1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
|
# Owner(s): ["module: scatter & gather ops"]
from itertools import product
from functools import partial
import numpy as np
import torch
from torch.testing._internal.common_device_type import (
instantiate_device_type_tests,
dtypes,
)
from torch.testing._internal.common_utils import (
TestCase,
run_tests,
gradcheck,
parametrize,
)
reductions = ["max", "mean", "min", "sum", "prod"]
def get_default_value(initial_value, reduction):
if initial_value is not None:
return initial_value
if reduction == "max":
return -float("Inf")
elif reduction == "mean":
return float("nan")
elif reduction == "min":
return float("Inf")
elif reduction == "sum":
return 0.0
elif reduction == "prod":
return 1.0
class TestSegmentReductions(TestCase):
def _test_common(
self,
reduction,
device,
dtype,
unsafe,
axis,
initial_value,
data_arr,
lengths_arr,
expected_arr,
expected_grad_arr,
check_backward,
lengths_dtype=torch.int,
):
lengths = torch.tensor(lengths_arr, device=device, dtype=lengths_dtype)
# generate offsets from lengths
zeros_shape = list(lengths.shape)
zeros_shape[-1] = 1
offsets = torch.cat((lengths.new_zeros(zeros_shape), lengths), -1).cumsum_(-1)
data = torch.tensor(
data_arr,
device=device,
dtype=dtype,
requires_grad=True,
)
expected_result = torch.tensor(expected_arr, device=device, dtype=dtype)
expected_grad = torch.tensor(expected_grad_arr, device=device, dtype=dtype)
for mode in ['lengths', 'offsets']:
segment_reduce_kwargs = dict(
axis=axis,
unsafe=unsafe,
initial=initial_value)
if (mode == 'lengths'):
segment_reduce_kwargs['lengths'] = lengths
else:
segment_reduce_kwargs['offsets'] = offsets
actual_result = torch.segment_reduce(
data=data,
reduce=reduction,
**segment_reduce_kwargs
)
self.assertEqual(
expected_result, actual_result, rtol=1e-02, atol=1e-05, equal_nan=True
)
if not check_backward:
return
# Test backward
actual_result.sum().backward()
self.assertEqual(
expected_grad, data.grad, rtol=1e-02, atol=1e-05, equal_nan=True
)
data = data.clone().detach().requires_grad_(True)
# gradcheck does not work well with bfloat16 or fp16 cpu types
# also there is small numerical difference with fp32
if dtype not in [torch.half, torch.bfloat16, torch.float]:
# gradcheck does not like "nan" input, setting to random 10
d_non_nan = np.nan_to_num(data_arr, nan=10)
new_data = torch.tensor(
# [10 if v == float("nan") else v for v in data],
d_non_nan,
device=device,
dtype=dtype,
requires_grad=True,
)
self.assertTrue(
gradcheck(
lambda x: torch.segment_reduce(
data=x,
reduce=reduction,
**segment_reduce_kwargs
),
(new_data,),
)
)
@dtypes(
*product(
(torch.half, torch.bfloat16, torch.float, torch.double),
(torch.int, torch.int64),
)
)
def test_simple_1d(self, device, dtypes):
val_dtype, length_type = dtypes
lengths = [1, 2, 3, 0]
data = [1, float("nan"), 3, 4, 5, 5]
for reduction in reductions:
for initial in [0, None]:
check_backward = True if initial is not None else False
initial_value = initial
default_value = get_default_value(initial_value, reduction)
if reduction == "max":
expected_result = [1, float("nan"), 5, default_value]
expected_grad = [1, 1, 0, 0, 0.5, 0.5]
elif reduction == "mean":
expected_result = [1, float("nan"), 4.666, default_value]
expected_grad = [1.0, 0.5, 0.5, 0.333, 0.333, 0.333]
elif reduction == "min":
if initial is not None:
initial_value = 1000 # some high number
