.encoder_rnn(inputs)#UNRELEASED! 42 | 43 | hidden = torch.cat((hidden_1[0,...], hidden_1[1,...], hidden_1[2,...], hidden_1[3,...]),1) 44 | 45 | return hidden 46 | 47 | 48 | class Lambda(nn.Module): 49 | def __init__(self,ZDIMS, hidden_size_layer_1, hidden_size_layer_2, softplus): 50 | super(Lambda, self).__init__() 51 | 52 | self.hid_dim = hidden_size_layer_1*4 53 | self.latent_length = ZDIMS 54 | self.softplus = softplus 55 | 56 | self.hidden_to_mean = nn.Linear(self.hid_dim, self.latent_length) 57 | self.hidden_to_logvar = nn.Linear(self.hid_dim, self.latent_length) 58 | 59 | if self.softplus == True: 60 | print("Using a softplus activation to ensures that the variance is parameterized as non-negative and activated by a smooth function") 61 | self.softplus_fn = nn.Softplus() 62 | 63 | def forward(self, hidden): 64 | 65 | self.mean = self.hidden_to_mean(hidden) 66 | if self.softplus == True: 67 | self.logvar = self.softplus_fn(self.hidden_to_logvar(hidden)) 68 | else: 69 | self.logvar = self.hidden_to_logvar(hidden) 70 | 71 | if self.training: 72 | std = torch.exp(0.5 * self.logvar) 73 | eps = torch.randn_like(std) 74 | return eps.mul(std).add_(self.mean), self.mean, self.logvar 75 | else: 76 | return self.mean, self.mean, self.logvar 77 | 78 | 79 | class Decoder(nn.Module): 80 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES, hidden_size_rec, dropout_rec): 81 | super(Decoder,self).__init__() 82 | 83 | self.num_features = NUM_FEATURES 84 | self.sequence_length = TEMPORAL_WINDOW 85 | self.hidden_size = hidden_size_rec 86 | self.latent_length = ZDIMS 87 | self.n_layers = 1 88 | self.dropout = dropout_rec 89 | self.bidirectional = True 90 | 91 | self.rnn_rec = nn.GRU(self.latent_length, hidden_size=self.hidden_size, num_layers=self.n_layers, 92 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=self.bidirectional) 93 | 94 | self.hidden_factor = (2 if self.bidirectional else 1) * self.n_layers # NEW 95 | 96 | self.latent_to_hidden = nn.Linear(self.latent_length,self.hidden_size * self.hidden_factor) # NEW 97 | self.hidden_to_output = nn.Linear(self.hidden_size*(2 if self.bidirectional else 1), self.num_features) 98 | 99 | def forward(self, inputs, z): 100 | batch_size = inputs.size(0) # NEW 101 | 102 | hidden = self.latent_to_hidden(z) # NEW 103 | 104 | hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size) # NEW 105 | 106 | decoder_output, _ = self.rnn_rec(inputs, hidden) 107 | prediction = self.hidden_to_output(decoder_output) 108 | 109 | return prediction 110 | 111 | 112 | class Decoder_Future(nn.Module): 113 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_STEPS, hidden_size_pred, dropout_pred): 114 | super(Decoder_Future,self).__init__() 115 | 116 | self.num_features = NUM_FEATURES 117 | self.future_steps = FUTURE_STEPS 118 | self.sequence_length = TEMPORAL_WINDOW 119 | self.hidden_size = hidden_size_pred 120 | self.latent_length = ZDIMS 121 | self.n_layers = 1 122 | self.dropout = dropout_pred 123 | self.bidirectional = True 124 | 125 | self.rnn_pred = nn.GRU(self.latent_length, hidden_size=self.hidden_size, num_layers=self.n_layers, 126 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=self.bidirectional) 127 | 128 | self.hidden_factor = (2 if self.bidirectional else 1) * self.n_layers # NEW 129 | 130 | self.latent_to_hidden = nn.Linear(self.latent_length,self.hidden_size * self.hidden_factor) 131 | self.hidden_to_output = nn.Linear(self.hidden_size*2, self.num_features) 132 | 133 | def forward(self, inputs, z): 134 | batch_size = inputs.size(0) 135 | 136 | hidden = self.latent_to_hidden(z) 137 | hidden = hidden.view(self.hidden_factor, batch_size, self.hidden_size) 138 | 139 | inputs = inputs[:,:self.future_steps,:] 140 | decoder_output, _ = self.rnn_pred(inputs, hidden) 141 | 142 | prediction = self.hidden_to_output(decoder_output) 143 | 144 | return prediction 145 | 146 | 147 | class RNN_VAE(nn.Module): 148 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 149 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 150 | dropout_rec, dropout_pred, softplus): 151 | super(RNN_VAE,self).__init__() 152 | 153 | self.FUTURE_DECODER = FUTURE_DECODER 154 | self.seq_len = int(TEMPORAL_WINDOW / 2) 155 | self.encoder = Encoder(NUM_FEATURES, hidden_size_layer_1, hidden_size_layer_2, dropout_encoder) 156 | self.lmbda = Lambda(ZDIMS, hidden_size_layer_1, hidden_size_layer_2, softplus) 157 | self.decoder = Decoder(self.seq_len,ZDIMS,NUM_FEATURES, hidden_size_rec, dropout_rec) 158 | if FUTURE_DECODER: 159 | self.decoder_future = Decoder_Future(self.seq_len,ZDIMS,NUM_FEATURES,FUTURE_STEPS, hidden_size_pred, 160 | dropout_pred) 161 | 162 | def forward(self,seq): 163 | 164 | """ Encode input sequence """ 165 | h_n = self.encoder(seq) 166 | 167 | """ Compute the latent state via reparametrization trick """ 168 | z, mu, logvar = self.lmbda(h_n) 169 | ins = z.unsqueeze(2).repeat(1, 1, self.seq_len) 170 | ins = ins.permute(0,2,1) 171 | 172 | """ Predict the future of the sequence from the latent state""" 173 | prediction = self.decoder(ins, z) 174 | 175 | if self.FUTURE_DECODER: 176 | future = self.decoder_future(ins, z) 177 | return prediction, future, z, mu, logvar 178 | else: 179 | return prediction, z, mu, logvar 180 | 181 | 182 | #---------------------------------------------------------------------------------------- 183 | # LEGACY MODEL | 184 | #---------------------------------------------------------------------------------------- 185 | 186 | 187 | """ MODEL """ 188 | class Encoder_LEGACY(nn.Module): 189 | def __init__(self, NUM_FEATURES, hidden_size_layer_1, hidden_size_layer_2, dropout_encoder): 190 | super(Encoder_LEGACY, self).__init__() 191 | 192 | self.input_size = NUM_FEATURES 193 | self.hidden_size = hidden_size_layer_1 194 | self.hidden_size_2 = hidden_size_layer_2 195 | self.n_layers = 1 196 | self.dropout = dropout_encoder 197 | 198 | self.rnn_1 = nn.GRU(input_size=self.input_size, hidden_size=self.hidden_size, num_layers=self.n_layers, 199 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=True) 200 | 201 | self.rnn_2 = nn.GRU(input_size=self.hidden_size*2, hidden_size=self.hidden_size_2, num_layers=self.n_layers, 202 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=True) 203 | 204 | def forward(self, inputs): 205 | outputs_1, hidden_1 = self.rnn_1(inputs) 206 | outputs_2, hidden_2 = self.rnn_2(outputs_1) 207 | 208 | h_n_1 = torch.cat((hidden_1[0,...], hidden_1[1,...]), 1) 209 | h_n_2 = torch.cat((hidden_2[0,...], hidden_2[1,...]), 1) 210 | 211 | h_n = torch.cat((h_n_1, h_n_2), 1) 212 | 213 | return h_n 214 | 215 | 216 | class Lambda_LEGACY(nn.Module): 217 | def __init__(self,ZDIMS, hidden_size_layer_1, hidden_size_layer_2): 218 | super(Lambda_LEGACY, self).