/
glad_batch.py
168 lines (124 loc) · 6.73 KB
/
glad_batch.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
import logging
import numpy as np
from ..common.utils import Timer, logger, dir_create, configure_logger
from ..common.gen_samples import read_anomaly_dataset
from ..aad.aad_globals import get_first_vals_not_marked
from .afss import get_glad_command_args, GladOpts
from .glad_support import get_afss_model, GLADEnsembleLimeExplainer, set_results_dir, set_random_seeds, \
SequentialResults
from .glad_test_support import plot_afss_scores, plot_weighted_scores, plot_glad_relevance_regions, \
prepare_loda_ensemble
"""
An implementation of:
GLAD: *GL*ocalized *A*nomaly *D*etection via Active Feature Space Suppression
python -m ad_examples.glad.glad_batch --log_file=temp/glad/glad_batch.log --debug --dataset=toy2 --n_epochs=200 --afss_bias_prob=0.50 --train_batch_size=25 --budget=60 --afss_nodes=0 --afss_max_labeled_reps=5 --loda_debug --plot
"""
supported_ensemble_types = ["loda"]
def afss_active_learn_ensemble(x, y, ensemble, opts):
data_2D = x.shape[1] == 2
plot = opts.plot and data_2D
# populate labels as some dummy value (-1) initially
y_labeled = np.ones(x.shape[0], dtype=int) * -1
scores = ensemble.get_scores(x)
xx = yy = None
afss = get_afss_model(opts, n_output=ensemble.m)
afss.init_network(x, prime_network=True)
baseline_scores = afss.get_weighted_scores(x, scores)
baseline_queried = np.argsort(-baseline_scores)
baseline_found = np.cumsum(y[baseline_queried[np.arange(opts.budget)]])
logger.debug("baseline found:\n%s" % (str(list(baseline_found))))
queried = [] # labeled instances
for i in range(opts.budget):
tm = Timer()
a_scores = afss.get_weighted_scores(x, scores)
ordered_indexes = np.argsort(-a_scores)
items = get_first_vals_not_marked(ordered_indexes, queried, start=0, n=1)
queried.extend(items)
hf = np.array(queried, dtype=int)
y_labeled[items] = y[items]
afss.update_afss(x, y_labeled, hf, scores, tau=opts.afss_tau)
if plot and ensemble.m < 5:
xx, yy = plot_afss_scores(x, y, ensemble, afss, selected=x[hf], cmap='jet', xx=xx, yy=yy,
name="_f%d_after" % (i+1), dataset=opts.dataset, outpath=opts.results_dir)
logger.debug(tm.message("finished budget %d:" % (i+1)))
if plot:
xx, yy = plot_weighted_scores(x, y, ensemble, afss, selected=None, xx=xx, yy=yy, contour_levels=20,
name="_feedback_after", n_anoms=opts.n_anoms,
dataset=opts.dataset, outpath=opts.results_dir)
if opts.explain:
# get the next unlabeled instance and try to explain its anomaly score
explainer = GLADEnsembleLimeExplainer(x, y, ensemble, afss, feature_names=["x", "y"])
a_scores = afss.get_weighted_scores(x, scores)
ordered_indexes = np.argsort(-a_scores)
items = get_first_vals_not_marked(ordered_indexes, queried, start=0, n=1)
# why did GLAD assign a high anomaly score to the instance in its current state?
explanation, best_member, member_relevance = explainer.explain(x[items[0]])
if explanation is not None:
logger.debug("\nExplain inst: %d %s (%s); best: %d %s\n%s" %
(items[0], "" if x.shape[1] != 2 else str(x[items[0]]),
"anomaly" if y[items[0]] == 1 else "nominal",
best_member, str(member_relevance), str(explanation.as_list())))
if plot:
max_rank = 1
plot_glad_relevance_regions(x, y, ensemble, afss, selected=x[items],
max_rank=max_rank, name="rel_regions_r%d" % max_rank,
dataset=opts.dataset, outpath=opts.results_dir)
afss.close_session()
# the number of anomalies discovered within the budget while incorporating feedback
# logger.debug("queried:\n%s" % str(queried))
# logger.debug("y_labeled:\n%s" % str(list(y_labeled[queried])))
found = np.cumsum(y[queried])
logger.debug("GLAD found:\n%s" % (str(list(found))))
# make queried indexes 1-indexed
queried = np.array(queried, dtype=int) + 1
baseline_queried = np.array(baseline_queried[0:opts.budget], dtype=int) + 1
results = SequentialResults(num_seen=found, num_seen_baseline=baseline_found,
queried_indexes=queried,
queried_indexes_baseline=baseline_queried)
return results
def glad_active_learn(opts):
set_results_dir(opts)
dir_create(opts.results_dir)
opts.plot = opts.plot and opts.reruns == 1 # just in case...
logger.debug("feedback budget: %d, batch_size: %d, afss_nodes: %s" %
(opts.budget, opts.train_batch_size, opts.afss_nodes))
if opts.ensemble_type not in supported_ensemble_types:
raise ValueError("Unsupported ensemble type '%s'. Supported ensemble types: %s." %
(opts.ensemble_type, str(supported_ensemble_types)))
x, y = read_anomaly_dataset(opts.dataset, datafile=opts.datafile)
logger.debug("dataset: %s, shape: %s" % (opts.dataset, str(x.shape)))
all_results = SequentialResults()
orig_randseed = opts.randseed
ensembles = []
for i in range(opts.reruns):
tm = Timer()
opts.randseed = orig_randseed + i
opts.runidx = i+1
set_random_seeds(opts.randseed, opts.randseed+1, opts.randseed+2)
ensemble = prepare_loda_ensemble(x, mink=opts.loda_mink, maxk=opts.loda_maxk,
debug=opts.loda_debug and x.shape[1] == 2, m=4)
logger.debug("#LODA projections: %d" % ensemble.m)
ensembles.append(ensemble)
if not opts.ensemble_only:
results = afss_active_learn_ensemble(x, y, ensemble, opts)
all_results.merge(results)
logger.debug(tm.message("completed run %d/%d:" % (i+1, opts.reruns)))
if not opts.ensemble_only:
all_results.write_to_csv(opts)
else:
all_results = None
return x, y, all_results, ensembles
if __name__ == "__main__":
logger = logging.getLogger(__name__)
dir_create("./temp/glad")
args = get_glad_command_args(debug=False, debug_args=["--debug",
"--plot",
"--dataset=toy",
"--budget=1",
"--n_anoms=30",
"--log_file=temp/glad/glad_batch.log"])
# print "log file: %s" % args.log_file
configure_logger(args)
opts = GladOpts(args)
logger.debug("running: %s" % opts.str_opts())
glad_active_learn(opts)