├── .gitignore ├── Functions ├── act.m ├── activation_bp.m ├── activation_ff.m ├── activation_init.m ├── dact.m └── loss_handle.m ├── LICENSE ├── Layer ├── activation │ ├── activation_bp.m │ ├── activation_ff.m │ ├── activation_init.m │ ├── activation_init_cpu.m │ └── activation_init_gpu.m ├── dense │ ├── dense_bp.m │ ├── dense_bp_gpu.m │ ├── dense_ff.m │ ├── dense_ff_gpu.m │ ├── dense_init_cpu.m │ └── dense_init_gpu.m ├── dropout │ ├── drop.m │ ├── dropout_bp.m │ ├── dropout_bp_gpu.m │ ├── dropout_ff.m │ ├── dropout_ff_gpu.m │ ├── dropout_init_cpu.m │ └── dropout_init_gpu.m ├── lstm │ ├── lstm_bp.m │ ├── lstm_bp_gpu.m │ ├── lstm_ff.m │ ├── lstm_ff_gpu.m │ ├── lstm_init_cpu.m │ └── lstm_init_gpu.m ├── tensor │ ├── tensor_init.m │ ├── tensor_init_cpu.m │ └── tensor_init_gpu.m └── tensor_init_gpu.m ├── Model ├── eval_loss.m ├── layer_optimize.m ├── model_evaluate.m ├── model_init.m ├── model_load.m ├── model_predict.m ├── model_save.m └── model_train.m ├── README.md ├── cudnn_LSTM_FF.m ├── cumexhelp.h ├── example ├── char_rnn │ ├── Christ2FSM.m │ ├── char_rnn.m │ ├── textgenerate.m │ └── txt2seq.m ├── keras_test_lstm.py ├── test_lstm.asv ├── test_lstm.m └── test_mlp.m ├── matDL_128X128.ico ├── mat_cudnn.cu ├── mat_cudnn.h ├── mat_cudnn_test.cu ├── mat_cudnn_test.h └── mat_cudnn_test.mexw64 /.gitignore: -------------------------------------------------------------------------------- 1 | *.mat 2 | *.txt 3 | -------------------------------------------------------------------------------- /Functions/act.m: -------------------------------------------------------------------------------- 1 | function y=act(x,fun) 2 | switch fun 3 | case 'sigmoid' 4 | y = 1./(1+exp(-x)); 5 | return 6 | case 'tanh' 7 | y=tanh(x); 8 | return 9 | case 'softmax' 10 | E=exp(x- max(x,[],2)); 11 | y = E./ sum(E,2) ; 12 | return 13 | case 'Relu' 14 | y=x.*(x>0); 15 | return 16 | case 'linear' 17 | y=x; 18 | return 19 | end 20 | end -------------------------------------------------------------------------------- /Functions/activation_bp.m: -------------------------------------------------------------------------------- 1 | function layer=activation_bp(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | layer.e=next_layer.dx; 7 | end 8 | layer.dx=layer.e.*layer.dact(layer.output); 9 | end 10 | -------------------------------------------------------------------------------- /Functions/activation_ff.m: -------------------------------------------------------------------------------- 1 | function layer=activation_ff(layer,prelayer) 2 | if ~isequal(size(prelayer.output),layer.input_shape) 3 | error('Shape unmatched!') 4 | end 5 | layer.output=layer.act(prelayer.output); 6 | end -------------------------------------------------------------------------------- /Functions/activation_init.m: -------------------------------------------------------------------------------- 1 | function layer= activation_init( prelayer,act_fun,flag,loss ) 2 | %% Basic layer attributes 3 | layer.type='activation'; 4 | layer.trainable=0; 5 | layer.flag=flag; 6 | layer.prelayer_type=prelayer.type; 7 | 8 | layer.batch=1; 9 | layer.epoch=1; 10 | 11 | layer.input_shape=prelayer.output_shape; 12 | layer.output_shape=prelayer.output_shape; 13 | 14 | % layer.input=prelayer.output; 15 | layer.output=prelayer.output; 16 | 17 | if ~strcmpi(layer.prelayer_type,'input')&&flag 18 | layer.dx=layer.output; 19 | end 20 | layer.e=layer.output; 21 | 22 | if nargin>2 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.act=@(x)act(x,act_fun); 27 | layer.dact=@(x)dact(x,act_fun); 28 | layer.ff=@(layer,prelayer)activation_ff(layer,prelayer); 29 | layer.bp=@(layer,next_layer)activation_bp(layer,next_layer); 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.act_fun=act_fun; 34 | end 35 | 36 | -------------------------------------------------------------------------------- /Functions/dact.m: -------------------------------------------------------------------------------- 1 | function dx=dact(y,fun) 2 | switch fun 3 | case 'sigmoid' 4 | dx = y .* (1 - y); 5 | return 6 | case 'tanh' 7 | dx=1-y.^2; 8 | return 9 | case 'Relu' 10 | dx=(y>single(0)); 11 | return 12 | case 'linear' 13 | dx = y; 14 | return 15 | case 'softmax' 16 | dx=y; 17 | return 18 | end 19 | end -------------------------------------------------------------------------------- /Functions/loss_handle.m: -------------------------------------------------------------------------------- 1 | function [f,df]=loss_handle(type) 2 | syms y_true y_pred num 3 | switch type 4 | case 'mse' 5 | symsf(y_true,y_pred)=(y_true-y_pred).^2; 6 | f=matlabFunction(symsf); 7 | df=matlabFunction(diff(symsf,y_pred)); 8 | return 9 | case 'cross_entropy' 10 | symsf(y_true,y_pred)=-1.*sum(y_true.*(y_pred)+(1-y_true).*log(1-y_pred)); 11 | f=matlabFunction(symsf); 12 | df=matlabFunction(diff(symsf,y_pred)); 13 | return 14 | case 'categorical_cross_entropy' 15 | symsf(y_true,y_pred)=-1.*y_true.*log(y_pred); 16 | f=matlabFunction(symsf); 17 | df=@(y_true,y_pre)y_pre-y_true; 18 | end 19 | end 20 | -------------------------------------------------------------------------------- /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 | {one line to give the program's name and a brief idea of what it does.} 635 | Copyright (C) {year} {name of author} 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 | {project} Copyright (C) {year} {fullname} 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 | -------------------------------------------------------------------------------- /Layer/activation/activation_bp.m: -------------------------------------------------------------------------------- 1 | function layer=activation_bp(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | layer.e=next_layer.dx; 7 | end 8 | layer.dx=layer.e.*layer.dact(layer.output); 9 | end 10 | -------------------------------------------------------------------------------- /Layer/activation/activation_ff.m: -------------------------------------------------------------------------------- 1 | function layer=activation_ff(layer,prelayer) 2 | if ~isequal(size(prelayer.output),layer.input_shape) 3 | error('Shape unmatched!') 4 | end 5 | layer.output=layer.act(prelayer.output); 6 | end -------------------------------------------------------------------------------- /Layer/activation/activation_init.m: -------------------------------------------------------------------------------- 1 | function layer= activation_init( prelayer,act_fun,flag,loss ) 2 | %% Basic layer attributes 3 | layer.type='activation'; 4 | layer.trainable=0; 5 | layer.flag=flag; 6 | layer.prelayer_type=prelayer.type; 7 | 8 | layer.batch=1; 9 | layer.epoch=1; 10 | 11 | layer.input_shape=prelayer.output_shape; 12 | layer.output_shape=prelayer.output_shape; 13 | 14 | % layer.input=prelayer.output; 15 | layer.output=prelayer.output; 16 | 17 | if ~strcmpi(layer.prelayer_type,'input')&&flag 18 | layer.dx=layer.output; 19 | end 20 | layer.e=layer.output; 21 | 22 | if nargin>3 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.act=@(x)act(x,act_fun); 27 | layer.dact=@(x)dact(x,act_fun); 28 | layer.ff=@(layer,prelayer)activation_ff(layer,prelayer); 29 | layer.bp=@(layer,next_layer)activation_bp(layer,next_layer); 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.act_fun=act_fun; 34 | end 35 | 36 | -------------------------------------------------------------------------------- /Layer/activation/activation_init_cpu.m: -------------------------------------------------------------------------------- 1 | function layer= activation_init_cpu( prelayer,act_fun,flag,loss ) 2 | %% Basic layer attributes 3 | layer.type='activation'; 4 | layer.trainable=0; 5 | layer.flag=flag; 6 | layer.prelayer_type=prelayer.type; 7 | 8 | layer.batch=1; 9 | layer.epoch=1; 10 | 11 | layer.input_shape=prelayer.output_shape; 12 | layer.output_shape=prelayer.output_shape; 13 | 14 | % layer.input=prelayer.output; 15 | layer.output=prelayer.output; 16 | 17 | if ~strcmpi(layer.prelayer_type,'input')&&flag 18 | layer.dx=layer.output; 19 | end 20 | layer.e=layer.output; 21 | 22 | if nargin>3 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.act=@(x)act(x,act_fun); 27 | layer.dact=@(x)dact(x,act_fun); 28 | layer.ff=@(layer,prelayer)activation_ff(layer,prelayer); 29 | layer.bp=@(layer,next_layer)activation_bp(layer,next_layer); 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.act_fun=act_fun; 34 | end 35 | 36 | -------------------------------------------------------------------------------- /Layer/activation/activation_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer= activation_init_gpu( prelayer,act_fun,flag,loss ) 2 | %% Basic layer attributes 3 | layer.type='activation'; 4 | layer.trainable=0; 5 | layer.flag=flag; 6 | layer.prelayer_type=prelayer.type; 7 | 8 | layer.batch=1; 9 | layer.epoch=1; 10 | 11 | layer.input_shape=prelayer.output_shape; 12 | layer.output_shape=prelayer.output_shape; 13 | 14 | % layer.input=prelayer.output; 15 | layer.output=prelayer.output; 16 | 17 | if ~strcmpi(layer.prelayer_type,'input')&&flag 18 | layer.dx=layer.output; 19 | end 20 | layer.e=layer.output; 21 | 22 | if nargin>3 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.act=@(x)act(x,act_fun); 27 | layer.dact=@(x)dact(x,act_fun); 28 | layer.ff=@(layer,prelayer)activation_ff(layer,prelayer); 29 | layer.bp=@(layer,next_layer)activation_bp(layer,next_layer); 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.act_fun=act_fun; 34 | end 35 | 36 | -------------------------------------------------------------------------------- /Layer/dense/dense_bp.m: -------------------------------------------------------------------------------- 1 | function layer =dense_bp(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | layer.e=next_layer.dx; 7 | end 8 | if layer.timedistributed 9 | layer.dW=reshape(permute(layer.input,[2,1,3]),layer.weights_dim,[])*reshape(permute(layer.e,[2,1,3]),4*hiddensize,[])'; 10 | if ~isequal(layer.prelayer_type,'input') 11 | layer.dx(:)=mult_3d(layer.e,layer.W(1:end-1,:)'); 12 | end 13 | else 14 | layer.dW=layer.input'*layer.e; 15 | if ~isequal(layer.prelayer_type,'input') 16 | layer.dx=layer.e*layer.W(1:end-1,:)'; 17 | end 18 | end 19 | end 20 | function a=sq(a) 21 | a=reshape(a,size(a,1),[]); 22 | end 23 | function c=mult_3d(a,b) 24 | shape=size(a); 25 | timestep=shape(end); 26 | dim=shape(2); 27 | batchsize=shape(1); 28 | c=permute(reshape((reshape(permute(a,[2,1,3]),dim,[])'*b)',[dim,batchsize,timestep]),[2,1,3]); 29 | end -------------------------------------------------------------------------------- /Layer/dense/dense_bp_gpu.m: -------------------------------------------------------------------------------- 1 | function layer =dense_bp_gpu(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | layer.e=next_layer.dx; 7 | end 8 | layer.dW=sq(layer.e)*sq(layer.input)'; 9 | if ~isequal(layer.prelayer_type,'input') 10 | if layer.timedistributed 11 | layer.dx(:)=layer.W(:,1:end-1)'*sq(layer.e); 12 | else 13 | layer.dx=layer.W(:,1:end-1)'*layer.e; 14 | end 15 | end 16 | end 17 | function a=sq(a) 18 | a=reshape(a,size(a,1),[]); 19 | end 20 | -------------------------------------------------------------------------------- /Layer/dense/dense_ff.m: -------------------------------------------------------------------------------- 1 | function layer=dense_ff(layer,prelayer) 2 | if isequal(class(prelayer),'struct') 3 | if ~isequal(size(prelayer.output),layer.input_shape) 4 | error('Shape unmatched!') 