-
-
Notifications
You must be signed in to change notification settings - Fork 1.9k
Expand file tree
/
Copy pathreductions.py
More file actions
1813 lines (1542 loc) · 53.5 KB
/
reductions.py
File metadata and controls
1813 lines (1542 loc) · 53.5 KB
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
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import annotations
import builtins
import inspect
import math
import operator
import warnings
from collections.abc import Iterable
from functools import partial, reduce
from itertools import product, repeat
from numbers import Integral, Number
from operator import mul
import numpy as np
from packaging.version import Version
from tlz import accumulate, drop, pluck
import dask.array as da
from dask.array import chunk
from dask.array.core import (
Array,
_concatenate2,
handle_out,
implements,
unknown_chunk_message,
)
from dask.array.creation import arange, diagonal
from dask.array.dispatch import divide_lookup, nannumel_lookup, numel_lookup
from dask.array.numpy_compat import NUMPY_GE_200
from dask.array.utils import (
array_safe,
asarray_safe,
is_arraylike,
meta_from_array,
validate_axis,
)
from dask.array.wrap import ones, zeros
from dask.base import tokenize
from dask.highlevelgraph import HighLevelGraph
from dask.utils import apply, deepmap, derived_from
try:
import numbagg
except ImportError:
numbagg = None
if da._array_expr_enabled():
from dask.array._array_expr import _tree_reduce, reduction
else:
from dask.array._reductions_generic import _tree_reduce, reduction
def divide(a, b, dtype=None):
key = lambda x: getattr(x, "__array_priority__", float("-inf"))
f = divide_lookup.dispatch(type(builtins.max(a, b, key=key)))
return f(a, b, dtype=dtype)
def numel(x, **kwargs):
return numel_lookup(x, **kwargs)
def nannumel(x, **kwargs):
return nannumel_lookup(x, **kwargs)
@derived_from(np)
def sum(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is None:
dtype = getattr(np.zeros(1, dtype=a.dtype).sum(), "dtype", object)
result = reduction(
a,
chunk.sum,
chunk.sum,
axis=axis,
keepdims=keepdims,
dtype=dtype,
split_every=split_every,
out=out,
)
return result
@derived_from(np)
def prod(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is not None:
dt = dtype
else:
dt = getattr(np.ones((1,), dtype=a.dtype).prod(), "dtype", object)
return reduction(
a,
chunk.prod,
chunk.prod,
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
out=out,
)
@implements(np.min, np.amin)
@derived_from(np)
def min(a, axis=None, keepdims=False, split_every=None, out=None):
return reduction(
a,
chunk_min,
chunk.min,
combine=chunk_min,
axis=axis,
keepdims=keepdims,
dtype=a.dtype,
split_every=split_every,
out=out,
)
def chunk_min(x, axis=None, keepdims=None):
"""Version of np.min which ignores size 0 arrays"""
if x.size == 0:
return array_safe([], x, ndmin=x.ndim, dtype=x.dtype)
else:
return np.min(x, axis=axis, keepdims=keepdims)
@implements(np.max, np.amax)
@derived_from(np)
def max(a, axis=None, keepdims=False, split_every=None, out=None):
return reduction(
a,
chunk_max,
chunk.max,
combine=chunk_max,
axis=axis,
keepdims=keepdims,
dtype=a.dtype,
split_every=split_every,
out=out,
)
def chunk_max(x, axis=None, keepdims=None):
"""Version of np.max which ignores size 0 arrays"""
if x.size == 0:
return array_safe([], x, ndmin=x.ndim, dtype=x.dtype)
else:
return np.max(x, axis=axis, keepdims=keepdims)
@derived_from(np)
def any(a, axis=None, keepdims=False, split_every=None, out=None):
return reduction(
a,
chunk.any,
chunk.any,
axis=axis,
keepdims=keepdims,
dtype="bool",
split_every=split_every,
out=out,
)
@derived_from(np)
def all(a, axis=None, keepdims=False, split_every=None, out=None):
return reduction(
a,
chunk.all,
chunk.all,
axis=axis,
keepdims=keepdims,
dtype="bool",
split_every=split_every,
out=out,
)
@derived_from(np)
def nansum(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is not None:
dt = dtype
else:
dt = getattr(chunk.nansum(np.ones((1,), dtype=a.dtype)), "dtype", object)
return reduction(
a,
chunk.nansum,
chunk.sum,
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
out=out,
)
@derived_from(np)
def nanprod(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is not None:
dt = dtype
else:
dt = getattr(chunk.nansum(np.ones((1,), dtype=a.dtype)), "dtype", object)
return reduction(
a,
chunk.nanprod,
chunk.prod,
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
out=out,
)
@derived_from(np)
def nancumsum(x, axis, dtype=None, out=None, *, method="sequential"):
"""Dask added an additional keyword-only argument ``method``.
