TatWai Chong | bef907a | 2024-01-23 09:40:37 -0800 | [diff] [blame] | 1 | # Copyright (c) 2020-2024, ARM Limited. |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 2 | # SPDX-License-Identifier: Apache-2.0 |
Jerry Ge | c444836 | 2024-03-21 17:33:55 +0000 | [diff] [blame^] | 3 | import os |
| 4 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 5 | import numpy as np |
| 6 | import tensorflow as tf |
| 7 | from frameworks.tensor_gen import TGen |
| 8 | |
Jerry Ge | c444836 | 2024-03-21 17:33:55 +0000 | [diff] [blame^] | 9 | os.environ["TF_USE_LEGACY_KERAS"] = "1" |
| 10 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 11 | |
| 12 | class TBuilder: |
| 13 | """The member functions build the tensorflow operators into small networks |
| 14 | for our tests""" |
| 15 | |
| 16 | def __init__(self): |
| 17 | pass |
| 18 | |
| 19 | def fake_quant(tensor, tensor_scale, name): |
| 20 | """Helper function for quantizing with a scaling parameters structure.""" |
| 21 | return tf.quantization.fake_quant_with_min_max_args( |
| 22 | tensor, |
| 23 | min=tensor_scale.min, |
| 24 | max=tensor_scale.max, |
| 25 | num_bits=tensor_scale.num_bits, |
| 26 | narrow_range=tensor_scale.narrow_range, |
| 27 | name=name, |
| 28 | ) |
| 29 | |
| 30 | def fake_quant_params(tensor, min, max, scaling, name): |
| 31 | """Helper function for quantizing with individual scaling parameters.""" |
| 32 | return tf.quantization.fake_quant_with_min_max_args( |
| 33 | tensor, |
| 34 | min=min, |
| 35 | max=max, |
| 36 | num_bits=scaling.num_bits, |
| 37 | narrow_range=scaling.narrow_range, |
| 38 | name=name, |
| 39 | ) |
| 40 | |
| 41 | class Add: |
| 42 | def __init__(self, name): |
| 43 | self.result_name = name |
| 44 | |
| 45 | def eval(self, a, b): |
| 46 | return tf.add(a, b, name=self.result_name) |
| 47 | |
| 48 | class Sub: |
| 49 | def __init__(self, name): |
| 50 | self.result_name = name |
| 51 | |
| 52 | def eval(self, a, b): |
| 53 | return tf.subtract(a, b, name=self.result_name) |
| 54 | |
| 55 | class Mul: |
| 56 | def __init__(self, name): |
| 57 | self.result_name = name |
| 58 | |
| 59 | def eval(self, a, b): |
| 60 | return tf.multiply(a, b, name=self.result_name) |
| 61 | |
| 62 | class Exp: |
| 63 | def __init__(self, name): |
| 64 | self.result_name = name |
| 65 | |
| 66 | def eval(self, a): |
| 67 | return tf.exp(a, name=self.result_name) |
| 68 | |
| 69 | class Rcp: |
| 70 | def __init__(self, name): |
| 71 | self.result_name = name |
| 72 | |
| 73 | def eval(self, a): |
| 74 | return tf.math.reciprocal(a, name=self.result_name) |
| 75 | |
| 76 | class Relu: |
| 77 | def __init__(self, name): |
| 78 | self.result_name = name |
| 79 | |
| 80 | def eval(self, a): |
| 81 | return tf.nn.relu(a, name=self.result_name) |
| 82 | |
Jerry Ge | 9391243 | 2022-07-22 10:29:13 -0700 | [diff] [blame] | 83 | class Relu1: |
| 84 | def __init__(self, name): |
| 85 | self.result_name = name |
| 86 | |
| 87 | def eval(self, a): |
| 88 | # TF doesn't have relu_n1_to_1 operator, |
| 89 | # use min and max as a workaround |
| 90 | # alternatively, we can use clip_by_value |
| 91 | return tf.math.minimum(1.0, tf.math.maximum(-1.0, a)) |
| 92 | |
Jerry Ge | 2eea5bf | 2022-10-11 16:27:05 +0000 | [diff] [blame] | 93 | class Relu0To1: |
| 94 | def __init__(self, name): |
| 95 | self.result_name = name |
| 96 | |
| 97 | def eval(self, a): |
| 98 | # TF doesn't have relu_0_to_1 operator, |
| 99 | # use min and max as a workaround |
| 100 | # alternatively, we can use clip_by_value |
| 101 | return tf.math.minimum(1.0, tf.math.maximum(0.0, a)) |
| 102 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 103 | class Relu6: |
| 104 | def __init__(self, name): |
| 105 | self.result_name = name |
| 106 | |
| 107 | def eval(self, a): |
| 108 | return tf.nn.relu6(a, name=self.result_name) |
| 109 | |
| 110 | class LeakyRelu: |
| 111 | def __init__(self, alpha, name): |
| 112 | self.alpha = alpha |
| 113 | self.result_name = name |
| 114 | |
| 115 | def eval(self, a): |
| 116 | return tf.nn.leaky_relu(a, alpha=self.alpha, name=self.result_name) |
| 117 | |
TatWai Chong | 41a04fe | 2022-11-03 21:44:32 +0000 | [diff] [blame] | 118 | class Prelu: |
| 119 | def __init__(self, name): |
| 120 | self.result_name = name |
| 121 | self.prelu = tf.keras.layers.PReLU( |
| 122 | alpha_initializer=tf.keras.initializers.RandomNormal( |
| 123 | mean=0.0, stddev=1.0 |
| 124 | ) |
| 125 | ) |
| 126 | |
| 127 | def eval(self, a): |
| 128 | return self.prelu(a) |
| 129 | |
TatWai Chong | 473eb38 | 2022-08-02 04:21:30 +0000 | [diff] [blame] | 130 | class Gelu: |
| 131 | def __init__(self, name): |
| 132 | self.result_name = name |
| 133 | |
| 134 | def eval(self, a): |
| 135 | return tf.nn.gelu(a, name=self.result_name) |
| 136 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 137 | class Concat: |
| 138 | def __init__(self, axis, name): |
| 139 | self.axis = axis |
| 140 | self.result_name = name |
| 141 | |
| 142 | def eval(self, a, b): |
Won Jeon | f9c0cee | 2023-09-18 16:32:45 -0700 | [diff] [blame] | 143 | return ( |
| 144 | tf.concat([a, b], self.axis, name=self.result_name) |
| 145 | if a.shape != () |
| 146 | else tf.stack([a, b], name=self.result_name) |
| 147 | ) |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 148 | |
| 149 | class BitwiseAnd: |
| 150 | def __init__(self, name): |
| 151 | self.result_name = name |
| 152 | |
| 153 | def eval(self, a, b): |
| 154 | return tf.bitwise.bitwise_and(a, b, name=self.result_name) |
| 155 | |
| 156 | class BitwiseOr: |
| 157 | def __init__(self, name): |
| 158 | self.result_name = name |
| 159 | |
| 160 | def eval(self, a, b): |
| 161 | return tf.bitwise.bitwise_or(a, b, name=self.result_name) |
| 162 | |
| 163 | class BitwiseNot: |
| 164 | def __init__(self, name): |
| 165 | self.result_name = name |
| 166 | |
| 167 | def eval(self, a): |
| 168 | return tf.bitwise.invert(a, name=self.result_name) |
| 169 | |
| 170 | class BitwiseXor: |