5 | end 6 | if layer.timedistributed 7 | layer.input(:,1:end-1,:)=prelayer.output; 8 | else 9 | layer.input(:,1:end-1)=prelayer.output; 10 | end 11 | else 12 | if layer.timedistributed 13 | layer.input(:,1:end-1,:)=prelayer; 14 | else 15 | layer.input(:,1:end-1)=prelayer; 16 | end 17 | end 18 | if layer.timedistributed 19 | layer.output(:)=mult_3d(layer.input,layer.W); 20 | else 21 | layer.output=layer.input*layer.W; 22 | end 23 | end 24 | function c=mult_3d(a,b) 25 | shape=size(a); 26 | timestep=shape(end); 27 | dim=shape(2); 28 | batchsize=shape(1); 29 | c=permute(reshape((reshape(permute(a,[2,1,3]),dim,[])'*b)',[dim,batchsize,timestep]),[2,1,3]); 30 | end -------------------------------------------------------------------------------- /Layer/dense/dense_ff_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=dense_ff_gpu(layer,prelayer) 2 | if isequal(class(prelayer),'struct') 3 | if ~isequal(size(prelayer.output),layer.input_shape) 4 | error('Shape unmatched!') 5 | end 6 | if layer.timedistributed 7 | layer.input(1:end-1,:,:)=prelayer.output; 8 | else 9 | layer.input(1:end-1,:)=prelayer.output; 10 | end 11 | else 12 | if layer.timedistributed 13 | layer.input(1:end-1,:,:)=prelayer; 14 | else 15 | layer.input(1:end-1,:)=prelayer; 16 | end 17 | end 18 | if layer.timedistributed 19 | layer.output(:)=layer.W*sq(layer.input); 20 | else 21 | layer.output=layer.W*layer.input; 22 | end 23 | end 24 | function a=sq(a) 25 | a=reshape(a,size(a,1),[]); 26 | end 27 | -------------------------------------------------------------------------------- /Layer/dense/dense_init_cpu.m: -------------------------------------------------------------------------------- 1 | function layer=dense_init_cpu(prelayer,hiddensize ,flag,loss) 2 | %% Basic layer attributes 3 | %Input tensor sahpe 4 | layer.trainable=1; 5 | layer.flag=flag; 6 | layer.input_shape=prelayer.output_shape; 7 | if numel(prelayer.output_shape)>2 8 | layer.timedistributed=1; 9 | layer.output_shape=[layer.input_shape(1),hiddensize,layer.input_shape(end)]; 10 | else 11 | layer.timedistributed=0; 12 | layer.output_shape=[layer.input_shape(1),hiddensize]; 13 | end 14 | dim=prelayer.output_shape(2); 15 | batchsize=prelayer.output_shape(1); 16 | layer.type='dense'; 17 | layer.prelayer_type=prelayer.type; 18 | layer.hiddensize=hiddensize; 19 | layer.batchsize=batchsize; 20 | layer.batch=1; 21 | layer.epoch=1; 22 | %% Dense layer attributes 23 | %W contains weights bias 24 | layer.weights_dim=dim+1; 25 | layer.W=(rand([layer.weights_dim,hiddensize],'single')-0.5)./100; 26 | if layer.timedistributed 27 | layer.input=ones([layer.input_shape(1),layer.input_shape(2)+1,layer.input_shape(3)],'single'); 28 | else 29 | layer.input=ones([layer.input_shape(1),layer.input_shape(2)+1],'single'); 30 | end 31 | layer.output=zeros(layer.output_shape,'single'); 32 | if ~strcmpi(layer.prelayer_type,'input')&&flag 33 | layer.dx=zeros(layer.input_shape,'single'); 34 | end 35 | layer.e=layer.output; 36 | if nargin>3&&flag 37 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 38 | layer.loss=[]; 39 | end 40 | layer.ff=@(layer,prelayer)dense_ff(layer,prelayer); 41 | layer.bp=@(layer,next_layer)dense_bp(layer,next_layer); 42 | 43 | layer.configs.type=layer.type; 44 | layer.configs.input_shape=layer.input_shape; 45 | layer.configs.output_shape=layer.output_shape; 46 | layer.configs.hiddensize=layer.hiddensize; 47 | layer.configs.W=size(layer.W); 48 | end 49 | -------------------------------------------------------------------------------- /Layer/dense/dense_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=dense_init_gpu(prelayer,hiddensize ,flag,loss) 2 | %% Basic layer attributes 3 | %Input tensor sahpe 4 | layer.trainable=1; 5 | layer.flag=flag; 6 | layer.input_shape=prelayer.output_shape; 7 | if numel(prelayer.output_shape)>2 8 | layer.timedistributed=1; 9 | layer.output_shape=[layer.input_shape(1),hiddensize,layer.input_shape(end)]; 10 | else 11 | layer.timedistributed=0; 12 | layer.output_shape=[layer.input_shape(1),hiddensize]; 13 | end 14 | dim=prelayer.output_shape(2); 15 | batchsize=prelayer.output_shape(1); 16 | layer.type='dense'; 17 | layer.prelayer_type=prelayer.type; 18 | layer.hiddensize=hiddensize; 19 | layer.batchsize=batchsize; 20 | layer.batch=1; 21 | layer.epoch=1; 22 | %% Dense layer attributes 23 | %W contains weights bias 24 | layer.weights_dim=dim+1; 25 | layer.W=(rand([layer.weights_dim,hiddensize],'single','gpuArray')-0.5)./100; 26 | if layer.timedistributed 27 | layer.input=ones([layer.input_shape(1),layer.input_shape(2)+1,layer.input_shape(3)],'single','gpuArray'); 28 | else 29 | layer.input=ones([layer.input_shape(1),layer.input_shape(2)+1],'single','gpuArray'); 30 | end 31 | layer.output=zeros(layer.output_shape,'single','gpuArray'); 32 | if ~strcmpi(layer.prelayer_type,'input')&&flag 33 | layer.dx=zeros(layer.input_shape,'single','gpuArray'); 34 | end 35 | layer.e=layer.output; 36 | if nargin>3&&flag 37 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 38 | layer.loss=[]; 39 | end 40 | layer.ff=@(layer,prelayer)dense_ff(layer,prelayer); 41 | layer.bp=@(layer,next_layer)dense_bp(layer,next_layer); 42 | 43 | layer.configs.type=layer.type; 44 | layer.configs.input_shape=layer.input_shape; 45 | layer.configs.output_shape=layer.output_shape; 46 | layer.configs.hiddensize=layer.hiddensize; 47 | layer.configs.W=size(layer.W); 48 | end 49 | -------------------------------------------------------------------------------- /Layer/dropout/drop.m: -------------------------------------------------------------------------------- 1 | function [mask,mask_index]=drop(mask,drop_rate) 2 | mask_index=randperm(numel(mask),floor(numel(mask)*drop_rate)); 3 | mask(mask_index)=0; 4 | end -------------------------------------------------------------------------------- /Layer/dropout/dropout_bp.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_bp(layer,next_layer) 2 | layer.dx=next_layer.dx.*layer.mask; 3 | end -------------------------------------------------------------------------------- /Layer/dropout/dropout_bp_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_bp_gpu(layer,next_layer) 2 | layer.dx=next_layer.dx.*layer.mask; 3 | end -------------------------------------------------------------------------------- /Layer/dropout/dropout_ff.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_ff(layer,prelayer) 2 | if layer.flag 3 | [layer.mask,layer.mask_index]=layer.drop(layer.mask,layer.drop_rate); 4 | layer.output=prelayer.output.*layer.mask; 5 | else 6 | layer.output=prelayer.output*layer.drop_rate; 7 | end 8 | end -------------------------------------------------------------------------------- /Layer/dropout/dropout_ff_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_ff_gpu(layer,prelayer) 2 | if layer.flag 3 | [layer.mask,layer.mask_index]=layer.drop(layer.mask,layer.drop_rate); 4 | layer.output=prelayer.output.*layer.mask; 5 | else 6 | layer.output=prelayer.output*layer.drop_rate; 7 | end 8 | end -------------------------------------------------------------------------------- /Layer/dropout/dropout_init_cpu.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_init_cpu(prelayer,drop_rate ,flag,loss) 2 | %% Basic layer attributes 3 | layer.trainable=0; 4 | layer.flag=flag; 5 | layer.input_shape=prelayer.output_shape; 6 | batchsize=prelayer.output_shape(1); 7 | layer.type='dropout'; 8 | layer.prelayer_type=prelayer.type; 9 | layer.output_shape=layer.input_shape; 10 | layer.batchsize=batchsize; 11 | layer.batch=1; 12 | layer.epoch=1; 13 | %% Dropout layer attributes 14 | layer.drop_rate=drop_rate; 15 | if layer.flag 16 | layer.mask=ones(layer.output_shape,'single'); 17 | end 18 | layer.output=zeros(layer.output_shape,'single'); 19 | if ~strcmpi(layer.prelayer_type,'input')&&flag 20 | layer.dx=zeros(layer.input_shape,'single'); 21 | end 22 | if nargin>3&&flag 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.drop=@(mask,drop_rate)drop(mask,drop_rate); 27 | layer.ff=@(layer,prelayer)dropout_ff(layer,prelayer); 28 | layer.bp=@(layer,next_layer)dropout_bp(layer,next_layer); 29 | 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.drop_rate=layer.drop_rate; 34 | end 35 | -------------------------------------------------------------------------------- /Layer/dropout/dropout_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=dropout_init_gpu(prelayer,drop_rate ,flag,loss) 2 | %% Basic layer attributes 3 | layer.trainable=0; 4 | layer.flag=flag; 5 | layer.input_shape=prelayer.output_shape; 6 | batchsize=prelayer.output_shape(1); 7 | layer.type='dropout'; 8 | layer.prelayer_type=prelayer.type; 9 | layer.output_shape=layer.input_shape; 10 | layer.batchsize=batchsize; 11 | layer.batch=1; 12 | layer.epoch=1; 13 | %% Dropout layer attributes 14 | layer.drop_rate=drop_rate; 15 | if layer.flag 16 | layer.mask=ones(layer.output_shape,'single','gpuArray'); 17 | end 18 | layer.output=zeros(layer.output_shape,'single','gpuArray'); 19 | if ~strcmpi(layer.prelayer_type,'input')&&flag 20 | layer.dx=zeros(layer.input_shape,'single','gpuArray'); 21 | end 22 | if nargin>3&&flag 23 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 24 | layer.loss=[]; 25 | end 26 | layer.drop=@(mask,drop_rate)drop(mask,drop_rate); 27 | layer.ff=@(layer,prelayer)dropout_ff(layer,prelayer); 28 | layer.bp=@(layer,next_layer)dropout_bp(layer,next_layer); 29 | 30 | layer.configs.type=layer.type; 31 | layer.configs.input_shape=layer.input_shape; 32 | layer.configs.output_shape=layer.output_shape; 33 | layer.configs.drop_rate=layer.drop_rate; 34 | end 35 | -------------------------------------------------------------------------------- /Layer/lstm/lstm_bp.m: -------------------------------------------------------------------------------- 1 | function layer=lstm_bp(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | if layer.return_sequence 7 | layer.e=next_layer.dx; 8 | else 9 | layer.e(:,end,:)=next_layer.dx; 10 | end 11 | end 12 | timestep=layer.timestep; 13 | hiddensize=layer.hiddensize; 14 | batchsize=layer.batchsize; 15 | dim=layer.input_shape(2); 16 | r_x=1:dim+1; 17 | r_h=dim+1+(1:hiddensize); 18 | r_ifo=1:3*hiddensize; 19 | r_f=1:hiddensize; 20 | r_i=hiddensize+1:2*hiddensize; 21 | r_o=2*hiddensize+1:3*hiddensize; 22 | r_tc=3*hiddensize+1:4*hiddensize; 23 | %% Backpropagation through time 24 | for t=timestep:-1:2 25 | % d_h(t) = e(t) + d_a(t+1)*W 26 | layer.dh(:,:,t)=layer.e(:,:,t)+layer.dma(:,:,t+1)*layer.W(r_h,:)'; 27 | % d_c(t) = d_h(t) .