default_value = get_default_value(initial_value, reduction)
expected_result = [1, float("nan"), 4, default_value]
expected_grad = [1.0, 1.0, 0, 1, 0, 0]
elif reduction == "sum":
expected_result = [1, float("nan"), 14, default_value]
expected_grad = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
elif reduction == "prod":
if initial is not None:
initial_value = 2 # 0 initial_value will zero out everything for prod
default_value = get_default_value(initial_value, reduction)
expected_result = [2, float("nan"), 200, default_value]
expected_grad = [2.0, 6.0, float("nan"), 50.0, 40.0, 40.0]
else:
expected_result = [1, float("nan"), 100, default_value]
expected_grad = [1.0, 3.0, float("nan"), 25.0, 20.0, 20.0]
for axis in [0, -1]:
for unsafe in [True, False]:
self._test_common(
reduction,
device,
val_dtype,
unsafe,
axis,
initial_value,
data,
lengths,
expected_result,
expected_grad,
check_backward,
length_type,
)
@dtypes(
*product(
(torch.half, torch.bfloat16, torch.float, torch.double),
(torch.int, torch.int64),
)
)
def test_multi_d_simple(self, device, dtypes):
val_dtype, length_type = dtypes
axis = 0
lengths = [1, 2, 3, 0]
data = [[1, 1], [float("nan"), 1], [3, float("nan")], [4, 1], [3, 2], [2, 3]]
for reduction in reductions:
for initial in [0, None]:
check_backward = True if initial is not None else False
initial_value = initial
default_value = get_default_value(initial_value, reduction)
if reduction == "max":
expected_result = [
[1, 1],
[float("nan"), float("nan")],
[4, 3],
[default_value, default_value],
]
expected_grad = [
[1, 1],
[1, 0],
[0, 1],
[1, 0],
[0, 0],
[0, 1],
]
elif reduction == "mean":
expected_result = [
[1, 1],
[float("nan"), float("nan")],
[3, 2],
[default_value, default_value],
]
expected_grad = [
[1.0, 1.0],
[0.5, 0.5],
[0.5, 0.5],
[0.333, 0.333],
[0.333, 0.333],
[0.333, 0.333],
]
elif reduction == "min":
if initial is not None:
initial_value = 1000 # some high number
default_value = get_default_value(initial_value, reduction)
expected_result = [
[1, 1],
[float("nan"), float("nan")],
[2, 1],
[default_value, default_value],
]
expected_grad = [
[1.0, 1.0],
[1, 0],
[0, 1],
[0, 1],
[0, 0],
[1, 0],
]
elif reduction == "sum":
expected_result = [
[1, 1],
[float("nan"), float("nan")],
[9, 6],
[default_value, default_value],
]
expected_grad = [
[1.0, 1.0],
[1.0, 1.0],
[1.0, 1.0],
[1.0, 1.0],
[1.0, 1.0],
[1.0, 1.0],
]
elif reduction == "prod":
if initial is not None:
initial_value = 2 # 0 initial_value will zero out everything for prod
default_value = get_default_value(initial_value, reduction)
expected_result = [
[2, 2],
[float("nan"), float("nan")],
[48, 12],
[default_value, default_value],
]
expected_grad = [
[2.0, 2.0],
[6.0, float("nan")],
[float("nan"), 2.0],
[12.0, 12.0],
[16.0, 6.0],
[24.0, 4.0],
]
else:
expected_result = [
[1, 1],
[float("nan"), float("nan")],
[24, 6],
[default_value, default_value],
]
expected_grad = [
[1.0, 1.0],
[3.0, float("nan")],
[float("nan"), 1.0],
[6.0, 6.0],
[8.0, 3.0],
[12.0, 2.0],
]
for unsafe in [True, False]:
self._test_common(
reduction,
device,
val_dtype,
unsafe,
axis,
initial_value,
data,
lengths,
expected_result,
expected_grad,
check_backward,
)
@dtypes(
*product(
(torch.half, torch.bfloat16, torch.float, torch.double),
(torch.int, torch.int64),
)
)
@parametrize("reduce", ['sum', 'prod', 'min', 'max', 'mean'])
def test_pytorch_scatter_test_cases(self, device, dtypes, reduce):