| 171 | def __init__(self, name): |
| 172 | self.result_name = name |
| 173 | |
| 174 | def eval(self, a, b): |
| 175 | return tf.bitwise.bitwise_xor(a, b, name=self.result_name) |
| 176 | |
| 177 | class LogicalAnd: |
| 178 | def __init__(self, name): |
| 179 | self.result_name = name |
| 180 | |
| 181 | def eval(self, a, b): |
| 182 | return tf.math.logical_and(a, b, name=self.result_name) |
| 183 | |
| 184 | class LogicalOr: |
| 185 | def __init__(self, name): |
| 186 | self.result_name = name |
| 187 | |
| 188 | def eval(self, a, b): |
| 189 | return tf.math.logical_or(a, b, name=self.result_name) |
| 190 | |
| 191 | class LogicalNot: |
| 192 | def __init__(self, name): |
| 193 | self.result_name = name |
| 194 | |
| 195 | def eval(self, a): |
| 196 | return tf.math.logical_not(a, name=self.result_name) |
| 197 | |
| 198 | class ReduceAny: |
| 199 | def __init__(self, axis_list, keepdims, name): |
| 200 | self.axis_list = axis_list |
| 201 | self.keepdims = keepdims |
| 202 | self.result_name = name |
| 203 | |
| 204 | def eval(self, a): |
| 205 | return tf.math.reduce_any( |
| 206 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 207 | ) |
| 208 | |
| 209 | class ReduceAll: |
| 210 | def __init__(self, axis_list, keepdims, name): |
| 211 | self.axis_list = axis_list |
| 212 | self.keepdims = keepdims |
| 213 | self.result_name = name |
| 214 | |
| 215 | def eval(self, a): |
| 216 | return tf.math.reduce_all( |
| 217 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 218 | ) |
| 219 | |
| 220 | class ReduceMin: |
| 221 | def __init__(self, axis_list, keepdims, name): |
| 222 | self.axis_list = axis_list |
| 223 | self.keepdims = keepdims |
| 224 | self.result_name = name |
| 225 | |
| 226 | def eval(self, a): |
| 227 | return tf.math.reduce_min( |
| 228 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 229 | ) |
| 230 | |
| 231 | class ReduceMax: |
| 232 | def __init__(self, axis_list, keepdims, name): |
| 233 | self.axis_list = axis_list |
| 234 | self.keepdims = keepdims |
| 235 | self.result_name = name |
| 236 | |
| 237 | def eval(self, a): |
| 238 | return tf.math.reduce_max( |
| 239 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 240 | ) |
| 241 | |
| 242 | class ReduceSum: |
| 243 | def __init__(self, axis_list, keepdims, name): |
| 244 | self.axis_list = axis_list |
| 245 | self.keepdims = keepdims |
| 246 | self.result_name = name |
| 247 | |
| 248 | def eval(self, a): |
| 249 | return tf.math.reduce_sum( |
| 250 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 251 | ) |
| 252 | |
| 253 | class ReduceMean: |
| 254 | def __init__(self, axis_list, keepdims, name): |
| 255 | self.axis_list = axis_list |
| 256 | self.keepdims = keepdims |
| 257 | self.result_name = name |
| 258 | |
| 259 | def eval(self, a): |
| 260 | return tf.math.reduce_mean( |
| 261 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 262 | ) |
| 263 | |
| 264 | class ReduceProduct: |
| 265 | def __init__(self, axis_list, keepdims, name): |
| 266 | self.axis_list = axis_list |
| 267 | self.keepdims = keepdims |
| 268 | self.result_name = name |
| 269 | |
| 270 | def eval(self, a): |
| 271 | return tf.math.reduce_prod( |
| 272 | a, self.axis_list, keepdims=self.keepdims, name=self.result_name |
| 273 | ) |
| 274 | |
| 275 | class Min: |
| 276 | def __init__(self, name): |
| 277 | self.result_name = name |
| 278 | |
| 279 | def eval(self, a, b): |
| 280 | return tf.math.minimum(a, b, name=self.result_name) |
| 281 | |
| 282 | class Max: |
| 283 | def __init__(self, name): |
| 284 | self.result_name = name |
| 285 | |
| 286 | def eval(self, a, b): |
| 287 | return tf.math.maximum(a, b, name=self.result_name) |
| 288 | |
| 289 | class Pow: |
| 290 | def __init__(self, name): |
| 291 | self.result_name = name |
| 292 | |
| 293 | def eval(self, a, b): |
| 294 | return tf.math.pow(a, b, name=self.result_name) |
| 295 | |
| 296 | class Abs: |
| 297 | def __init__(self, name): |
| 298 | self.result_name = name |
| 299 | |
| 300 | def eval(self, a): |
| 301 | return tf.math.abs(a, name=self.result_name) |
| 302 | |
| 303 | class Ceil: |
| 304 | def __init__(self, name): |
| 305 | self.result_name = name |
| 306 | |
| 307 | def eval(self, a): |
| 308 | return tf.math.ceil(a, name=self.result_name) |
| 309 | |
| 310 | class Floor: |
| 311 | def __init__(self, name): |
| 312 | self.result_name = name |
| 313 | |
| 314 | def eval(self, a): |
| 315 | return tf.math.floor(a, name=self.result_name) |
| 316 | |
| 317 | class Log: |
| 318 | def __init__(self, name): |
| 319 | self.result_name = name |
| 320 | |
| 321 | def eval(self, a): |
| 322 | return tf.math.log(a, name=self.result_name) |
| 323 | |
| 324 | class Negate: |
| 325 | def __init__(self, name): |
| 326 | self.result_name = name |
| 327 | |
| 328 | def eval(self, a): |
| 329 | return tf.math.negative(a, name=self.result_name) |
| 330 | |
| 331 | class Rsqrt: |
| 332 | def __init__(self, name): |
| 333 | self.result_name = name |
| 334 | |
| 335 | def eval(self, a): |
| 336 | return tf.math.rsqrt(a, name=self.result_name) |
| 337 | |
TatWai Chong | d713a4d | 2022-11-10 13:54:28 -0800 | [diff] [blame] | 338 | class Sign: |
| 339 | def __init__(self, name): |
| 340 | self.result_name = name |
| 341 | |
| 342 | def eval(self, a): |
| 343 | return tf.math.sign(a, name=self.result_name) |
| 344 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 345 | class Sigmoid: |
| 346 | def __init__(self, name): |
| 347 | self.result_name = name |
| 348 | |
| 349 | def eval(self, a): |
| 350 | return tf.math.sigmoid(a, name=self.result_name) |
| 351 | |
| 352 | class Tanh: |
| 353 | def __init__(self, name): |
| 354 | self.result_name = name |
| 355 | |
| 356 | def eval(self, a): |
| 357 | return tf.math.tanh(a, name=self.result_name) |
| 358 | |
Won Jeon | 78155c6 | 2023-06-10 00:20:04 +0000 | [diff] [blame] | 359 | class Erf: |
| 360 | # tfl.ops cannot be generated right now. |