__init__() 219 | 220 | self.hid_dim = hidden_size_layer_1*2 + hidden_size_layer_2*2 221 | self.latent_length = ZDIMS 222 | 223 | self.hidden_to_linear = nn.Linear(self.hid_dim, self.hid_dim) 224 | self.hidden_to_mean = nn.Linear(self.hid_dim, self.latent_length) 225 | self.hidden_to_logvar = nn.Linear(self.hid_dim, self.latent_length) 226 | 227 | self.softplus = nn.Softplus() 228 | 229 | def forward(self, cell_output): 230 | self.latent_mean = self.hidden_to_mean(cell_output) 231 | 232 | # based on Pereira et al 2019: 233 | # "The SoftPlus function ensures that the variance is parameterized as non-negative and activated 234 | # by a smooth function 235 | self.latent_logvar = self.softplus(self.hidden_to_logvar(cell_output)) 236 | 237 | if self.training: 238 | std = self.latent_logvar.mul(0.5).exp_() 239 | eps = Variable(std.data.new(std.size()).normal_()) 240 | return eps.mul(std).add_(self.latent_mean), self.latent_mean, self.latent_logvar 241 | else: 242 | return self.latent_mean, self.latent_mean, self.latent_logvar 243 | 244 | 245 | class Decoder_LEGACY(nn.Module): 246 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES, hidden_size_rec, dropout_rec): 247 | super(Decoder_LEGACY,self).__init__() 248 | 249 | self.num_features = NUM_FEATURES 250 | self.sequence_length = TEMPORAL_WINDOW 251 | self.hidden_size = hidden_size_rec 252 | self.latent_length = ZDIMS 253 | self.n_layers = 1 254 | self.dropout = dropout_rec 255 | 256 | self.rnn_rec = nn.GRU(self.latent_length, hidden_size=self.hidden_size, num_layers=self.n_layers, 257 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=False) 258 | 259 | self.hidden_to_output = nn.Linear(self.hidden_size, self.num_features) 260 | 261 | def forward(self, inputs): 262 | decoder_output, _ = self.rnn_rec(inputs) 263 | prediction = self.hidden_to_output(decoder_output) 264 | 265 | return prediction 266 | 267 | class Decoder_Future_LEGACY(nn.Module): 268 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_STEPS, hidden_size_pred, dropout_pred): 269 | super(Decoder_Future_LEGACY,self).__init__() 270 | 271 | self.num_features = NUM_FEATURES 272 | self.future_steps = FUTURE_STEPS 273 | self.sequence_length = TEMPORAL_WINDOW 274 | self.hidden_size = hidden_size_pred 275 | self.latent_length = ZDIMS 276 | self.n_layers = 1 277 | self.dropout = dropout_pred 278 | 279 | self.rnn_pred = nn.GRU(self.latent_length, hidden_size=self.hidden_size, num_layers=self.n_layers, 280 | bias=True, batch_first=True, dropout=self.dropout, bidirectional=True) 281 | 282 | self.hidden_to_output = nn.Linear(self.hidden_size*2, self.num_features) 283 | 284 | def forward(self, inputs): 285 | inputs = inputs[:,:self.future_steps,:] 286 | decoder_output, _ = self.rnn_pred(inputs) 287 | prediction = self.hidden_to_output(decoder_output) 288 | 289 | return prediction 290 | 291 | 292 | class RNN_VAE_LEGACY(nn.Module): 293 | def __init__(self,TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 294 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 295 | dropout_rec, dropout_pred, softplus): 296 | super(RNN_VAE_LEGACY,self).__init__() 297 | 298 | self.FUTURE_DECODER = FUTURE_DECODER 299 | self.seq_len = int(TEMPORAL_WINDOW / 2) 300 | self.encoder = Encoder_LEGACY(NUM_FEATURES, hidden_size_layer_1, hidden_size_layer_2, dropout_encoder) 301 | self.lmbda = Lambda_LEGACY(ZDIMS, hidden_size_layer_1, hidden_size_layer_2) 302 | self.decoder = Decoder_LEGACY(self.seq_len,ZDIMS,NUM_FEATURES, hidden_size_rec, dropout_rec) 303 | if FUTURE_DECODER: 304 | self.decoder_future = Decoder_Future_LEGACY(self.seq_len,ZDIMS,NUM_FEATURES,FUTURE_STEPS, hidden_size_pred, 305 | dropout_pred) 306 | 307 | def forward(self,seq): 308 | 309 | """ Encode input sequence """ 310 | h_n = self.encoder(seq) 311 | 312 | """ Compute the latent state via reparametrization trick """ 313 | latent, mu, logvar = self.lmbda(h_n) 314 | z = latent.unsqueeze(2).repeat(1, 1, self.seq_len) 315 | z = z.permute(0,2,1) 316 | 317 | """ Predict the future of the sequence from the latent state""" 318 | prediction = self.decoder(z) 319 | 320 | if self.FUTURE_DECODER: 321 | future = self.decoder_future(z) 322 | return prediction, future, latent, mu, logvar 323 | else: 324 | return prediction, latent, mu, logvar 325 | 326 | -------------------------------------------------------------------------------- /vame/model/rnn_vae.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import torch 13 | from torch import nn 14 | import torch.utils.data as Data 15 | from torch.autograd import Variable 16 | from torch.optim.lr_scheduler import StepLR, ReduceLROnPlateau 17 | 18 | import os 19 | import numpy as np 20 | from pathlib import Path 21 | 22 | from vame.util.auxiliary import read_config 23 | from vame.model.dataloader import SEQUENCE_DATASET 24 | from vame.model.rnn_model import RNN_VAE, RNN_VAE_LEGACY 25 | 26 | # make sure torch uses cuda for GPU computing 27 | use_gpu = torch.cuda.is_available() 28 | if use_gpu: 29 | print("Using CUDA") 30 | print('GPU active:',torch.cuda.is_available()) 31 | print('GPU used:',torch.cuda.get_device_name(0)) 32 | else: 33 | torch.device("cpu") 34 | 35 | def reconstruction_loss(x, x_tilde, reduction): 36 | mse_loss = nn.MSELoss(reduction=reduction) 37 | rec_loss = mse_loss(x_tilde,x) 38 | return rec_loss 39 | 40 | def future_reconstruction_loss(x, x_tilde, reduction): 41 | mse_loss = nn.MSELoss(reduction=reduction) 42 | rec_loss = mse_loss(x_tilde,x) 43 | return rec_loss 44 | 45 | def cluster_loss(H, kloss, lmbda, batch_size): 46 | gram_matrix = (H.T @ H) / batch_size 47 | _ ,sv_2, _ = torch.svd(gram_matrix) 48 | sv = torch.sqrt(sv_2[:kloss]) 49 | loss = torch.sum(sv) 50 | return lmbda*loss 51 | 52 | 53 | def kullback_leibler_loss(mu, logvar): 54 | # see Appendix B from VAE paper: 55 | # Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014 56 | # https://arxiv.org/abs/1312.6114 57 | # 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2) 58 | # I'm using torch.mean() here as the sum() version depends on the size of the latent vector 59 | KLD = -0.5 * torch.mean(1 + logvar - mu.pow(2) - logvar.exp()) 60 | return KLD 61 | 62 | 63 | def kl_annealing(epoch, kl_start, annealtime, function): 64 | """ 65 | Annealing of Kullback-Leibler loss to let the model learn first 66 | the reconstruction of the data before the KL loss term gets introduced. 67 | """ 68 | if epoch > kl_start: 69 | if function == 'linear': 70 | new_weight = min(1, (epoch-kl_start)/(annealtime)) 71 | 72 | elif function == 'sigmoid': 73 | new_weight = float(1/(1+np.exp(-0.9*(epoch-annealtime)))) 74 | else: 75 | raise NotImplementedError('currently only "linear" and "sigmoid" are implemented') 76 | 77 | return new_weight 78 | 79 | else: 80 | new_weight = 0 81 | return new_weight 82 | 83 | 84 | def gaussian(ins, is_training, seq_len, std_n=0.8): 85 | if is_training: 86 | emp_std = ins.std(1)*std_n 87 | emp_std = emp_std.unsqueeze(2).repeat(1, 1, seq_len) 88 | emp_std = emp_std.permute(0,2,1) 89 | noise = Variable(ins.data.new(ins.size()).normal_(0, 1)) 90 | return ins + (noise*emp_std) 91 | return ins 92 | 93 | 94 | def train(train_loader, epoch, model, optimizer, anneal_function, BETA, kl_start, 95 | annealtime, seq_len, future_decoder, future_steps, scheduler, mse_red, 96 | mse_pred, kloss, klmbda, bsize, noise): 97 | model.train() # toggle model to train mode 98 | train_loss = 0.0 99 | mse_loss = 0.0 100 | kullback_loss = 0.0 101 | kmeans_losses = 0.0 102 | fut_loss = 0.0 103 | loss = 0.0 104 | seq_len_half = int(seq_len / 2) 105 | 106 | for idx, data_item in enumerate(train_loader): 107 | data_item = Variable(data_item) 108 | data_item = data_item.permute(0,2,1) 109 | 110 | if use_gpu: 111 | data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').cuda() 112 | fut = data_item[:,seq_len_half:seq_len_half+future_steps,:].type('torch.FloatTensor').cuda() 113 | else: 114 | data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').to() 115 | fut = data_item[:,seq_len_half:seq_len_half+future_steps,:].type('torch.FloatTensor').to() 116 | if noise == True: 117 | data_gaussian = gaussian(data,True,seq_len_half) 118 | else: 119 | data_gaussian = data 120 | 121 | if future_decoder: 122 | data_tilde, future, latent, mu, logvar = model(data_gaussian) 123 | 124 | rec_loss = reconstruction_loss(data, data_tilde, mse_red) 125 | fut_rec_loss = future_reconstruction_loss(fut, future, mse_pred) 126 | kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize) 127 | kl_loss = kullback_leibler_loss(mu, logvar) 128 | kl_weight = kl_annealing(epoch, kl_start, annealtime, anneal_function) 129 | loss = rec_loss + fut_rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss 130 | fut_loss += fut_rec_loss.item() 131 | 132 | else: 133 | data_tilde, latent, mu, logvar = model(data_gaussian) 134 | 135 | rec_loss = reconstruction_loss(data, data_tilde, mse_red) 136 | kl_loss = kullback_leibler_loss(mu, logvar) 137 | kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize) 138 | kl_weight = kl_annealing(epoch, kl_start, annealtime, anneal_function) 139 | loss = rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss 140 | 141 | optimizer.zero_grad() 142 | loss.backward() 143 | optimizer.step() 144 | 145 | # torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm = 5) 146 | 147 | train_loss += loss.item() 148 | mse_loss += rec_loss.item() 149 | kullback_loss += kl_loss.item() 150 | kmeans_losses += kmeans_loss.item() 151 | 152 | # if idx % 1000 == 0: 153 | # print('Epoch: %d. loss: %.4f' %(epoch, loss.item())) 154 | 155 | scheduler.step(loss) #be sure scheduler is called before optimizer in >1.1 pytorch 156 | 157 | if future_decoder: 158 | print('Train loss: {:.3f}, MSE-Loss: {:.3f}, MSE-Future-Loss {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}, weight: {:.2f}'.format(train_loss / idx, 159 | mse_loss /idx, fut_loss/idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx, kl_weight)) 160 | else: 161 | print('Train loss: {:.3f}, MSE-Loss: {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}, weight: {:.2f}'.format(train_loss / idx, 162 | mse_loss /idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx, kl_weight)) 163 | 164 | return kl_weight, train_loss/idx, kl_weight*kmeans_losses/idx, kullback_loss/idx, mse_loss/idx, fut_loss/idx 165 | 166 | 167 | def test(test_loader, epoch, model, optimizer, BETA, kl_weight, seq_len, mse_red, kloss, klmbda, future_decoder, bsize): 168 | model.eval() # toggle model to inference mode 169 | test_loss = 0.0 170 | mse_loss = 0.0 171 | kullback_loss = 0.0 172 | kmeans_losses = 0.0 173 | loss = 0.0 174 | seq_len_half = int(seq_len / 2) 175 | 176 | with torch.no_grad(): 177 | for idx, data_item in enumerate(test_loader): 178 | # we're only going to infer, so no autograd at all required 179 | data_item = Variable(data_item) 180 | data_item = data_item.permute(0,2,1) 181 | if use_gpu: 182 | data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').cuda() 183 | else: 184 | data = data_item[:,:seq_len_half,:].type('torch.FloatTensor').to() 185 | 186 | if future_decoder: 187 | recon_images, _, latent, mu, logvar = model(data) 188 | rec_loss = reconstruction_loss(data, recon_images, mse_red) 189 | kl_loss = kullback_leibler_loss(mu, logvar) 190 | kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize) 191 | loss = rec_loss + BETA*kl_weight*kl_loss+ kl_weight*kmeans_loss 192 | 193 | else: 194 | recon_images, latent, mu, logvar = model(data) 195 | rec_loss = reconstruction_loss(data, recon_images, mse_red) 196 | kl_loss = kullback_leibler_loss(mu, logvar) 197 | kmeans_loss = cluster_loss(latent.T, kloss, klmbda, bsize) 198 | loss = rec_loss + BETA*kl_weight*kl_loss + kl_weight*kmeans_loss 199 | 200 | # torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm = 5) 201 | 202 | test_loss += loss.item() 203 | mse_loss += rec_loss.item() 204 | kullback_loss += kl_loss.item() 205 | kmeans_losses += kmeans_loss 206 | 207 | print('Test loss: {:.3f}, MSE-Loss: {:.3f}, KL-Loss: {:.3f}, Kmeans-Loss: {:.3f}'.format(test_loss / idx, 208 | mse_loss /idx, BETA*kl_weight*kullback_loss/idx, kl_weight*kmeans_losses/idx)) 209 | 210 | return mse_loss /idx, test_loss/idx, kl_weight*kmeans_losses 211 | 212 | 213 | def train_model(config): 214 | config_file = Path(config).resolve() 215 | cfg = read_config(config_file) 216 | legacy = cfg['legacy'] 217 | model_name = cfg['model_name'] 218 | pretrained_weights = cfg['pretrained_weights'] 219 | pretrained_model = cfg['pretrained_model'] 220 | fixed = cfg['egocentric_data'] 221 | 222 | print("Train Variational Autoencoder - model name: %s \n" %model_name) 223 | if not os.path.exists(os.path.join(cfg['project_path'],'model','best_model',"")): 224 | os.mkdir(os.path.join(cfg['project_path'],'model','best_model',"")) 225 | os.mkdir(os.path.join(cfg['project_path'],'model','best_model','snapshots',"")) 226 | os.mkdir(os.path.join(cfg['project_path'],'model','model_losses',"")) 227 | 228 | # make sure torch uses cuda for GPU computing 229 | use_gpu = torch.cuda.is_available() 230 | if use_gpu: 231 | print("Using CUDA") 232 | print('GPU active:',torch.cuda.is_available()) 233 | print('GPU used: ',torch.cuda.get_device_name(0)) 234 | else: 235 | torch.device("cpu") 236 | print("warning, a GPU was not found... proceeding with CPU (slow!) \n") 237 | #raise NotImplementedError('GPU Computing is required!') 238 | 239 | """ HYPERPARAMTERS """ 240 | # General 241 | CUDA = use_gpu 242 | SEED = 19 243 | TRAIN_BATCH_SIZE = cfg['batch_size'] 244 | TEST_BATCH_SIZE = int(cfg['batch_size']/4) 245 | EPOCHS = cfg['max_epochs'] 246 | ZDIMS = cfg['zdims'] 247 | BETA = cfg['beta'] 248 | SNAPSHOT = cfg['model_snapshot'] 249 | LEARNING_RATE = cfg['learning_rate'] 250 | NUM_FEATURES = cfg['num_features'] 251 | if fixed == False: 252 | NUM_FEATURES = NUM_FEATURES - 2 253 | TEMPORAL_WINDOW = cfg['time_window']*2 254 | FUTURE_DECODER = cfg['prediction_decoder'] 255 | FUTURE_STEPS = cfg['prediction_steps'] 256 | 257 | # RNN 258 | hidden_size_layer_1 = cfg['hidden_size_layer_1'] 259 | hidden_size_layer_2 = cfg['hidden_size_layer_2'] 260 | hidden_size_rec = cfg['hidden_size_rec'] 261 | hidden_size_pred = cfg['hidden_size_pred'] 262 | dropout_encoder = cfg['dropout_encoder'] 263 | dropout_rec = cfg['dropout_rec'] 264 | dropout_pred = cfg['dropout_pred'] 265 | noise = cfg['noise'] 266 | scheduler_step_size = cfg['scheduler_step_size'] 267 | softplus = cfg['softplus'] 268 | 269 | # Loss 270 | MSE_REC_REDUCTION = cfg['mse_reconstruction_reduction'] 271 | MSE_PRED_REDUCTION = cfg['mse_prediction_reduction'] 272 | KMEANS_LOSS = cfg['kmeans_loss'] 273 | KMEANS_LAMBDA = cfg['kmeans_lambda'] 274 | KL_START = cfg['kl_start'] 275 | ANNEALTIME = cfg['annealtime'] 276 | anneal_function = cfg['anneal_function'] 277 | optimizer_scheduler = cfg['scheduler'] 278 | 279 | BEST_LOSS = 999999 280 | convergence = 0 281 | print('Latent Dimensions: %d, Time window: %d, Batch Size: %d, Beta: %d, lr: %.4f\n' %(ZDIMS, cfg['time_window'], TRAIN_BATCH_SIZE, BETA, LEARNING_RATE)) 282 | 283 | # simple logging of diverse losses 284 | train_losses = [] 285 | test_losses = [] 286 | kmeans_losses = [] 287 | kl_losses = [] 288 | weight_values = [] 289 | mse_losses = [] 290 | fut_losses = [] 291 | 292 | torch.manual_seed(SEED) 293 | 294 | if legacy == False: 295 | RNN = RNN_VAE 296 | else: 297 | RNN = RNN_VAE_LEGACY 298 | if CUDA: 299 | torch.cuda.manual_seed(SEED) 300 | model = RNN(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 301 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 302 | dropout_rec, dropout_pred, softplus).cuda() 303 | else: #cpu support ... 