Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1 | # Copyright (c) 2020-2022, ARM Limited. |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | import numpy as np |
| 4 | import tensorflow as tf |
| 5 | from frameworks.tensor_gen import TGen |
| 6 | |
| 7 | |
| 8 | class TBuilder: |
| 9 | """The member functions build the tensorflow operators into small networks |
| 10 | for our tests""" |
| 11 | |
| 12 | def __init__(self): |
| 13 | pass |
| 14 | |
| 15 | def fake_quant(tensor, tensor_scale, name): |
| 16 | """Helper function for quantizing with a scaling parameters structure.""" |
| 17 | return tf.quantization.fake_quant_with_min_max_args( |
| 18 | tensor, |
| 19 | min=tensor_scale.min, |
| 20 | max=tensor_scale.max, |
| 21 | num_bits=tensor_scale.num_bits, |
| 22 | narrow_range=tensor_scale.narrow_range, |
| 23 | name=name, |
| 24 | ) |
| 25 | |
| 26 | def fake_quant_params(tensor, min, max, scaling, name): |
| 27 | """Helper function for quantizing with individual scaling parameters.""" |
| 28 | return tf.quantization.fake_quant_with_min_max_args( |
| 29 | tensor, |
| 30 | min=min, |
| 31 | max=max, |
| 32 | num_bits=scaling.num_bits, |
| 33 | narrow_range=scaling.narrow_range, |
| 34 | name=name, |
| 35 | ) |
| 36 | |
| 37 | class Add: |
| 38 | def __init__(self, name): |
| 39 | self.result_name = name |
| 40 | |
| 41 | def eval(self, a, b): |
| 42 | return tf.add(a, b, name=self.result_name) |
| 43 | |
| 44 | class Sub: |
| 45 | def __init__(self, name): |
| 46 | self.result_name = name |
| 47 | |
| 48 | def eval(self, a, b): |
| 49 | return tf.subtract(a, b, name=self.result_name) |
| 50 | |
| 51 | class Mul: |
| 52 | def __init__(self, name): |
| 53 | self.result_name = name |
| 54 | |
| 55 | def eval(self, a, b): |
| 56 | return tf.multiply(a, b, name=self.result_name) |
| 57 | |
| 58 | class Exp: |
| 59 | def __init__(self, name): |
| 60 | self.result_name = name |
| 61 | |
| 62 | def eval(self, a): |
| 63 | return tf.exp(a, name=self.result_name) |
| 64 | |
| 65 | class Rcp: |
| 66 | def __init__(self, name): |
| 67 | self.result_name = name |
| 68 | |
| 69 | def eval(self, a): |
| 70 | return tf.math.reciprocal(a, name=self.result_name) |
| 71 | |
| 72 | class Relu: |
| 73 | def __init__(self, name): |
| 74 | self.result_name = name |
| 75 | |
| 76 | def eval(self, a): |
| 77 | return tf.nn.relu(a, name=self.result_name) |
| 78 | |
| 79 | class Relu6: |
| 80 | def __init__(self, name): |
| 81 | self.result_name = name |
| 82 | |
| 83 | def eval(self, a): |
| 84 | return tf.nn.relu6(a, name=self.result_name) |
| 85 | |
| 86 | class LeakyRelu: |
| 87 | def __init__(self, alpha, name): |
| 88 | self.alpha = alpha |
| 89 | self.result_name = name |
| 90 | |
| 91 | def eval(self, a): |
| 92 | return tf.nn.leaky_relu(a, alpha=self.alpha, name=self.result_name) |
| 93 | |
| 94 | class Concat: |
| 95 | def __init__(self, axis, name): |
| 96 | self.axis = axis |
| 97 | self.result_name = name |
| 98 | |
| 99 | def eval(self, a, b): |
| 100 | return tf.concat([a, b], self.axis, name=self.result_name) |
| 101 | |
| 102 | class BitwiseAnd: |
| 103 | def __init__(self, name): |
| 104 | self.result_name = name |
| 105 | |
| 106 | def eval(self, a, b): |
| 107 | return tf.bitwise.bitwise_and(a, b, name=self.result_name) |
| 108 | |
| 109 | class BitwiseOr: |
| 110 | def __init__(self, name): |
| 111 | self.result_name = name |
| 112 | |
| 113 | def eval(self, a, b): |
| 114 | return tf.bitwise.bitwise_or(a, b, name=self.result_name) |
| 115 | |
| 116 | class BitwiseNot: |
| 117 | def __init__(self, name): |
| 118 | self.result_name = name |
| 119 | |
| 120 | def eval(self, a): |
| 121 | return tf.bitwise.invert(a, name=self.result_name) |
| 122 | |
| 123 | class BitwiseXor: |
| 124 | def __init__(self, name): |
| 125 | self.result_name = name |
| 126 | |
| 127 | def eval(self, a, b): |
| 128 | return tf.bitwise.bitwise_xor(a, b, name=self.result_name) |
| 129 | |
| 130 | class LogicalAnd: |
| 131 | def __init__(self, name): |
| 132 | self.result_name = name |
| 133 | |
