-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluations.py
170 lines (133 loc) · 6.11 KB
/
evaluations.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
"""
This script was modified from https://github.com/ZhaoJ9014/face.evoLVe.PyTorch
"""
import os
import cv2
import bcolz
import numpy as np
import tqdm
from sklearn.model_selection import KFold
from utils import l2_norm
from scipy import spatial
from numpy import dot
from numpy.linalg import norm
from sklearn.metrics.pairwise import cosine_similarity
def get_val_pair(path, name):
carray = bcolz.carray(rootdir=os.path.join(path, name), mode='r')
issame = np.load('{}/{}_list.npy'.format(path, name))
return carray, issame
def get_val_data(data_path):
"""get validation data"""
lfw, lfw_issame = get_val_pair(data_path, 'lfw_align_112/lfw')
agedb_30, agedb_30_issame = get_val_pair(data_path, 'agedb_align_112/agedb_30')
cfp_fp, cfp_fp_issame = get_val_pair(data_path, 'cfp_align_112/cfp_fp')
return lfw, agedb_30, cfp_fp, lfw_issame, agedb_30_issame, cfp_fp_issame
def ccrop_batch(imgs):
assert len(imgs.shape) == 4
resized_imgs = np.array([cv2.resize(img, (128, 128)) for img in imgs])
ccropped_imgs = resized_imgs[:, 8:-8, 8:-8, :]
return ccropped_imgs
def hflip_batch(imgs):
assert len(imgs.shape) == 4
return imgs[:, :, ::-1, :]
def calculate_accuracy(threshold, dist, actual_issame):
predict_issame = np.less(dist, threshold)
tp = np.sum(np.logical_and(predict_issame, actual_issame))
fp = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
tn = np.sum(np.logical_and(np.logical_not(predict_issame),
np.logical_not(actual_issame)))
fn = np.sum(np.logical_and(np.logical_not(predict_issame), actual_issame))
tpr = 0 if (tp + fn == 0) else float(tp) / float(tp + fn)
fpr = 0 if (fp + tn == 0) else float(fp) / float(fp + tn)
acc = float(tp + tn) / dist.size
return tpr, fpr, acc
def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame,
nrof_folds=10):
assert (embeddings1.shape[0] == embeddings2.shape[0])
assert (embeddings1.shape[1] == embeddings2.shape[1])
nrof_pairs = min(len(actual_issame), embeddings1.shape[0])
nrof_thresholds = len(thresholds)
k_fold = KFold(n_splits=nrof_folds, shuffle=False)
tprs = np.zeros((nrof_folds, nrof_thresholds))
fprs = np.zeros((nrof_folds, nrof_thresholds))
accuracy = np.zeros((nrof_folds))
best_thresholds = np.zeros((nrof_folds))
indices = np.arange(nrof_pairs)
best_tprs = np.zeros(nrof_folds)
best_fprs = np.zeros(nrof_folds)
# Euclidean Distance
# diff = np.subtract(embeddings1, embeddings2)
# dist = np.sum(np.square(diff), 1)
# Cosine Similarity
# diff = dot(embeddings1, embeddings2.T)/(norm(embeddings1)*norm(embeddings2))
diff = dot(embeddings1, embeddings2.T)
dist = 1 - np.diag(diff)
# dist = 1/np.diag(diff)
for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
# Find the best threshold for the fold
acc_train = np.zeros((nrof_thresholds))
for threshold_idx, threshold in enumerate(thresholds):
_, _, acc_train[threshold_idx] = calculate_accuracy(
threshold, dist[train_set], actual_issame[train_set])
best_threshold_index = np.argmax(acc_train)
best_thresholds[fold_idx] = thresholds[best_threshold_index]
for threshold_idx, threshold in enumerate(thresholds):
tprs[fold_idx, threshold_idx], fprs[fold_idx, threshold_idx], _ = \
calculate_accuracy(threshold,
dist[test_set],
actual_issame[test_set])
_, _, accuracy[fold_idx] = calculate_accuracy(
thresholds[best_threshold_index],
dist[test_set],
actual_issame[test_set])
## tpr and fpr for best thresholds
best_tprs[fold_idx], best_fprs[fold_idx], _ = calculate_accuracy(
thresholds[best_threshold_index],
dist[test_set],
actual_issame[test_set])
tpr = np.mean(tprs, 0)
fpr = np.mean(fprs, 0)
# print("tpr: ", tpr)
# print("fpr: ", fpr)
# print("best_tpr: ", best_tprs)
# print("best_fpr: ", best_fprs)
# return tpr, fpr, accuracy, best_thresholds
return best_tprs, best_fprs, tpr, fpr, accuracy, best_thresholds
def evaluate(embeddings, actual_issame, nrof_folds=10):
# Calculate evaluation metrics
# thresholds = np.arange(0, 3, 0.005)
thresholds = np.arange(0, 4, 0.01)
embeddings1 = embeddings[0::2]
embeddings2 = embeddings[1::2]
best_tprs, best_fprs, tpr, fpr, accuracy, best_thresholds = calculate_roc(
thresholds, embeddings1, embeddings2, np.asarray(actual_issame),
nrof_folds=nrof_folds)
# return tpr, fpr, accuracy, best_thresholds
return best_tprs, best_fprs, tpr, fpr, accuracy, best_thresholds
def perform_val(embedding_size, batch_size, model,
carray, issame, nrof_folds=10, is_ccrop=False, is_flip=True):
"""perform val"""
embeddings = np.zeros([len(carray), embedding_size])
for idx in tqdm.tqdm(range(0, len(carray), batch_size)):
batch = carray[idx:idx + batch_size]
# print("batch_imgs: ", batch)
batch = np.transpose(batch, [0, 2, 3, 1]) * 0.5 + 0.5
# print("batch_imgs2: ", batch)
if is_ccrop:
batch = ccrop_batch(batch)
if is_flip:
fliped = hflip_batch(batch)
# print("output_batch: ", model(batch))
# print("flipped_batch: ", model(fliped))
emb_batch = model(batch) + model(fliped)
# print("emb_batch: ", emb_batch)
embeddings[idx:idx + batch_size] = l2_norm(emb_batch)
# print("embeddings: ", l2_norm(emb_batch))
else:
batch = ccrop_batch(batch)
emb_batch = model(batch)
embeddings[idx:idx + batch_size] = l2_norm(emb_batch)
best_tprs, best_fprs, tpr, fpr, accuracy, best_thresholds = evaluate(
embeddings, issame, nrof_folds)
# return accuracy.mean(), best_thresholds.mean()
return best_tprs.mean(), best_fprs.mean(), tpr.mean(), fpr.mean(), accuracy.mean(), best_thresholds.mean()