val_dtype, length_dtype = dtypes
# zero-length segments are filled with reduction inits contrary to pytorch_scatter.
tests = [
{
'src': [1, 2, 3, 4, 5, 6],
'index': [0, 0, 1, 1, 1, 3],
'indptr': [0, 2, 5, 5, 6],
'sum': [3, 12, 0, 6],
'prod': [2, 60, 1, 6],
'mean': [1.5, 4, float('nan'), 6],
'min': [1, 3, float('inf'), 6],
'max': [2, 5, -float('inf'), 6],
},
{
'src': [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]],
'index': [0, 0, 1, 1, 1, 3],
'indptr': [0, 2, 5, 5, 6],
'sum': [[4, 6], [21, 24], [0, 0], [11, 12]],
'prod': [[3, 8], [315, 480], [1, 1], [11, 12]],
'mean': [[2, 3], [7, 8], [float('nan'), float('nan')], [11, 12]],
'min': [[1, 2], [5, 6], [float('inf'), float('inf')], [11, 12]],
'max': [[3, 4], [9, 10], [-float('inf'), -float('inf')], [11, 12]],
},
{
'src': [[1, 3, 5, 7, 9, 11], [2, 4, 6, 8, 10, 12]],
'index': [[0, 0, 1, 1, 1, 3], [0, 0, 0, 1, 1, 2]],
'indptr': [[0, 2, 5, 5, 6], [0, 3, 5, 6, 6]],
'sum': [[4, 21, 0, 11], [12, 18, 12, 0]],
'prod': [[3, 315, 1, 11], [48, 80, 12, 1]],
'mean': [[2, 7, float('nan'), 11], [4, 9, 12, float('nan')]],
'min': [[1, 5, float('inf'), 11], [2, 8, 12, float('inf')]],
'max': [[3, 9, -float('inf'), 11], [6, 10, 12, -float('inf')]],
},
{
'src': [[[1, 2], [3, 4], [5, 6]], [[7, 9], [10, 11], [12, 13]]],
'index': [[0, 0, 1], [0, 2, 2]],
'indptr': [[0, 2, 3, 3], [0, 1, 1, 3]],
'sum': [[[4, 6], [5, 6], [0, 0]], [[7, 9], [0, 0], [22, 24]]],
'prod': [[[3, 8], [5, 6], [1, 1]], [[7, 9], [1, 1], [120, 143]]],
'mean': [[[2, 3], [5, 6], [float('nan'), float('nan')]],
[[7, 9], [float('nan'), float('nan')], [11, 12]]],
'min': [[[1, 2], [5, 6], [float('inf'), float('inf')]],
[[7, 9], [float('inf'), float('inf')], [10, 11]]],
'max': [[[3, 4], [5, 6], [-float('inf'), -float('inf')]],
[[7, 9], [-float('inf'), -float('inf')], [12, 13]]],
},
{
'src': [[1, 3], [2, 4]],
'index': [[0, 0], [0, 0]],
'indptr': [[0, 2], [0, 2]],
'sum': [[4], [6]],
'prod': [[3], [8]],
'mean': [[2], [3]],
'min': [[1], [2]],
'max': [[3], [4]],
},
{
'src': [[[1, 1], [3, 3]], [[2, 2], [4, 4]]],
'index': [[0, 0], [0, 0]],
'indptr': [[0, 2], [0, 2]],
'sum': [[[4, 4]], [[6, 6]]],
'prod': [[[3, 3]], [[8, 8]]],
'mean': [[[2, 2]], [[3, 3]]],
'min': [[[1, 1]], [[2, 2]]],
'max': [[[3, 3]], [[4, 4]]],
},
]
for test in tests:
data = torch.tensor(test['src'], dtype=val_dtype, device=device, requires_grad=True)
indptr = torch.tensor(test['indptr'], dtype=length_dtype, device=device)
dim = indptr.ndim - 1
# calculate lengths from indptr
lengths = torch.diff(indptr, dim=dim)
expected = torch.tensor(test[reduce], dtype=val_dtype, device=device)
actual_result = torch.segment_reduce(
data=data,
reduce=reduce,
lengths=lengths,
axis=dim,
unsafe=True,
)
self.assertEqual(actual_result, expected)
# test offsets
actual_result = torch.segment_reduce(
data=data,
reduce=reduce,
offsets=indptr,
axis=dim,
unsafe=True,
)
self.assertEqual(actual_result, expected)
if val_dtype == torch.float64:
def fn(x, mode='lengths'):
initial = 1
# supply initial values to prevent gradcheck from failing for 0 length segments
# where nan/inf are reduction identities that produce nans when calculating the numerical jacobian
if reduce == 'min':
