├── .flake8 ├── .gitignore ├── .mypy.ini ├── .pylintrc ├── .style.yapf ├── LICENSE ├── README.md ├── fastc ├── __init__.py ├── classifiers │ ├── centroids.py │ ├── embeddings.py │ ├── interface.py │ ├── loader.py │ └── logistic_regression.py ├── fastc.py ├── kernels.py ├── pooling.py └── template.py ├── install.sh ├── misc └── fastc.svg ├── pyproject.toml ├── requirements-dev.txt ├── requirements.txt ├── server ├── config.yaml.example ├── docker │ ├── Dockerfile │ └── docker-compose.yaml ├── main.py ├── requirements.txt ├── resources │ ├── inference.py │ └── version.py └── scripts │ ├── start-docker.sh │ └── start-server.sh └── upload.sh /.flake8: -------------------------------------------------------------------------------- 1 | [flake8] 2 | max-line-length = 79 3 | ignore = W503 4 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Custom 2 | config.yaml 3 | .dccache 4 | 5 | # Byte-compiled / optimized / DLL files 6 | __pycache__/ 7 | *.py[cod] 8 | *$py.class 9 | 10 | # C extensions 11 | *.so 12 | 13 | # Distribution / packaging 14 | .Python 15 | build/ 16 | develop-eggs/ 17 | dist/ 18 | downloads/ 19 | eggs/ 20 | .eggs/ 21 | lib/ 22 | lib64/ 23 | parts/ 24 | sdist/ 25 | var/ 26 | wheels/ 27 | *.egg-info/ 28 | .installed.cfg 29 | *.egg 30 | MANIFEST 31 | 32 | # PyInstaller 33 | # Usually these files are written by a python script from a template 34 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 35 | *.manifest 36 | *.spec 37 | 38 | # Installer logs 39 | pip-log.txt 40 | pip-delete-this-directory.txt 41 | 42 | # Unit test / coverage reports 43 | htmlcov/ 44 | .tox/ 45 | .coverage 46 | .coverage.* 47 | .cache 48 | nosetests.xml 49 | coverage.xml 50 | *.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 | 63 | # Flask stuff: 64 | instance/ 65 | .webassets-cache 66 | 67 | # Scrapy stuff: 68 | .scrapy 69 | 70 | # Sphinx documentation 71 | docs/_build/ 72 | 73 | # PyBuilder 74 | target/ 75 | 76 | # Jupyter Notebook 77 | .ipynb_checkpoints 78 | 79 | # pyenv 80 | .python-version 81 | 82 | # celery beat schedule file 83 | celerybeat-schedule 84 | 85 | # SageMath parsed files 86 | *.sage.py 87 | 88 | # Environments 89 | .env 90 | .venv 91 | env/ 92 | venv/ 93 | ENV/ 94 | env.bak/ 95 | venv.bak/ 96 | 97 | # Spyder project settings 98 | .spyderproject 99 | .spyproject 100 | 101 | # Rope project settings 102 | .ropeproject 103 | 104 | # mkdocs documentation 105 | /site 106 | 107 | # mypy 108 | .mypy_cache/ 109 | -------------------------------------------------------------------------------- /.mypy.ini: -------------------------------------------------------------------------------- 1 | [mypy] 2 | disable_error_code = attr-defined 3 | ignore_missing_imports = True 4 | -------------------------------------------------------------------------------- /.pylintrc: -------------------------------------------------------------------------------- 1 | disable=E1101 2 | -------------------------------------------------------------------------------- /.style.yapf: -------------------------------------------------------------------------------- 1 | [style] 2 | based_on_style=pep8 3 | column_limit=79 4 | split_before_arithmetic_operator=true 5 | split_before_logical_operator=true 6 | split_before_named_assigns=true 7 | split_before_first_argument=true 8 | allow_split_before_dict_value=false 9 | dedent_closing_brackets=true 10 | -------------------------------------------------------------------------------- /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 |
2 | Logo 3 |

Unattended Lightweight Text Classifiers with LLM Embeddings

4 |
5 |
6 | 7 |

8 | PyPi 9 | License 10 |

11 | 12 | 13 | # Key features 14 | - **Suitable for limited-memory CPU execution:** Use efficient distilled models such as [`deepset/tinyroberta-6l-768d`](https://huggingface.co/deepset/tinyroberta-6l-768d) for embedding generation. 15 | - **Logistic Regression and Nearest Centroid classification:** Bypass the need for fine-tuning by utilizing LLM embeddings to efficiently categorize texts using either logistic regression or the nearest centroid through cosine similarity. 16 | - **Efficient Parallel Execution:** Run hundreds of classifiers concurrently with minimal overhead by sharing the same model for embedding generation. 17 | 18 | # Installation 19 | ```bash 20 | pip install -U fastc 21 | ``` 22 | 23 | # Train a model 24 | You can train a text classifier with just a few lines of code: 25 | ```python 26 | from fastc import Fastc 27 | 28 | tuples = [ 29 | ("I just got a promotion! Feeling fantastic.", 'positive'), 30 | ("Today was terrible. I lost my wallet and missed the bus.", 'negative'), 31 | ("I had a great time with my friends at the party.", 'positive'), 32 | ("I'm so frustrated with the traffic jam this morning.", 'negative'), 33 | ("My vacation was wonderful and relaxing.", 'positive'), 34 | ("I didn't get any sleep last night because of the noise.", 'negative'), 35 | ("I'm so excited for the concert tonight!", 'positive'), 36 | ("I'm disappointed with the service at the restaurant.", 'negative'), 37 | ("The weather is beautiful and I enjoyed my walk.", 'positive'), 38 | ("I had a bad day. Nothing went right.", 'negative'), 39 | ("I'm thrilled to announce that we are expecting a baby!", 'positive'), 40 | ("I feel so lonely and sad today.", 'negative'), 41 | ("My team won the championship! We are