method : {'sequential', 'blelloch'}, optional
Choose which method to use to perform the cumsum. Default is 'sequential'.
* 'sequential' performs the cumsum of each prior block before the current block.
* 'blelloch' is a work-efficient parallel cumsum. It exposes parallelism by
first taking the sum of each block and combines the sums via a binary tree.
This method may be faster or more memory efficient depending on workload,
scheduler, and hardware. More benchmarking is necessary.
"""
return cumreduction(
chunk.nancumsum,
operator.add,
0,
x,
axis,
dtype,
out=out,
method=method,
preop=np.nansum,
)
@derived_from(np)
def nancumprod(x, axis, dtype=None, out=None, *, method="sequential"):
"""Dask added an additional keyword-only argument ``method``.
method : {'sequential', 'blelloch'}, optional
Choose which method to use to perform the cumprod. Default is 'sequential'.
* 'sequential' performs the cumprod of each prior block before the current block.
* 'blelloch' is a work-efficient parallel cumprod. It exposes parallelism by first
taking the product of each block and combines the products via a binary tree.
This method may be faster or more memory efficient depending on workload,
scheduler, and hardware. More benchmarking is necessary.
"""
return cumreduction(
chunk.nancumprod,
operator.mul,
1,
x,
axis,
dtype,
out=out,
method=method,
preop=np.nanprod,
)
@derived_from(np)
def nanmin(a, axis=None, keepdims=False, split_every=None, out=None):
if np.isnan(a.size):
raise ValueError(f"Arrays chunk sizes are unknown. {unknown_chunk_message}")
if a.size == 0:
raise ValueError(
"zero-size array to reduction operation fmin which has no identity"
)
return reduction(
a,
_nanmin_skip,
_nanmin_skip,
axis=axis,
keepdims=keepdims,
dtype=a.dtype,
split_every=split_every,
out=out,
)
def _nanmin_skip(x_chunk, axis, keepdims):
if x_chunk.size > 0:
with warnings.catch_warnings():
warnings.filterwarnings(
"ignore", "All-NaN slice encountered", RuntimeWarning
)
return np.nanmin(x_chunk, axis=axis, keepdims=keepdims)
else:
return asarray_safe(
np.array([], dtype=x_chunk.dtype), like=meta_from_array(x_chunk)
)
@derived_from(np)
def nanmax(a, axis=None, keepdims=False, split_every=None, out=None):
if np.isnan(a.size):
raise ValueError(f"Arrays chunk sizes are unknown. {unknown_chunk_message}")
if a.size == 0:
raise ValueError(
"zero-size array to reduction operation fmax which has no identity"
)
return reduction(
a,
_nanmax_skip,
_nanmax_skip,
axis=axis,
keepdims=keepdims,
dtype=a.dtype,
split_every=split_every,
out=out,
)
def _nanmax_skip(x_chunk, axis, keepdims):
if x_chunk.size > 0:
with warnings.catch_warnings():
warnings.filterwarnings(
"ignore", "All-NaN slice encountered", RuntimeWarning
)
return np.nanmax(x_chunk, axis=axis, keepdims=keepdims)
else:
return asarray_safe(
np.array([], dtype=x_chunk.dtype), like=meta_from_array(x_chunk)
)
def mean_chunk(
x, sum=chunk.sum, numel=numel, dtype="f8", computing_meta=False, **kwargs
):
if computing_meta:
return x
n = numel(x, dtype=dtype, **kwargs)
