├── .github └── workflows │ └── CI.yml ├── .gitignore ├── Dockerfile ├── LICENSE ├── README.md ├── data └── config.json ├── pipeline.sh ├── requirements.txt ├── scripts └── generate_time_series.py ├── setup.py ├── tests └── unit │ ├── test_data_utils.py │ └── test_model.py └── time_series_forecasting ├── __init__.py ├── data_utils.py ├── evaluation.py ├── model.py ├── plot_images.py └── training.py /.github/workflows/CI.yml: -------------------------------------------------------------------------------- 1 | # This is a basic workflow to help you get started with Actions 2 | 3 | name: CI 4 | 5 | # Controls when the action will run. 6 | on: 7 | # Triggers the workflow on push or pull request events but only for the main branch 8 | push: 9 | branches: [ main ] 10 | pull_request: 11 | branches: [ main ] 12 | 13 | # Allows you to run this workflow manually from the Actions tab 14 | workflow_dispatch: 15 | 16 | # A workflow run is made up of one or more jobs that can run sequentially or in parallel 17 | jobs: 18 | # This workflow contains a single job called "build" 19 | build: 20 | # The type of runner that the job will run on 21 | runs-on: ubuntu-latest 22 | 23 | # Steps represent a sequence of tasks that will be executed as part of the job 24 | steps: 25 | # Checks-out your repository under $GITHUB_WORKSPACE, so your job can access it 26 | - uses: actions/checkout@v2 27 | 28 | - name: Lint with flake8 29 | run: | 30 | pip install flake8 31 | # stop the build if there are Python syntax errors or undefined names 32 | flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics 33 | # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide 34 | flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics 35 | 36 | - name: Build docker image 37 | run: docker build . -t time_series_forecasting 38 | 39 | - name: Run tests 40 | run: docker run time_series_forecasting sh -c "pytest" 41 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | FROM python:3.8-slim 2 | 3 | WORKDIR /app 4 | 5 | COPY ./time_series_forecasting ./time_series_forecasting 6 | COPY requirements.txt requirements.txt 7 | COPY setup.py setup.py 8 | 9 | COPY tests tests 10 | 11 | RUN pip install torch==1.8.1+cpu -f https://download.pytorch.org/whl/torch_stable.html --no-cache-dir 12 | RUN pip install . --no-cache-dir 13 | -------------------------------------------------------------------------------- /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 | # time_series_forcasting 2 | Code for https://towardsdatascience.com/how-to-use-transformer-networks-to-build-a-forecasting-model-297f9270e630 3 | ### Install (GPU) 4 | 5 | ``` 6 | conda create -n py38 python=3.8 7 | conda activate py38 8 | conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c conda-forge 9 | conda install -c conda-forge jupyterlab 10 | conda install -c conda-forge matplotlib 11 | git clone https://github.com/CVxTz/time_series_forecasting 12 | cd time_series_forecasting 13 | pip install . 14 | ``` 15 | 16 | ### Run 17 | 18 | ``` 19 | bash pipeline.sh 20 | ``` 21 | -------------------------------------------------------------------------------- /data/config.json: -------------------------------------------------------------------------------- 1 | { 2 | "features": [ 3 | "day_of_month", 4 | "day_of_year", 5 | "month", 6 | "week_of_year", 7 | "year" 8 | ], 9 | "target": "views", 10 | "group_by_key": "article", 11 | "lag_features": [ 12 | "views_lag_1" 13 | ] 14 | } -------------------------------------------------------------------------------- /pipeline.sh: -------------------------------------------------------------------------------- 1 | python scripts/generate_time_series.py 2 | 3 | python time_series_forecasting/data_utils.py --csv_path "data/data.csv" \ 4 | --out_path "data/processed_data.csv" \ 5 | --config_path "data/config.json" 6 | 7 | # Train 8 | # Trains a model and saves the model in models/ts_models/ 9 | 10 | python time_series_forecasting/training.py --data_csv_path "data/processed_data.csv" \ 