* o(t) * tanh'(c(t)) 28 | layer.dsc(:,:,t)=layer.dh(:,:,t).*layer.mb(:,r_o,t).*layer.dact_h(layer.sc(:,:,t)); 29 | %db_o(t) = d_h(t) * bc(t) 30 | layer.dmb(:,r_o,t)=layer.dh(:,:,t).*layer.bc(:,:,t); 31 | % db_i(t) = d_c(t) .* tc(t) 32 | layer.dmb(:,r_i,t)=layer.dsc(:,:,t).*layer.mb(:,r_tc,t); 33 | % db_tc(t) = db_c(t) .* i(t) 34 | layer.dmb(:,r_tc,t)=layer.dsc(:,:,t).*layer.mb(:,r_i,t); 35 | % db_f(t) = db_c(t) .* c(t-1) 36 | layer.dmb(:,r_f,t)=layer.dsc(:,:,t).*layer.sc(:,:,t-1); 37 | %da=act'(b).*db 38 | layer.dma(:,r_ifo,t)=layer.dact_f(layer.mb(:,r_ifo,t)).*layer.dmb(:,r_ifo,t); 39 | layer.dma(:,r_tc,t)=layer.dact_tc(layer.mb(:,r_tc,t)).*layer.dmb(:,r_tc,t); 40 | end 41 | t=1; 42 | layer.dh(:,:,t)=layer.e(:,:,t)+layer.dma(:,:,t+1)*layer.W(r_h,:)'; 43 | layer.dsc(:,:,t)=layer.dh(:,:,t).*layer.mb(:,r_o,t).*layer.dact_h(layer.sc(:,:,t)); 44 | layer.dmb(:,r_o,t)=layer.dh(:,:,t).*layer.bc(:,:,t); 45 | layer.dmb(:,r_i,t)=layer.dsc(:,:,t).*layer.mb(:,r_tc,t); 46 | layer.dmb(:,r_tc,t)=layer.dsc(:,:,t).*layer.mb(:,r_i,t); 47 | layer.dma(:,r_ifo,t)=layer.dact_f(layer.mb(:,r_ifo,t)).*layer.dmb(:,r_ifo,t); 48 | layer.dma(:,r_tc,t)=layer.dact_tc(layer.mb(:,r_tc,t)).*layer.dmb(:,r_tc,t); 49 | 50 | if ~isequal(layer.prelayer_type,'input') 51 | layer.dx(:)=mult_3d(layer.dma(:,:,1:end-1),layer.W(:,r_x(1:end-1))'); 52 | end 53 | %layer.dma(:,r_f,2:end)=layer.dma(:,r_f,2:end)./(timestep-1); 54 | %layer.dma(:,hiddensize+1:end,:)=layer.dma(:,hiddensize+1:end,:)./timestep; 55 | layer.dW=reshape(permute(layer.xh,[2,1,3]),layer.weights_dim,[])*reshape(permute(layer.dma,[2,1,3]),4*hiddensize,[])'; 56 | end 57 | function c=mult_3d(a,b) 58 | input_shape=size(a); 59 | output_dim=size(b,2); 60 | timestep=input_shape(end); 61 | input_dim=input_shape(2); 62 | batchsize=input_shape(1); 63 | c=permute(reshape((reshape(permute(a,[2,1,3]),input_dim,[])'*b)',[output_dim,batchsize,timestep]),[2,1,3]); 64 | end -------------------------------------------------------------------------------- /Layer/lstm/lstm_bp_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=lstm_bp_gpu(layer,next_layer) 2 | if isequal(class(next_layer),'struct') 3 | if ~isequal(size(next_layer.dx),layer.output_shape) 4 | error('Shape unmatched!') 5 | end 6 | if layer.return_sequence 7 | layer.e=next_layer.dx; 8 | else 9 | layer.e(:,end,:)=next_layer.dx; 10 | end 11 | end 12 | timestep=layer.timestep; 13 | hiddensize=layer.hiddensize; 14 | batchsize=layer.batchsize; 15 | dim=layer.input_shape(1); 16 | r_x=1:dim+1; 17 | r_h=dim+1+(1:hiddensize); 18 | r_ifo=1:3*hiddensize; 19 | r_f=1:hiddensize; 20 | r_i=hiddensize+1:2*hiddensize; 21 | r_o=2*hiddensize+1:3*hiddensize; 22 | r_tc=3*hiddensize+1:4*hiddensize; 23 | %% Backpropagation through time 24 | for t=timestep:-1:2 25 | % d_h(t) = e(t) + d_a(t+1)*W 26 | layer.dh(:,t,:)=sq(layer.e(:,t,:))+layer.W(:,r_h)'*sq(layer.dma(:,t+1,:)); 27 | % d_c(t) = d_h(t) .* o(t) * tanh'(c(t)) 28 | layer.dsc(:,t,:)=layer.dh(:,t,:).*layer.mb(r_o,t,:).*layer.dact_h(layer.sc(:,t,:)); 29 | %db_o(t) = d_h(t) * bc(t) 30 | layer.dmb(r_o,t,:)=layer.dh(:,t,:).*layer.bc(:,t,:); 31 | % db_i(t) = d_c(t) .* tc(t) 32 | layer.dmb(r_i,t,:)=layer.dsc(:,t,:).*layer.mb(r_tc,t,:); 33 | % db_tc(t) = db_c(t) .* i(t) 34 | layer.dmb(r_tc,t,:)=layer.dsc(:,t,:).*layer.mb(r_i,t,:); 35 | % db_f(t) = db_c(t) .* c(t-1) 36 | layer.dmb(r_f,t,:)=layer.dsc(:,t,:).*layer.sc(:,t-1,:); 37 | %da=act'(b).*db 38 | layer.dma(r_ifo,t,:)=layer.dact_f(layer.mb(r_ifo,t,:)).*layer.dmb(r_ifo,t,:); 39 | layer.dma(r_tc,t,:)=layer.dact_tc(layer.mb(r_tc,t,:)).*layer.dmb(r_tc,t,:); 40 | end 41 | t=1; 42 | layer.dh(:,t,:)=sq(layer.e(:,t,:))+layer.W(:,r_h)'*sq(layer.dma(:,t+1,:)); 43 | layer.dsc(:,t,:)=layer.dh(:,t,:).*layer.mb(r_o,t,:).*layer.dact_h(layer.sc(:,t,:)); 44 | layer.dmb(r_o,t,:)=layer.dh(:,t,:).*layer.bc(:,t,:); 45 | layer.dmb(r_i,t,:)=layer.dsc(:,t,:).*layer.mb(r_tc,t,:); 46 | layer.dmb(r_tc,t,:)=layer.dsc(:,t,:).*layer.mb(r_i,t,:); 47 | layer.dma(r_ifo,t,:)=layer.dact_f(layer.mb(r_ifo,t,:)).*layer.dmb(r_ifo,t,:); 48 | layer.dma(r_tc,t,:)=layer.dact_tc(layer.mb(r_tc,t,:)).*layer.dmb(r_tc,t,:); 49 | 50 | layer.dma(r_f,2:end,:)=layer.dma(r_f,2:end,:)./(timestep-1); 51 | layer.dma(hiddensize+1:end,:)=layer.dma(hiddensize+1:end,:)./timestep; 52 | layer.dW=layer.dma(:,:)*layer.xh(:,:)'./batchsize; 53 | if ~isequal(layer.prelayer_type,'input') 54 | layer.dx(:)=layer.W(:,r_x(1:end-1))'*sq(layer.dma(:,1:end-1,:))./batchsize; 55 | end 56 | end 57 | function a=sq(a) 58 | a=reshape(a,size(a,1),[]); 59 | end -------------------------------------------------------------------------------- /Layer/lstm/lstm_ff.m: -------------------------------------------------------------------------------- 1 | function layer=lstm_ff(layer,prelayer) 2 | timestep=layer.timestep; 3 | hiddensize=layer.hiddensize; 4 | dim=layer.input_shape(2); 5 | r_x=1:dim+1;%range of x and bias 6 | r_h=dim+1+(1:hiddensize);%range of h 7 | r_ifo=1:3*hiddensize;%range of forget,input and output gates 8 | r_f=1:hiddensize;%range of forget gate 9 | r_i=hiddensize+1:2*hiddensize;%~input gate 10 | r_o=2*hiddensize+1:3*hiddensize;%~output gate 11 | r_tc=3*hiddensize+1:4*hiddensize;%range of tilde c gate 12 | %the xh is a 2d tensor contain x,bias,and h,((r_x)-1,1:end-1,:) is the area of x 13 | %assign value from input tensor 14 | if isequal(class(prelayer),'struct') 15 | if ~isequal(size(prelayer.output),layer.input_shape) 16 | error('Shape unmatched!') 17 | end 18 | layer.xh(:,r_x(1:end-1),1:end-1)=prelayer.output; 19 | else 20 | layer.xh(:,r_x(1:end-1),1:end-1)=prelayer; 21 | end 22 | %compute all x(t)*W_x+bias in one time at first 23 | layer.maX(:)=mult_3d(layer.xh(:,r_x,1:end-1),layer.W(r_x,:)); 24 | 25 | %% Feed forward 26 | %t=1 27 | layer.ma( :,:,1)=layer.maX( :,:,1); 28 | layer.mb( :,r_ifo,1)=layer.act_f(layer.ma(:,r_ifo,1)); 29 | layer.mb( :,r_tc,1)=layer.act_tc(layer.ma(:,r_tc,1)); 30 | layer.sc( :,:,1)=layer.mb(:,r_i,1).*layer.mb(:,r_tc,1); 31 | layer.bc( :,:,1)=layer.act_h(layer.sc( :,:,1)); 32 | layer.xh(:,r_h,2)=layer.bc(:,:,1).*layer.mb(:,r_o,1); 33 | %t>1 34 | for t=2:timestep 35 | % a(t) = W_x * x(t) + W_h * h(t-1) 36 | layer.ma( :,:,t)=layer.maX( :,:,t)+layer.xh( :,r_h,t)*layer.W(r_h,:); 37 | %b(t)=act(a(t)) 38 | %The active functions of i,f,o gates are sigmoid,compute in one time 39 | layer.mb( :,r_ifo,t)=layer.act_f(layer.ma( :,r_ifo,t)); 40 | %The active function of tc gate is tanh 41 | layer.mb( :,r_tc,t)=layer.act_tc(layer.ma( :,r_tc,t)); 42 | % c(t) = f(t) * c(t-1) + i(t) * tc(t) 43 | layer.sc( :,:,t)=layer.sc( :,:,t-1).*layer.mb( :,r_f,t)+layer.mb( :,r_i,t).*layer.mb( :,r_tc,t); 44 | %tanh(c(t)) 45 | layer.bc( :,:,t)=layer.act_h(layer.sc( :,:,t)); 46 | % h(t) = o(t) * tanh(c(t)) 47 | layer.xh( :,r_h,t+1)=layer.bc( :,:,t).*layer.mb( :,r_o,t); 48 | if layer.return_sequence 49 | layer.output=layer.xh(:,r_h,2:end); 50 | else 51 | layer.output=sq(layer.xh(:,r_h,1)); 52 | end 53 | end 54 | end 55 | function c=mult_3d(a,b) 56 | input_shape=size(a); 57 | output_dim=size(b,2); 58 | timestep=input_shape(end); 59 | input_dim=input_shape(2); 60 | batchsize=input_shape(1); 61 | c=permute(reshape((reshape(permute(a,[2,1,3]),input_dim,[])'*b)',[output_dim,batchsize,timestep]),[2,1,3]); 62 | end -------------------------------------------------------------------------------- /Layer/lstm/lstm_ff_gpu.m: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/QuantumLiu/matDL/ce75b21dd97e9d58c074d50915d5fc000ee46afb/Layer/lstm/lstm_ff_gpu.m -------------------------------------------------------------------------------- /Layer/lstm/lstm_init_cpu.m: -------------------------------------------------------------------------------- 1 | function layer=lstm_init_cpu(prelayer,hiddensize,return_sequence,flag,loss) 2 | %% Basic layer attributes 3 | %Input tensor sahpe 4 | layer.input_shape=prelayer.output_shape; 5 | layer.trainable=1; 6 | layer.flag=flag; 7 | 8 | dim=prelayer.output_shape(2); 9 | timestep=prelayer.output_shape(3); 10 | batchsize=prelayer.output_shape(1); 11 | if nargin<3 12 | return_sequence=1; 13 | end 14 | layer.return_sequence=return_sequence; 15 | if return_sequence 16 | %Output tensor shape 17 | layer.output_shape=[batchsize,hiddensize,timestep]; 18 | else 19 | layer.output_shape=[batchsize,hiddensize]; 20 | end 21 | %The type of the layer 22 | layer.type='lstm'; 23 | %conected layer type 24 | layer.prelayer_type=prelayer.type; 25 | %The hiddensize of the layer 26 | layer.hiddensize=hiddensize; 27 | 28 | layer.batch=1; 29 | layer.epoch=1; 30 | %% lstm layer attributes 31 | %Timestep 32 | layer.timestep=timestep; 33 | layer.batchsize=batchsize; 34 | %n is the number of unrolled timesteps in one batch 35 | layer.n=batchsize*timestep; 36 | %Put x(t) and h(t) in one array 37 | layer.xh=ones([batchsize,dim+1+hiddensize,timestep+1],'single'); 38 | %W is the weights of all four gates and bias 39 | layer.weights_dim=dim+1+hiddensize; 40 | layer.W=(rand([layer.weights_dim,4*hiddensize],'single')-0.5)./100; 41 | %Compute the value of x_t*wx_t for all ts in one time 42 | layer.maX=zeros([batchsize,4*hiddensize,timestep],'single'); 43 | %value before activited 44 | layer.ma=layer.maX; 45 | %value activited 46 | layer.mb=layer.maX; 47 | %sc:state of cell 48 | layer.sc=zeros([batchsize,hiddensize,timestep],'single'); 49 | layer.bc=layer.sc; 50 | %The output tensor and error 51 | layer.output=zeros(layer.output_shape,'single'); 52 | layer.e=layer.sc; 53 | if layer.flag 54 | %diffs 55 | layer.dW=zeros(size(layer.W),'single'); 56 | layer.dma=zeros([batchsize,4*hiddensize,timestep+1],'single'); 57 | layer.dmb=layer.dma; 58 | layer.dsc=layer.sc; 59 | layer.dh=layer.dsc; 60 | end 61 | if ~strcmpi(layer.prelayer_type,'input')&&layer.flag 62 | layer.dx=zeros(layer.input_shape,'single'); 63 | end 64 | if nargin>4 65 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 66 | layer.loss=[]; 67 | end 68 | %% methods 69 | layer.act_f =@(x)act(x,'sigmoid'); % active function for gate 70 | layer.act_tc =@(x)act(x, 'tanh'); % active function for tc 71 | layer.act_h = @(x)act(x, 'tanh'); 72 | 73 | layer.dact_f= @(x)dact(x,'sigmoid'); 74 | layer.dact_tc =@(x)dact(x, 'tanh'); % active function for tc 75 | layer.dact_h = @(x)dact(x, 'tanh'); 76 | layer.ff=@(layer,prelayer)lstm_ff(layer,prelayer); 77 | layer.bp=@(layer,next_layer)lstm_bp(layer,next_layer); 78 | 79 | layer.configs.type=layer.type; 80 | layer.configs.input_shape=layer.input_shape; 81 | layer.configs.output_shape=layer.output_shape; 82 | layer.configs.hiddensize=layer.hiddensize; 83 | layer.configs.W=size(layer.W); 84 | end -------------------------------------------------------------------------------- /Layer/lstm/lstm_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=lstm_init_gpu(prelayer,hiddensize,return_sequence,flag,loss) 2 | %% Basic layer