the champions.", 'positive'), 42 | ("I can't stand my job anymore, it's so stressful.", 'negative'), 43 | ("I love spending time with my family during the holidays.", 'positive'), 44 | ("My computer crashed and I lost all my work.", 'negative'), 45 | ("I'm proud of my achievements this year.", 'positive'), 46 | ("I'm exhausted and overwhelmed with everything.", 'positive'), 47 | ] 48 | ``` 49 | 50 | ## Classification Kernels 51 | ### Nearest Centroid 52 | ```python 53 | model = Fastc( 54 | embeddings_model='microsoft/deberta-base', 55 | kernel=Kernels.NEAREST_CENTROID, 56 | ) 57 | 58 | model.load_dataset(tuples) 59 | model.train() 60 | ``` 61 | 62 | ### Logistic Regression 63 | ```python 64 | from fastc import Kernels 65 | 66 | model = Fastc( 67 | embeddings_model='microsoft/deberta-base', 68 | kernel=Kernels.LOGISTIC_REGRESSION, 69 | # cross_validation_splits=5, 70 | # cross_validation_repeats=3, 71 | # iterations=100, 72 | # parameters={...}, 73 | # seed=1984, 74 | ) 75 | 76 | model.load_dataset(tuples) 77 | model.train() 78 | ``` 79 | 80 | ## Pooling Strategies 81 | The implemented pooling strategies are: 82 | - `MEAN` (default) 83 | - `MEAN_MASKED` 84 | - `MAX` 85 | - `MAX_MASKED` 86 | - `CLS` 87 | - `SUM` 88 | - `ATTENTION_WEIGHTED` 89 | 90 | ```python 91 | from fastc import Pooling 92 | 93 | model = Fastc( 94 | embeddings_model='microsoft/deberta-base', 95 | pooling=Pooling.MEAN_MASKED, 96 | ) 97 | 98 | model.load_dataset(tuples) 99 | model.train() 100 | ``` 101 | 102 | ## Templates and Instruct Models 103 | You can use instruct templates with instruct models such as `intfloat/multilingual-e5-large-instruct`. Other models may also improve in performance by using templates, even if they were not explicitly trained with them. 104 | 105 | ```python 106 | from fastc import ModelTemplates, Fastc, Template 107 | 108 | # template_text = 'Instruct: {instruction}\nQuery: {text}' 109 | template_text = ModelTemplates.E5_INSTRUCT 110 | 111 | model = Fastc( 112 | embeddings_model='intfloat/multilingual-e5-large-instruct', 113 | template=Template( 114 | template_text, 115 | instruction='Classify as positive or negative' 116 | ), 117 | ) 118 | ``` 119 | 120 | # Save, load and export models 121 | After training, you can save the model for future use: 122 | ```python 123 | model.save_model('./sentiment-classifier/') 124 | ``` 125 | 126 | ## Publish a model to HuggingFace 127 | > [!IMPORTANT] 128 | > Log in to HuggingFace first with `huggingface-cli login` 129 | 130 | ```python 131 | model.push_to_hub( 132 | 'braindao/sentiment-classifier', 133 | tags=['sentiment-analysis'], 134 | languages=['multilingual'], 135 | private=False, 136 | ) 137 | ``` 138 | 139 | ## Load an existing model 140 | You can load a pre-trained model either from a directory or from HuggingFace: 141 | ```python 142 | # From a directory 143 | model = Fastc('./sentiment-classifier/') 144 | 145 | # From HuggingFace 146 | model = Fastc('braindao/sentiment-classifier') 147 | ``` 148 | 149 | # Class prediction 150 | ```python 151 | sentences = [ 152 | 'I am feeling well.', 153 | 'I am in pain.', 154 | ] 155 | 156 | # Single prediction 157 | scores = model.predict_one(sentences[0]) 158 | print(scores['label']) 159 | 160 | # Batch predictions 161 | scores_list = model.predict(sentences) 162 | for scores in scores_list: 163 | print(scores['label']) 164 | ``` 165 | 166 | # Inference Server 167 | 168 | To launch the dockerized inference server, use the following script: 169 | ```bash 170 | ./server/scripts/start-docker.sh 171 | ``` 172 | 173 | Alternatively, on the host machine: 174 | ```bash 175 | ./server/scripts/start-server.sh 176 | ``` 177 | 178 | In both cases, an HTTP API will be available, listening on the `fastc-server` *[hashport](https://github.com/labteral/hashport)* `53256`. 179 | 180 | ## Inference 181 | 182 | To classify text, use `POST /` with a JSON payload such as: 183 | ```json 184 | { 185 | "model": "braindao/tinyroberta-6l-768d-language-identifier-en-es-ko-zh-fastc-lr", 186 | "text": "오늘 저녁에 친구들과 함께 pizza를 먹을 거예요." 187 | } 188 | ``` 189 | 190 | Response: 191 | ```json 192 | { 193 | "label": "ko", 194 | "scores": { 195 | "en": 1.0146501463135055e-08, 196 | "es": 6.806091549848057e-09, 197 | "ko": 0.9999852640487916, 198 | "zh": 1.471899861513275e-05 199 | } 200 | } 201 | ``` 202 | 203 | ## Version 204 | 205 | To check the `fastc` version, use `GET /version`: 206 | 207 | Response: 208 | ```json 209 | { 210 | "version": "2.2407.0" 211 | } 212 | ``` 213 | -------------------------------------------------------------------------------- /fastc/__init__.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | import sys 5 | import warnings 6 | from os import environ as env 7 | 8 | from .fastc import Fastc # noqa: F401 9 | from .fastc import Kernels # noqa: F401 10 | from .fastc import Pooling, SentenceClassifier # noqa: F401 11 | from .template import ModelTemplates, Template # noqa: F401 12 | 13 | # Backwards compatibility 14 | ModelTypes = Kernels 15 | 16 | env['TOKENIZERS_PARALLELISM'] = 'true' 17 | 18 | if not sys.warnoptions: 19 | warnings.simplefilter("ignore") 20 | env["PYTHONWARNINGS"] = "ignore::UserWarning" 21 | -------------------------------------------------------------------------------- /fastc/classifiers/centroids.