| 134 | def eval(self, a, b): |
| 135 | return tf.math.logical_and(a, b, name=self.result_name) |
| 136 | |
| 137 | class LogicalOr: |
| 138 | def __init__(self, name): |
| 139 | self.result_name = name |
| 140 | |
| 141 | def eval(self, a, b): |
| 142 | return tf.math.logical_or(a, b, name=self.result_name) |
| 143 | |
| 144 | class LogicalNot: |
| 145 | def __init__(self, name): |
| 146 | self.result_name = name |
| 147 | |
| 148 | def eval(self, a): |
| 149 | return tf.math.logical_not(a, name=self.result_name) |
| 150 | |
| 151 | class ReduceAny: |
| 152 | def __init__(self, axis_list, keepdims, name): |
| 153 | self.axis_list = axis_list |
| 154 | self.keepdims = keepdims |
| 155 | self.result_name = name |
| 156 | |
| 157 | def eval(self, a): |
| 158 | return tf.math.reduce_any( |
| 159 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 160 | ) |
| 161 | |
| 162 | class ReduceAll: |
| 163 | def __init__(self, axis_list, keepdims, name): |
| 164 | self.axis_list = axis_list |
| 165 | self.keepdims = keepdims |
| 166 | self.result_name = name |
| 167 | |
| 168 | def eval(self, a): |
| 169 | return tf.math.reduce_all( |
| 170 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 171 | ) |
| 172 | |
| 173 | class ReduceMin: |
| 174 | def __init__(self, axis_list, keepdims, name): |
| 175 | self.axis_list = axis_list |
| 176 | self.keepdims = keepdims |
| 177 | self.result_name = name |
| 178 | |
| 179 | def eval(self, a): |
| 180 | return tf.math.reduce_min( |
| 181 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 182 | ) |
| 183 | |
| 184 | class ReduceMax: |
| 185 | def __init__(self, axis_list, keepdims, name): |
| 186 | self.axis_list = axis_list |
| 187 | self.keepdims = keepdims |
| 188 | self.result_name = name |
| 189 | |
| 190 | def eval(self, a): |
| 191 | return tf.math.reduce_max( |
| 192 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 193 | ) |
| 194 | |
| 195 | class ReduceSum: |
| 196 | def __init__(self, axis_list, keepdims, name): |
| 197 | self.axis_list = axis_list |
| 198 | self.keepdims = keepdims |
| 199 | self.result_name = name |
| 200 | |
| 201 | def eval(self, a): |
| 202 | return tf.math.reduce_sum( |
| 203 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 204 | ) |
| 205 | |
| 206 | class ReduceMean: |
| 207 | def __init__(self, axis_list, keepdims, name): |
| 208 | self.axis_list = axis_list |
| 209 | self.keepdims = keepdims |
| 210 | self.result_name = name |
| 211 | |
| 212 | def eval(self, a): |
| 213 | return tf.math.reduce_mean( |
| 214 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 215 | ) |
| 216 | |
| 217 | class ReduceProduct: |
| 218 | def __init__(self, axis_list, keepdims, name): |
| 219 | self.axis_list = axis_list |
| 220 | self.keepdims = keepdims |
| 221 | self.result_name = name |
| 222 | |
| 223 | def eval(self, a): |
| 224 | return tf.math.reduce_prod( |
| 225 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 226 | ) |
| 227 | |
| 228 | class Min: |
| 229 | def __init__(self, name): |
| 230 | self.result_name = name |
| 231 | |
| 232 | def eval(self, a, b): |
| 233 | return tf.math.minimum(a, b, name=self.result_name) |
| 234 | |
| 235 | class Max: |
| 236 | def __init__(self, name): |
| 237 | self.result_name = name |
| 238 | |
| 239 | def eval(self, a, b): |
| 240 | return tf.math.maximum(a, b, name=self.result_name) |
| 241 | |
| 242 | class Pow: |
| 243 | def __init__(self, name): |
| 244 | self.result_name = name |
| 245 | |
| 246 | def eval(self, a, b): |
| 247 | return tf.math.pow(a, b, name=self.result_name) |
| 248 | |
| 249 | class Abs: |
| 250 | def __init__(self, name): |
| 251 | self.result_name = name |
| 252 | |
| 253 | def eval(self, a): |
| 254 | return tf.math.abs(a, name=self.result_name) |
| 255 | |
| 256 | class Ceil: |
| 257 | def __init__(self, name): |
| 258 | self.result_name = name |
| 259 | |
| 260 | def eval(self, a): |
| 261 | return tf.math.ceil(a, name=self.result_name) |
| 262 | |
| 263 | class Floor: |
| 264 | def __init__(self, name): |
| 265 | self.result_name = name |
| 266 | |