attributes 3 | %Input tensor sahpe 4 | layer.input_shape=prelayer.output_shape; 5 | layer.trainable=1; 6 | layer.flag=flag; 7 | 8 | dim=prelayer.output_shape(2); 9 | timestep=prelayer.output_shape(3); 10 | batchsize=prelayer.output_shape(1); 11 | if nargin<3 12 | return_sequence=1; 13 | end 14 | layer.return_sequence=return_sequence; 15 | if return_sequence 16 | %Output tensor shape 17 | layer.output_shape=[batchsize,hiddensize,timestep]; 18 | else 19 | layer.output_shape=[batchsize,hiddensize]; 20 | end 21 | %The type of the layer 22 | layer.type='lstm'; 23 | %conected layer type 24 | layer.prelayer_type=prelayer.type; 25 | %The hiddensize of the layer 26 | layer.hiddensize=hiddensize; 27 | 28 | layer.batch=1; 29 | layer.epoch=1; 30 | %% lstm layer attributes 31 | %Timestep 32 | layer.timestep=timestep; 33 | layer.batchsize=batchsize; 34 | %n is the number of unrolled timesteps in one batch 35 | layer.n=batchsize*timestep; 36 | %Put x(t) and h(t) in one array 37 | layer.xh=ones([batchsize,dim+1+hiddensize,timestep+1],'single','gpuArray'); 38 | %W is the weights of all four gates and bias 39 | layer.weights_dim=dim+1+hiddensize; 40 | layer.W=(rand([layer.weights_dim,4*hiddensize],'single','gpuArray')-0.5)./100; 41 | %Compute the value of x_t*wx_t for all ts in one time 42 | layer.maX=zeros([batchsize,4*hiddensize,timestep],'single','gpuArray'); 43 | %value before activited 44 | layer.ma=layer.maX; 45 | %value activited 46 | layer.mb=layer.maX; 47 | %sc:state of cell 48 | layer.sc=zeros([batchsize,hiddensize,timestep],'single','gpuArray'); 49 | layer.bc=layer.sc; 50 | %The output tensor and error 51 | layer.output=zeros(layer.output_shape,'single','gpuArray'); 52 | layer.e=layer.sc; 53 | if layer.flag 54 | %diffs 55 | layer.dW=zeros(size(layer.W),'single','gpuArray'); 56 | layer.dma=zeros([batchsize,4*hiddensize,timestep+1],'single','gpuArray'); 57 | layer.dmb=layer.dma; 58 | layer.dsc=layer.sc; 59 | layer.dh=layer.dsc; 60 | end 61 | if ~strcmpi(layer.prelayer_type,'input')&&layer.flag 62 | layer.dx=zeros(layer.input_shape,'single','gpuArray'); 63 | end 64 | if nargin>4 65 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 66 | layer.loss=[]; 67 | end 68 | %% methods 69 | layer.act_f =@(x)act(x,'sigmoid'); % active function for gate 70 | layer.act_tc =@(x)act(x, 'tanh'); % active function for tc 71 | layer.act_h = @(x)act(x, 'tanh'); 72 | 73 | layer.dact_f= @(x)dact(x,'sigmoid'); 74 | layer.dact_tc =@(x)dact(x, 'tanh'); % active function for tc 75 | layer.dact_h = @(x)dact(x, 'tanh'); 76 | layer.ff=@(layer,prelayer)lstm_ff(layer,prelayer); 77 | layer.bp=@(layer,next_layer)lstm_bp(layer,next_layer); 78 | 79 | layer.configs.type=layer.type; 80 | layer.configs.input_shape=layer.input_shape; 81 | layer.configs.output_shape=layer.output_shape; 82 | layer.configs.hiddensize=layer.hiddensize; 83 | layer.configs.W=size(layer.W); 84 | end -------------------------------------------------------------------------------- /Layer/tensor/tensor_init.m: -------------------------------------------------------------------------------- 1 | function layer=tensor_init(input_shape,type,loss) 2 | %% A tensor layer ,can be a input layer or a ouyput alyer 3 | %% Basic layer attributes 4 | %Input tensor sahpe 5 | layer.input_shape=input_shape; 6 | %Output tensor shape 7 | layer.output_shape=input_shape; 8 | %The type of the layer 9 | layer.type=type; 10 | if nargin>2 11 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 12 | layer.loss=[]; 13 | end 14 | layer.configs.type=layer.type; 15 | layer.configs.input_shape=layer.input_shape; 16 | layer.configs.output_shape=layer.output_shape; 17 | end -------------------------------------------------------------------------------- /Layer/tensor/tensor_init_cpu.m: -------------------------------------------------------------------------------- 1 | function layer=tensor_init_gpu(input_shape,type,loss) 2 | %% A tensor layer ,can be a input layer or a ouyput alyer 3 | %% Basic layer attributes 4 | %Input tensor sahpe 5 | layer.input_shape=input_shape; 6 | %Output tensor shape 7 | layer.output_shape=input_shape; 8 | %The type of the layer 9 | layer.type=type; 10 | if nargin>2 11 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 12 | layer.loss=[]; 13 | end 14 | layer.configs.type=layer.type; 15 | layer.configs.input_shape=layer.input_shape; 16 | layer.configs.output_shape=layer.output_shape; 17 | end -------------------------------------------------------------------------------- /Layer/tensor/tensor_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=tensor_init_gpu(input_shape,type,loss) 2 | %% A tensor layer ,can be a input layer or a ouyput alyer 3 | %% Basic layer attributes 4 | %Input tensor sahpe 5 | layer.input_shape=input_shape; 6 | %Output tensor shape 7 | layer.output_shape=input_shape; 8 | %The type of the layer 9 | layer.type=type; 10 | if nargin>2 11 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 12 | layer.loss=[]; 13 | end 14 | layer.configs.type=layer.type; 15 | layer.configs.input_shape=layer.input_shape; 16 | layer.configs.output_shape=layer.output_shape; 17 | end -------------------------------------------------------------------------------- /Layer/tensor_init_gpu.m: -------------------------------------------------------------------------------- 1 | function layer=tensor_init_gpu(input_shape,type,loss) 2 | %% A tensor layer ,can be a input layer or a ouyput alyer 3 | %% Basic layer attributes 4 | %Input tensor sahpe 5 | layer.input_shape=input_shape; 6 | %Output tensor shape 7 | layer.output_shape=input_shape; 8 | %The type of the layer 9 | layer.type=type; 10 | if nargin>2 11 | [layer.loss_f,layer.loss_df]=loss_handle(loss); 12 | layer.loss=[]; 13 | end 14 | layer.configs.type=layer.type; 15 | layer.configs.input_shape=layer.input_shape; 16 | layer.configs.output_shape=layer.output_shape; 17 | end -------------------------------------------------------------------------------- /Model/eval_loss.m: -------------------------------------------------------------------------------- 1 | function [outputlayer,loss]=eval_loss(outputlayer,y_true,flag) 2 | dim=size(y_true,2); 3 | loss=dim*feval(@(x)mean(x(:)),outputlayer.loss_f(single(y_true),outputlayer.output)); 4 | if flag 5 | outputlayer.loss=[outputlayer.loss,loss]; 6 | if isequal(outputlayer.type,'lstm')&& ~outputlayer.return_sequence 7 | outputlayer.e(:,:,end)=outputlayer.loss_df(y_true,outputlayer.output); 8 | else 9 | outputlayer.e=outputlayer.loss_df(single(y_true),outputlayer.output); 10 | end 11 | end 12 | end 13 | -------------------------------------------------------------------------------- /Model/layer_optimize.m: -------------------------------------------------------------------------------- 1 | function layer=layer_optimize(layer,pars,batch,epoch) 2 | if nargin <2 3 | pars.opt='sgd'; 4 | end 5 | switch pars.opt 6 | case 'sgd' 7 | if pars.momentum >0 8 | if batch==1 9 | layer.vW=pars.learningrate*layer.dW; 10 | else 11 | layer.vW=pars.momentum*layer.vW+pars.learningrate*layer.dW; 12 | end 13 | layer.W=layer.W-layer.vW; 14 | else 15 | layer.W=layer.W-pars.learningrate*layer.dW; 16 | end 17 | end 18 | layer.batch=batch; 19 | layer.epoch=epoch; 20 | end -------------------------------------------------------------------------------- /Model/model_evaluate.m: -------------------------------------------------------------------------------- 1 | function mean_loss=model_evaluate(model,x,y_true) 2 | y_pred=model.predict(model,x); 3 | dim=size(y_true,2); 4 | mean_loss=dim*feval(@(x)mean(x(:)),model.layers{end-1}.loss_f(single(y_true),y_pred)); 5 | end -------------------------------------------------------------------------------- /Model/model_init.m: -------------------------------------------------------------------------------- 1 | function model=model_init(input_shape,configs ,flag,optimizer,device) 2 | switch nargin 3 | case 2 4 | flag=0; 5 | device='cpu'; 6 | case 3 7 | if flag 8 | optimizer.type='sgd'; 9 | optimizer.momentum=0; 10 | optimizer.learningrate=0.01; 11 | end 12 | device='cpu'; 13 | case 4 14 | device='cpu'; 15 | end 16 | if nargin<3 17 | flag=0; 18 | end 19 | model.flag=flag; 20 | if nargin<4&&flag 21 | optimizer.type='sgd'; 22 | optimizer.momentum=0; 23 | optimizer.learningrate=0.01; 24 | end 25 | model.layers=cell(1,length(configs)+1); 26 | model.layers{1}=tensor_init(input_shape,'input'); 27 | switch device 28 | case 'cpu' 29 | for l=2:length(model.layers) 30 | model.layers{l}=layer_init_cpu(model.layers{l-1},configs{l-1},flag); 31 | end 32 | case 'gpu' 33 | for l=2:length(model.layers) 34 | model.layers{l}=layer_init_gpu(model.layers{l-1},configs{l-1},flag); 35 | end 36 | end 37 | model.layers=[model.layers,0]; 38 | for l=1:length(model.layers)-1 39 | disp(['layer ' ,num2str(l),' :']); 40 | disp(model.layers{l}.configs); 41 | end 42 | 43 | model.input_shape=model.layers{1}.input_shape(2:end); 44 | model.output_shape=model.layers{end-1}.output_shape(2:end); 45 | model.batchsize=input_shape(1); 46 | model.loss=[]; 47 | model.configs=configs; 48 | if flag 49 | model.optimizer=optimizer; 50 | model.optimize=@(layer,optimizer,batch,epoch)layer_optimize(layer,optimizer,batch,epoch); 51 | end 52 | model.eval_loss=@(outputlayer,y_true,flag)eval_loss(outputlayer,y_true,flag); 53 | model.predict=@(model,x)model_predict(model,x); 54 | model.save=@(model,filename)model_save(model,filename); 55 | model.evaluate=@(model,x,y_true)model_evaluate(model,x,y_true); 56 | if flag 57 | model.train=@(model,x,y,nb_epoch,verbose,filename)model_train(model,x,y,nb_epoch,verbose,filename); 58 | end 59 | end 60 | function layer=layer_init_gpu(prelayer,config,flag) 61 | switch config.type 62 | case 'lstm' 63 | if isfield(config,'loss') 64 | layer=lstm_init_gpu(prelayer,config.hiddensize,config.return_sequence,flag,config.loss); 65 | else 66 | layer=lstm_init_gpu(prelayer,config.hiddensize,config.return_sequence,flag); 67 | end 68 | case 'dense' 69 | if isfield(config,'loss') 70 | layer=dense_init_gpu(prelayer,config.hiddensize,flag,config.loss); 71 | else 72 | layer=dense_init_gpu(prelayer,config.hiddensize,flag); 73 | end 74 | case 'activation' 75 | if isfield(config,'loss') 76 | layer=activation_init_gpu(prelayer,config.act_fun,flag,config.loss); 77 | else 78 | layer=activation_init_gpu(prelayer,config.act_fun,flag); 79 | end 80 | case 'dropout' 81 | if isfield(config,'loss') 82 | layer=dropout_init_gpu(prelayer,config.drop_rate,flag,config.loss); 83 | else 84 | layer=dropout_init_gpu(prelayer,config.drop_rate,flag); 85 | end 86 | end 87 | end 88 | function layer=layer_init_cpu(prelayer,config,flag) 89 | switch config.type 90 | case 'lstm' 91 | if isfield(config,'loss') 92 | layer=lstm_init_cpu(prelayer,config.hiddensize,config.return_sequence,flag,config.loss); 93 | else 94 | layer=lstm_init_cpu(prelayer,config.hiddensize,config.return_sequence,flag); 95 | end 96 | case 'dense' 97 | if isfield(config,'loss') 