-
Notifications
You must be signed in to change notification settings - Fork 18
/
steps.py
675 lines (565 loc) · 27.8 KB
/
steps.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
"""
FIDDLE Preprocessing steps
1. Pre-filter
2. Transform
3. Post-filter
"""
try:
from .helpers import *
except:
from helpers import *
import time
import json
import joblib
import multiprocessing
def pre_filter(df, threshold, df_population, args):
T = int(args.T)
theta_1 = args.theta_1
df_population = args.df_population
# Remove rows not in population
print('Remove rows not in population')
df = df[df['ID'].isin(df_population.index)]
# Remove rows with t outside of [0, T)
print('Remove rows with t outside of [0, {}]'.format(T))
df = df[pd.isnull(df[t_col]) | ((0 <= df[t_col]) & (df[t_col] < T))]
# Data table should not contain duplicate rows with any numerical values
# Check for inconsistencies
var_names = [v for v, ty in args.value_type_override.items() if 'hierarchical' in ty.lower() or 'categorical' in ty.lower()]
df_tmp = df[~df[var_col].isin(var_names)]
dups = df_tmp.duplicated(subset=[ID_col, t_col, var_col], keep=False)
df_dups = df_tmp[dups]
if any(dups) and any(is_numeric(v) for v in df_dups[val_col] if not pd.isnull(v)):
print(df_dups.head())
raise Exception('Inconsistent numerical values recorded')
# Remove variables that occur too rarely as defined by the threshold
print('Remove rare variables (<= {})'.format(threshold))
## Calculate overall occurrence rate of each variable based on IDs
df_count = calculate_variable_counts(df, df_population) # (N x |var|) table of counts
df_bool = df_count.astype(bool) # convert counts to boolean
## Keep variables that are recorded for more than threshold fraction of IDs
variables_keep = df_bool.columns[df_bool.mean(axis=0) > threshold]
df_out = df[df[var_col].isin(variables_keep)]
assert set(variables_keep) == set(df_out[var_col].unique())
variables = sorted(df_bool.columns)
variables_remove = sorted(set(variables) - set(variables_keep))
print('Total variables :', len(variables))
print('Rare variables :', len(variables_remove))
print('Remaining variables :', len(variables_keep))
print('# rows (original) :', len(df))
print('# rows (filtered) :', len(df_out))
return df_out
def parse_variable_data_type(df_data, args):
# 1. parse hierarchical values (e.g. ICD codes) into strings
# 2. automatically detect value types, respecting user override, and set dtypes in DataFrames
# 3. pre-map duplicated non-numerical values into multiple categorical variables
output_dir = args.output_dir
df = df_data
assert val_col in df.columns
print_header('*) Detecting and parsing value types', char='-')
## 1. Hierarchical values
var_names = [v for v, ty in args.value_type_override.items() if 'hierarchical' in ty.lower()]
if len(var_names) == 0: # No hierarchical values
pass
else:
print('Parsing hierarchical values')
for var_name in var_names:
var_type = args.value_type_override[var_name]
df_var = df.loc[df[var_col] == var_name, val_col]
if var_type.lower() == 'hierarchical_icd':
# need to figure out ICD version
raise NotImplementedError
elif var_type.lower() == 'hierarchical_icd9':
df_var = df_var.apply(lambda s: map_icd_hierarchy(s, version=9))
elif var_type.lower() == 'hierarchical_icd10':
df_var = df_var.apply(lambda s: map_icd_hierarchy(s, version=10))
else:
df_var = df_var.apply(lambda s: s.split(hierarchical_sep))
# Assign mapped values back to original df
df.loc[df[var_col] == var_name, val_col] = df_var
# Only encode selected levels
df_nonhier = df[~df[var_col].isin(var_names)]
df_hier = df[df[var_col].isin(var_names)]
df_hier_levels = []
for hier_level in args.hierarchical_levels:
# encode level if available
df_hier_level = df_hier.copy()
df_hier_level[val_col] = df_hier_level[val_col].apply(lambda h: h[min(hier_level, len(h))])
df_hier_levels.append(df_hier_level)
df_hier_levels = pd.concat(df_hier_levels).drop_duplicates()
# Combine hierarchical and non-hierarchical data
df = pd.concat([df_nonhier, df_hier_levels])
## 2. Detect value types
data_types = []
