| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | module athena__batchnorm1d_layer | ||
| 2 | !! Module containing implementation of 0D and 1D batch normalisation layers | ||
| 3 | !! | ||
| 4 | !! This module implements batch normalisation for 3D convolutional layers, | ||
| 5 | !! normalizing activations across the batch dimension. | ||
| 6 | !! | ||
| 7 | !! Mathematical operation (training): | ||
| 8 | !! \[ \mu_\mathcal{B} = \frac{1}{m}\sum_{i=1}^{m} x_i \] | ||
| 9 | !! \[ \sigma^2_\mathcal{B} = \frac{1}{m}\sum_{i=1}^{m} (x_i - \mu_\mathcal{B})^2 \] | ||
| 10 | !! \[ \hat{x}_i = \frac{x_i - \mu_\mathcal{B}}{\sqrt{\sigma^2_\mathcal{B} + \epsilon}} \] | ||
| 11 | !! \[ y_i = \gamma \hat{x}_i + \beta \] | ||
| 12 | !! | ||
| 13 | !! where \(\gamma, \beta\) are learnable parameters, \(\epsilon\) is stability constant | ||
| 14 | !! | ||
| 15 | !! Inference: uses running statistics | ||
| 16 | !! \(\mu_{\text{running}}, \sigma^2_{\text{running}}\) from training | ||
| 17 | !! | ||
| 18 | !! Benefits: Reduces internal covariate shift, enables higher learning rates, | ||
| 19 | !! acts as regularisation, reduces dependence on initialisation | ||
| 20 | !! Reference: Ioffe & Szegedy (2015), ICML | ||
| 21 | use coreutils, only: real32, stop_program, print_warning | ||
| 22 | use athena__base_layer, only: batch_layer_type, base_layer_type | ||
| 23 | use athena__misc_types, only: base_init_type, & | ||
| 24 | onnx_node_type, onnx_initialiser_type, onnx_tensor_type | ||
| 25 | use diffstruc, only: array_type | ||
| 26 | use athena__diffstruc_extd, only: batchnorm_array_type, & | ||
| 27 | batchnorm, batchnorm_inference | ||
| 28 | implicit none | ||
| 29 | |||
| 30 | |||
| 31 | private | ||
| 32 | |||
| 33 | public :: batchnorm1d_layer_type | ||
| 34 | public :: read_batchnorm1d_layer | ||
| 35 | |||
| 36 | |||
| 37 | type, extends(batch_layer_type) :: batchnorm1d_layer_type | ||
| 38 | !! Type for 0D or 1D batch normalisation layer with overloaded procedures | ||
| 39 | contains | ||
| 40 | procedure, pass(this) :: set_hyperparams => set_hyperparams_batchnorm1d | ||
| 41 | !! Set hyperparameters for 1D batch normalisation layer | ||
| 42 | procedure, pass(this) :: read => read_batchnorm1d | ||
| 43 | !! Read 1D batch normalisation layer from file | ||
| 44 | |||
| 45 | procedure, pass(this) :: forward => forward_batchnorm1d | ||
| 46 | !! Forward propagation derived type handler | ||
| 47 | |||
| 48 | final :: finalise_batchnorm1d | ||
| 49 | !! Finalise 1D batch normalisation layer | ||
| 50 | end type batchnorm1d_layer_type | ||
| 51 | |||
| 52 | |||
| 53 | interface batchnorm1d_layer_type | ||
| 54 | !! Interface for setting up the 1D batch normalisation layer | ||
| 55 | module function layer_setup( & | ||
| 56 | input_shape, & | ||
| 57 | num_channels, num_inputs, & | ||
| 58 | momentum, epsilon, & | ||
| 59 | gamma_init_mean, gamma_init_std, & | ||
| 60 | beta_init_mean, beta_init_std, & | ||
| 61 | gamma_initialiser, beta_initialiser, & | ||
| 62 | moving_mean_initialiser, moving_variance_initialiser, & | ||
| 63 | verbose & | ||
| 64 | ) result(layer) | ||
| 65 | !! Set up the 1D batch normalisation layer | ||
| 66 | integer, dimension(:), optional, intent(in) :: input_shape | ||
| 67 | !! Input shape | ||
| 68 | integer, optional, intent(in) :: num_channels, num_inputs | ||
| 69 | !! Number of channels and inputs | ||
| 70 | real(real32), optional, intent(in) :: momentum, epsilon | ||
| 71 | !! Momentum and epsilon | ||
| 72 | real(real32), optional, intent(in) :: gamma_init_mean, gamma_init_std | ||
| 73 | !! Gamma initialisation mean and standard deviation | ||
| 74 | real(real32), optional, intent(in) :: beta_init_mean, beta_init_std | ||
| 75 | !! Beta initialisation mean and standard deviation | ||
| 76 | class(*), optional, intent(in) :: & | ||
| 77 | gamma_initialiser, beta_initialiser, & | ||
| 78 | moving_mean_initialiser, moving_variance_initialiser | ||
| 79 | !! Initialisers | ||
| 80 | integer, optional, intent(in) :: verbose | ||
| 81 | !! Verbosity level | ||
| 82 | type(batchnorm1d_layer_type) :: layer | ||
