-
Notifications
You must be signed in to change notification settings - Fork 89
/
_sfa.py
1085 lines (900 loc) · 36.1 KB
/
_sfa.py
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
# -*- coding: utf-8 -*-
"""Symbolic Fourier Approximation (SFA) Transformer.
Configurable SFA transform for discretising time series into words.
"""
__author__ = ["MatthewMiddlehurst", "patrickzib"]
__all__ = ["SFA"]
import math
import os
import sys
import warnings
import numpy as np
from joblib import Parallel, delayed
from numba import NumbaTypeSafetyWarning, njit, types
from numba.typed import Dict
from sklearn.feature_selection import f_classif
from sklearn.preprocessing import KBinsDiscretizer
from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
from aeon.transformations.collection import BaseCollectionTransformer
from aeon.utils.validation.panel import check_X
# The binning methods to use: equi-depth, equi-width, information gain or kmeans
binning_methods = {
"equi-depth",
"equi-width",
"information-gain",
"information-gain-mae",
"kmeans",
}
class SFA(BaseCollectionTransformer):
"""Symbolic Fourier Approximation (SFA) Transformer.
Overview: for each series:
run a sliding window across the series
for each window
shorten the series with DFT
discretise the shortened series into bins set by MFC
form a word from these discrete values
by default SFA produces a single word per series (window_size=0)
if a window is used, it forms a histogram of counts of words.
Parameters
----------
word_length: int, default = 8
length of word to shorten window to (using PAA)
alphabet_size: int, default = 4
number of values to discretise each value to
window_size: int, default = 12
size of window for sliding. Input series
length for whole series transform
norm: boolean, default = False
mean normalise words by dropping first fourier coefficient
binning_method: {"equi-depth", "equi-width", "information-gain",
"information-gain-mae", "kmeans"}, default="equi-depth"
the binning method used to derive the breakpoints.
anova: boolean, default = False
If True, the Fourier coefficient selection is done via a one-way
ANOVA test. If False, the first Fourier coefficients are selected.
Only applicable if labels are given
bigrams: boolean, default = False
whether to create bigrams of SFA words
skip_grams: boolean, default = False
whether to create skip-grams of SFA words
remove_repeat_words: boolean, default = False
whether to use numerosity reduction (default False)
levels: int, default = 1
Number of spatial pyramid levels
save_words: boolean, default = False
whether to save the words generated for each series (default False)
n_jobs: int, optional, default = 1
The number of jobs to run in parallel for both `transform`.
``-1`` means using all processors.
Attributes
----------
words: []
breakpoints: = []
num_insts = 0
num_atts = 0
References
----------
.. [1] Schäfer, Patrick, and Mikael Högqvist. "SFA: a symbolic fourier approximation
and index for similarity search in high dimensional datasets." Proceedings of the
15th international conference on extending database technology. 2012.
"""
_tags = {
"univariate-only": True,
"scitype:instancewise": False,
"fit_is_empty": False,
"requires_y": True,
"y_inner_mtype": "numpy1D",
}
def __init__(
self,
word_length=8,
alphabet_size=4,
window_size=12,
norm=False,
binning_method="equi-depth",
anova=False,
bigrams=False,
skip_grams=False,
remove_repeat_words=False,
levels=1,
lower_bounding=True,
save_words=False,
keep_binning_dft=False,
use_fallback_dft=False,
typed_dict=False,
n_jobs=1,
random_state=None,
):
self.words = []
self.breakpoints = []
# we cannot select more than window_size many letters in a word
offset = 2 if norm else 0
self.dft_length = window_size - offset if anova is True else word_length
# make dft_length an even number (same number of reals and imags)
self.dft_length = self.dft_length + self.dft_length % 2
self.support = np.array(list(range(word_length)))
self.word_length = word_length
self.alphabet_size = alphabet_size
self.window_size = window_size
self.norm = norm
self.lower_bounding = lower_bounding
self.inverse_sqrt_win_size = (
1.0 / math.sqrt(window_size) if not lower_bounding else 1.0
)
self.remove_repeat_words = remove_repeat_words
self.save_words = save_words
self.keep_binning_dft = keep_binning_dft
self.binning_dft = None
self.levels = levels
self.binning_method = binning_method
self.anova = anova
self.bigrams = bigrams
self.skip_grams = skip_grams
self.use_fallback_dft = use_fallback_dft
self._use_fallback_dft = (
use_fallback_dft if word_length < window_size - offset else True
)
self.typed_dict = typed_dict
# we will disable typed_dict if numba is disabled
self._typed_dict = typed_dict and not os.environ.get("NUMBA_DISABLE_JIT") == "1"
self.n_jobs = n_jobs
self.random_state = random_state
self.n_instances = 0
self.series_length = 0
self.letter_bits = 0
self.letter_max = 0
self.word_bits = 0
self.max_bits = 0
self.level_bits = 0
self.level_max = 0
super(SFA, self).__init__()
def _fit(self, X, y=None):
"""Calculate word breakpoints using MCB or IGB.