| 267 | def eval(self, a): |
| 268 | return tf.math.floor(a, name=self.result_name) |
| 269 | |
| 270 | class Log: |
| 271 | def __init__(self, name): |
| 272 | self.result_name = name |
| 273 | |
| 274 | def eval(self, a): |
| 275 | return tf.math.log(a, name=self.result_name) |
| 276 | |
| 277 | class Negate: |
| 278 | def __init__(self, name): |
| 279 | self.result_name = name |
| 280 | |
| 281 | def eval(self, a): |
| 282 | return tf.math.negative(a, name=self.result_name) |
| 283 | |
| 284 | class Rsqrt: |
| 285 | def __init__(self, name): |
| 286 | self.result_name = name |
| 287 | |
| 288 | def eval(self, a): |
| 289 | return tf.math.rsqrt(a, name=self.result_name) |
| 290 | |
| 291 | class Sigmoid: |
| 292 | def __init__(self, name): |
| 293 | self.result_name = name |
| 294 | |
| 295 | def eval(self, a): |
| 296 | return tf.math.sigmoid(a, name=self.result_name) |
| 297 | |
| 298 | class Tanh: |
| 299 | def __init__(self, name): |
| 300 | self.result_name = name |
| 301 | |
| 302 | def eval(self, a): |
| 303 | return tf.math.tanh(a, name=self.result_name) |
| 304 | |
| 305 | class Square: |
| 306 | def __init__(self, name): |
| 307 | self.result_name = name |
| 308 | |
| 309 | def eval(self, a): |
| 310 | return tf.math.square(a, name=self.result_name) |
| 311 | |
| 312 | class SquaredDifference: |
| 313 | def __init__(self, name): |
| 314 | self.result_name = name |
| 315 | |
| 316 | def eval(self, a, b): |
| 317 | return tf.math.squared_difference(a, b, name=self.result_name) |
| 318 | |
| 319 | class Equal: |
| 320 | def __init__(self, name): |
| 321 | self.result_name = name |
| 322 | |
| 323 | def eval(self, a, b): |
| 324 | return tf.math.equal(a, b, name=self.result_name) |
| 325 | |
| 326 | class GreaterEqual: |
| 327 | def __init__(self, name): |
| 328 | self.result_name = name |
| 329 | |
| 330 | def eval(self, a, b): |
| 331 | return tf.math.greater_equal(a, b, name=self.result_name) |
| 332 | |
| 333 | class Greater: |
| 334 | def __init__(self, name): |
| 335 | self.result_name = name |
| 336 | |
| 337 | def eval(self, a, b): |
| 338 | return tf.math.greater(a, b, name=self.result_name) |
| 339 | |
| 340 | class Less: |
| 341 | def __init__(self, name): |
| 342 | self.result_name = name |
| 343 | |
| 344 | def eval(self, a, b): |
| 345 | return tf.math.less(a, b, name=self.result_name) |
| 346 | |
| 347 | class LessEqual: |
| 348 | def __init__(self, name): |
| 349 | self.result_name = name |
| 350 | |
| 351 | def eval(self, a, b): |
| 352 | return tf.math.less_equal(a, b, name=self.result_name) |
| 353 | |
| 354 | class Conv2d: |
| 355 | def __init__(self, weight, strides, padding, dilations, name): |
| 356 | self.weight = weight |
| 357 | self.strides = strides |
| 358 | self.padding = padding |
| 359 | self.dilations = dilations |
| 360 | self.result_name = name |
| 361 | |
| 362 | def eval(self, input): |
| 363 | return tf.nn.conv2d( |
| 364 | input, |
| 365 | self.weight, |
| 366 | self.strides, |
| 367 | self.padding, |
| 368 | data_format="NHWC", |
| 369 | dilations=self.dilations, |
| 370 | name=self.result_name, |
| 371 | ) |
| 372 | |
| 373 | class Conv2dRelu: |
| 374 | def __init__(self, weight, name): |
| 375 | self.weight = weight |
| 376 | self.result_name = name |
| 377 | |
| 378 | def eval(self, input): |
| 379 | conv2d = tf.nn.conv2d( |
| 380 | input, |
| 381 | self.weight, |
| 382 | [1, 1, 1, 1], |
| 383 | "SAME", |
| 384 | data_format="NHWC", |
| 385 | dilations=[1, 1, 1, 1], |
| 386 | name="conv2d", |
| 387 | ) |
| 388 | return tf.nn.relu(conv2d, name=self.result_name) |
| 389 | |
| 390 | class Conv2dRelu6: |
| 391 | def __init__(self, weight, name): |
| 392 | self.weight = weight |
| 393 | self.result_name = name |
| 394 | |
| 395 | def eval(self, input): |
| 396 | conv2d = tf.nn.conv2d( |
| 397 | input, |
| 398 | self.weight, |
| 399 | [1, 1, 1, 1], |
| 400 | "SAME", |
| 401 | data_format="NHWC", |
| 402 | dilations=[1, 1, 1, 1], |
| 403 | name="conv2d", |
| 404 | ) |
| 405 | return tf.nn.relu6(conv2d, name=self.result_name) |
| 406 | |
