├── HOWTO.md ├── LICENSE ├── README.md ├── img ├── spectogram.png ├── val_acc.png └── val_f1.png ├── requirements.txt ├── research_seed ├── README.md ├── __init__.py └── audio_classification │ ├── __init__.py │ ├── anaconda3 │ └── bin │ │ └── wandb │ ├── cnn_lflb.py │ ├── cnn_rnn.py │ ├── cnn_sweep.py │ ├── cnn_trainer.py │ └── model_testing.py └── setup.py /HOWTO.md: -------------------------------------------------------------------------------- 1 | ### First use 2 | Clone the repo 3 | 4 | ```python 5 | git clone https://github.com/williamFalcon/pytorch-lightning-conference-seed 6 | ``` 7 | 8 | Install the package so you can access everything use package references 9 | ```python 10 | cd pytorch-lightning-conference-seed 11 | pip install -e . 12 | 13 | # now you can do: 14 | from research_seed import Whatever 15 | ``` 16 | 17 | ### Running LightningModules 18 | A [LightningModule](https://pytorch-lightning.readthedocs.io/en/latest/LightningModule/RequiredTrainerInterface/) has the core logic to your research code. This includes: 19 | - Dataloaders (train, test, val) 20 | - Optimizers 21 | - Training loop actions 22 | - Validation loop actions 23 | 24 | To run the module, feed it to the [Trainer](https://pytorch-lightning.readthedocs.io/en/latest/Trainer/) which handles mixed precision, checkpointing, multi-node, multi-GPU, etc... 25 | This makes your research contribution very clear and abstracted from the engineering details. 26 | 27 | To run the MNIST example in this package use the following command 28 | ```python 29 | # On CPU 30 | python lightning_modules/train.py 31 | 32 | # On multiple GPUs (4 gpus here) 33 | python lightning_modules/train.py --gpus '0,1,2,3' 34 | 35 | # On multiple nodes (16 gpus here) 36 | python lightning_modules/train.py --gpus '0,1,2,3' --nodes 4 37 | ``` 38 | 39 | ### How to use this seed for research 40 | For each project define a new LightningModule. 41 | 42 | ```python 43 | import pytorch_lightning as pl 44 | 45 | class CoolerNotBERT(pl.LightningModule): 46 | def __init__(self): 47 | self.net = ... 48 | 49 | def training_step(self, batch, batch_nb): 50 | # do some other cool task 51 | # return loss 52 | ``` 53 | 54 | If you have variations of the same project it makes sense to use the same Module 55 | ```python 56 | class BERT(pl.LightningModule): 57 | def __init__(self, model_name, task): 58 | self.task = task 59 | 60 | if model_name == 'transformer': 61 | self.net = Transformer() 62 | elif model_name == 'my_cool_version': 63 | self.net = MyCoolVersion() 64 | 65 | def training_step(self, batch, batch_nb): 66 | if self.task == 'standard_bert': 67 | # do standard bert training with self.net... 68 | # return loss 69 | 70 | if self.task == 'my_cool_task': 71 | # do my own version with self.net 72 | # return loss 73 | ``` 74 | 75 | Then decide which to run using the trainer. 76 | ```python 77 | if use_bert: 78 | model = BERT() 79 | else: 80 | model = CoolerNotBERT() 81 | 82 | trainer = Trainer(gpus=[0, 1, 2, 3], use_amp=True) 83 | trainer.fit(model) 84 | ``` 85 | 86 | ### Trainer 87 | It's recommended that you have a single trainer per lightning module. However, you can also use a single trainer for all your LightningModules. 88 | 89 | Check out the [MNIST example](https://github.com/williamFalcon/pytorch-lightning-conference-seed/tree/master/research_seed/mnist). 90 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU AFFERO GENERAL PUBLIC LICENSE 2 | Version 3, 19 November 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 Affero General Public License is a free, copyleft license for 11 | software and other kinds of works, specifically designed to ensure 12 | cooperation with the community in the case of network server software. 13 | 14 | The licenses for most software and other practical works are designed 15 | to take away your freedom to share and change the works. By contrast, 16 | our General Public Licenses are intended to guarantee your freedom to 17 | share and change all versions of a program--to make sure it remains free 18 | software for all its users. 19 | 20 | When we speak of free software, we are referring to freedom, not 21 | price. Our General Public Licenses are designed to make sure that you 22 | have the freedom to distribute copies of free software (and charge for 23 | them if you wish), that you receive source code or can get it if you 24 | want it, that you can change the software or use pieces of it in new 25 | free programs, and that you know you can do these things. 26 | 27 | Developers that use our General Public Licenses protect your rights 28 | with two steps: (1) assert copyright on the software, and (2) offer 29 | you this License which gives you legal permission to copy, distribute 30 | and/or modify the software. 31 | 32 | A secondary benefit of defending all users' freedom is that 33 | improvements made in alternate versions of the program, if they 34 | receive widespread use, become available for other developers to 35 | incorporate. Many developers of free software are heartened and 36 | encouraged by the resulting cooperation. However, in the case of 37 | software used on network servers, this result may fail to come about. 38 | The GNU General Public License permits making a modified version and 39 | letting the public access it on a server without ever releasing its 40 | source code to the public. 41 | 42 | The GNU Affero General Public License is designed specifically to 43 | ensure that, in such cases, the modified source code becomes available 44 | to the community. It requires the operator of a network server to 45 | provide the source code of the modified version running there to the 46 | users of that server. Therefore, public use of a modified version, on 47 | a publicly accessible server, gives the public access to the source 48 | code of the modified version. 49 | 50 | An older license, called the Affero General Public License and 51 | published by Affero, was designed to accomplish similar goals. This is 52 | a different license, not a version of the Affero GPL, but Affero has 53 | released a new version of the Affero GPL which permits relicensing under 54 | this license. 55 | 56 | The precise terms and conditions for copying, distribution and 57 | modification follow. 58 | 59 | TERMS AND CONDITIONS 60 | 61 | 0. Definitions. 62 | 63 | "This License" refers to version 3 of the GNU Affero General Public License. 64 | 65 | "Copyright" also means copyright-like laws that apply to other kinds of 66 | works, such as semiconductor masks. 67 | 68 | "The Program" refers to any copyrightable work licensed under this 69 | License. Each licensee is addressed as "you". "Licensees" and 70 | "recipients" may be individuals or organizations. 71 | 72 | To "modify" a work means to copy from or adapt all or part of the work 73 | in a fashion requiring copyright permission, other than the making of an 74 | exact copy. The resulting work is called a "modified version" of the 75 | earlier work or a work "based on" the earlier work. 76 | 77 | A "covered work" means either the unmodified Program or a work based 78 | on the Program. 79 | 80 | To "propagate" a work means to do anything with it that, without 81 | permission, would make you directly or secondarily liable for 82 | infringement under applicable copyright law, except executing it on a 83 | computer or modifying a private copy. Propagation includes copying, 84 | distribution (with or without modification), making available to the 85 | public, and in some countries other activities as well. 86 | 87 | To "convey" a work means any kind of propagation that enables other 88 | parties to make or receive copies. Mere interaction with a user through 89 | a computer network, with no transfer of a copy, is not conveying. 90 | 91 | An interactive user interface displays "Appropriate Legal Notices" 92 | to the extent that it includes a convenient and prominently visible 93 | feature that (1) displays an appropriate copyright notice, and (2) 94 | tells the user that there is no warranty for the work (except to the 95 | extent that warranties are provided), that licensees may convey the 96 | work under this License, and how to view a copy of this License. If 97 | the interface presents a list of user commands or options, such as a 98 | menu, a prominent item in the list meets this criterion. 99 | 100 | 1. Source Code. 101 | 102 | The "source code" for a work means the preferred form of the work 103 | for making modifications to it. "Object code" means any non-source 104 | form of a work. 105 | 106 | A "Standard Interface" means an interface that either is an official 107 | standard defined by a recognized standards body, or, in the case of 108 | interfaces specified for a particular programming language, one that 109 | is widely used among developers working in that language. 