-
Notifications
You must be signed in to change notification settings - Fork 5
/
test_preds.py
125 lines (112 loc) · 4.1 KB
/
test_preds.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
"""
Test the angular errors of predictions by (plain or weighted) median pooling of local estimations.
The local estimations is produced by running
python solver.py --test_only --gs-test-set 0
python solver.py --test_only --gs-test-set 1
python solver.py --test_only --gs-test-set 2
"""
import numpy as np
import math
import os.path as osp
from argparse import ArgumentParser
from gs568_dataset import GS568Dataset
RAD2DEG = 180. / math.pi
def compute_angular_error(pred, target):
# pred = [r/g, b/g], target = [r, g, b]
assert pred.shape == (2,) and target.shape == (3,)
ip = pred[0] * target[0] + 1. * target[1] + pred[1] * target[2]
norm = np.sqrt((pred[0]*pred[0] + pred[1]*pred[1] + 1.) * \
(target[0]*target[0] + target[1]*target[1] + target[2]*target[2]))
return math.acos(ip/norm) * RAD2DEG
def get_median(preds):
return np.median(preds, axis=0)
def get_confidence_weights(img, locs):
weights = np.zeros(locs.shape[0], dtype=np.float32)
minchn = np.min(img, axis=-1)
for i in range(locs.shape[0]):
weights[i] = np.mean(minchn[locs[i][0]:locs[i][1], locs[i][2]:locs[i][3]])
return weights / np.max(weights)
def get_weighted_median(preds, weights):
n, c = preds.shape
weighted_median = np.zeros(c)
for i in range(c):
temp = zip(preds[:,i], weights)
sorted_pw = sorted(temp, key=lambda x: x[0])
sorted_p, sorted_w = zip(*sorted_pw)
cum_w = np.cumsum(sorted_w)
med_w = cum_w[-1] *0.5
index = np.searchsorted(cum_w, med_w)
weighted_median[i] = (sorted_p[index-1] * (cum_w[index] - med_w) \
+ sorted_p[index] * (med_w - cum_w[index-1])) / sorted_w[index]
return weighted_median
def get_selected_preds(hyp_preds, logits):
n_preds = hyp_preds.shape[0]
sel_idx = np.argmax(logits, axis=1)
return hyp_preds[range(n_preds), sel_idx]
def test_split(args, test_set_id):
args['gs_test_set'] = test_set_id
ds = GS568Dataset('test', args, to_uv=False)
n_imgs = ds.size()
ae_list = list()
for i in range(n_imgs):
img, illum = ds.get(np.array(i))
name = ds.get_name(i)
data = np.load(osp.join(args['pred_dir'], name+'.npz'))
hyp_preds = data['preds']
logits = data['logits']
preds = get_selected_preds(hyp_preds, logits)
if args['weighted_median']:
weights = get_confidence_weights(img, data['locs'])
pred = get_weighted_median(preds, weights)
else:
pred = get_median(preds)
ae = compute_angular_error(pred, illum)
print('%s %.4f' % (name, ae),
pred, illum / illum[1])
ae_list.append(ae)
return ae_list
def print_errors(ae_all):
e_avg = np.mean(ae_all)
e_med = np.median(ae_all)
e_min = np.min(ae_all)
e_max = np.max(ae_all)
e_tri = (np.percentile(ae_all, 25) + np.percentile(ae_all, 75) + e_med * 2) / 4
e_p95 = np.percentile(ae_all, 95)
p25 = np.percentile(ae_all, 25)
p75 = np.percentile(ae_all, 75)
e_l25 = np.mean(ae_all[ae_all <= p25])
e_h25 = np.mean(ae_all[ae_all >= p75])
err_all = {'avg':e_avg,
'med':e_med,
'tri':e_tri,
'p95':e_p95,
'min':e_min,
'max':e_max,
'l25':e_l25,
'h25':e_h25}
for (k,v) in err_all.items():
print(k, ":", v)
return err_all
def main():
ps = get_parser()
args = vars(ps.parse_args())
#
test_set_id = args['test_set_id']
ae_list = list()
if test_set_id in [0, 1, 2]:
ae_list += test_split(args, test_set_id)
else:
for i in range(3):
ae_list += test_split(args, i)
ae_arr = np.array(ae_list)
print_errors(ae_arr)
def get_parser(ps=None):
if ps is None: ps = ArgumentParser()
ps = GS568Dataset.get_parser(ps)
g = ps.add_argument_group('test_preds')
g.add_argument('--pred-dir', type=str, default='preds/')
g.add_argument('--test-set-id', type=int, default=-1)
g.add_argument('--weighted-median', action='store_true', default=False)
return ps
if __name__ == '__main__':
main()