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | from typing import Dict, Generator, List 5 | 6 | import torch 7 | import torch.nn.functional as F 8 | 9 | from ..kernels import Kernels 10 | from ..template import Template 11 | from .embeddings import Pooling 12 | from .interface import ClassifierInterface 13 | 14 | 15 | class NearestCentroidClassifier(ClassifierInterface): 16 | def __init__( 17 | self, 18 | embeddings_model: str, 19 | template: Template, 20 | pooling: Pooling, 21 | label_names_by_id: Dict[int, str], 22 | model_data: Dict[int, List[float]] = None, 23 | ): 24 | super().__init__( 25 | embeddings_model=embeddings_model, 26 | template=template, 27 | pooling=pooling, 28 | label_names_by_id=label_names_by_id, 29 | ) 30 | 31 | self._centroids = {} 32 | self._normalized_centroids = {} 33 | 34 | if model_data is not None: 35 | self._load_model(model_data) 36 | 37 | def _load_model(self, centroids: Dict): 38 | self._centroids = { 39 | self._label_ids_by_name[label]: torch.tensor(centroid) 40 | for label, centroid in centroids.items() 41 | } 42 | self._normalized_centroids = { 43 | label: self._normalize(centroid) 44 | for label, centroid in self._centroids.items() 45 | } 46 | 47 | def train(self): 48 | if self._texts_by_label is None: 49 | raise ValueError("Dataset is not loaded.") 50 | 51 | for label, texts in self._texts_by_label.items(): 52 | texts = [self._template.format(text) for text in texts] 53 | embeddings = list(self.embeddings_model.get_embeddings( 54 | texts=texts, 55 | pooling=self._pooling, 56 | title='Generating embeddings [{}]'.format( 57 | self._label_names_by_id[label], 58 | ), 59 | show_progress=True, 60 | )) 61 | embeddings = torch.stack(embeddings) 62 | centroid = torch.mean(embeddings, dim=0) 63 | self._centroids[label] = centroid 64 | self._normalized_centroids[label] = self._normalize(centroid) 65 | 66 | def predict( 67 | self, 68 | texts: List[str], 69 | ) -> Generator[Dict[int, float], None, None]: 70 | if self._normalized_centroids is None: 71 | raise ValueError("Model is not trained.") 72 | 73 | if not isinstance(texts, list): 74 | raise TypeError("Input must be a list of strings.") 75 | 76 | texts = [self._template.format(text) for text in texts] 77 | embeddings = self.embeddings_model.get_embeddings( 78 | texts=texts, 79 | pooling=self._pooling, 80 | ) 81 | 82 | normalized_embeddings = [ 83 | self._normalize(embedding) 84 | for embedding in embeddings 85 | ] 86 | 87 | for text_embedding in normalized_embeddings: 88 | dot_products = { 89 | label: torch.dot(text_embedding, centroid).item() 90 | for label, centroid in self._normalized_centroids.items() 91 | } 92 | 93 | dot_product_values = torch.tensor(list(dot_products.values())) 94 | softmax_scores = F.softmax(dot_product_values, dim=0).tolist() 95 | 96 | scores = { 97 | self._label_names_by_id[label]: score 98 | for label, score in zip(dot_products.keys(), softmax_scores) 99 | } 100 | 101 | result = { 102 | 'label': max(scores, key=scores.get), 103 | 'scores': scores, 104 | } 105 | 106 | yield result 107 | 108 | def _get_info(self): 109 | info = super()._get_info() 110 | info['model']['kernel'] = Kernels.NEAREST_CENTROID.value 111 | info['model']['data'] = { 112 | self._label_names_by_id[key]: value.tolist() 113 | for key, value in self._centroids.items() 114 | } 115 | return info 116 | -------------------------------------------------------------------------------- /fastc/classifiers/embeddings.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | from typing import Generator, List, Optional 5 | 6 | import torch 7 | import torch.nn.functional as F 8 | from tqdm import tqdm 9 | from transformers import AutoModel, AutoTokenizer 10 | 11 | from ..pooling import Pooling 12 | 13 | 14 | class EmbeddingsModel: 15 | _instances = {} 16 | 17 | def __new__( 18 | cls, 19 | model_name, 20 | output_attentions, 21 | ): 22 | if ( 23 | model_name not in cls._instances 24 | or ( 25 | output_attentions 26 | and not cls._instances[model_name]._output_attentions 27 | ) 28 | ): 29 | instance = super(EmbeddingsModel, cls).__new__(cls) 30 | cls._instances[model_name] = instance 31 | instance._initialized = False 32 | 33 | return cls._instances[model_name] 34 | 35 | def __init__( 36 | self, 37 | model_name, 38 | output_attentions, 39 | ): 40 | if not self._initialized: 41 | self.model_name = model_name 42 | self._output_attentions = output_attentions 43 | self._tokenizer = AutoTokenizer.from_pretrained(model_name) 44 | self._model = AutoModel.from_pretrained( 45 | model_name, 46 | output_attentions=output_attentions, 47 | ) 48 | self._model.eval() 49 | self._initialized = True 50 | 51 | @torch.no_grad() 52 | def get_embeddings( 53 | self, 54 | texts: List[str], 55 | pooling: Pooling, 56 | title: Optional[str] = None, 57 | show_progress: bool = False, 58 | ) -> Generator[torch.Tensor, None, None]: 59 | for text in tqdm( 60 | texts, 61 | desc=title, 62 | unit='text', 63 | disable=not show_progress, 64 | ): 65 | inputs = self._tokenizer( 66 | text, 67 | return_tensors='pt', 68 | padding=True, 69 | truncation=True, 70 | ) 71 | outputs = self._model(**inputs) 72 | attention_mask = inputs['attention_mask'] 73 | 74 | if pooling == Pooling.MEAN: 75 | sentence_embeddings = outputs.last_hidden_state.mean(dim=1) 76 | yield from sentence_embeddings 77 | 78 | elif pooling == Pooling.MEAN_MASKED: 79 | last_hidden_state = outputs.last_hidden_state.masked_fill( 80 | ~attention_mask[..., None].bool(), 81 | 0.0, 82 | ) 83 | sentence_embeddings = ( 84 | last_hidden_state.sum(dim=1) 85 | / attention_mask.sum(dim=1) 86 | [..., None] 87 | ) 88 | yield from sentence_embeddings 89 | 90 | elif pooling == Pooling.ATTENTION_WEIGHTED: 91 | attention_weights = outputs.attentions[-1] 92 | attention_weights = attention_weights.masked_fill( 93 | ~attention_mask[:, None, None, :].bool(), 94 | 0.0, 95 | ) 96 | attention_weights = attention_weights.mean(dim=1) 97 | attention_weights = F.normalize(attention_weights, p=1, dim=-1) 98 | sentence_embeddings = torch.einsum( 99 | 'bsl,bsd->bsd', 100 | attention_weights, 101 | outputs.last_hidden_state, 102 | ).sum(dim=1) 103 | yield from sentence_embeddings 104 | 105 | elif pooling == Pooling.CLS: 106 | sentence_embeddings = outputs.last_hidden_state[:, 0] 107 | yield from sentence_embeddings 108 | 109 | elif pooling == Pooling.MAX: 110 | sentence_embeddings = torch.max( 111 | outputs.last_hidden_state, 112 | dim=1, 113 | )[0] 114 | yield from sentence_embeddings 115 | 116 | elif pooling == Pooling.MAX_MASKED: 117 | last_hidden_state = outputs.last_hidden_state.masked_fill( 118 | ~attention_mask[..., None].bool(), 119 | float('-inf'), 120 | ) 121 | sentence_embeddings = torch.max(last_hidden_state, dim=1)[0] 122 | yield from sentence_embeddings 123 | 124 | elif pooling == Pooling.SUM: 125 | sentence_embeddings = outputs.last_hidden_state.sum(dim=1) 126 | yield from sentence_embeddings 127 | 128 | else: 129 | raise ValueError("Unsupported pooling strategy.") 130 | -------------------------------------------------------------------------------- /fastc/classifiers/interface.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | import json 4 | import os 5 | from typing import Dict, Generator, List, Tuple 6 | 7 | import torch 8 | import torch.nn.functional as F 9 | from huggingface_hub import HfApi 10 | 11 | from ..pooling import ATTENTION_POOLING_STRATEGIES 12 | from ..template import Template 13 | from .embeddings import EmbeddingsModel, Pooling 14 | 15 | FASTC_FORMAT_VERSION = 2.0 16 | 17 | 18 | class ClassifierInterface: 19 | def __init__( 20 | self, 21 | embeddings_model: str, 22 | template: Template, 23 | pooling: Pooling, 24 | label_names_by_id: Dict[int, str], 25 | ): 26 | output_attentions = False 27 | if pooling in ATTENTION_POOLING_STRATEGIES: 28 | output_attentions = True 29 | self._embeddings_model = EmbeddingsModel( 30 | embeddings_model, 31 | output_attentions=output_attentions, 32 | ) 33 | self._template = template 34 | self._pooling = pooling 35 | self._texts_by_label = None 36 | self._label_names_by_id = label_names_by_id 37 | 38 | self._label_ids_by_name = None 39 | if label_names_by_id is not None: 40 | self._label_ids_by_name = { 41 | v: k for k, v in label_names_by_id.items() 42 | } 43 | 44 | @staticmethod 45 | def _normalize(tensor: torch.Tensor) -> torch.Tensor: 46 | return F.normalize(tensor, p=2, dim=-1) 47 | 48 | def load_dataset(self, dataset: List[Tuple[str, str]]): 49 | if not isinstance(dataset, list): 50 | raise TypeError('Dataset must be a list of tuples.') 51 | 52 | texts_by_label = {} 53 | label_names_by_id = {} 54 | label_index = 0 55 | 56 | for text, label in dataset: 57 | if label not in label_names_by_id: 58 | label_names_by_id[label] = label_index 59 | label_index += 1 60 | 61 | label_id = label_names_by_id[label] 62 | 63 | if label_id not in texts_by_label: 64 | texts_by_label[label_id] = [] 65 | 66 | texts_by_label[label_id].append(text) 67 | 68 | self._texts_by_label = texts_by_label 69 | self._label_names_by_id = {v: k for k, v in label_names_by_id.items()} 70 | 71 | @property 72 | def embeddings_model(self): 73 | return self._embeddings_model 74 | 75 | def train(self): 76 | raise NotImplementedError 77 | 78 | def predict(self, texts: List[str]) -> Generator[Dict[int, float], None, None]: # noqa: E501 79 | raise NotImplementedError 80 | 81 | def predict_one(self, text: str) -> Dict[int, float]: 82 | return list(self.predict([text]))[0] 83 | 84 | def _get_info(self): 85 | return { 86 | 'version': FASTC_FORMAT_VERSION, 87 | 'model': { 88 | 'embeddings': self._embeddings_model_name, 89 | 'pooling': self._pooling.value, 90 | 'template': { 91 | 'text': self._template._template, 92 | 'variables': self._template._variables, 93 | }, 94 | 'labels': {v: k for k, v in self._label_names_by_id.items()}, 95 | }, 96 | } 97 | 98 | def save_model( 99 | self, 100 | path: str, 101 | description: str = None, 102 | ): 103 | os.makedirs(path, exist_ok=True) 104 | 105 | model_info = self._get_info() 106 | if description is not None: 107 | model_info['description'] = description 108 | 109 | with open(os.path.join(path, 'config.json'), 'w') as f: 110 | json.dump( 111 | model_info, 112 | f, 113 | indent=4, 114 | ensure_ascii=False, 115 | ) 116 | 117 | @property 118 | def _embeddings_model_name(self): 119 | return self._embeddings_model._model.name_or_path 120 | 121 | def push_to_hub( 122 | self, 123 | repo_id: str, 124 | tags: List[str] = None, 125 | languages: List[str] = None, 126 | private: bool = False, 127 | ): 128 | if tags is None: 129 | tags = [] 130 | tags = ['fastc', 'fastc-{}'.format(FASTC_FORMAT_VERSION)] + tags 131 | 132 | self.save_model('/tmp/fastc') 133 | 134 | api = HfApi() 135 | 136 | api.create_repo( 137 | repo_id=repo_id, 138 | repo_type='model', 139 | private=private, 140 | exist_ok=True, 141 | ) 142 | 143 | readme = ( 144 | '---\n' 145 | 'base_model: {}\n' 146 | ).format(self._embeddings_model_name) 147 | 148 | if languages is not None: 149 | readme += 'language:\n' 150 | for language in languages: 151 | readme += '- {}\n'.format(language) 152 | 153 | readme += 'tags:\n' 154 | for tag in tags: 155 | readme += '- {}\n'.format(tag) 156 | 157 | readme += '---\n\n' 158 | 159 | repo_name = repo_id.split('/')[1] 160 | readme += ( 161 | '# {}\n\n' 162 | '## Install fastc\n' 163 | '```bash\npip install fastc\n```\n\n' 164 | '## Model Inference\n' 165 | '```python\n' 166 | 'from fastc import Fastc\n\n' 167 | 'model = Fastc(\'{}\')\n' 168 | 'label = model.predict_one(text)[\'label\']\n' 169 | '```' 170 | ).format(repo_name, repo_id) 171 | 172 | readme_path = '/tmp/fastc/README.md' 173 | model_path = '/tmp/fastc/config.json' 174 | 175 | with open(readme_path, 'w') as readme_file: 176 | readme_file.write(readme) 177 | 178 | for file_path in [readme_path, model_path]: 179 | base_name = os.path.basename(file_path) 180 | api.upload_file( 181 | repo_id=repo_id, 182 | repo_type='model', 183 | path_or_fileobj=file_path, 184 | path_in_repo=base_name, 185 | commit_message='Updated {} via fastc'.format(base_name), 186 | ) 187 | os.remove(file_path) 188 | -------------------------------------------------------------------------------- /fastc/classifiers/loader.py: -------------------------------------------------------------------------------- 1 | import sys 2 | import threading 3 | import time 4 | 5 | 6 | class Loader: 7 | def __init__(self, description=''): 8 | self.description = description 9 | self.is_running = False 10 | self.animation_thread = None 11 | self.chars = ['|', '/', '-', '\\'] 12 | 13 | def animate(self): 14 | while self.is_running: 15 | for char in self.chars: 16 | sys.stdout.write(f'\r{self.description} {char}') 17 | sys.stdout.flush() 18 | time.sleep(0.1) 19 | 20 | def start(self): 21 | self.is_running = True 22 | self.animation_thread = threading.Thread(target=self.animate) 23 | self.animation_thread.start() 24 | 25 | def stop(self): 26 | self.is_running = False 27 | if self.animation_thread: 28 | self.animation_thread.join() 29 | sys.stdout.write('\r' + ' ' * (len(self.description) + 2) + '\r') 30 | sys.stdout.flush() 31 | -------------------------------------------------------------------------------- /fastc/classifiers/logistic_regression.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | import base64 5 | import io 6 | import warnings 7 | from typing import Dict, Generator, List, Union 8 | 9 | import joblib 10 | from scipy.stats import loguniform 11 | from sklearn.exceptions import ConvergenceWarning 12 | from sklearn.linear_model import LogisticRegression 13 | from sklearn.model_selection import RandomizedSearchCV, RepeatedStratifiedKFold 14 | from sklearn.pipeline import make_pipeline 15 | from sklearn.preprocessing import StandardScaler 16 | 17 | from ..kernels import Kernels 18 | from ..template import Template 19 | from .embeddings import Pooling 20 | from .interface import ClassifierInterface 21 | from .loader import Loader 22 | 23 | 24 | class LogisticRegressionClassifier(ClassifierInterface): 25 | def __init__( 26 | self, 27 | embeddings_model: str, 28 | template: Template, 29 | pooling: Pooling, 30 | label_names_by_id: Dict[int, str], 31 | model_data: Dict[int, List[float]] = None, 32 | cross_validation_splits: int = None, 33 | cross_validation_repeats: int = None, 34 | iterations: int = None, 35 | parameters: Union[Dict, List] = None, 36 | seed: int = None, 37 | ): 38 | super().__init__( 39 | embeddings_model=embeddings_model, 40 | template=template, 41 | pooling=pooling, 42 | label_names_by_id=label_names_by_id, 43 | ) 44 | 45 | self._lr_pipeline = make_pipeline( 46 | StandardScaler(), 47 | LogisticRegression() 48 | ) 49 | 50 | if cross_validation_splits is None: 51 | cross_validation_splits = 5 52 | self._cross_validation_splits = cross_validation_splits 53 | 54 | if cross_validation_repeats is None: 55 | cross_validation_repeats = 3 56 | self._cross_validation_repeats = cross_validation_repeats 57 | 58 | if iterations is None: 59 | iterations = 100 60 | self._iterations = iterations 61 | 62 | self._parameters = parameters 63 | self._seed = seed 64 | 65 | if model_data is None: 66 | return 67 | 68 | self._load_model(model_data) 69 | 70 | def _load_model( 71 | self, 72 | model_data: str, 73 | ): 74 | buffer = io.BytesIO(base64.b64decode(model_data)) 75 | self._model = joblib.load(buffer) 76 | 77 | def train(self): 78 | X = [] 79 | y = [] 80 | 81 | for label, texts in self._texts_by_label.items(): 82 | texts = [self._template.format(text) for text in texts] 83 | embeddings = list(self.embeddings_model.get_embeddings( 84 | texts=texts, 85 | pooling=self._pooling, 86 | title='Generating embeddings [{}]'.format( 87 | self._label_names_by_id[label], 88 | ), 89 | show_progress=True, 90 | )) 91 | 92 | normalized_embeddings = [ 93 | self._normalize(embedding) 94 | for embedding in embeddings 95 | ] 96 | 97 | X.extend(normalized_embeddings) 98 | y.extend([label] * len(texts)) 99 | 100 | common_params = { 101 | 'C': loguniform(1e-6, 1e3), 102 | 'max_iter': [3000], 103 | 'tol': loguniform(1e-6, 1e-2), 104 | 'class_weight': [None, 'balanced'], 105 | 'warm_start': [True, False], 106 | 'fit_intercept': [True, False], 107 | 'intercept_scaling': [0.001, 0.01, 0.1, 0.2, 0.5, 1, 