110 | 111 | The "System Libraries" of an executable work include anything, other 112 | than the work as a whole, that (a) is included in the normal form of 113 | packaging a Major Component, but which is not part of that Major 114 | Component, and (b) serves only to enable use of the work with that 115 | Major Component, or to implement a Standard Interface for which an 116 | implementation is available to the public in source code form. A 117 | "Major Component", in this context, means a major essential component 118 | (kernel, window system, and so on) of the specific operating system 119 | (if any) on which the executable work runs, or a compiler used to 120 | produce the work, or an object code interpreter used to run it. 121 | 122 | The "Corresponding Source" for a work in object code form means all 123 | the source code needed to generate, install, and (for an executable 124 | work) run the object code and to modify the work, including scripts to 125 | control those activities. However, it does not include the work's 126 | System Libraries, or general-purpose tools or generally available free 127 | programs which are used unmodified in performing those activities but 128 | which are not part of the work. For example, Corresponding Source 129 | includes interface definition files associated with source files for 130 | the work, and the source code for shared libraries and dynamically 131 | linked subprograms that the work is specifically designed to require, 132 | such as by intimate data communication or control flow between those 133 | subprograms and other parts of the work. 134 | 135 | The Corresponding Source need not include anything that users 136 | can regenerate automatically from other parts of the Corresponding 137 | Source. 138 | 139 | The Corresponding Source for a work in source code form is that 140 | same work. 141 | 142 | 2. Basic Permissions. 143 | 144 | All rights granted under this License are granted for the term of 145 | copyright on the Program, and are irrevocable provided the stated 146 | conditions are met. This License explicitly affirms your unlimited 147 | permission to run the unmodified Program. The output from running a 148 | covered work is covered by this License only if the output, given its 149 | content, constitutes a covered work. This License acknowledges your 150 | rights of fair use or other equivalent, as provided by copyright law. 151 | 152 | You may make, run and propagate covered works that you do not 153 | convey, without conditions so long as your license otherwise remains 154 | in force. You may convey covered works to others for the sole purpose 155 | of having them make modifications exclusively for you, or provide you 156 | with facilities for running those works, provided that you comply with 157 | the terms of this License in conveying all material for which you do 158 | not control copyright. Those thus making or running the covered works 159 | for you must do so exclusively on your behalf, under your direction 160 | and control, on terms that prohibit them from making any copies of 161 | your copyrighted material outside their relationship with you. 162 | 163 | Conveying under any other circumstances is permitted solely under 164 | the conditions stated below. Sublicensing is not allowed; section 10 165 | makes it unnecessary. 166 | 167 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 168 | 169 | No covered work shall be deemed part of an effective technological 170 | measure under any applicable law fulfilling obligations under article 171 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 172 | similar laws prohibiting or restricting circumvention of such 173 | measures. 174 | 175 | When you convey a covered work, you waive any legal power to forbid 176 | circumvention of technological measures to the extent such circumvention 177 | is effected by exercising rights under this License with respect to 178 | the covered work, and you disclaim any intention to limit operation or 179 | modification of the work as a means of enforcing, against the work's 180 | users, your or third parties' legal rights to forbid circumvention of 181 | technological measures. 182 | 183 | 4. Conveying Verbatim Copies. 184 | 185 | You may convey verbatim copies of the Program's source code as you 186 | receive it, in any medium, provided that you conspicuously and 187 | appropriately publish on each copy an appropriate copyright notice; 188 | keep intact all notices stating that this License and any 189 | non-permissive terms added in accord with section 7 apply to the code; 190 | keep intact all notices of the absence of any warranty; and give all 191 | recipients a copy of this License along with the Program. 192 | 193 | You may charge any price or no price for each copy that you convey, 194 | and you may offer support or warranty protection for a fee. 195 | 196 | 5. Conveying Modified Source Versions. 197 | 198 | You may convey a work based on the Program, or the modifications to 199 | produce it from the Program, in the form of source code under the 200 | terms of section 4, provided that you also meet all of these conditions: 201 | 202 | a) The work must carry prominent notices stating that you modified 203 | it, and giving a relevant date. 204 | 205 | b) The work must carry prominent notices stating that it is 206 | released under this License and any conditions added under section 207 | 7. This requirement modifies the requirement in section 4 to 208 | "keep intact all notices". 209 | 210 | c) You must license the entire work, as a whole, under this 211 | License to anyone who comes into possession of a copy. This 212 | License will therefore apply, along with any applicable section 7 213 | additional terms, to the whole of the work, and all its parts, 214 | regardless of how they are packaged. This License gives no 215 | permission to license the work in any other way, but it does not 216 | invalidate such permission if you have separately received it. 217 | 218 | d) If the work has interactive user interfaces, each must display 219 | Appropriate Legal Notices; however, if the Program has interactive 220 | interfaces that do not display Appropriate Legal Notices, your 221 | work need not make them do so. 222 | 223 | A compilation of a covered work with other separate and independent 224 | works, which are not by their nature extensions of the covered work, 225 | and which are not combined with it such as to form a larger program, 226 | in or on a volume of a storage or distribution medium, is called an 227 | "aggregate" if the compilation and its resulting copyright are not 228 | used to limit the access or legal rights of the compilation's users 229 | beyond what the individual works permit. Inclusion of a covered work 230 | in an aggregate does not cause this License to apply to the other 231 | parts of the aggregate. 232 | 233 | 6. Conveying Non-Source Forms. 234 | 235 | You may convey a covered work in object code form under the terms 236 | of sections 4 and 5, provided that you also convey the 237 | machine-readable Corresponding Source under the terms of this License, 238 | in one of these ways: 239 | 240 | a) Convey the object code in, or embodied in, a physical product 241 | (including a physical distribution medium), accompanied by the 242 | Corresponding Source fixed on a durable physical medium 243 | customarily used for software interchange. 244 | 245 | b) Convey the object code in, or embodied in, a physical product 246 | (including a physical distribution medium), accompanied by a 247 | written offer, valid for at least three years and valid for as 248 | long as you offer spare parts or customer support for that product 249 | model, to give anyone who possesses the object code either (1) a 250 | copy of the Corresponding Source for all the software in the 251 | product that is covered by this License, on a durable physical 252 | medium customarily used for software interchange, for a price no 253 | more than your reasonable cost of physically performing this 254 | conveying of source, or (2) access to copy the 255 | Corresponding Source from a network server at no charge. 256 | 257 | c) Convey individual copies of the object code with a copy of the 258 | written offer to provide the Corresponding Source. This 259 | alternative is allowed only occasionally and noncommercially, and 260 | only if you received the object code with such an offer, in accord 261 | with subsection 6b. 262 | 263 | d) Convey the object code by offering access from a designated 264 | place (gratis or for a charge), and offer equivalent access to the 265 | Corresponding Source in the same way through the same place at no 266 | further charge. You need not require recipients to copy the 267 | Corresponding Source along with the object code. If the place to 268 | copy the object code is a network server, the Corresponding Source 269 | may be on a different server (operated by you or a third party) 270 | that supports equivalent copying facilities, provided you maintain 271 | clear directions next to the object code saying where to find the 272 | Corresponding Source. Regardless of what server hosts the 273 | Corresponding Source, you remain obligated to ensure that it is 274 | available for as long as needed to satisfy these requirements. 