304 | torch.cuda.manual_seed(SEED) 305 | model = RNN(TEMPORAL_WINDOW,ZDIMS,NUM_FEATURES,FUTURE_DECODER,FUTURE_STEPS, hidden_size_layer_1, 306 | hidden_size_layer_2, hidden_size_rec, hidden_size_pred, dropout_encoder, 307 | dropout_rec, dropout_pred, softplus).to() 308 | 309 | if pretrained_weights: 310 | try: 311 | print("Loading pretrained weights from model: %s\n" %os.path.join(cfg['project_path'],'model','best_model',pretrained_model+'_'+cfg['Project']+'.pkl')) 312 | model.load_state_dict(torch.load(os.path.join(cfg['project_path'],'model','best_model',pretrained_model+'_'+cfg['Project']+'.pkl'))) 313 | KL_START = 0 314 | ANNEALTIME = 1 315 | except: 316 | print("No file found at %s\n" %os.path.join(cfg['project_path'],'model','best_model',pretrained_model+'_'+cfg['Project']+'.pkl')) 317 | try: 318 | print("Loading pretrained weights from %s\n" %pretrained_model) 319 | model.load_state_dict(torch.load(pretrained_model)) 320 | KL_START = 0 321 | ANNEALTIME = 1 322 | except: 323 | print("Could not load pretrained model. Check file path in config.yaml.") 324 | 325 | """ DATASET """ 326 | trainset = SEQUENCE_DATASET(os.path.join(cfg['project_path'],"data", "train",""), data='train_seq.npy', train=True, temporal_window=TEMPORAL_WINDOW) 327 | testset = SEQUENCE_DATASET(os.path.join(cfg['project_path'],"data", "train",""), data='test_seq.npy', train=False, temporal_window=TEMPORAL_WINDOW) 328 | 329 | train_loader = Data.DataLoader(trainset, batch_size=TRAIN_BATCH_SIZE, shuffle=True, drop_last=True) 330 | test_loader = Data.DataLoader(testset, batch_size=TEST_BATCH_SIZE, shuffle=True, drop_last=True) 331 | 332 | optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE, amsgrad=True) 333 | 334 | if optimizer_scheduler: 335 | print('Scheduler step size: %d, Scheduler gamma: %.2f\n' %(scheduler_step_size, cfg['scheduler_gamma'])) 336 | # Thanks to @alexcwsmith for the optimized scheduler contribution 337 | scheduler = ReduceLROnPlateau(optimizer, 'min', factor=cfg['scheduler_gamma'], patience=cfg['scheduler_step_size'], threshold=1e-3, threshold_mode='rel', verbose=True) 338 | else: 339 | scheduler = StepLR(optimizer, step_size=scheduler_step_size, gamma=1, last_epoch=-1) 340 | 341 | print("Start training... ") 342 | for epoch in range(1,EPOCHS): 343 | print("Epoch: %d" %epoch) 344 | weight, train_loss, km_loss, kl_loss, mse_loss, fut_loss = train(train_loader, epoch, model, optimizer, 345 | anneal_function, BETA, KL_START, 346 | ANNEALTIME, TEMPORAL_WINDOW, FUTURE_DECODER, 347 | FUTURE_STEPS, scheduler, MSE_REC_REDUCTION, 348 | MSE_PRED_REDUCTION, KMEANS_LOSS, KMEANS_LAMBDA, 349 | TRAIN_BATCH_SIZE, noise) 350 | 351 | current_loss, test_loss, test_list = test(test_loader, epoch, model, optimizer, 352 | BETA, weight, TEMPORAL_WINDOW, MSE_REC_REDUCTION, 353 | KMEANS_LOSS, KMEANS_LAMBDA, FUTURE_DECODER, TEST_BATCH_SIZE) 354 | 355 | # logging losses 356 | train_losses.append(train_loss) 357 | test_losses.append(test_loss) 358 | kmeans_losses.append(km_loss) 359 | kl_losses.append(kl_loss) 360 | weight_values.append(weight) 361 | mse_losses.append(mse_loss) 362 | fut_losses.append(fut_loss) 363 | 364 | # save best model 365 | if weight > 0.99 and current_loss <= BEST_LOSS: 366 | BEST_LOSS = current_loss 367 | print("Saving model!") 368 | 369 | if use_gpu: 370 | torch.save(model.state_dict(), os.path.join(cfg['project_path'],"model", "best_model",model_name+'_'+cfg['Project']+'.pkl')) 371 | 372 | else: 373 | torch.save(model.state_dict(), os.path.join(cfg['project_path'],"model", "best_model",model_name+'_'+cfg['Project']+'.pkl')) 374 | 375 | convergence = 0 376 | else: 377 | convergence += 1 378 | 379 | if epoch % SNAPSHOT == 0: 380 | print("Saving model snapshot!\n") 381 | torch.save(model.state_dict(), os.path.join(cfg['project_path'],'model','best_model','snapshots',model_name+'_'+cfg['Project']+'_epoch_'+str(epoch)+'.pkl')) 382 | 383 | if convergence > cfg['model_convergence']: 384 | print('Finished training...') 385 | print('Model converged. Please check your model with vame.evaluate_model(). \n' 386 | 'You can also re-run vame.trainmodel() to further improve your model. \n' 387 | 'Make sure to set _pretrained_weights_ in your config.yaml to "true" \n' 388 | 'and plug your current model name into _pretrained_model_. \n' 389 | 'Hint: Set "model_convergence" in your config.yaml to a higher value. \n' 390 | '\n' 391 | 'Next: \n' 392 | 'Use vame.pose_segmentation() to identify behavioral motifs in your dataset!') 393 | #return 394 | break 395 | 396 | # save logged losses 397 | np.save(os.path.join(cfg['project_path'],'model','model_losses','train_losses_'+model_name), train_losses) 398 | np.save(os.path.join(cfg['project_path'],'model','model_losses','test_losses_'+model_name), test_losses) 399 | np.save(os.path.join(cfg['project_path'],'model','model_losses','kmeans_losses_'+model_name), kmeans_losses) 400 | np.save(os.path.join(cfg['project_path'],'model','model_losses','kl_losses_'+model_name), kl_losses) 401 | np.save(os.path.join(cfg['project_path'],'model','model_losses','weight_values_'+model_name), weight_values) 402 | np.save(os.path.join(cfg['project_path'],'model','model_losses','mse_train_losses_'+model_name), mse_losses) 403 | np.save(os.path.join(cfg['project_path'],'model','model_losses','mse_test_losses_'+model_name), current_loss) 404 | np.save(os.path.join(cfg['project_path'],'model','model_losses','fut_losses_'+model_name), fut_losses) 405 | 406 | print("\n") 407 | 408 | if convergence < cfg['model_convergence']: 409 | print('Finished training...') 410 | print('Model seems to have not reached convergence. You may want to check your model \n' 411 | 'with vame.evaluate_model(). If your satisfied you can continue. \n' 412 | 'Use vame.pose_segmentation() to identify behavioral motifs! \n' 413 | 'OPTIONAL: You can re-run vame.train_model() to improve performance.') 