| 361 | # https://github.com/tensorflow/tensorflow/issues/60809 |
| 362 | def __init__(self, name): |
| 363 | self.result_name = name |
| 364 | |
| 365 | def eval(self, a): |
| 366 | return tf.math.erf(a, name=self.result_name) |
| 367 | |
Luke Hutton | 4160186 | 2022-12-06 17:29:15 +0000 | [diff] [blame] | 368 | class Sin: |
| 369 | def __init__(self, name): |
| 370 | self.result_name = name |
| 371 | |
| 372 | def eval(self, a): |
| 373 | return tf.math.sin(a, name=self.result_name) |
| 374 | |
| 375 | class Cos: |
| 376 | def __init__(self, name): |
| 377 | self.result_name = name |
| 378 | |
| 379 | def eval(self, a): |
| 380 | return tf.math.cos(a, name=self.result_name) |
| 381 | |
Luke Hutton | 2138a19 | 2022-12-15 11:01:39 +0000 | [diff] [blame] | 382 | class Atan2: |
| 383 | def __init__(self, name): |
| 384 | self.result_name = name |
| 385 | |
| 386 | def eval(self, a, b): |
| 387 | return tf.math.atan2(a, b, name=self.result_name) |
| 388 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 389 | class Square: |
| 390 | def __init__(self, name): |
| 391 | self.result_name = name |
| 392 | |
| 393 | def eval(self, a): |
| 394 | return tf.math.square(a, name=self.result_name) |
| 395 | |
| 396 | class SquaredDifference: |
| 397 | def __init__(self, name): |
| 398 | self.result_name = name |
| 399 | |
| 400 | def eval(self, a, b): |
| 401 | return tf.math.squared_difference(a, b, name=self.result_name) |
| 402 | |
| 403 | class Equal: |
| 404 | def __init__(self, name): |
| 405 | self.result_name = name |
| 406 | |
| 407 | def eval(self, a, b): |
| 408 | return tf.math.equal(a, b, name=self.result_name) |
| 409 | |
| 410 | class GreaterEqual: |
| 411 | def __init__(self, name): |
| 412 | self.result_name = name |
| 413 | |
| 414 | def eval(self, a, b): |
| 415 | return tf.math.greater_equal(a, b, name=self.result_name) |
| 416 | |
| 417 | class Greater: |
| 418 | def __init__(self, name): |
| 419 | self.result_name = name |
| 420 | |
| 421 | def eval(self, a, b): |
| 422 | return tf.math.greater(a, b, name=self.result_name) |
| 423 | |
| 424 | class Less: |
| 425 | def __init__(self, name): |
| 426 | self.result_name = name |
| 427 | |
| 428 | def eval(self, a, b): |
| 429 | return tf.math.less(a, b, name=self.result_name) |
| 430 | |
| 431 | class LessEqual: |
| 432 | def __init__(self, name): |
| 433 | self.result_name = name |
| 434 | |
| 435 | def eval(self, a, b): |
| 436 | return tf.math.less_equal(a, b, name=self.result_name) |
| 437 | |
| 438 | class Conv2d: |
| 439 | def __init__(self, weight, strides, padding, dilations, name): |
| 440 | self.weight = weight |
| 441 | self.strides = strides |
| 442 | self.padding = padding |
| 443 | self.dilations = dilations |
| 444 | self.result_name = name |
| 445 | |
| 446 | def eval(self, input): |
| 447 | return tf.nn.conv2d( |
| 448 | input, |
| 449 | self.weight, |
| 450 | self.strides, |
| 451 | self.padding, |
| 452 | data_format="NHWC", |
| 453 | dilations=self.dilations, |
| 454 | name=self.result_name, |
| 455 | ) |
| 456 | |
| 457 | class Conv2dRelu: |
| 458 | def __init__(self, weight, name): |
| 459 | self.weight = weight |
| 460 | self.result_name = name |
| 461 | |
| 462 | def eval(self, input): |
| 463 | conv2d = tf.nn.conv2d( |
| 464 | input, |
| 465 | self.weight, |
| 466 | [1, 1, 1, 1], |
| 467 | "SAME", |
| 468 | data_format="NHWC", |
| 469 | dilations=[1, 1, 1, 1], |
| 470 | name="conv2d", |
| 471 | ) |
| 472 | return tf.nn.relu(conv2d, name=self.result_name) |
| 473 | |
| 474 | class Conv2dRelu6: |
| 475 | def __init__(self, weight, name): |
| 476 | self.weight = weight |
| 477 | self.result_name = name |
| 478 | |
| 479 | def eval(self, input): |
| 480 | conv2d = tf.nn.conv2d( |
| 481 | input, |
| 482 | self.weight, |
| 483 | [1, 1, 1, 1], |
| 484 | "SAME", |
| 485 | data_format="NHWC", |
| 486 | dilations=[1, 1, 1, 1], |
| 487 | name="conv2d", |
| 488 | ) |
| 489 | return tf.nn.relu6(conv2d, name=self.result_name) |
| 490 | |
| 491 | class Conv2dReluN1To1: |
| 492 | def __init__(self, weight, name): |
| 493 | self.weight = weight |
| 494 | self.result_name = name |
| 495 | |
| 496 | def eval(self, input): |
| 497 | conv2d = tf.nn.conv2d( |
| 498 | input, |
| 499 | self.weight, |
| 500 | [1, 1, 1, 1], |
| 501 | "SAME", |
| 502 | data_format="NHWC", |
| 503 | dilations=[1, 1, 1, 1], |
| 504 | name="conv2d", |
| 505 | ) |
| 506 | return tf.clip_by_value(conv2d, -1.0, 1.0, name=self.result_name) |
| 507 | |
| 508 | class Conv2dTanh: |
| 509 | def __init__(self, weight, name): |
| 510 | self.weight = weight |
| 511 | self.result_name = name |
| 512 | |
| 513 | def eval(self, input): |
| 514 | conv2d = tf.nn.conv2d( |
| 515 | input, |
| 516 | self.weight, |
| 517 | [1, 1, 1, 1], |
| 518 | "SAME", |
| 519 | data_format="NHWC", |
| 520 | dilations=[1, 1, 1, 1], |
| 521 | name="conv2d", |
| 522 | ) |
| 523 | return tf.math.tanh(conv2d, name=self.result_name) |
| 524 | |
| 525 | class Conv2dWithBias: |
| 526 | def __init__(self, weight, bias, strides, padding, dilations, name): |
| 527 | self.weight = weight |
| 528 | self.bias = bias |
| 529 | self.strides = strides |
| 530 | self.padding = padding |
| 531 | self.dilations = dilations |
| 532 | self.result_name = name |
| 533 | |
| 534 | def eval(self, input): |
| 535 | conv2d_op = tf.nn.conv2d( |
| 536 | input, |
| 537 | self.weight, |
| 538 | self.strides, |
| 539 | self.padding, |
| 540 | data_format="NHWC", |
| 541 | dilations=self.dilations, |
| 542 | name="conv2d", |
| 543 | ) |
| 544 | bias_add_op = tf.nn.bias_add( |
| 545 | conv2d_op, self.bias, data_format="NHWC", name=self.result_name |
| 546 | ) |
| 547 | return bias_add_op |
| 548 | |
TatWai Chong | fd62905 | 2022-07-25 04:01:58 +0000 | [diff] [blame] | 549 | class Conv3d: |
| 550 | def __init__(self, weight, strides, padding, dilations, name): |
| 551 | self.weight = weight |
| 552 | self.strides = strides |
| 553 | self.padding = padding |
| 554 | self.dilations = dilations |