275 | 276 | e) Convey the object code using peer-to-peer transmission, provided 277 | you inform other peers where the object code and Corresponding 278 | Source of the work are being offered to the general public at no 279 | charge under subsection 6d. 280 | 281 | A separable portion of the object code, whose source code is excluded 282 | from the Corresponding Source as a System Library, need not be 283 | included in conveying the object code work. 284 | 285 | A "User Product" is either (1) a "consumer product", which means any 286 | tangible personal property which is normally used for personal, family, 287 | or household purposes, or (2) anything designed or sold for incorporation 288 | into a dwelling. In determining whether a product is a consumer product, 289 | doubtful cases shall be resolved in favor of coverage. For a particular 290 | product received by a particular user, "normally used" refers to a 291 | typical or common use of that class of product, regardless of the status 292 | of the particular user or of the way in which the particular user 293 | actually uses, or expects or is expected to use, the product. A product 294 | is a consumer product regardless of whether the product has substantial 295 | commercial, industrial or non-consumer uses, unless such uses represent 296 | the only significant mode of use of the product. 297 | 298 | "Installation Information" for a User Product means any methods, 299 | procedures, authorization keys, or other information required to install 300 | and execute modified versions of a covered work in that User Product from 301 | a modified version of its Corresponding Source. The information must 302 | suffice to ensure that the continued functioning of the modified object 303 | code is in no case prevented or interfered with solely because 304 | modification has been made. 305 | 306 | If you convey an object code work under this section in, or with, or 307 | specifically for use in, a User Product, and the conveying occurs as 308 | part of a transaction in which the right of possession and use of the 309 | User Product is transferred to the recipient in perpetuity or for a 310 | fixed term (regardless of how the transaction is characterized), the 311 | Corresponding Source conveyed under this section must be accompanied 312 | by the Installation Information. But this requirement does not apply 313 | if neither you nor any third party retains the ability to install 314 | modified object code on the User Product (for example, the work has 315 | been installed in ROM). 316 | 317 | The requirement to provide Installation Information does not include a 318 | requirement to continue to provide support service, warranty, or updates 319 | for a work that has been modified or installed by the recipient, or for 320 | the User Product in which it has been modified or installed. Access to a 321 | network may be denied when the modification itself materially and 322 | adversely affects the operation of the network or violates the rules and 323 | protocols for communication across the network. 324 | 325 | Corresponding Source conveyed, and Installation Information provided, 326 | in accord with this section must be in a format that is publicly 327 | documented (and with an implementation available to the public in 328 | source code form), and must require no special password or key for 329 | unpacking, reading or copying. 330 | 331 | 7. Additional Terms. 332 | 333 | "Additional permissions" are terms that supplement the terms of this 334 | License by making exceptions from one or more of its conditions. 335 | Additional permissions that are applicable to the entire Program shall 336 | be treated as though they were included in this License, to the extent 337 | that they are valid under applicable law. If additional permissions 338 | apply only to part of the Program, that part may be used separately 339 | under those permissions, but the entire Program remains governed by 340 | this License without regard to the additional permissions. 341 | 342 | When you convey a copy of a covered work, you may at your option 343 | remove any additional permissions from that copy, or from any part of 344 | it. (Additional permissions may be written to require their own 345 | removal in certain cases when you modify the work.) You may place 346 | additional permissions on material, added by you to a covered work, 347 | for which you have or can give appropriate copyright permission. 348 | 349 | Notwithstanding any other provision of this License, for material you 350 | add to a covered work, you may (if authorized by the copyright holders of 351 | that material) supplement the terms of this License with terms: 352 | 353 | a) Disclaiming warranty or limiting liability differently from the 354 | terms of sections 15 and 16 of this License; or 355 | 356 | b) Requiring preservation of specified reasonable legal notices or 357 | author attributions in that material or in the Appropriate Legal 358 | Notices displayed by works containing it; or 359 | 360 | c) Prohibiting misrepresentation of the origin of that material, or 361 | requiring that modified versions of such material be marked in 362 | reasonable ways as different from the original version; or 363 | 364 | d) Limiting the use for publicity purposes of names of licensors or 365 | authors of the material; or 366 | 367 | e) Declining to grant rights under trademark law for use of some 368 | trade names, trademarks, or service marks; or 369 | 370 | f) Requiring indemnification of licensors and authors of that 371 | material by anyone who conveys the material (or modified versions of 372 | it) with contractual assumptions of liability to the recipient, for 373 | any liability that these contractual assumptions directly impose on 374 | those licensors and authors. 375 | 376 | All other non-permissive additional terms are considered "further 377 | restrictions" within the meaning of section 10. If the Program as you 378 | received it, or any part of it, contains a notice stating that it is 379 | governed by this License along with a term that is a further 380 | restriction, you may remove that term. If a license document contains 381 | a further restriction but permits relicensing or conveying under this 382 | License, you may add to a covered work material governed by the terms 383 | of that license document, provided that the further restriction does 384 | not survive such relicensing or conveying. 385 | 386 | If you add terms to a covered work in accord with this section, you 387 | must place, in the relevant source files, a statement of the 388 | additional terms that apply to those files, or a notice indicating 389 | where to find the applicable terms. 390 | 391 | Additional terms, permissive or non-permissive, may be stated in the 392 | form of a separately written license, or stated as exceptions; 393 | the above requirements apply either way. 394 | 395 | 8. Termination. 396 | 397 | You may not propagate or modify a covered work except as expressly 398 | provided under this License. Any attempt otherwise to propagate or 399 | modify it is void, and will automatically terminate your rights under 400 | this License (including any patent licenses granted under the third 401 | paragraph of section 11). 402 | 403 | However, if you cease all violation of this License, then your 404 | license from a particular copyright holder is reinstated (a) 405 | provisionally, unless and until the copyright holder explicitly and 406 | finally terminates your license, and (b) permanently, if the copyright 407 | holder fails to notify you of the violation by some reasonable means 408 | prior to 60 days after the cessation. 409 | 410 | Moreover, your license from a particular copyright holder is 411 | reinstated permanently if the copyright holder notifies you of the 412 | violation by some reasonable means, this is the first time you have 413 | received notice of violation of this License (for any work) from that 414 | copyright holder, and you cure the violation prior to 30 days after 415 | your receipt of the notice. 416 | 417 | Termination of your rights under this section does not terminate the 418 | licenses of parties who have received copies or rights from you under 419 | this License. If your rights have been terminated and not permanently 420 | reinstated, you do not qualify to receive new licenses for the same 421 | material under section 10. 422 | 423 | 9. Acceptance Not Required for Having Copies. 424 | 425 | You are not required to accept this License in order to receive or 426 | run a copy of the Program. Ancillary propagation of a covered work 427 | occurring solely as a consequence of using peer-to-peer transmission 428 | to receive a copy likewise does not require acceptance. However, 429 | nothing other than this License grants you permission to propagate or 430 | modify any covered work. These actions infringe copyright if you do 431 | not accept this License. Therefore, by modifying or propagating a 432 | covered work, you indicate your acceptance of this License to do so. 433 | 434 | 10. Automatic Licensing of Downstream Recipients. 435 | 436 | Each time you convey a covered work, the recipient automatically 437 | receives a license from the original licensors, to run, modify and 438 | propagate that work, subject to this License. You are not responsible 439 | for enforcing compliance by third parties with this License. 440 | 441 | An "entity transaction" is a transaction transferring control of an 442 | organization, or substantially all assets of one, or subdividing an 443 | organization, or merging organizations. If propagation of a covered 444 | work results from an entity transaction, each party to that 445 | transaction who receives a copy of the work also receives whatever 446 | licenses to the work the party's predecessor in interest had or could 447 | give under the previous paragraph, plus a right to possession of the 448 | Corresponding Source of the work from the predecessor in interest, if 449 | the predecessor has it or can get it with reasonable efforts. 