total = sum(x, dtype=dtype, **kwargs)
return {"n": n, "total": total}
def mean_combine(
pairs,
sum=chunk.sum,
numel=numel,
dtype="f8",
axis=None,
computing_meta=False,
**kwargs,
):
if not isinstance(pairs, list):
pairs = [pairs]
ns = deepmap(lambda pair: pair["n"], pairs) if not computing_meta else pairs
n = _concatenate2(ns, axes=axis).sum(axis=axis, **kwargs)
if computing_meta:
return n
totals = deepmap(lambda pair: pair["total"], pairs)
total = _concatenate2(totals, axes=axis).sum(axis=axis, **kwargs)
return {"n": n, "total": total}
def mean_agg(pairs, dtype="f8", axis=None, computing_meta=False, **kwargs):
ns = deepmap(lambda pair: pair["n"], pairs) if not computing_meta else pairs
n = _concatenate2(ns, axes=axis)
n = np.sum(n, axis=axis, dtype=dtype, **kwargs)
if computing_meta:
return n
totals = deepmap(lambda pair: pair["total"], pairs)
total = _concatenate2(totals, axes=axis).sum(axis=axis, dtype=dtype, **kwargs)
with np.errstate(divide="ignore", invalid="ignore"):
return divide(total, n, dtype=dtype)
@derived_from(np)
def mean(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is not None:
dt = dtype
elif a.dtype == object:
dt = object
else:
dt = getattr(np.mean(np.zeros(shape=(1,), dtype=a.dtype)), "dtype", object)
return reduction(
a,
mean_chunk,
mean_agg,
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
combine=mean_combine,
out=out,
concatenate=False,
)
@derived_from(np)
def nanmean(a, axis=None, dtype=None, keepdims=False, split_every=None, out=None):
if dtype is not None:
dt = dtype
else:
dt = getattr(np.mean(np.ones(shape=(1,), dtype=a.dtype)), "dtype", object)
return reduction(
a,
partial(mean_chunk, sum=chunk.nansum, numel=nannumel),
mean_agg,
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
out=out,
concatenate=False,
combine=partial(mean_combine, sum=chunk.nansum, numel=nannumel),
)
def moment_chunk(
A,
order=2,
sum=chunk.sum,
numel=numel,
dtype="f8",
computing_meta=False,
implicit_complex_dtype=False,
**kwargs,
):
if computing_meta:
return A
n = numel(A, **kwargs)
n = n.astype(np.int64)
if implicit_complex_dtype:
total = sum(A, **kwargs)
else:
total = sum(A, dtype=dtype, **kwargs)
with np.errstate(divide="ignore", invalid="ignore"):
u = total / n
d = A - u
if np.issubdtype(A.dtype, np.complexfloating):
d = np.abs(d)
xs = [sum(d**i, dtype=dtype, **kwargs) for i in range(2, order + 1)]
M = np.stack(xs, axis=-1)
return {"total": total, "n": n, "M": M}
def _moment_helper(Ms, ns, inner_term, order, sum, axis, kwargs):
M = Ms[..., order - 2].sum(axis=axis, **kwargs) + sum(
ns * inner_term**order, axis=axis, **kwargs
)
for k in range(1, order - 1):
coeff = math.factorial(order) / (math.factorial(k) * math.factorial(order - k))
M += coeff * sum(Ms[..., order - k - 2] * inner_term**k, axis=axis, **kwargs)
return M
def moment_combine(
pairs,
order=2,
ddof=0,
dtype="f8",
sum=np.sum,
axis=None,
computing_meta=False,
**kwargs,
):
if not isinstance(pairs, list):
pairs = [pairs]
kwargs["dtype"] = None
kwargs["keepdims"] = True
ns = deepmap(lambda pair: pair["n"], pairs) if not computing_meta else pairs