98 | layer=dense_init_cpu(prelayer,config.hiddensize,flag,config.loss); 99 | else 100 | layer=dense_init_cpu(prelayer,config.hiddensize,flag); 101 | end 102 | case 'activation' 103 | if isfield(config,'loss') 104 | layer=activation_init_cpu(prelayer,config.act_fun,flag,config.loss); 105 | else 106 | layer=activation_init_cpu(prelayer,config.act_fun,flag); 107 | end 108 | case 'dropout' 109 | if isfield(config,'loss') 110 | layer=dropout_init_cpu(prelayer,config.drop_rate,flag,config.loss); 111 | else 112 | layer=dropout_init_cpu(prelayer,config.drop_rate,flag); 113 | end 114 | end 115 | end -------------------------------------------------------------------------------- /Model/model_load.m: -------------------------------------------------------------------------------- 1 | function model=model_load(minimodel,batch_size,flag,optimizer,device) 2 | if nargin<2 3 | batch_size=32; 4 | end 5 | if nargin<3 6 | flag=0; 7 | end 8 | if nargin<4&&flag 9 | optimizer.type='sgd'; 10 | optimizer.momentum=0; 11 | optimizer.learningrate=0.01; 12 | elseif ~flag 13 | optimizer=[]; 14 | end 15 | if isequal(class(minimodel),'char') 16 | load(minimodel); 17 | end 18 | model=model_init([batch_size,minimodel.input_shape],minimodel.configs,flag,optimizer,device); 19 | end -------------------------------------------------------------------------------- /Model/model_predict.m: -------------------------------------------------------------------------------- 1 | function y_pred=model_predict(model,x) 2 | batchsize=model.batchsize; 3 | shape_x=size(x); 4 | nb_batch=floor(shape_x(1)/batchsize); 5 | m=mod(shape_x(1),batchsize); 6 | y_pred=zeros([shape_x(1),model.output_shape],'single'); 7 | for batch=1:nb_batch 8 | %% ff 9 | if numel(shape_x)==2 10 | model.layers{1}=x((batch-1)*batchsize+1:batch*batchsize,:); 11 | elseif numel(shape_x)==3 12 | model.layers{1}=x((batch-1)*batchsize+1:batch*batchsize,:,:); 13 | else 14 | error('The number of dims of input data must be 2/3'); 15 | end 16 | for l=2:length(model.layers)-1 17 | model.layers{l}=model.layers{l}.ff(model.layers{l},model.layers{l-1}); 18 | end 19 | if numel(size(y_pred))>2 20 | y_pred((batch-1)*batchsize+1:batch*batchsize,:,:)=gather(model.layers{end-1}.output); 21 | else 22 | y_pred((batch-1)*batchsize+1:batch*batchsize,:)=gather(model.layers{end-1}.output); 23 | end 24 | end 25 | if m 26 | if numel(shape_x)==2 27 | model.layers{1}=x(end-batchsize+1:end,:); 28 | elseif numel(shape_x)==3 29 | model.layers{1}=x(end-batchsize+1:end,:,:); 30 | else 31 | error('The number of dims of input data must be 2/3'); 32 | end 33 | for l=2:length(model.layers)-1 34 | model.layers{l}=model.layers{l}.ff(model.layers{l},model.layers{l-1}); 35 | end 36 | if numel(size(y_pred))>2 37 | y_pred(end-batchsize+1:end,:,:)=gather(model.layers{end-1}.output); 38 | else 39 | y_pred(end-batchsize+1:end,:,:)=gather(model.layers{end-1}.output); 40 | end 41 | end 42 | end -------------------------------------------------------------------------------- /Model/model_save.m: -------------------------------------------------------------------------------- 1 | function model_save(model,filename) 2 | % if nargin<3 3 | % batchsize=model.input_shape(end); 4 | % end 5 | minimodel.input_shape=model.input_shape; 6 | minimodel.output_shape=model.output_shape; 7 | minimodel.configs=model.configs; 8 | for l=2:length(model.layers)-1 9 | if model.layers{l}.trainable 10 | minimodel.Ws{l}=gather(model.layers{l}.W); 11 | end 12 | end 13 | save(filename,'minimodel','-v7.3'); 14 | end 15 | 16 | -------------------------------------------------------------------------------- /Model/model_train.m: -------------------------------------------------------------------------------- 1 | function model=model_train(model,x,y,nb_epoch,verbose,filename) 2 | if nargin<5 3 | verbose=0; 4 | end 5 | if nargin<6 6 | filename=0; 7 | end 8 | batchsize=model.batchsize; 9 | shape_x=size(x); 10 | shape_y=size(y); 11 | g_batch=1; 12 | nb_batch=floor(shape_x(1)/batchsize)*nb_epoch; 13 | if verbose 14 | h = waitbar(g_batch/nb_batch,'Training model'); 15 | end 16 | model.epoch_loss=[]; 17 | model.batch_loss=[]; 18 | if verbose>=2 19 | f_epoch=figure('Name',' epochs loss'); 20 | f_batch=figure('Name',' batches loss'); 21 | end 22 | for epoch=1:nb_epoch 23 | batch=1; 24 | tic; 25 | epoch_batch_loss=[]; 26 | while batch*batchsize<=shape_x(1) 27 | %% ff 28 | if numel(shape_x)==2 29 | model.layers{1}=x((batch-1)*batchsize+1:batch*batchsize,:); 30 | elseif numel(shape_x)==3 31 | model.layers{1}=x((batch-1)*batchsize+1:batch*batchsize,:,:); 32 | else 33 | error('The number of dims of input data must be 2/3'); 34 | end 35 | for l=2:length(model.layers)-1 36 | model.layers{l}=model.layers{l}.ff(model.layers{l},model.layers{l-1}); 37 | end 38 | %% eval 39 | if numel(shape_y)==2 40 | model.layers{end-1}=model.eval_loss(model.layers{end-1},y((batch-1)*batchsize+1:batch*batchsize,:),model.flag); 41 | elseif numel(shape_y)==3 42 | model.layers{end-1}=model.eval_loss(model.layers{end-1},y((batch-1)*batchsize+1:batch*batchsize,:,:),model.flag); 43 | else 44 | error('The number of dims of output data must be 2/3'); 45 | end 46 | epoch_batch_loss=[epoch_batch_loss,model.layers{end-1}.loss(end)]; 47 | cu_epoch_loss=mean(epoch_batch_loss(:)); 48 | model.batch_loss=model.layers{end-1}.loss; 49 | if verbose>=3 50 | set(0,'CurrentFigure',f_batch); 51 | plot(model.batch_loss,'r-');hold off; 52 | end 53 | if verbose 54 | pro=num2str(100*g_batch/nb_batch); 55 | message=['Training model ','Epoch: ',num2str(epoch),'/',num2str(nb_epoch), ' Progress: ',pro,'%',' loss: ',num2str(cu_epoch_loss)]; 56 | waitbar(g_batch/nb_batch,h,message); 57 | end 58 | %% bp 59 | for l=length(model.layers)-1:-1:2 60 | if model.layers{l}.trainable 61 | model.layers{l}=model.optimize(model.layers{l}.bp(model.layers{l},model.layers{l+1}),model.optimizer,batch,epoch); 62 | else 63 | model.layers{l}=model.layers{l}.bp(model.layers{l},model.layers{l+1}); 64 | end 65 | end 66 | batch=batch+1; 67 | g_batch=g_batch+1; 68 | end 69 | toc 70 | model.epoch_loss=[model.epoch_loss,cu_epoch_loss]; 71 | if verbose>=2 72 | set(0,'CurrentFigure',f_epoch); 73 | plot(model.epoch_loss,'r-'); 74 | set(0,'CurrentFigure',f_batch); 75 | plot(model.batch_loss,'r-'); 76 | end 77 | end 78 | if filename 79 | model.save(model,filename); 80 | end 81 | delete(h); 82 | end -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # matDL 2 | ![matDL icon](./matDL_128X128.ico) 3 | v0.42 BETA 4 | A lightweight MATLAB deeplearning toolbox,based on gpuArray. 5 | One of the fastest matlab's RNN libs. 6 | ## Performance 7 | model:A LSTM model has [1024,1024,1024] hidensizes and 10 8 | timestep with a 256 dims input. 9 | Device: i7-4710hq,GTX940m 10 | matDL: 60sec/epoch Keras(1.2.2,Tensorflow backend,cudnn5.1): 29sec/epoch 11 | ## Features 12 | High parallel Implementation. 13 | 14 | 15 | * Concatance the weights of 4 gates to **W** and the values of **x** and **h** of every timesteps in a batch to a 3D tensor **xh**.Compute **x*W** for every timesteps of every samples in a batch at one time. 16 | * Compute the activated values of **input,forget ,ouput gates** at one time. 17 | 18 | OOP style 19 | * Use `struct` type to define a **layer** class and a **model** class.Define **ff**, **bp**, **optimize** methods by using a `FunctionHandle`. 20 | 21 | ## APIs 22 | ### Model 23 | * A `model` is a set of `layers`,`data` and `optimizer`. 24 | * build 25 | * `model=model_init(input_shape,configs ,flag,optimizer)` 26 | * arguments: 27 | * `input_shape` : a `vector`,`[input_dim,batchsize]` or `[input_dim,timestep,batchsize]` 28 | * `configs` : `cell` ,configures of each layers 29 | * `flag` : `bool` ,0 is predict model,1 is trrain model 30 | * `optimizer` : `struct` ,keywords: `opt`(type of optimizer) ,`learningrate` 31 | * attributes : 32 | * `model.input_shape` 33 | * `model.output_shape` 34 | * `model.batchsize` 35 | * `model.configs` 36 | * `model.flag` 37 | * `model.layers` 38 | * `model.optimizer` (if `flag`) 39 | * `model.loss` 40 | * methods: 41 | * private: 42 | * `model.eval_loss=@(outputlayer,y_true,flag)eval_loss(outputlayer,y_true,flag)` 43 | * `model.optimize=@(layer,optimizer,batch,epoch)layer_optimize(layer,optimizer,batch,epoch)` 44 | * public: 45 | * `model.train=@(model,x,y,nb_epoch,verbose,filename)model_train(model,x,y,nb_epoch,verbose,filename)` 46 | * `model=model.train(model,x,y,nb_epoch,verbose,filename)` 47 | * arguments: 48 | * `model` : self 49 | * `x`:input,shape:[dim,timestep,nb_samples],or [dim,nb_samples] 50 | * `y`:targets 51 | * `nb_epoch`: how many epochs you want to train 52 | * `verbose` :0,1,2,3,0 means no waitbar an figure,1 means showing waitbar only,2 means showing waitbar and plotting figures every epoch,3 means showing waitbar and plotting figures every epoch an batch. 53 | * `model.predict=@(model,x)model_predict(model,x)` 54 | * `y=model.predict(model,x)` 55 | * `model.evaluate=@(model,x,y_true)model_evaluate(model,x,y_true)` 56 | * `mean_loss=model.evaluate(model,x,y_true)` 57 | * `model.save=@(filename)model_save(model,filename)` 58 | * `model.save(filename)` 59 | * Save layers weigths and configs to a`.mat` file. 60 | * reload: 61 | * `model=model_load(minimodel,batch_size,flag,optimizer)` 62 | * `minimodel` is the minimodel saved by `model.save()`,can be a `struct` variable or a `string` of filename. 63 | * **example**: 64 | x=rand(100,10,3200,'single','gpuArray'); 65 | y=(zeros(512,10,3200'single','gpuArray')); 66 | y(1,:,:)=1; 67 | %% Define a model which has 2 lstm layers with 512 hiddenunits,and a timedistrbuted dense layer with 512 hiddenunits 68 | input_shape=[100,10,64];%input dim is 100,timestep is 10,batchsize is 64 69 | hiddensizes=[512,512,512]; 70 | for l=1:length(hiddensize) 71 | configs{l}.type='lstm'; 72 | configs{l}.hiddensize=hiddensize(l); 73 | configs{l}.return_sequence=1; 74 | end 75 | configs{l+1}.type='activation'; 76 | configs{l+1}.act_fun='softmax'; 77 | configs{l+1}.loss='categorical_cross_entropy'; 78 | optimizer.learningrate=0.1; 79 | optimizer.momentum=0.2; 80 | optimizer.opt='sgd'; 81 | model=model_init(input_shape,configs,1,optimizer); 82 | %% Train the model 83 | model=model.train(model,x,y,nb_epoch,3,'example/minimodel_f.mat'); 84 | or 85 | `test_lstm(50,[512,512,512],256,10,64,5);` 86 | 87 | 88 | ### Layers 89 | #### Layer class: 90 | * attributes: 91 | * `type` : `string`,type of the layer,available types:`input`,`dense`,`lstm`,`activation` 92 | * `prelayer_type` : `string`,type of the previous layer,available types:`input`,`dense`,`lstm`,`activation` 93 | * `trainable` : `bool`,is the layer trainable 94 | * `flag` : train model or predict model 95 | * `configs` :configures of the layer 96 | * `input_shape` : `vector`,`[input_dim,batchsize]` or `[input_dim,timestep,batchsize]` 97 | * `output_shape` : `vector`,`[hiddensize,batchsize]`or`[hiddensize,timestep,batchsize]` 98 | * `batch` : `int`,how many batches have been passed 99 | * `epoch` : same to `batch` 100 | * methods: 101 | * `layer=**layer_init(prelayer,loss,kwgrs)` 102 | * Built and init a layer.If the layer is a `input` layer,`prelayer` argument should be `input_shape` 103 | * `layer=layer.ff(layer,prelayer)` 104 | * `layer=layer.bp(layer,nextlayer)` 105 | ##### LSTM layer(layer) 106 | * `layer=lstm_init_gpu(prelayer,hiddensize,return_sequence,flag,loss)` 107 | * A LSTM(**Long-Short Term Memory unit - Hochreiter 1997**) layer,see [there]:http://deeplearning.net/tutorial/lstm.html for a step-by-step description of the algorithm. 