450 | 451 | You may not impose any further restrictions on the exercise of the 452 | rights granted or affirmed under this License. For example, you may 453 | not impose a license fee, royalty, or other charge for exercise of 454 | rights granted under this License, and you may not initiate litigation 455 | (including a cross-claim or counterclaim in a lawsuit) alleging that 456 | any patent claim is infringed by making, using, selling, offering for 457 | sale, or importing the Program or any portion of it. 458 | 459 | 11. Patents. 460 | 461 | A "contributor" is a copyright holder who authorizes use under this 462 | License of the Program or a work on which the Program is based. The 463 | work thus licensed is called the contributor's "contributor version". 464 | 465 | A contributor's "essential patent claims" are all patent claims 466 | owned or controlled by the contributor, whether already acquired or 467 | hereafter acquired, that would be infringed by some manner, permitted 468 | by this License, of making, using, or selling its contributor version, 469 | but do not include claims that would be infringed only as a 470 | consequence of further modification of the contributor version. For 471 | purposes of this definition, "control" includes the right to grant 472 | patent sublicenses in a manner consistent with the requirements of 473 | this License. 474 | 475 | Each contributor grants you a non-exclusive, worldwide, royalty-free 476 | patent license under the contributor's essential patent claims, to 477 | make, use, sell, offer for sale, import and otherwise run, modify and 478 | propagate the contents of its contributor version. 479 | 480 | In the following three paragraphs, a "patent license" is any express 481 | agreement or commitment, however denominated, not to enforce a patent 482 | (such as an express permission to practice a patent or covenant not to 483 | sue for patent infringement). To "grant" such a patent license to a 484 | party means to make such an agreement or commitment not to enforce a 485 | patent against the party. 486 | 487 | If you convey a covered work, knowingly relying on a patent license, 488 | and the Corresponding Source of the work is not available for anyone 489 | to copy, free of charge and under the terms of this License, through a 490 | publicly available network server or other readily accessible means, 491 | then you must either (1) cause the Corresponding Source to be so 492 | available, or (2) arrange to deprive yourself of the benefit of the 493 | patent license for this particular work, or (3) arrange, in a manner 494 | consistent with the requirements of this License, to extend the patent 495 | license to downstream recipients. "Knowingly relying" means you have 496 | actual knowledge that, but for the patent license, your conveying the 497 | covered work in a country, or your recipient's use of the covered work 498 | in a country, would infringe one or more identifiable patents in that 499 | country that you have reason to believe are valid. 500 | 501 | If, pursuant to or in connection with a single transaction or 502 | arrangement, you convey, or propagate by procuring conveyance of, a 503 | covered work, and grant a patent license to some of the parties 504 | receiving the covered work authorizing them to use, propagate, modify 505 | or convey a specific copy of the covered work, then the patent license 506 | you grant is automatically extended to all recipients of the covered 507 | work and works based on it. 508 | 509 | A patent license is "discriminatory" if it does not include within 510 | the scope of its coverage, prohibits the exercise of, or is 511 | conditioned on the non-exercise of one or more of the rights that are 512 | specifically granted under this License. You may not convey a covered 513 | work if you are a party to an arrangement with a third party that is 514 | in the business of distributing software, under which you make payment 515 | to the third party based on the extent of your activity of conveying 516 | the work, and under which the third party grants, to any of the 517 | parties who would receive the covered work from you, a discriminatory 518 | patent license (a) in connection with copies of the covered work 519 | conveyed by you (or copies made from those copies), or (b) primarily 520 | for and in connection with specific products or compilations that 521 | contain the covered work, unless you entered into that arrangement, 522 | or that patent license was granted, prior to 28 March 2007. 523 | 524 | Nothing in this License shall be construed as excluding or limiting 525 | any implied license or other defenses to infringement that may 526 | otherwise be available to you under applicable patent law. 527 | 528 | 12. No Surrender of Others' Freedom. 529 | 530 | If conditions are imposed on you (whether by court order, agreement or 531 | otherwise) that contradict the conditions of this License, they do not 532 | excuse you from the conditions of this License. If you cannot convey a 533 | covered work so as to satisfy simultaneously your obligations under this 534 | License and any other pertinent obligations, then as a consequence you may 535 | not convey it at all. For example, if you agree to terms that obligate you 536 | to collect a royalty for further conveying from those to whom you convey 537 | the Program, the only way you could satisfy both those terms and this 538 | License would be to refrain entirely from conveying the Program. 539 | 540 | 13. Remote Network Interaction; Use with the GNU General Public License. 541 | 542 | Notwithstanding any other provision of this License, if you modify the 543 | Program, your modified version must prominently offer all users 544 | interacting with it remotely through a computer network (if your version 545 | supports such interaction) an opportunity to receive the Corresponding 546 | Source of your version by providing access to the Corresponding Source 547 | from a network server at no charge, through some standard or customary 548 | means of facilitating copying of software. This Corresponding Source 549 | shall include the Corresponding Source for any work covered by version 3 550 | of the GNU General Public License that is incorporated pursuant to the 551 | following paragraph. 552 | 553 | Notwithstanding any other provision of this License, you have 554 | permission to link or combine any covered work with a work licensed 555 | under version 3 of the GNU General Public License into a single 556 | combined work, and to convey the resulting work. The terms of this 557 | License will continue to apply to the part which is the covered work, 558 | but the work with which it is combined will remain governed by version 559 | 3 of the GNU General Public License. 560 | 561 | 14. Revised Versions of this License. 562 | 563 | The Free Software Foundation may publish revised and/or new versions of 564 | the GNU Affero General Public License from time to time. Such new versions 565 | will be similar in spirit to the present version, but may differ in detail to 566 | address new problems or concerns. 567 | 568 | Each version is given a distinguishing version number. If the 569 | Program specifies that a certain numbered version of the GNU Affero General 570 | Public License "or any later version" applies to it, you have the 571 | option of following the terms and conditions either of that numbered 572 | version or of any later version published by the Free Software 573 | Foundation. If the Program does not specify a version number of the 574 | GNU Affero General Public License, you may choose any version ever published 575 | by the Free Software Foundation. 576 | 577 | If the Program specifies that a proxy can decide which future 578 | versions of the GNU Affero General Public License can be used, that proxy's 579 | public statement of acceptance of a version permanently authorizes you 580 | to choose that version for the Program. 581 | 582 | Later license versions may give you additional or different 583 | permissions. However, no additional obligations are imposed on any 584 | author or copyright holder as a result of your choosing to follow a 585 | later version. 586 | 587 | 15. Disclaimer of Warranty. 588 | 589 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 590 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 591 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 592 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 593 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 594 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 595 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 596 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 597 | 598 | 16. Limitation of Liability. 599 | 600 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 601 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 602 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 603 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 604 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 605 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 606 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 607 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 608 | SUCH DAMAGES. 