Parameters
----------
X : 3d numpy array, input time series.
y : array_like, target values (optional, ignored).
Returns
-------
self: object
"""
if self.alphabet_size < 2:
raise ValueError("Alphabet size must be an integer greater than 2")
if self.word_length < 1:
raise ValueError("Word length must be an integer greater than 1")
if (
self.binning_method == "information-gain"
or self.binning_method == "information-gain-mae"
) and y is None:
raise ValueError(
"Class values must be provided for information gain binning"
)
if self.binning_method not in binning_methods:
raise TypeError("binning_method must be one of: ", binning_methods)
if self._typed_dict != self.typed_dict:
warnings.warn(
"Typed dict is not supported when numba is disabled.", stacklevel=2
)
self.letter_bits = math.ceil(math.log2(self.alphabet_size))
self.letter_max = pow(2, self.letter_bits) - 1
self.word_bits = self.word_length * self.letter_bits
self.max_bits = (
self.word_bits * 2 if self.bigrams or self.skip_grams else self.word_bits
)
if self._typed_dict and self.max_bits > 64:
raise ValueError(
"Typed Dictionaries can only handle 64 bit words. "
"ceil(log2(alphabet_size)) * word_length must be less than or equal "
"to 64."
"With bi-grams or skip-grams enabled, this bit limit is 32."
)
if self._typed_dict and self.levels > 15:
raise ValueError(
"Typed Dictionaries can only handle 15 levels "
"(this is way to many anyway)."
)
X = check_X(X, enforce_univariate=True, coerce_to_numpy=True)
X = X.squeeze(1)
if self.levels > 1:
quadrants = 0
for i in range(self.levels):
quadrants += pow(2, i)
self.level_bits = math.ceil(math.log2(quadrants))
self.level_max = pow(2, self.level_bits) - 1
self.n_instances, self.series_length = X.shape
self.breakpoints = self._binning(X, y)
self._is_fitted = True
return self
def _transform(self, X, y=None):
"""Transform data into SFA words.
Parameters
----------
X : 3d numpy array, input time series.
y : array_like, target values (optional, ignored).