ns = _concatenate2(ns, axes=axis)
n = ns.sum(axis=axis, **kwargs)
if computing_meta:
return n
totals = _concatenate2(deepmap(lambda pair: pair["total"], pairs), axes=axis)
Ms = _concatenate2(deepmap(lambda pair: pair["M"], pairs), axes=axis)
total = totals.sum(axis=axis, **kwargs)
with np.errstate(divide="ignore", invalid="ignore"):
if np.issubdtype(total.dtype, np.complexfloating):
mu = divide(total, n)
inner_term = np.abs(divide(totals, ns) - mu)
else:
mu = divide(total, n, dtype=dtype)
inner_term = divide(totals, ns, dtype=dtype) - mu
xs = [
_moment_helper(Ms, ns, inner_term, o, sum, axis, kwargs)
for o in range(2, order + 1)
]
M = np.stack(xs, axis=-1)
return {"total": total, "n": n, "M": M}
def moment_agg(
pairs,
order=2,
ddof=0,
dtype="f8",
sum=np.sum,
axis=None,
computing_meta=False,
**kwargs,
):
if not isinstance(pairs, list):
pairs = [pairs]
kwargs["dtype"] = dtype
# To properly handle ndarrays, the original dimensions need to be kept for
# part of the calculation.
keepdim_kw = kwargs.copy()
keepdim_kw["keepdims"] = True
keepdim_kw["dtype"] = None
ns = deepmap(lambda pair: pair["n"], pairs) if not computing_meta else pairs
ns = _concatenate2(ns, axes=axis)
n = ns.sum(axis=axis, **keepdim_kw)
if computing_meta:
return n
totals = _concatenate2(deepmap(lambda pair: pair["total"], pairs), axes=axis)
Ms = _concatenate2(deepmap(lambda pair: pair["M"], pairs), axes=axis)
mu = divide(totals.sum(axis=axis, **keepdim_kw), n)
with np.errstate(divide="ignore", invalid="ignore"):
if np.issubdtype(totals.dtype, np.complexfloating):
inner_term = np.abs(divide(totals, ns) - mu)
else:
inner_term = divide(totals, ns, dtype=dtype) - mu
inner_term = np.where(ns == 0, 0, inner_term)
M = _moment_helper(Ms, ns, inner_term, order, sum, axis, kwargs)
denominator = n.sum(axis=axis, **kwargs) - ddof
# taking care of the edge case with empty or all-nans array with ddof > 0
if isinstance(denominator, Number):
if denominator < 0:
denominator = np.nan
elif denominator is not np.ma.masked:
denominator[denominator < 0] = np.nan
return divide(M, denominator, dtype=dtype)
def moment(
a, order, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None
):
"""Calculate the nth centralized moment.
Parameters
----------
a : Array
Data over which to compute moment
order : int
Order of the moment that is returned, must be >= 2.
axis : int, optional
Axis along which the central moment is computed. The default is to
compute the moment of the flattened array.
dtype : data-type, optional
Type to use in computing the moment. For arrays of integer type the
default is float64; for arrays of float types it is the same as the
array type.
keepdims : bool, optional
If this is set to True, the axes which are reduced are left in the
result as dimensions with size one. With this option, the result
will broadcast correctly against the original array.
ddof : int, optional
"Delta Degrees of Freedom": the divisor used in the calculation is
N - ddof, where N represents the number of elements. By default
ddof is zero.