| 83 | !! Instance of the 1D batch normalisation layer | ||
| 84 | end function layer_setup | ||
| 85 | end interface batchnorm1d_layer_type | ||
| 86 | |||
| 87 | |||
| 88 | |||
| 89 | contains | ||
| 90 | |||
| 91 | !############################################################################### | ||
| 92 | − | subroutine finalise_batchnorm1d(this) | |
| 93 | !! Finalise 1D batch normalisation layer | ||
| 94 | implicit none | ||
| 95 | |||
| 96 | ! Arguments | ||
| 97 | type(batchnorm1d_layer_type), intent(inout) :: this | ||
| 98 | !! Instance of the 1D batch normalisation layer | ||
| 99 | |||
| 100 | − | if(allocated(this%mean)) deallocate(this%mean) | |
| 101 | − | if(allocated(this%variance)) deallocate(this%variance) | |
| 102 | − | if(allocated(this%input_shape)) deallocate(this%input_shape) | |
| 103 | − | if(allocated(this%output)) deallocate(this%output) | |
| 104 | |||
| 105 | − | end subroutine finalise_batchnorm1d | |
| 106 | !############################################################################### | ||
| 107 | |||
| 108 | |||
| 109 | !##############################################################################! | ||
| 110 | ! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * ! | ||
| 111 | !##############################################################################! | ||
| 112 | |||
| 113 | |||
| 114 | !############################################################################### | ||
| 115 | − | module function layer_setup( & | |
| 116 | − | input_shape, & | |
| 117 | num_channels, num_inputs, & | ||
| 118 | momentum, epsilon, & | ||
| 119 | gamma_init_mean, gamma_init_std, & | ||
| 120 | beta_init_mean, beta_init_std, & | ||
| 121 | gamma_initialiser, beta_initialiser, & | ||
| 122 | moving_mean_initialiser, moving_variance_initialiser, & | ||
| 123 | verbose & | ||
| 124 | − | ) result(layer) | |
| 125 | !! Set up the 1D batch normalisation layer | ||
| 126 | use athena__initialiser, only: initialiser_setup | ||
| 127 | implicit none | ||
| 128 | |||
| 129 | ! Arguments | ||
| 130 | integer, dimension(:), optional, intent(in) :: input_shape | ||
| 131 | !! Input shape | ||
| 132 | integer, optional, intent(in) :: num_channels, num_inputs | ||
| 133 | !! Number of channels and inputs | ||
| 134 | real(real32), optional, intent(in) :: momentum, epsilon | ||
| 135 | !! Momentum and epsilon | ||
| 136 | real(real32), optional, intent(in) :: gamma_init_mean, gamma_init_std | ||
| 137 | !! Gamma initialisation mean and standard deviation | ||
| 138 | real(real32), optional, intent(in) :: beta_init_mean, beta_init_std | ||
| 139 | !! Beta initialisation mean and standard deviation | ||
| 140 | class(*), optional, intent(in) :: & | ||
| 141 | gamma_initialiser, beta_initialiser, & | ||
| 142 | moving_mean_initialiser, moving_variance_initialiser | ||
| 143 | !! Initialisers | ||
| 144 | integer, optional, intent(in) :: verbose | ||
| 145 | !! Verbosity level | ||
| 146 | |||
| 147 | type(batchnorm1d_layer_type) :: layer | ||
| 148 | !! Instance of the 1D batch normalisation layer | ||
| 149 | |||
| 150 | ! Local variables | ||
| 151 | integer :: verbose_ = 0 | ||
| 152 | !! Verbosity level | ||
| 153 | character(256) :: err_msg | ||
| 154 | !! Error message | ||
| 155 | class(base_init_type), allocatable :: & | ||
| 156 | − | gamma_initialiser_, beta_initialiser_, & | |
| 157 | − | moving_mean_initialiser_, moving_variance_initialiser_ | |
| 158 | !! Initialisers | ||
| 159 | |||
| 160 | |||
| 161 | − | if(present(verbose)) verbose_ = verbose | |
| 162 | |||
| 163 | !--------------------------------------------------------------------------- | ||
| 164 | ! Set up momentum and epsilon | ||
| 165 | !--------------------------------------------------------------------------- | ||
| 166 | − | if(present(momentum))then | |
| 167 | − | layer%momentum = momentum | |
| 168 | else | ||
| 169 | − | layer%momentum = 0._real32 | |
| 170 | end if | ||
| 171 | − | if(present(epsilon))then | |
| 172 | − | layer%epsilon = epsilon | |
| 173 | else | ||
| 174 | − | layer%epsilon = 1.E-5_real32 | |
| 175 | end if | ||
| 176 | |||
| 177 | |||