2, 5, 10], 108 | } 109 | 110 | if self._seed is not None: 111 | common_params['random_state'] = [self._seed] 112 | 113 | compatible_params = [ 114 | { 115 | 'solver': ['liblinear'], 116 | 'penalty': ['l1'], 117 | 'dual': [False], 118 | }, 119 | { 120 | 'solver': ['liblinear'], 121 | 'penalty': ['l2'], 122 | 'dual': [True, False], 123 | }, 124 | { 125 | 'solver': ['saga'], 126 | 'penalty': ['l1', 'l2'], 127 | 'l1_ratio': loguniform(1e-2, 1) 128 | }, 129 | { 130 | 'solver': ['saga'], 131 | 'penalty': ['elasticnet'], 132 | 'l1_ratio': loguniform(1e-2, 1) 133 | }, 134 | { 135 | 'solver': ['saga'], 136 | 'penalty': [None], 137 | }, 138 | { 139 | 'solver': ['newton-cg'], 140 | 'penalty': ['l2'], 141 | }, 142 | { 143 | 'solver': ['newton-cg'], 144 | 'penalty': [None], 145 | }, 146 | { 147 | 'solver': ['lbfgs'], 148 | 'penalty': ['l2'], 149 | }, 150 | { 151 | 'solver': ['lbfgs'], 152 | 'penalty': [None], 153 | }, 154 | { 155 | 'solver': ['sag'], 156 | 'penalty': ['l2'], 157 | }, 158 | { 159 | 'solver': ['sag'], 160 | 'penalty': [None], 161 | }, 162 | ] 163 | 164 | def prefix_params(params: Dict): 165 | return { 166 | 'logisticregression__' + key: value 167 | for key, value in params.items() 168 | } 169 | 170 | if self._parameters is None: 171 | parameters = [] 172 | for item in compatible_params: 173 | new_item = item | common_params 174 | 175 | if new_item['penalty'] == [None]: 176 | del new_item['C'] 177 | 178 | if new_item['penalty'] != ['elasticnet']: 179 | if 'l1_ratio' in new_item: 180 | del new_item['l1_ratio'] 181 | 182 | parameters.append(prefix_params(new_item)) 183 | 184 | random_search = RandomizedSearchCV( 185 | self._lr_pipeline, 186 | param_distributions=parameters, 187 | n_iter=self._iterations, 188 | cv=RepeatedStratifiedKFold( 189 | n_splits=self._cross_validation_splits, 190 | n_repeats=self._cross_validation_repeats, 191 | random_state=self._seed, 192 | ), 193 | scoring='accuracy', 194 | n_jobs=-1, 195 | verbose=0, 196 | random_state=self._seed, 197 | ) 198 | 199 | loader = Loader("Training") 200 | loader.start() 201 | with warnings.catch_warnings(): 202 | warnings.simplefilter("ignore", ConvergenceWarning) 203 | warnings.simplefilter("ignore", ConvergenceWarning) 204 | random_search.fit(X, y) 205 | loader.stop() 206 | 207 | self._model = random_search.best_estimator_ 208 | 209 | def predict( 210 | self, 211 | texts: List[str], 212 | ) -> Generator[Dict[int, float], None, None]: 213 | texts = [self._template.format(text) for text in texts] 214 | embeddings = self.embeddings_model.get_embeddings( 215 | texts=texts, 216 | pooling=self._pooling, 217 | ) 218 | 219 | normalized_embeddings = [ 220 | self._normalize(embedding) 221 | for embedding in embeddings 222 | ] 223 | 224 | for text_embedding in normalized_embeddings: 225 | probabilities = self._model.predict_proba([text_embedding])[0] 226 | 227 | scores = { 228 | self._label_names_by_id[label]: probability 229 | for label, probability in enumerate(probabilities) 230 | } 231 | 232 | result = { 233 | 'label': max(scores, key=scores.get), 234 | 'scores': scores, 235 | } 236 | 237 | yield result 238 | 239 | def _get_info(self): 240 | info = super()._get_info() 241 | info['model']['kernel'] = Kernels.LOGISTIC_REGRESSION.value 242 | 243 | buffer = io.BytesIO() 244 | joblib.dump(self._model, buffer, protocol=5) 245 | buffer.seek(0) 246 | model_base64 = base64.b64encode(buffer.read()).decode('utf-8') 247 | 248 | info['model']['data'] = model_base64 249 | return info 250 | -------------------------------------------------------------------------------- /fastc/fastc.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | import json 5 | import os 6 | 7 | from huggingface_hub import hf_hub_download 8 | from transformers import logging 9 | 10 | from .classifiers.centroids import NearestCentroidClassifier 11 | from .classifiers.embeddings import Pooling 12 | from .classifiers.logistic_regression import LogisticRegressionClassifier 13 | from .kernels import Kernels 14 | from .template import Template 15 | 16 | logging.set_verbosity_error() 17 | 18 | 19 | class Fastc: 20 | def __new__( 21 | cls, 22 | model: str = None, 23 | embeddings_model: str = None, 24 | kernel: Kernels = None, 25 | model_type: Kernels = None, # Backwards compatibility 26 | template: str = None, 27 | pooling: Pooling = None, 28 | **kwargs, 29 | ): 30 | model_data = None 31 | label_names_by_id = None 32 | 33 | # Backwards compatibility 34 | if kernel is None and model_type is not None: 35 | kernel = model_type 36 | 37 | if model is not None: 38 | config = cls._get_config(model) 39 | model_config = config['model'] 40 | 41 | try: 42 | kernel = Kernels.from_value(model_config['kernel']) 43 | except KeyError: 44 | # Backwards compatibility 45 | kernel = Kernels.from_value(model_config['type']) 46 | 47 | model_data = model_config['data'] 48 | embeddings_model = model_config['embeddings'] 49 | 50 | labels = model_config.get('labels') 51 | if labels is None: 52 | # Backwards compatibility 53 | label_names_by_id = {label: label for label in model_data.keys()} # noqa: E501 54 | else: 55 | label_names_by_id = {v: k for k, v in labels.items()} 56 | 57 | pooling = Pooling.from_value(model_config.get( 58 | 'pooling', 59 | Pooling.MEAN.value, # Backwards compatibility 60 | )) 61 | 62 | if 'template' in model_config: 63 | template_text = model_config['template']['text'] 