| 407 | class Conv2dReluN1To1: |
| 408 | def __init__(self, weight, name): |
| 409 | self.weight = weight |
| 410 | self.result_name = name |
| 411 | |
| 412 | def eval(self, input): |
| 413 | conv2d = tf.nn.conv2d( |
| 414 | input, |
| 415 | self.weight, |
| 416 | [1, 1, 1, 1], |
| 417 | "SAME", |
| 418 | data_format="NHWC", |
| 419 | dilations=[1, 1, 1, 1], |
| 420 | name="conv2d", |
| 421 | ) |
| 422 | return tf.clip_by_value(conv2d, -1.0, 1.0, name=self.result_name) |
| 423 | |
| 424 | class Conv2dTanh: |
| 425 | def __init__(self, weight, name): |
| 426 | self.weight = weight |
| 427 | self.result_name = name |
| 428 | |
| 429 | def eval(self, input): |
| 430 | conv2d = tf.nn.conv2d( |
| 431 | input, |
| 432 | self.weight, |
| 433 | [1, 1, 1, 1], |
| 434 | "SAME", |
| 435 | data_format="NHWC", |
| 436 | dilations=[1, 1, 1, 1], |
| 437 | name="conv2d", |
| 438 | ) |
| 439 | return tf.math.tanh(conv2d, name=self.result_name) |
| 440 | |
| 441 | class Conv2dWithBias: |
| 442 | def __init__(self, weight, bias, strides, padding, dilations, name): |
| 443 | self.weight = weight |
| 444 | self.bias = bias |
| 445 | self.strides = strides |
| 446 | self.padding = padding |
| 447 | self.dilations = dilations |
| 448 | self.result_name = name |
| 449 | |
| 450 | def eval(self, input): |
| 451 | conv2d_op = tf.nn.conv2d( |
| 452 | input, |
| 453 | self.weight, |
| 454 | self.strides, |
| 455 | self.padding, |
| 456 | data_format="NHWC", |
| 457 | dilations=self.dilations, |
| 458 | name="conv2d", |
| 459 | ) |
| 460 | bias_add_op = tf.nn.bias_add( |
| 461 | conv2d_op, self.bias, data_format="NHWC", name=self.result_name |
| 462 | ) |
| 463 | return bias_add_op |
| 464 | |
| 465 | class DepthwiseConv2d: |
| 466 | def __init__(self, weight, strides, padding, dilations, name): |
| 467 | self.weight = weight |
| 468 | self.strides = strides |
| 469 | self.padding = padding |
| 470 | self.dilations = dilations |
| 471 | self.result_name = name |
| 472 | |
| 473 | def eval(self, input): |
| 474 | dws_conv2d = tf.nn.depthwise_conv2d( |
| 475 | input, |
| 476 | self.weight, |
| 477 | self.strides, |
| 478 | self.padding, |
| 479 | data_format="NHWC", |
| 480 | dilations=self.dilations, |
| 481 | name="dws_conv2d", |
| 482 | ) |
| 483 | return tf.identity(dws_conv2d, name=self.result_name) |
| 484 | |
| 485 | class DepthwiseConv2dWithBias: |
| 486 | def __init__(self, weight, bias, strides, padding, dilations, name): |
| 487 | self.weight = weight |
| 488 | self.bias = bias |
| 489 | self.strides = strides |
| 490 | self.padding = padding |
| 491 | self.dilations = dilations |
| 492 | self.result_name = name |
| 493 | |
| 494 | def eval(self, input): |
| 495 | dws_conv2d = tf.nn.depthwise_conv2d( |
| 496 | input, |
| 497 | self.weight, |
| 498 | self.strides, |
| 499 | self.padding, |
| 500 | data_format="NHWC", |
| 501 | dilations=self.dilations, |
| 502 | name="dws_conv2d", |
| 503 | ) |
| 504 | bias_add_op = tf.nn.bias_add( |
| 505 | dws_conv2d, self.bias, data_format="NHWC", name=self.result_name |
| 506 | ) |
| 507 | return bias_add_op |
| 508 | |
| 509 | class TransposeConv2d: |
| 510 | def __init__(self, weight, output_shape, strides, padding, name): |
| 511 | self.weight = weight |
| 512 | self.output_shape = output_shape |
| 513 | self.strides = strides |
| 514 | self.padding = padding |
| 515 | self.result_name = name |
| 516 | |
| 517 | def eval(self, input): |
| 518 | return tf.nn.conv2d_transpose( |
| 519 | input, |
| 520 | self.weight, |
| 521 | self.output_shape, |
| 522 | self.strides, |
| 523 | self.padding, |
| 524 | data_format="NHWC", |
| 525 | name=self.result_name, |
| 526 | ) |
| 527 | |
| 528 | class Argmax: |
| 529 | def __init__(self, axis, name): |
| 530 | self.axis = axis |
| 531 | self.result_name = name |
| 532 | |
| 533 | def eval(self, a): |
| 534 | return tf.argmax(a, self.axis, output_type=tf.int32, name=self.result_name) |
| 535 | |
| 536 | class AvgPool2d: |
| 537 | def __init__(self, strides, kernel_size, padding, name): |
| 538 | self.strides = strides |
| 539 | self.kernel_size = kernel_size |