| 178 | !--------------------------------------------------------------------------- | ||
| 179 | ! Set up initialiser mean and standard deviations | ||
| 180 | !--------------------------------------------------------------------------- | ||
| 181 | − | if(present(gamma_init_mean)) layer%gamma_init_mean = gamma_init_mean | |
| 182 | − | if(present(gamma_init_std)) layer%gamma_init_std = gamma_init_std | |
| 183 | − | if(present(beta_init_mean)) layer%beta_init_mean = beta_init_mean | |
| 184 | − | if(present(beta_init_std)) layer%beta_init_std = beta_init_std | |
| 185 | |||
| 186 | |||
| 187 | !--------------------------------------------------------------------------- | ||
| 188 | ! Define gamma and beta initialisers | ||
| 189 | !--------------------------------------------------------------------------- | ||
| 190 | − | if(present(gamma_initialiser))then | |
| 191 | − | gamma_initialiser_ = initialiser_setup(gamma_initialiser) | |
| 192 | end if | ||
| 193 | − | if(present(beta_initialiser))then | |
| 194 | − | beta_initialiser_ = initialiser_setup(beta_initialiser) | |
| 195 | end if | ||
| 196 | − | if(present(moving_mean_initialiser))then | |
| 197 | − | moving_mean_initialiser_ = initialiser_setup(moving_mean_initialiser) | |
| 198 | end if | ||
| 199 | − | if(present(moving_variance_initialiser))then | |
| 200 | − | moving_variance_initialiser_ = initialiser_setup(moving_variance_initialiser) | |
| 201 | end if | ||
| 202 | |||
| 203 | |||
| 204 | !--------------------------------------------------------------------------- | ||
| 205 | ! Set hyperparameters | ||
| 206 | !--------------------------------------------------------------------------- | ||
| 207 | call layer%set_hyperparams( & | ||
| 208 | momentum = layer%momentum, epsilon = layer%epsilon, & | ||
| 209 | gamma_init_mean = layer%gamma_init_mean, & | ||
| 210 | gamma_init_std = layer%gamma_init_std, & | ||
| 211 | beta_init_mean = layer%beta_init_mean, & | ||
| 212 | beta_init_std = layer%beta_init_std, & | ||
| 213 | gamma_initialiser = gamma_initialiser_, & | ||
| 214 | beta_initialiser = beta_initialiser_, & | ||
| 215 | moving_mean_initialiser = moving_mean_initialiser_, & | ||
| 216 | moving_variance_initialiser = moving_variance_initialiser_, & | ||
| 217 | verbose = verbose_ & | ||
| 218 | − | ) | |
| 219 | |||
| 220 | |||
| 221 | !--------------------------------------------------------------------------- | ||
| 222 | ! Initialise layer shape | ||
| 223 | !--------------------------------------------------------------------------- | ||
| 224 | − | if(present(input_shape))then | |
| 225 | − | if(present(num_channels).or.present(num_inputs))then | |
| 226 | write(err_msg,'(A)') & | ||
| 227 | "both input_shape and num_channels/num_inputs present" // & | ||
| 228 | achar(13) // achar(10) // & | ||
| 229 | − | "These represent the same parameter, so are conflicting" | |
| 230 | − | call stop_program(err_msg) | |
| 231 | − | return | |
| 232 | end if | ||
| 233 | − | if(size(input_shape).eq.1)then | |
| 234 | − | call layer%init(input_shape= [ 1, input_shape ] ) | |
| 235 | else | ||
| 236 | − | call layer%init(input_shape= input_shape) | |
| 237 | end if | ||
| 238 | − | elseif(present(num_channels).and.present(num_inputs))then | |
| 239 | − | call layer%init(input_shape=[num_inputs, num_channels]) | |
| 240 | − | elseif(present(num_channels))then | |
| 241 | − | call layer%init(input_shape=[1, num_channels]) | |
| 242 | − | elseif(present(num_inputs))then | |
| 243 | − | call layer%init(input_shape=[num_inputs, 1]) | |
| 244 | end if | ||
| 245 | |||
| 246 | − | end function layer_setup | |
| 247 | !############################################################################### | ||
| 248 | |||
| 249 | |||
| 250 | !############################################################################### | ||
| 251 | − | subroutine set_hyperparams_batchnorm1d( & | |
| 252 | this, & | ||
| 253 | momentum, epsilon, & | ||
| 254 | gamma_init_mean, gamma_init_std, & | ||
| 255 | beta_init_mean, beta_init_std, & | ||
| 256 | gamma_initialiser, beta_initialiser, & | ||
| 257 | moving_mean_initialiser, moving_variance_initialiser, & | ||
