-
Notifications
You must be signed in to change notification settings - Fork 0
/
validation.py
254 lines (205 loc) · 10.6 KB
/
validation.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
import torch
import numpy as np
import cv2
import pickle
from sklearn.neighbors import NearestNeighbors
import torchvision.transforms as T
import os
from PIL import Image
import matplotlib.pyplot as plt
from sklearn.cluster import KMeans
from sklearn.neighbors import KNeighborsClassifier
from lshash import LSHash
import models
import model4
import embedding
import config
import data
class validation():
def __init__(self,embedding_mode,similarity_mode):
self.SimMode = similarity_mode
self.EmbMode = embedding_mode
self.DEVICE = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
self.ENCODER = model4.ConvEncoder()
self.ENCODER.load_state_dict(torch.load(config.ENC_STATE, map_location=self.DEVICE))
self.ENCODER.eval()
self.ENCODER.to(self.DEVICE)
self.DATASET = data.FolderDataset(config.DATASET_PATH, config.TRANSFORMS)
self.DATASET_EMB, self.IMG_PATH_LIST = self.get_embedding_dataset()
if(similarity_mode == "LSH"):
k = 10 # hash size
L = 5 # number of tables
d = self.DATASET_EMB.shape[1] # Dimension of Feature vector
self.lsh = LSHash(hash_size=k, input_dim=d, num_hashtables=L)
def load_image_tensor(self,image_path):
"""
Definition: Load image from the path, resize, normalize and convert to the tensor
"""
image_tensor = config.TRANSFORMS(Image.open(image_path))
image_tensor = image_tensor.unsqueeze(0)
return image_tensor
def image_embedding(self,image_path):
"""
Definition: Create query image embedding according to the choice of Embedding mode
if embedding mode == Flatten; The representation matrix obtained as the result of encoder
is directly converted to a 1-dimensional vector.
if embedding mode == Max_Pool; The one-dimensional vector is obtained by taking the highest
value in each chanel of the representation matrix.
(Like Global Max Pooling)
"""
image_tensor = self.load_image_tensor(image_path)
image_tensor = image_tensor.to(self.DEVICE)
with torch.no_grad():
image_embedding = self.ENCODER(image_tensor)
print(image_embedding.shape)
if(self.EmbMode == "Flatten"):
image_embedding_numpy = image_embedding.cpu().detach().numpy()
flattened_embedding = image_embedding_numpy.reshape((image_embedding_numpy.shape[0], -1))
return flattened_embedding
elif (self.EmbMode == "Max_Pool"):
Max_Pool_embedding=torch.amax(image_embedding, dim=(2, 3)).cpu().detach().numpy()
return Max_Pool_embedding
else:
return "Embedding Mode not sellected"
def get_embedding_dataset(self):
"""
Definition: Create the embedding of all images in dataset according to the choice of Embedding mode
if embedding mode == Flatten; The representation matrix obtained as the result of encoder
is directly converted to a 1-dimensional vector.
if embedding mode == Max_Pool; The one-dimensional vector is obtained by taking the highest
value in each chanel of the representation matrix.
(Like Global Max Pooling)
"""
""" Get Representation Matrix of all images in dataset """
path = os.getcwd()
allfiles = os.listdir(path)
if( config.IDX_SAVE in allfiles):
img_index = torch.load(config.IDX_SAVE , map_location=self.DEVICE)
emb = torch.load(config.TOTAL_EMB, map_location=self.DEVICE)
else:
Dataset_loader = torch.utils.data.DataLoader(self.DATASET, batch_size=32)
emb,img_index = embedding.create_embedding(self.ENCODER, Dataset_loader, config.EMBEDDING_SHAPE_MODEL, self.DEVICE)
torch.save(emb, config.TOTAL_EMB)
torch.save(img_index, config.IDX_SAVE)
if(self.EmbMode == "Max_Pool"):
max_emb =torch.amax(emb, dim=(2, 3))
numpy_max_embedding = max_emb.cpu().detach().numpy()
#num_images = numpy_max_embedding.shape[0]
#np.save(config.EMB_SAVE_MP, numpy_max_embedding)
#final_embedding = np.load(config.EMB_SAVE_MP)
return numpy_max_embedding,img_index
elif (self.EmbMode == "Flatten"):
