-
Notifications
You must be signed in to change notification settings - Fork 1
/
triplets_generator.py
223 lines (184 loc) · 8.93 KB
/
triplets_generator.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
import os
import numpy as np
import random
from keras.preprocessing import image
from keras.applications.resnet50 import preprocess_input
import sqlite3
import json, dbutils
#TODO
#ukladani do db pro vyhledavani
imgs_per_class=10000
def get_subdirectories(a_dir):
return [name for name in os.listdir(a_dir)
if os.path.isdir(os.path.join(a_dir, name))]
def img_to_np(path,x=224,y=224):
nparray = image.load_img(path,target_size=(x,y))
nparray = image.img_to_array(nparray)
nparray = np.expand_dims(nparray, axis=0)
nparray = preprocess_input(nparray)
nparray = np.squeeze(nparray)
return nparray
class DataGenerator(object):
'Generates data for Keras'
def __init__(self,model,graph, dim_x = 224, dim_y = 224, batch_size = 10, dataset_path = './places365-dataset/20_classes'):
'Initialization'
self.dim_x = dim_x
self.dim_y = dim_y
self.batch_size = batch_size
self.dataset_path = dataset_path
self.model=model
self.graph=graph
self.conn=sqlite3.connect('representations.db')
dbutils.create()
def generate(self):
'Generates batches of samples'
# Infinite loop
while 1:
# Generate order of exploration of dataset
#indexes = self.__get_exploration_order(list_IDs)
image_IDs = self.__make_triplets()
# Generate batches
imax = int(len(image_IDs)/self.batch_size)
for i in range(imax):
# Find list of IDs
#list_IDs_temp = [list_IDs[k] for k in indexes[i*self.batch_size:(i+1)*self.batch_size]]
image_IDS_temp = image_IDs[i*self.batch_size:(i+1)*self.batch_size]
# Generate data
#X, y = self.__data_generation(labels, list_IDs_temp)
X = self.__data_generation(image_IDS_temp)
#y_anch = np.ones((self.batch_size, 1)) # not used by triple loss function
#y_pos = np.ones((self.batch_size, 1)) # not used by triple loss function
#y_neg = np.ones((self.batch_size, 1)) # not used by triple loss function
y_stacked = np.ones((self.batch_size,2, 1)) # not used by triple loss function
yield X,y_stacked#,y_anch,y_pos,y_neg]
def __get_subdirectories(self, a_dir):
return [name for name in os.listdir(a_dir)
if os.path.isdir(os.path.join(a_dir, name))]
def _img_to_np(self,path):
nparray = image.load_img(os.path.join(self.dataset_path, path ),
target_size=(self.dim_y, self.dim_x))
nparray = image.img_to_array(nparray)
nparray = np.expand_dims(nparray, axis=0)
nparray = preprocess_input(nparray)
nparray = np.squeeze(nparray)
return nparray
def __make_triplets(self):
np.random.seed(0)
classes = self.__get_subdirectories(self.dataset_path)
all_triplets = []
for Id, c in enumerate(classes):
print ("generating triplets for %s (%s/%s)"%(c,Id+1,len(classes)))
pos_dir = os.path.join(self.dataset_path, c)
imgs_pos = os.listdir(pos_dir)
class_triplets = []
anchor_batch = []
positive_batch = []
negative_batch = []
anchors=[]
positives=[]
negatives=[]
for idx in range(0,len(imgs_pos),2):
if not idx%500:
print ("%s/%s"%(idx,min(len(imgs_pos),imgs_per_class)))
if idx>=imgs_per_class:
break
anchor=img_to_np(os.path.join(self.dataset_path, c + '/' + imgs_pos[idx]))
try:
positive=img_to_np(os.path.join(self.dataset_path, c + '/' + imgs_pos[idx+1]))
except IndexError:
idx=-1
positive=img_to_np(os.path.join(self.dataset_path, c + '/' + imgs_pos[idx+1]))
rand_class = random.choice([x for x in classes if x != c]) # choose a different class randomly
#print ("Positive: %s"%c)
#print ("rand_class: %s"%rand_class)
neg = random.choice(os.listdir(os.path.join(self.dataset_path, rand_class)))
negative1 = img_to_np(os.path.join(self.dataset_path, rand_class + '/' + neg))
