-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathpicksamples.py
280 lines (250 loc) · 12.4 KB
/
picksamples.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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
import os
import numpy as np
import predict
import random
class PickSamples():
def __init__(self, exp=0, percent=[0.5, 0.1, 0.1, 0.1, 0.1, 0.1], pace=0, alpha=[0.9, 0.9, 0.87, 0.85, 0.8, 0.8],
ent_threshold=-4.2, diff_threshold=1.5, ent_pick_per=0.1, random_pick=False,
soft=False,root='.', max_step=40000):
self.max_step = max_step
self.root = root
self.exp = exp
self.root_image = os.path.join(self.root, 'images/txt{}'.format(self.exp))
self.root_MAE = os.path.join(self.root, 'MAE/mae{}'.format(self.exp))
self.root_entropy = os.path.join(self.root, 'entropy/E{}'.format(self.exp))
self.checkdir(self.root_image)
self.checkdir(self.root_MAE)
self.checkdir(self.root_entropy)
self.percent = percent
self.pace = pace
self.ent_threshold = ent_threshold
self.diff_threshold = diff_threshold
self.ent_pick_per = ent_pick_per
self.prefix_pick = 'trainPick'
self.prefix_left = 'trainLeft'
self.prefix_MAE = 'MAEOnTrainLeft'
self.prefix_pick_ent = 'pick_ent'
self.prefix_ent = 'entropy'
self.alpha = alpha
self.random_pick = random_pick
self.soft = soft
self.fn_traintxt0 = './images/MORPH-train.txt'
train_images = self.readtxt(self.fn_traintxt0)
self.pace_samples = [int(p*len(train_images)) for p in self.percent]
print('percent: ', self.percent)
print('pace_samples: ', self.pace_samples)
imgs = []
labels = []
for t in train_images:
img, label = t.split(' ')
imgs.append(img)
labels.append(float(label))
self.dict_train = dict(zip(imgs, labels))
def checkdir(self, tmp_dir):
if not os.path.isdir(tmp_dir):
os.makedirs(tmp_dir)
def readtxt(self, fn):
with open(fn, 'r') as f:
lines = f.readlines()
return lines
def savetxt(self, fn, lines):
with open(fn, 'w') as f:
f.writelines(lines)
def get_prefix(self, phase):
if phase == 'pick' :
prefix = self.prefix_pick
root = self.root_image
elif phase == 'left':
prefix = self.prefix_left
root = self.root_image
elif phase == 'pick_ent':
prefix = self.prefix_pick_ent
root = self.root_image
elif phase == 'MAE':
prefix = self.prefix_MAE
root = self.root_MAE
elif phase == 'entropy':
prefix = self.prefix_ent
root = self.root + '/'
else:
raise NameError('illegal phase in get prefix')
return (prefix, root)
def get_fn(self, pace, phase='pick'):
prefix, root = self.get_prefix(phase)
if 'entropy' in prefix:
fn = 'Entropy.txt'
fn = root + fn
else:
fn = prefix + str(int(pace)) + '.txt'
fn = os.path.join(root, fn)
return fn
def predict(self, pace=0, soft=False):
para_dict = {}
exp = str(self.exp)
if soft:
para_dict['predict'] = os.path.join('./MAE/mae{}'.format(exp), 'MAEOnTrainPick{}.txt'.format(pace))
para_dict['test'] = os.path.join('./images/txt{}'.format(exp), 'trainPick{}.txt'.format(pace))
para_dict['model'] = os.path.join('./checkpoints/M{}'.format(exp), '{}VGG_iter_{}.caffemodel'.format(pace-1, self.max_step))
para_dict['deploy'] = os.path.join('./tmp/Exp{}'.format(exp), '{}VGG-deploy.prototxt'.format(pace-1))