# Collect the unique values of each variable
# values_by_variable: dict(variable_name -> [value1, value2, ...])
d = df[[var_col, val_col]].drop_duplicates().sort_values(by=[var_col, val_col])
values_by_variable = defaultdict(list)
for n,v in zip(d[var_col], d[val_col]):
values_by_variable[n].append(v)
# Determine type of each variable
for variable, values in sorted(values_by_variable.items()):
# Manual override type in config
if variable in args.value_type_override:
data_types.append((variable, args.value_type_override[variable]))
# Force categorical values to be a string
if args.value_type_override[variable] == 'Categorical' and \
any(is_numeric(v) for v in values if not pd.isnull(v)):
m_var = df[var_col] == variable
df.loc[m_var, val_col] = df.loc[m_var, val_col].apply(lambda s: '_' + str(s))
else:
if len(values) == 1 and pd.isnull(values[0]):
data_types.append((variable, 'None'))
elif all(is_numeric(v) for v in values if not pd.isnull(v)):
data_types.append((variable, 'Numeric'))
elif any(is_numeric(v) for v in values if not pd.isnull(v)):
data_types.append((variable, 'Numeric + Categorical'))
else:
data_types.append((variable, 'Categorical'))
df_types = pd.DataFrame(data_types, columns=['variable_name', 'value_type'])
df_types[var_col] = df_types[var_col].astype(str)
df_types = df_types.set_index(var_col)
fpath = output_dir + 'value_types.csv'
df_types.to_csv(fpath, quoting=1)
print('Saved as:', fpath)
## 3. Pre-map duplicated non-numerical values to separate variables
var_names = [v for v, ty in data_types if 'numeric' not in ty.lower() and 'none' not in ty.lower()]
df_non_num = df[df[var_col].isin(var_names)].copy()
dup_ = df_non_num.duplicated(subset=[ID_col, t_col, var_col], keep=False)
df_non_num_dup = df_non_num[dup_].copy()
dup_var_names = df_non_num_dup[var_col].unique()
df_non_num_dup[var_col] = df_non_num_dup[var_col].astype(str) + ':' + df_non_num_dup[val_col].astype(str)
df_non_num_dup[val_col] = 1
df_non_num.loc[dup_, :] = df_non_num_dup
df.loc[df[var_col].isin(var_names), :] = df_non_num
return df, df_types['value_type']
def split_by_timestamp_type(df):
print_header('*) Separate time-invariant and time-dependent', char='-')
variables_inv = df[pd.isnull(df[t_col])][var_col].unique() # Invariant variables have t = NULL
df_time_invariant = df[df[var_col].isin(variables_inv)]
df_time_series = df[~df[var_col].isin(variables_inv)]
print('Variables (time-invariant):', len(variables_inv))
print('Variables (time-dependent):', df[var_col].nunique() - len(variables_inv))
print('# rows (time-invariant):', len(df_time_invariant))
print('# rows (time-dependent):', len(df_time_series))
return df_time_invariant, df_time_series
def process_time_invariant(df_data_time_invariant, args):
if len(df_data_time_invariant) == 0:
return None, None, None
output_dir = args.output_dir
df_population = args.df_population
theta_2 = args.theta_2
##############
print_header('2-A) Transform time-invariant data', char='-')
dir_path = output_dir + '/'
start_time = time.time()
## Create Nxd^ table
df_time_invariant = transform_time_invariant_table(df_data_time_invariant, df_population)
df_time_invariant[[]].to_csv(dir_path + 'S.ID.csv')
print('Time elapsed: %f seconds' % (time.time() - start_time))
## Discretize
S_all, S_all_feature_names, S_discretization_bins = map_time_invariant_features(df_time_invariant, args)
sparse.save_npz(dir_path + 'S_all.npz', S_all)
json.dump(list(S_all_feature_names), open(dir_path + 'S_all.feature_names.json', 'w'), sort_keys=True)
json.dump(S_discretization_bins, open(dir_path + 'S_all.discretization.json', 'w'))
print('Time elapsed: %f seconds' % (time.time() - start_time))
if args.postfilter:
##############
print_header('3-A) Post-filter time-invariant data', char='-')