Returns
-------
List of dictionaries containing SFA words
"""
X = X.squeeze(1)
with warnings.catch_warnings():
warnings.simplefilter("ignore", category=NumbaTypeSafetyWarning)
transform = Parallel(n_jobs=self.n_jobs, prefer="threads")(
delayed(self._transform_case)(
X[i, :],
supplied_dft=self.binning_dft[i] if self.keep_binning_dft else None,
)
for i in range(X.shape[0])
)
dim, words = zip(*transform)
if self.save_words:
self.words = list(words)
# cant pickle typed dict
if self._typed_dict and self.n_jobs != 1:
nl = [None] * len(dim)
for i, pdict in enumerate(dim):
ndict = (
Dict.empty(
key_type=types.UniTuple(types.int64, 2), value_type=types.uint32
)
if self.levels > 1
else Dict.empty(key_type=types.int64, value_type=types.uint32)
)
for key, val in pdict.items():
ndict[key] = val
nl[i] = pdict
dim = nl
bags = [None]
bags[0] = list(dim)
return bags
def _transform_case(self, X, supplied_dft=None):
if supplied_dft is None:
dfts = self._mft(X)
else:
dfts = supplied_dft
if self._typed_dict:
bag = (
Dict.empty(
key_type=types.UniTuple(types.int64, 2), value_type=types.uint32
)
if self.levels > 1
else Dict.empty(key_type=types.int64, value_type=types.uint32)
)
else:
bag = {}
last_word = -1
repeat_words = 0
words = (
np.zeros(dfts.shape[0], dtype=np.int64)
if self.word_bits <= 64
else [0 for _ in range(dfts.shape[0])]
)
for window in range(dfts.shape[0]):
word_raw = (
SFA._create_word(
dfts[window],
self.word_length,
self.alphabet_size,
self.breakpoints,
self.letter_bits,
)
if self.word_bits <= 64
else self._create_word_large(
dfts[window],
)
)
words[window] = word_raw
repeat_word = (
self._add_to_pyramid(
bag, word_raw, last_word, window - int(repeat_words / 2)
)
if self.levels > 1
else self._add_to_bag(bag, word_raw, last_word)
)
if repeat_word:
repeat_words += 1
else:
last_word = word_raw
repeat_words = 0
if self.bigrams:
if window - self.window_size >= 0:
bigram = self._create_bigram_words(
word_raw, words[window - self.window_size]
)
if self.levels > 1:
if self._typed_dict:
bigram = (bigram, -1)
else:
bigram = (bigram << self.level_bits) | 0
bag[bigram] = bag.get(bigram, 0) + 1
if self.skip_grams:
# creates skip-grams, skipping every (s-1)-th word in-between
for s in range(2, 4):
if window - s * self.window_size >= 0:
skip_gram = self._create_bigram_words(
word_raw, words[window - s * self.window_size]
)
if self.levels > 1:
if self._typed_dict:
skip_gram = (skip_gram, -1)
else:
skip_gram = (skip_gram << self.level_bits) | 0
bag[skip_gram] = bag.get(skip_gram, 0) + 1
# cant pickle typed dict
if self._typed_dict and self.n_jobs != 1:
pdict = dict()
for key, val in bag.items():
pdict[key] = val
bag = pdict
return [
bag,
words if self.save_words else [],
]
def _binning(self, X, y=None):
num_windows_per_inst = math.ceil(self.series_length / self.window_size)
dft = np.array(
[
self._binning_dft(X[i, :], num_windows_per_inst)
for i in range(self.n_instances)
]
)
if self.keep_binning_dft:
self.binning_dft = dft
dft = dft.reshape(len(X) * num_windows_per_inst, self.dft_length)
if y is not None:
y = np.repeat(y, num_windows_per_inst)
if self.anova and y is not None:
non_constant = np.where(
~np.isclose(dft.var(axis=0), np.zeros_like(dft.shape[1]))
)[0]
# select word-length many indices with best f-score
if self.word_length <= non_constant.size:
f, _ = f_classif(dft[:, non_constant], y)
self.support = non_constant[np.argsort(-f)][: self.word_length]
# sort remaining indices
# self.support = np.sort(self.support)
# select the Fourier coefficients with highest f-score
dft = dft[:, self.support]
self.dft_length = np.max(self.support) + 1
self.dft_length = self.dft_length + self.dft_length % 2 # even