11 | --feature_target_names_path "data/config.json" \ 12 | --output_json_path "models/trained_config.json" \ 13 | --log_dir "models/ts_views_logs" \ 14 | --model_dir "models/ts_views_models" 15 | 16 | python time_series_forecasting/evaluation.py --data_csv_path "data/processed_data.csv" \ 17 | --feature_target_names_path "data/config.json" \ 18 | --trained_json_path "data/trained_config.json" \ 19 | --eval_json_path "data/eval.json" \ 20 | --data_for_visualization_path "data/visualization.json" 21 | 22 | python time_series_forecasting/plot_images.py -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | pytorch_lightning==1.2.8 2 | pytest==6.2.3 3 | numpy==1.19.2 4 | pandas==1.2.4 5 | scikit_learn==0.24.1 6 | tqdm==4.60.0 -------------------------------------------------------------------------------- /scripts/generate_time_series.py: -------------------------------------------------------------------------------- 1 | import random 2 | 3 | import pandas as pd 4 | from tqdm import tqdm 5 | import numpy as np 6 | from uuid import uuid4 7 | 8 | periods = [7, 14, 28, 30] 9 | 10 | 11 | def get_init_df(): 12 | 13 | date_rng = pd.date_range(start="2015-01-01", end="2020-01-01", freq="D") 14 | 15 | dataframe = pd.DataFrame(date_rng, columns=["timestamp"]) 16 | 17 | dataframe["index"] = range(dataframe.shape[0]) 18 | 19 | dataframe["article"] = uuid4().hex 20 | 21 | return dataframe 22 | 23 | 24 | def set_amplitude(dataframe): 25 | 26 | max_step = random.randint(90, 365) 27 | max_amplitude = random.uniform(0.1, 1) 28 | offset = random.uniform(-1, 1) 29 | 30 | phase = random.randint(-1000, 1000) 31 | 32 | amplitude = ( 33 | dataframe["index"] 34 | .apply(lambda x: max_amplitude * (x % max_step + phase) / max_step + offset) 35 | .values 36 | ) 37 | 38 | if random.random() < 0.5: 39 | amplitude = amplitude[::-1] 40 | 41 | dataframe["amplitude"] = amplitude 42 | 43 | return dataframe 44 | 45 | 46 | def set_offset(dataframe): 47 | 48 | max_step = random.randint(15, 45) 49 | max_offset = random.uniform(-1, 1) 50 | base_offset = random.uniform(-1, 1) 51 | 52 | phase = random.randint(-1000, 1000) 53 | 54 | offset = ( 55 | dataframe["index"] 56 | .apply( 57 | lambda x: max_offset * np.cos(x * 2 * np.pi / max_step + phase) 58 | + base_offset 59 | ) 60 | .values 61 | ) 62 | 63 | if random.random() < 0.5: 64 | offset = offset[::-1] 65 | 66 | dataframe["offset"] = offset 67 | 68 | return dataframe 69 | 70 | 71 | def generate_time_series(dataframe): 72 | 73 | clip_val = random.uniform(0.3, 1) 74 | 75 | period = random.choice(periods) 76 | 77 | phase = random.randint(-1000, 1000) 78 | 79 | dataframe["views"] = dataframe.apply( 80 | lambda x: np.clip( 81 | np.cos(x["index"] * 2 * np.pi / period + phase), -clip_val, clip_val 82 | ) 83 | * x["amplitude"] 84 | + x["offset"], 85 | axis=1, 86 | ) + np.random.normal( 87 | 0, dataframe["amplitude"].abs().max() / 10, size=(dataframe.shape[0],) 88 | ) 89 | 90 | return dataframe 91 | 92 | 93 | def generate_df(): 94 | dataframe = get_init_df() 95 | dataframe = set_amplitude(dataframe) 96 | dataframe = set_offset(dataframe) 97 | dataframe = generate_time_series(dataframe) 98 | return dataframe 99 | 100 | 101 | if __name__ == "__main__": 102 | 103 | import matplotlib.pyplot as plt 104 | 105 | dataframes = [] 106 | 107 | for _ in tqdm(range(20000)): 108 | df = generate_df() 109 | 110 | # fig = plt.figure() 111 | # plt.plot(df[-120:]["index"], df[-120:]["views"]) 112 | # plt.show() 113 | 114 | dataframes.append(df) 115 | 116 | all_data = pd.concat(dataframes, ignore_index=True) 117 | 118 | all_data.to_csv("data/data.csv", index=False) 119 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | from setuptools import find_packages 4 | from setuptools import setup 5 | 6 | path = os.path.abspath(os.path.dirname(__file__)) 7 | 8 | try: 9 | with open(os.path.join(path, "requirements.txt"), encoding="utf-8") as f: 10 | REQUIRED = f.read().split("\n") 11 | except FileNotFoundError: 12 | REQUIRED = [] 13 | 14 | setup( 15 | name="time_series_forecasting", 16 | version="0.1", 17 | description="Time series forecasting", 18 | author="TimeSeries", 19 | url="https://github.com/CVxTz/time_series_forecasting", 20 | install_requires=REQUIRED, 21 | classifiers=[ 22 | "Intended Audience :: Developers", 23 | "Intended Audience :: Education", 24 | "Intended Audience :: Science/Research", 25 | "Programming Language :: Python :: 3", 26 | "Programming Language :: Python :: 3.6", 27 | "Topic :: Software Development :: Libraries", 28 | "Topic :: Software Development :: Libraries :: Python Modules", 29 | ], 30 | packages=find_packages(exclude=("example", "app", "data", "docker", "tests")), 31 | ) 32 | -------------------------------------------------------------------------------- /tests/unit/test_data_utils.py: -------------------------------------------------------------------------------- 1 | import pandas as pd 2 | import pytest 3 | 4 | from time_series_forecasting.data_utils import ( 5 | add_date_cols, 6 | add_basic_lag_features, 7 | ) 8 | 9 | 10 | def data(start_date="2018-01-01"): 11 | df = pd.DataFrame( 12 | { 13 | "key": ["A"] * 100 + ["B"] * 200, 14 | "value": list(range(100)) + list(range(1000, 1200)), 15 | "date": pd.date_range(start_date, periods=100, freq="D").tolist() 16 | + pd.date_range(start_date, periods=200, freq="D").tolist(), 17 | } 18 | ) 19 | return df 20 | 21 | 22 | @pytest.fixture 23 | def tr_data(): 24 | return data("2018-01-01") 25 | 26 | 27 | @pytest.fixture 28 | def te_data(): 29 | return data("2020-01-01") 30 | 31 | 32 | def test_add_date_cols(tr_data): 33 | df, new_cols = add_date_cols(tr_data, date_col="date") 34 | assert new_cols == ["day_of_month", "day_of_year", "month", "week_of_year", "year"] 35 | 36 | 37 | def test_add_basic_lag_features(tr_data): 38 | df, new_cols = add_basic_lag_features( 39 | tr_data, 40 | group_by_cols=["key"], 41 | col_names=["value", "date"], 42 | horizons=[0, 1, 2], 43 | fill_na=False, 44 | ) 45 | 46 | diff_date_0 = (df["date_lag_0"] - df["date"]).dropna().dt.days 47 | diff_date_1 = (df["date_lag_1"] - df["date"]).dropna().dt.days 48 | 49 | diff_value_2 = (df["value_lag_2"] - df["value"]).dropna() 50 | 51 | assert new_cols == [ 52 | "value_lag_0", 53 | "date_lag_0", 54 | "value_lag_1", 55 | "date_lag_1", 56 | "value_lag_2", 57 | "date_lag_2", 58 | ] 59 | 60 | assert (diff_date_0 == 0).all() 61 | assert (diff_date_1 == -1).all() 62 | assert (diff_value_2 == -2).all() 63 | -------------------------------------------------------------------------------- /tests/unit/test_model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from time_series_forecasting.model import TimeSeriesForcasting, smape_loss 4 | 5 | 6 | def test_smape_loss(): 7 | target = torch.arange(1, 100) 8 | y_pred = target + 10 9 | 10 | loss = smape_loss(y_pred=y_pred, target=target).item() 11 | 12 | assert loss == 0.29728013277053833 13 | 14 | 15 | def test_model(): 16 | source = torch.rand(size=(32, 16, 9)) 17 | target_in = torch.rand(size=(32, 16, 8)) 18 | target_out = torch.rand(size=(32, 16, 1)) 19 | 20 | ts = TimeSeriesForcasting(n_encoder_inputs=9, n_decoder_inputs=8) 21 | 22 | pred = ts((source, target_in)) 23 | 24 | ts.training_step((source, target_in, target_out), batch_idx=1) 25 | 26 | assert pred.size() == torch.Size([32, 16, 1]) 27 | -------------------------------------------------------------------------------- /time_series_forecasting/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/CVxTz/time_series_forecasting/26de501ec28bab1153a8a7abb2ad664d446ada57/time_series_forecasting/__init__.py -------------------------------------------------------------------------------- /time_series_forecasting/data_utils.py: -------------------------------------------------------------------------------- 1 | import json 2 | from typing import List 3 | import pandas as pd 4 | from pathlib import Path 5 | import numpy as np 6 | 7 | 8 | def add_date_cols(dataframe: pd.DataFrame, date_col: str = "timestamp"): 9 | """ 10 | add time features like month, week of the year ... 