-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathanalyze_results.py
1254 lines (1039 loc) · 51.5 KB
/
analyze_results.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
import os
import shutil
import subprocess
import click
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
import matplotlib.cm as cm
import pandas as pd
import numpy as np
from util import *
from datasets import THE_DATASETS, parse_dataset
from pprint import pprint
from sklearn.metrics import roc_auc_score, auc, accuracy_score
from sklearn.neighbors import KernelDensity
DETECT_ALGS = ["ocscore", "ambig", "iforest", "lof", "mlloo"]
ADV_SETS = ["lt", "kan", "ver", "cub"]
CONF_BUCKETS = np.linspace(0.0, 1.0, 4)
CONF_BUCKETS[-1] = 1.0001 # make last conf_hi inclusive
CONF_BUCKETS = list(zip(CONF_BUCKETS, CONF_BUCKETS[1:]))
TABLE_DIR = os.environ["LATEX_TABLE_PATH"]
FIG_DIR = os.environ["LATEX_FIGURE_PATH"]
ALG_LS = {
"ocscore": "-",
"ambig": "--",
"iforest": ":",
"lof": (5, (10, 3)),
"mlloo": "-.",
}
SET_HATCH = {
"test": "",
"lt": "|||",
"kan": "///",
"ver": "---",
"cub": "\\\\\\",
}
SAVEFIGS=True
configure_matplotlib()
def collect_aucs_for_dataset(dataset, model_type, N, ratio, nfolds, cache_dir, seed):
d, num_trees, tree_depth, lr = dataset
d.seed = seed
d.nfolds = nfolds
#d.load_dataset()
#d.minmax_normalize()
aucs = {adv_set: {detect_alg: {"values": [], "mean": 0.0, "std": 0.0} for
detect_alg in DETECT_ALGS} for adv_set in ADV_SETS}
print(d.name())
for fold in range(nfolds):
report_name = get_report_name(d, seed, fold, N, ratio, model_type,
num_trees, tree_depth, lr, cache_dir,
special="")
if not os.path.exists(report_name):
print("(!) skipping", report_name)
continue
report = load(report_name)
for adv_set in ADV_SETS:
for detect_alg in DETECT_ALGS:
aucs[adv_set][detect_alg]["values"].append(report[adv_set][detect_alg]["auc"])
for adv_set in ADV_SETS:
for detect_alg in DETECT_ALGS:
vs = aucs[adv_set][detect_alg]["values"]
aucs[adv_set][detect_alg]["mean"] = np.mean(vs)
aucs[adv_set][detect_alg]["std"] = np.std(vs)
print()
return aucs
def collect_aucs(model_type, N, ratio, nfolds, cache_dir, seed):
aucs = {}
for dname in THE_DATASETS:
dataset = parse_dataset(dname)
aucs[dname] = collect_aucs_for_dataset(dataset, model_type, N, ratio,
nfolds, cache_dir, seed)
return aucs
# per confidence bucket, compute AUC
def collect_aucs_per_conf_for_dataset(dataset, model_type, N, ratio, nfolds,
cache_dir, seed):
d, num_trees, tree_depth, lr = dataset
d.seed = seed
d.nfolds = nfolds
d.load_dataset()
d.minmax_normalize()
aucs_per_conf = {detect_alg: [{"values": []} for i in
range(len(CONF_BUCKETS))] for detect_alg in
DETECT_ALGS}
for fold in range(nfolds):
report_name = get_report_name(d, seed, fold, N, ratio, model_type,
num_trees, tree_depth, lr, cache_dir,
special="")
if not os.path.exists(report_name):
print("(!) skipping", report_name)
continue
model, meta, at = get_model(d, model_type, fold, lr, num_trees,
tree_depth,
groot_epsilon=0.5*INPUT_DELTA[d.name()])
report = load(report_name)
sample_indices = report["sample_indices"]
Xsample = d.X.iloc[sample_indices, :].to_numpy()
Xall = Xsample
Sall = {}
ydetect = np.zeros(len(sample_indices), dtype=bool)
ysource = np.array(["test"] * len(sample_indices))
for adv_set in ADV_SETS:
#for adv_set in ["cub"]:
advs_fname = report[f"{adv_set}_fname"]
advs = load(advs_fname)
Xadv = np.array([x["adv_example"] for x in advs])
Xall = np.vstack((Xall, Xadv))
ydetect = np.hstack((ydetect, np.ones(Xadv.shape[0], dtype=bool)))
ysource = np.hstack((ysource, [adv_set] * Xadv.shape[0]))
for detect_alg in DETECT_ALGS:
S = report[adv_set][detect_alg]["S"]
Nsample = report["Nsample"]
if detect_alg not in Sall:
Sall[detect_alg] = S
else:
Sall[detect_alg] = np.hstack((Sall[detect_alg], S[Nsample:]))
assert len(ydetect) == Xall.shape[0]
for detect_alg in DETECT_ALGS:
assert len(Sall[detect_alg]) == Xall.shape[0]
pred_prob = at.predict_proba(Xall)
conf = np.abs((pred_prob * 2.0) - 1.0)
for i, (conf_lo, conf_hi) in enumerate(CONF_BUCKETS):
#conf_sample = (conf_lo <= conf) & (conf < conf_hi)
#conf_sample = (conf_lo <= conf)
conf_sample = (conf < conf_hi)
if conf_sample.sum() == 0:
print(f"SAMPLE EMPTY for conf [{conf_lo:.2f}, {conf_hi:.2f})")
ydetect_sample = ydetect[conf_sample]
ysource_sample = ysource[conf_sample]