108 | * aviliable configures: 109 | * `config.hiddensize` : `int`(`double`),number of hidden units(output dim) 110 | * `config.return_sequence` :`bool`(`double`),return sequences or not.if `return_sequences`,output will be a 3D tensor with shape (hiddensize,timestep,batchsize). Else ,a 2D tensor with shape (hiddensize,batchsize). 111 | * `config.loss` : `string`,type of loss function.Optional,only be used if the layer is an ouput layer. 112 | * **example** 113 | 114 | 115 | 116 | -------------------------------------------------------------------------------- /cudnn_LSTM_FF.m: -------------------------------------------------------------------------------- 1 | [e,n]=loadlibrary('C:\projects\mexcuda\matcudnn\mat_cudnn_test','mat_cudnn_test.h','includepath','C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\include','addheader','cudnn.h','addheader','cuda_runtime.h'); 2 | ax=ones([128,256,20],'single','gpuArray'); 3 | reserve=libpointer('voidPtr'); 4 | tic; 5 | for i=1:100 6 | calllib('mat_cudnn_test','MAT_CUDNN_RNN_LSTM_FF',ax,reserve); 7 | end 8 | toc; 9 | -------------------------------------------------------------------------------- /cumexhelp.h: -------------------------------------------------------------------------------- 1 | #ifndef CUMEXHELP 2 | #define CUMEXHELP 3 | #include 4 | #include 5 | #include 6 | #include 7 | #include 8 | #include 9 | #include "gpu/mxGPUArray.h" 10 | #endif -------------------------------------------------------------------------------- /example/char_rnn/Christ2FSM.m: -------------------------------------------------------------------------------- 1 | function fsmgospel=Christ2FSM(bibel) 2 | dict{1}={'god','God','LORD','Lord','lord','holy','Holy',' Amen','heaven','Heaven','hell','Hell','angle','Angle'... 3 | ,'demon','Demon','christ','Christ','water','Water'}; 4 | dict{2}={'monster','Monster','FSM','FSM','FSM','yummy','Yummy',' RAmen','plate','Plate','sewer','Sewer','ball','Ball'... 5 | ,'fork','Fork','pasta','Pasta','soip','Soup'}; 6 | if ~length(dict{1})==length(dict{2}) 7 | error('Keywords dict length unmatched'); 8 | end 9 | fsmgospel=bibel; 10 | for i=1:length(dict{1}) 11 | old=dict{1}{i}; 12 | new=dict{2}{i}; 13 | fsmgospel=strrep(fsmgospel,old,new); 14 | end -------------------------------------------------------------------------------- /example/char_rnn/char_rnn.m: -------------------------------------------------------------------------------- 1 | function char_rnn(data_filename,hiddensize,timestep,batch_size,nb_epoch) 2 | load(data_filename,'x'); 3 | load(data_filename,'y'); 4 | x=reshape(x(:,1:(timestep*batch_size)*floor(length(x)/(timestep*batch_size))),size(x,1),timestep,[]); 5 | y=reshape(y(:,1:(timestep*batch_size)*floor(length(y)/(timestep*batch_size))),size(y,1),timestep,[]); 6 | y=squeeze(y(:,end,:)); 7 | input_shape=[size(x,1),timestep,batch_size]; 8 | for l=1:length(hiddensize)-2 9 | configs{l}.type='lstm';configs{l}.hiddensize=hiddensize(l);configs{l}.return_sequence=1; 10 | end 11 | l=l+1; 12 | configs{l}.type='dropout';configs{l}.drop_rate=0.5; 13 | l=l+1; 14 | configs{l}.type='lstm';configs{l}.hiddensize=hiddensize(l-1);configs{l}.return_sequence=0; 15 | configs{l+1}.type='dense';configs{l+1}.hiddensize=hiddensize(l); 16 | configs{l+2}.type='activation';configs{l+2}.act_fun='softmax';configs{l+2}.loss='categorical_cross_entropy'; 17 | optimizer.learningrate=0.001; 18 | optimizer.momentum=0; 19 | optimizer.opt='sgd'; 20 | model=model_init(input_shape,configs,1,optimizer); 21 | profile on; 22 | model=model.train(model,x,y,nb_epoch,3,'example/minimodel_f.mat'); 23 | profile report; 24 | end 25 | -------------------------------------------------------------------------------- /example/char_rnn/textgenerate.m: -------------------------------------------------------------------------------- 1 | function text=textgenerate(model,dic,term) 2 | seed=double('In the beginning Monster created the plate and the earth.Ge1:2 And the earth was without form, and void; and darkness was upon the face of the deep.'); 3 | seed=seed(1:50) 4 | end 5 | function index=sample(pred,temp) 6 | pred=exp(log(double(pred))./temp); 7 | [~,index]=max(pred,[],1); 8 | end -------------------------------------------------------------------------------- /example/char_rnn/txt2seq.m: -------------------------------------------------------------------------------- 1 | function [x,y,dic]=txt2seq(text,threshold) 2 | if nargin<2 3 | threshold=50000; 4 | end 5 | if exist(text,'file') 6 | text=cell2mat(importdata(text)'); 7 | end 8 | [dic,~,index]=unique(double(text)); 9 | for i=1:length(dic) 10 | if numel(find(index==i))<=length(text)/threshold 11 | text(index==i)=','; 12 | end 13 | end 14 | [dic,~,index]=unique(double(text)); 15 | seq=zeros(length(dic),length(index),'int8'); 16 | for i=1:length(index) 17 | seq(index(i),i)=1; 18 | end 19 | x=seq(:,1:end-1); 20 | y=seq(:,2:end); 21 | end 22 | 23 | -------------------------------------------------------------------------------- /example/keras_test_lstm.py: -------------------------------------------------------------------------------- 1 | from keras.models import Sequential 2 | from keras.layers import LSTM 3 | import numpy as np 4 | import time 5 | def main(nb_batch=100,hiddensize=512,input_dim=100,timestep=10,batch_size=32,nb_epoch=1): 6 | x=np.ones((nb_batch*batch_size,timestep,input_dim)).astype('float32') 7 | y=np.ones((nb_batch*batch_size,timestep,hiddensize)) 8 | model = Sequential() 9 | model.add(LSTM(implementation=2,output_dim=hiddensize, input_shape=(timestep,input_dim),return_sequences=True)) 10 | #============================================================================== 11 | # model.add(LSTM(output_dim=hiddensize,return_sequences=True)) 12 | # model.add(LSTM(output_dim=hiddensize,return_sequences=True)) 13 | #============================================================================== 14 | model.compile(loss='mse',optimizer='sgd') 15 | start=time.time() 16 | model.predict(x=x,batch_size=batch_size) 17 | duration=time.time()-start 18 | print('Duration: ',duration,' sec') 19 | if __name__ == "__main__": 20 | main(nb_batch=100,hiddensize=512,input_dim=256,timestep=20,batch_size=128,nb_epoch=1) 21 | -------------------------------------------------------------------------------- /example/test_lstm.asv: -------------------------------------------------------------------------------- 1 | function test_lstm(nb_batch,hiddensizes,input_dim,timestep,batch_size,nb_epoch) 2 | optimizer.learningrate=0.01; 3 | optimizer.momentum=0; 4 | optimizer.opt='sgd'; 5 | x=rand(input_dim,timestep,batch_size*nb_batch,'single','gpuArray'); 6 | y=(zeros(hiddensizes(end),timestep,batch_size*nb_batch,'single','gpuArray')); 7 | y(1,:,:)=1; 8 | input_shape=[input_dim,timestep,batch_size]; 9 | for l=1:length(hiddensizes) 10 | configs{l}.type='lstm';configs{l}.hiddensize=hiddensizes(l);configs{l}.return_sequence=1; 11 | 12 | end 13 | configs{l+1}.type='activation';configs{l+1}.act_fun='softmax';configs{l+1}.loss='categorical_cross_entropy'; 14 | model=model_init(input_shape,configs,1,optimizer); 15 | profile on; 16 | model=model.train(x,y,nb_epoch,2,0); 17 | profile report; 18 | end -------------------------------------------------------------------------------- /example/test_lstm.m: -------------------------------------------------------------------------------- 1 | function test_lstm(nb_batch,hiddensizes,input_dim,timestep,batch_size,nb_epoch) 2 | optimizer.learningrate=0.01; 3 | optimizer.momentum=0.2; 4 | optimizer.opt='sgd'; 5 | x=sin(ones(batch_size*nb_batch,input_dim,timestep,'single','gpuArray')+5); 6 | y=(zeros(batch_size*nb_batch,hiddensizes(end),timestep,'single','gpuArray')); 7 | y(:,1,:)=1; 8 | input_shape=[batch_size,input_dim,timestep]; 9 | for l=1:length(hiddensizes) 10 | configs{l}.type='lstm';configs{l}.hiddensize=hiddensizes(l);configs{l}.return_sequence=1; 11 | end 12 | configs{l+1}.type='dropout';configs{l+1}.drop_rate=0.5; 13 | configs{l+2}.type='activation';configs{l+2}.act_fun='softmax';configs{l+2}.loss='categorical_cross_entropy'; 14 | model=model_init(input_shape,configs,1,optimizer); 15 | profile on; 16 | model=model.train(model,x,y,nb_epoch,3,0); 17 | %loss=model.evaluate(model,x,y); 18 | %disp(loss); 19 | profile report; 20 | end -------------------------------------------------------------------------------- /example/test_mlp.m: -------------------------------------------------------------------------------- 1 | function test_mlp(nb_batch,hiddensize,input_dim,batch_size,nb_epoch) 2 | input_shape=[batch_size,input_dim]; 3 | l=1; 4 | for i=1:length(hiddensize) 5 | configs{l}.type='dense';configs{l}.hiddensize=hiddensize(i); 6 | l=l+1; 7 | configs{l}.type='activation';configs{l}.act_fun='Relu'; 8 | l=l+1; 9 | configs{l}.type='dropout';configs{l}.drop_rate=0.5; 10 | end 11 | configs{l+1}.type='activation';configs{l+1}.act_fun='softmax';configs{l+1}.loss='categorical_cross_entropy'; 12 | optimizer.learningrate=0.01; 13 | optimizer.momentum=0.5; 14 | optimizer.opt='sgd'; 15 | model=model_init(input_shape,configs,1,optimizer); 16 | x=rand(batch_size*nb_batch,input_dim); 17 | y=(zeros(batch_size*nb_batch,hiddensize(end))); 18 | y(:,1,:)=1; 19 | model=model.train(model,x,y,nb_epoch,3,0);%not save 20 | loss=model.evaluate(model,x,y); 21 | disp(loss); 22 | y_pred=model.predict(model,x); 23 | end -------------------------------------------------------------------------------- /matDL_128X128.ico: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/QuantumLiu/matDL/ce75b21dd97e9d58c074d50915d5fc000ee46afb/matDL_128X128.ico -------------------------------------------------------------------------------- /mat_cudnn.cu: -------------------------------------------------------------------------------- 1 | #include"mat_cudnn.h" 2 | void GET_GPU_CONST_PTR(mxArray const *arrayPtr,float const *dataPtr) 3 | { 4 | dataPtr=(float const *)(mxGPUGetData(mxGPUCreateFromMxArray(arrayPtr))); 5 | } 6 | void GET_GPU_PTR(mxArray const *arrayPtr,float *dataPtr) 7 | { 8 | dataPtr=(float *)(mxGPUGetData(mxGPUCreateFromMxArray(arrayPtr))); 9 | } 10 | void MAT_CUDNN_LSTM_FF(mxArray const *x_array,mxArray const *w_array,void **reserveSpace,int* minibatch,int* hiddenSize,int* inputSize,int* seqLength ) 11 | { // ------------------------- 12 | // Create cudnn context 13 | // ------------------------- 14 | mxInitGPU(); 15 | cudnnHandle_t cudnnHandle; 16 | cudnnErrCheck(cudnnCreate(&cudnnHandle)); 17 | 18 | cudnnTensorDescriptor_t *xDesc, *yDesc, *dxDesc, *dyDesc; 19 | cudnnTensorDescriptor_t hxDesc, cxDesc; 20 | cudnnTensorDescriptor_t hyDesc, cyDesc; 21 | cudnnTensorDescriptor_t dhxDesc, dcxDesc; 22 | cudnnTensorDescriptor_t dhyDesc, dcyDesc; 23 | 24 | xDesc = (cudnnTensorDescriptor_t*)malloc(*seqLength * sizeof(cudnnTensorDescriptor_t)); 25 | yDesc = (cudnnTensorDescriptor_t*)malloc(*seqLength * sizeof(cudnnTensorDescriptor_t)); 26 | dxDesc = (cudnnTensorDescriptor_t*)malloc(*seqLength * sizeof(cudnnTensorDescriptor_t)); 27 | dyDesc = (cudnnTensorDescriptor_t*)malloc(*seqLength * sizeof(cudnnTensorDescriptor_t)); 28 | 29 | int dimA[3]; 30 | int strideA[3]; 31 | // In this example dimA[1] is constant across the whole sequence 32 | // This isn't required, all that is required is that it does not increase. 