64 | template_variables = model_config['template']['variables'] 65 | template = Template(template_text, **template_variables) 66 | 67 | if embeddings_model is None: 68 | embeddings_model = 'deepset/tinyroberta-6l-768d' 69 | 70 | if kernel is None: 71 | kernel = Kernels.LOGISTIC_REGRESSION 72 | 73 | if template is None: 74 | template = Template() 75 | 76 | if pooling is None: 77 | pooling = Pooling.DEFAULT 78 | 79 | classifier_kwargs = { 80 | 'embeddings_model': embeddings_model, 81 | 'model_data': model_data, 82 | 'template': template, 83 | 'pooling': pooling, 84 | 'label_names_by_id': label_names_by_id, 85 | **kwargs, 86 | } 87 | 88 | if kernel == Kernels.LOGISTIC_REGRESSION: 89 | return LogisticRegressionClassifier(**classifier_kwargs) 90 | 91 | if kernel == Kernels.NEAREST_CENTROID: 92 | return NearestCentroidClassifier(**classifier_kwargs) 93 | 94 | # Backwards compatibility 95 | if kernel == Kernels.CENTROIDS: 96 | return NearestCentroidClassifier(**classifier_kwargs) 97 | 98 | raise ValueError("Unsupported model type {}".format(kernel)) 99 | 100 | @staticmethod 101 | def _get_config(model: str): 102 | if os.path.isdir(model): 103 | file_path = os.path.join(model, 'config.json') 104 | elif os.path.isfile(model): 105 | file_path = model 106 | else: 107 | file_path = hf_hub_download( 108 | repo_id=model, 109 | filename='config.json' 110 | ) 111 | with open(file_path, 'r') as model_file: 112 | model = json.load(model_file) 113 | return model 114 | 115 | 116 | # Backwards compatibility 117 | SentenceClassifier = Fastc 118 | -------------------------------------------------------------------------------- /fastc/kernels.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | from enum import Enum 5 | 6 | 7 | class Kernels(Enum): 8 | NEAREST_CENTROID = 'nearest-centroid' 9 | LOGISTIC_REGRESSION = 'logistic-regression' 10 | 11 | # Backwards compatibility 12 | CENTROIDS = 'centroids' 13 | 14 | @classmethod 15 | def from_value(cls, value: str) -> 'Kernels': 16 | try: 17 | return cls._value2member_map_[value] 18 | except KeyError: 19 | raise ValueError(f"{value} is not a valid model type value") 20 | -------------------------------------------------------------------------------- /fastc/pooling.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | from enum import Enum 5 | 6 | 7 | class Pooling(Enum): 8 | MEAN = 'mean' 9 | MEAN_MASKED = 'mean-masked' 10 | MAX = 'max' 11 | MAX_MASKED = 'max-masked' 12 | CLS = 'cls' 13 | SUM = 'sum' 14 | ATTENTION_WEIGHTED = 'attention-weighted' 15 | DEFAULT = MEAN 16 | 17 | @classmethod 18 | def from_value(cls, value: str) -> 'Pooling': 19 | try: 20 | return cls._value2member_map_[value] 21 | except KeyError: 22 | raise ValueError(f"{value} is not a valid pooling value") 23 | 24 | 25 | ATTENTION_POOLING_STRATEGIES = set([ 26 | Pooling.ATTENTION_WEIGHTED, 27 | ]) 28 | -------------------------------------------------------------------------------- /fastc/template.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | class ModelTemplates: 5 | E5_INSTRUCT = 'Instruct: {instruction}\nQuery: {text}' 6 | E5_QUERY = 'query: {text}' 7 | E5_PASSAGE = 'passage: {text}' 8 | DEFAULT = '{text}' 9 | 10 | 11 | class Template: 12 | def __init__( 13 | self, 14 | template: str = None, 15 | **kwargs, 16 | ): 17 | if template is None: 18 | template = ModelTemplates.DEFAULT 19 | self._template = template 20 | self._variables = kwargs 21 | 22 | def format(self, text: str) -> str: 23 | return self._template.format(text=text, **self._variables) 24 | -------------------------------------------------------------------------------- /install.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | cd $(dirname $0) 3 | pip install -e . 4 | -------------------------------------------------------------------------------- /misc/fastc.svg: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 | 63 | 64 | 65 | 66 | 67 | 68 | 69 | 70 | 71 | 72 | 73 | 74 | 75 | 76 | 77 | 78 | 79 | 80 | 81 | 82 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [build-system] 2 | requires = ["setuptools>=42", "wheel"] 3 | build-backend = "setuptools.build_meta" 4 | 5 | [project] 6 | name = "fastc" 7 | version = "2.2407.0" 8 | description = "Unattended Lightweight Text Classifiers with State-of-the-Art LLM Embeddings" 9 | authors = [ 10 | {name = "Rodrigo Martínez (brunneis)", email = "dev@brunneis.com"} 11 | ] 12 | license = {text = "GNU General Public License v3 (GPLv3)"} 13 | readme = "README.md" 14 | requires-python = ">=3.10" 15 | classifiers = [ 16 | "Development Status :: 4 - Beta", 17 | "Environment :: Console", 18 | "Intended Audience :: Developers", 19 | "License :: OSI Approved :: GNU General Public License v3 (GPLv3)", 20 | "Operating System :: POSIX :: Linux", 21 | "Programming Language :: Python :: 3.10", 22 | "Programming Language :: Python :: Implementation :: PyPy", 23 | "Topic :: Software Development :: Libraries :: Python Modules" 24 | ] 25 | dependencies = [ 26 | "transformers", 27 | "huggingface-hub", 28 | "torch", 29 | "tqdm", 30 | "scikit-learn", 31 | "joblib", 32 | ] 33 | 34 | [project.urls] 35 | Repository = "https://github.com/EveripediaNetwork/fastc" 36 | 37 | [project.optional-dependencies] 38 | dev = [ 39 | "build", 40 | "installer", 41 | "wheel", 42 | "twine", 43 | ] 44 | 45 | [tool.setuptools.packages.find] 46 | where = ["."] 