## Filter
S, S_feature_names, S_feature_aliases = post_filter_time_invariant(S_all, S_all_feature_names, theta_2)
print('Time elapsed: %f seconds' % (time.time() - start_time))
## Save output
print()
print('Output')
print('S: shape={}, density={:.3f}'.format(S.shape, S.density))
sparse.save_npz(dir_path + 'S.npz', S)
with open(dir_path + 'S.feature_names.json', 'w') as f:
json.dump(list(S_feature_names), f, sort_keys=True)
with open(dir_path + 'S.feature_aliases.json', 'w') as f:
json.dump(S_feature_aliases, f, sort_keys=True)
print('Total time: %f seconds' % (time.time() - start_time))
print('', flush=True)
return S, S_feature_names, S_feature_aliases
else:
return S_all, S_all_feature_names, None
def process_time_dependent(df_data_time_series, args):
if len(df_data_time_series) == 0:
return None, None, None
output_dir = args.output_dir
theta_2 = args.theta_2
##############
print_header('2-B) Transform time-dependent data', char='-')
dir_path = output_dir + '/'
start_time = time.time()
## Create NxLxD^ table
df_time_series, dtypes_time_series = transform_time_series_table(df_data_time_series, args)
print('Time elapsed: %f seconds' % (time.time() - start_time))
## Save intermediate files
joblib.dump(df_time_series, output_dir + 'df_time_series.joblib')
joblib.dump(dtypes_time_series, output_dir + 'dtypes_time_series.joblib')
df_time_series[[]].to_csv(dir_path + 'X.ID,t_range.csv')
## Map variables to features
X_all, X_all_feature_names, X_discretization_bins = map_time_series_features(df_time_series, dtypes_time_series, args)
sparse.save_npz(dir_path + 'X_all.npz', X_all)
json.dump(list(X_all_feature_names), open(dir_path + 'X_all.feature_names.json', 'w'), sort_keys=True)
json.dump(X_discretization_bins, open(dir_path + 'X_all.discretization.json', 'w'))
print('Time elapsed: %f seconds' % (time.time() - start_time))
if args.postfilter:
##############
print_header('3-B) Post-filter time-dependent data', char='-')
print(X_all.shape, X_all.density)
## Filter features
X, X_feature_names, X_feature_aliases = post_filter_time_series(X_all, X_all_feature_names, theta_2, args)
print(X.shape, X.density)
print('Time elapsed: %f seconds' % (time.time() - start_time))
## Save output
print()
print('Output')
print('X: shape={}, density={:.3f}'.format(X.shape, X.density))
sparse.save_npz(dir_path + 'X.npz', X)
with open(dir_path + 'X.feature_names.json', 'w') as f:
json.dump(list(X_feature_names), f, sort_keys=True)
with open(dir_path + 'X.feature_aliases.json', 'w') as f:
json.dump(X_feature_aliases, f, sort_keys=True)
print('Total time: %f seconds' % (time.time() - start_time))
print('', flush=True)
return X, X_feature_names, X_feature_aliases
else:
return X_all, X_all_feature_names, None
######
# Time-invariant routines
######
def transform_time_invariant_table(df_in, df_population):
df_in = df_in.copy()
# Recorded Value (np.nan if not recorded)
df_value = pd.pivot_table(df_in, val_col, ID_col, var_col, 'last', np.nan)
df_value = df_value.reindex(index=df_population.index, fill_value=np.nan)
df_value.columns = [str(col) + '_value' for col in df_value.columns]
print('(N \u00D7 ^d) table :\t', df_value.shape)
print('number of missing entries :\t', '{} out of {} total'.format(df_value.isna().sum().sum(), df_value.size))
return df_value
def map_time_invariant_features(df, args):
# Categorical -> binary features
# Numeric -> binary/float-valued features
discretization_bins = None
if args.discretize:
discretization_bins = args.S_discretization_bins
if discretization_bins is None:
discretization_bins = [compute_bin_edges(df[col], q=5) for col in df.columns]
discretization_bins = dict(discretization_bins)
out = [smart_qcut_dummify(df[col], discretization_bins[col], use_ordinal_encoding=args.use_ordinal_encoding) for col in df.columns]