414 | -------------------------------------------------------------------------------- /vame/util/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 0.1 Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | import sys 12 | sys.dont_write_bytecode = True 13 | 14 | from vame.util.auxiliary import * 15 | -------------------------------------------------------------------------------- /vame/util/align_egocentrical.py: -------------------------------------------------------------------------------- 1 | """ 2 | Variational Animal Motion Embedding 0.1 Toolbox 3 | © K. Luxem & J. Kürsch & P. Bauer, Department of Cellular Neuroscience 4 | Leibniz Institute for Neurobiology, Magdeburg, Germany 5 | 6 | https://github.com/LINCellularNeuroscience/VAME 7 | Licensed under GNU General Public License v3.0 8 | """ 9 | 10 | import os 11 | import cv2 as cv 12 | import numpy as np 13 | import pandas as pd 14 | import tqdm 15 | import glob 16 | 17 | from pathlib import Path 18 | from vame.util.auxiliary import read_config 19 | 20 | #Returns cropped image using rect tuple 21 | def crop_and_flip(rect, src, points, ref_index): 22 | #Read out rect structures and convert 23 | center, size, theta = rect 24 | center, size = tuple(map(int, center)), tuple(map(int, size)) 25 | #Get rotation matrix 26 | M = cv.getRotationMatrix2D(center, theta, 1) 27 | 28 | #shift DLC points 29 | x_diff = center[0] - size[0]//2 30 | y_diff = center[1] - size[1]//2 31 | 32 | dlc_points_shifted = [] 33 | 34 | for i in points: 35 | point=cv.transform(np.array([[[i[0], i[1]]]]),M)[0][0] 36 | 37 | point[0] -= x_diff 38 | point[1] -= y_diff 39 | 40 | dlc_points_shifted.append(point) 41 | 42 | # Perform rotation on src image 43 | dst = cv.warpAffine(src.astype('float32'), M, src.shape[:2]) 44 | out = cv.getRectSubPix(dst, size, center) 45 | 46 | #check if flipped correctly, otherwise flip again 47 | if dlc_points_shifted[ref_index[1]][0] >= dlc_points_shifted[ref_index[0]][0]: 48 | rect = ((size[0]//2,size[0]//2),size,180) 49 | center, size, theta = rect 50 | center, size = tuple(map(int, center)), tuple(map(int, size)) 51 | #Get rotation matrix 52 | M = cv.getRotationMatrix2D(center, theta, 1) 53 | 54 | 55 | #shift DLC points 56 | x_diff = center[0] - size[0]//2 57 | y_diff = center[1] - size[1]//2 58 | 59 | points = dlc_points_shifted 60 | dlc_points_shifted = [] 61 | 62 | for i in points: 63 | point=cv.transform(np.array([[[i[0], i[1]]]]),M)[0][0] 64 | 65 | point[0] -= x_diff 66 | point[1] -= y_diff 67 | 68 | dlc_points_shifted.append(point) 69 | 70 | # Perform rotation on src image 71 | dst = cv.warpAffine(out.astype('float32'), M, out.shape[:2]) 72 | out = cv.getRectSubPix(dst, size, center) 73 | 74 | return out, dlc_points_shifted 75 | 76 | 77 | #Helper function to return indexes of nans 78 | def nan_helper(y): 79 | return np.isnan(y), lambda z: z.nonzero()[0] 80 | 81 | 82 | #Interpolates all nan values of given array 83 | def interpol(arr): 84 | 85 | y = np.transpose(arr) 86 | 87 | nans, x = nan_helper(y[0]) 88 | y[0][nans]= np.interp(x(nans), x(~nans), y[0][~nans]) 89 | nans, x = nan_helper(y[1]) 90 | y[1][nans]= np.interp(x(nans), x(~nans), y[1][~nans]) 91 | 92 | arr = np.transpose(y) 93 | 94 | return arr 95 | 96 | def background(path_to_file,filename,video_format='.mp4',num_frames=1000): 97 | """ 98 | Compute background image from fixed camera 99 | """ 100 | import scipy.ndimage 101 | capture = cv.VideoCapture(os.path.join(path_to_file,'videos',filename+video_format)) 102 | 103 | if not capture.isOpened(): 104 | raise Exception("Unable to open video file: {0}".format(os.path.join(path_to_file,'videos',filename+video_format))) 105 | 106 | frame_count = int(capture.get(cv.CAP_PROP_FRAME_COUNT)) 107 | ret, frame = capture.read() 108 | 109 | height, width, _ = frame.shape 110 | frames = np.zeros((height,width,num_frames)) 111 | 112 | for i in tqdm.tqdm(range(num_frames), disable=not True, desc='Compute background image for video %s' %filename): 113 | rand = np.random.choice(frame_count, replace=False) 114 | capture.set(1,rand) 115 | ret, frame = capture.read() 116 | gray = cv.cvtColor(frame, cv.COLOR_BGR2GRAY) 117 | frames[...,i] = gray 118 | 119 | print('Finishing up!') 120 | medFrame = np.median(frames,2) 121 | background = scipy.ndimage.median_filter(medFrame, (5,5)) 122 | 123 | # np.save(path_to_file+'videos/'+'background/'+filename+'-background.npy',background) 124 | 125 | capture.release() 126 | return background 127 | 128 | 129 | def align_mouse(path_to_file,filename,video_format,crop_size, pose_list, 130 | pose_ref_index, confidence, pose_flip_ref,bg,frame_count,use_video=True): 131 | 132 | #returns: list of cropped images (if video is used) and list of cropped DLC points 133 | # 134 | #parameters: 135 | #path_to_file: directory 136 | #filename: name of video file without format 137 | #video_format: format of video file 138 | #crop_size: tuple of x and y crop size 139 | #dlc_list: list of arrays containg corresponding x and y DLC values 140 | #dlc_ref_index: indices of 2 lists in dlc_list to align mouse along 141 | #dlc_flip_ref: indices of 2 lists in dlc_list to flip mouse if flip was false 142 | #bg: background image to subtract 143 | #frame_count: number of frames to align 144 | #use_video: boolean if video should be cropped or DLC points only 145 | 146 | images = [] 147 | points = [] 148 | 149 | for i in pose_list: 150 | for j in i: 151 | if j[2] <= confidence: 152 | j[0],j[1] = np.nan, np.nan 153 | 154 | 155 | for i in pose_list: 156 | i = interpol(i) 157 | 158 | if use_video: 159 | capture = cv.VideoCapture(os.path.join(path_to_file,'videos',filename+video_format)) 160 | 161 | if not capture.isOpened(): 162 | raise Exception("Unable to open video file: {0}".format(os.path.join(path_to_file,'videos',filename+video_format))) 163 | 164 | for idx in tqdm.tqdm(range(frame_count), disable=not True, desc='Align frames'): 165 | 166 | if use_video: 167 | #Read frame 168 | try: 169 | ret, frame = capture.read() 170 | frame = cv.cvtColor(frame, cv.COLOR_BGR2GRAY) 171 | frame = frame - bg 172 | frame[frame <= 0] = 0 173 | except: 174 | print("Couldn't find a frame in capture.read(). #Frame: %d" %idx) 175 | continue 176 | else: 177 | frame=np.zeros((1,1)) 178 | 179 | #Read coordinates and add border 180 | pose_list_bordered = [] 181 | 182 | for i in pose_list: 183 | pose_list_bordered.append((int(i[idx][0]+crop_size[0]),int(i[idx][1]+crop_size[1]))) 184 | 185 | img = cv.copyMakeBorder(frame, crop_size[1], crop_size[1], crop_size[0], crop_size[0], cv.BORDER_CONSTANT, 0) 186 | 187 | punkte = [] 188 | for i in pose_ref_index: 189 | coord = [] 190 | coord.append(pose_list_bordered[i][0]) 191 | coord.append(pose_list_bordered[i][1]) 192 | punkte.append(coord) 193 | punkte = [punkte] 194 | punkte = np.asarray(punkte) 195 | 196 | #calculate minimal rectangle around snout and tail 197 | rect = cv.minAreaRect(punkte) 198 | 199 | #change size in rect tuple structure to be equal to crop_size 200 | lst = list(rect) 201 | lst[1] = crop_size 202 | rect = tuple(lst) 203 | 204 | center, size, theta = rect 205 | 206 | #crop image 207 | out, shifted_points = crop_and_flip(rect, img,pose_list_bordered,pose_flip_ref) 208 | 209 | if use_video: #for memory optimization, just save images when video is used. 210 | images.append(out) 211 | points.append(shifted_points) 212 | 213 | if use_video: 214 | capture.release() 215 | 216 | time_series = np.zeros((len(pose_list)*2,frame_count)) 217 | for i in range(frame_count): 218 | idx = 0 219 | for j in range(len(pose_list)): 220 | time_series[idx:idx+2,i] = points[i][j] 221 | idx += 2 222 | 223 | return images, points, time_series 224 | 225 | 226 | #play aligned video 227 | def play_aligned_video(a, n, frame_count): 228 | colors = [(255,0,0),(0,255,0),(0,0,255),(255,255,0),(255,0,255),(0,255,255),(0,0,0),(255,255,255)] 229 | 230 | for i in range(frame_count): 231 | # Capture frame-by-frame 232 | ret, frame = True,a[i] 233 | if ret == True: 234 | 235 | # Display the resulting frame 236 | frame = cv.cvtColor(frame.astype('uint8')*255, cv.COLOR_GRAY2BGR) 237 | im_color = cv.applyColorMap(frame, cv.COLORMAP_JET) 238 | 239 | for c,j in enumerate(n[i]): 240 | cv.circle(im_color,(j[0], j[1]), 5, colors[c], -1) 241 | 242 | cv.imshow('Frame',im_color) 243 | 244 | # Press Q on