-
Notifications
You must be signed in to change notification settings - Fork 10
/
test_ensemble_probs.py
144 lines (121 loc) · 4.74 KB
/
test_ensemble_probs.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
import argparse
import cPickle as pickle
import collections
import numpy as np
label_names = ['trigger_chans', 'trigger_funcs', 'action_chans',
'action_funcs', 'tc+ac', 'tc+tf+ac+af']
def compute_accuracy_from_probs(probs, labels):
pred_labels = np.vstack([np.argmax(arr, axis=1) for arr in probs]).T
correct = pred_labels == labels
correct = np.concatenate(
(correct, np.all(correct[:, [0, 2]], axis=1)[:, np.newaxis],
np.all(correct[:, [1, 3]], axis=1)[:, np.newaxis]),
axis=1)
for i, name in enumerate(label_names):
print name, np.sum(correct[:, i]) / float(len(correct[:, i]))
print
def compute_F1(args, probs_trigger, probs_action, labels, raw_data):
precision = 0
recall = 0
tot_precision = 0
tot_recall = 0
input_data = pickle.load(open(args.data))
labelers = input_data['labelers']
train_params = input_data['train_params']
param_num = input_data['param_num']
for i in xrange(len(labels)):
pred_trigger = np.argmax(probs_trigger[i])
pred_trigger_name = labelers['trigger_funcs'].classes_[pred_trigger]
pred_trigger_channel= pred_trigger_name[:pred_trigger_name.find('.')]
trigger_name = labelers['trigger_funcs'].classes_[labels[i][1]]
trigger_channel = trigger_name[:trigger_name.find('.')]
pred_action = np.argmax(probs_action[i])
pred_action_name = labelers['action_funcs'].classes_[pred_action]
pred_action_channel= pred_action_name[:pred_action_name.find('.')]
action_name = labelers['action_funcs'].classes_[labels[i][3]]
action_channel = action_name[:action_name.find('.')]
tot_precision += 1
tot_recall += 1
precision += 1
recall += 1
tot_precision += 1
tot_recall += 1
if pred_trigger_channel == trigger_channel:
precision += 1
recall += 1
tot_precision += 1
tot_recall += 1
if pred_trigger_name == trigger_name:
precision += 1
recall += 1
tot_precision += param_num['trigger'+'/'+pred_trigger_name] * 2
tot_recall += len(raw_data[i]['params'][0]) * 2
if pred_trigger_name == trigger_name:
precision += raw_data[i]['correct_trigger_param'] * 2 + raw_data[i]['semi_correct_trigger_param']
recall += raw_data[i]['correct_trigger_param'] * 2 + raw_data[i]['semi_correct_trigger_param']
tot_precision += 1
tot_recall += 1
precision += 1
recall += 1
tot_precision += 1
tot_recall += 1
if pred_action_channel == action_channel:
precision += 1
recall += 1
tot_precision += 1
tot_recall += 1
if pred_action_name == action_name:
precision += 1
recall += 1
tot_precision += param_num['action'+'/'+pred_action_name] * 2
tot_recall += len(raw_data[i]['params'][1]) * 2
if pred_action_name == action_name:
precision += raw_data[i]['correct_action_param'] * 2 + raw_data[i]['semi_correct_action_param']
recall += raw_data[i]['correct_action_param'] * 2 + raw_data[i]['semi_correct_action_param']
precision = precision * 1.0 / tot_precision
recall = recall * 1.0 / tot_recall
print 'F1 score: ', 2 * precision * recall / (precision + recall)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--res', nargs='+')
parser.add_argument('--data')
parser.add_argument('--F1', action='store_true')
args = parser.parse_args()
# section -> label type -> list?
all_probs = collections.defaultdict(lambda: [[] for i in xrange(4)])
input_data = pickle.load(open(args.data))
test_data = input_data['test']
for path in args.res:
data = pickle.load(open(path))
print path
print '================='
for sect in data['probs']:
probs = [np.array(ls) for ls in data['probs'][sect]]
for label_type, prob in enumerate(probs):
all_probs[sect][label_type].append(prob)
labels = np.array(data['labels'][sect]).T
print sect
print '--------------------'
compute_accuracy_from_probs(probs, labels)
if args.F1:
raw_data = []
for item in test_data:
if sect == 'test' or (('tags' in item) and (sect[sect.find('-')+1:] in item['tags'])):
raw_data.append(item)
compute_F1(args, probs[1], probs[3], labels, raw_data)
print 'averaged'
print '=========================='
# import IPython
# IPython.embed()
for sect in all_probs:
probs = [np.mean(arrs, axis=0) for arrs in all_probs[sect]]
labels = np.array(data['labels'][sect]).T
print sect
print '--------------------'
compute_accuracy_from_probs(probs, labels)
if args.F1:
raw_data = []
for item in test_data:
if sect == 'test' or (('tags' in item) and (sect[sect.find('-')+1:] in item['tags'])):
raw_data.append(item)
compute_F1(args, probs[1], probs[3], labels, raw_data)