| 540 | self.padding = padding |
| 541 | self.result_name = name |
| 542 | |
| 543 | def eval(self, input): |
| 544 | return tf.nn.avg_pool2d( |
| 545 | input, |
| 546 | strides=self.strides, |
| 547 | ksize=self.kernel_size, |
| 548 | padding=self.padding, |
| 549 | data_format="NHWC", |
| 550 | name=self.result_name, |
| 551 | ) |
| 552 | |
| 553 | class MaxPool2d: |
| 554 | def __init__(self, strides, kernel_size, padding, name): |
| 555 | self.strides = strides |
| 556 | self.kernel_size = kernel_size |
| 557 | self.padding = padding |
| 558 | self.result_name = name |
| 559 | |
| 560 | def eval(self, input): |
| 561 | return tf.nn.max_pool2d( |
| 562 | input, |
| 563 | strides=self.strides, |
| 564 | ksize=self.kernel_size, |
| 565 | padding=self.padding, |
| 566 | data_format="NHWC", |
| 567 | name=self.result_name, |
| 568 | ) |
| 569 | |
| 570 | class Reshape: |
| 571 | def __init__(self, shape, name): |
| 572 | self.shape = shape |
| 573 | self.result_name = name |
| 574 | |
| 575 | def eval(self, a): |
| 576 | reshape_op = tf.reshape(a, self.shape) |
| 577 | return tf.identity(reshape_op, name=self.result_name) |
| 578 | |
| 579 | class Transpose: |
| 580 | def __init__(self, perm, name): |
| 581 | self.perm = perm |
| 582 | self.result_name = name |
| 583 | |
| 584 | def eval(self, a): |
| 585 | return tf.transpose(a, self.perm, name=self.result_name) |
| 586 | |
| 587 | class Slice: |
| 588 | def __init__(self, begin, size, name): |
| 589 | self.begin = begin |
| 590 | self.size = size |
| 591 | self.result_name = name |
| 592 | |
| 593 | def eval(self, a): |
| 594 | return tf.slice(a, begin=self.begin, size=self.size, name=self.result_name) |
| 595 | |
| 596 | class StridedSlice: |
| 597 | def __init__( |
| 598 | self, |
| 599 | begin, |
| 600 | end, |
| 601 | strides, |
| 602 | begin_mask, |
| 603 | end_mask, |
| 604 | ellipsis_mask, |
| 605 | new_axis_mask, |
| 606 | shrink_axis_mask, |
| 607 | name, |
| 608 | ): |
| 609 | self.begin = begin |
| 610 | self.end = end |
| 611 | self.strides = strides |
| 612 | self.begin_mask = begin_mask |
| 613 | self.end_mask = end_mask |
| 614 | self.ellipsis_mask = ellipsis_mask |
| 615 | self.new_axis_mask = new_axis_mask |
| 616 | self.shrink_axis_mask = shrink_axis_mask |
| 617 | self.result_name = name |
| 618 | |
| 619 | def eval(self, a): |
| 620 | return tf.strided_slice( |
| 621 | a, |
| 622 | begin=self.begin, |
| 623 | end=self.end, |
| 624 | strides=self.strides, |
| 625 | begin_mask=self.begin_mask, |
| 626 | end_mask=self.end_mask, |
| 627 | ellipsis_mask=self.ellipsis_mask, |
| 628 | new_axis_mask=self.new_axis_mask, |
| 629 | shrink_axis_mask=self.shrink_axis_mask, |
| 630 | name=self.result_name, |
| 631 | ) |
| 632 | |
| 633 | class Select: |
| 634 | def __init__(self, name): |
| 635 | self.result_name = name |
| 636 | |
| 637 | def eval(self, selector, a, b): |
| 638 | return tf.where(condition=selector, x=a, y=b, name=self.result_name) |
| 639 | |
| 640 | class Addn: |
| 641 | def __init__(self, name): |
| 642 | self.result_name = name |
| 643 | |
| 644 | def eval(self, a, b, c, d): |
| 645 | return tf.add_n([a, b, c, d], name=self.result_name) |
| 646 | |
| 647 | class Concatv2: |
| 648 | def __init__(self, axis, name): |
| 649 | self.axis = axis |
| 650 | self.result_name = name |
| 651 | |
| 652 | def eval(self, a, b, c, d): |
| 653 | return tf.concat([a, b, c, d], axis=self.axis, name=self.result_name) |
| 654 | |
| 655 | class Stack: |
| 656 | def __init__(self, axis, name): |
| 657 | self.axis = axis |
| 658 | self.result_name = name |
| 659 | |
| 660 | def eval(self, a, b, c, d): |
| 661 | return tf.stack([a, b, c, d], axis=self.axis, name=self.result_name) |
| 662 | |
| 663 | class Unstack: |
| 664 | def __init__(self, axis, name): |
| 665 | self.axis = axis |
| 666 | self.result_name = name |
| 667 | |
| 668 | def eval(self, a): |
| 669 | unstack_op = tf.unstack(a, axis=self.axis, name="unstack_op") |
| 670 | result_count = a.shape[self.axis] |
| 671 | |
| 672 | if result_count == 1: |