if not soft:
para_dict['predict'] = os.path.join('./MAE/mae{}'.format(exp), '{}MAEOnTrainLeft0.txt'.format(pace-1))
para_dict['test'] = os.path.join('./images/txt{}'.format(exp), 'trainLeft0.txt')
para_dict['model'] = os.path.join('./checkpoints/M{}'.format(exp), '{}VGG_iter_{}.caffemodel'.format(pace-1, self.max_step))
para_dict['deploy'] = os.path.join('./tmp/Exp{}'.format(exp), '{}VGG-deploy.prototxt'.format(pace-1))
diff = predict.Predict(para_dict)
def pick(self, pace=0):
'''
pace represent the txt need to be generated
'''
pick,left,pick_soft,left_soft,pred_sort,pick_ent = [],[],[],[],[],[]
if pace == 0:
pick = []
left = self.readtxt(self.fn_traintxt0)
if self.soft:
for line in left:
img, label = line.strip('\n').split(' ')
left_soft.append(img + ' ' + label + ' ' + '10000' + '\n')
else:
fn_train_previous = self.get_fn(pace-1, phase='pick')
fn_predictMAE = self.get_fn(pace-1, phase='MAE')
fn_train_pick = self.get_fn(pace, phase='pick')
fn_train_left = self.get_fn(pace, phase='left')
print('fn_train_previous: %s, Length: %d' % (fn_train_previous, len(self.readtxt(fn_train_previous))))
print('fn_predictMAE: %s, Length: %d' % (fn_predictMAE, len(self.readtxt(fn_predictMAE))))
print('fn_train_pick: %s' % (fn_train_pick))
print('fn_train_left: %s' % (fn_train_left))
pred = self.readtxt(fn_predictMAE)
fn_entropy = os.path.join(self.root_entropy, 'entropy{}.txt'.format(pace-1))
entropy = self.readtxt(fn_entropy)
assert len(entropy) == len(pred), 'entropy do not equal to pred %d vs %d' % (len(entropy), len(pred))
for i, p in enumerate(pred):
diff = float(p.split(':')[-1])
if diff > self.diff_threshold:
diff = self.diff_threshold
img = p.split(':')[1].split(',')[0][1:]
ent = float(entropy[i].split(':')[-1])
if self.ent_threshold < 0:
if ent < self.ent_threshold:
ent = self.ent_threshold
diff = diff - self.alpha[pace-1] * ent
pred_sort.append((img, diff))
pred_sort.sort(key=lambda x:x[1])
print('pred_sort: %d' % len(pred_sort))
for i in range(len(pred_sort)):
p = pred_sort[i][0]
line = p + ' ' + str(self.dict_train[p]) + '\n'
if i < self.pace_samples[pace-1]:
pick.append(line)
else:
left.append(line)
print('pace pick: %d' % len(pick))
tem = self.readtxt(fn_train_previous)
# Curriculum Reconstruction
if self.ent_pick_per > 0:
if self.random_pick:
lines = self.readtxt(self.fn_traintxt0)
random.shuffle(lines)
if self.ent_pick_per < 1:
len_lim = int(len(idx_ent) * self.ent_pick_per)
else:
len_lim = self.ent_pick_per
pick_ent = lines[:len_lim]
else:
ent_all_txt_ = './entropy/E{}/entropyAll{}.txt'.format(self.exp, pace-1)
fn_predictMAE = './MAE/mae{}/{}MAEOnTrainLeft0.txt'.format(self.exp, pace-1)
entropy2, pred2 = [], []
if os.path.exists(ent_all_txt_) and os.path.exists(fn_predictMAE):
entropy2 = self.readtxt(ent_all_txt_)
pred2 = self.readtxt(fn_predictMAE)
else:
self.predict(pace=pace)
entropy2 = self.readtxt('./Entropy.txt')
pred2 = self.readtxt(fn_predictMAE)
self.savetxt(ent_all_txt_, entropy2)
assert len(entropy2) == len(pred2), 'entropy do not equal to pred %d vs %d' % (len(entropy), len(pred))
ent_sort = []