609 | 610 | 17. Interpretation of Sections 15 and 16. 611 | 612 | If the disclaimer of warranty and limitation of liability provided 613 | above cannot be given local legal effect according to their terms, 614 | reviewing courts shall apply local law that most closely approximates 615 | an absolute waiver of all civil liability in connection with the 616 | Program, unless a warranty or assumption of liability accompanies a 617 | copy of the Program in return for a fee. 618 | 619 | END OF TERMS AND CONDITIONS 620 | 621 | How to Apply These Terms to Your New Programs 622 | 623 | If you develop a new program, and you want it to be of the greatest 624 | possible use to the public, the best way to achieve this is to make it 625 | free software which everyone can redistribute and change under these terms. 626 | 627 | To do so, attach the following notices to the program. It is safest 628 | to attach them to the start of each source file to most effectively 629 | state the exclusion of warranty; and each file should have at least 630 | the "copyright" line and a pointer to where the full notice is found. 631 | 632 | 633 | Copyright (C) 634 | 635 | This program is free software: you can redistribute it and/or modify 636 | it under the terms of the GNU Affero General Public License as published 637 | by the Free Software Foundation, either version 3 of the License, or 638 | (at your option) any later version. 639 | 640 | This program is distributed in the hope that it will be useful, 641 | but WITHOUT ANY WARRANTY; without even the implied warranty of 642 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 643 | GNU Affero General Public License for more details. 644 | 645 | You should have received a copy of the GNU Affero General Public License 646 | along with this program. If not, see . 647 | 648 | Also add information on how to contact you by electronic and paper mail. 649 | 650 | If your software can interact with users remotely through a computer 651 | network, you should also make sure that it provides a way for users to 652 | get its source. For example, if your program is a web application, its 653 | interface could display a "Source" link that leads users to an archive 654 | of the code. There are many ways you could offer source, and different 655 | solutions will be better for different programs; see section 13 for the 656 | specific requirements. 657 | 658 | You should also get your employer (if you work as a programmer) or school, 659 | if any, to sign a "copyright disclaimer" for the program, if necessary. 660 | For more information on this, and how to apply and follow the GNU AGPL, see 661 | . 662 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | 2 |
3 | 4 | # Speech2dCNN_LSTM 5 | 6 | 7 | 8 | 11 |
12 | 13 | ## Description 14 | A pytorch implementation of [Speech emotion recognition using deep 1D & 2D CNN LSTM networks](https://www.sciencedirect.com/science/article/abs/pii/S1746809418302337) using pytorch lighting and wandb sweep for hyperparameter finding. I'm not affiliated with the authors of the paper. 15 | 16 | ![Example of spectogram image used as input](/img/spectogram.png) 17 | ## How to run 18 | First, install dependencies 19 | ```bash 20 | # clone project 21 | git clone https://github.com/RicardoP0/Speech2dCNN_LSTM.git 22 | 23 | # install project 24 | cd Speech2dCNN_LSTM 25 | pip install -e . 26 | pip install requirements.txt 27 | ``` 28 | Next, navigate to [CNN+LSTM](https://github.com/RicardoP0/Speech2dCNN_LSTM/tree/master/research_seed/audio_classification) and run it. 29 | ```bash 30 | # module folder 31 | cd research_seed/audio_classification/ 32 | 33 | # run module 34 | python cnn_trainer.py 35 | ``` 36 | 37 | ## Main Contribution 38 | 39 | - [CNN+LSTM](https://github.com/RicardoP0/Speech2dCNN_LSTM/tree/master/research_seed/audio_classification) 40 | 41 | ## Results 42 | 43 | Validation accuracy reaches 0.4 and a F1 value of 0.3 using 8 classes. 44 | ![Validation accuracy on 8 classes](/img/val_acc.png) 45 | ![F1 on 8 classes](/img/val_f1.png) 46 | -------------------------------------------------------------------------------- /img/spectogram.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RicardoP0/Speech2dCNN_LSTM/78b9cb6a483c0acbce08cda6bf769d9015716ba8/img/spectogram.png -------------------------------------------------------------------------------- /img/val_acc.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RicardoP0/Speech2dCNN_LSTM/78b9cb6a483c0acbce08cda6bf769d9015716ba8/img/val_acc.png -------------------------------------------------------------------------------- /img/val_f1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RicardoP0/Speech2dCNN_LSTM/78b9cb6a483c0acbce08cda6bf769d9015716ba8/img/val_f1.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | adabound==0.0.5 2 | torchaudio==0.4.0a0+719bcc7 3 | librosa==0.7.1 4 | pytorch_lightning==0.7.1 5 | scikit_image==0.16.2 6 | matplotlib==3.1.3 7 | torchvision==0.5.0 8 | numpy==1.22.0 9 | wandb==0.8.32 10 | pandas==1.0.3 11 | torch==1.4.0 12 | Pillow==9.3.0 13 | scikit_learn==0.22.2.post1 14 | skimage==0.0 15 | torch==1.4.0 16 | 17 | -------------------------------------------------------------------------------- /research_seed/README.md: -------------------------------------------------------------------------------- 1 | ## Research Seed Folder 2 | 3 | 4 | ##### cnn_trainer.py 5 | Runs your LightningModule. Abstracts training loop, distributed training, etc... 6 | 7 | ##### cnn_sweep.py 8 | Code for hyperparameter finding using wandb's sweep. 9 | 10 | ##### cnn_lflb.py and cnn_rnn.py 11 | Code for cnn and cnn + lstm models 12 | -------------------------------------------------------------------------------- /research_seed/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RicardoP0/Speech2dCNN_LSTM/78b9cb6a483c0acbce08cda6bf769d9015716ba8/research_seed/__init__.py -------------------------------------------------------------------------------- /research_seed/audio_classification/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RicardoP0/Speech2dCNN_LSTM/78b9cb6a483c0acbce08cda6bf769d9015716ba8/research_seed/audio_classification/__init__.py -------------------------------------------------------------------------------- /research_seed/audio_classification/anaconda3/bin/wandb: -------------------------------------------------------------------------------- 1 | #!/home/ricardo/anaconda3/bin/python 2 | # -*- coding: utf-8 -*- 3 | import re 4 | import sys 5 | from wandb.cli import cli 6 | if __name__ == '__main__': 7 | sys.argv[0] = re.sub(r'(-script\.pyw|\.exe)?$', '', sys.argv[0]) 8 | sys.exit(cli()) 9 | -------------------------------------------------------------------------------- /research_seed/audio_classification/cnn_lflb.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file defines the core research contribution 3 | """ 4 | # %% 5 | import os 6 | import torch 7 | import torch.nn.functional as F 8 | import torch.nn as nn 9 | from torch.utils.data import DataLoader 10 | from torchvision.datasets import MNIST 11 | import torchvision.transforms as transforms 12 | from argparse import ArgumentParser 13 | from collections import OrderedDict 14 | from research_seed.audio_classification.datasets.iemocap_spect import IEMOCAPSpectDataset 15 | import pytorch_lightning as pl 16 | import numpy as np 17 | import sklearn.metrics as metrics 18 | from adabound import AdaBound 19 | 20 | import torchvision.models as models 21 | 22 | 23 | class LFLBlock(nn.Module): 24 | def __init__(self, inp_ch, out_ch, conv_k, conv_s, pool_k, pool_s, p_dropout): 25 | super(LFLBlock, self).__init__() 26 | 27 | self.conv = nn.Conv2d(inp_ch, out_ch, conv_k, conv_s, padding=(1, 2)) 28 | self.batch_nm = nn.BatchNorm2d(out_ch) 29 | self.pool = nn.MaxPool2d(pool_k, pool_s) 30 | 31 | self.dropout = nn.Dropout2d(p=p_dropout) # AlphaDropout 32 | self.actv = nn.ELU() 33 | 34 | def forward(self, x): 35 | 36 | x = self.conv(x) 37 | 38 | x = self.actv(x) 39 | x = self.pool(x) 40 | x = self.dropout(x) 41 | x = self.batch_nm(x) 42 | 43 | return x 44 | 45 | 46 | class CNN_LFLB(pl.LightningModule): 47 | 48 | def __init__(self, hparams): 49 | super(CNN_LFLB, self).__init__() 50 | # not the best model... 