11 | :param dataframe: 12 | :param date_col: 13 | :return: 14 | """ 15 | 16 | dataframe[date_col] = pd.to_datetime(dataframe[date_col], format="%Y-%m-%d") 17 | 18 | dataframe["day_of_month"] = dataframe[date_col].dt.day / 31 19 | dataframe["day_of_year"] = dataframe[date_col].dt.dayofyear / 365 20 | dataframe["month"] = dataframe[date_col].dt.month / 12 21 | dataframe["week_of_year"] = dataframe[date_col].dt.isocalendar().week / 53 22 | dataframe["year"] = (dataframe[date_col].dt.year - 2015) / 5 23 | 24 | return dataframe, ["day_of_month", "day_of_year", "month", "week_of_year", "year"] 25 | 26 | 27 | def add_basic_lag_features( 28 | dataframe: pd.DataFrame, 29 | group_by_cols: List, 30 | col_names: List, 31 | horizons: List, 32 | fill_na=True, 33 | ): 34 | """ 35 | Computes simple lag features 36 | :param dataframe: 37 | :param group_by_cols: 38 | :param col_names: 39 | :param horizons: 40 | :param fill_na: 41 | :return: 42 | """ 43 | group_by_data = dataframe.groupby(by=group_by_cols) 44 | 45 | new_cols = [] 46 | 47 | for horizon in horizons: 48 | dataframe[[a + "_lag_%s" % horizon for a in col_names]] = group_by_data[ 49 | col_names 50 | ].shift(periods=horizon) 51 | new_cols += [a + "_lag_%s" % horizon for a in col_names] 52 | 53 | if fill_na: 54 | dataframe[new_cols] = dataframe[new_cols].fillna(0) 55 | 56 | return dataframe, new_cols 57 | 58 | 59 | def process_df(dataframe: pd.DataFrame, target_col: str = "views"): 60 | 61 | """ 62 | :param dataframe: 63 | :param target_col: 64 | :return: 65 | """ 66 | 67 | dataframe, new_cols = add_date_cols(dataframe, date_col="timestamp") 68 | dataframe, lag_cols = add_basic_lag_features( 69 | dataframe, group_by_cols=["article"], col_names=[target_col], horizons=[1] 70 | ) 71 | 72 | return dataframe, new_cols 73 | 74 | 75 | if __name__ == "__main__": 76 | 77 | import argparse 78 | 79 | parser = argparse.ArgumentParser() 80 | parser.add_argument("--csv_path") 81 | parser.add_argument("--out_path") 82 | parser.add_argument("--config_path") 83 | args = parser.parse_args() 84 | 85 | data = pd.read_csv(args.csv_path) 86 | 87 | data, cols = process_df(data) 88 | 89 | data.to_csv(args.out_path, index=False) 90 | 91 | config = { 92 | "features": cols, 93 | "target": "views", 94 | "group_by_key": "article", 95 | "lag_features": ["views_lag_1"], 96 | } 97 | 98 | with open(args.config_path, "w") as f: 99 | json.dump(config, f, indent=4) 100 | -------------------------------------------------------------------------------- /time_series_forecasting/evaluation.py: -------------------------------------------------------------------------------- 1 | import json 2 | from typing import Optional 3 | import numpy as np 4 | import pandas as pd 5 | import torch 6 | from sklearn.metrics import mean_absolute_error 7 | from tqdm import tqdm 8 | 9 | from time_series_forecasting.model import TimeSeriesForcasting 10 | from time_series_forecasting.training import split_df, Dataset 11 | 12 | 13 | def smape(true, pred): 14 | """ 15 | Symmetric mean absolute percentage error 16 | :param true: 17 | :param pred: 18 | :return: 19 | """ 20 | true = np.array(true) 21 | pred = np.array(pred) 22 | 23 | smape_val = ( 24 | 100 25 | / pred.size 26 | * np.sum(2 * (np.abs(true - pred)) / (np.abs(pred) + np.abs(true) + 1e-8)) 27 | ) 28 | 29 | return smape_val 30 | 31 | 32 | def evaluate_regression(true, pred): 33 | """ 34 | eval mae + smape 35 | :param true: 36 | :param pred: 37 | :return: 38 | """ 39 | 40 | return {"smape": smape(true, pred), "mae": mean_absolute_error(true, pred)} 41 | 42 | 43 | def evaluate( 44 | data_csv_path: str, 45 | feature_target_names_path: str, 46 | trained_json_path: str, 47 | eval_json_path: str, 48 | horizon_size: int = 30, 49 | data_for_visualization_path: Optional[str] = None, 50 | ): 51 | """ 52 | Evaluates the model on the last 8 labeled weeks of the data. 53 | Compares the model to a simple baseline : prediction the last known value 54 | :param data_csv_path: 55 | :param