initial = 1000
elif reduce == 'max':
initial = -1000
segment_reduce_args = {x, reduce}
segment_reduce_kwargs = dict(axis=dim, unsafe=True, initial=initial)
if mode == 'lengths':
segment_reduce_kwargs[mode] = lengths
elif mode == 'offsets':
segment_reduce_kwargs[mode] = indptr
return torch.segment_reduce(*segment_reduce_args, **segment_reduce_kwargs)
self.assertTrue(gradcheck(partial(fn, mode='lengths'), (data.clone().detach().requires_grad_(True))))
self.assertTrue(gradcheck(partial(fn, mode='offsets'), (data.clone().detach().requires_grad_(True))))
@dtypes(
*product(
(torch.half, torch.bfloat16, torch.float, torch.double),
(torch.int, torch.int64),
)
)
def test_multi_d(self, device, dtypes):
val_dtype, length_type = dtypes
axis = 0
lengths = [0, 2, 3, 0]
data = np.arange(50).reshape(5, 2, 5).tolist()
expected_grad = []
# TODO: calculate grad and check correctness
check_backward = False
for reduction in reductions:
initial_value = 0
if reduction == "max":
expected_result = [
np.full((2, 5), initial_value).tolist(),
np.max(data[:2], axis=0).tolist(),
np.max(data[2:], axis=0).tolist(),
np.full((2, 5), initial_value).tolist(),
]
elif reduction == "mean":
expected_result = [
np.full((2, 5), initial_value).tolist(),
np.mean(data[:2], axis=0).tolist(),
np.mean(data[2:], axis=0).tolist(),
np.full((2, 5), initial_value).tolist(),
]
elif reduction == "min":
initial_value = 1000 # some high number
expected_result = [
np.full((2, 5), initial_value).tolist(),
np.min(data[:2], axis=0).tolist(),
np.min(data[2:], axis=0).tolist(),
np.full((2, 5), initial_value).tolist(),
]
elif reduction == "sum":
expected_result = [
np.full((2, 5), initial_value).tolist(),
np.sum(data[:2], axis=0).tolist(),
np.sum(data[2:], axis=0).tolist(),
np.full((2, 5), initial_value).tolist(),
]
elif reduction == "prod":
initial_value = 1
expected_result = [
np.full((2, 5), initial_value).tolist(),
np.prod(data[:2], axis=0).tolist(),
np.prod(data[2:], axis=0).tolist(),
np.full((2, 5), initial_value).tolist(),
]
for unsafe in [True, False]:
self._test_common(
reduction,
device,
val_dtype,
unsafe,
axis,
initial_value,
data,
lengths,
expected_result,
expected_grad,
check_backward,
)
@dtypes(torch.int, torch.int64)
def test_unsafe_flag(self, device, dtype):
length_type = dtype
lengths = torch.tensor([0, 2, 3, 0], device=device, dtype=length_type)
data = torch.arange(6, dtype=torch.float, device=device)
# test for error on 1-D lenghts
with self.assertRaisesRegex(RuntimeError, "Expected all rows of lengths along axis"):
torch.segment_reduce(data, 'sum', lengths=lengths, axis=0, unsafe=False)
# test for error on multi-D lengths
nd_lengths = torch.tensor([[0, 3, 3, 0], [2, 3, 0, 0]], dtype=length_type, device=device)
nd_data = torch.arange(12, dtype=torch.float, device=device).reshape(2, 6)
with self.assertRaisesRegex(RuntimeError, "Expected all rows of lengths along axis"):
torch.segment_reduce(nd_data, 'sum', lengths=nd_lengths, axis=1, unsafe=False)
instantiate_device_type_tests(TestSegmentReductions, globals())
if __name__ == "__main__":
run_tests()
|