51 | 52 | self.hparams = hparams 53 | self.num_classes = hparams.num_classes 54 | 55 | self.lflb1 = LFLBlock(inp_ch=1, out_ch=64, conv_k=3, 56 | conv_s=1, pool_k=2, pool_s=2, p_dropout=self.hparams.dropout_1) 57 | self.lflb2 = LFLBlock(inp_ch=64, out_ch=64, conv_k=3, 58 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_2) 59 | self.lflb3 = LFLBlock(inp_ch=64, out_ch=128, conv_k=3, 60 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_3) 61 | self.lflb4 = LFLBlock(inp_ch=128, out_ch=128, conv_k=3, 62 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_3) 63 | 64 | self.fc1 = nn.Linear(256, 64) 65 | self.fc2 = nn.Linear(64, self.num_classes) 66 | 67 | def forward(self, x): 68 | x = self.lflb1(x) 69 | x = self.lflb2(x) 70 | x = self.lflb3(x) 71 | x = self.lflb4(x) 72 | 73 | x = x.view(x.shape[0], -1) 74 | x = F.relu(self.fc1(x)) 75 | x = self.fc2(x) 76 | 77 | return x 78 | 79 | def training_step(self, batch, batch_idx): 80 | # REQUIRED 81 | x, y = batch 82 | y_hat = self.forward(x) 83 | loss_val = F.cross_entropy(y_hat, y) 84 | with torch.no_grad(): 85 | y_pred = torch.max(F.softmax(y_hat, dim=1), 1)[1] 86 | acc = metrics.accuracy_score(y.cpu(), y_pred.cpu()) 87 | tqdm_dict = {'train_loss': loss_val, 'train_acc': acc} 88 | output = OrderedDict({ 89 | 'loss': loss_val, 90 | 'progress_bar': tqdm_dict, 91 | 'log': tqdm_dict 92 | }) 93 | return output 94 | 95 | def validation_step(self, batch, batch_idx): 96 | # OPTIONAL 97 | x, y = batch 98 | 99 | with torch.no_grad(): 100 | y_hat = self.forward(x) 101 | y_pred = torch.max(F.softmax(y_hat, dim=1), 1)[1] 102 | 103 | acc = metrics.accuracy_score(y.cpu(), y_pred.cpu()) 104 | f1 = metrics.f1_score(y.cpu(), y_pred.cpu(), average='macro') 105 | loss_val = F.cross_entropy(y_hat, y) 106 | 107 | output = OrderedDict({'val_loss': loss_val, 'val_f1': f1, 'val_acc': acc}) 108 | 109 | return output 110 | 111 | def validation_end(self, outputs): 112 | # OPTIONAL 113 | tqdm_dict = {} 114 | 115 | for metric_name in ["val_loss", "val_f1", "val_acc"]: 116 | metric_total = 0 117 | 118 | for output in outputs: 119 | metric_value = output[metric_name] 120 | 121 | # reduce manually when using dp 122 | if self.trainer.use_dp or self.trainer.use_ddp2: 123 | metric_value = torch.mean(metric_value) 124 | 125 | metric_total += metric_value 126 | 127 | tqdm_dict[metric_name] = metric_total / len(outputs) 128 | 129 | result = {'progress_bar': tqdm_dict, 'log': tqdm_dict,'val_loss': tqdm_dict["val_loss"]} 130 | 131 | return result 132 | 133 | def configure_optimizers(self): 134 | # REQUIRED 135 | # can return multiple optimizers and learning_rate schedulers 136 | 137 | # ,weight_decay=0.01)#torch.optim.Adam(self.parameters(), lr=self.hparams.learning_rate) 138 | return AdaBound(self.parameters(), lr=self.hparams.learning_rate_init, 139 | final_lr=self.hparams.learning_rate_final,weight_decay=self.hparams.weight_decay) 140 | 141 | def train_dataloader(self): 142 | # REQUIRED 143 | transform = transforms.Compose([transforms.ToTensor()]) 144 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='train', transform=transform, num_classes=self.num_classes), 145 | batch_size=32, num_workers=2, pin_memory=True, 146 | shuffle=True) 147 | 148 | def val_dataloader(self): 149 | # OPTIONAL 150 | 151 | transform = transforms.Compose([transforms.ToTensor()]) 152 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='val', transform=transform, num_classes=self.num_classes), 153 | batch_size=32, num_workers=2, pin_memory=True, 154 | shuffle=True) 155 | 156 | def test_dataloader(self): 157 | # OPTIONAL 158 | transform = transforms.Compose([transforms.ToTensor()]) 159 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='test', transform=transform, num_classes=self.num_classes), 160 | batch_size=32, num_workers=2, pin_memory=True, 161 | shuffle=True) 162 | 163 | @staticmethod 164 | def add_model_specific_args(parent_parser): 165 | """ 166 | Specify the hyperparams for this LightningModule 167 | """ 168 | # MODEL specific 169 | parser = ArgumentParser(parents=[parent_parser]) 170 | parser.add_argument('--learning_rate_init', default=7e-4, type=float) 171 | parser.add_argument('--learning_rate_final', default=0.02, type=float) 172 | parser.add_argument('--batch_size', default=32, type=int) 173 | parser.add_argument('--dropout_1', default=0.6, type=float) 174 | parser.add_argument('--dropout_2', default=0.3, type=float) 175 | parser.add_argument('--dropout_3', default=0.2, type=float) 176 | parser.add_argument('--weight_decay', default=0.0, type=float) 177 | 178 | 179 | # training specific (for this model) 180 | parser.add_argument('--max_nb_epochs', default=10000, type=int) 181 | 182 | # data 183 | parser.add_argument( 184 | '--data_root', default='../datasets/RAVDESS/SOUND_SPECT/', type=str) 185 | parser.add_argument( 186 | '--num_classes', dest='num_classes', default=8, type=int) 187 | return parser 188 | -------------------------------------------------------------------------------- /research_seed/audio_classification/cnn_rnn.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file defines the core research contribution 3 | """ 4 | # %% 5 | import os 6 | import torch 7 | import torch.nn.functional as F 8 | import torch.nn as nn 9 | from torch.utils.data import DataLoader 10 | from torchvision.datasets import MNIST 11 | import torchvision.transforms as transforms 12 | from argparse import ArgumentParser 13 | from collections import OrderedDict 14 | from research_seed.audio_classification.datasets.iemocap_spect import IEMOCAPSpectDataset 15 | import pytorch_lightning as pl 16 | import numpy as np 17 | import sklearn.metrics as metrics 18 | from adabound import AdaBound 19 | 20 | import torchvision.models as models 21 | 22 | 23 | class LSTMBlock(nn.Module): 24 | def __init__(self, input_size=300, hidden_size=256, num_layers=2, bidirectional=True, dropout=0.0, num_classes=8): 25 | super(LSTMBlock, self).__init__() 26 | 27 | self.input_size = input_size 28 | self.num_layers = num_layers # RNN hidden layers 29 | self.hidden_size = hidden_size # RNN hidden nodes 30 | self.num_classes = num_classes 31 | if bidirectional: 32 | self.num_directions = 2 33 | else: 34 | self.num_directions = 1 35 | 36 | self.LSTM = nn.LSTM( 37 | input_size=self.input_size, 38 | hidden_size=self.hidden_size, 39 | num_layers=self.num_layers, 40 | bidirectional=bidirectional, 41 | dropout=dropout, 42 | # input & output has batch size as 1s dimension. e.g. (batch, time_step, input_size) 43 | batch_first=True, 44 | ) 45 | 46 | self.fc1 = nn.Linear(self.hidden_size * self.num_directions, self.num_classes) 47 | 48 | def forward(self, x): 49 | 50 | self.LSTM.flatten_parameters() 51 | # print(x.shape) 52 | 53 | RNN_out, (h_n, h_c) = self.LSTM(x, None) 54 | # out" will give you access to all hidden states in the sequence 55 | """ h_n shape ((num_layers * num_directions, batch, hidden_size)), h_c shape (n_layers, batch, hidden_size) """ 56 | """ None represents zero initial hidden state. RNN_out has shape=(batch, time_step, output_size) """ 57 | 58 | x = self.fc1(RNN_out[:,-1,:]) # choose RNN_out at the last time step and activations in both directions 59 | 60 | return x 61 | 62 | 63 | class LFLBlock(nn.Module): 64 | def __init__(self, inp_ch, out_ch, conv_k, conv_s, pool_k, pool_s, p_dropout): 65 | super(LFLBlock, self).__init__() 66 | 67 | self.conv = nn.Conv2d(inp_ch, out_ch, conv_k, conv_s, padding=(1, 2)) 68 | self.batch_nm = nn.BatchNorm2d(out_ch) 69 | self.pool = nn.MaxPool2d(pool_k, pool_s) 70 | self.dropout = nn.Dropout2d(p=p_dropout) # AlphaDropout 71 | self.actv = nn.ELU() 72 | 73 | def forward(self, x): 74 | 75 | x = self.conv(x) 76 | 77 | x = self.actv(x) 78 | x = self.pool(x) 79 | x = self.dropout(x) 80 | x = self.batch_nm(x) 81 | 82 | return x 83 | 84 | 85 | class CNN_RNN(pl.LightningModule): 86 | 87 | def __init__(self, hparams): 88 | super(CNN_RNN, self).__init__() 89 | 90 | self.hparams = hparams 91 | self.num_classes = hparams.num_classes 92 | self.bidirectional = bool(hparams.bidirectional) 93 | self.num_layers_rnn = hparams.num_layers_rnn 94 | self.dropout_rnn = hparams.dropout_rnn 95 | self.num_layers_rnn = hparams.num_layers_rnn # RNN hidden layers 96 | self.hidden_size_rnn = hparams.hidden_size_rnn # RNN hidden nodes 97 | 98 | self.lflb1 = LFLBlock(inp_ch=1, out_ch=64, conv_k=3, 99 | conv_s=1, pool_k=2, pool_s=2, p_dropout=self.hparams.dropout_1) 100 | self.lflb2 = LFLBlock(inp_ch=64, out_ch=64, conv_k=3, 101 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_2) 102 | self.lflb3 = LFLBlock(inp_ch=64, out_ch=128, conv_k=3, 103 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_3) 104 | self.lflb4 = LFLBlock(inp_ch=128, out_ch=128, conv_k=3, 105 | conv_s=1, pool_k=4, pool_s=4, p_dropout=self.hparams.dropout_3) 106 | 107 | self.rnn = LSTMBlock(input_size=128, hidden_size=self.hidden_size_rnn,dropout=self.dropout_rnn, num_classes=self.num_classes, 108 | bidirectional=self.bidirectional, num_layers=self.num_layers_rnn, 109 | ) 110 | 111 | 112 | def forward(self, x): 113 | x = self.lflb1(x) 114 | x = self.lflb2(x) 115 | x = self.lflb3(x) 116 | x = self.lflb4(x) 117 | 118 | x = x.permute(0, 3, 1, 2) 119 | x = x.view(x.shape[0], x.shape[1], -1) 120 | 121 | x = self.rnn(x) 122 | 123 | return x 124 | 125 | def training_step(self, batch, batch_idx): 126 | # REQUIRED 127 | x, y = batch 128 | y_hat = self.forward(x) 129 | loss_val = F.cross_entropy(y_hat, y) 130 | with torch.no_grad(): 131 | y_pred = torch.max(F.softmax(y_hat, dim=1), 1)[1] 132 | acc = metrics.accuracy_score(y.cpu(), y_pred.cpu()) 133 | tqdm_dict = {'train_loss': loss_val, 'train_acc': acc} 134 | 135 | output = OrderedDict({ 136 | 'loss': loss_val, 137 | 'progress_bar': tqdm_dict, 138 | 'log': tqdm_dict 139 | }) 140 | return output 141 | 142 | def accuracy(self, y_true, y_pred): 143 | with torch.no_grad(): 144 | acc = (y_true == y_pred).sum().to(torch.float32) 145 | acc /= y_pred.shape[0] 146 | 147 | return acc 148 | 149 | def validation_step(self, batch, batch_idx): 150 | # OPTIONAL 151 | x, y = batch 152 | 153 | y_hat = self.forward(x) 154 | 155 | with torch.no_grad(): 156 | y_pred = torch.max(F.softmax(y_hat, dim=1), 1)[1] 157 | acc = metrics.accuracy_score(y.cpu(), y_pred.cpu()) 158 | f1 = metrics.f1_score(y.cpu(), y_pred.cpu(), average='macro') 159 | loss_val = F.cross_entropy(y_hat, y) 160 | 161 | output = OrderedDict( 162 | {'val_loss': loss_val, 'val_f1': f1, 'val_acc': acc}) 163 | 164 | return output 165 | 166 | def validation_end(self, outputs): 167 | # OPTIONAL 168 | tqdm_dict = {} 169 | 170 | for metric_name in ["val_loss", "val_f1", "val_acc"]: 171 | metric_total = 0 172 | 173 | for output in outputs: 174 | metric_value = output[metric_name] 175 | 176 | # reduce manually when using dp 177 | if self.trainer.use_dp or self.trainer.use_ddp2: 178 | metric_value = torch.mean(metric_value) 179 | 180 | metric_total += metric_value 181 | 182 | tqdm_dict[metric_name] = metric_total / len(outputs) 183 | 184 | result = {'progress_bar': tqdm_dict, 'log': tqdm_dict, 185 | 'val_loss': tqdm_dict["val_loss"]} 186 | 187 | return result 188 | 189 | def configure_optimizers(self): 190 | # REQUIRED 191 | # can return multiple optimizers and learning_rate schedulers 192 | 193 | return AdaBound(self.parameters(), lr=self.hparams.learning_rate_init, 194 | final_lr=self.hparams.learning_rate_final, weight_decay=self.hparams.weight_decay) 195 | 196 | def train_dataloader(self): 197 | # REQUIRED 198 | transform = transforms.Compose([transforms.ToTensor()]) 199 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='train', transform=transform, num_classes=self.num_classes), 200 | batch_size=32, num_workers=2, pin_memory=True, 201 | shuffle=True) 202 | 203 | def val_dataloader(self): 204 | # OPTIONAL 205 | 206 | transform = transforms.Compose([transforms.ToTensor()]) 207 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='val', transform=transform, num_classes=self.num_classes), 208 | batch_size=32, num_workers=2, pin_memory=True, 209 | shuffle=True) 210 | 211 | def test_dataloader(self): 212 | # OPTIONAL 213 | transform = transforms.Compose([transforms.ToTensor()]) 214 | return DataLoader(IEMOCAPSpectDataset(self.hparams.data_root, set_type='test', transform=transform, num_classes=self.num_classes), 215 | batch_size=32, num_workers=2, pin_memory=True, 216 | shuffle=True) 217 | 218 | @staticmethod 219 | def add_model_specific_args(parent_parser): 220 | """ 221 | Specify the hyperparams for this LightningModule 222 | """ 223 | # MODEL specific 224 | parser = ArgumentParser(parents=[parent_parser]) 225 | 226 | parser.add_argument('--learning_rate_init', 227 | default=0.0002898, type=float) 228 | parser.add_argument('--learning_rate_final', 229 | default=0.01435, type=float) 230 | parser.add_argument('--batch_size', default=32, type=int) 231 | parser.add_argument('--weight_decay', default=0.004566, type=float) 232 | #cnn 233 | parser.add_argument('--dropout_1', default=0.5424, type=float) 234 | parser.add_argument('--dropout_2', default=0.257, type=float) 235 | parser.add_argument('--dropout_3', default=0.558, type=float) 236 | #rnn 237 | parser.add_argument('--bidirectional', default=1, type=int) 238 | parser.add_argument('--num_layers_rnn', default=2, type=int) 239 | parser.add_argument('--dropout_rnn', default=0.0, type=float) 240 | parser.add_argument('--hidden_size_rnn', default=256, type=int) 241 | 242 | 243 | 244 | # training specific (for this model) 245 | parser.add_argument('--max_nb_epochs', default=10000, type=int) 246 | 247 | # data 248 | parser.add_argument( 249 | '--data_root', default='../datasets/RAVDESS/SOUND_SPECT/', type=str) 250 | parser.add_argument( 251 | '--num_classes', dest='num_classes', default=8, type=int) 252 | return parser 253 | -------------------------------------------------------------------------------- /research_seed/audio_classification/cnn_sweep.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file runs the main training/val loop, etc... using Lightning Trainer 3 | """ 4 | # %% 5 | import os 6 | import sys 7 | os.chdir('../../') 8 | current_path = os.path.abspath('.') 9 | sys.path.append(current_path) 10 | 11 | import logging 12 | from shutil import copyfile 13 | import numpy as np 14 | import pandas as pd 15 | from sklearn.metrics import confusion_matrix, classification_report 16 | import torch.nn.functional as F 17 | import torch 18 | from research_seed.audio_classification.datasets.iemocap_spect import IEMOCAPSpectDataset 19 | from pytorch_lightning.callbacks import EarlyStopping 20 | import wandb 21 | from pytorch_lightning.loggers import WandbLogger 22 | from research_seed.audio_classification.cnn_rnn import CNN_RNN 23 | from research_seed.audio_classification.cnn_lflb import CNN_LFLB 24 | from research_seed.audio_classification.cnn_spect import CNN_SPECT 25 | from argparse import ArgumentParser 26 | from pytorch_lightning import Trainer 27 | 28 | 29 | logging.basicConfig( 30 | level=logging.INFO, 31 | format="%(asctime)s [%(levelname)s] %(message)s", 32 | handlers=[ 33 | logging.FileHandler("debug.log"), 34 | logging.StreamHandler() 35 | ] 36 | ) 37 | 38 | 39 | def report(model, wandb_logger): 40 | # https://donatstudios.com/CsvToMarkdownTable 41 | 42 | model.eval() 43 | model = model.cpu() 44 | y_pred = [] 45 | y_true = [] 46 | 47 | for x, y in model.val_dataloader(): 48 | 49 | res = torch.max(F.softmax(model(x), dim=1), 1)[1].numpy() 50 | y_pred.extend(res) 51 | y_true.extend(y.numpy()) 52 | 53 | unique_label = np.unique([y_true, y_pred]) 54 | cmtx = pd.DataFrame( 55 | confusion_matrix(y_true, y_pred, labels=unique_label), 56 | index=['true:{:}'.format(x) for x in unique_label], 57 | columns=['pred:{:}'.format(x) for x in unique_label] 58 | ) 59 | 60 | report = pd.DataFrame(classification_report( 61 | y_true, y_pred, output_dict=True)) 62 | print(cmtx, report) 63 | wreport = [] 64 | tmp = [str(item) for item in report.values[0]] 65 | tmp.insert(0, 'precision') 66 | wreport.append(tmp) 67 | tmp = [str(item) for item in report.values[1]] 68 | tmp.insert(0, 'recall') 69 | wreport.append(tmp) 70 | tmp = [str(item) for item in report.values[2]] 71 | tmp.insert(0, 'f1-score') 72 | wreport.append(tmp) 73 | tmp = [str(item) for item in report.values[3]] 74 | tmp.insert(0, 'support') 75 | wreport.append(tmp) 76 | 77 | hreport = report.columns 78 | hreport = hreport.insert(0, '') 79 | 80 | wandb_logger.log_metrics({'confusion_matrix': wandb.plots.HeatMap(unique_label, unique_label, cmtx.values, show_text=True), 81 | 'classification_report': wandb.Table(data=wreport, columns=hreport.values)}) 82 | 83 | 84 | def main(hparams, network): 85 | # init module 86 | 87 | model = network(hparams) 88 | print(model.hparams) 89 | project_folder = 'audio_emotion_team' 90 | wandb_logger = WandbLogger( 91 | name='lflb_dropout_rnn', project=project_folder, entity='thesis', offline=False) 92 | 93 | early_stop_callback = EarlyStopping( 94 | monitor='val_loss', 95 | min_delta=0.00, 96 | patience=20, 97 | verbose=False, 98 | mode='min' 99 | ) 100 | 101 | # most basic trainer, uses good defaults 102 | trainer = Trainer( 103 | max_nb_epochs=hparams.max_nb_epochs, 104 | gpus=hparams.gpus, 105 | nb_gpu_nodes=hparams.nodes, 106 | logger=wandb_logger, 107 | #weights_summary='full', 108 | early_stop_callback=early_stop_callback, 109 | #profiler=True, 110 | benchmark=True, 111 | #log_gpu_memory='all' 112 | 113 | ) 114 | wandb_logger.experiment 115 | wandb_logger.watch(model) 116 | 117 | trainer.fit(model) 118 | 119 | 120 | if __name__ == '__main__': 121 | parser = ArgumentParser(add_help=False) 122 | parser.add_argument('--gpus', type=str, default=1) 123 | parser.add_argument('--nodes', type=int, default=1) 124 | 125 | network = CNN_RNN#CNN_LFLB # CNN_RNN 126 | # give the module a chance to add own params 127 | # good practice to define LightningModule speficic params in the module 128 | parser = network.add_model_specific_args(parser) 129 | 130 | # parse params 131 | #print(os.getcwd()) 132 | hparams = parser.parse_args() 133 | 134 | main(hparams, network) 135 | 136 | 137 | # %% 138 | -------------------------------------------------------------------------------- /research_seed/audio_classification/cnn_trainer.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file runs the main training/val loop, etc... using Lightning Trainer 3 | """ 4 | # %% 5 | import os 6 | import sys 7 | current_path = os.path.abspath('.') 