-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy patheval.py
203 lines (177 loc) · 6.58 KB
/
eval.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
import sys
import os
import os.path as osp
import numpy as np
import torch
import torch.nn.functional as F
import cv2
from tqdm import tqdm
from sklearn import metrics
from base_container import BaseContainer
class NetworkTester(BaseContainer):
def __init__(self):
super().__init__()
self.init_evaluation_container()
self.loaders = []
self.v2f = []
self.test_set = self.Dataset_val(self.args.evaluation.dataset, split='test', sample_mode='id')
self.num_id = len(self.test_set)
def gen_feat(self, save_path):
print('Generating features...')
self.model.eval()
name2feat = dict()
name2id = dict()
name2gender = dict()
for i in tqdm(range(self.num_id)):
sample_img, sample_aud = self.test_set.getitem_all(i)
imgs, img_filenames, img_gender = array2tensor_img(sample_img)
auds, aud_filenames, aud_gender = array2tensor_aud(sample_aud)
with torch.no_grad():
imgs = self.model.normalize(imgs.cuda())
auds = auds.cuda()
img_feats = self.model.F(imgs)
aud_feats = self.model.V(auds)
if 'wang' in save_path:
img_feats = self.model.F.shared_fc(img_feats)
aud_feats = self.model.F.shared_fc(aud_feats)
for j in range(len(img_feats)):
name2feat[img_filenames[j]] = img_feats[j]
name2id[img_filenames[j]] = i
name2gender[img_filenames[j]] = img_gender[j]
for j in range(len(aud_feats)):
name2feat[aud_filenames[j]] = aud_feats[j]
name2id[aud_filenames[j]] = i
name2gender[aud_filenames[j]] = aud_gender[j]
save = {
'name2feat': name2feat,
'name2id': name2id,
'name2gender': name2gender
}
torch.save(save, save_path)
print('Done.')
def load_feat(self):
save_path = self.args.evaluation.save_feat
# if not osp.exists(save_path):
self.gen_feat(save_path)
res = torch.load(save_path)
self.name2feat = res['name2feat']
self.name2id = res['name2id']
self.name2gender = res['name2gender']
def eval(self):
args = self.args.evaluation.dataset
self.load_feat()
if not isinstance(args.eval_triplet_test_all, list):
with open(osp.join(args.list_dir, args.eval_triplet_test_all + '.txt'), 'r') as f:
args.eval_triplet_test_all = f.read().splitlines()
all_res = []
for l in args.eval_triplet_test_all:
with open(osp.join(args.list_dir, l + '.txt'), 'r') as f:
eval_list = f.read().splitlines()
eval_list = [i.split(' ') for i in eval_list]
if l.startswith('match'):
res = self.eval_match(eval_list)
elif l.startswith('verify'):
res = self.eval_verify(eval_list)
elif l.startswith('reterival'):
res = self.eval_retrival(eval_list)
print('%s: %.6f'%(l, res))
all_res.append(res)
with open(self.args.evaluation.save_result, 'w') as f:
for i in range(len(all_res)):
f.write(args.eval_triplet_test_all[i] + ' ' + str(all_res[i]) + '\n')
def eval_match(self, eval_list):
N = len(eval_list[0]) - 1
cnt_pos = 0
cnt_neg = 0
for i in tqdm(range(len(eval_list))):
cand = []
for j in range(N):
cand.append(self.name2feat[eval_list[i][j+1]])
cand = torch.stack(cand)
x = self.name2feat[eval_list[i][0]]
score = cosine_distance(x.unsqueeze(0), cand)
ID = self.name2id[eval_list[i][0]]
if score.argmax() == 0:
cnt_pos += 1
else:
cnt_neg += 1
return cnt_pos / (cnt_pos + cnt_neg)
def eval_verify(self, eval_list):
pred = []
target = []
for i in tqdm(eval_list):
x = self.name2feat[i[0]]
cand = self.name2feat[i[1]]
score = cosine_distance(x, cand)
pred.append(float(score))
target.append(int(self.name2id[i[0]] == self.name2id[i[1]]))
fpr, tpr, thresholds = metrics.roc_curve(target, pred, pos_label=1)
return metrics.auc(fpr, tpr)
def eval_retrival(self, eval_list):
assert len(eval_list) == 2
mAP = []
x = torch.stack([self.name2feat[i] for i in eval_list[0]], 0)
cand = torch.stack([self.name2feat[i] for i in eval_list[1]], 0)
x_id = torch.tensor([self.name2id[i] for i in eval_list[0]])
cand_id = torch.tensor([self.name2id[i] for i in eval_list[1]])
for i in tqdm(range(len(x))):
score = cosine_distance(x[i].unsqueeze(0), cand)
r = torch.argsort(-score)
ID = cand_id[r]
ap = compute_ap((ID == x_id[i]))
mAP.append(ap)
return np.array(mAP).mean()
def array2tensor_img(sample):
imgs = []
filenames = []
gender = []
for i in range(len(sample)):
imgs.append(sample[i]['image'])
filenames.append(sample[i]['image_filename'])
gender.append(sample[i]['gender'])
imgs = torch.stack(imgs, 0)
return imgs, filenames, gender
def array2tensor_aud(sample):
auds = []
filenames = []
gender = []
for i in range(len(sample)):
auds.append(sample[i]['audio'])
filenames.append(sample[i]['audio_filename'])
gender.append(sample[i]['gender'])
auds = torch.stack(auds, 0)
return auds, filenames, gender
def Euclidean_distance(x, y):
return ((x - y)**2).sum(-1)
def cosine_distance(x, y):
return (F.normalize(x, dim=-1) * F.normalize(y, dim=-1)).sum(-1)
def compute_ap(label):
old_recall = 0.0
ap = 0.0
intersect_size = 0.0
nz = label.nonzero().squeeze(1)
n_true = len(nz)
for i in nz:
i = int(i)
intersect_size += 1
recall = intersect_size / n_true
precision = intersect_size / (i+1)
ap += (recall - old_recall) * precision
old_recall = recall
return ap
def to_cuda(sample):
if isinstance(sample, list):
return [to_cuda(i) for i in sample]
elif isinstance(sample, dict):
for key in sample.keys():
sample[key] = to_cuda(sample[key])
return sample
elif isinstance(sample, torch.Tensor):
return sample.cuda()
else:
return sample
def main():
tester = NetworkTester()
tester.eval()
if __name__ == "__main__":
main()