-
Notifications
You must be signed in to change notification settings - Fork 3
/
lfw_eval.py
131 lines (106 loc) · 4.62 KB
/
lfw_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
from __future__ import print_function
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
torch.backends.cudnn.bencmark = True
import os,sys,cv2,random,datetime
import argparse
import numpy as np
import zipfile
#from dataset import ImageDataset
from matlab_cp2tform import get_similarity_transform_for_cv2
import CNN_R
def alignment(src_img,src_pts):
ref_pts = [ [30.2946, 51.6963],[65.5318, 51.5014],
[48.0252, 71.7366],[33.5493, 92.3655],[62.7299, 92.2041] ]
crop_size = (96, 112)
src_pts = np.array(src_pts).reshape(5,2)
s = np.array(src_pts).astype(np.float32)
r = np.array(ref_pts).astype(np.float32)
tfm = get_similarity_transform_for_cv2(s, r)
face_img = cv2.warpAffine(src_img, tfm, crop_size)
return face_img
# 10 folds. 5400 imgs for training, 600 imgs for testing
def KFold(n=6000, n_folds=10, shuffle=False):
folds = []
base = list(range(n))
for i in range(n_folds):
test = base[int(i*n/n_folds):int((i+1)*n/n_folds)]
train = list(set(base)-set(test))
folds.append([train,test])
return folds
def eval_acc(threshold, diff):
y_true = []
y_predict = []
for d in diff:
same = 1 if float(d[2]) > threshold else 0
y_predict.append(same)
y_true.append(int(d[3]))
y_true = np.array(y_true)
y_predict = np.array(y_predict)
accuracy = 1.0*np.count_nonzero(y_true==y_predict)/len(y_true)
return accuracy
def find_best_threshold(thresholds, predicts):
best_threshold = best_acc = 0
for threshold in thresholds:
accuracy = eval_acc(threshold, predicts)
if accuracy >= best_acc:
best_acc = accuracy
best_threshold = threshold
return best_threshold
parser = argparse.ArgumentParser(description='PyTorch sphereface lfw')
parser.add_argument('--net','-n', default='sphere64a', type=str)
parser.add_argument('--lfw', default='../datasets/lfw.zip', type=str)
parser.add_argument('--model','-m', default='./weights/CNN_R_sphere64a_14.pth', type=str)
args = parser.parse_args()
predicts=[]
net = getattr(CNN_R,args.net)()
net.load_state_dict(torch.load(args.model))
net.cuda()
net.eval()
net.feature = True # produce features instead similarities
zfile = zipfile.ZipFile(args.lfw)
landmark = {}
with open('../datasets/lfw_landmark.txt') as f:
landmark_lines = f.readlines()
for line in landmark_lines:
l = line.replace('\n','').split('\t')
landmark[l[0]] = [int(k) for k in l[1:]]
with open('../datasets/pairs.txt') as f:
pairs_lines = f.readlines()[1:]
for i in range(6000):
p = pairs_lines[i].replace('\n','').split('\t')
if 3==len(p): # the two imgs of the same person
sameflag = 1
name1 = p[0]+'/'+p[0]+'_'+'{:04}.jpg'.format(int(p[1]))
name2 = p[0]+'/'+p[0]+'_'+'{:04}.jpg'.format(int(p[2]))
if 4==len(p): # the two imggs of two diffrent person
sameflag = 0
name1 = p[0]+'/'+p[0]+'_'+'{:04}.jpg'.format(int(p[1]))
name2 = p[2]+'/'+p[2]+'_'+'{:04}.jpg'.format(int(p[3]))
img1 = alignment(cv2.imdecode(np.frombuffer(zfile.read('lfw/'+name1),np.uint8),1),landmark[name1])
img2 = alignment(cv2.imdecode(np.frombuffer(zfile.read('lfw/'+name2),np.uint8),1),landmark[name2])
imglist = [img1,cv2.flip(img1,1),img2,cv2.flip(img2,1)] # flip horizontally
for i in range(len(imglist)):
imglist[i] = imglist[i].transpose(2, 0, 1).reshape((1,3,112,96)) # BGR2RGB
imglist[i] = (imglist[i]-127.5)/128.0
img = np.vstack(imglist)
img = Variable(torch.from_numpy(img).float(),volatile=True).cuda()
output = net(img)
f = output.data
f1,f2 = f[0],f[2] # original features. but should be the concatenating of ori and aug features?
cosdistance = f1.dot(f2)/(f1.norm()*f2.norm()+1e-5)
predicts.append('{}\t{}\t{}\t{}\n'.format(name1,name2,cosdistance,sameflag))
accuracy = []
thd = []
folds = KFold(n=6000, n_folds=10, shuffle=False)
thresholds = np.arange(-1.0, 1.0, 0.005) # find the best thresholds
predicts = np.array(list(map(lambda line:line.strip('\n').split(), predicts)))
for idx, (train, test) in enumerate(folds):
best_thresh = find_best_threshold(thresholds, predicts[train]) # 5400 imgs to find the best thresh
accuracy.append(eval_acc(best_thresh, predicts[test])) # 600 imgs to test the performance
thd.append(best_thresh) # 10 different best thresh for training sets(5400 imgs)
print('Fold {:d}: Acc:{:.4f}'.format(idx+1, accuracy[idx]))
print('Model:{:s} LFWACC={:.4f} std={:.4f} thd={:.4f}'.format(args.model, np.mean(accuracy), np.std(accuracy), np.mean(thd)))