Returns
-------
moment : Array
References
----------
.. [1] Pebay, Philippe (2008), "Formulas for Robust, One-Pass Parallel
Computation of Covariances and Arbitrary-Order Statistical Moments",
Technical Report SAND2008-6212, Sandia National Laboratories.
"""
if not isinstance(order, Integral) or order < 0:
raise ValueError("Order must be an integer >= 0")
if order < 2:
reduced = a.sum(axis=axis) # get reduced shape and chunks
if order == 0:
# When order equals 0, the result is 1, by definition.
return ones(
reduced.shape, chunks=reduced.chunks, dtype="f8", meta=reduced._meta
)
# By definition the first order about the mean is 0.
return zeros(
reduced.shape, chunks=reduced.chunks, dtype="f8", meta=reduced._meta
)
if dtype is not None:
dt = dtype
else:
dt = getattr(np.var(np.ones(shape=(1,), dtype=a.dtype)), "dtype", object)
implicit_complex_dtype = dtype is None and np.iscomplexobj(a)
return reduction(
a,
partial(
moment_chunk, order=order, implicit_complex_dtype=implicit_complex_dtype
),
partial(moment_agg, order=order, ddof=ddof),
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
out=out,
concatenate=False,
combine=partial(moment_combine, order=order),
)
@derived_from(np)
def var(a, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None):
if dtype is not None:
dt = dtype
else:
dt = getattr(np.var(np.ones(shape=(1,), dtype=a.dtype)), "dtype", object)
implicit_complex_dtype = dtype is None and np.iscomplexobj(a)
return reduction(
a,
partial(moment_chunk, implicit_complex_dtype=implicit_complex_dtype),
partial(moment_agg, ddof=ddof),
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
combine=moment_combine,
name="var",
out=out,
concatenate=False,
)
@derived_from(np)
def nanvar(
a, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None
):
if dtype is not None:
dt = dtype
else:
dt = getattr(np.var(np.ones(shape=(1,), dtype=a.dtype)), "dtype", object)
implicit_complex_dtype = dtype is None and np.iscomplexobj(a)
return reduction(
a,
partial(
moment_chunk,
sum=chunk.nansum,
numel=nannumel,
implicit_complex_dtype=implicit_complex_dtype,
),
partial(moment_agg, sum=np.sum, ddof=ddof),
axis=axis,
keepdims=keepdims,
dtype=dt,
split_every=split_every,
combine=partial(moment_combine, sum=np.nansum),
out=out,
concatenate=False,
)
def _sqrt(a):
if isinstance(a, np.ma.masked_array) and not a.shape and a.mask.all():
return np.ma.masked
return np.sqrt(a)
def safe_sqrt(a):
"""A version of sqrt that properly handles scalar masked arrays.
To mimic ``np.ma`` reductions, we need to convert scalar masked arrays that
have an active mask to the ``np.ma.masked`` singleton. This is properly
handled automatically for reduction code, but not for ufuncs. We implement
a simple version here, since calling `np.ma.sqrt` everywhere is
significantly more expensive.
"""
if hasattr(a, "_elemwise"):
return a._elemwise(_sqrt, a)
return _sqrt(a)
@derived_from(np)
def std(a, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None):
result = safe_sqrt(
var(
a,
axis=axis,
dtype=dtype,
keepdims=keepdims,
ddof=ddof,
split_every=split_every,
out=out,
)
)
if dtype and dtype != result.dtype:
result = result.astype(dtype)
return result
@derived_from(np)
def nanstd(
a, axis=None, dtype=None, keepdims=False, ddof=0, split_every=None, out=None
):
result = safe_sqrt(
nanvar(
a,
axis=axis,
dtype=dtype,
keepdims=keepdims,
ddof=ddof,
split_every=split_every,
out=out,
)
)
if dtype and dtype != result.dtype:
result = result.astype(dtype)
return result
def _arg_combine(data, axis, argfunc, keepdims=False):
"""Merge intermediate results from ``arg_*`` functions"""
if isinstance(data, dict):