print(f"conf {conf_lo:.1f}",
"#normal", len(ydetect_sample),
"#adversarial", int(sum(ydetect_sample)))
if np.all(ydetect_sample == ydetect_sample[0]):
print(f"WARNING: all the same ydetect for conf [{conf_lo:.2f}, {conf_hi:.2f})",
ydetect_sample[0], d.name(), "fold", fold)
for detect_alg in DETECT_ALGS:
S = Sall[detect_alg][conf_sample]
if np.all(ydetect_sample == ydetect_sample[0]):
auc = 1.0
else:
auc = roc_auc_score(ydetect_sample, S)
aucs_per_conf[detect_alg][i]["values"].append(auc)
aucs_per_conf[detect_alg][i]["N"] = len(ydetect_sample)
aucs_per_conf[detect_alg][i]["Nadv"] = int(sum(ydetect_sample))
for i in range(len(CONF_BUCKETS)):
for detect_alg in DETECT_ALGS:
vs = aucs_per_conf[detect_alg][i]["values"]
aucs_per_conf[detect_alg][i]["mean"] = np.mean(vs)
aucs_per_conf[detect_alg][i]["std"] = np.std(vs)
return aucs_per_conf
def collect_aucs_per_conf(model_type, N, ratio, nfolds, cache_dir, seed):
aucs_per_conf = {}
#for dname in THE_DATASETS:
for dname in ["phoneme", "spambase", "mnist2v4", "calhouse"]:
dataset = parse_dataset(dname)
aucs_per_conf[dname] = collect_aucs_per_conf_for_dataset(dataset,
model_type, N,
ratio, nfolds,
cache_dir,
seed)
return aucs_per_conf
def collect_accs_fixed_threshold_for_dataset(dataset, model_type, N, ratio,
nfolds, cache_dir, seed):
d, num_trees, tree_depth, lr = dataset
d.seed = seed
d.nfolds = nfolds
d.load_dataset()
d.minmax_normalize()
accs = {detect_alg: {} for detect_alg in DETECT_ALGS}
accs["nadvs"] = {}
for fold in range(nfolds):
report_name = get_report_name(d, seed, fold, N, ratio, model_type,
num_trees, tree_depth, lr, cache_dir,
special="")
if not os.path.exists(report_name):
print("(!) skipping", report_name)
continue
model, meta, at = get_model(d, model_type, fold, lr, num_trees,
tree_depth,
groot_epsilon=0.5*INPUT_DELTA[d.name()])
report = load(report_name)
sample_indices = report["sample_indices"]
Nsample = len(sample_indices)
Xsample = d.X.iloc[sample_indices, :].to_numpy()
Xall = Xsample
Xbase_all = None
Sall = {}
ydetect = np.zeros(len(sample_indices), dtype=bool)
ysource = np.array(["test"] * len(sample_indices))
# Step 1: collect all the S'sses for all the folds and for all
# adversarial sets per detection method
for adv_set in ADV_SETS:
advs_fname = report[f"{adv_set}_fname"]
advs = load(advs_fname)
Xadv = np.array([x["adv_example"] for x in advs])
Xbase = np.array([x["base_example"] for x in advs])
Xall = np.vstack((Xall, Xadv))
Xbase_all = np.vstack((Xbase_all, Xbase)) if Xbase_all is not None else Xbase
ydetect = np.hstack((ydetect, np.ones(Xadv.shape[0], dtype=bool)))
ysource = np.hstack((ysource, [adv_set] * Xadv.shape[0]))
for detect_alg in DETECT_ALGS:
S = report[adv_set][detect_alg]["S"]
results[adv_set][detect_alg]["S"] = S
if detect_alg not in Sall:
Sall[detect_alg] = S
else:
Sall[detect_alg] = np.hstack((Sall[detect_alg], S[Nsample:]))
assert len(ydetect) == Xall.shape[0]
for detect_alg in DETECT_ALGS:
assert len(Sall[detect_alg]) == Xall.shape[0]
# Compute the confidence of the ensemble for each example
pred_prob = at.predict_proba(Xall)
conf = np.abs(pred_prob - 0.5) + 0.5
conf_perm = np.argsort(conf)
# Compute linf deltas between original examples and the adversarial examples
delta = np.sum(np.abs(Xall[Nsample:, :]-Xbase_all), axis=1)
# Step 2: Fix the threshold
# fix threshold, then check accuracy for each point
Nall = Xall.shape[0]
sample_ratio = Nsample/Nall
for detect_alg in DETECT_ALGS:
S = Sall[detect_alg]
threshold = np.quantile(S, sample_ratio)
ydetect_pred = S > threshold
ydetect_acc = ydetect_pred == ydetect
window_size = 100
#windows = np.arange(window_size)[np.newaxis, :]\
# + np.arange(len(S)-window_size+1)[:, np.newaxis]
sliding_acc = np.lib.stride_tricks.sliding_window_view(
ydetect_acc[conf_perm], window_size).mean(axis=1)
sliding_conf = np.lib.stride_tricks.sliding_window_view(
conf[conf_perm], window_size).mean(axis=1)
#skipper = np.arange(1, len(sliding_acc), 4)
#sliding_acc = sliding_acc[skipper]
#sliding_conf = sliding_conf[skipper]
#sliding_acc = np.convolve(sliding_acc, np.ones(5)/5, mode="same")
# sliding window on confidence
if "sliding_acc" not in accs[detect_alg]:
accs[detect_alg]["sliding_acc"] = np.zeros((nfolds, len(sliding_acc)))
accs[detect_alg]["sliding_conf"] = np.zeros((nfolds, len(sliding_acc)))