| 258 | verbose ) | ||
| 259 | !! Set hyperparameters for 1D batch normalisation layer | ||
| 260 | use athena__initialiser, only: initialiser_setup | ||
| 261 | implicit none | ||
| 262 | |||
| 263 | ! Arguments | ||
| 264 | class(batchnorm1d_layer_type), intent(inout) :: this | ||
| 265 | !! Instance of the 1D batch normalisation layer | ||
| 266 | real(real32), intent(in) :: momentum, epsilon | ||
| 267 | !! Momentum and epsilon | ||
| 268 | real(real32), intent(in) :: gamma_init_mean, gamma_init_std | ||
| 269 | !! Gamma initialisation mean and standard deviation | ||
| 270 | real(real32), intent(in) :: beta_init_mean, beta_init_std | ||
| 271 | !! Beta initialisation mean and standard deviation | ||
| 272 | class(base_init_type), allocatable, intent(in) :: & | ||
| 273 | gamma_initialiser, beta_initialiser | ||
| 274 | !! Gamma and beta initialisers | ||
| 275 | class(base_init_type), allocatable, intent(in) :: & | ||
| 276 | moving_mean_initialiser, moving_variance_initialiser | ||
| 277 | !! Moving mean and variance initialisers | ||
| 278 | integer, optional, intent(in) :: verbose | ||
| 279 | !! Verbosity level | ||
| 280 | |||
| 281 | |||
| 282 | − | this%name = "batchnorm1d" | |
| 283 | − | this%type = "batc" | |
| 284 | − | this%input_rank = 2 | |
| 285 | − | this%output_rank = 2 | |
| 286 | − | this%use_bias = .true. | |
| 287 | − | this%momentum = momentum | |
| 288 | − | this%epsilon = epsilon | |
| 289 | − | if(allocated(this%kernel_init)) deallocate(this%kernel_init) | |
| 290 | − | if(.not.allocated(gamma_initialiser))then | |
| 291 | − | this%kernel_init = initialiser_setup('ones') | |
| 292 | else | ||
| 293 | − | allocate(this%kernel_init, source=gamma_initialiser) | |
| 294 | end if | ||
| 295 | − | if(allocated(this%bias_init)) deallocate(this%bias_init) | |
| 296 | − | if(.not.allocated(beta_initialiser))then | |
| 297 | − | this%bias_init = initialiser_setup('zeros') | |
| 298 | else | ||
| 299 | − | allocate(this%bias_init, source=beta_initialiser) | |
| 300 | end if | ||
| 301 | − | if(.not.allocated(moving_mean_initialiser))then | |
| 302 | − | this%moving_mean_init = initialiser_setup('zeros') | |
| 303 | else | ||
| 304 | − | this%moving_mean_init = moving_mean_initialiser | |
| 305 | end if | ||
| 306 | − | if(.not.allocated(moving_variance_initialiser))then | |
| 307 | − | this%moving_variance_init = initialiser_setup('ones') | |
| 308 | else | ||
| 309 | − | this%moving_variance_init = moving_variance_initialiser | |
| 310 | end if | ||
| 311 | − | this%gamma_init_mean = gamma_init_mean | |
| 312 | − | this%gamma_init_std = gamma_init_std | |
| 313 | − | this%beta_init_mean = beta_init_mean | |
| 314 | − | this%beta_init_std = beta_init_std | |
| 315 | − | this%kernel_init%mean = this%gamma_init_mean | |
| 316 | − | this%kernel_init%std = this%gamma_init_std | |
| 317 | − | this%bias_init%mean = this%beta_init_mean | |
| 318 | − | this%bias_init%std = this%beta_init_std | |
| 319 | − | if(present(verbose))then | |
| 320 | − | if(abs(verbose).gt.0)then | |
| 321 | write(*,'("BATCHNORM1D gamma (kernel) initialiser: ",A)') & | ||
| 322 | − | trim(this%kernel_init%name) | |
| 323 | write(*,'("BATCHNORM1D beta (bias) initialiser: ",A)') & | ||
| 324 | − | trim(this%bias_init%name) | |
| 325 | write(*,'("BATCHNORM1D moving mean initialiser: ",A)') & | ||
| 326 | − | trim(this%moving_mean_init%name) | |
| 327 | write(*,'("BATCHNORM1D moving variance initialiser: ",A)') & | ||
| 328 | − | trim(this%moving_variance_init%name) | |
| 329 | end if | ||
| 330 | end if | ||
| 331 | |||
| 332 | − | end subroutine set_hyperparams_batchnorm1d | |
| 333 | !############################################################################### | ||
| 334 | |||
| 335 | |||
| 336 | !##############################################################################! | ||
| 337 | ! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * ! | ||
| 338 | !##############################################################################! | ||
| 339 | |||
| 340 | |||
| 341 | !############################################################################### | ||