time_invariant_features = pd.concat(out, axis=1)
feature_names_all = time_invariant_features.columns.values
sdf = time_invariant_features.astype(pd.SparseDtype(int, fill_value=0))
S_ = sparse.COO(sdf.sparse.to_coo())
else:
# Split a mixed column into numeric and string columns
for col in df.columns:
col_data = df[col]
col_is_numeric = [is_numeric(v) for v in col_data if not pd.isnull(v)]
if not all(col_is_numeric) and any(col_is_numeric): # have mixed type values
numeric_mask = col_data.apply(is_numeric)
df[col+'_str'] = df[col].copy()
df.loc[~numeric_mask, col] = np.nan
df.loc[numeric_mask, col+'_str'] = np.nan
out = [smart_dummify_impute(df[col]) for col in df.columns]
time_invariant_features = pd.concat(out, axis=1)
feature_names_all = time_invariant_features.columns.values
sdf = time_invariant_features.astype(pd.SparseDtype(float, fill_value=0))
S_ = sparse.COO(sdf.sparse.to_coo())
print()
print('Output')
print('S_all, binary features :\t', S_.shape)
return S_, feature_names_all, discretization_bins
def post_filter_time_invariant(S_, S_feature_names_all, threshold):
# Filter features (optional)
assert S_.shape[1] == len(S_feature_names_all)
feature_names_0 = S_feature_names_all
S0 = S_.to_scipy_sparse()
print('Original :', len(feature_names_0))
## Remove nearly-constant features (with low variance)
## a binary feature is removed if =0 (or =1) for >th fraction of examples
## i.e., variance <= (th * (1 - th))
sel_rare = VarianceThreshold(threshold=(threshold * (1 - threshold)))
S1 = sel_rare.fit_transform(S0)
feature_names_1 = feature_names_0[sel_rare.get_support()]
print('Nearly-constant:', len(feature_names_0) - len(feature_names_1))
## Keep only first of pairwise perfectly correlated features
sel_corr = CorrelationSelector()
S2 = sel_corr.fit_transform(S1)
feature_names_2 = feature_names_1[sel_corr.get_support()]
feature_aliases = sel_corr.get_feature_aliases(feature_names_1)
print('Correlated :', len(feature_names_1) - len(feature_names_2))
S = sparse.COO(S2)
feature_names = feature_names_2
assert S.shape[1] == len(feature_names)
return S, feature_names, feature_aliases
######
# Time-series routines
######
def func_encode_single_time_series(i, g, variables, variables_num_freq, T, dt, stats_functions, impute=True):
try:
assert g.index.nunique() == 1
assert g.index.unique()[0] == i
# non-frequent
variables_non = sorted(set(variables) - set(variables_num_freq))
if len(variables_non) > 0:
variables_non = sorted(set(variables) - set(variables_num_freq))
df_j = pivot_event_table(g).reindex(columns=variables_non).sort_index()
df_values_j = most_recent_values(df_j, variables, T, dt)
df_out = df_values_j
if len(variables_num_freq) > 0:
# frequent
# we're only producing mask, ffill, and statistics if the data is measured frequently enough
df_i = pivot_event_table(g).reindex(columns=variables_num_freq).sort_index()
mask_i = presence_mask(df_i, variables_num_freq, T, dt)
delta_t_i = get_delta_time(mask_i)
df_i = impute_ffill(df_i, variables_num_freq, T, dt, mask_i)
df_stats_i = summary_statistics(df_i, variables_num_freq, stats_functions, T, dt)
df_values_i = most_recent_values(df_i, variables, T, dt)
if impute:
check_imputed_output(df_values_i)
check_imputed_output(df_stats_i)
df_out = df_out.join([mask_i, delta_t_i, df_values_i, df_stats_i])
except:
print(i)
raise Exception(i)
return i, df_out
def divide_chunks(l, n):
# looping till length l
for i in range(0, len(l), n):
yield l[i:i + n]
def form_batches_of_examples(df_in, args, batch_size):
grouped = df_in.set_index(ID_col)
IDs = list(grouped.index.unique())
batches_IDs = list(divide_chunks(IDs, batch_size))
batches = [grouped.loc[chunk] for chunk in batches_IDs]
return batches, batches_IDs
def process_batch_time_series(first_arg):
batch, batch_IDs, args = first_arg