numpy_embedding = emb.cpu().detach().numpy()
num_images = numpy_embedding.shape[0]
flattened_embedding = numpy_embedding.reshape((num_images, -1))
#np.save(config.EMB_SAVE, flattened_embedding)
#final_embedding = np.load(config.EMB_SAVE)
return flattened_embedding, img_index
else:
return "Embedding Mode not sellected"
def compute_similar_images(self,image_path, num_images, num_cluster=None):
img_embedding = self.image_embedding(image_path)
path = os.getcwd()
allfiles = os.listdir(path)
if (self.SimMode == "NN_cosine"):
if (config.KNN_NN_COSINE in allfiles):
knn=pickle.load(open(config.KNN_NN_COSINE, 'rb'))
else:
knn = NearestNeighbors(n_neighbors=num_images, metric="cosine")
knn.fit(self.DATASET_EMB)
knnPickle = open(config.KNN_NN_COSINE, 'wb')
pickle.dump(knn, knnPickle)
_, indices = knn.kneighbors(img_embedding)
indices_list = indices.tolist()
return indices_list
elif (self.SimMode == "KMeans"):
if(config.KMEANS in allfiles):
kmeans = pickle.load(open(config.KMEANS, 'rb'))
else:
kmeans = KMeans(n_clusters = num_cluster, random_state=0).fit(self.DATASET_EMB)
kmeansPickle = open(config.KMEANS, 'wb')
pickle.dump(kmeans, kmeansPickle)
labels=kmeans.labels_
if(config.KNN_KMEANS in allfiles):
knn=pickle.load(open(config.KNN_KMEANS, 'rb'))
else:
knn = KNeighborsClassifier(n_neighbors=num_images,algorithm='ball_tree',n_jobs=-1)
knn.fit(self.DATASET_EMB,np.array(labels))
knnPickle = open(config.KNN_KMEANS, 'wb')
pickle.dump(knn, knnPickle)
_,res = knn.kneighbors(img_embedding,return_distance=True,n_neighbors=num_images)
res_list =res.tolist()
return res_list
elif (self.SimMode == "LSH"):
num_images = self.DATASET_EMB.shape[0] # size: number of images + 1 First embedding is dummy embedding
"""create embedding function start with one dummy representation matrix"""
for i in range(num_images-1):
self.lsh.index(self.DATASET_EMB[i+1])
list_r=self.lsh.query(img_embedding[0],num_results=num_images,distance_func="euclidean")
print("LSH Result:")
print(list_r)
return list_r
else:
return "Similarity mode not sellected"
def return_similar_images(self,indices_list,path):
if(self.SimMode == "LSH"):
path_list=[]
for j in range(config.NUM_IMG):
for key in indices_list[j]:
print(type(np.asarray(key)))
index=np.where((self.DATASET_EMB == np.asarray(key)).all(axis=1))
print(index)
for i in index[0]:
if self.IMG_PATH_LIST[i-1] not in path_list:
file_path = os.path.join(config.DATASET_PATH,self.IMG_PATH_LIST[i-1])
path_list.append(file_path)
# image = cv2.imread(file_path)
# plt.imshow(image)
# plt.show()
# fig = plt.figure(figsize=(8, 8))
# for i in range(1,len(path_list)+1):
# image = cv2.imread(file_path)
# fig.add_subplot(2,4,i)
# plt.imshow(image)
# plt.show()
print(path_list)
return path_list
else:
img_list=[]
indices = indices_list[0]
for index in indices:
if index == 0:
# index 0 is a dummy embedding.
pass
else:
transforms=T.ToPILImage();
a=self.DATASET[int(index)-1][0]
img = transforms(a)
img_list.append(img)
print(self.IMG_PATH_LIST[int(index)-1])
# plt.imshow(img)
# plt.show()
fig = plt.figure(figsize=(8, 8))
for i in range(1,len(img_list)+1):
fig.add_subplot(2,4,i)
plt.imshow(img_list[i-1])
#plt.savefig(path)
plt.show()
return img_list
def load_autoencoder_state(self):
# Load state of encoder
self.ENCODER.load_state_dict(torch.load(config.ENC_STATE, map_location=self.DEVICE))
self.ENCODER.eval()
self.ENCODER.to(self.DEVICE)
#Load state of decoder
# self.DECODER.load_state_dict(torch.load(config.DEC_STATE, map_location=self.DEVICE))
# self.DECODER.eval()
# self.DECODER.to(self.DEVICE)
def result(self):
indices_list = self.compute_similar_images(config.IMAGE_PATH,config.NUM_IMG, 50)
return indices_list
if __name__ == "__main__":
val = validation("Max_Pool","KMeans")
#val.load_autoencoder_state()
img_list = val.result()
val.return_similar_images(img_list,"Result_dataset2/Model4/Flatten-NNcosine")