| 342 | − | subroutine read_batchnorm1d(this, unit, verbose) | |
| 343 | !! Read 1D batch normalisation layer from file | ||
| 344 | use athena__tools_infile, only: assign_val, assign_vec, move | ||
| 345 | use coreutils, only: to_lower, to_upper, icount | ||
| 346 | use athena__initialiser, only: initialiser_setup | ||
| 347 | implicit none | ||
| 348 | |||
| 349 | ! Arguments | ||
| 350 | class(batchnorm1d_layer_type), intent(inout) :: this | ||
| 351 | !! Instance of the 1D batch normalisation layer | ||
| 352 | integer, intent(in) :: unit | ||
| 353 | !! File unit | ||
| 354 | integer, optional, intent(in) :: verbose | ||
| 355 | !! Verbosity level | ||
| 356 | |||
| 357 | ! Local variables | ||
| 358 | integer :: stat, verbose_ = 0 | ||
| 359 | !! File status and verbosity level | ||
| 360 | integer :: i, j, k, c, itmp1, iline, final_line | ||
| 361 | !! Temporary integers and loop indices | ||
| 362 | integer :: num_channels | ||
| 363 | !! Number of channels | ||
| 364 | real(real32) :: momentum = 0._real32, epsilon = 1.E-5_real32 | ||
| 365 | !! Momentum and epsilon | ||
| 366 | − | class(base_init_type), allocatable :: gamma_initialiser, beta_initialiser | |
| 367 | !! Initialisers | ||
| 368 | class(base_init_type), allocatable :: & | ||
| 369 | − | moving_mean_initialiser, moving_variance_initialiser | |
| 370 | !! Moving mean and variance initialisers | ||
| 371 | character(14) :: gamma_initialiser_name='', beta_initialiser_name='' | ||
| 372 | !! Initialisers | ||
| 373 | character(14) :: & | ||
| 374 | moving_mean_initialiser_name='', & | ||
| 375 | moving_variance_initialiser_name='' | ||
| 376 | !! Moving mean and variance initialisers | ||
| 377 | character(256) :: buffer, tag, err_msg | ||
| 378 | !! Buffer, tag, and error message | ||
| 379 | integer, dimension(2) :: input_shape | ||
| 380 | !! Input shape | ||
| 381 | − | real(real32), allocatable, dimension(:) :: data_list | |
| 382 | !! Data list | ||
| 383 | integer, dimension(2) :: param_lines | ||
| 384 | !! Lines where parameters are found | ||
| 385 | |||
| 386 | |||
| 387 | ! Initialise optional arguments | ||
| 388 | !--------------------------------------------------------------------------- | ||
| 389 | − | if(present(verbose)) verbose_ = verbose | |
| 390 | |||
| 391 | |||
| 392 | ! Loop over tags in layer card | ||
| 393 | !--------------------------------------------------------------------------- | ||
| 394 | − | iline = 0 | |
| 395 | − | param_lines = 0 | |
| 396 | − | final_line = 0 | |
| 397 | − | tag_loop: do | |
| 398 | |||
| 399 | ! Check for end of file | ||
| 400 | !------------------------------------------------------------------------ | ||
| 401 | − | read(unit,'(A)',iostat=stat) buffer | |
| 402 | − | if(stat.ne.0)then | |
| 403 | write(err_msg,'("file encountered error (EoF?) before END ",A)') & | ||
| 404 | − | to_upper(this%name) | |
| 405 | − | call stop_program(err_msg) | |
| 406 | − | return | |
| 407 | end if | ||
| 408 | − | if(trim(adjustl(buffer)).eq."") cycle tag_loop | |
| 409 | |||
| 410 | ! Check for end of layer card | ||
| 411 | !------------------------------------------------------------------------ | ||
| 412 | − | if(trim(adjustl(buffer)).eq."END "//to_upper(trim(this%name)))then | |
| 413 | − | final_line = iline | |
| 414 | − | backspace(unit) | |
| 415 | − | exit tag_loop | |
| 416 | end if | ||
| 417 | − | iline = iline + 1 | |
| 418 | |||
| 419 | − | tag = trim(adjustl(buffer)) | |
| 420 | − | if(scan(buffer,"=").ne.0) tag=trim(tag(:scan(tag,"=")-1)) | |
| 421 | |||
| 422 | ! Read parameters from save file | ||
| 423 | !------------------------------------------------------------------------ | ||
| 424 | − | select case(trim(tag)) | |
| 425 | case("INPUT_SHAPE") | ||
| 426 | − | call assign_vec(buffer, input_shape, itmp1) | |
| 427 | case("MOMENTUM") | ||
| 428 | − | call assign_val(buffer, momentum, itmp1) | |
| 429 | case("EPSILON") | ||
| 430 | − | call assign_val(buffer, epsilon, itmp1) | |
| 431 | case("NUM_CHANNELS") | ||
| 432 | − | call assign_val(buffer, num_channels, itmp1) | |
| 433 | − | write(0,*) "NUM_CHANNELS and INPUT_SHAPE are conflicting parameters" | |