| 555 | self.result_name = name |
| 556 | |
| 557 | def eval(self, input): |
| 558 | return tf.nn.conv3d( |
| 559 | input, |
| 560 | self.weight, |
| 561 | self.strides, |
| 562 | self.padding, |
| 563 | data_format="NDHWC", |
| 564 | dilations=self.dilations, |
| 565 | name=self.result_name, |
| 566 | ) |
| 567 | |
| 568 | class Conv3dWithBias: |
| 569 | def __init__(self, weight, bias, strides, padding, dilations, name): |
| 570 | self.weight = weight |
| 571 | self.bias = bias |
| 572 | self.strides = strides |
| 573 | self.padding = padding |
| 574 | self.dilations = dilations |
| 575 | self.result_name = name |
| 576 | |
| 577 | def eval(self, input): |
| 578 | conv3d_op = tf.nn.conv3d( |
| 579 | input, |
| 580 | self.weight, |
| 581 | self.strides, |
| 582 | self.padding, |
| 583 | data_format="NDHWC", |
| 584 | dilations=self.dilations, |
| 585 | name="conv3d", |
| 586 | ) |
| 587 | bias_add_op = tf.nn.bias_add(conv3d_op, self.bias, name=self.result_name) |
| 588 | return bias_add_op |
| 589 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 590 | class DepthwiseConv2d: |
| 591 | def __init__(self, weight, strides, padding, dilations, name): |
| 592 | self.weight = weight |
| 593 | self.strides = strides |
| 594 | self.padding = padding |
| 595 | self.dilations = dilations |
| 596 | self.result_name = name |
| 597 | |
| 598 | def eval(self, input): |
| 599 | dws_conv2d = tf.nn.depthwise_conv2d( |
| 600 | input, |
| 601 | self.weight, |
| 602 | self.strides, |
| 603 | self.padding, |
| 604 | data_format="NHWC", |
| 605 | dilations=self.dilations, |
| 606 | name="dws_conv2d", |
| 607 | ) |
| 608 | return tf.identity(dws_conv2d, name=self.result_name) |
| 609 | |
| 610 | class DepthwiseConv2dWithBias: |
| 611 | def __init__(self, weight, bias, strides, padding, dilations, name): |
| 612 | self.weight = weight |
| 613 | self.bias = bias |
| 614 | self.strides = strides |
| 615 | self.padding = padding |
| 616 | self.dilations = dilations |
| 617 | self.result_name = name |
| 618 | |
| 619 | def eval(self, input): |
| 620 | dws_conv2d = tf.nn.depthwise_conv2d( |
| 621 | input, |
| 622 | self.weight, |
| 623 | self.strides, |
| 624 | self.padding, |
| 625 | data_format="NHWC", |
| 626 | dilations=self.dilations, |
| 627 | name="dws_conv2d", |
| 628 | ) |
| 629 | bias_add_op = tf.nn.bias_add( |
| 630 | dws_conv2d, self.bias, data_format="NHWC", name=self.result_name |
| 631 | ) |
| 632 | return bias_add_op |
| 633 | |
| 634 | class TransposeConv2d: |
| 635 | def __init__(self, weight, output_shape, strides, padding, name): |
| 636 | self.weight = weight |
| 637 | self.output_shape = output_shape |
| 638 | self.strides = strides |
| 639 | self.padding = padding |
| 640 | self.result_name = name |
| 641 | |
| 642 | def eval(self, input): |
| 643 | return tf.nn.conv2d_transpose( |
| 644 | input, |
| 645 | self.weight, |
| 646 | self.output_shape, |
| 647 | self.strides, |
| 648 | self.padding, |
| 649 | data_format="NHWC", |
| 650 | name=self.result_name, |
| 651 | ) |
| 652 | |
| 653 | class Argmax: |
| 654 | def __init__(self, axis, name): |
| 655 | self.axis = axis |
| 656 | self.result_name = name |
| 657 | |
| 658 | def eval(self, a): |
| 659 | return tf.argmax(a, self.axis, output_type=tf.int32, name=self.result_name) |
| 660 | |
| 661 | class AvgPool2d: |
| 662 | def __init__(self, strides, kernel_size, padding, name): |
| 663 | self.strides = strides |
| 664 | self.kernel_size = kernel_size |
| 665 | self.padding = padding |
| 666 | self.result_name = name |
| 667 | |
| 668 | def eval(self, input): |
| 669 | return tf.nn.avg_pool2d( |
| 670 | input, |
| 671 | strides=self.strides, |
| 672 | ksize=self.kernel_size, |
| 673 | padding=self.padding, |
| 674 | data_format="NHWC", |
| 675 | name=self.result_name, |
| 676 | ) |
| 677 | |
| 678 | class MaxPool2d: |
| 679 | def __init__(self, strides, kernel_size, padding, name): |
| 680 | self.strides = strides |
| 681 | self.kernel_size = kernel_size |
| 682 | self.padding = padding |
| 683 | self.result_name = name |
| 684 | |
| 685 | def eval(self, input): |
| 686 | return tf.nn.max_pool2d( |
| 687 | input, |
| 688 | strides=self.strides, |
| 689 | ksize=self.kernel_size, |
| 690 | padding=self.padding, |
| 691 | data_format="NHWC", |
| 692 | name=self.result_name, |
| 693 | ) |
| 694 | |
| 695 | class Reshape: |
| 696 | def __init__(self, shape, name): |
| 697 | self.shape = shape |
| 698 | self.result_name = name |
| 699 | |
| 700 | def eval(self, a): |
| 701 | reshape_op = tf.reshape(a, self.shape) |
| 702 | return tf.identity(reshape_op, name=self.result_name) |
| 703 | |
| 704 | class Transpose: |
| 705 | def __init__(self, perm, name): |
| 706 | self.perm = perm |
| 707 | self.result_name = name |
| 708 | |
| 709 | def eval(self, a): |
| 710 | return tf.transpose(a, self.perm, name=self.result_name) |
| 711 | |
| 712 | class Slice: |
| 713 | def __init__(self, begin, size, name): |
| 714 | self.begin = begin |
| 715 | self.size = size |
| 716 | self.result_name = name |
| 717 | |
| 718 | def eval(self, a): |
| 719 | return tf.slice(a, begin=self.begin, size=self.size, name=self.result_name) |
| 720 | |
| 721 | class StridedSlice: |
| 722 | def __init__( |
| 723 | self, |
| 724 | begin, |
| 725 | end, |
| 726 | strides, |
| 727 | begin_mask, |
| 728 | end_mask, |
| 729 | ellipsis_mask, |
| 730 | new_axis_mask, |
| 731 | shrink_axis_mask, |
| 732 | name, |
| 733 | ): |
| 734 | self.begin = begin |
| 735 | self.end = end |
| 736 | self.strides = strides |
| 737 | self.begin_mask = begin_mask |
| 738 | self.end_mask = end_mask |
| 739 | self.ellipsis_mask = ellipsis_mask |
| 740 | self.new_axis_mask = new_axis_mask |
| 741 | self.shrink_axis_mask = shrink_axis_mask |
| 742 | self.result_name = name |
| 743 | |
| 744 | def eval(self, a): |
| 745 | return tf.strided_slice( |
| 746 | a, |
| 747 | begin=self.begin, |
| 748 | end=self.end, |
| 749 | strides=self.strides, |