| 673 | return tf.identity(unstack_op[0], name=self.result_name) |
| 674 | |
| 675 | sums = [] |
| 676 | for i in range(result_count): |
| 677 | sums.append( |
| 678 | tf.math.reduce_sum(unstack_op[i], name="reduce_{}".format(i)) |
| 679 | ) |
| 680 | return tf.stack(sums, 0, name=self.result_name) |
| 681 | |
| 682 | class Pad: |
| 683 | def __init__(self, padding, name): |
| 684 | self.padding = padding |
| 685 | self.result_name = name |
| 686 | |
| 687 | def eval(self, a): |
| 688 | return tf.pad( |
| 689 | a, |
| 690 | self.padding, |
| 691 | mode="CONSTANT", |
| 692 | constant_values=0, |
| 693 | name=self.result_name, |
| 694 | ) |
| 695 | |
| 696 | class ExpandDims: |
| 697 | def __init__(self, axis, name): |
| 698 | self.axis = axis |
| 699 | self.result_name = name |
| 700 | |
| 701 | def eval(self, a): |
| 702 | return tf.expand_dims(a, self.axis, name=self.result_name) |
| 703 | |
| 704 | class Shape: |
| 705 | def __init__(self, name): |
| 706 | self.result_name = name |
| 707 | |
| 708 | def eval(self, a): |
| 709 | return tf.shape(a, name=self.result_name) |
| 710 | |
| 711 | class Rank: |
| 712 | def __init__(self, name): |
| 713 | self.result_name = name |
| 714 | |
| 715 | def eval(self, a): |
| 716 | return tf.rank(a, name=self.result_name) |
| 717 | |
| 718 | class Fill: |
| 719 | def __init__(self, shape, value, name): |
| 720 | self.shape = shape |
| 721 | self.value = value |
| 722 | self.result_name = name |
| 723 | |
| 724 | def eval(self, a): |
| 725 | return tf.fill(self.shape, self.value, name=self.result_name) |
| 726 | |
| 727 | class Elu: |
| 728 | def __init__(self, name): |
| 729 | self.result_name = name |
| 730 | |
| 731 | def eval(self, a): |
| 732 | return tf.nn.elu(a, name=self.result_name) |
| 733 | |
| 734 | class Softmax: |
| 735 | def __init__(self, name): |
| 736 | self.result_name = name |
| 737 | |
| 738 | def eval(self, a): |
| 739 | return tf.nn.softmax(a, name=self.result_name) |
| 740 | |
| 741 | class LogSoftmax: |
| 742 | def __init__(self, name): |
| 743 | self.result_name = name |
| 744 | |
| 745 | def eval(self, a): |
| 746 | return tf.nn.log_softmax(a, name=self.result_name) |
| 747 | |
| 748 | class MatMul: |
| 749 | def __init__(self, name): |
| 750 | self.result_name = name |
| 751 | |
| 752 | def eval(self, a, b): |
| 753 | return tf.linalg.matmul(a, b, name=self.result_name) |
| 754 | |
| 755 | class AddScalar: |
| 756 | def __init__(self, name): |
| 757 | self.result_name = name |
| 758 | |
| 759 | def eval(self, a): |
| 760 | return tf.add(a, 1, name=self.result_name) |
| 761 | |
| 762 | class Add1d: |
| 763 | def __init__(self, name): |
| 764 | self.result_name = name |
| 765 | |
| 766 | def eval(self, a, b): |
| 767 | if len(b.shape) > 1: |
| 768 | b_1d = tf.reduce_sum(b, axis=list(range(0, len(b.shape) - 1, 1))) |
| 769 | else: |
| 770 | b_1d = b |
| 771 | return tf.add(a, b_1d, name=self.result_name) |
| 772 | |
| 773 | class Split: |
| 774 | def __init__(self, num_splits, axis, name): |
| 775 | self.num_splits = num_splits |
| 776 | self.axis = axis |
| 777 | self.result_name = name |
| 778 | |
| 779 | def eval(self, a): |
| 780 | # The split op generates a list of outputs. Since we have difficulty |
| 781 | # serializing a list or array of Numpy arrays, we will reduce each of |
| 782 | # the results |
| 783 | |
| 784 | if not isinstance(self.num_splits, list): |
| 785 | split_op = tf.split( |
| 786 | a, num_or_size_splits=self.num_splits, axis=self.axis, name="split" |
| 787 | ) |
| 788 | result_count = self.num_splits |
| 789 | else: |
| 790 | num_split = np.asarray(self.num_splits, dtype=np.int32) |
| 791 | split_vec_op = tf.compat.v1.constant( |
| 792 | num_split, |
| 793 | shape=num_split.shape, |
| 794 | dtype=tf.int32, |
| 795 | name="const_split_vec", |
| 796 | ) |
| 797 | split_op = tf.split( |
| 798 | a, num_or_size_splits=split_vec_op, axis=self.axis, name="split" |
| 799 | ) |
| 800 | result_count = num_split.shape[0] |
| 801 | |
| 802 | sums = [] |
| 803 | for i in range(result_count): |