variables, variables_num_freq = args.variables, args.variables_num_freq
out = dict(
func_encode_single_time_series(i, batch.loc[i:i], variables, variables_num_freq, args.T, args.dt, args.stats_functions)
for i in batch_IDs
)
return out
def transform_time_series_table(df_in, args):
output_dir = args.output_dir
theta_freq = args.theta_freq
stats_functions = args.stats_functions
N, L = args.N, args.L
df_population = args.df_population
parallel = args.parallel
## TODO: asserts shape of df_in
# Determine all unique variable names
variables = get_unique_variables(df_in)
assert df_in[var_col].nunique() == len(variables)
print('Total variables :', len(variables))
# Determine frequent variables -> we'll calculate statistics, mask, and delta_time only on these
variables_num_freq = get_frequent_numeric_variables(df_in, variables, theta_freq, args)
print('Frequent variables :', list(variables_num_freq))
print('{} = {}'.format('M\u2081', len(variables_num_freq)))
print('{} = {}'.format('M\u2082', len(variables) - len(variables_num_freq)))
print('{} = {} {}'.format('k ', len(stats_functions), stats_functions))
print()
print('Transforming each example...')
args.variables = variables
args.variables_num_freq = variables_num_freq
# Encode time series table for each patient
if args.parallel:
batches, batches_IDs = form_batches_of_examples(df_in, args, batch_size=args.batch_size)
print('Batches of size {}: '.format(args.batch_size), len(batches))
pool = multiprocessing.Pool(args.n_jobs)
out = list(tqdm(pool.imap_unordered(
process_batch_time_series,
zip(batches, batches_IDs, [args]*len(batches))), total=len(batches)
))
pool.close()
pool.join()
out = dict((key, d[key]) for d in out for key in d)
print()
print('Parallel processing done', flush=True)
else:
grouped = list(df_in.groupby(ID_col))
out = dict(
func_encode_single_time_series(i, g.set_index(ID_col), variables, variables_num_freq, args.T, args.dt, args.stats_functions)
for i, g in tqdm(grouped[:N])
)
# Handle IDs not in the table
df_original = list(out.values())[0]
df_copy = pd.DataFrame().reindex_like(df_original)
for i, j in df_original.dtypes.iteritems():
if i.endswith('_mask'):
assert j == bool
df_copy[i] = False
df_copy[i] = df_copy[i].astype(bool)
if i.endswith('_delta_time'):
df_copy[i] = 0
df_copy[i] = df_copy[i].astype(int)
if j == 'object':
df_copy[i] = df_copy[i].astype('object')
for ID in df_population.index.values[:N]:
if ID not in out:
out[ID] = df_copy.copy()
out = {ID: out[ID] for ID in df_population.index.values[:N]}
assert len(out) == N
D_timeseries = out
# check each example have identical LxD table structure
ID0 = sorted(D_timeseries.keys())[0]
df0 = D_timeseries[ID0]
for ID, df_i in D_timeseries.items():
pd.testing.assert_index_equal(df_i.index, df0.index)
pd.testing.assert_index_equal(df_i.columns, df0.columns)
D_timeseries = out
D_ = len(list(D_timeseries.values())[0].columns)
# (N*L)xD^ table
## Create MultiIndex of (ID, time_bin)
index = sum([
[(ID, t_) for t_ in list(df_.index)]
for ID, df_ in sorted(D_timeseries.items())
], [])
index = pd.Index(index, names=['ID', 't_range'])
assert len(index) == N * L
## Assume all dataframes have the same columns, used after concatenation
columns = list(sorted(D_timeseries.items())[0][1].columns)
columns = np.array(columns)
dtypes = sorted(D_timeseries.items())[0][1].dtypes
## Convert each df to a numpy array
## Concatenate **sorted** numpy arrays (faster than calling pd.concat)
feature_values = [(ID, df_.to_numpy()) for ID, df_ in sorted(D_timeseries.items())]
time_series = np.concatenate([feat_val[1] for feat_val in feature_values])
assert time_series.shape == (len(index), len(columns))
df_time_series = pd.DataFrame(data=time_series, index=index, columns=columns)
# Print metadata
print('DONE: Transforming each example...')