accs[detect_alg]["sliding_acc"][fold, :] = sliding_acc
accs[detect_alg]["sliding_conf"][fold, :] = sliding_conf
# how many adversarial examples in the window?
#sliding_nadv = np.lib.stride_tricks.sliding_window_view(
# ydetect[conf_perm], window_size).mean(axis=1)
#sliding_conf = np.lib.stride_tricks.sliding_window_view(
# conf[conf_perm], window_size).mean(axis=1)
#if "sliding_acc" not in accs["nadvs"]:
# accs["nadvs"]["sliding_acc"] = np.zeros((nfolds, len(sliding_acc)))
# accs["nadvs"]["sliding_conf"] = np.zeros((nfolds, len(sliding_acc)))
#accs["nadvs"]["sliding_acc"][fold, :] = sliding_nadv
#accs["nadvs"]["sliding_conf"][fold, :] = sliding_conf
nadv = np.cumsum(ydetect[conf_perm]) / (Xall.shape[0]-Nsample) # how many adversarials have we seen?
nadv = np.cumsum(np.ones(len(ydetect))) / Xall.shape[0] # how many examples have we seen?
if "sliding_acc" not in accs["nadvs"]:
accs["nadvs"]["sliding_acc"] = np.zeros((nfolds, len(nadv)))
accs["nadvs"]["sliding_conf"] = np.zeros((nfolds, len(nadv)))
accs["nadvs"]["sliding_acc"][fold, :] = nadv
accs["nadvs"]["sliding_conf"][fold, :] = conf[conf_perm]
for detect_alg in accs.keys():
skipper = np.arange(1, accs[detect_alg]["sliding_acc"].shape[1], 5)
accs[detect_alg]["acc_mean"] = np.mean(accs[detect_alg]["sliding_acc"], axis=0)[skipper]
accs[detect_alg]["acc_std"] = np.std(accs[detect_alg]["sliding_acc"], axis=0)[skipper]
accs[detect_alg]["conf"] = np.mean(accs[detect_alg]["sliding_conf"], axis=0)[skipper]
return accs
def collect_accs_fixed_threshold(model_type, N, ratio, nfolds, cache_dir, seed):
accs = {}
#for dname in THE_DATASETS:
#for dname in ["covtype", "spambase", "mnist2v4", "calhouse", "ijcnn1"]:
for dname in ["covtype", "mnist2v4"]:
dataset = parse_dataset(dname)
accs[dname] = collect_accs_fixed_threshold_for_dataset(dataset,
model_type, N,
ratio, nfolds,
cache_dir, seed)
return accs
def collect_results_dataset(dataset, model_type, N, ratio, nfolds, cache_dir, seed):
d, num_trees, tree_depth, lr = dataset
d.seed = seed
d.nfolds = nfolds
d.load_dataset()
d.minmax_normalize()
per_set_all = []
per_alg_all = []
per_set_alg_all = []
conf_sample = []
for fold in range(nfolds):
per_alg = {detect_alg: {} for detect_alg in DETECT_ALGS}
per_set = {adv_set: {} for adv_set in ADV_SETS}
per_set_alg = {adv_set: {detect_alg: {}
for detect_alg in DETECT_ALGS}
for adv_set in ADV_SETS}
per_alg_all.append(per_alg)
per_set_all.append(per_set)
per_set_alg_all.append(per_set_alg)
report_name = get_report_name(d, seed, fold, N, ratio, model_type,
num_trees, tree_depth, lr, cache_dir,
special="")
if not os.path.exists(report_name):
print("(!) skipping", report_name)
continue
model, meta, at = get_model(d, model_type, fold, lr, num_trees,
tree_depth,
groot_epsilon=0.5*INPUT_DELTA[d.name()])
report = load(report_name)
sample_indices = report["sample_indices"]
Nsample = len(sample_indices)
Xsample = d.X.iloc[sample_indices, :].to_numpy()
pred_prob_sample = at.predict_proba(Xsample)
conf_sample.append(np.abs(pred_prob_sample - 0.5) + 0.5)
Xadv, Xbase = {}, {}
per_alg["ocscore"]["setup_time"] = report["refset_time"]
per_alg["iforest"]["setup_time"] = report["iforest_time"]
per_alg["ambig"]["setup_time"] = 0.0
per_alg["mlloo"]["setup_time"] = 0.0
if "lof_time" in report:
per_alg["lof"]["setup_time"] = report["lof_time"]
# per adv set
for adv_set in ADV_SETS:
advs_fname = report[f"{adv_set}_fname"]
advs = load(advs_fname)
Xadv[adv_set] = np.array([x["adv_example"] for x in advs]).astype(np.float32)
Xbase[adv_set] = np.array([x["base_example"] for x in advs]).astype(np.float32)
# DELTA
absdiff = np.abs(Xadv[adv_set]-Xbase[adv_set]).max(axis=1)
per_set[adv_set]["delta"] = absdiff
# CONFIDENCE
pred_prob = at.predict_proba(Xadv[adv_set])
conf = np.abs(pred_prob - 0.5) + 0.5
per_set[adv_set]["pred_prob"] = pred_prob
per_set[adv_set]["conf"] = conf
# collect aucs and scores S per adv_set and per algorithm
for adv_set in ADV_SETS:
for detect_alg in DETECT_ALGS:
per_set_alg[adv_set][detect_alg]["auc"] = report[adv_set][detect_alg]["auc"]
per_set_alg[adv_set][detect_alg]["time"] = report[adv_set][detect_alg]["time"]
Sfull = report[adv_set][detect_alg]["S"]
Sadv = Sfull[Nsample:]
per_set_alg[adv_set][detect_alg]["S"] = Sadv
sample_ratio_per_set = (Nsample+1)/len(Sfull)
per_set_alg[adv_set][detect_alg]["threshold_per_set"] =\
np.quantile(Sfull, sample_ratio_per_set)
# scores per algorithm on the test set sample
for detect_alg in DETECT_ALGS:
S = report["ver"][detect_alg]["S"][:Nsample]
per_alg[detect_alg]["Ssample"] = S
# determine fixed threshold on all adversarial examples for each detection algorithm
Xall = np.vstack([Xsample] + list(Xadv.values()))
is_adv = np.hstack([np.zeros(Nsample), np.ones(Xall.shape[0]-Nsample)])
sample_ratio = (Nsample+1)/Xall.shape[0]
for detect_alg in DETECT_ALGS:
Ssample = per_alg[detect_alg]["Ssample"]
Sall = np.hstack([Ssample] + [per_set_alg[s][detect_alg]["S"] for s in ADV_SETS])
thrs = np.quantile(Sall, sample_ratio)
per_alg[detect_alg]["threshold"] = thrs
# collect aggregated stats over all 4 sets
per_alg[detect_alg]["acc_aggr"] = np.mean(is_adv == (Sall >= thrs))
per_alg[detect_alg]["auc_aggr"] = roc_auc_score(is_adv, Sall)
per_alg[detect_alg]["is_adv_pred_sample"] = (Ssample >= thrs)
# collect accs per adv_set and per algorithm using threshold
for adv_set in ADV_SETS:
for detect_alg in DETECT_ALGS:
S = per_set_alg[adv_set][detect_alg]["S"]
Ssample = per_alg[detect_alg]["Ssample"]
thrs = per_alg[detect_alg]["threshold"]
is_adv_pred = (S >= thrs)
per_set_alg[adv_set][detect_alg]["is_adv_pred"] = is_adv_pred
per_set_alg[adv_set][detect_alg]["acc"] = np.mean(is_adv_pred)
thrs_per_set = per_set_alg[adv_set][detect_alg]["threshold_per_set"]
per_set_alg[adv_set][detect_alg]["acc_per_set"] = np.mean(S >= thrs_per_set)
per_set_alg[adv_set][detect_alg]["acc_sample_per_set"] = np.mean(Ssample < thrs_per_set)
# performance on test set sample
for detect_alg in DETECT_ALGS:
S = per_alg[detect_alg]["Ssample"]
thrs = per_alg[detect_alg]["threshold"]
per_alg[detect_alg]["acc_sample"] = np.mean(S < thrs)
## END FOLDS FOR-LOOP
# collect everything over the folds
per_alg = {detect_alg: {} for detect_alg in DETECT_ALGS}
per_set = {adv_set: {} for adv_set in ADV_SETS}
per_set_alg = {adv_set: {detect_alg: {}
for detect_alg in DETECT_ALGS}
for adv_set in ADV_SETS}
for adv_set in ADV_SETS:
for detect_alg in DETECT_ALGS:
for k in ["auc", "acc", "threshold_per_set", "acc_per_set",
"acc_sample_per_set", "time"]:
per_set_alg[adv_set][detect_alg][f"{k}_mean"] =\
np.mean([x[adv_set][detect_alg][k] for x in per_set_alg_all])
per_set_alg[adv_set][detect_alg][f"{k}_std"] =\
np.std([x[adv_set][detect_alg][k] for x in per_set_alg_all])
for adv_set in ADV_SETS:
for k in ["delta"]:
per_set[adv_set][f"{k}_mean"] = np.mean([x[adv_set][k].mean()
for x in per_set_all])
per_set[adv_set][f"{k}_std"] = np.std([x[adv_set][k].mean()
for x in per_set_all])
for detect_alg in DETECT_ALGS:
for k in ["acc_sample", "threshold", "acc_aggr", "auc_aggr"]:
per_alg[detect_alg][f"{k}_mean"] = np.mean([x[detect_alg][k].mean()
for x in per_alg_all])
per_alg[detect_alg][f"{k}_std"] = np.std([x[detect_alg][k].mean()
for x in per_alg_all])
per_alg[detect_alg]["setup_time_mean"] = np.mean([x[detect_alg]["setup_time"]
for x in per_alg_all])
# combine all is_adv_pred's over all folds, together with conf & delta
conf_sample = np.hstack(conf_sample)
delta_sample = np.zeros(conf_sample.shape)
per_confdelta = {
"conf": conf_sample,
"delta": delta_sample,
"set": ["test"] * len(conf_sample),
"is_adv_pred": {
detect_alg: np.hstack([pa[detect_alg]["is_adv_pred_sample"] for pa in per_alg_all])
for detect_alg in DETECT_ALGS
}
}
for fold in range(nfolds):
ps = per_set_all[fold]
psa = per_set_alg_all[fold]
# adversarial SETS for fold
for adv_set in ADV_SETS:
set_arr = [adv_set] * len(ps[adv_set]["conf"])
per_confdelta["conf"] = np.hstack((per_confdelta["conf"], ps[adv_set]["conf"]))
per_confdelta["delta"] = np.hstack((per_confdelta["delta"], ps[adv_set]["delta"]))