#rand_class = random.choice([x for x in classes if x != c]) # choose a different class randomly
#neg = random.choice(os.listdir(os.path.join(self.dataset_path, rand_class)))
#negative2 = self._img_to_np(rand_class + '/' + neg)
anchor_batch.append(anchor)
anchors.append(imgs_pos[idx])
positive_batch.append(positive)
positives.append(imgs_pos[idx+1])
negative_batch.append(negative1)
negatives.append(rand_class + '/'+ neg)
#negative_batch.append(negative2)
print ("Calculating representations for %s"%c)
with self.graph.as_default():
preds = self.model.predict(
[np.asarray(anchor_batch), np.asarray(positive_batch), np.asarray(negative_batch)])
# print (preds)
# print (preds.shape)
#preds_pos = np.concatenate(np.asarray(preds[1]),np.asarray(preds[2]))
preds_anch = np.asarray(preds[0])
preds_pos = np.asarray(preds[1])
preds_neg = np.asarray(preds[2])
#print (preds_pos)
#print (preds_neg)
#print (preds_anch.shape)
#print (preds_pos.shape)
#print (preds_neg.shape)
print ("Calculating distances for %s"%c)
#anch_pairs=[("%s/%s"%(c,matrix[0]),matrix[1]) for matrix in zip(imgs_pos,preds_anch)]
#pos_pairs=[("%s/%s"%(c,matrix[0]),matrix[1]) for matrix in zip(positives,preds_pos)]
#print (anch_pairs)
conn = sqlite3.connect('representations.db')
cur = conn.cursor()
for name, matrix in zip(anchors,preds_anch):
name="%s/%s"%(c,name)
cur.execute("REPLACE INTO images VALUES (?,?,?)", (None,name, json.dumps(matrix.tolist())))
for name, matrix in zip(positives,preds_pos):
name="%s/%s"%(c,name)
cur.execute("REPLACE INTO images VALUES (?,?,?)", (None,name, json.dumps(matrix.tolist())))
conn.commit()
for i,anch in enumerate(preds_anch):
least_sim_pos_idx,most_sim_neg_idx=self._least_similar(preds_anch[i],preds_pos,preds_neg)
""" Takes two images from the same class/folder and one image from the next class/folder """
#print(imgs_pos[i])
class_triplets.append([c+'/'+imgs_pos[i], c+'/'+positives[least_sim_pos_idx], negatives[most_sim_neg_idx]])
all_triplets += class_triplets
triplets = np.array(all_triplets)
np.save("triplets.npy",triplets)
np.random.shuffle(triplets)
#print (triplets)
#print (triplets.shape)
return triplets
# get the least similar image from the same class
def _least_similar(self, anch, preds_pos, preds_neg):
def euclidean_distance(x,y):
return np.linalg.norm(x-y)
least_sim_pos=preds_pos[0]
least_sim_pos_idx=0
least_sim_dist=euclidean_distance(anch,least_sim_pos)
for i,candidate in enumerate(preds_pos):
if euclidean_distance(anch,candidate)>least_sim_dist:
least_sim_pos = candidate
least_sim_pos_idx=i
least_sim_dist = euclidean_distance(anch, least_sim_pos)
most_sim_neg = preds_pos[0]
most_sim_neg_idx = 0
most_dist = euclidean_distance(anch, most_sim_neg)
for i,candidate in enumerate(preds_neg):
if euclidean_distance(anch, candidate) < most_dist:
most_sim_neg = candidate
most_sim_neg_idx=i
most_dist = euclidean_distance(anch, most_sim_neg)
#print (least_sim_dist,most_dist)
return least_sim_pos_idx,most_sim_neg_idx
# print (preds)
# print ("______________________________________")
#print ((image_anch,filtered_pos[np.argmax(preds_pos,axis=0)]))
#return filtered_pos[np.argmax(preds,axis=0)]
def __data_generation(self, image_IDs):
anchor_batch = []
positive_batch = []
negative_batch = []
for img_path in image_IDs:
#print (img_path)
anchor = img_to_np(os.path.join(self.dataset_path, img_path[0]))
positive = img_to_np(os.path.join(self.dataset_path, img_path[1]))
negative = img_to_np(os.path.join(self.dataset_path, img_path[2]))
anchor_batch.append(anchor)
positive_batch.append(positive)
negative_batch.append(negative)
return [np.array(anchor_batch), np.array(positive_batch), np.array(negative_batch)]