## Freq: Count missing entries using mask
ts_mask = df_time_series[[col for col in df_time_series if col.endswith('_mask')]]
ts_mask.columns = [col.replace('_mask', '') for col in ts_mask.columns]
print('(freq) number of missing entries :\t',
'{} out of {}={} total'.format(
(1-ts_mask).astype(int).sum().sum(),
'\u00D7'.join(str(i) for i in [N,L,ts_mask.shape[1]]), ts_mask.size))
## Freq: Count imputed entries using mask and dt
ts_delta_time = df_time_series[[col for col in df_time_series if col.endswith('_delta_time')]]
ts_delta_time.columns = [col.replace('_delta_time', '') for col in ts_delta_time.columns]
imputed = (1-ts_mask).astype(bool) & (ts_delta_time > 0)
print('(freq) number of imputed entries :\t',
'{}'.format(imputed.sum().sum(), ts_delta_time.size))
imputed.sum().rename('count').to_csv(output_dir + '/' + 'freq_imputed.csv')
not_imputed = (1-ts_mask).astype(bool) & (ts_delta_time == 0)
print('(freq) number of not imputed entries :\t',
'{}'.format(not_imputed.sum().sum(), ts_delta_time.size))
not_imputed.sum().rename('count').to_csv(output_dir + '/' + 'freq_not_imputed.csv')
## Non-Freq: Count missing entries
non_freq_cols = sorted([c + '_value' for c in set(variables) - set(variables_num_freq)])
non_freqs = df_time_series[non_freq_cols]
print('(non-freq) number of missing entries :\t',
'{} out of {}={} total'.format(
non_freqs.isna().sum().sum(),
'\u00D7'.join(str(i) for i in [N,L,non_freqs.shape[1]]), non_freqs.size))
print()
print('(N \u00D7 L \u00D7 ^D) table :\t', (N, L, len(columns)))
return df_time_series, dtypes
def map_time_series_features(df_time_series, dtypes, args):
N, L = args.N, args.L
df_time_series = df_time_series.dropna(axis='columns', how='all').sort_index()
print('Discretizing features...')
ts_mask = select_dtype(df_time_series, 'mask', dtypes)
ts_mixed = select_dtype(df_time_series, '~mask', dtypes)
assert len(ts_mixed.columns) + len(ts_mask.columns) == len(df_time_series.columns)
ts_feature_mask = ts_mask.astype(int)
ts_mixed_cols = [ts_mixed[col] for col in ts_mixed.columns]
print()
discretization_bins = None
if args.discretize:
dtype = int
print('Processing', len(ts_mixed_cols), 'non-boolean variable columns...')
discretization_bins = args.X_discretization_bins
if discretization_bins is None:
print(' Computing bin edges for numeric variables...')
discretization_bins = [compute_bin_edges(col_data, q=5) for col_data in tqdm(ts_mixed_cols)]
discretization_bins = dict(discretization_bins)
else:
print(' Usng predetermined bin edges for numeric variables...')