per_confdelta["set"] = np.hstack((per_confdelta["set"], set_arr))
pred = per_confdelta["is_adv_pred"]
for detect_alg in DETECT_ALGS:
pred[detect_alg] = np.hstack((
pred[detect_alg],
psa[adv_set][detect_alg]["is_adv_pred"]))
return per_set, per_alg, per_set_alg, per_confdelta
def collect_results(dnames, model_type, N, ratio, nfolds, cache_dir, seed):
per_set, per_alg, per_set_alg, per_confdelta = {}, {}, {}, {}
for dname in dnames:
#for dname in ["phoneme", "covtype", "mnist2v4", "ijcnn1", "webspam", "calhouse"]:
#for dname in ["mnist2v4", "covtype", "ijcnn1", "webspam"]:
#for dname in ["phoneme", "mnist2v4"]:
dataset = parse_dataset(dname)
ps, pa, psa, pcd = collect_results_dataset(dataset, model_type, N,
ratio, nfolds, cache_dir,
seed)
per_set[dname] = ps
per_alg[dname] = pa
per_set_alg[dname] = psa
per_confdelta[dname] = pcd
return per_set, per_alg, per_set_alg, per_confdelta
def display_results(per_set, per_alg, per_set_alg, model_type):
dnames = list(per_set.keys())
df_auc_per_alg = pd.DataFrame("-", index=USED_DATASETS, columns=DETECT_ALGS)
df_acc_per_alg = pd.DataFrame("-", index=USED_DATASETS, columns=DETECT_ALGS)
df_auc_per_set = pd.DataFrame("-", index=DETECT_ALGS, columns=ADV_SETS)
set_kinds = ["low confidence", "high confidence"]
df_auc_per_kind = pd.DataFrame("-", index=DETECT_ALGS, columns=set_kinds)
index = pd.MultiIndex.from_product([DETECT_ALGS, ADV_SETS], names=["Algorithm", "Set"])
df_auc = pd.DataFrame("-", index=index, columns=USED_DATASETS)
df_time = pd.DataFrame("-", index=USED_DATASETS, columns=DETECT_ALGS)
df_setup_time = pd.DataFrame("-", index=USED_DATASETS, columns=DETECT_ALGS)
df_delta = pd.DataFrame("-", index=USED_DATASETS, columns=ADV_SETS)
for dname in dnames:
ps = per_set[dname]
pa = per_alg[dname]
psa = per_set_alg[dname]
aucmax = max(pa[da]["auc_aggr_mean"] for da in DETECT_ALGS)
aucmax_ps = {s:max(psa[s][da]["auc_mean"] for da in DETECT_ALGS) for s in ADV_SETS}
for detect_alg in pa.keys():
v = pa[detect_alg]["auc_aggr_mean"]
e = pa[detect_alg]["auc_aggr_std"]
df_auc_per_alg.loc[dname, detect_alg] = fmt_val_std(v, e, aucmax)
df_acc_per_alg.loc[dname, detect_alg] = pa[detect_alg]["acc_aggr_mean"]
v = np.sum([psa[a][detect_alg]["time_mean"] for a in psa.keys()]) / (5*2500) * 1000
e = np.sum([psa[a][detect_alg]["time_std"] for a in psa.keys()]) / (5*2500) * 1000
#df_time.loc[dname, detect_alg] = f"{v:1.2f}{{\\tiny±{e:1.2f}}}"
df_time.loc[dname, detect_alg] = f"{v:1.3f}"
v = pa[detect_alg]["setup_time_mean"]
df_setup_time.loc[dname, detect_alg] = f"{v:1.3f}"
for adv_set in ADV_SETS:
v = psa[adv_set][detect_alg]["auc_mean"]
e = psa[adv_set][detect_alg]["auc_std"]
df_auc.loc[(detect_alg, adv_set), dname] = fmt_val_std(v, e, aucmax_ps[adv_set])
for adv_set in ADV_SETS:
v = ps[adv_set]["delta_mean"]
#e = ps[adv_set]["delta_std"]
#df_delta.loc[dname, adv_set] = fmt_val_std(v, e, 99.0)
df_delta.loc[dname, adv_set] = f"{v:1.4f}".lstrip("0")
# Per set performance of each alg, averaged over all datasets
for adv_set in ADV_SETS:
aucs = {}
for detect_alg in DETECT_ALGS:
vs = [per_set_alg[d][adv_set][detect_alg]["auc_mean"] for d in dnames]
aucs[detect_alg] = vs
aucmax = max(np.mean(v) for v in aucs.values())
for detect_alg, vs in aucs.items():
v = np.mean(aucs[detect_alg])
e = np.std(aucs[detect_alg])
df_auc_per_set.loc[detect_alg, adv_set] = fmt_val_std(v, e, aucmax)
# Per set 'kind' performance of each alg, averaged over all datasets
for kind, adv_sets in zip(set_kinds, [["lt", "kan"], ["ver", "cub"]]):
aucs = {}
for detect_alg in DETECT_ALGS:
vs = []
for adv_set in adv_sets:
vs += [per_set_alg[d][adv_set][detect_alg]["auc_mean"]
for d in dnames]
aucs[detect_alg] = vs
aucmax = max(np.mean(v) for v in aucs.values())
for detect_alg, vs in aucs.items():
v = np.mean(aucs[detect_alg])
e = np.std(aucs[detect_alg])
df_auc_per_kind.loc[detect_alg, kind] = fmt_val_std(v, e, aucmax)
print("\nauc_aggr AUC aggregated over all sets")
print(df_auc_per_alg)