keyboard to exit 245 | if cv.waitKey(25) & 0xFF == ord('q'): 246 | break 247 | 248 | # Break the loop 249 | else: 250 | break 251 | cv.destroyAllWindows() 252 | 253 | 254 | def alignment(path_to_file, filename, pose_ref_index, video_format, crop_size, confidence, use_video=False, check_video=False): 255 | 256 | #read out data 257 | dataFile = glob.glob(os.path.join(path_to_file,'videos','pose_estimation',filename+'*')) 258 | if len(dataFile)>1: 259 | raise AssertionError("Multiple data files match video {}".format(filename)) 260 | else: 261 | dataFile=dataFile[0] 262 | if dataFile.endswith('.csv'): 263 | data = pd.read_csv(dataFile, skiprows = 2, index_col=0) 264 | elif dataFile.endswith('.h5'): 265 | data = pd.read_hdf(dataFile) 266 | data_mat = pd.DataFrame.to_numpy(data) 267 | # data_mat = data_mat[:,1:] 268 | 269 | # get the coordinates for alignment from data table 270 | pose_list = [] 271 | 272 | for i in range(int(data_mat.shape[1]/3)): 273 | pose_list.append(data_mat[:,i*3:(i+1)*3]) 274 | 275 | #list of reference coordinate indices for alignment 276 | #0: snout, 1: forehand_left, 2: forehand_right, 277 | #3: hindleft, 4: hindright, 5: tail 278 | 279 | pose_ref_index = pose_ref_index 280 | 281 | #list of 2 reference coordinate indices for avoiding flipping 282 | pose_flip_ref = pose_ref_index 283 | 284 | if use_video: 285 | #compute background 286 | bg = background(path_to_file,filename,video_format) 287 | capture = cv.VideoCapture(os.path.join(path_to_file,'videos',filename+video_format)) 288 | if not capture.isOpened(): 289 | raise Exception("Unable to open video file: {0}".format(os.path.join(path_to_file,'videos',filename+video_format))) 290 | 291 | frame_count = int(capture.get(cv.CAP_PROP_FRAME_COUNT)) 292 | capture.release() 293 | else: 294 | bg = 0 295 | frame_count = len(data) # Change this to an abitrary number if you first want to test the code 296 | 297 | 298 | frames, n, time_series = align_mouse(path_to_file, filename, video_format, crop_size, pose_list, pose_ref_index, 299 | confidence, pose_flip_ref, bg, frame_count, use_video) 300 | 301 | if check_video: 302 | play_aligned_video(frames, n, frame_count) 303 | 304 | return time_series, frames 305 | 306 | 307 | def egocentric_alignment(config, pose_ref_index=[0,5], crop_size=(300,300), use_video=False, video_format='.mp4', check_video=False): 308 | """ Happy aligning """ 309 | #config parameters 310 | config_file = Path(config).resolve() 311 | cfg = read_config(config_file) 312 | 313 | path_to_file = cfg['project_path'] 314 | filename = cfg['video_sets'] 315 | confidence = cfg['pose_confidence'] 316 | video_format=video_format 317 | crop_size=crop_size 318 | 319 | if cfg['egocentric_data'] == True: 320 | raise ValueError("The config.yaml indicates that the data is not egocentric. Please check the parameter egocentric_data") 321 | 322 | # call function and save into your VAME data folder 323 | for file in filename: 324 | print("Aligning data %s, Pose confidence value: %.2f" %(file, confidence)) 325 | egocentric_time_series, frames = alignment(path_to_file, file, pose_ref_index, video_format, crop_size, 326 | confidence, use_video=use_video, check_video=check_video) 327 | np.save(os.path.join(path_to_file,'data',file,file+'-PE-seq.npy'), egocentric_time_series) 328 | # np.save(os.path.join(path_to_file,'data/',file,"",file+'-PE-seq.npy', egocentric_time_series)) 329 | 330 | print("Your data is now ine right format and you can call vame.create_trainset()") 331 | 332 | -------------------------------------------------------------------------------- /vame/util/auxiliary.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | 11 | The following code is adapted from: 12 | 13 | DeepLabCut2.0 Toolbox (deeplabcut.org) 14 | © A. & M. Mathis Labs 15 | https://github.com/AlexEMG/DeepLabCut 16 | Please see AUTHORS for contributors. 17 | https://github.com/AlexEMG/DeepLabCut/blob/master/AUTHORS 18 | Licensed under GNU Lesser General Public License v3.0 19 | """ 20 | 21 | import os, yaml 22 | from pathlib import Path 23 | import ruamel.yaml 24 | 25 | 26 | def create_config_template(): 27 | """ 28 | Creates a template for config.yaml file. This specific order is preserved while saving as yaml file. 29 | """ 30 | import ruamel.yaml 31 | yaml_str = """\ 32 | # Project configurations 33 | Project: 34 | model_name: 35 | n_cluster: 36 | pose_confidence: 37 | \n 38 | # Project path and videos 39 | project_path: 40 | video_sets: 41 | \n 42 | # Data 43 | all_data: 44 | \n 45 | # Creation of train set: 46 | egocentric_data: 47 | robust: 48 | iqr_factor: 49 | axis: 50 | savgol_filter: 51 | savgol_length: 52 | savgol_order: 53 | test_fraction: 54 | \n 55 | # RNN model general hyperparameter: 56 | pretrained_model: 57 | pretrained_weights: 58 | num_features: 59 | batch_size: 60 | max_epochs: 61 | model_snapshot: 62 | model_convergence: 63 | transition_function: 64 | beta: 65 | beta_norm: 66 | zdims: 67 | learning_rate: 68 | time_window: 69 | prediction_decoder: 70 | prediction_steps: 71 | noise: 72 | scheduler: 73 | scheduler_step_size: 74 | scheduler_gamma: 75 | #Note the optimal scheduler threshold below can vary greatly (from .1-.0001) between experiments. 76 | #You are encouraged to read the torch.optim.ReduceLROnPlateau docs to understand the threshold to use. 77 | scheduler_threshold: 78 | softplus: 79 | \n 80 | # Segmentation: 81 | parameterization: 82 | hmm_trained: False 83 | load_data: 84 | individual_parameterization: 85 | random_state_kmeans: 86 | n_init_kmeans: 87 | \n 88 | # Video writer: 89 | length_of_motif_video: 90 | \n 91 | # UMAP parameter: 92 | min_dist: 93 | n_neighbors: 94 | random_state: 95 | num_points: 96 | \n 97 | # ONLY CHANGE ANYTHING BELOW IF YOU ARE FAMILIAR WITH RNN MODELS 98 | # RNN encoder hyperparamter: 99 | hidden_size_layer_1: 100 | hidden_size_layer_2: 101 | dropout_encoder: 102 | \n 103 | # RNN reconstruction hyperparameter: 104 | hidden_size_rec: 105 | dropout_rec: 106 | n_layers: 107 | \n 108 | # RNN prediction hyperparamter: 109 | hidden_size_pred: 110 | dropout_pred: 111 | \n 112 | # RNN loss hyperparameter: 113 | mse_reconstruction_reduction: 114 | mse_prediction_reduction: 115 | kmeans_loss: 116 | kmeans_lambda: 117 | anneal_function: 118 | kl_start: 119 | annealtime: 120 | \n 121 | # Legacy mode 122 | legacy: 123 | """ 124 | ruamelFile = ruamel.yaml.YAML() 125 | cfg_file = ruamelFile.load(yaml_str) 126 | return(cfg_file,ruamelFile) 127 | 128 | 129 | def read_config(configname): 130 | """ 131 | Reads structured config file defining a project. 132 | """ 133 | ruamelFile = ruamel.yaml.YAML() 134 | path = Path(configname) 135 | if os.path.exists(path): 136 | try: 137 | with open(path, "r") as f: 138 | cfg = ruamelFile.load(f) 139 | curr_dir = os.path.dirname(configname) 140 | if cfg["project_path"] != curr_dir: 141 | cfg["project_path"] = curr_dir 142 | write_config(configname, cfg) 143 | except Exception as err: 144 | if len(err.args) > 2: 145 | if ( 146 | err.args[2] 147 | == "could not determine a constructor for the tag '!!python/tuple'" 148 | ): 149 | with open(path, "r") as ymlfile: 150 | cfg = yaml.load(ymlfile, Loader=yaml.SafeLoader) 151 | write_config(configname, cfg) 152 | else: 153 | raise 154 | 155 | else: 156 | raise FileNotFoundError( 157 | "Config file is not found. Please make sure that the file exists and/or that you passed the path of the config file correctly!" 