| 434 | − | write(0,*) "NUM_CHANNELS will be ignored" | |
| 435 | case("GAMMA_INITIALISER", "KERNEL_INITIALISER") | ||
| 436 | − | if(param_lines(1).ne.0)then | |
| 437 | − | write(err_msg,'("GAMMA and GAMMA_INITIALISER defined. Using GAMMA only.")') | |
| 438 | − | call print_warning(err_msg) | |
| 439 | end if | ||
| 440 | − | call assign_val(buffer, gamma_initialiser_name, itmp1) | |
| 441 | case("BETA_INITIALISER", "BIAS_INITIALISER") | ||
| 442 | − | if(param_lines(2).ne.0)then | |
| 443 | − | write(err_msg,'("BETA and BETA_INITIALISER defined. Using BETA only.")') | |
| 444 | − | call print_warning(err_msg) | |
| 445 | end if | ||
| 446 | − | call assign_val(buffer, beta_initialiser_name, itmp1) | |
| 447 | case("MOVING_MEAN_INITIALISER") | ||
| 448 | − | call assign_val(buffer, moving_mean_initialiser_name, itmp1) | |
| 449 | case("MOVING_VARIANCE_INITIALISER") | ||
| 450 | − | call assign_val(buffer, moving_variance_initialiser_name, itmp1) | |
| 451 | case("GAMMA") | ||
| 452 | − | gamma_initialiser_name = 'zeros' | |
| 453 | − | param_lines(1) = iline | |
| 454 | case("BETA") | ||
| 455 | − | beta_initialiser_name = 'zeros' | |
| 456 | − | param_lines(2) = iline | |
| 457 | case default | ||
| 458 | ! Don't look for "e" due to scientific notation of numbers | ||
| 459 | ! ... i.e. exponent (E+00) | ||
| 460 | − | if(scan(to_lower(trim(adjustl(buffer))),& | |
| 461 | 'abcdfghijklmnopqrstuvwxyz').eq.0)then | ||
| 462 | − | cycle tag_loop | |
| 463 | − | elseif(tag(:3).eq.'END')then | |
| 464 | − | cycle tag_loop | |
| 465 | end if | ||
| 466 | write(err_msg,'("Unrecognised line in input file: ",A)') & | ||
| 467 | − | trim(adjustl(buffer)) | |
| 468 | − | call stop_program(err_msg) | |
| 469 | − | return | |
| 470 | end select | ||
| 471 | end do tag_loop | ||
| 472 | − | gamma_initialiser = initialiser_setup(gamma_initialiser_name) | |
| 473 | − | beta_initialiser = initialiser_setup(beta_initialiser_name) | |
| 474 | − | moving_mean_initialiser = initialiser_setup(moving_mean_initialiser_name) | |
| 475 | − | moving_variance_initialiser = initialiser_setup(moving_variance_initialiser_name) | |
| 476 | |||
| 477 | |||
| 478 | ! Set hyperparameters and initialise layer | ||
| 479 | !--------------------------------------------------------------------------- | ||
| 480 | − | num_channels = input_shape(size(input_shape)) | |
| 481 | call this%set_hyperparams( & | ||
| 482 | momentum = momentum, & | ||
| 483 | epsilon = epsilon, & | ||
| 484 | gamma_init_mean = this%gamma_init_mean, & | ||
| 485 | gamma_init_std = this%gamma_init_std, & | ||
| 486 | beta_init_mean = this%beta_init_mean, & | ||
| 487 | beta_init_std = this%beta_init_std, & | ||
| 488 | gamma_initialiser = gamma_initialiser, & | ||
| 489 | beta_initialiser = beta_initialiser, & | ||
| 490 | moving_mean_initialiser = moving_mean_initialiser, & | ||
| 491 | moving_variance_initialiser = moving_variance_initialiser, & | ||
| 492 | verbose = verbose_ & | ||
| 493 | − | ) | |
| 494 | − | call this%init(input_shape = input_shape) | |
| 495 | |||
| 496 | |||
| 497 | ! Check if WEIGHTS card was found | ||
| 498 | !--------------------------------------------------------------------------- | ||
| 499 | − | allocate(data_list(num_channels), source=0._real32) | |
| 500 | − | do i = 2, 1, -1 | |
| 501 | − | if(param_lines(i).eq.0) cycle | |
| 502 | − | call move(unit, param_lines(i) - iline, iostat=stat) | |
| 503 | − | iline = param_lines(i) + 1 | |
| 504 | − | c = 1 | |
| 505 | − | k = 1 | |
| 506 | − | data_list = 0._real32 | |
| 507 | − | data_concat_loop: do while(c.le.num_channels) | |
| 508 | − | iline = iline + 1 | |
| 509 | − | read(unit,'(A)',iostat=stat) buffer | |
| 510 | − | if(stat.ne.0) exit data_concat_loop | |
| 511 | − | k = icount(buffer) | |
| 512 | − | read(buffer,*,iostat=stat) (data_list(j),j=c,c+k-1) | |
| 513 | − | c = c + k | |
| 514 | end do data_concat_loop | ||
| 515 | − | read(unit,'(A)',iostat=stat) buffer | |
| 516 | − | select case(i) | |
| 517 | case(1) ! gamma | ||
| 518 | − | this%params(1)%val(1:this%num_channels,1) = data_list | |