with open(os.path.join(TABLE_DIR, f"auc_aggr_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_auc_per_alg.T.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"l"*df_auc_per_alg.shape[0])
print("\nauc_aggr AUC aggregated over all datasets")
print(df_auc_per_set)
print(df_auc_per_kind)
with open(os.path.join(TABLE_DIR, f"auc_aggr_per_set_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_auc_per_kind.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"c"*df_auc_per_kind.shape[1])
print("\nDF AUC")
print(df_auc)
with open(os.path.join(TABLE_DIR, f"auc_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_auc.to_latex(buf=f, longtable=False, escape=False,
column_format="ll"+"l"*df_auc.shape[1])
print("\nDF TIME")
print(df_time)
with open(os.path.join(TABLE_DIR, f"time_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_time.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"l"*df_time.shape[1])
print("\nDF SETUP TIME")
print(df_setup_time)
with open(os.path.join(TABLE_DIR, f"setup_time_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_setup_time.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"l"*df_setup_time.shape[1])
print("\nDF DELTA")
print(df_delta)
with open(os.path.join(TABLE_DIR, f"delta_table_{model_type}.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_delta.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"l"*df_delta.shape[1])
def dataset_prop_table(N, ratio, nfolds, cache_dir, seed):
columns_tex = ["$\#F$", "$n$", "$\\eta$", "$M$", "$d_T$", "class balance"]
columns = ["nfeat", "nexamples", "lr", "ntrees", "tree_depth", "classbal"]
df_dsets = pd.DataFrame("-", index=USED_DATASETS, columns=columns)
columns2_tex = ["$\#F$", "$n$", "$\\eta$", "$M$", "tree depth", "class balance"]
model_types = ["xgb", "rf", "groot"]
columns2 = pd.MultiIndex.from_product([model_types,
["acc train", "acc test", "refset"]],
names=["", ""])
df_metrics = pd.DataFrame(0.0, index=USED_DATASETS, columns=columns2)
for dname in USED_DATASETS:
#for dname in ["mnist2v4", "fmnist2v4"]:
d, num_trees, tree_depth, lr = parse_dataset(dname)
d.seed = seed
d.nfolds = nfolds
d.load_dataset()
d.minmax_normalize()
df_dsets.loc[dname, "nfeat"] = d.X.shape[1]
df_dsets.loc[dname, "nexamples"] = f"{d.X.shape[0]/1000:.1f}k"
df_dsets.loc[dname, "lr"] = f"{lr:.1f}"
df_dsets.loc[dname, "ntrees"] = num_trees
df_dsets.loc[dname, "tree_depth"] = tree_depth
df_dsets.loc[dname, "classbal"] = f"{np.mean(d.y==0)*100:.0f}\\%"
#for model_type in model_types:
# for fold in range(nfolds):
# model, meta, at = get_model(d, model_type, fold, lr, num_trees,
# tree_depth,
# groot_epsilon=0.5*INPUT_DELTA[d.name()])
# full_refset0, full_refset1, refset_time = get_refset(d, at, fold)
# xtrain, ytrain, xtest, ytest = d.train_and_test_set(fold)
# xtrain, ytrain = xtrain.to_numpy(), ytrain.to_numpy()
# xtest, ytest = xtest.to_numpy(), ytest.to_numpy()
# acc_train = accuracy_score(ytrain, at.eval(xtrain) > 0.0)
# acc_test = accuracy_score(ytest, at.eval(xtest) > 0.0)
# refset_size = full_refset0.shape[0] + full_refset1.shape[0]
# print(f"REFSET {dname} {model_type} {fold}: {refset_time:.1f}s, "\
# f"{full_refset0.shape[0]} + {full_refset1.shape[0]}, "\
# f"{acc_train*100:.1f}%, {meta['metric'][1]*100:.1f}%, {acc_test*100:.1f}%")
# df_metrics.loc[dname, (model_type, "refset")] += refset_size
# df_metrics.loc[dname, (model_type, "acc train")] += acc_train
# df_metrics.loc[dname, (model_type, "acc test")] += acc_test
print("\nDF DATASET PROPERTIES")
print(df_dsets)
with open(os.path.join(TABLE_DIR, f"datasets_table.tex"), "w") as f:
#df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
df_dsets.columns = columns_tex
df_dsets.to_latex(buf=f, longtable=False, escape=False,
column_format="l"+"r"*df_dsets.shape[1])
#print("\nDF DATASET PROPERTIES 2")
#df_metrics /= nfolds
#for mt in model_types:
# df_metrics.loc[:, (mt, "acc train")] = [f"{x*100:.1f}\\%"
# for x in df_metrics.loc[:, (mt, "acc train")]]
# df_metrics.loc[:, (mt, "acc test")] = [f"{x*100:.1f}\\%"
# for x in df_metrics.loc[:, (mt, "acc test")]]
# df_metrics.loc[:, (mt, "refset")] = [f"{x/1000:.1f}k"
# for x in df_metrics.loc[:, (mt, "refset")]]
#df_metrics.columns = pd.MultiIndex.from_product([[f"\\bf {mt.upper()}" for mt in model_types],
# ["acc. train", "acc. test", "$|R|$"]],
# names=["", ""])
#print(df_metrics)
#with open(os.path.join(TABLE_DIR, f"datasets_metrics_table.tex"), "w") as f:
# #df_auc_per_alg.columns = [f"{{{x}}}" for x in df_auc_per_alg.columns]
# #df_metrics.columns = columns2_tex
# df_metrics.to_latex(buf=f, longtable=False, escape=False,
# column_format="l"+"r"*df_metrics.shape[1])
def plot_confdelta(per_confdelta, name, model_type):
dnames = per_confdelta.keys()
fig, axs = plt.subplots(2, len(dnames),
figsize=(6.2 * 3.0/4.0, 1.8),
sharex=True,
gridspec_kw={'height_ratios':[2, 1]})
fig.subplots_adjust(left=0.08, bottom=0.25, right=0.98, hspace=0.2,
wspace=0.35, top=0.78)
print("plot_confdelta", dnames)
for ax, axl, dname in zip(axs[0, :], axs[1, :], dnames):
pcd = per_confdelta[dname]
conf = pcd["conf"]
#conf = pcd["delta"]
#confq0 = np.linspace(min(conf), max(conf), 20)
#confq1 = np.quantile(conf, np.linspace(0, 1, 101))
#confq = np.sort(np.hstack((confq0, confq1)))
confq = np.quantile(conf, np.linspace(0, 1, 201))
confq[-1] += 0.001 # make inclusive
intervals = list(zip(confq, confq[20:]))
xs = []
xs2 = [] # average conf in window instead of mid
ns = []
ns_adv = []
ws = [] # number of ex in window
is_adv = pcd["set"] != "test"
n_adv = sum(is_adv)
for vlo, vhi in intervals:
vmid = vlo + (vhi-vlo)/2.0
xs.append(vmid)
mask = (conf < vhi)
ns.append(mask.mean())
ns_adv.append((mask & is_adv).sum() / n_adv)
mask = (vlo <= conf) & (conf < vhi)
ws.append(mask.mean())
xs2.append(conf[mask].mean())
ns[0] = 0
#axl.plot(ns, xs, label="Confidence", color="gray", ls="--")
axl.plot(ns, xs2, label="Confidence", color="gray", ls="-")
#axlt = axl.twinx()
#axlt.plot(ns, ns_adv, label="Fraction Adv", color="gray", ls=":")
#axlt.set_ylim([0.0, 1.05])
#axl.plot(xs, ns, label="#ex seen", color="gray", ls="-")
#axl.plot(ns, ns_adv, label="#advs", color="lightgray", ls=":")
#ax.plot(ns, ws, label="#window", color="lightgray", ls="-")
#print("window sizes", dname, np.round(ws, 3))
lines = []
for detect_alg in DETECT_ALGS:
is_adv_pred = pcd["is_adv_pred"][detect_alg]
is_correct = is_adv == is_adv_pred
acc_per_conf = []
for vlo, vhi in intervals:
mask = (vlo <= conf) & (conf < vhi)
acc_per_conf.append(is_correct[mask].mean())
l, = ax.plot(ns, acc_per_conf, label=detect_alg, ls=ALG_LS[detect_alg])
lines.append(l)
#axl.set_xlabel("Confidence")
axl.set_xlabel("fraction of\nexamples")
axl.text(0, 0, "fraction of examples")
#xticks = np.linspace(0, 1, 5)
#xticklabels = np.quantile(confq, xticks).round(2)
#ax.set_xticks(xticks)
#ax.set_xticklabels(xticklabels)
ax.set_title(dname)
ax.set_xlim([0.0, 1.0])
axl.set_xlim([0.0, 1.0])
ax.set_ylim([0.0, 1.05])
axl.set_xticks([0.0, 0.2, 0.4, 0.6, 0.8, 1.0])
axl.set_xticklabels([".0", ".2", ".4", ".6", ".8", "1"])
ax.set_xticks(axl.get_xticks())
ax.set_yticks([0.0, 0.5, 1.0])
ax.set_yticklabels(["0", ".5", "1"])
axl.set_yticks([0.5, 1.0])
axl.set_yticklabels([".5", "1"])
#ax.set_xticklabels([""] * len(axl.get_xticks()))
#ax.grid(visible=True, axis="x", which="major", color="lightgray")
#axl.grid(visible=True, axis="x", which="major", color="lightgray")
axl.set_ylim([0.5, 1.05])
#fig, ax = plt.subplots()
#ax.set_title(dname)
#for adv_set in ADV_SETS:
# mask = pcd["set"] == adv_set
# mask0 = mask & (is_adv_pred == False)
# mask1 = mask & (is_adv_pred == True)
# x = pcd["conf"][mask0]
# y = pcd["delta"][mask0]
# s, = ax.plot(x, y, "x", label=adv_set)
# ax.set_xscale("log")
# ax.set_yscale("log")
# x = pcd["conf"][mask1]
# y = pcd["delta"][mask1]
# ax.plot(x, y, "d", label=adv_set, color=s.get_color())
#ax.legend()
#axs[0, 0].legend()
fig.legend(handles=lines, ncol=len(lines), fancybox=False, loc="upper center")
axs[0, 0].set_ylabel("Accuracy")
#axs[1, 0].legend()
axs[1, 0].set_ylabel("Conf.")