158 | ) 159 | return cfg 160 | 161 | def write_config(configname,cfg): 162 | """ 163 | Write structured config file. 164 | """ 165 | with open(configname, 'w') as cf: 166 | ruamelFile = ruamel.yaml.YAML() 167 | cfg_file,ruamelFile = create_config_template() 168 | for key in cfg.keys(): 169 | cfg_file[key]=cfg[key] 170 | 171 | ruamelFile.dump(cfg_file, cf) 172 | 173 | def update_config(config): 174 | config_file = Path(config).resolve() 175 | cfg = read_config(config_file) 176 | 177 | project = cfg['Project'] 178 | project_path = cfg['project_path'] 179 | video_names = [] 180 | for file in cfg['video_sets']: 181 | video_names.append(file) 182 | 183 | flag = input("ATTENTION! You are about to overwrite your current config.yaml. If you did changes, " 184 | "back up your current version and compare to the updated version. Do you want to continue? (yes/no)") 185 | 186 | if flag == 'yes': 187 | cfg_file,ruamelFile = create_config_template() 188 | 189 | cfg_file['Project']=str(project) 190 | cfg_file['project_path']=str(project_path)+'/' 191 | cfg_file['test_fraction']=.1 192 | cfg_file['video_sets']=video_names 193 | cfg_file['all_data']='yes' 194 | cfg_file['load_data']='-PE-seq-clean' 195 | cfg_file['anneal_function']='linear' 196 | cfg_file['batch_size']=256 197 | cfg_file['max_epochs']=500 198 | cfg_file['transition_function']='GRU' 199 | cfg_file['beta']=1 200 | cfg_file['zdims']=30 201 | cfg_file['learning_rate']=5e-4 202 | cfg_file['time_window']=30 203 | cfg_file['prediction_decoder']=1 204 | cfg_file['prediction_steps']=15 205 | cfg_file['model_convergence']=50 206 | cfg_file['model_snapshot']=50 207 | cfg_file['num_features']=12 208 | cfg_file['savgol_filter']=True 209 | cfg_file['savgol_length']=5 210 | cfg_file['savgol_order']=2 211 | cfg_file['hidden_size_layer_1']=256 212 | cfg_file['hidden_size_layer_2']=256 213 | cfg_file['dropout_encoder']=0 214 | cfg_file['hidden_size_rec']=256 215 | cfg_file['dropout_rec']=0 216 | cfg_file['hidden_size_pred']=256 217 | cfg_file['dropout_pred']=0 218 | cfg_file['kl_start']=2 219 | cfg_file['annealtime']=4 220 | cfg_file['mse_reconstruction_reduction']='sum' 221 | cfg_file['mse_prediction_reduction']='sum' 222 | cfg_file['kmeans_loss']=cfg_file['zdims'] 223 | cfg_file['kmeans_lambda']=0.1 224 | cfg_file['scheduler']=1 225 | cfg_file['length_of_motif_video'] = 1000 226 | cfg_file['noise'] = False 227 | cfg_file['scheduler_step_size'] = 100 228 | cfg_file['legacy'] = False 229 | cfg_file['individual_parameterization'] = False 230 | cfg_file['random_state_kmeans'] = 42 231 | cfg_file['n_init_kmeans'] = 15 232 | cfg_file['model_name']='VAME' 233 | cfg_file['n_cluster'] = 15 234 | cfg_file['pretrained_weights'] = False 235 | cfg_file['pretrained_model']='None' 236 | cfg_file['min_dist'] = 0.1 237 | cfg_file['n_neighbors'] = 200 238 | cfg_file['random_state'] = 42 239 | cfg_file['num_points'] = 30000 240 | cfg_file['scheduler_gamma'] = 0.2 241 | cfg_file['scheduler_threshold'] = .1 242 | cfg_file['softplus'] = False 243 | cfg_file['pose_confidence'] = 0.99 244 | cfg_file['iqr_factor'] = 4 245 | cfg_file['robust'] = True 246 | cfg_file['beta_norm'] = False 247 | cfg_file['n_layers'] = 1 248 | cfg_file['axis'] = 'None' 249 | cfg_file['egocentric_data'] = True 250 | cfg_file['parameterization'] = 'kmeans' 251 | 252 | projconfigfile=os.path.join(str(project_path),'config.yaml') 253 | # Write dictionary to yaml config file 254 | write_config(projconfigfile,cfg_file) 255 | 256 | print("Your config.yaml has been updated.") 257 | else: 258 | print("No changes have been applied.") 259 | -------------------------------------------------------------------------------- /vame/util/csv_to_npy.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import numpy as np 14 | import pandas as pd 15 | 16 | from pathlib import Path 17 | from vame.util.auxiliary import read_config 18 | 19 | 20 | #Helper function to return indexes of nans 21 | def nan_helper(y): 22 | return np.isnan(y), lambda z: z.nonzero()[0] 23 | 24 | #Interpolates all nan values of given array 25 | def interpol(arr): 26 | 27 | y = np.transpose(arr) 28 | 29 | nans, x = nan_helper(y[0]) 30 | y[0][nans]= np.interp(x(nans), x(~nans), y[0][~nans]) 31 | nans, x = nan_helper(y[1]) 32 | y[1][nans]= np.interp(x(nans), x(~nans), y[1][~nans]) 33 | 34 | arr = np.transpose(y) 35 | 36 | return arr 37 | 38 | def csv_to_numpy(config): 39 | """ 40 | This is a function to convert your pose-estimation.csv file to a numpy array. 41 | 42 | Note that this code is only useful for data which is a priori egocentric, i.e. head-fixed 43 | or otherwise restrained animals. 44 | 45 | example use: 46 | vame.csv_to_npy('pathto/your/config/yaml', 'path/toYourFolderwithCSV/') 47 | """ 48 | config_file = Path(config).resolve() 49 | cfg = read_config(config_file) 50 | 51 | path_to_file = cfg['project_path'] 52 | filename = cfg['video_sets'] 53 | confidence = cfg['pose_confidence'] 54 | if cfg['egocentric_data'] == False: 55 | raise ValueError("The config.yaml indicates that the data is not egocentric. Please check the parameter egocentric_data") 56 | 57 | for file in filename: 58 | print(file) 59 | # Read in your .csv file, skip the first two rows and create a numpy array 60 | data = pd.read_csv(os.path.join(path_to_file,"videos","pose_estimation",file+'.csv'), skiprows = 3, header=None) 61 | data_mat = pd.DataFrame.to_numpy(data) 62 | data_mat = data_mat[:,1:] 63 | 64 | pose_list = [] 65 | 66 | # get the number of bodyparts, their x,y-position and the confidence from DeepLabCut 67 | for i in range(int(data_mat.shape[1]/3)): 68 | pose_list.append(data_mat[:,i*3:(i+1)*3]) 69 | 70 | # find low confidence and set them to NaN 71 | for i in pose_list: 72 | for j in i: 73 | if j[2] <= confidence: 74 | j[0],j[1] = np.nan, np.nan 75 | 76 | # interpolate NaNs 77 | for i in pose_list: 78 | i = interpol(i) 79 | 80 | positions = np.concatenate(pose_list, axis=1) 81 | final_positions = np.zeros((data_mat.shape[0], int(data_mat.shape[1]/3)*2)) 82 | 83 | jdx = 0 84 | idx = 0 85 | for i in range(int(data_mat.shape[1]/3)): 86 | final_positions[:,idx:idx+2] = positions[:,jdx:jdx+2] 87 | jdx += 3 88 | idx += 2 89 | 90 | # save the final_positions array with np.save() 91 | np.save(os.path.join(path_to_file,'data',file,file+"-PE-seq.npy"), final_positions.T) 92 | print("conversion from DeepLabCut csv to numpy complete...") 93 | 94 | print("Your data is now in right format and you can call vame.create_trainset()") 95 | -------------------------------------------------------------------------------- /vame/util/gif_pose_helper.