# Array type doesn't support structured arrays (e.g., CuPy),
# therefore `data` is stored in a `dict`.
assert data["vals"].ndim == data["arg"].ndim
axis = (
None
if len(axis) == data["vals"].ndim or data["vals"].ndim == 1
else axis[0]
)
else:
axis = None if len(axis) == data.ndim or data.ndim == 1 else axis[0]
vals = data["vals"]
arg = data["arg"]
if axis is None:
local_args = argfunc(vals, axis=axis, keepdims=keepdims)
vals = vals.ravel()[local_args]
arg = arg.ravel()[local_args]
else:
local_args = argfunc(vals, axis=axis)
inds = list(np.ogrid[tuple(map(slice, local_args.shape))])
inds.insert(axis, local_args)
inds = tuple(inds)
vals = vals[inds]
arg = arg[inds]
if keepdims:
vals = np.expand_dims(vals, axis)
arg = np.expand_dims(arg, axis)
return arg, vals
def arg_chunk(func, argfunc, x, axis, offset_info):
arg_axis = None if len(axis) == x.ndim or x.ndim == 1 else axis[0]
vals = func(x, axis=arg_axis, keepdims=True)
arg = argfunc(x, axis=arg_axis, keepdims=True)
if x.ndim > 0:
if arg_axis is None:
offset, total_shape = offset_info
ind = np.unravel_index(arg.ravel()[0], x.shape)
total_ind = tuple(o + i for (o, i) in zip(offset, ind))
arg[:] = np.ravel_multi_index(total_ind, total_shape)
else:
arg += offset_info
if isinstance(vals, np.ma.masked_array):
if "min" in argfunc.__name__:
fill_value = np.ma.minimum_fill_value(vals)
else:
fill_value = np.ma.maximum_fill_value(vals)
vals = np.ma.filled(vals, fill_value)
try:
result = np.empty_like(
vals, shape=vals.shape, dtype=[("vals", vals.dtype), ("arg", arg.dtype)]
)
except TypeError:
# Array type doesn't support structured arrays (e.g., CuPy)
result = dict()
result["vals"] = vals
result["arg"] = arg
return result
def arg_combine(argfunc, data, axis=None, **kwargs):
arg, vals = _arg_combine(data, axis, argfunc, keepdims=True)
try:
result = np.empty_like(
vals, shape=vals.shape, dtype=[("vals", vals.dtype), ("arg", arg.dtype)]
)
except TypeError:
# Array type doesn't support structured arrays (e.g., CuPy).
result = dict()
result["vals"] = vals
result["arg"] = arg
return result
def arg_agg(argfunc, data, axis=None, keepdims=False, **kwargs):
return _arg_combine(data, axis, argfunc, keepdims=keepdims)[0]
def nanarg_agg(argfunc, data, axis=None, keepdims=False, **kwargs):
arg, vals = _arg_combine(data, axis, argfunc, keepdims=keepdims)
if np.any(np.isnan(vals)):
raise ValueError("All NaN slice encountered")
return arg
def arg_reduction(
x, chunk, combine, agg, axis=None, keepdims=False, split_every=None, out=None
):
"""Generic function for argreduction.
Parameters
----------
x : Array
chunk : callable
Partialed ``arg_chunk``.
combine : callable
Partialed ``arg_combine``.
agg : callable
Partialed ``arg_agg``.