| 750 | begin_mask=self.begin_mask, |
| 751 | end_mask=self.end_mask, |
| 752 | ellipsis_mask=self.ellipsis_mask, |
| 753 | new_axis_mask=self.new_axis_mask, |
| 754 | shrink_axis_mask=self.shrink_axis_mask, |
| 755 | name=self.result_name, |
| 756 | ) |
| 757 | |
| 758 | class Select: |
| 759 | def __init__(self, name): |
| 760 | self.result_name = name |
| 761 | |
| 762 | def eval(self, selector, a, b): |
| 763 | return tf.where(condition=selector, x=a, y=b, name=self.result_name) |
| 764 | |
| 765 | class Addn: |
| 766 | def __init__(self, name): |
| 767 | self.result_name = name |
| 768 | |
| 769 | def eval(self, a, b, c, d): |
| 770 | return tf.add_n([a, b, c, d], name=self.result_name) |
| 771 | |
| 772 | class Concatv2: |
| 773 | def __init__(self, axis, name): |
| 774 | self.axis = axis |
| 775 | self.result_name = name |
| 776 | |
| 777 | def eval(self, a, b, c, d): |
Won Jeon | f9c0cee | 2023-09-18 16:32:45 -0700 | [diff] [blame] | 778 | return ( |
| 779 | tf.concat([a, b, c, d], axis=self.axis, name=self.result_name) |
| 780 | if a.shape != () |
| 781 | else tf.stack([a, b, c, d], name=self.result_name) |
| 782 | ) |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 783 | |
| 784 | class Stack: |
| 785 | def __init__(self, axis, name): |
| 786 | self.axis = axis |
| 787 | self.result_name = name |
| 788 | |
| 789 | def eval(self, a, b, c, d): |
| 790 | return tf.stack([a, b, c, d], axis=self.axis, name=self.result_name) |
| 791 | |
| 792 | class Unstack: |
| 793 | def __init__(self, axis, name): |
| 794 | self.axis = axis |
| 795 | self.result_name = name |
| 796 | |
| 797 | def eval(self, a): |
| 798 | unstack_op = tf.unstack(a, axis=self.axis, name="unstack_op") |
| 799 | result_count = a.shape[self.axis] |
| 800 | |
| 801 | if result_count == 1: |
| 802 | return tf.identity(unstack_op[0], name=self.result_name) |
| 803 | |
| 804 | sums = [] |
| 805 | for i in range(result_count): |
| 806 | sums.append( |
| 807 | tf.math.reduce_sum(unstack_op[i], name="reduce_{}".format(i)) |
| 808 | ) |
| 809 | return tf.stack(sums, 0, name=self.result_name) |
| 810 | |
TatWai Chong | f7008da | 2022-09-09 09:35:40 +0000 | [diff] [blame] | 811 | class MirrorPad: |
| 812 | def __init__(self, padding, mode, name): |
| 813 | self.padding = padding |
| 814 | self.mode = mode |
| 815 | self.result_name = name |
| 816 | |
| 817 | def eval(self, a): |
| 818 | return tf.pad( |
| 819 | a, |
| 820 | self.padding, |
| 821 | mode=self.mode, |
| 822 | constant_values=0, |
| 823 | name=self.result_name, |
| 824 | ) |
| 825 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 826 | class Pad: |
TatWai Chong | 2226f90 | 2023-02-22 18:38:01 -0800 | [diff] [blame] | 827 | def __init__(self, padding, pad_const, name): |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 828 | self.padding = padding |
TatWai Chong | 2226f90 | 2023-02-22 18:38:01 -0800 | [diff] [blame] | 829 | self.pad_const = pad_const |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 830 | self.result_name = name |
| 831 | |
| 832 | def eval(self, a): |
| 833 | return tf.pad( |
| 834 | a, |
| 835 | self.padding, |
| 836 | mode="CONSTANT", |
TatWai Chong | 2226f90 | 2023-02-22 18:38:01 -0800 | [diff] [blame] | 837 | constant_values=self.pad_const, |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 838 | name=self.result_name, |
| 839 | ) |
| 840 | |
| 841 | class ExpandDims: |
| 842 | def __init__(self, axis, name): |
| 843 | self.axis = axis |
| 844 | self.result_name = name |
| 845 | |
| 846 | def eval(self, a): |
| 847 | return tf.expand_dims(a, self.axis, name=self.result_name) |
| 848 | |
| 849 | class Shape: |
| 850 | def __init__(self, name): |
| 851 | self.result_name = name |
| 852 | |
| 853 | def eval(self, a): |
| 854 | return tf.shape(a, name=self.result_name) |
| 855 | |
| 856 | class Rank: |
| 857 | def __init__(self, name): |
| 858 | self.result_name = name |
| 859 | |
| 860 | def eval(self, a): |
| 861 | return tf.rank(a, name=self.result_name) |
| 862 | |
| 863 | class Fill: |
| 864 | def __init__(self, shape, value, name): |
| 865 | self.shape = shape |
| 866 | self.value = value |
| 867 | self.result_name = name |
| 868 | |
| 869 | def eval(self, a): |
| 870 | return tf.fill(self.shape, self.value, name=self.result_name) |
| 871 | |
| 872 | class Elu: |
| 873 | def __init__(self, name): |
| 874 | self.result_name = name |
| 875 | |
| 876 | def eval(self, a): |
| 877 | return tf.nn.elu(a, name=self.result_name) |
| 878 | |
| 879 | class Softmax: |
| 880 | def __init__(self, name): |
| 881 | self.result_name = name |
| 882 | |
| 883 | def eval(self, a): |
| 884 | return tf.nn.softmax(a, name=self.result_name) |
| 885 | |
| 886 | class LogSoftmax: |
| 887 | def __init__(self, name): |
| 888 | self.result_name = name |
| 889 | |
| 890 | def eval(self, a): |
| 891 | return tf.nn.log_softmax(a, name=self.result_name) |
| 892 | |
Jerry Ge | 28811d9 | 2023-12-05 00:53:26 +0000 | [diff] [blame] | 893 | class DynamicLinear: |
| 894 | def __init__(self, dynamic_input_shape, name): |
| 895 | self.result_name = name |
| 896 | self.model = tf.keras.Sequential( |
| 897 | [ |
| 898 | tf.keras.layers.Input(shape=dynamic_input_shape), |
| 899 | tf.keras.layers.Dense(units=5), |
| 900 | ] |
| 901 | ) |
| 902 | |
| 903 | def eval(self, a): |
| 904 | return self.model(a) |
| 905 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 906 | class MatMul: |
| 907 | def __init__(self, name): |
| 908 | self.result_name = name |
| 909 | |
| 910 | def eval(self, a, b): |
| 911 | return tf.linalg.matmul(a, b, name=self.result_name) |
| 912 | |
| 913 | class AddScalar: |
| 914 | def __init__(self, name): |
| 915 | self.result_name = name |
| 916 | |
| 917 | def eval(self, a): |
| 918 | return tf.add(a, 1, name=self.result_name) |
| 919 | |
| 920 | class Add1d: |
| 921 | def __init__(self, name): |
| 922 | self.result_name = name |
| 923 | |
| 924 | def eval(self, a, b): |
| 925 | if len(b.shape) > 1: |