8 | sys.path.append(current_path) 9 | 10 | import logging 11 | from shutil import copyfile 12 | import numpy as np 13 | import pandas as pd 14 | from sklearn.metrics import confusion_matrix, classification_report 15 | import torch.nn.functional as F 16 | import torch 17 | from research_seed.audio_classification.datasets.iemocap_spect import IEMOCAPSpectDataset 18 | from pytorch_lightning.callbacks import EarlyStopping 19 | import wandb 20 | from pytorch_lightning.loggers import WandbLogger 21 | from research_seed.audio_classification.cnn_rnn import CNN_RNN 22 | from research_seed.audio_classification.cnn_lflb import CNN_LFLB 23 | 24 | from argparse import ArgumentParser 25 | from pytorch_lightning import Trainer 26 | 27 | 28 | logging.basicConfig( 29 | level=logging.INFO, 30 | format="%(asctime)s [%(levelname)s] %(message)s", 31 | handlers=[ 32 | logging.FileHandler("debug.log"), 33 | logging.StreamHandler() 34 | ] 35 | ) 36 | 37 | 38 | def report(model, wandb_logger): 39 | 40 | model.eval() 41 | model = model.cpu() 42 | y_pred = [] 43 | y_true = [] 44 | 45 | for x, y in model.val_dataloader(): 46 | 47 | res = torch.max(F.softmax(model(x), dim=1), 1)[1].numpy() 48 | y_pred.extend(res) 49 | y_true.extend(y.numpy()) 50 | 51 | unique_label = np.unique([y_true, y_pred]) 52 | cmtx = pd.DataFrame( 53 | confusion_matrix(y_true, y_pred, labels=unique_label), 54 | index=['true:{:}'.format(x) for x in unique_label], 55 | columns=['pred:{:}'.format(x) for x in unique_label] 56 | ) 57 | 58 | report = pd.DataFrame(classification_report( 59 | y_true, y_pred, output_dict=True)) 60 | print(cmtx, report) 61 | wreport = [] 62 | tmp = [str(item) for item in report.values[0]] 63 | tmp.insert(0, 'precision') 64 | wreport.append(tmp) 65 | tmp = [str(item) for item in report.values[1]] 66 | tmp.insert(0, 'recall') 67 | wreport.append(tmp) 68 | tmp = [str(item) for item in report.values[2]] 69 | tmp.insert(0, 'f1-score') 70 | wreport.append(tmp) 71 | tmp = [str(item) for item in report.values[3]] 72 | tmp.insert(0, 'support') 73 | wreport.append(tmp) 74 | 75 | hreport = report.columns 76 | hreport = hreport.insert(0, '') 77 | 78 | wandb_logger.log_metrics({'confusion_matrix': wandb.plots.HeatMap(unique_label, unique_label, cmtx.values, show_text=True), 79 | 'classification_report': wandb.Table(data=wreport, columns=hreport.values)}) 80 | 81 | 82 | def main(hparams, network): 83 | # init module 84 | 85 | model = network(hparams) 86 | project_folder = 'audio_emotion_team' 87 | wandb_logger = WandbLogger( 88 | name='lflb_dropout_rnn', project=project_folder, entity='thesis', offline=False) 89 | 90 | early_stop_callback = EarlyStopping( 91 | monitor='val_loss', 92 | min_delta=0.00, 93 | patience=20, 94 | verbose=False, 95 | mode='min' 96 | ) 97 | 98 | # most basic trainer, uses good defaults 99 | trainer = Trainer( 100 | max_nb_epochs=hparams.max_nb_epochs, 101 | gpus=hparams.gpus, 102 | nb_gpu_nodes=hparams.nodes, 103 | logger=wandb_logger, 104 | weights_summary='full', 105 | early_stop_callback=early_stop_callback, 106 | profiler=True, 107 | benchmark=True, 108 | log_gpu_memory='all' 109 | 110 | ) 111 | 112 | wandb_logger.experiment.config.update( 113 | {'dataset': 'IEMOCAP_SPECT_GS_8s_512h_2048n'}) 114 | wandb_logger.watch(model) 115 | 116 | trainer.fit(model) 117 | # load best model 118 | exp_folder = project_folder + '/version_'+wandb_logger.experiment.id 119 | model_file = os.listdir(exp_folder + '/checkpoints')[0] 120 | # eval and upload best model 121 | model = network.load_from_checkpoint( 122 | exp_folder+'/checkpoints/' + model_file) 123 | report(model, wandb_logger) 124 | copyfile(exp_folder+'/checkpoints/' + model_file, 125 | wandb_logger.experiment.dir+'/model.ckpt') 126 | wandb_logger.experiment.save('model.ckpt') 127 | 128 | 129 | if __name__ == '__main__': 130 | parser = ArgumentParser(add_help=False) 131 | parser.add_argument('--gpus', type=str, default=1) 132 | parser.add_argument('--nodes', type=int, default=1) 133 | 134 | network = CNN_RNN#CNN_LFLB # CNN_RNN 135 | parser = network.add_model_specific_args(parser) 136 | 137 | # parse params 138 | print(os.getcwd()) 139 | hparams = parser.parse_args(["--data_root", "../datasets/IEMOCAP/SOUND_SPECT_GS_8s_512h_2048n/", '--max_nb_epochs', '10000', 140 | '--num_classes', '8']) 141 | 142 | main(hparams, network) 143 | 144 | 145 | # %% 146 | -------------------------------------------------------------------------------- /research_seed/audio_classification/model_testing.py: -------------------------------------------------------------------------------- 1 | """ 2 | This file runs the main training/val loop, etc... using Lightning Trainer 3 | """ 4 | #%% 5 | import os 6 | import sys 7 | current_path = os.path.abspath('.') 8 | sys.path.append(current_path) 9 | from pytorch_lightning import Trainer 10 | from pytorch_lightning.profiler import Profiler 11 | from argparse import ArgumentParser 12 | 13 | from research_seed.audio_classification.cnn_spect import CNN_SPECT 14 | from research_seed.audio_classification.cnn_lflb import CNN_LFLB 15 | from research_seed.audio_classification.cnn_rnn import CNN_RNN 16 | from pytorch_lightning.loggers import WandbLogger 17 | import wandb 18 | from pytorch_lightning.callbacks import EarlyStopping 19 | from research_seed.audio_classification.datasets.iemocap_spect import IEMOCAPSpectDataset 20 | 21 | import torch 22 | import torch.nn.functional as F 23 | from sklearn.metrics import confusion_matrix, classification_report 24 | import pandas as pd 25 | import numpy as np 26 | 27 | from shutil import copyfile 28 | import sys 29 | import logging 30 | # ... 31 | logging.basicConfig( 32 | level=logging.INFO, 33 | format="%(asctime)s [%(levelname)s] %(message)s", 34 | handlers=[ 35 | logging.FileHandler("debug.log"), 36 | logging.StreamHandler() 37 | ] 38 | ) 39 | def report(model, wandb_logger): 40 | #https://donatstudios.com/CsvToMarkdownTable 41 | 42 | model.eval() 43 | model = model.cpu() 44 | y_pred = [] 45 | y_true = [] 46 | 47 | for x,y in model.val_dataloader(): 48 | 49 | res = torch.max(F.softmax(model(x), dim=1),1)[1].numpy() 50 | y_pred.extend(res) 51 | y_true.extend(y.numpy()) 52 | break 53 | 54 | 55 | 56 | unique_label = np.unique([y_true, y_pred]) 57 | cmtx = pd.DataFrame( 58 | confusion_matrix(y_true, y_pred, labels=unique_label), 59 | index=['true:{:}'.format(x) for x in unique_label], 60 | columns=['pred:{:}'.format(x) for x in unique_label] 61 | ) 62 | 63 | report = pd.DataFrame(classification_report(y_true,y_pred, output_dict=True)) 64 | 65 | wreport = [] 66 | tmp = [str(item) for item in report.values[0]] 67 | tmp.insert(0,'precision') 68 | wreport.append(tmp) 69 | tmp = [str(item) for item in report.values[1]] 70 | tmp.insert(0,'recall') 71 | wreport.append(tmp) 72 | tmp = [str(item) for item in report.values[2]] 73 | tmp.insert(0,'f1-score') 74 | wreport.append(tmp) 75 | tmp = [str(item) for item in report.values[3]] 76 | tmp.insert(0,'support') 77 | wreport.append(tmp) 78 | 79 | print(report,cmtx) 80 | 81 | 82 | 83 | hreport = report.columns 84 | hreport = hreport.insert(0,'') 85 | 86 | if wandb_logger: 87 | wandb_logger.log_metrics({'confusion_matrix': wandb.plots.HeatMap(unique_label, unique_label, cmtx.values, show_text=True), 88 | 'classification_report':wandb.Table(data=wreport, columns=hreport.values)}) 89 | def main(hparams, network): 90 | # init module 91 | debugging = True 92 | 93 | project_folder = 'test' 94 | model = network(hparams) 95 | #wandb_logger = WandbLogger(name='test',offline=True,project=project_folder,entity='ricardop0') 96 | 97 | 98 | early_stop_callback = EarlyStopping( 99 | monitor='val_loss', 100 | min_delta=0.00, 101 | patience=5, 102 | verbose=False, 103 | mode='min' 104 | ) 105 | 106 | # most basic trainer, uses good defaults 107 | trainer = Trainer( 108 | max_nb_epochs=hparams.max_nb_epochs, 109 | gpus=hparams.gpus, 110 | nb_gpu_nodes=hparams.nodes, 111 | fast_dev_run=debugging, 112 | weights_summary='full', 113 | early_stop_callback=early_stop_callback, 114 | profiler=True, 115 | benchmark=True, 116 | log_gpu_memory='all', 117 | overfit_pct = 0.1, 118 | #logger=wandb_logger, 119 | 120 | ) 121 | 122 | trainer.fit(model) 123 | # id = wandb_logger.experiment.id 124 | # print(id) 125 | # os.environ["WANDB_RUN_ID"] = id 126 | # wandb_logger = WandbLogger(name='test',offline=True,project=project_folder,entity='ricardop0') 127 | # wandb_logger.experiment 128 | # print(wandb_logger.experiment.id) 129 | # # load best model 130 | # exp_folder = project_folder + '/version_'+wandb_logger.experiment.id 131 | # model_file = os.listdir(exp_folder + '/checkpoints')[0] 132 | # # eval and upload best model 133 | # model = network.load_from_checkpoint( 134 | # exp_folder+'/checkpoints/' + model_file) 135 | # report(model, wandb_logger) 136 | # copyfile(exp_folder+'/checkpoints/' + model_file, 137 | # wandb_logger.experiment.dir+'/model.ckpt') 138 | # wandb_logger.experiment.save('model.ckpt') 139 | # print('here') 140 | # wandb_logger.finalize() 141 | 142 | #print(wandb_logger.experiment.config) 143 | #print(wandb_logger.experiment.id) 144 | #exp_folder = 'audio_class/version_'+wandb_logger.experiment.id 145 | #model_file = os.listdir(exp_folder + '/checkpoints')[0] 146 | #model = CNN_LFLB.load_from_checkpoint(exp_folder+'/checkpoints/'+ model_file) 147 | #report(model, None) 148 | #wandb_logger.experiment.save(exp_folder+'/checkpoints/'+ model_file) 149 | 150 | 151 | 152 | if __name__ == '__main__': 153 | 154 | parser = ArgumentParser(add_help=False) 155 | parser.add_argument('--gpus', type=str, default=1) 156 | parser.add_argument('--nodes', type=int, default=1) 157 | 158 | # give the module a chance to add own params 159 | # good practice to define LightningModule speficic params in the module 160 | network = CNN_RNN#CNN_LFLB#CNN_RNN 161 | parser = network.add_model_specific_args(parser) 162 | 163 | # parse params 164 | print(os.getcwd()) 165 | hparams = parser.parse_args(["--data_root", "../datasets/IEMOCAP/SOUND_SPECT_GS_8s_512h_2048n/", '--max_nb_epochs', '50', 166 | '--num_classes', '6']) 167 | 168 | main(hparams, network) 169 | 170 | 171 | 172 | #%% 173 | 174 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | 3 | from setuptools import setup, find_packages 4 | 5 | setup(name='Speech2dCNN_LSTM', 6 | version='0.0.1', 7 | description='A pytorch implementation of Speech emotion recognition using deep 1D & 2D CNN LSTM networks', 8 | author='', 9 | author_email='', 10 | url='https://github.com/RicardoP0/Speech2dCNN_LSTM.git', 11 | install_requires=[ 12 | 'pytorch-lightning' 13 | ], 14 | packages=find_packages() 15 | ) 16 | 17 | --------------------------------------------------------------------------------