if self.binning_method == "information-gain":
return self._igb(dft, y)
if self.binning_method == "information-gain-mae":
return self._igb_mae(dft, y)
elif self.binning_method == "kmeans":
return self._k_bins_discretizer(dft)
else:
return self._mcb(dft)
def _k_bins_discretizer(self, dft):
encoder = KBinsDiscretizer(
n_bins=self.alphabet_size, strategy=self.binning_method
)
encoder.fit(dft)
breaks = encoder.bin_edges_
breakpoints = np.zeros((self.word_length, self.alphabet_size))
for letter in range(self.word_length):
for bp in range(1, len(breaks[letter]) - 1):
breakpoints[letter][bp - 1] = breaks[letter][bp]
breakpoints[:, self.alphabet_size - 1] = sys.float_info.max
return breakpoints
def _mcb(self, dft):
num_windows_per_inst = math.ceil(self.series_length / self.window_size)
total_num_windows = int(self.n_instances * num_windows_per_inst)
breakpoints = np.zeros((self.word_length, self.alphabet_size))
for letter in range(self.word_length):
res = [round(dft[i][letter] * 100) / 100 for i in range(total_num_windows)]
column = np.sort(np.array(res))
bin_index = 0
# use equi-depth binning
if self.binning_method == "equi-depth":
target_bin_depth = total_num_windows / self.alphabet_size
for bp in range(self.alphabet_size - 1):
bin_index += target_bin_depth
breakpoints[letter][bp] = column[int(bin_index)]
# use equi-width binning aka equi-frequency binning
elif self.binning_method == "equi-width":
target_bin_width = (column[-1] - column[0]) / self.alphabet_size
for bp in range(self.alphabet_size - 1):
breakpoints[letter][bp] = (bp + 1) * target_bin_width + column[0]
breakpoints[:, self.alphabet_size - 1] = sys.float_info.max
return breakpoints
def _igb(self, dft, y):
breakpoints = np.zeros((self.word_length, self.alphabet_size))
clf = DecisionTreeClassifier(
criterion="entropy",
max_depth=int(np.floor(np.log2(self.alphabet_size))),
max_leaf_nodes=self.alphabet_size,
random_state=self.random_state,
)
for i in range(self.word_length):
clf.fit(dft[:, i][:, None], y)
threshold = clf.tree_.threshold[clf.tree_.children_left != -1]
for bp in range(len(threshold)):
breakpoints[i][bp] = threshold[bp]
for bp in range(len(threshold), self.alphabet_size):
breakpoints[i][bp] = sys.float_info.max
return np.sort(breakpoints, axis=1)
def _igb_mae(self, dft, y):
breakpoints = np.zeros((self.word_length, self.alphabet_size))
clf = DecisionTreeRegressor(
criterion="friedman_mse",
max_depth=int(np.floor(np.log2(self.alphabet_size))),
max_leaf_nodes=self.alphabet_size,
random_state=self.random_state,
)
for i in range(self.word_length):
clf.fit(dft[:, i][:, None], y)
threshold = clf.tree_.threshold[clf.tree_.children_left != -1]
for bp in range(len(threshold)):
breakpoints[i][bp] = threshold[bp]
for bp in range(len(threshold), self.alphabet_size):
breakpoints[i][bp] = sys.float_info.max
return np.sort(breakpoints, axis=1)
def _binning_dft(self, series, num_windows_per_inst):
# Splits individual time series into windows and returns the DFT for
# each
split = np.split(
series,
np.linspace(
self.window_size,
self.window_size * (num_windows_per_inst - 1),
num_windows_per_inst - 1,
dtype=np.int_,
),
)
start = self.series_length - self.window_size
split[-1] = series[start : self.series_length]
result = np.zeros((len(split), self.dft_length), dtype=np.float64)
for i, row in enumerate(split):
result[i] = (
self._discrete_fourier_transform(
row,
self.dft_length,
self.norm,
self.inverse_sqrt_win_size,
self.lower_bounding,
)
if self._use_fallback_dft
else self._fast_fourier_transform(row)
)
return result
def _fast_fourier_transform(self, series):
"""Perform a discrete fourier transform using the fast fourier transform.