print(' Discretizing variables to binary features')
if args.parallel:
pool = multiprocessing.Pool(args.n_jobs)
out = list(tqdm(pool.imap_unordered(
smart_qcut_dummify_parallel,
[(col_data, discretization_bins[col_data.name], args.use_ordinal_encoding) for col_data in ts_mixed_cols]), total=len(ts_mixed_cols)
))
pool.close()
pool.join()
else:
out = [smart_qcut_dummify(col_data, discretization_bins[col_data.name], use_ordinal_encoding=args.use_ordinal_encoding) for col_data in tqdm(ts_mixed_cols)]
else:
dtype = float
df = ts_mixed.copy()
# Split a mixed column into numeric and string columns
for col in df.columns:
col_data = df[col]
col_is_numeric = [is_numeric(v) for v in col_data if not pd.isnull(v)]
if not all(col_is_numeric) and any(col_is_numeric): # have mixed type values
numeric_mask = col_data.apply(is_numeric)
df[col+'_str'] = df[col].copy()
df.loc[~numeric_mask, col] = np.nan
df.loc[numeric_mask, col+'_str'] = np.nan
ts_mixed_cols = [df[col] for col in df.columns]
print('Discretizing categorical features...')
if args.parallel:
pool = multiprocessing.Pool(args.n_jobs)
out = list(tqdm(pool.imap_unordered(
smart_dummify_impute, [(col_data) for col_data in ts_mixed_cols]), total=len(ts_mixed_cols)
))
pool.close()
pool.join()
else:
out = [smart_dummify_impute(col_data) for col_data in tqdm(ts_mixed_cols)]
out = [ts_feature_mask, *out]
D_all = sum(len(df_i.columns) for df_i in out)
X_all_feature_names = np.asarray(sum([list(df_i.columns) for df_i in out], []))
X_dense = np.concatenate([df_i.values for df_i in out], axis=1).astype(dtype)
X_all = sparse.COO(X_dense)
print('Finished discretizing features')
assert X_all.shape[0] == N * L
X_all = X_all.reshape((N, L, D_all))
print()
print('Output')
print('X_all: shape={}, density={:.3f}'.format(X_all.shape, X_all.density))
return X_all, X_all_feature_names, discretization_bins
def post_filter_time_series(X_all, feature_names_all, threshold, args):
N, L = args.N, args.L
assert X_all.shape[0] == N
assert X_all.shape[1] == L
# assert X_all.dtype == int
start_time = time.time()
X0 = X_all
feature_names_0 = feature_names_all
print('Original :', len(feature_names_0))
## Remove nearly-constant features (with low variance)
sel_const = FrequencyThreshold_temporal(threshold=threshold, L=L)
sel_const.fit(X0.reshape((N*L, -1)))
m_ts_const = sel_const.get_support()
assert len(m_ts_const) == X0.shape[-1]
X1 = X0[:, :, m_ts_const]
feature_names_1 = feature_names_0[m_ts_const]
print('Nearly-constant:', len(feature_names_0) - len(feature_names_1))
print('*** time: ', time.time() - start_time)
## Keep only first of pairwise perfectly correlated features
sel_ts_corr = CorrelationSelector()
sel_ts_corr.fit(X1.reshape((N*L, -1)))
m_ts_corr = sel_ts_corr.get_support()
assert len(m_ts_corr) == X1.shape[-1]
X2 = X1[:, :, m_ts_corr]
feature_names_2 = feature_names_1[m_ts_corr]
feature_aliases = sel_ts_corr.get_feature_aliases(feature_names_1)
print('Correlated :', len(feature_names_1) - len(feature_names_2))
print('*** time: ', time.time() - start_time)
X = sparse.COO(X2)
feature_names = feature_names_2
assert X.shape == (N, L, len(feature_names))
## Save output
print()
print('Output')
print('X: shape={}, density={:.3f}'.format(X.shape, X.density))
return X, feature_names, feature_aliases