| 926 | b_1d = tf.reduce_sum(b, axis=list(range(0, len(b.shape) - 1, 1))) |
| 927 | else: |
| 928 | b_1d = b |
| 929 | return tf.add(a, b_1d, name=self.result_name) |
| 930 | |
| 931 | class Split: |
| 932 | def __init__(self, num_splits, axis, name): |
| 933 | self.num_splits = num_splits |
| 934 | self.axis = axis |
| 935 | self.result_name = name |
| 936 | |
| 937 | def eval(self, a): |
| 938 | # The split op generates a list of outputs. Since we have difficulty |
| 939 | # serializing a list or array of Numpy arrays, we will reduce each of |
| 940 | # the results |
| 941 | |
| 942 | if not isinstance(self.num_splits, list): |
| 943 | split_op = tf.split( |
| 944 | a, num_or_size_splits=self.num_splits, axis=self.axis, name="split" |
| 945 | ) |
| 946 | result_count = self.num_splits |
| 947 | else: |
| 948 | num_split = np.asarray(self.num_splits, dtype=np.int32) |
| 949 | split_vec_op = tf.compat.v1.constant( |
| 950 | num_split, |
| 951 | shape=num_split.shape, |
| 952 | dtype=tf.int32, |
| 953 | name="const_split_vec", |
| 954 | ) |
| 955 | split_op = tf.split( |
| 956 | a, num_or_size_splits=split_vec_op, axis=self.axis, name="split" |
| 957 | ) |
| 958 | result_count = num_split.shape[0] |
| 959 | |
| 960 | sums = [] |
| 961 | for i in range(result_count): |
| 962 | sums.append(tf.math.reduce_sum(split_op[i], name="reduce_{}".format(i))) |
| 963 | return tf.stack(sums, 0, name=self.result_name) |
| 964 | |
| 965 | class Tile: |
| 966 | def __init__(self, multiples, name): |
| 967 | self.multiples = multiples |
| 968 | self.result_name = name |
| 969 | |
| 970 | def eval(self, a): |
| 971 | t = tf.tile(a, self.multiples, name="tile") |
| 972 | return tf.identity(t, name=self.result_name) |
| 973 | |
| 974 | class Reverse: |
| 975 | def __init__(self, axis, name): |
| 976 | self.axis = axis |
| 977 | self.result_name = name |
| 978 | |
| 979 | def eval(self, a): |
| 980 | return tf.reverse(a, [self.axis], name=self.result_name) |
| 981 | |
| 982 | class Gather: |
| 983 | def __init__(self, indices, batch_dims, axis, name): |
| 984 | self.indices = indices |
| 985 | self.batch_dims = batch_dims |
| 986 | self.axis = axis |
| 987 | self.result_name = name |
| 988 | |
| 989 | def eval(self, a): |
| 990 | return tf.gather( |
| 991 | a, |
| 992 | self.indices, |
| 993 | batch_dims=self.batch_dims, |
| 994 | axis=self.axis, |
| 995 | name=self.result_name, |
| 996 | ) |
| 997 | |
| 998 | class GatherNd: |
| 999 | def __init__(self, indices, name): |
| 1000 | self.indices = indices |
| 1001 | self.result_name = name |
| 1002 | |
| 1003 | def eval(self, a): |
| 1004 | return tf.gather_nd(a, self.indices, name=self.result_name) |
| 1005 | |
| 1006 | class ScatterNd: |
| 1007 | def __init__(self, shape, indices_shape, N, rng, name): |
| 1008 | self.shape = shape |
| 1009 | self.indices_shape = indices_shape |
| 1010 | self.N = N |
| 1011 | self.rng = rng |
| 1012 | self.result_name = name |
| 1013 | |
| 1014 | def eval(self, a): |
| 1015 | |
| 1016 | # This operator is special. The indices and updates tensors really need |
| 1017 | # to be created together, but in the current structure of this tool there |
| 1018 | # is no way to do that before now. The number of updates is determined by |
| 1019 | # the indices, so we can really only create that after indices; but we |
| 1020 | # don't know the type at that time. |
| 1021 | # |
| 1022 | # Shapes are guaranteed deterministic, but we'll use our rng |
| 1023 | # copied from the arggen stage. It's possible that index and |
| 1024 | # update *values* will be non-deterministic. |
| 1025 | # |
| 1026 | # We take the tensor_tensor simply to get the dtype. |
| 1027 | |
| 1028 | shape_const = tf.constant(self.shape, tf.int32) |
| 1029 | |
| 1030 | updates_shape = list(self.indices_shape[:-1]) |
| 1031 | updates_shape.extend(self.shape[self.indices_shape[-1] :]) |
| 1032 | |
| 1033 | updates_const = tf.constant(TGen.getRand(updates_shape, a.dtype, self.rng)) |
| 1034 | |
| 1035 | indices = np.zeros(self.indices_shape, dtype=np.int32) |
| 1036 | |
| 1037 | # We need to generate the random indices tensor based on the |
| 1038 | # limits of 'shape' for each dimension. Surely, there is a faster |
| 1039 | # vectorized way to do this, but the tensors are fairly small so we |
| 1040 | # will do this one element at a time. Each element needs to be sized based |
| 1041 | # on the size of the last dimension. |
| 1042 | for idx in np.ndindex(indices.shape): |
| 1043 | indices[idx] = self.rng.integers(0, self.shape[idx[-1]], size=1)[0] |
| 1044 | # print('{} {}'.format(idx, indices[idx])) |
| 1045 | |
| 1046 | indices_const = tf.constant(indices, dtype=tf.int32) |
| 1047 | |
| 1048 | return tf.scatter_nd( |
| 1049 | indices=indices_const, |
| 1050 | updates=updates_const, |
| 1051 | shape=shape_const, |
| 1052 | name=self.result_name, |
| 1053 | ) |
| 1054 | |
| 1055 | class SpaceToBatch: |
| 1056 | def __init__(self, block_shape, padding, name): |
| 1057 | self.block_shape = block_shape |
| 1058 | self.padding = padding |
| 1059 | self.result_name = name |
| 1060 | |
| 1061 | def eval(self, a): |
| 1062 | return tf.space_to_batch( |
| 1063 | a, self.block_shape, self.padding, name=self.result_name |
| 1064 | ) |
| 1065 | |
TatWai Chong | bef907a | 2024-01-23 09:40:37 -0800 | [diff] [blame] | 1066 | class DynamicSpaceToBatch: |
| 1067 | def __init__(self, block_shape, padding, dynamic_input_shape, name): |
| 1068 | self.result_name = name |
| 1069 | |
| 1070 | dynamic_input_shape_with_batch = list(dynamic_input_shape) |
| 1071 | dynamic_input_shape_no_batch = dynamic_input_shape_with_batch[1:] |
| 1072 | dynamic_input_shape_no_batch = tuple(dynamic_input_shape_no_batch) |
| 1073 | |
| 1074 | self.model = tf.keras.Sequential( |
| 1075 | [ |
| 1076 | tf.keras.layers.Input(shape=dynamic_input_shape_no_batch), |
| 1077 | tf.keras.layers.Lambda( |
| 1078 | lambda x: tf.space_to_batch(x, block_shape, padding, name=None) |