if self.norm is True, then the first term of the DFT is ignored
Input
-------
X : The training input samples. array-like or sparse matrix of
shape = [n_samps, num_atts]
Returns
-------
1D array of fourier term, real_0,imag_0, real_1, imag_1 etc, length
num_atts or
num_atts-2 if if self.norm is True
"""
# first two are real and imaginary parts
start = 2 if self.norm else 0
s = np.std(series)
std = s if s > 1e-8 else 1
X_fft = np.fft.rfft(series)
reals = np.real(X_fft)
imags = np.imag(X_fft)
length = start + self.dft_length
dft = np.empty((length,), dtype=reals.dtype)
dft[0::2] = reals[: np.uint32(length / 2)]
dft[1::2] = imags[: np.uint32(length / 2)]
if self.lower_bounding:
dft[1::2] = dft[1::2] * -1 # lower bounding
dft *= self.inverse_sqrt_win_size / std
return dft[start:]
@staticmethod
@njit(fastmath=True, cache=True)
def _discrete_fourier_transform(
series,
dft_length,
norm,
inverse_sqrt_win_size,
lower_bounding,
apply_normalising_factor=True,
cut_start_if_norm=True,
):
"""Perform a discrete fourier transform using standard O(n^2) transform.
if self.norm is True, then the first term of the DFT is ignored
Input
-------
X : The training input samples. array-like or sparse matrix of
shape = [n_samps, num_atts]
Returns
-------
1D array of fourier term, real_0,imag_0, real_1, imag_1 etc, length
num_atts or
num_atts-2 if if self.norm is True
"""
start = 2 if norm else 0
output_length = start + dft_length
if cut_start_if_norm:
c = int(start / 2)
else:
c = 0
start = 0
dft = np.zeros(output_length - start)
for i in range(c, int(output_length / 2)):
for n in range(len(series)):
dft[(i - c) * 2] += series[n] * math.cos(
2 * math.pi * n * i / len(series)
)
dft[(i - c) * 2 + 1] += -series[n] * math.sin(
2 * math.pi * n * i / len(series)
)
if apply_normalising_factor:
if lower_bounding:
dft[1::2] = dft[1::2] * -1 # lower bounding
std = np.std(series)
if std == 0:
std = 1
dft *= inverse_sqrt_win_size / std
return dft
def _mft(self, series):
start_offset = 2 if self.norm else 0
length = self.dft_length + start_offset + self.dft_length % 2
end = max(1, len(series) - self.window_size + 1)
phis = SFA._get_phis(self.window_size, length)
stds = SFA._calc_incremental_mean_std(series, end, self.window_size)
transformed = np.zeros((end, length))
# first run with dft
if self._use_fallback_dft:
mft_data = self._discrete_fourier_transform(
series[0 : self.window_size],
self.dft_length,
self.norm,
self.inverse_sqrt_win_size,
self.lower_bounding,
apply_normalising_factor=False,
cut_start_if_norm=False,
)
else:
X_fft = np.fft.rfft(series[: self.window_size])
reals = np.real(X_fft)
imags = np.imag(X_fft)
mft_data = np.empty((length,), dtype=reals.dtype)
mft_data[0::2] = reals[: np.uint32(length / 2)]
mft_data[1::2] = imags[: np.uint32(length / 2)]
transformed[0] = mft_data * self.inverse_sqrt_win_size / stds[0]
# other runs using mft
# moved to external method to use njit
SFA._iterate_mft(
series,
mft_data,
phis,
self.window_size,
stds,
transformed,
self.inverse_sqrt_win_size,
)
if self.lower_bounding:
transformed[:, 1::2] = transformed[:, 1::2] * -1 # lower bounding
return (
transformed[:, start_offset:][:, self.support]
if self.anova
else transformed[:, start_offset:]
)
@staticmethod
@njit(fastmath=True, cache=True)
def _get_phis(window_size, length):
phis = np.zeros(length)
for i in range(int(length / 2)):
phis[i * 2] += math.cos(2 * math.pi * (-i) / window_size)
phis[i * 2 + 1] += -math.sin(2 * math.pi * (-i) / window_size)
return phis
@staticmethod
@njit(fastmath=True, cache=True)
def _iterate_mft(
series, mft_data, phis, window_size, stds, transformed, inverse_sqrt_win_size
):
for i in range(1, len(transformed)):
for n in range(0, len(mft_data), 2):
# only compute needed indices
real = mft_data[n] + series[i + window_size - 1] - series[i - 1]
imag = mft_data[n + 1]
mft_data[n] = real * phis[n] - imag * phis[n + 1]
mft_data[n + 1] = real * phis[n + 1] + phis[n] * imag
normalising_factor = inverse_sqrt_win_size / stds[i]
transformed[i] = mft_data * normalising_factor
def _shorten_bags(self, word_len):
if self.save_words is False:
raise ValueError(
"Words from transform must be saved using save_word to shorten bags."