33 | for (int i = 0; i < *seqLength; i++) { 34 | cudnnErrCheck(cudnnCreateTensorDescriptor(&xDesc[i])); 35 | cudnnErrCheck(cudnnCreateTensorDescriptor(&yDesc[i])); 36 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dxDesc[i])); 37 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dyDesc[i])); 38 | 39 | dimA[0] = *miniBatch; 40 | dimA[1] = *inputSize; 41 | dimA[2] = 1; 42 | 43 | strideA[0] = dimA[2] * dimA[1]; 44 | strideA[1] = dimA[2]; 45 | strideA[2] = 1; 46 | 47 | cudnnErrCheck(cudnnSetTensorNdDescriptor(xDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 48 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dxDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 49 | 50 | dimA[0] = *miniBatch; 51 | dimA[1] = *hiddenSize; 52 | dimA[2] = 1; 53 | 54 | strideA[0] = dimA[2] * dimA[1]; 55 | strideA[1] = dimA[2]; 56 | strideA[2] = 1; 57 | 58 | cudnnErrCheck(cudnnSetTensorNdDescriptor(yDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 59 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dyDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 60 | } 61 | dimA[0] = 1; 62 | dimA[1] = *miniBatch; 63 | dimA[2] = *hiddenSize; 64 | 65 | strideA[0] = dimA[2] * dimA[1]; 66 | strideA[1] = dimA[2]; 67 | strideA[2] = 1; 68 | 69 | cudnnErrCheck(cudnnCreateTensorDescriptor(&hxDesc)); 70 | cudnnErrCheck(cudnnCreateTensorDescriptor(&cxDesc)); 71 | cudnnErrCheck(cudnnCreateTensorDescriptor(&hyDesc)); 72 | cudnnErrCheck(cudnnCreateTensorDescriptor(&cyDesc)); 73 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dhxDesc)); 74 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dcxDesc)); 75 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dhyDesc)); 76 | cudnnErrCheck(cudnnCreateTensorDescriptor(&dcyDesc)); 77 | 78 | cudnnErrCheck(cudnnSetTensorNdDescriptor(hxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 79 | cudnnErrCheck(cudnnSetTensorNdDescriptor(cxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 80 | cudnnErrCheck(cudnnSetTensorNdDescriptor(hyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 81 | cudnnErrCheck(cudnnSetTensorNdDescriptor(cyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 82 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dhxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 83 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dcxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 84 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dhyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 85 | cudnnErrCheck(cudnnSetTensorNdDescriptor(dcyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 86 | // ------------------------- 87 | // Set up the dropout descriptor (needed for the RNN descriptor) 88 | // ------------------------- 89 | unsigned long long seed = 1337ull; // Pick a seed. 90 | 91 | cudnnDropoutDescriptor_t dropoutDesc; 92 | cudnnErrCheck(cudnnCreateDropoutDescriptor(&dropoutDesc)); 93 | 94 | // How much memory does dropout need for states? 95 | // These states are used to generate random numbers internally 96 | // and should not be freed until the RNN descriptor is no longer used 97 | size_t stateSize; 98 | void *states; 99 | cudnnErrCheck(cudnnDropoutGetStatesSize(cudnnHandle, &stateSize)); 100 | 101 | cudaErrCheck(cudaMalloc(&states, stateSize)); 102 | float dropout=0; 103 | cudnnErrCheck(cudnnSetDropoutDescriptor(dropoutDesc, 104 | cudnnHandle, 105 | dropout, 106 | states, 107 | stateSize, 108 | seed)); 109 | 110 | // ------------------------- 111 | // Set up the RNN descriptor 112 | // ------------------------- 113 | cudnnRNNDescriptor_t rnnDesc; 114 | cudnnRNNMode_t RNNMode; 115 | 116 | cudnnErrCheck(cudnnCreateRNNDescriptor(&rnnDesc)); 117 | 118 | RNNMode = CUDNN_LSTM; 119 | 120 | cudnnErrCheck(cudnnSetRNNDescriptor(rnnDesc, 121 | hiddenSize, 122 | numLayers, 123 | dropoutDesc, 124 | CUDNN_LINEAR_INPUT, // We can also skip the input matrix transformation 125 | CUDNN_UNIDIRECTIONAL, 126 | RNNMode, 127 | CUDNN_DATA_FLOAT)); 128 | // ------------------------- 129 | // Set up parameters 130 | // ------------------------- 131 | // This needs to be done after the rnn descriptor is set as otherwise 132 | // we don't know how many parameters we have to allocate 133 | void *w; 134 | void *dw; 135 | 136 | cudnnFilterDescriptor_t wDesc, dwDesc; 137 | 138 | cudnnErrCheck(cudnnCreateFilterDescriptor(&wDesc)); 139 | cudnnErrCheck(cudnnCreateFilterDescriptor(&dwDesc)); 140 | 141 | size_t weightsSize; 142 | cudnnErrCheck(cudnnGetRNNParamsSize(cudnnHandle, rnnDesc, xDesc[0], &weightsSize, CUDNN_DATA_FLOAT)); 143 | 144 | int dimW[3]; 145 | dimW[0] = weightsSize / sizeof(float); 146 | dimW[1] = 1; 147 | dimW[2] = 1; 148 | 149 | cudnnErrCheck(cudnnSetFilterNdDescriptor(wDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dimW)); 150 | cudnnErrCheck(cudnnSetFilterNdDescriptor(dwDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dimW)); 151 | 152 | cudaErrCheck(cudaMalloc((void**)&w, weightsSize)); 153 | cudaErrCheck(cudaMalloc((void**)&dw, weightsSize)); 154 | 155 | 156 | // ------------------------- 157 | // Set up work space and reserved memory 158 | // ------------------------- 159 | void *workspace; 160 | 161 | size_t workSize; 162 | size_t reserveSize; 163 | 164 | // Need for every pass 165 | cudnnErrCheck(cudnnGetRNNWorkspaceSize(cudnnHandle, rnnDesc, seqLength, xDesc, &workSize)); 166 | // Only needed in training, shouldn't be touched between passes. 167 | cudnnErrCheck(cudnnGetRNNTrainingReserveSize(cudnnHandle, rnnDesc, seqLength, xDesc, &reserveSize)); 168 | 169 | cudaErrCheck(cudaMalloc((void**)&workspace, workSize)); 170 | cudaErrCheck(cudaMalloc((void**)&reserveSpace, reserveSize)); 171 | // Weights 172 | int numLinearLayers = 0; 173 | for (int layer = 0; layer < numLayers * (bidirectional ? 2 : 1); layer++) { 174 | for (int linLayerID = 0; linLayerID < numLinearLayers; linLayerID++) { 175 | cudnnFilterDescriptor_t linLayerMatDesc; 176 | cudnnErrCheck(cudnnCreateFilterDescriptor(&linLayerMatDesc)); 177 | float *linLayerMat; 178 | 179 | cudnnErrCheck(cudnnGetRNNLinLayerMatrixParams( cudnnHandle, 180 | rnnDesc, 181 | layer, 182 | xDesc[0], 183 | wDesc, 184 | w, 185 | linLayerID, 186 | linLayerMatDesc, 187 | (void**)&linLayerMat)); 188 | 189 | cudnnDataType_t dataType; 190 | cudnnTensorFormat_t format; 191 | int nbDims; 192 | int filterDimA[3]; 193 | cudnnErrCheck(cudnnGetFilterNdDescriptor(linLayerMatDesc, 194 | 3, 195 | &dataType, 196 | &format, 197 | &nbDims, 198 | filterDimA)); 199 | 200 | initGPUData(linLayerMat, filterDimA[0] * filterDimA[1] * filterDimA[2], 1.f / (float)(filterDimA[0] * filterDimA[1] * filterDimA[2])); 201 | 202 | cudnnErrCheck(cudnnDestroyFilterDescriptor(linLayerMatDesc)); 203 | 204 | cudnnFilterDescriptor_t linLayerBiasDesc; 205 | cudnnErrCheck(cudnnCreateFilterDescriptor(&linLayerBiasDesc)); 206 | float *linLayerBias; 207 | 208 | cudnnErrCheck(cudnnGetRNNLinLayerBiasParams( cudnnHandle, 209 | rnnDesc, 210 | layer, 211 | xDesc[0], 212 | wDesc, 213 | w, 214 | linLayerID, 215 | linLayerBiasDesc, 216 | (void**)&linLayerBias)); 217 | 218 | cudnnErrCheck(cudnnGetFilterNdDescriptor(linLayerBiasDesc, 219 | 3, 220 | &dataType, 221 | &format, 222 | &nbDims, 223 | filterDimA)); 224 | 225 | initGPUData(linLayerBias, filterDimA[0] * filterDimA[1] * filterDimA[2], 1.f); 226 | 227 | cudnnErrCheck(cudnnDestroyFilterDescriptor(linLayerBiasDesc)); 228 | } 229 | } 230 | // ********************************************************************************************************* 231 | // At this point all of the setup is done. We now need to pass through the RNN. 232 | // ********************************************************************************************************* 233 | 234 | 235 | 236 | cudaErrCheck(cudaDeviceSynchronize()); 237 | 238 | cudaEvent_t start, stop; 239 | float timeForward, timeBackward1, timeBackward2; 240 | cudaErrCheck(cudaEventCreate(&start)); 241 | cudaErrCheck(cudaEventCreate(&stop)); 242 | 243 | cudaErrCheck(cudaEventRecord(start)); 244 | 245 | // If we're not training we use this instead 246 | // cudnnErrCheck(cudnnRNNForwardInference(cudnnHandle, 247 | // rnnDesc, 248 | // xDesc, 249 | // x, 250 | // hxDesc, 251 | // hx, 252 | // cxDesc, 253 | // cx, 254 | // wDesc, 255 | // w, 256 | // yDesc, 257 | // y, 258 | // hyDesc, 259 | // hy, 260 | // cyDesc, 261 | // cy, 262 | // workspace, 263 | // workSize)); 264 | 265 | cudnnErrCheck(cudnnRNNForwardTraining(cudnnHandle, 266 | rnnDesc, 267 | seqLength, 268 | xDesc, 269 | x, 270 | hxDesc, 271 | hx, 272 | cxDesc, 273 | cx, 274 | wDesc, 275 | w, 276 | yDesc, 277 | y, 278 | hyDesc, 279 | hy, 280 | cyDesc, 281 | cy, 282 | workspace, 283 | workSize, 284 | reserveSpace, 285 | reserveSize)); 286 | 287 | 288 | } -------------------------------------------------------------------------------- /mat_cudnn.h: -------------------------------------------------------------------------------- 1 | #define EXPORT_FCNS 2 | #ifndef SHRHELP 3 | #include "shrhelp.h" 4 | #endif 5 | 6 | #include "cumexhelp.h" 7 | 8 | #ifndef MATCUDNN 9 | #define MATCUDNN 10 | #endif 11 | 12 | #ifdef __cplusplus 13 | #include 14 | extern "C" 15 | { 16 | #endif 17 | 18 | EXPORTED_FUNCTION void MAT_CUDNN_test(void* x); 19 | 20 | #ifdef __cplusplus 21 | } 22 | #endif 23 | -------------------------------------------------------------------------------- /mat_cudnn_test.cu: -------------------------------------------------------------------------------- 1 | #include "cumexhelp.h" 2 | #define EXPORT_FCNS 3 | #ifndef SHRHELP 4 | #include "shrhelp.h" 5 | #endif 6 | #include "mat_cudnn_test.h" 7 | 8 | #define cudaErrCheck(stat) { cudaErrCheck_((stat)); } 9 | void cudaErrCheck_(cudaError_t stat) { 10 | if (stat != cudaSuccess) { 11 | mexPrintf("CUDA Error: %s\n", cudaGetErrorString(stat)); 12 | mexErrMsgTxt("CUDA Error"); 13 | } 14 | } 15 | #define cudnnErrCheck(stat) { cudnnErrCheck_((stat)); } 16 | void