| 804 | sums.append(tf.math.reduce_sum(split_op[i], name="reduce_{}".format(i))) |
| 805 | return tf.stack(sums, 0, name=self.result_name) |
| 806 | |
| 807 | class Tile: |
| 808 | def __init__(self, multiples, name): |
| 809 | self.multiples = multiples |
| 810 | self.result_name = name |
| 811 | |
| 812 | def eval(self, a): |
| 813 | t = tf.tile(a, self.multiples, name="tile") |
| 814 | return tf.identity(t, name=self.result_name) |
| 815 | |
| 816 | class Reverse: |
| 817 | def __init__(self, axis, name): |
| 818 | self.axis = axis |
| 819 | self.result_name = name |
| 820 | |
| 821 | def eval(self, a): |
| 822 | return tf.reverse(a, [self.axis], name=self.result_name) |
| 823 | |
| 824 | class Gather: |
| 825 | def __init__(self, indices, batch_dims, axis, name): |
| 826 | self.indices = indices |
| 827 | self.batch_dims = batch_dims |
| 828 | self.axis = axis |
| 829 | self.result_name = name |
| 830 | |
| 831 | def eval(self, a): |
| 832 | return tf.gather( |
| 833 | a, |
| 834 | self.indices, |
| 835 | batch_dims=self.batch_dims, |
| 836 | axis=self.axis, |
| 837 | name=self.result_name, |
| 838 | ) |
| 839 | |
| 840 | class GatherNd: |
| 841 | def __init__(self, indices, name): |
| 842 | self.indices = indices |
| 843 | self.result_name = name |
| 844 | |
| 845 | def eval(self, a): |
| 846 | return tf.gather_nd(a, self.indices, name=self.result_name) |
| 847 | |
| 848 | class ScatterNd: |
| 849 | def __init__(self, shape, indices_shape, N, rng, name): |
| 850 | self.shape = shape |
| 851 | self.indices_shape = indices_shape |
| 852 | self.N = N |
| 853 | self.rng = rng |
| 854 | self.result_name = name |
| 855 | |
| 856 | def eval(self, a): |
| 857 | |
| 858 | # This operator is special. The indices and updates tensors really need |
| 859 | # to be created together, but in the current structure of this tool there |
| 860 | # is no way to do that before now. The number of updates is determined by |
| 861 | # the indices, so we can really only create that after indices; but we |
| 862 | # don't know the type at that time. |
| 863 | # |
| 864 | # Shapes are guaranteed deterministic, but we'll use our rng |
| 865 | # copied from the arggen stage. It's possible that index and |
| 866 | # update *values* will be non-deterministic. |
| 867 | # |
| 868 | # We take the tensor_tensor simply to get the dtype. |
| 869 | |
| 870 | shape_const = tf.constant(self.shape, tf.int32) |
| 871 | |
| 872 | updates_shape = list(self.indices_shape[:-1]) |
| 873 | updates_shape.extend(self.shape[self.indices_shape[-1] :]) |
| 874 | |
| 875 | updates_const = tf.constant(TGen.getRand(updates_shape, a.dtype, self.rng)) |
| 876 | |
| 877 | indices = np.zeros(self.indices_shape, dtype=np.int32) |
| 878 | |
| 879 | # We need to generate the random indices tensor based on the |
| 880 | # limits of 'shape' for each dimension. Surely, there is a faster |
| 881 | # vectorized way to do this, but the tensors are fairly small so we |
| 882 | # will do this one element at a time. Each element needs to be sized based |
| 883 | # on the size of the last dimension. |
| 884 | for idx in np.ndindex(indices.shape): |
| 885 | indices[idx] = self.rng.integers(0, self.shape[idx[-1]], size=1)[0] |
| 886 | # print('{} {}'.format(idx, indices[idx])) |
| 887 | |
| 888 | indices_const = tf.constant(indices, dtype=tf.int32) |
| 889 | |
| 890 | return tf.scatter_nd( |
| 891 | indices=indices_const, |
| 892 | updates=updates_const, |
| 893 | shape=shape_const, |
| 894 | name=self.result_name, |
| 895 | ) |
| 896 | |
| 897 | class SpaceToBatch: |
| 898 | def __init__(self, block_shape, padding, name): |
| 899 | self.block_shape = block_shape |
| 900 | self.padding = padding |
| 901 | self.result_name = name |
| 902 | |
| 903 | def eval(self, a): |
| 904 | return tf.space_to_batch( |
| 905 | a, self.block_shape, self.padding, name=self.result_name |
| 906 | ) |
| 907 | |
| 908 | class BatchToSpace: |
| 909 | def __init__(self, block_shape, cropping, name): |
| 910 | self.block_shape = block_shape |
| 911 | self.cropping = cropping |
| 912 | self.result_name = name |