for e in entropy2:
ent = float(e.split(':')[-1])
ent_sort.append(ent)
ent_sort_np = np.array(ent_sort)
idx_ent = np.argsort(-ent_sort_np)
if self.ent_pick_per < 1:
len_lim = int(len(idx_ent) * self.ent_pick_per)
else:
len_lim = self.ent_pick_per
idx_ent_pick = idx_ent[:len_lim]
for i in range(idx_ent_pick.shape[0]):
idx = idx_ent_pick[i]
p = pred2[idx]
img = p.split(':')[1].split(',')[0][1:]
label = str(int(self.dict_train[img]))
line = img + ' ' + label + '\n'
pick_ent.append(line)
l1, l2, l3 = len(pick), len(tem), len(pick_ent)
pick = pick + tem + pick_ent
print('pick{} = pace_pick{} + previous{} + ent_pick{}'.format(l1+l2+l3, l1,l2,l3))
fn_save_pick = self.get_fn(pace, phase='pick')
fn_save_left = self.get_fn(pace, phase='left')
fn_save_pick_ent = self.get_fn(pace, phase='pick_ent')
self.savetxt(fn_save_pick_ent, pick_ent)
self.savetxt(fn_save_pick, pick)
self.savetxt(fn_save_left, left)
# Mixture Weighting
if self.soft and pace > 0:
ent_all_txt_ = './entropy/E{}/entropyAll{}.txt'.format(self.exp, pace-1)
fn_predictMAE = './MAE/mae{}/{}MAEOnTrainLeft0.txt'.format(self.exp, pace-1)
ent_all = self.readtxt(ent_all_txt_)
mae_all = self.readtxt(fn_predictMAE)
print('mae_all: ', len(mae_all))
assert len(ent_all) == len(mae_all), 'entropy do not equal to pred %d vs %d' % (len(ent_all), len(mae_all))
mae_pick, ent_pick, img_all = [], [], []
for mae in mae_all:
img = mae.split(':')[1].split(',')[0][1:]
img_all.append(img)
for p in pick:
img = p.split(' ')[0]
idx = img_all.index(img)
mae_pick.append(mae_all[idx])
ent_pick.append(ent_all[idx])
print('mae_pick: ', len(mae_pick))
assert len(ent_pick) == len(mae_pick), 'entropy do not equal to pred %d vs %d' % (len(ent_pick), len(mae_pick))
for i in range(len(mae_pick)):
diff = float(mae_pick[i].split(':')[-1])
img = mae_pick[i].split(':')[1].split(',')[0][1:]
ent = float(ent_pick[i].split(':')[-1])
if self.ent_threshold < 0:
if ent < self.ent_threshold:
ent = self.ent_threshold
diff = diff - self.alpha[pace-1] * ent
pick_soft.append((img, diff))
pick_soft.sort(key=lambda x:x[1])
num_pick = len(pick_soft)
lambda_0 = pick_soft[-1][1]
lambda_1 = pick_soft[int(num_pick*0.9)][1]
epsilon = 1 / (1/lambda_1 - 1/lambda_0)
pick_new = []
for i, (img, diff) in enumerate(pick_soft):
label = str(int(self.dict_train[img]))
if i < num_pick*0.9:
weight = 10000
else:
weight = int(10000*(epsilon / diff - epsilon / lambda_0))
pick_new.append(img + ' ' + label + ' ' + str(weight) + '\n')
pick_soft = pick_new
if self.soft:
fn_pick_soft = './images/txt{}/trainPick_soft{}.txt'.format(self.exp, pace)
self.savetxt(fn_pick_soft, pick_soft)
print('new pick: %d' % len(pick_soft))
print('entropy pick %d' % len(pick_ent))
print('new left: %d' % len(left))
if pace == 0:
fn_left_soft = './images/txt{}/trainLeft_soft{}.txt'.format(self.exp, pace)
self.savetxt(fn_left_soft, left_soft)
return (fn_left_soft, fn_pick_soft)
else:
return (fn_save_left, fn_pick_soft)
else:
print('new pick: %d' % len(pick))
print('entropy pick %d' % len(pick_ent))
print('new left: %d' % len(left))
return (fn_save_left, fn_save_pick)