feature_target_names_path: 56 | :param trained_json_path: 57 | :param eval_json_path: 58 | :param horizon_size: 59 | :param data_for_visualization_path: 60 | :return: 61 | """ 62 | data = pd.read_csv(data_csv_path) 63 | 64 | with open(trained_json_path) as f: 65 | model_json = json.load(f) 66 | 67 | model_path = model_json["best_model_path"] 68 | 69 | with open(feature_target_names_path) as f: 70 | feature_target_names = json.load(f) 71 | 72 | target = feature_target_names["target"] 73 | 74 | data_train = data[~data[target].isna()] 75 | 76 | grp_by_train = data_train.groupby(by=feature_target_names["group_by_key"]) 77 | 78 | groups = list(grp_by_train.groups) 79 | 80 | full_groups = [ 81 | grp for grp in groups if grp_by_train.get_group(grp).shape[0] > horizon_size 82 | ] 83 | 84 | val_data = Dataset( 85 | groups=full_groups, 86 | grp_by=grp_by_train, 87 | split="val", 88 | features=feature_target_names["features"], 89 | target=feature_target_names["target"], 90 | ) 91 | 92 | model = TimeSeriesForcasting( 93 | n_encoder_inputs=len(feature_target_names["features"]) + 1, 94 | n_decoder_inputs=len(feature_target_names["features"]) + 1, 95 | lr=1e-4, 96 | dropout=0.5, 97 | ) 98 | model.load_state_dict(torch.load(model_path)["state_dict"]) 99 | 100 | model.eval() 101 | 102 | gt = [] 103 | baseline_last_known_values = [] 104 | neural_predictions = [] 105 | 106 | data_for_visualization = [] 107 | 108 | for i, group in tqdm(enumerate(full_groups[:100])): 109 | time_series_data = {"history": [], "ground_truth": [], "prediction": []} 110 | 111 | df = grp_by_train.get_group(group) 112 | src, trg = split_df(df, split="val") 113 | 114 | time_series_data["history"] = src[target].tolist()[-120:] 115 | time_series_data["ground_truth"] = trg[target].tolist() 116 | 117 | last_known_value = src[target].values[-1] 118 | 119 | trg["last_known_value"] = last_known_value 120 | 121 | gt += trg[target].tolist() 122 | baseline_last_known_values += trg["last_known_value"].tolist() 123 | 124 | src, trg_in, _ = val_data[i] 125 | 126 | src, trg_in = src.unsqueeze(0), trg_in.unsqueeze(0) 127 | 128 | with torch.no_grad(): 129 | prediction = model((src, trg_in[:, :1, :])) 130 | for j in range(1, horizon_size): 131 | last_prediction = prediction[0, -1] 132 | trg_in[:, j, -1] = last_prediction 133 | prediction = model((src, trg_in[:, : (j + 1), :])) 134 | 135 | trg[target + "_predicted"] = (prediction.squeeze().numpy()).tolist() 136 | 137 | neural_predictions += trg[target + "_predicted"].tolist() 138 | 139 | time_series_data["prediction"] = trg[target + "_predicted"].tolist() 140 | 141 | data_for_visualization.append(time_series_data) 142 | 143 | baseline_eval = evaluate_regression(gt, baseline_last_known_values) 144 | model_eval = evaluate_regression(gt, neural_predictions) 145 | 146 | eval_dict = { 147 | "Baseline_MAE": baseline_eval["mae"], 148 | "Baseline_SMAPE": baseline_eval["smape"], 149 | "Model_MAE": model_eval["mae"], 150 | "Model_SMAPE": model_eval["smape"], 151 | } 152 | 153 | if eval_json_path is not None: 154 | with open(eval_json_path, "w") as f: 155 | json.dump(eval_dict, f, indent=4) 156 | 157 | if data_for_visualization_path is not None: 158 | with open(data_for_visualization_path, "w") as f: 159 | json.dump(data_for_visualization, f, indent=4) 160 | 161 | for k, v in eval_dict.items(): 162 | print(k, v) 163 | 164 | return eval_dict 165 | 166 | 167 | if __name__ == "__main__": 168 | import argparse 169 | 170 | parser = argparse.ArgumentParser() 171 | parser.add_argument("--data_csv_path") 172 | parser.add_argument("--feature_target_names_path") 173 | parser.add_argument("--trained_json_path") 174 | parser.add_argument("--eval_json_path", default=None) 175 | parser.add_argument("--data_for_visualization_path", default=None) 176 | args = parser.parse_args() 177 | 178 | evaluate( 179 | data_csv_path=args.data_csv_path, 180 | feature_target_names_path=args.feature_target_names_path, 181 | trained_json_path=args.trained_json_path, 182 | eval_json_path=args.eval_json_path, 