)
if word_len > self.word_length:
word_len = self.word_length
if self._typed_dict:
warnings.simplefilter("ignore", category=NumbaTypeSafetyWarning)
dim = Parallel(n_jobs=self.n_jobs, prefer="threads")(
delayed(self._shorten_case)(word_len, i) for i in range(len(self.words))
)
# cant pickle typed dict
if self._typed_dict and self.n_jobs != 1:
nl = [None] * len(dim)
for i, pdict in enumerate(dim):
ndict = (
Dict.empty(
key_type=types.UniTuple(types.int64, 2), value_type=types.uint32
)
if self.levels > 1
else Dict.empty(key_type=types.int64, value_type=types.uint32)
)
for key, val in pdict.items():
ndict[key] = val
nl[i] = pdict
dim = nl
new_bags = [None]
new_bags[0] = list(dim)
return new_bags
def _shorten_case(self, word_len, i):
if self._typed_dict:
new_bag = (
Dict.empty(
key_type=types.Tuple((types.int64, types.int16)),
value_type=types.uint32,
)
if self.levels > 1
else Dict.empty(key_type=types.int64, value_type=types.uint32)
)
else:
new_bag = {}
last_word = -1
repeat_words = 0
for window, word in enumerate(self.words[i]):
new_word = self._shorten_words(word, word_len)
repeat_word = (
self._add_to_pyramid(
new_bag, new_word, last_word, window - int(repeat_words / 2)
)
if self.levels > 1
else self._add_to_bag(new_bag, new_word, last_word)
)
if repeat_word:
repeat_words += 1
else:
last_word = new_word
repeat_words = 0
if self.bigrams:
if window - self.window_size >= 0:
bigram = self._create_bigram_words(
new_word,
self._shorten_words(
self.words[i][window - self.window_size], word_len
),
)
if self.levels > 1:
if self._typed_dict:
bigram = (bigram, -1)
else:
bigram = (bigram << self.level_bits) | 0
new_bag[bigram] = new_bag.get(bigram, 0) + 1
if self.skip_grams:
# creates skip-grams, skipping every (s-1)-th word in-between
for s in range(2, 4):
if window - s * self.window_size >= 0:
skip_gram = self._create_bigram_words(
new_word,
self._shorten_words(
self.words[i][window - s * self.window_size],
word_len,
),
)
if self.levels > 1:
if self._typed_dict:
skip_gram = (skip_gram, -1)
else:
skip_gram = (skip_gram << self.level_bits) | 0
new_bag[skip_gram] = new_bag.get(skip_gram, 0) + 1
# cant pickle typed dict
if self._typed_dict and self.n_jobs != 1:
pdict = dict()
for key, val in new_bag.items():
pdict[key] = val
new_bag = pdict
return new_bag
def _add_to_bag(self, bag, word, last_word):
if self.remove_repeat_words and word == last_word:
return True
# store the histogram of word counts
bag[word] = bag.get(word, 0) + 1
return False
def _add_to_pyramid(self, bag, word, last_word, window_ind):
if self.remove_repeat_words and word == last_word:
return True
start = 0
for i in range(self.levels):
if self._typed_dict:
new_word, num_quadrants = SFA._add_level_typed(
word, start, i, window_ind, self.window_size, self.series_length
)
else:
new_word, num_quadrants = (
SFA._add_level(
word,
start,
i,
window_ind,
self.window_size,
self.series_length,
self.level_bits,
)
if self.word_bits + self.level_bits <= 64
else self._add_level_large(
word,
start,
i,
window_ind,
)
)
bag[new_word] = bag.get(new_word, 0) + num_quadrants
start += num_quadrants
return False
@staticmethod