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # -*- coding: utf-8 -*- 3 | """ 4 | Variational Animal Motion Embedding 1.0-alpha Toolbox 5 | © K. Luxem & P. Bauer, Department of Cellular Neuroscience 6 | Leibniz Institute for Neurobiology, Magdeburg, Germany 7 | 8 | https://github.com/LINCellularNeuroscience/VAME 9 | Licensed under GNU General Public License v3.0 10 | """ 11 | 12 | import os 13 | import h5py 14 | import tqdm 15 | import scipy 16 | import cv2 as cv 17 | import numpy as np 18 | import pandas as pd 19 | import matplotlib.pyplot as plt 20 | 21 | 22 | #Returns cropped image using rect tuple 23 | def crop_and_flip(rect, src, points, ref_index): 24 | #Read out rect structures and convert 25 | center, size, theta = rect 26 | center, size = tuple(map(int, center)), tuple(map(int, size)) 27 | #Get rotation matrix 28 | M = cv.getRotationMatrix2D(center, theta, 1) 29 | 30 | #shift DLC points 31 | x_diff = center[0] - size[0]//2 32 | y_diff = center[1] - size[1]//2 33 | 34 | dlc_points_shifted = [] 35 | 36 | for i in points: 37 | point=cv.transform(np.array([[[i[0], i[1]]]]),M)[0][0] 38 | 39 | point[0] -= x_diff 40 | point[1] -= y_diff 41 | 42 | dlc_points_shifted.append(point) 43 | 44 | # Perform rotation on src image 45 | dst = cv.warpAffine(src.astype('float32'), M, src.shape[:2]) 46 | out = cv.getRectSubPix(dst, size, center) 47 | 48 | #check if flipped correctly, otherwise flip again 49 | if dlc_points_shifted[ref_index[1]][0] >= dlc_points_shifted[ref_index[0]][0]: 50 | rect = ((size[0]//2,size[0]//2),size,180) 51 | center, size, theta = rect 52 | center, size = tuple(map(int, center)), tuple(map(int, size)) 53 | #Get rotation matrix 54 | M = cv.getRotationMatrix2D(center, theta, 1) 55 | 56 | 57 | #shift DLC points 58 | x_diff = center[0] - size[0]//2 59 | y_diff = center[1] - size[1]//2 60 | 61 | points = dlc_points_shifted 62 | dlc_points_shifted = [] 63 | 64 | for i in points: 65 | point=cv.transform(np.array([[[i[0], i[1]]]]),M)[0][0] 66 | 67 | point[0] -= x_diff 68 | point[1] -= y_diff 69 | 70 | dlc_points_shifted.append(point) 71 | 72 | # Perform rotation on src image 73 | dst = cv.warpAffine(out.astype('float32'), M, out.shape[:2]) 74 | out = cv.getRectSubPix(dst, size, center) 75 | 76 | return out, dlc_points_shifted 77 | 78 | 79 | def background(path_to_file,filename,file_format='.mp4',num_frames=1000): 80 | """ 81 | Compute background image from fixed camera 82 | """ 83 | 84 | capture = cv.VideoCapture(os.path.join(path_to_file,"videos",filename+file_format)) 85 | 86 | if not capture.isOpened(): 87 | raise Exception("Unable to open video file: {0}".format(os.path.join(path_to_file,"videos",filename+file_format))) 88 | 89 | frame_count = int(capture.get(cv.CAP_PROP_FRAME_COUNT)) 90 | ret, frame = capture.read() 91 | 92 | height, width, _ = frame.shape 93 | frames = np.zeros((height,width,num_frames)) 94 | 95 | for i in tqdm.tqdm(range(num_frames), disable=not True, desc='Compute background image for video %s' %filename): 96 | rand = np.random.choice(frame_count, replace=False) 97 | capture.set(1,rand) 98 | ret, frame = capture.read() 99 | gray = cv.cvtColor(frame, cv.COLOR_BGR2GRAY) 100 | frames[...,i] = gray 101 | 102 | print('Finishing up!') 103 | medFrame = np.median(frames,2) 104 | background = scipy.ndimage.median_filter(medFrame, (5,5)) 105 | 106 | np.save(os.path.join(path_to_file,"videos",filename+'-background.npy'),background) 107 | 108 | capture.release() 109 | return background 110 | 111 | 112 | def get_rotation_matrix(adjacent, opposite, crop_size=(300, 300)): 113 | 114 | tan_alpha = np.abs(opposite) / np.abs(adjacent) 115 | alpha = np.arctan(tan_alpha) 116 | alpha = np.rad2deg(alpha) 117 | 118 | if adjacent < 0 and opposite > 0: 119 | alpha = 180-alpha 120 | 121 | if adjacent < 0 and opposite < 0: 122 | alpha = -(180-alpha) 123 | 124 | if adjacent > 0 and opposite < 0: 125 | alpha = -alpha 126 | 127 | rot_mat = cv.getRotationMatrix2D((crop_size[0] // 2, crop_size[1] // 2),alpha, 1) 128 | 129 | return rot_mat 130 | 131 | 132 | #Helper function to return indexes of nans 133 | def nan_helper(y): 134 | return np.isnan(y), lambda z: z.nonzero()[0] 135 | 136 | 137 | #Interpolates all nan values of given array 138 | def interpol(arr): 139 | 140 | y = np.transpose(arr) 141 | 142 | nans, x = nan_helper(y[0]) 143 | y[0][nans]= np.interp(x(nans), x(~nans), y[0][~nans]) 144 | nans, x = nan_helper(y[1]) 145 | y[1][nans]= np.interp(x(nans), x(~nans), y[1][~nans]) 146 | 147 | arr = np.transpose(y) 148 | 149 | return arr 150 | 151 | 152 | def get_animal_frames(cfg, filename, pose_ref_index, start, length, subtract_background, file_format='.mp4', crop_size=(300, 300)): 153 | path_to_file = cfg['project_path'] 154 | time_window = cfg['time_window'] 155 | lag = int(time_window / 2) 156 | #read out data 157 | data = pd.read_csv(os.path.join(path_to_file,"videos","pose_estimation",filename+'.csv'), skiprows = 2) 158 | data_mat = pd.DataFrame.to_numpy(data) 159 | data_mat = data_mat[:,1:] 160 | 161 | # get the coordinates for alignment from data table 162 | pose_list = [] 163 | 164 | for i in range(int(data_mat.shape[1]/3)): 165 | pose_list.append(data_mat[:,i*3:(i+1)*3]) 166 | 167 | #list of reference coordinate indices for alignment 168 | #0: snout, 1: forehand_left, 2: forehand_right, 169 | #3: hindleft, 4: hindright, 5: tail 170 | 171 | pose_ref_index = pose_ref_index 172 | 173 | #list of 2 reference coordinate indices for avoiding flipping 174 | pose_flip_ref = pose_ref_index 175 | 176 | # compute background 177 | if subtract_background == True: 178 | try: 179 | print("Loading background image ...") 180 | bg = np.load(os.path.join(path_to_file,"videos",filename+'-background.npy')) 181 | except: 182 | print("Can't find background image... Calculate background image...") 183 | bg = background(path_to_file,filename, file_format) 184 | 185 | images = [] 186 | points = [] 187 | 188 | for i in pose_list: 189 | for j in i: 190 | if j[2] <= 0.8: 191 | j[0],j[1] = np.nan, np.nan 192 | 193 | 194 | for i in pose_list: 195 | i = interpol(i) 196 | 197 | capture = cv.VideoCapture(os.path.join(path_to_file,"videos",filename+file_format)) 198 | if not capture.isOpened(): 199 | raise Exception("Unable to open video file: {0}".format(os.path.join(path_to_file,"videos",filename++file_format))) 200 | 201 | for idx in tqdm.tqdm(range(length), disable=not True, desc='Align frames'): 202 | try: 203 | capture.set(1,idx+start+lag) 204 | ret, frame = capture.read() 205 | frame = cv.cvtColor(frame, cv.COLOR_BGR2GRAY) 206 | if subtract_background == True: 207 | frame = frame - bg 208 | frame[frame <= 0] = 0 209 | except: 210 | print("Couldn't find a frame in capture.read(). #Frame: %d" %idx+start+lag) 211 | continue 212 | 213 | #Read coordinates and add border 214 | pose_list_bordered = [] 215 | 216 | for i in pose_list: 217 | pose_list_bordered.append((int(i[idx+start+lag][0]+crop_size[0]),int(i[idx+start+lag][1]+crop_size[1]))) 218 | 219 | img = cv.copyMakeBorder(frame, crop_size[1], crop_size[1], crop_size[0], crop_size[0], cv.BORDER_CONSTANT, 0) 220 | 221 | punkte = [] 222 | for i in pose_ref_index: 223 | coord = [] 224 | coord.append(pose_list_bordered[i][0]) 225 | coord.append(pose_list_bordered[i][1]) 226 | punkte.append(coord) 227 | punkte = [punkte] 228 | punkte = np.asarray(punkte) 229 | 230 | #calculate minimal rectangle around snout and tail 231 | rect = cv.minAreaRect(punkte) 232 | 233 | #change size in rect tuple structure to be equal to crop_size 234 | lst = list(rect) 235 | lst[1] = crop_size 236 | rect = tuple(lst) 237 | 238 | center, size, theta = rect 239 | 240 | #crop image 241 | out, shifted_points = crop_and_flip(rect, img,pose_list_bordered,pose_flip_ref) 242 | 243 | images.append(out) 244 | points.append(shifted_points) 245 | 246 | capture.release() 247 | return images 248 | --------------------------------------------------------------------------------