| 1079 | ), |
| 1080 | ] |
| 1081 | ) |
| 1082 | |
| 1083 | def eval(self, a): |
| 1084 | return self.model(a) |
| 1085 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1086 | class BatchToSpace: |
| 1087 | def __init__(self, block_shape, cropping, name): |
| 1088 | self.block_shape = block_shape |
| 1089 | self.cropping = cropping |
| 1090 | self.result_name = name |
| 1091 | |
| 1092 | def eval(self, a): |
| 1093 | # transpose to swap depth and batch first. this could avoid adding new shape |
| 1094 | block_rank = len(self.block_shape) |
| 1095 | perm = [len(a.shape) - 1] |
| 1096 | for i in range(block_rank): |
| 1097 | perm.append(i + 1) |
| 1098 | perm.append(0) |
| 1099 | transpose_op = tf.transpose(a, perm) |
| 1100 | return tf.batch_to_space( |
| 1101 | transpose_op, self.block_shape, self.cropping, name=self.result_name |
| 1102 | ) |
| 1103 | |
Jerry Ge | 28811d9 | 2023-12-05 00:53:26 +0000 | [diff] [blame] | 1104 | class DynamicBatchToSpace: |
| 1105 | def __init__(self, block_shape, cropping, dynamic_input_shape, name): |
| 1106 | self.result_name = name |
| 1107 | |
| 1108 | dynamic_input_shape_with_batch = list(dynamic_input_shape) |
| 1109 | dynamic_input_shape_no_batch = dynamic_input_shape_with_batch[1:] |
| 1110 | dynamic_input_shape_no_batch = tuple(dynamic_input_shape_no_batch) |
| 1111 | |
| 1112 | self.model = tf.keras.Sequential( |
| 1113 | [ |
| 1114 | tf.keras.layers.Input(shape=dynamic_input_shape_no_batch), |
| 1115 | tf.keras.layers.Lambda( |
| 1116 | lambda x: tf.batch_to_space(x, block_shape, cropping, name=None) |
| 1117 | ), |
| 1118 | ] |
| 1119 | ) |
| 1120 | |
| 1121 | def eval(self, a): |
| 1122 | return self.model(a) |
| 1123 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1124 | class SpaceToDepth: |
| 1125 | def __init__(self, block_shape, name): |
| 1126 | self.block_shape = block_shape |
| 1127 | self.result_name = name |
| 1128 | |
| 1129 | def eval(self, a): |
| 1130 | return tf.nn.space_to_depth(a, self.block_shape, name=self.result_name) |
| 1131 | |
Jerry Ge | 28811d9 | 2023-12-05 00:53:26 +0000 | [diff] [blame] | 1132 | class DynamicSpaceToDepth: |
| 1133 | def __init__(self, dynamic_input_shape, name): |
| 1134 | self.result_name = name |
| 1135 | |
| 1136 | dynamic_input_shape_with_batch = list(dynamic_input_shape) |
| 1137 | dynamic_input_shape_no_batch = dynamic_input_shape_with_batch[1:] |
| 1138 | dynamic_input_shape_no_batch = tuple(dynamic_input_shape_no_batch) |
| 1139 | |
| 1140 | self.model = tf.keras.Sequential( |
| 1141 | [ |
| 1142 | tf.keras.layers.Input(shape=dynamic_input_shape_no_batch), |
| 1143 | tf.keras.layers.Lambda( |
| 1144 | lambda x: tf.nn.space_to_depth( |
| 1145 | x, 2, data_format="NHWC", name=None |
| 1146 | ) |
| 1147 | ), |
| 1148 | ] |
| 1149 | ) |
| 1150 | |
| 1151 | def eval(self, a): |
| 1152 | return self.model(a) |
| 1153 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1154 | class DepthToSpace: |
| 1155 | def __init__(self, block_shape, name): |
| 1156 | self.block_shape = block_shape |
| 1157 | self.result_name = name |
| 1158 | |
| 1159 | def eval(self, a): |
| 1160 | return tf.nn.depth_to_space(a, self.block_shape, name=self.result_name) |
| 1161 | |
Jerry Ge | 28811d9 | 2023-12-05 00:53:26 +0000 | [diff] [blame] | 1162 | class DynamicDepthToSpace: |
| 1163 | def __init__(self, dynamic_input_shape, name): |
| 1164 | self.result_name = name |
| 1165 | |
| 1166 | dynamic_input_shape_with_batch = list(dynamic_input_shape) |
| 1167 | dynamic_input_shape_no_batch = dynamic_input_shape_with_batch[1:] |
| 1168 | dynamic_input_shape_no_batch = tuple(dynamic_input_shape_no_batch) |
| 1169 | |
| 1170 | self.model = tf.keras.Sequential( |
| 1171 | [ |
| 1172 | tf.keras.layers.Input(shape=dynamic_input_shape_no_batch), |
| 1173 | tf.keras.layers.Lambda( |
| 1174 | lambda x: tf.nn.depth_to_space( |
| 1175 | x, 2, data_format="NHWC", name=None |
| 1176 | ) |
| 1177 | ), |
| 1178 | ] |
| 1179 | ) |
| 1180 | |
| 1181 | def eval(self, a): |
| 1182 | return self.model(a) |
| 1183 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1184 | class OneHot: |
| 1185 | def __init__(self, depth, axis, name): |
| 1186 | self.depth = depth |
| 1187 | self.axis = axis |
| 1188 | self.result_name = name |
| 1189 | |
| 1190 | def eval(self, indices, on_value, off_value): |
| 1191 | return tf.one_hot( |
| 1192 | indices, |
| 1193 | self.depth, |
| 1194 | on_value, |
| 1195 | off_value, |
| 1196 | self.axis, |
| 1197 | on_value.dtype, |
| 1198 | self.result_name, |
| 1199 | ) |
| 1200 | |
| 1201 | class Fakequant: |
| 1202 | def __init__(self, num_bits, narrow_range, name): |
| 1203 | self.num_bits = num_bits |
| 1204 | self.narrow_range = narrow_range |
| 1205 | self.result_name = name |
| 1206 | |
| 1207 | def eval(self, a): |
| 1208 | return tf.quantization.fake_quant_with_min_max_args( |
| 1209 | a, |
| 1210 | min=-2.0, |
| 1211 | max=2.0, |
| 1212 | num_bits=self.num_bits, |
| 1213 | narrow_range=self.narrow_range, |
| 1214 | name=self.result_name, |
| 1215 | ) |
| 1216 | |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1217 | class Resize: |
| 1218 | def __init__(self, mode, align, half, scale, name): |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1219 | self.result_name = name |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1220 | self.mode = mode |
| 1221 | self.align = align |
| 1222 | self.half = half |
| 1223 | self.scale = scale |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1224 | |
| 1225 | def eval(self, a): |
| 1226 | out_shape = [] |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1227 | out_shape.append(a.shape[1] * self.scale) |
| 1228 | out_shape.append(a.shape[2] * self.scale) |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1229 | |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1230 | tf_resize_dict = ( |
| 1231 | {"tf_resize_func": tf.compat.v1.image.resize_nearest_neighbor} |
| 1232 | if (self.mode == "nearest") |