| 519 | − | if(trim(adjustl(buffer)).ne."END GAMMA")then | |
| 520 | write(err_msg,'("END GAMMA not where expected: ",A)') & | ||
| 521 | − | trim(adjustl(buffer)) | |
| 522 | − | call stop_program(err_msg) | |
| 523 | − | return | |
| 524 | end if | ||
| 525 | case(2) ! beta | ||
| 526 | − | this%params(1)%val(this%num_channels+1:this%num_channels*2,1) = & | |
| 527 | − | data_list | |
| 528 | − | if(trim(adjustl(buffer)).ne."END BETA")then | |
| 529 | write(err_msg,'("END BETA not where expected: ",A)') & | ||
| 530 | − | trim(adjustl(buffer)) | |
| 531 | − | call stop_program(err_msg) | |
| 532 | − | return | |
| 533 | end if | ||
| 534 | end select | ||
| 535 | end do | ||
| 536 | − | deallocate(data_list) | |
| 537 | |||
| 538 | |||
| 539 | ! Check for end of layer card | ||
| 540 | !--------------------------------------------------------------------------- | ||
| 541 | − | call move(unit, final_line - iline, iostat=stat) | |
| 542 | − | read(unit,'(A)') buffer | |
| 543 | − | if(trim(adjustl(buffer)).ne."END "//to_upper(trim(this%name)))then | |
| 544 | − | write(0,*) trim(adjustl(buffer)) | |
| 545 | − | write(err_msg,'("END ",A," not where expected")') to_upper(this%name) | |
| 546 | − | call stop_program(err_msg) | |
| 547 | − | return | |
| 548 | end if | ||
| 549 | |||
| 550 | − | end subroutine read_batchnorm1d | |
| 551 | !############################################################################### | ||
| 552 | |||
| 553 | |||
| 554 | !############################################################################### | ||
| 555 | − | function read_batchnorm1d_layer(unit, verbose) result(layer) | |
| 556 | !! Read 1D batch normalisation layer from file and return layer | ||
| 557 | implicit none | ||
| 558 | |||
| 559 | ! Arguments | ||
| 560 | integer, intent(in) :: unit | ||
| 561 | !! File unit | ||
| 562 | integer, optional, intent(in) :: verbose | ||
| 563 | !! Verbosity level | ||
| 564 | class(base_layer_type), allocatable :: layer | ||
| 565 | !! Allocatable instance of the base layer | ||
| 566 | |||
| 567 | ! Local variables | ||
| 568 | integer :: verbose_ = 0 | ||
| 569 | !! Verbosity level | ||
| 570 | |||
| 571 | − | if(present(verbose)) verbose_ = verbose | |
| 572 | − | allocate(layer, source=batchnorm1d_layer_type()) | |
| 573 | − | call layer%read(unit, verbose=verbose_) | |
| 574 | |||
| 575 | − | end function read_batchnorm1d_layer | |
| 576 | !############################################################################### | ||
| 577 | |||
| 578 | |||
| 579 | !############################################################################### | ||
| 580 | − | subroutine build_from_onnx_batchnorm1d( & | |
| 581 | − | this, node, initialisers, value_info, verbose & | |
| 582 | ) | ||
| 583 | !! Read ONNX attributes for 1D batch normalisation layer | ||
| 584 | use athena__initialiser_data, only: data_init_type | ||
| 585 | implicit none | ||
| 586 | |||
| 587 | ! Arguments | ||
| 588 | class(batchnorm1d_layer_type), intent(inout) :: this | ||
| 589 | !! Instance of the 1D batch normalisation layer | ||
| 590 | type(onnx_node_type), intent(in) :: node | ||
| 591 | !! ONNX node information | ||
| 592 | type(onnx_initialiser_type), dimension(:), intent(in) :: initialisers | ||
| 593 | !! ONNX initialiser information | ||
| 594 | type(onnx_tensor_type), dimension(:), intent(in) :: value_info | ||
| 595 | !! ONNX value info | ||
| 596 | integer, intent(in) :: verbose | ||
| 597 | !! Verbosity level | ||
| 598 | |||
| 599 | ! Local variables | ||
| 600 | integer :: i | ||
| 601 | !! Loop index | ||
| 602 | real(real32) :: epsilon, momentum | ||
| 603 | !! Epsilon and momentum values | ||
| 604 | character(256) :: val | ||
| 605 | !! Attribute value | ||
| 606 | − | class(base_init_type), allocatable :: gamma_initialiser, beta_initialiser | |
| 607 | class(base_init_type), allocatable :: & | ||
| 608 | − | moving_mean_initialiser, moving_variance_initialiser | |
| 609 | |||
| 610 | ! Set default values | ||
| 611 | − | epsilon = 1.E-5_real32 | |
| 612 | − | momentum = 0.9_real32 | |
| 613 | |||
| 614 | ! Parse ONNX attributes | ||
| 615 | − | do i = 1, size(node%attributes) | |