if SAVEFIGS:
figname = f"acc_per_conf_{name}_{model_type}"
fig.savefig(os.path.join("figures", f"{figname}.svg"))
subprocess.run(["/home/laurens/repos/dotfiles/scripts/svg2latex", f"{figname}.svg"], cwd="figures")
shutil.copyfile(os.path.join("figures", f"{figname}.pdf_tex"),
os.path.join(FIG_DIR, f"{figname}.pdf_tex"))
shutil.copyfile(os.path.join("figures", f"{figname}.pdf"),
os.path.join(FIG_DIR, f"{figname}.pdf"))
plt.show()
def plot_confdist(per_confdelta, model_type): # all in one plot --> busy hard to read
dnames = per_confdelta.keys()
#fig, axs = plt.subplots(1, len(dnames), figsize=(20, 5))
#fig.subplots_adjust(left=0.02, right=0.98)
cmap = plt.get_cmap("tab10")
all_conf = np.array([])
all_sets = np.array([])
#for ax, dname in zip(axs.ravel(), dnames):
for k, dname in enumerate(dnames):
pcd = per_confdelta[dname]
conf = pcd["conf"]
sets = pcd["set"]
all_conf = np.hstack((all_conf, conf))
all_sets = np.hstack((all_sets, sets))
bins = np.linspace(0.5, 1.0, 21)
x = np.linspace(0.5, 1.0, 101)
bottom = np.zeros(x.shape)
#ax=axs.ravel()[k]
#for u, s in enumerate(ADV_SETS + ["test"]):
# conf_s = conf[sets == s]
# hist, bin_edges = np.histogram(conf_s, bins=bins)
# hist = hist / len(conf)
# kwargs = {"facecolor": cmap(u)}
# if s == "test":
# kwargs["facecolor"] = "gray"
# kde = KernelDensity(kernel='gaussian', bandwidth=0.03).fit(conf_s.reshape(-1, 1))
# y = np.exp(kde.score_samples(x.reshape(-1, 1))) / (len(ADV_SETS)+1)
# #ax.bar(bins[:-1], hist, width=0.05, align="edge", bottom=bottom,
# # label=s, hatch=SET_HATCH[s],
# # linewidth=0.5, edgecolor="black", **kwargs)
# #ax.plot(bins[:-1], hist+bottom, label=s)
# ax.fill_between(x, bottom, bottom+y, label=s, edgecolor="black",
# linewidth=0.5, hatch=SET_HATCH[s], **kwargs)
# #print(f"{s:3}", hist)
# bottom += y
#ax.set_title(dname)
#ax.set_xlabel("Confidence")
#ax.set_ylabel("Density")
#ax.legend()
fig2, ax = plt.subplots(figsize=(1.8, 1.8))
fig2.subplots_adjust(left=0.15, bottom=0.19, right=0.98, top=0.96)
conf = all_conf
sets = all_sets
bottom = np.zeros(x.shape)
for u, s in enumerate(ADV_SETS + ["test"]):
conf_s = conf[sets == s]
hist, bin_edges = np.histogram(conf_s, bins=bins)
hist = hist / len(conf)
kwargs = {"facecolor": cmap(u)}
if s == "test":
kwargs["facecolor"] = "gray"
kde = KernelDensity(kernel='epanechnikov', bandwidth=0.02).fit(conf_s.reshape(-1, 1))
#y = kde.score_samples(x.reshape(-1, 1))
y = np.exp(kde.score_samples(x.reshape(-1, 1)))
#ax.bar(bins[:-1], hist, width=0.05, align="edge", bottom=bottom,
# label=s, hatch=SET_HATCH[s],
# linewidth=0.5, edgecolor="black", **kwargs)
#ax.plot(bins[:-1], hist+bottom, label=s)
ax.fill_between(x, bottom, bottom+y, label=s, edgecolor="black",
linewidth=0.5, hatch=SET_HATCH[s], **kwargs)
#print(f"{s:3}", hist)
bottom += y
ax.set_xlabel("Confidence")
ax.set_ylabel("Density")
ax.set_xlim([0.5, 1])
ax.set_ylim([0.0, ax.get_ylim()[1]])
ax.set_yticklabels([""] * len(ax.get_yticks()))
ax.legend(loc="upper center")
if SAVEFIGS:
figname = f"conf_density_{model_type}"
fig2.savefig(os.path.join("figures", f"{figname}.svg"))
subprocess.run(["/home/laurens/repos/dotfiles/scripts/svg2latex", f"{figname}.svg"], cwd="figures")
shutil.copyfile(os.path.join("figures", f"{figname}.pdf_tex"),
os.path.join(FIG_DIR, f"{figname}.pdf_tex"))
shutil.copyfile(os.path.join("figures", f"{figname}.pdf"),
os.path.join(FIG_DIR, f"{figname}.pdf"))
plt.show()
def plot_confdist2(per_confdelta, model_type): # 5 different subplots
dnames = per_confdelta.keys()
cmap = plt.get_cmap("tab10")
all_conf = np.array([])
all_sets = np.array([])
for k, dname in enumerate(dnames):
pcd = per_confdelta[dname]
conf = pcd["conf"]
sets = pcd["set"]
all_conf = np.hstack((all_conf, conf))
all_sets = np.hstack((all_sets, sets))
bins = np.linspace(0.5, 1.0, 21)
x = np.linspace(0.5, 1.0, 101)
bottom = np.zeros(x.shape)
fig, axs = plt.subplots(1, 5, figsize=(6.2 * 3.0/4.0, 0.9), sharey=True)
fig.subplots_adjust(left=0.09, bottom=0.36, right=0.99, top=0.85, wspace=0.14)
conf = all_conf
sets = all_sets
bottom = np.zeros(x.shape)
for u, (ax, s) in enumerate(zip(axs.ravel(), ["test"] + ADV_SETS)):
conf_s = conf[sets == s]
hist, bin_edges = np.histogram(conf_s, bins=bins)
hist = hist / len(conf)
kwargs = {"facecolor": cmap(u)}
label = s
if s == "test":
label = "normal"
kwargs["facecolor"] = "gray"
kde = KernelDensity(kernel="gaussian", bandwidth=0.02).fit(conf_s.reshape(-1, 1))
#y = kde.score_samples(x.reshape(-1, 1))
y = np.exp(kde.score_samples(x.reshape(-1, 1)))
#ax.bar(bins[:-1], hist, width=0.05, align="edge", bottom=bottom,
# label=s, hatch=SET_HATCH[s],
# linewidth=0.5, edgecolor="black", **kwargs)
#ax.plot(bins[:-1], hist+bottom, label=s)
ax.fill_between(x, bottom, bottom+y, label=label, edgecolor="black",
linewidth=0.5, **kwargs)
#print(f"{s:3}", hist)
#bottom += y
#ax.set_title(s)
ax.set_xlabel("Confidence")
#ax.set_ylabel("Density")
ax.set_xlim([0.5, 1])
ax.set_ylim([0.0, ax.get_ylim()[1]])
ax.set_xticks([0.5, 0.75, 1.0])
ax.set_xticklabels([".5", ".75", "1"])