47 | include = ["fastc*"] 48 | 49 | [tool.setuptools] 50 | package-data = {"fastc" = ["*"]} 51 | zip-safe = false 52 | -------------------------------------------------------------------------------- /requirements-dev.txt: -------------------------------------------------------------------------------- 1 | build 2 | installer 3 | wheel 4 | twine 5 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | transformers 2 | huggingface-hub 3 | torch 4 | tqdm 5 | scikit-learn 6 | joblib 7 | -------------------------------------------------------------------------------- /server/config.yaml.example: -------------------------------------------------------------------------------- 1 | --- 2 | download_on_demand: true 3 | cached_models: 4 | - 'braindao/tinyroberta-6l-768d-language-identifier-en-es-ko-zh-fastc' 5 | -------------------------------------------------------------------------------- /server/docker/Dockerfile: -------------------------------------------------------------------------------- 1 | FROM python:3.12-slim-bookworm 2 | WORKDIR /opt/app 3 | 4 | RUN \ 5 | apt-get update && \ 6 | apt-get -y upgrade && \ 7 | apt-get -y install libuv1 zlib1g && \ 8 | apt-get clean && \ 9 | pip install --upgrade pip 10 | 11 | COPY requirements.txt . 12 | RUN pip install -r requirements.txt 13 | 14 | COPY server/requirements.txt ./requirements-server.txt 15 | RUN pip install -r requirements-server.txt 16 | 17 | COPY . . 18 | RUN pip install . 19 | 20 | ENTRYPOINT [ "./server/scripts/start-server.sh" ] 21 | -------------------------------------------------------------------------------- /server/docker/docker-compose.yaml: -------------------------------------------------------------------------------- 1 | services: 2 | app: 3 | build: 4 | context: ../.. 5 | dockerfile: ./server/docker/Dockerfile 6 | volumes: 7 | - ../config.yaml:/opt/app/server/config.yaml 8 | - ${HOME}/.cache/huggingface/:/home/cortex/.cache/huggingface/ 9 | ports: 10 | - 53256:53256 11 | restart: always 12 | -------------------------------------------------------------------------------- /server/main.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | import falcon 5 | import yaml 6 | from resources.inference import InferenceResource 7 | from resources.version import VersionResource 8 | 9 | with open('./config.yaml', 'r') as stream: 10 | config = yaml.safe_load(stream) 11 | 12 | app = falcon.App( 13 | cors_enable=True, 14 | ) 15 | 16 | inference_resource = InferenceResource( 17 | model_names=config.get('cached_models', []), 18 | download_on_demand=config.get('download_on_demand', False) 19 | ) 20 | 21 | app.add_route('/', inference_resource) 22 | app.add_route('/version', VersionResource()) 23 | -------------------------------------------------------------------------------- /server/requirements.txt: -------------------------------------------------------------------------------- 1 | falcon 2 | socketify 3 | pyyaml 4 | -------------------------------------------------------------------------------- /server/resources/inference.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | 5 | from typing import List 6 | 7 | from fastc import Fastc 8 | 9 | 10 | class InferenceResource: 11 | def __init__( 12 | self, 13 | model_names: List[str] = None, 14 | download_on_demand: bool = False, 15 | ): 16 | self._classifiers = {} 17 | if model_names: 18 | for model_name in model_names: 19 | self._classifiers[model_name] = Fastc(model_name) 20 | self._download_on_demand = download_on_demand 21 | 22 | def on_post(self, request, response): 23 | payload = request.media 24 | model_name = payload.get('model') 25 | text = payload.get('text') 26 | 27 | if model_name not in self._classifiers: 28 | if self._download_on_demand: 29 | self._classifiers[model_name] = Fastc(model_name) 30 | else: 31 | response.status = 404 32 | response.media = { 33 | 'error': 'Model {} not found.'.format(model_name), 34 | } 35 | return 36 | 37 | result = self._classifiers[model_name].predict_one(text) 38 | response.media = result 39 | -------------------------------------------------------------------------------- /server/resources/version.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | 4 | from importlib import metadata 5 | 6 | import fastc 7 | 8 | 9 | class VersionResource: 10 | def on_get(self, _, response): 11 | response.media = { 12 | 'version': metadata.version(fastc.__name__), 13 | } 14 | -------------------------------------------------------------------------------- /server/scripts/start-docker.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | cd $(dirname $0)/../docker 3 | export COMPOSE_PROJECT_NAME="fastc-server" 4 | docker compose up -d --build --remove-orphans 5 | -------------------------------------------------------------------------------- /server/scripts/start-server.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | cd $(dirname $0)/.. 3 | 4 | if [ -z "$API_PROCESSES_COUNT" ]; then 5 | API_PROCESSES_COUNT=1 6 | fi 7 | 8 | python -m socketify main:app --host 0.0.0.0 --port 53256 --workers $API_PROCESSES_COUNT 9 | -------------------------------------------------------------------------------- /upload.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | set -e 3 | cd $(dirname $0) 4 | rm -rf dist/*.whl 5 | pip install -r requirements-dev.txt 6 | python -m build 7 | twine upload dist/*.whl 8 | --------------------------------------------------------------------------------