| 1233 | else {"tf_resize_func": tf.compat.v1.image.resize_bilinear} |
| 1234 | ) |
| 1235 | resize = tf_resize_dict["tf_resize_func"]( |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1236 | a, |
| 1237 | out_shape, |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1238 | align_corners=self.align, |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1239 | name="resize", |
TatWai Chong | 0cef07e | 2023-02-27 13:22:52 -0800 | [diff] [blame] | 1240 | half_pixel_centers=self.half, |
TatWai Chong | f732609 | 2022-06-08 12:17:14 -0700 | [diff] [blame] | 1241 | ) |
| 1242 | return tf.identity(resize, name=self.result_name) |
| 1243 | |
Jeremy Johnson | 015c355 | 2022-02-23 12:15:03 +0000 | [diff] [blame] | 1244 | class LeftShift: |
| 1245 | def __init__(self, shift, name): |
| 1246 | self.shift = shift |
| 1247 | self.result_name = name |
| 1248 | |
| 1249 | def eval(self, a): |
| 1250 | return tf.bitwise.left_shift(a, self.shift, name=self.result_name) |
| 1251 | |
| 1252 | class RightShift: |
| 1253 | def __init__(self, shift, name): |
| 1254 | self.shift = shift |
| 1255 | self.result_name = name |
| 1256 | |
| 1257 | def eval(self, a): |
| 1258 | return tf.bitwise.right_shift(a, self.shift, name=self.result_name) |
Jerry Ge | 9e94af8 | 2022-10-27 09:57:00 -0700 | [diff] [blame] | 1259 | |
| 1260 | class While: |
| 1261 | def __init__(self, name): |
| 1262 | self.result_name = name |
| 1263 | |
| 1264 | def while_cond(self, x): |
| 1265 | return tf.reduce_sum(x) < self.cap |
| 1266 | |
| 1267 | def while_body(self, x): |
| 1268 | return tf.add(x, tf.math.sigmoid(x)) |
| 1269 | |
| 1270 | def eval(self, a): |
| 1271 | self.cap = tf.cast( |
| 1272 | tf.constant( |
| 1273 | 2.0, |
| 1274 | shape=[ |
| 1275 | 1, |
| 1276 | ], |
| 1277 | ), |
| 1278 | a.dtype, |
| 1279 | ) |
| 1280 | |
| 1281 | result = tf.while_loop( |
| 1282 | self.while_cond, self.while_body, [a], name=self.result_name |
| 1283 | ) |
| 1284 | |
| 1285 | return result[0] |
| 1286 | |
Tai Ly | cf84bc9 | 2023-09-07 20:49:09 +0000 | [diff] [blame] | 1287 | class LSTM(tf.Module): |
Jerry Ge | 9e94af8 | 2022-10-27 09:57:00 -0700 | [diff] [blame] | 1288 | def __init__(self, name): |
| 1289 | self.result_name = name |
| 1290 | self.lstm = tf.keras.layers.LSTM( |
| 1291 | 2, |
| 1292 | activation="tanh", |
| 1293 | unroll=False, |
| 1294 | recurrent_activation="sigmoid", |
| 1295 | use_bias=True, |
| 1296 | recurrent_initializer="ones", |
| 1297 | kernel_initializer="ones", |
| 1298 | ) |
| 1299 | |
| 1300 | def eval(self, a): |
| 1301 | return self.lstm(a) |
| 1302 | |
Tai Ly | cf84bc9 | 2023-09-07 20:49:09 +0000 | [diff] [blame] | 1303 | class SLSTM(tf.Module): |
| 1304 | def __init__(self, name): |
| 1305 | self.result_name = name |
| 1306 | self.lstm = tf.keras.layers.LSTM( |
| 1307 | 2, |
| 1308 | stateful=True, |
| 1309 | activation="tanh", |
| 1310 | unroll=False, |
| 1311 | recurrent_activation="sigmoid", |
| 1312 | use_bias=True, |
| 1313 | recurrent_initializer="ones", |
| 1314 | kernel_initializer="ones", |
| 1315 | ) |
| 1316 | |
| 1317 | def eval(self, a): |
| 1318 | return self.lstm(a) |
| 1319 | |
Jerry Ge | 9e94af8 | 2022-10-27 09:57:00 -0700 | [diff] [blame] | 1320 | class GRU: |
| 1321 | def __init__(self, name): |
| 1322 | self.result_name = name |
| 1323 | self.lstm = tf.keras.layers.GRU( |
| 1324 | 2, |
| 1325 | recurrent_activation="sigmoid", |
| 1326 | use_bias=True, |
| 1327 | recurrent_initializer="ones", |
| 1328 | kernel_initializer="ones", |
| 1329 | ) |
| 1330 | |
| 1331 | def eval(self, a): |
| 1332 | return self.lstm(a) |
| 1333 | |
| 1334 | class RNN: |
| 1335 | def __init__(self, name): |
| 1336 | self.result_name = name |
| 1337 | basic_cell = tf.keras.layers.SimpleRNNCell( |
| 1338 | units=2, |
| 1339 | activation="sigmoid", |
| 1340 | use_bias=True, |
| 1341 | recurrent_initializer="ones", |
| 1342 | ) |
| 1343 | self.rnn = tf.keras.layers.RNN(basic_cell, unroll=False) |
| 1344 | |
| 1345 | def eval(self, a): |
| 1346 | return self.rnn(a) |
| 1347 | |
| 1348 | class FullyConnected: |
| 1349 | def __init__(self, name): |
| 1350 | self.result_name = name |
| 1351 | self.dense = tf.keras.layers.Dense(2) |
| 1352 | |
| 1353 | def eval(self, a): |
| 1354 | return self.dense(a) |
Luke Hutton | 261b7b6 | 2023-01-10 14:50:31 +0000 | [diff] [blame] | 1355 | |
| 1356 | class RFFT2d: |
| 1357 | def __init__(self, fft_length, name): |
| 1358 | self.fft_length = fft_length |
| 1359 | self.result_name = name |
| 1360 | |
| 1361 | def eval(self, a): |
| 1362 | return tf.signal.rfft2d(a, self.fft_length, name=self.result_name) |
Luke Hutton | 714aa60 | 2023-02-08 19:45:26 +0000 | [diff] [blame] | 1363 | |
| 1364 | class Real: |
| 1365 | def __init__(self, name): |
| 1366 | self.result_name = name |
| 1367 | |
| 1368 | def eval(self, a): |
| 1369 | return tf.math.real(a, name=self.result_name) |
| 1370 | |
| 1371 | class Imag: |
| 1372 | def __init__(self, name): |
| 1373 | self.result_name = name |
| 1374 | |
| 1375 | def eval(self, a): |
| 1376 | return tf.math.imag(a, name=self.result_name) |
Tai Ly | fe36fa9 | 2023-06-01 21:45:12 +0000 | [diff] [blame] | 1377 | |
| 1378 | class BroadcastTo: |
| 1379 | def __init__(self, shape, name): |
| 1380 | self.shape = shape |
| 1381 | self.result_name = name |
| 1382 | |
| 1383 | def eval(self, a): |
| 1384 | return tf.broadcast_to(a, shape=self.shape, name=self.result_name) |
Tai Ly | cf84bc9 | 2023-09-07 20:49:09 +0000 | [diff] [blame] | 1385 | |
| 1386 | class CallOnce(tf.Module): |
| 1387 | def __init__(self, name): |
| 1388 | print(tf.__version__) |
| 1389 | self.result_name = name |
| 1390 | self.var = tf.Variable([1.0]) |
| 1391 | |
| 1392 | @tf.function( |
| 1393 | input_signature=[ |
| 1394 | tf.TensorSpec( |
| 1395 | shape=[ |
| 1396 | 1, |
| 1397 | ], |
| 1398 | dtype=tf.float32, |
| 1399 | ) |
| 1400 | ] |
| 1401 | ) |
| 1402 | def eval(self, a): |
| 1403 | return self.var.assign([2.0]) |