@njit(fastmath=True, cache=True)
def _add_level(
word, start, level, window_ind, window_size, series_length, level_bits
):
num_quadrants = pow(2, level)
quadrant = start + int(
(window_ind + int(window_size / 2)) / int(series_length / num_quadrants)
)
return (word << level_bits) | quadrant, num_quadrants
def _add_level_large(self, word, start, level, window_ind):
num_quadrants = pow(2, level)
quadrant = start + int(
(window_ind + int(self.window_size / 2))
/ int(self.series_length / num_quadrants)
)
return (word << self.level_bits) | quadrant, num_quadrants
@staticmethod
@njit(fastmath=True, cache=True)
def _add_level_typed(word, start, level, window_ind, window_size, series_length):
num_quadrants = pow(2, level)
quadrant = start + int(
(window_ind + int(window_size / 2)) / int(series_length / num_quadrants)
)
return (word, quadrant), num_quadrants
@staticmethod
@njit(fastmath=True, cache=True)
def _create_word(dft, word_length, alphabet_size, breakpoints, letter_bits):
word = np.int64(0)
for i in range(word_length):
for bp in range(alphabet_size):
if dft[i] <= breakpoints[i][bp]:
word = (word << letter_bits) | bp
break
return word
def _create_word_large(self, dft):
word = 0
for i in range(self.word_length):
for bp in range(self.alphabet_size):
if dft[i] <= self.breakpoints[i][bp]:
word = (word << self.letter_bits) | bp
break
return word
@staticmethod
@njit(fastmath=True, cache=True)
def _calc_incremental_mean_std(series, end, window_size):
stds = np.zeros(end)
window = series[0:window_size]
series_sum = np.sum(window)
square_sum = np.sum(np.multiply(window, window))
r_window_length = 1 / window_size
mean = series_sum * r_window_length
buf = math.sqrt(square_sum * r_window_length - mean * mean)
stds[0] = buf if buf > 1e-8 else 1
for w in range(1, end):
series_sum += series[w + window_size - 1] - series[w - 1]
mean = series_sum * r_window_length
square_sum += (
series[w + window_size - 1] * series[w + window_size - 1]
- series[w - 1] * series[w - 1]
)
buf = math.sqrt(square_sum * r_window_length - mean * mean)
stds[w] = buf if buf > 1e-8 else 1
return stds
def _create_bigram_words(self, word_raw, word):
return (
SFA._create_bigram_word(
word,
word_raw,
self.word_bits,
)
if self.max_bits <= 64
else self._create_bigram_word_large(
word,
word_raw,
)
)
@staticmethod
@njit(fastmath=True, cache=True)
def _create_bigram_word(word, other_word, word_bits):
return (word << word_bits) | other_word
def _create_bigram_word_large(self, word, other_word):
return (word << self.word_bits) | other_word
def _shorten_words(self, word, word_len):
return (
SFA._shorten_word(word, self.word_length - word_len, self.letter_bits)
if self.word_bits <= 64
else self._shorten_word_large(word, self.word_length - word_len)
)
@staticmethod
@njit(fastmath=True, cache=True)
def _shorten_word(word, amount, letter_bits):
# shorten a word by set amount of letters
return word >> amount * letter_bits
def _shorten_word_large(self, word, amount):
# shorten a word by set amount of letters
return word >> amount * self.letter_bits
def bag_to_string(self, bag):
"""Convert a bag of SFA words into a string."""