| 913 | |
| 914 | def eval(self, a): |
| 915 | # transpose to swap depth and batch first. this could avoid adding new shape |
| 916 | block_rank = len(self.block_shape) |
| 917 | perm = [len(a.shape) - 1] |
| 918 | for i in range(block_rank): |
| 919 | perm.append(i + 1) |
| 920 | perm.append(0) |
| 921 | transpose_op = tf.transpose(a, perm) |
| 922 | return tf.batch_to_space( |
| 923 | transpose_op, self.block_shape, self.cropping, name=self.result_name |
| 924 | ) |
| 925 | |
| 926 | class SpaceToDepth: |
| 927 | def __init__(self, block_shape, name): |
| 928 | self.block_shape = block_shape |
| 929 | self.result_name = name |
| 930 | |
| 931 | def eval(self, a): |
| 932 | return tf.nn.space_to_depth(a, self.block_shape, name=self.result_name) |
| 933 | |
| 934 | class DepthToSpace: |
| 935 | def __init__(self, block_shape, name): |
| 936 | self.block_shape = block_shape |
| 937 | self.result_name = name |
| 938 | |
| 939 | def eval(self, a): |
| 940 | return tf.nn.depth_to_space(a, self.block_shape, name=self.result_name) |
| 941 | |
| 942 | class OneHot: |
| 943 | def __init__(self, depth, axis, name): |
| 944 | self.depth = depth |
| 945 | self.axis = axis |
| 946 | self.result_name = name |
| 947 | |
| 948 | def eval(self, indices, on_value, off_value): |
| 949 | return tf.one_hot( |
| 950 | indices, |
| 951 | self.depth, |
| 952 | on_value, |
| 953 | off_value, |
| 954 | self.axis, |
| 955 | on_value.dtype, |
| 956 | self.result_name, |
| 957 | ) |
| 958 | |
| 959 | class Fakequant: |
| 960 | def __init__(self, num_bits, narrow_range, name): |
| 961 | self.num_bits = num_bits |
| 962 | self.narrow_range = narrow_range |
| 963 | self.result_name = name |
| 964 | |
| 965 | def eval(self, a): |
| 966 | return tf.quantization.fake_quant_with_min_max_args( |
| 967 | a, |
| 968 | min=-2.0, |
| 969 | max=2.0, |
| 970 | num_bits=self.num_bits, |
| 971 | narrow_range=self.narrow_range, |
| 972 | name=self.result_name, |
| 973 | ) |
| 974 | |
| 975 | class ResizeNearest: |
| 976 | def __init__(self, name): |
| 977 | self.result_name = name |
| 978 | |
| 979 | def eval(self, a): |
| 980 | out_shape = [] |
| 981 | out_shape.append(a.shape[1] * 2) |
| 982 | out_shape.append(a.shape[2] * 2) |
| 983 | |
| 984 | # tf.image.resize() will overwrite the node name with result_name + |
| 985 | # '/BILINEAR' need to add extra identity to force output tensor name to |
| 986 | # result_name return tf.image.resize(a, out_shape, |
| 987 | # method=tf.image.ResizeMethod.NEAREST_NEIGHBOR, name=result_name) |
| 988 | resize = tf.image.resize( |
| 989 | a, |
| 990 | out_shape, |
| 991 | method=tf.image.ResizeMethod.NEAREST_NEIGHBOR, |
| 992 | name="resize", |
| 993 | ) |
| 994 | return tf.identity(resize, name=self.result_name) |
| 995 | |
| 996 | class ResizeBilinear: |
| 997 | def __init__(self, name): |
| 998 | self.result_name = name |
| 999 | |
| 1000 | def eval(self, a): |
| 1001 | out_shape = [] |
| 1002 | out_shape.append(a.shape[1] * 2) |
| 1003 | out_shape.append(a.shape[2] * 2) |
| 1004 | |
| 1005 | # tf.image.resize() will overwrite the node name with result_name + |
| 1006 | # '/BILINEAR' need to add extra identity to force output tensor name to |
| 1007 | # result_name return tf.image.resize(a, out_shape, |
| 1008 | # method=tf.image.ResizeMethod.NEAREST_NEIGHBOR, name=result_name) |
| 1009 | resize = tf.image.resize( |
| 1010 | a, out_shape, method=tf.image.ResizeMethod.BILINEAR, name="resize" |
| 1011 | ) |
| 1012 | return tf.identity(resize, name=self.result_name) |
| 1013 | |
| 1014 | class LeftShift: |
| 1015 | def __init__(self, shift, name): |
| 1016 | self.shift = shift |
| 1017 | self.result_name = name |
| 1018 | |
| 1019 | def eval(self, a): |
| 1020 | return tf.bitwise.left_shift(a, self.shift, name=self.result_name) |
| 1021 | |
| 1022 | class RightShift: |
| 1023 | def __init__(self, shift, name): |
| 1024 | self.shift = shift |
| 1025 | self.result_name = name |
| 1026 | |
| 1027 | def eval(self, a): |
| 1028 | return tf.bitwise.right_shift(a, self.shift, name=self.result_name) |