axis : int, optional
split_every : int or dict, optional
"""
if axis is None:
axis = tuple(range(x.ndim))
ravel = True
elif isinstance(axis, Integral):
axis = validate_axis(axis, x.ndim)
axis = (axis,)
ravel = x.ndim == 1
else:
raise TypeError(f"axis must be either `None` or int, got '{axis}'")
for ax in axis:
chunks = x.chunks[ax]
if len(chunks) > 1 and np.isnan(chunks).any():
raise ValueError(
"Arg-reductions do not work with arrays that have "
"unknown chunksizes. At some point in your computation "
"this array lost chunking information.\n\n"
"A possible solution is with \n"
" x.compute_chunk_sizes()"
)
# Map chunk across all blocks
name = f"arg-reduce-{tokenize(axis, x, chunk, combine, split_every)}"
old = x.name
keys = list(product(*map(range, x.numblocks)))
offsets = list(product(*(accumulate(operator.add, bd[:-1], 0) for bd in x.chunks)))
if ravel:
offset_info = zip(offsets, repeat(x.shape))
else:
offset_info = pluck(axis[0], offsets)
chunks = tuple((1,) * len(c) if i in axis else c for (i, c) in enumerate(x.chunks))
dsk = {
(name,) + k: (chunk, (old,) + k, axis, off)
for (k, off) in zip(keys, offset_info)
}
dtype = np.argmin(asarray_safe([1], like=meta_from_array(x)))
meta = None
if is_arraylike(dtype):
# This case occurs on non-NumPy types (e.g., CuPy), where the returned
# value is an ndarray rather than a scalar.
meta = dtype
dtype = meta.dtype
graph = HighLevelGraph.from_collections(name, dsk, dependencies=[x])
tmp = Array(graph, name, chunks, dtype=dtype, meta=meta)
result = _tree_reduce(
tmp,
agg,
axis,
keepdims=keepdims,
dtype=dtype,
split_every=split_every,
combine=combine,
)
return handle_out(out, result)
def _nanargmin(x, axis, **kwargs):
try:
return chunk.nanargmin(x, axis, **kwargs)
except ValueError:
return chunk.nanargmin(np.where(np.isnan(x), np.inf, x), axis, **kwargs)
def _nanargmax(x, axis, **kwargs):
try:
return chunk.nanargmax(x, axis, **kwargs)
except ValueError:
return chunk.nanargmax(np.where(np.isnan(x), -np.inf, x), axis, **kwargs)
@derived_from(np)
def argmax(a, axis=None, keepdims=False, split_every=None, out=None):
return arg_reduction(
a,
partial(arg_chunk, chunk.max, chunk.argmax),
partial(arg_combine, chunk.argmax),
partial(arg_agg, chunk.argmax),
axis=axis,
keepdims=keepdims,
split_every=split_every,
out=out,
)
@derived_from(np)
def argmin(a, axis=None, keepdims=False, split_every=None, out=None):
return arg_reduction(
a,
partial(arg_chunk, chunk.min, chunk.argmin),
partial(arg_combine, chunk.argmin),
partial(arg_agg, chunk.argmin),
axis=axis,
keepdims=keepdims,
split_every=split_every,
out=out,
)
@derived_from(np)
def nanargmax(a, axis=None, keepdims=False, split_every=None, out=None):
return arg_reduction(
a,
partial(arg_chunk, chunk.nanmax, _nanargmax),
partial(arg_combine, _nanargmax),
partial(nanarg_agg, _nanargmax),
axis=axis,
keepdims=keepdims,
split_every=split_every,
out=out,
)
@derived_from(np)
def nanargmin(a, axis=None, keepdims=False, split_every=None, out=None):
return arg_reduction(
a,
partial(arg_chunk, chunk.nanmin, _nanargmin),
partial(arg_combine, _nanargmin),
partial(nanarg_agg, _nanargmin),
axis=axis,
keepdims=keepdims,
split_every=split_every,
out=out,
)
def _prefixscan_combine(func, binop, pre, x, axis, dtype):
"""Combine results of a parallel prefix scan such as cumsum
Parameters
----------
func : callable
Cumulative function (e.g. ``np.cumsum``)
binop : callable
Associative function (e.g. ``add``)
pre : np.array
The value calculated in parallel from ``preop``.
For example, the sum of all the previous blocks.