cudnnErrCheck_(cudnnStatus_t stat) { 17 | if (stat != CUDNN_STATUS_SUCCESS) { 18 | mexPrintf( "cuDNN Error: %s\n", cudnnGetErrorString(stat)); 19 | mexErrMsgTxt("cuDNN Error"); 20 | } 21 | } 22 | __global__ void initGPUData_ker(float *data, int numElements, float value) { 23 | int tid = blockIdx.x * blockDim.x + threadIdx.x; 24 | if (tid < numElements) { 25 | data[tid] = value; 26 | } 27 | } 28 | void initGPUData(float *data, int numElements, float value) { 29 | dim3 gridDim; 30 | dim3 blockDim; 31 | 32 | blockDim.x = 1024; 33 | gridDim.x = (numElements + blockDim.x - 1) / blockDim.x; 34 | 35 | initGPUData_ker <<< gridDim, blockDim >>> (data, numElements, value); 36 | } 37 | // void GET_GPU_CONST_PTR(mxArray *arrayPtr,float const *dataPtr) 38 | // { 39 | // dataPtr=(float const *)(mxGPUGetDataReadOnly (mxGPUCreateFromMxArray(arrayPtr))); 40 | // } 41 | // void GET_GPU_PTR(mxArray *arrayPtr,float *dataPtr) 42 | // { 43 | // dataPtr=(float *)(mxGPUGetData(mxGPUCreateFromMxArray(arrayPtr))); 44 | // } 45 | 46 | EXPORTED_FUNCTION void MAT_CUDNN_test(void* x) 47 | { 48 | int(*seqLength)=10; 49 | int (*inputSize)=128; 50 | int (*miniBatch)=64; 51 | cudnnHandle_t cudnnHandle; 52 | cudnnErrCheck(cudnnCreate(&cudnnHandle)); 53 | cudaErrCheck(cudaMalloc((void**)&x,(*seqLength) * (*inputSize) * (*miniBatch) * sizeof(float))); 54 | cudnnDestroy(cudnnHandle); 55 | cudaFree(x); 56 | } 57 | EXPORTED_FUNCTION void MAT_CUDNN_RNN_LSTM_FF(mxArray const *ax,mxArray const *aw,mxArray *ah,mxArray *ac,int *hiddenSize,int *miniBatch，int *inputSize，int *seqLength,void *reserveSpace) 58 | { 59 | //int(*seqLength)=20; 60 | int numLayers=1; 61 | //int hiddenSize=256; 62 | //int (*inputSize)=128; 63 | //int (*miniBatch)=64; 64 | float dropout=0.0; 65 | bool bidirectional=0; 66 | int mode=2; 67 | cudnnHandle_t cudnnHandle; 68 | cudnnErrCheck(cudnnCreate(&cudnnHandle)); 69 | float const *x=(float const *)mxGPUGetDataReadOnly(mxGPUCreateFromMxArray(ax)); 70 | void *hx = NULL; 71 | void *cx = NULL; 72 | void *y; 73 | void *hy = NULL; 74 | void *cy = NULL; 75 | cudaErrCheck(cudaMalloc((void**)&hx, numLayers * (*hiddenSize) * (*miniBatch) * (bidirectional ? 2 : 1) * sizeof(float))); 76 | cudaErrCheck(cudaMalloc((void**)&cx, numLayers * (*hiddenSize) * (*miniBatch) * (bidirectional ? 2 : 1) * sizeof(float))); 77 | cudaErrCheck(cudaMalloc((void**)&y,(*seqLength) * (*hiddenSize) * (*miniBatch) * (bidirectional ? 2 : 1) * sizeof(float))); 78 | cudaErrCheck(cudaMalloc((void**)&hy, numLayers * (*hiddenSize) * (*miniBatch) * (bidirectional ? 2 : 1) * sizeof(float))); 79 | cudaErrCheck(cudaMalloc((void**)&cy, numLayers * (*hiddenSize) * (*miniBatch) * (bidirectional ? 2 : 1) * sizeof(float))); 80 | cudnnTensorDescriptor_t *xDesc, *yDesc; 81 | cudnnTensorDescriptor_t hxDesc, cxDesc; 82 | cudnnTensorDescriptor_t hyDesc, cyDesc; 83 | xDesc = (cudnnTensorDescriptor_t*)malloc((*seqLength) * sizeof(cudnnTensorDescriptor_t)); 84 | yDesc = (cudnnTensorDescriptor_t*)malloc((*seqLength) * sizeof(cudnnTensorDescriptor_t)); 85 | int dimA[3]; 86 | int strideA[3]; 87 | for (int i = 0; i <(*seqLength); i++) { 88 | cudnnErrCheck(cudnnCreateTensorDescriptor(&xDesc[i])); 89 | cudnnErrCheck(cudnnCreateTensorDescriptor(&yDesc[i])); 90 | 91 | dimA[0] = (*miniBatch); 92 | dimA[1] = (*inputSize); 93 | dimA[2] = 1; 94 | 95 | strideA[0] = dimA[2] * dimA[1]; 96 | strideA[1] = dimA[2]; 97 | strideA[2] = 1; 98 | 99 | cudnnErrCheck(cudnnSetTensorNdDescriptor(xDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 100 | 101 | dimA[0] = (*miniBatch); 102 | dimA[1] = bidirectional ? (*hiddenSize) * 2 : (*hiddenSize); 103 | dimA[2] = 1; 104 | 105 | strideA[0] = dimA[2] * dimA[1]; 106 | strideA[1] = dimA[2]; 107 | strideA[2] = 1; 108 | 109 | cudnnErrCheck(cudnnSetTensorNdDescriptor(yDesc[i], CUDNN_DATA_FLOAT, 3, dimA, strideA)); 110 | } 111 | 112 | 113 | dimA[0] = numLayers * (bidirectional ? 2 : 1); 114 | dimA[1] = (*miniBatch); 115 | dimA[2] = (*hiddenSize); 116 | 117 | strideA[0] = dimA[2] * dimA[1]; 118 | strideA[1] = dimA[2]; 119 | strideA[2] = 1; 120 | 121 | cudnnErrCheck(cudnnCreateTensorDescriptor(&hxDesc)); 122 | cudnnErrCheck(cudnnCreateTensorDescriptor(&cxDesc)); 123 | cudnnErrCheck(cudnnCreateTensorDescriptor(&hyDesc)); 124 | cudnnErrCheck(cudnnCreateTensorDescriptor(&cyDesc)); 125 | cudnnErrCheck(cudnnSetTensorNdDescriptor(hxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 126 | cudnnErrCheck(cudnnSetTensorNdDescriptor(cxDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 127 | cudnnErrCheck(cudnnSetTensorNdDescriptor(hyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 128 | cudnnErrCheck(cudnnSetTensorNdDescriptor(cyDesc, CUDNN_DATA_FLOAT, 3, dimA, strideA)); 129 | unsigned long long seed = 1337ull; // Pick a seed. 130 | 131 | cudnnDropoutDescriptor_t dropoutDesc; 132 | cudnnErrCheck(cudnnCreateDropoutDescriptor(&dropoutDesc)); 133 | 134 | // How much memory does dropout need for states? 135 | // These states are used to generate random numbers internally 136 | // and should not be freed until the RNN descriptor is no longer used 137 | size_t stateSize; 138 | void *states; 139 | cudnnErrCheck(cudnnDropoutGetStatesSize(cudnnHandle, &stateSize)); 140 | 141 | cudaErrCheck(cudaMalloc(&states, stateSize)); 142 | 143 | cudnnErrCheck(cudnnSetDropoutDescriptor(dropoutDesc, 144 | cudnnHandle, 145 | dropout, 146 | states, 147 | stateSize, 148 | seed)); 149 | 150 | // ------------------------- 151 | // Set up the RNN descriptor 152 | // ------------------------- 153 | cudnnRNNDescriptor_t rnnDesc; 154 | cudnnRNNMode_t RNNMode; 155 | 156 | cudnnErrCheck(cudnnCreateRNNDescriptor(&rnnDesc)); 157 | 158 | if (mode == 0) RNNMode = CUDNN_RNN_RELU; 159 | else if (mode == 1) RNNMode = CUDNN_RNN_TANH; 160 | else if (mode == 2) RNNMode = CUDNN_LSTM; 161 | else if (mode == 3) RNNMode = CUDNN_GRU; 162 | 163 | cudnnErrCheck(cudnnSetRNNDescriptor(rnnDesc, 164 | (*hiddenSize), 165 | numLayers, 166 | dropoutDesc, 167 | CUDNN_LINEAR_INPUT, // We can also skip the input matrix transformation 168 | bidirectional ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL, 169 | RNNMode, 170 | CUDNN_DATA_FLOAT)); 171 | void *w; 172 | cudnnFilterDescriptor_t wDesc; 173 | cudnnErrCheck(cudnnCreateFilterDescriptor(&wDesc)); 174 | size_t weightsSize; 175 | cudnnErrCheck(cudnnGetRNNParamsSize(cudnnHandle, rnnDesc, xDesc[0], &weightsSize, CUDNN_DATA_FLOAT)); 176 | 177 | int dimW[3]; 178 | dimW[0] = weightsSize / sizeof(float); 179 | dimW[1] = 1; 180 | dimW[2] = 1; 181 | 182 | cudnnErrCheck(cudnnSetFilterNdDescriptor(wDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, 3, dimW)); 183 | 184 | cudaErrCheck(cudaMalloc((void**)&w, weightsSize)); 185 | void *workspace; 186 | size_t workSize; 187 | size_t reserveSize; 188 | cudnnErrCheck(cudnnGetRNNWorkspaceSize(cudnnHandle, rnnDesc,(*seqLength), xDesc, &workSize)); 189 | // Only needed in training, shouldn't be touched between passes. 190 | cudnnErrCheck(cudnnGetRNNTrainingReserveSize(cudnnHandle, rnnDesc,(*seqLength), xDesc, &reserveSize)); 191 | 192 | cudaErrCheck(cudaMalloc((void**)&workspace, workSize)); 193 | cudaErrCheck(cudaMalloc((void**)&reserveSpace, reserveSize)); 194 | //if (hx != NULL) initGPUData((float*)hx, numLayers * hiddenSize * (*miniBatch) * (bidirectional ? 2 : 1), 1.f); 195 | //if (cx != NULL) initGPUData((float*)cx, numLayers * hiddenSize * (*miniBatch) * (bidirectional ? 2 : 1), 1.f); 196 | int numLinearLayers = 0; 197 | if (RNNMode == CUDNN_RNN_RELU || RNNMode == CUDNN_RNN_TANH) { 198 | numLinearLayers = 2; 199 | } 200 | else if (RNNMode == CUDNN_LSTM) { 201 | numLinearLayers = 8; 202 | } 203 | else if (RNNMode == CUDNN_GRU) { 204 | numLinearLayers = 6; 205 | } 206 | for (int layer = 0; layer < numLayers * (bidirectional ? 2 : 1); layer++) { 207 | for (int linLayerID = 0; linLayerID < numLinearLayers; linLayerID++) { 208 | cudnnFilterDescriptor_t linLayerMatDesc; 209 | cudnnErrCheck(cudnnCreateFilterDescriptor(&linLayerMatDesc)); 210 | float *linLayerMat; 211 | 212 | cudnnErrCheck(cudnnGetRNNLinLayerMatrixParams( cudnnHandle, 213 | rnnDesc, 214 | layer, 215 | xDesc[0], 216 | wDesc, 217 | w, 218 | linLayerID, 219 | linLayerMatDesc, 220 | (void**)&linLayerMat)); 221 | 222 | cudnnDataType_t dataType; 223 | cudnnTensorFormat_t format; 224 | int nbDims; 225 | int filterDimA[3]; 226 | cudnnErrCheck(cudnnGetFilterNdDescriptor(linLayerMatDesc, 227 | 3, 228 | &dataType, 229 | &format, 230 | &nbDims, 231 | filterDimA)); 232 | 233 | initGPUData(linLayerMat, filterDimA[0] * filterDimA[1] * filterDimA[2], 1.f / (float)(filterDimA[0] * filterDimA[1] * filterDimA[2])); 234 | 235 | cudnnErrCheck(cudnnDestroyFilterDescriptor(linLayerMatDesc)); 236 | 237 | cudnnFilterDescriptor_t linLayerBiasDesc; 238 | cudnnErrCheck(cudnnCreateFilterDescriptor(&linLayerBiasDesc)); 239 | float *linLayerBias; 240 | 241 | cudnnErrCheck(cudnnGetRNNLinLayerBiasParams( cudnnHandle, 242 | rnnDesc, 243 | layer, 244 | xDesc[0], 245 | wDesc, 246 | w, 247 | linLayerID, 248 | linLayerBiasDesc, 249 | (void**)&linLayerBias)); 250 | 251 | cudnnErrCheck(cudnnGetFilterNdDescriptor(linLayerBiasDesc, 252 | 3, 253 | &dataType, 254 | &format, 255 | &nbDims, 256 | filterDimA)); 257 | 258 | initGPUData(linLayerBias, filterDimA[0] * filterDimA[1] * filterDimA[2], 1.f); 259 | 260 | cudnnErrCheck(cudnnDestroyFilterDescriptor(linLayerBiasDesc)); 261 | } 262 | } 263 | cudaErrCheck(cudaDeviceSynchronize()); 264 | 265 | cudnnErrCheck(cudnnRNNForwardTraining(cudnnHandle, 266 | rnnDesc, 267 | (*seqLength), 268 | xDesc, 269 | x, 270 | hxDesc, 271 | hx, 272 | cxDesc, 273 | cx, 274 | wDesc, 275 | w, 276 | yDesc, 277 | y, 278 | hyDesc, 279 | hy, 280 | cyDesc, 281 | cy, 282 | workspace, 283 | workSize, 284 | reserveSpace, 285 | reserveSize)); 286 | 287 | 288 | 289 | cudaFree(hx); 290 | cudaFree(cx); 291 | cudaFree(y); 292 | cudaFree(hy); 293 | cudaFree(cy); 294 | cudaFree(workspace); 295 | cudaFree(reserveSpace); 296 | cudaFree(w); 297 | cudnnDestroy(cudnnHandle); 298 | 299 | } 300 | void mexFunction( int nlhs, mxArray *plhs[], 301 | int nrhs, const mxArray*prhs[] ) 302 | { 303 | } 304 | -------------------------------------------------------------------------------- /mat_cudnn_test.h: -------------------------------------------------------------------------------- 1 | #include "shrhelp.h" 2 | #ifndef CUMEXHELP 3 | #include 4 | #endif 5 | #ifdef __cplusplus 6 | #include 7 | extern "C" 8 | { 9 | #endif 10 | EXPORTED_FUNCTION void MAT_CUDNN_test(void* x); 11 | EXPORTED_FUNCTION void MAT_CUDNN_RNN_LSTM_FF(mxArray const *ax,void *reserveSpace); 12 | #ifdef __cplusplus 13 | } 14 | #endif -------------------------------------------------------------------------------- /mat_cudnn_test.mexw64: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/QuantumLiu/matDL/ce75b21dd97e9d58c074d50915d5fc000ee46afb/mat_cudnn_test.mexw64 --------------------------------------------------------------------------------