183 | data_for_visualization_path=args.data_for_visualization_path, 184 | ) 185 | -------------------------------------------------------------------------------- /time_series_forecasting/model.py: -------------------------------------------------------------------------------- 1 | import pytorch_lightning as pl 2 | import torch 3 | import torch.nn as nn 4 | from torch.nn import Linear 5 | 6 | 7 | def smape_loss(y_pred, target): 8 | loss = 2 * (y_pred - target).abs() / (y_pred.abs() + target.abs() + 1e-8) 9 | return loss.mean() 10 | 11 | 12 | def gen_trg_mask(length, device): 13 | mask = torch.tril(torch.ones(length, length, device=device)) == 1 14 | 15 | mask = ( 16 | mask.float() 17 | .masked_fill(mask == 0, float("-inf")) 18 | .masked_fill(mask == 1, float(0.0)) 19 | ) 20 | 21 | return mask 22 | 23 | 24 | class TimeSeriesForcasting(pl.LightningModule): 25 | def __init__( 26 | self, 27 | n_encoder_inputs, 28 | n_decoder_inputs, 29 | channels=512, 30 | dropout=0.1, 31 | lr=1e-4, 32 | ): 33 | super().__init__() 34 | 35 | self.save_hyperparameters() 36 | 37 | self.lr = lr 38 | self.dropout = dropout 39 | 40 | self.input_pos_embedding = torch.nn.Embedding(1024, embedding_dim=channels) 41 | self.target_pos_embedding = torch.nn.Embedding(1024, embedding_dim=channels) 42 | 43 | encoder_layer = nn.TransformerEncoderLayer( 44 | d_model=channels, 45 | nhead=8, 46 | dropout=self.dropout, 47 | dim_feedforward=4 * channels, 48 | ) 49 | decoder_layer = nn.TransformerDecoderLayer( 50 | d_model=channels, 51 | nhead=8, 52 | dropout=self.dropout, 53 | dim_feedforward=4 * channels, 54 | ) 55 | 56 | self.encoder = torch.nn.TransformerEncoder(encoder_layer, num_layers=8) 57 | self.decoder = torch.nn.TransformerDecoder(decoder_layer, num_layers=8) 58 | 59 | self.input_projection = Linear(n_encoder_inputs, channels) 60 | self.output_projection = Linear(n_decoder_inputs, channels) 61 | 62 | self.linear = Linear(channels, 1) 63 | 64 | self.do = nn.Dropout(p=self.dropout) 65 | 66 | def encode_src(self, src): 67 | src_start = self.input_projection(src).permute(1, 0, 2) 68 | 69 | in_sequence_len, batch_size = src_start.size(0), src_start.size(1) 70 | pos_encoder = ( 71 | torch.arange(0, in_sequence_len, device=src.device) 72 | .unsqueeze(0) 73 | .repeat(batch_size, 1) 74 | ) 75 | 76 | pos_encoder = self.input_pos_embedding(pos_encoder).permute(1, 0, 2) 77 | 78 | src = src_start + pos_encoder 79 | 80 | src = self.encoder(src) + src_start 81 | 82 | return src 83 | 84 | def decode_trg(self, trg, memory): 85 | 86 | trg_start = self.output_projection(trg).permute(1, 0, 2) 87 | 88 | out_sequence_len, batch_size = trg_start.size(0), trg_start.size(1) 89 | 90 | pos_decoder = ( 91 | torch.arange(0, out_sequence_len, device=trg.device) 92 | .unsqueeze(0) 93 | .repeat(batch_size, 1) 94 | ) 95 | pos_decoder = self.target_pos_embedding(pos_decoder).permute(1, 0, 2) 96 | 97 | trg = pos_decoder + trg_start 98 | 99 | trg_mask = gen_trg_mask(out_sequence_len, trg.device) 100 | 101 | out = self.decoder(tgt=trg, memory=memory, tgt_mask=trg_mask) + trg_start 102 | 103 | out = out.permute(1, 0, 2) 104 | 105 | out = self.linear(out) 106 | 107 | return out 108 | 109 | def forward(self, x): 110 | src, trg = x 111 | 112 | src = self.encode_src(src) 113 | 114 | out = self.decode_trg(trg=trg, memory=src) 115 | 116 | return out 117 | 118 | def training_step(self, batch, batch_idx): 119 | src, trg_in, trg_out = batch 120 | 121 | y_hat = self((src, trg_in)) 122 | 123 | y_hat = y_hat.view(-1) 124 | y = trg_out.view(-1) 125 | 126 | loss = smape_loss(y_hat, y) 127 | 128 | self.log("train_loss", loss) 129 | 130 | return loss 131 | 132 | def validation_step(self, batch, batch_idx): 133 | src, trg_in, trg_out = batch 134 | 135 | y_hat = self((src, trg_in)) 136 | 137 | y_hat = y_hat.view(-1) 138 | y = trg_out.view(-1) 139 | 140 | loss = smape_loss(y_hat, y) 141 | 142 | self.log("valid_loss", loss) 143 | 144 | return loss 145 | 146 | def test_step(self, batch, batch_idx): 147 | src, trg_in, trg_out = batch 148 | 149 | y_hat = self((src, trg_in)) 150 | 151 | y_hat = y_hat.view(-1) 152 | y = trg_out.view(-1) 153 | 154 | loss = smape_loss(y_hat, y) 155 | 156 | self.log("test_loss", loss) 157 | 158 | return loss 159 | 160 | def configure_optimizers(self): 161 | optimizer = torch.optim.Adam(self.parameters(), lr=self.lr) 162 | scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( 163 | optimizer, patience=10, factor=0.1 164 | ) 165 | return { 166 | "optimizer": optimizer, 167 | "lr_scheduler": scheduler, 168 | "monitor": "valid_loss", 169 | } 170 | 171 | 172 | if __name__ == "__main__": 173 | n_classes = 100 174 | 175 | source = torch.rand(size=(32, 16, 9)) 176 | target_in = torch.rand(size=(32, 16, 8)) 177 | target_out = torch.rand(size=(32, 16, 1)) 178 | 179 | ts = TimeSeriesForcasting(n_encoder_inputs=9, n_decoder_inputs=8) 180 | 181 | pred = ts((source, target_in)) 182 | 183 | print(pred.size()) 184 | 185 | ts.training_step((source, target_in, target_out), batch_idx=1) 186 | -------------------------------------------------------------------------------- /time_series_forecasting/plot_images.py: -------------------------------------------------------------------------------- 1 | import json 2 | import os 3 | 4 | import matplotlib.pyplot as plt 5 | 6 | 7 | if __name__ == "__main__": 8 | 9 | with open("data/visualization.json", "r") as f: 10 | data = json.load(f) 11 | 12 | os.makedirs("data/images", exist_ok=True) 13 | 14 | for i, sample in enumerate(data): 15 | hist_size = len(sample["history"]) 16 | gt_size = len(sample["ground_truth"]) 17 | plt.figure() 18 | plt.plot(range(hist_size), sample["history"], label="History") 19 | plt.plot( 20 | range(hist_size, hist_size + gt_size), sample["ground_truth"], label="Ground Truth" 21 | ) 22 | plt.plot( 23 | range(hist_size, hist_size + gt_size), sample["prediction"], label="Prediction" 24 | ) 25 | 26 | plt.xlabel("Time") 27 | 28 | plt.ylabel("Time Series") 29 | 30 | plt.legend() 31 | 32 | plt.savefig(f"data/images/{i}.png") 33 | plt.close() 34 | -------------------------------------------------------------------------------- /time_series_forecasting/training.py: -------------------------------------------------------------------------------- 1 | import json 2 | import random 3 | 4 | import numpy as np 5 | import pandas as pd 6 | import pytorch_lightning as pl 7 | import torch 8 | from pytorch_lightning.callbacks import ModelCheckpoint 9 | from pytorch_lightning.loggers import TensorBoardLogger 10 | from torch.utils.data import DataLoader 11 | 12 | from time_series_forecasting.model import TimeSeriesForcasting 13 | 14 | 15 | def split_df( 16 | df: pd.DataFrame, split: str, history_size: int = 120, horizon_size: int = 30 17 | ): 18 | """ 19 | Create a training / validation samples 20 | Validation samples are the last horizon_size rows 21 | 22 | :param df: 23 | :param split: 24 | :param history_size: 25 | :param horizon_size: 26 | :return: 27 | """ 28 | if split == "train": 29 | end_index = random.randint(horizon_size + 1, df.shape[0] - horizon_size) 30 | elif split in ["val", "test"]: 31 | end_index = df.shape[0] 32 | else: 33 | raise ValueError 34 | 35 | label_index = end_index - horizon_size 36 | start_index = max(0, label_index - history_size) 37 | 38 | history = df[start_index:label_index] 39 | targets = df[label_index:end_index] 40 | 41 | return history, targets 42 | 43 | 44 | def pad_arr(arr: np.ndarray, expected_size: int = 120): 45 | """ 46 | Pad top of array when there is not enough history 47 | :param arr: 48 | :param expected_size: 49 | :return: 50 | """ 51 | arr = np.pad(arr, [(expected_size - arr.shape[0], 0), (0, 0)], mode="edge") 52 | return arr 53 | 54 | 55 | def df_to_np(df): 56 | arr = np.array(df) 57 | arr = pad_arr(arr) 58 | return arr 59 | 60 | 61 | class Dataset(torch.utils.data.Dataset): 62 | def __init__(self, groups, grp_by, split, features, target): 63 | self.groups = groups 64 | self.grp_by = grp_by 65 | self.split = split 66 | self.features = features 67 | self.target = target 68 | 69 | def __len__(self): 70 | return len(self.groups) 71 | 72 | def __getitem__(self, idx): 73 | group = self.groups[idx] 74 | 75 | df = self.grp_by.get_group(group) 76 | 77 | src, trg = split_df(df, split=self.split) 78 | 79 | src = src[self.features + [self.target]] 80 | 81 | src = df_to_np(src) 82 | 83 | trg_in = trg[self.features + [f"{self.target}_lag_1"]] 84 | 85 | trg_in = np.array(trg_in) 86 | trg_out = np.array(trg[self.target]) 87 | 88 | src = torch.tensor(src, dtype=torch.float) 89 | trg_in = torch.tensor(trg_in, dtype=torch.float) 90 | trg_out = torch.tensor(trg_out, dtype=torch.float) 91 | 92 | return src, trg_in, trg_out 93 | 94 | 95 | def train( 96 | data_csv_path: str, 97 | feature_target_names_path: str, 98 | output_json_path: str, 99 | log_dir: str = "ts_logs", 100 | model_dir: str = "ts_models", 101 | batch_size: int = 32, 102 | epochs: int = 2000, 103 | horizon_size: int = 30, 104 | ): 105 | data = pd.read_csv(data_csv_path) 106 | 107 | with open(feature_target_names_path) as f: 108 | feature_target_names = json.load(f) 109 | 110 | data_train = data[~data[feature_target_names["target"]].isna()] 111 | 112 | grp_by_train = data_train.groupby(by=feature_target_names["group_by_key"]) 113 | 114 | groups = list(grp_by_train.groups) 115 | 116 | full_groups = [ 117 | grp for grp in groups if grp_by_train.get_group(grp).shape[0] > 2 * horizon_size 118 | ] 119 | 120 | train_data = Dataset( 121 | groups=full_groups, 122 | grp_by=grp_by_train, 123 | split="train", 124 | features=feature_target_names["features"], 125 | target=feature_target_names["target"], 126 | ) 127 | val_data = Dataset( 128 | groups=full_groups, 129 | grp_by=grp_by_train, 130 | split="val", 131 | features=feature_target_names["features"], 132 | target=feature_target_names["target"], 133 | ) 134 | 135 | print("len(train_data)", len(train_data)) 136 | print("len(val_data)", len(val_data)) 137 | 138 | train_loader = DataLoader( 139 | train_data, 140 | batch_size=batch_size, 141 | num_workers=10, 142 | shuffle=True, 143 | ) 144 | val_loader = DataLoader( 145 | val_data, 146 | batch_size=batch_size, 147 | num_workers=10, 148 | shuffle=False, 149 | ) 150 | 151 | model = TimeSeriesForcasting( 152 | n_encoder_inputs=len(feature_target_names["features"]) + 1, 153 | n_decoder_inputs=len(feature_target_names["features"]) + 1, 154 | lr=1e-5, 155 | dropout=0.1, 156 | ) 157 | 158 | logger = TensorBoardLogger( 159 | save_dir=log_dir, 160 | ) 161 | 162 | checkpoint_callback = ModelCheckpoint( 163 | monitor="valid_loss", 164 | mode="min", 165 | dirpath=model_dir, 166 | filename="ts", 167 | ) 168 | 169 | trainer = pl.Trainer( 170 | max_epochs=epochs, 171 | gpus=1, 172 | logger=logger, 173 | callbacks=[checkpoint_callback], 174 | ) 175 | trainer.fit(model, train_loader, val_loader) 176 | 177 | result_val = trainer.test(test_dataloaders=val_loader) 178 | 179 | output_json = { 180 | "val_loss": result_val[0]["test_loss"], 181 | "best_model_path": checkpoint_callback.best_model_path, 182 | } 183 | 184 | if output_json_path is not None: 185 | with open(output_json_path, "w") as f: 186 | json.dump(output_json, f, indent=4) 187 | 188 | return output_json 189 | 190 | 191 | if __name__ == "__main__": 192 | import argparse 193 | 194 | parser = argparse.ArgumentParser() 195 | parser.add_argument("--data_csv_path") 196 | parser.add_argument("--feature_target_names_path") 197 | parser.add_argument("--output_json_path", default=None) 198 | parser.add_argument("--log_dir") 199 | parser.add_argument("--model_dir") 200 | parser.add_argument("--epochs", type=int, default=2000) 201 | args = parser.parse_args() 202 | 203 | train( 204 | data_csv_path=args.data_csv_path, 205 | feature_target_names_path=args.feature_target_names_path, 206 | output_json_path=args.output_json_path, 207 | log_dir=args.log_dir, 208 | model_dir=args.model_dir, 209 | epochs=args.epochs, 210 | ) 211 | --------------------------------------------------------------------------------