| 616 | − | val = node%attributes(i)%val | |
| 617 | − | select case(trim(adjustl(node%attributes(i)%name))) | |
| 618 | case("epsilon") | ||
| 619 | − | read(val,*) epsilon | |
| 620 | case("momentum") | ||
| 621 | − | read(val,*) momentum | |
| 622 | case default | ||
| 623 | ! Do nothing | ||
| 624 | write(0,*) "WARNING: Unrecognised attribute in ONNX BATCHNORM1D & | ||
| 625 | − | &layer: ", trim(adjustl(node%attributes(i)%name)) | |
| 626 | end select | ||
| 627 | end do | ||
| 628 | |||
| 629 | ! Check for 4 initialisers: gamma, beta, mean, variance | ||
| 630 | − | if(size(initialisers).ne.4)then | |
| 631 | call stop_program("ONNX BATCHNORM1D layer requires 4 initialisers & | ||
| 632 | − | &(gamma, beta, mean, variance)") | |
| 633 | − | return | |
| 634 | end if | ||
| 635 | |||
| 636 | ! ONNX BatchNormalization order: gamma, beta, mean, variance | ||
| 637 | − | gamma_initialiser = data_init_type( data = initialisers(1)%data ) | |
| 638 | − | beta_initialiser = data_init_type( data = initialisers(2)%data ) | |
| 639 | − | moving_mean_initialiser = data_init_type( data = initialisers(3)%data ) | |
| 640 | − | moving_variance_initialiser = data_init_type( data = initialisers(4)%data ) | |
| 641 | |||
| 642 | call this%set_hyperparams( & | ||
| 643 | momentum = momentum, & | ||
| 644 | epsilon = epsilon, & | ||
| 645 | gamma_init_mean = 1.0_real32, & | ||
| 646 | gamma_init_std = 0.0_real32, & | ||
| 647 | beta_init_mean = 0.0_real32, & | ||
| 648 | beta_init_std = 0.0_real32, & | ||
| 649 | gamma_initialiser = gamma_initialiser, & | ||
| 650 | beta_initialiser = beta_initialiser, & | ||
| 651 | moving_mean_initialiser = moving_mean_initialiser, & | ||
| 652 | moving_variance_initialiser = moving_variance_initialiser, & | ||
| 653 | verbose = verbose & | ||
| 654 | − | ) | |
| 655 | |||
| 656 | − | end subroutine build_from_onnx_batchnorm1d | |
| 657 | !############################################################################### | ||
| 658 | |||
| 659 | |||
| 660 | !##############################################################################! | ||
| 661 | ! * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * ! | ||
| 662 | !##############################################################################! | ||
| 663 | |||
| 664 | |||
| 665 | !############################################################################### | ||
| 666 | − | subroutine forward_batchnorm1d(this, input) | |
| 667 | !! Forward propagation | ||
| 668 | implicit none | ||
| 669 | |||
| 670 | ! Arguments | ||
| 671 | class(batchnorm1d_layer_type), intent(inout) :: this | ||
| 672 | !! Instance of the 1D batch normalisation layer | ||
| 673 | class(array_type), dimension(:,:), intent(in) :: input | ||
| 674 | !! Input values | ||
| 675 | |||
| 676 | ! Local variables | ||
| 677 | class(batchnorm_array_type), pointer :: ptr | ||
| 678 | ! Pointer array | ||
| 679 | |||
| 680 | |||
| 681 | − | select case(this%inference) | |
| 682 | case(.true.) | ||
| 683 | ! Do not perform the drop operation | ||
| 684 | |||
| 685 | − | ptr => batchnorm_inference(input(1,1), this%params(1), & | |
| 686 | − | this%mean(:), this%variance(:), this%epsilon & | |
| 687 | − | ) | |
| 688 | |||
| 689 | case default | ||
| 690 | ! Perform the drop operation | ||
| 691 | ptr => batchnorm( & | ||
| 692 | − | input(1,1), this%params(1),& | |
| 693 | − | this%momentum, this%mean(:), this%variance(:), this%epsilon & | |
| 694 | − | ) | |
| 695 | |||
| 696 | end select | ||
| 697 | − | select type(output => this%output(1,1)) | |
| 698 | type is(batchnorm_array_type) | ||
| 699 | − | call output%assign_shallow(ptr) | |
| 700 | − | output%epsilon = ptr%epsilon | |
| 701 | − | output%mean = ptr%mean | |
| 702 | − | output%variance = ptr%variance | |
| 703 | end select | ||
| 704 | − | deallocate(ptr) | |
| 705 | − | this%output(1,1)%is_temporary = .false. | |
| 706 | |||
| 707 | − | end subroutine forward_batchnorm1d | |
| 708 | !############################################################################### | ||
| 709 | |||
| 710 | − | end module athena__batchnorm1d_layer | |
| 711 |