video_recommender.py
6.83 KB
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
from gensim.models import Word2Vec
import numpy as np
from numpy import dot
from numpy.linalg import norm
import pandas as pd
import math
import activation as ac
def recommend_videos(tags, segments, tag_model_path, video_model_path, top_k):
# 이 함수에서 모든걸 다 함
# tags는 label val 로 묶인 문자열 리스트임
# tags의 길이는 segment의 길이
# 비디오 벡터를 생성한 뒤 각 segment의 벡터와 비교해 가장 유사도 높은 segment의 인덱스 추출
# 그 후 그 인덱스 전후 2개 총 크기 5의 커널을 생성
# 벡터공간에서 비교한 후 추천할만한 영상의 아이디만 반환
#segments는 클래스 확률 클래스 확률... 일케 저장되어 있음
tag_vectors = Word2Vec.load(tag_model_path).wv
video_vectors = Word2Vec().wv.load(video_model_path)
error_tags = []
maxSimilarSegment = 0
maxSimilarity = -1
kernel = [np.zeros(100) for i in range(0,5)]
#우선은 비교를 뜰 입력 영상의 단일 비디오벡터를 구함
video_vector = np.zeros(100)
tag_preds =[]
for (tag, weight) in tags:
tag_preds.append(weight)
#print("tag preds = ",tag_preds)
tag_preds = ac.softmax(tag_preds)
for (tag, weight),pred in zip(tags,tag_preds):
print(tag,pred)
if tag in tag_vectors.vocab:
#float(weight)
video_vector = video_vector + (tag_vectors[tag] * float(pred))
else:
print("unknown",tag)
# Pass if tag is unknown
if tag not in error_tags:
error_tags.append(tag)
#각 세그먼트마다 비교를 떠서 인덱스를 저장
currentIndex = 0
for segment in segments:
segment_vector = np.zeros(100)
segTags = [segment[i] for i in range(0,len(segment),2)]
segProbs = ac.softmax([float(segment[i]) for i in range(1,len(segment),2)])
#print(segProbs)
for tag, weight in zip(segTags,segProbs):
if tag in tag_vectors.vocab:
#float(weight)
segment_vector = segment_vector + (tag_vectors[tag] * float(weight))
else:
# Pass if tag is unknown
if tag not in error_tags:
error_tags.append(tag)
#비디오 벡터와 세그먼트 벡터 비교
similarity = cos_sim(video_vector, segment_vector) #cos_sim(video_vector, segment_vector)#
#print(segTags,similarity)
if similarity >= maxSimilarity:
maxSimilarSegment = currentIndex
maxSimilarity = similarity
if maxSimilarSegment < int(len(kernel)/2):
maxSimilarSegment = int(len(kernel)/2)
elif maxSimilarSegment == len(segments) - int(len(kernel)/2):
maxSimilarSegment = len(segments) - int(len(kernel)/2) - 1
#세그먼트 인덱스 증가
currentIndex = currentIndex + 1
#print("len=============================================")
#print(len(kernel))
#print(maxSimilarSegment)
#커널 생성
for k in range (0,len(kernel)):
segment = segments[maxSimilarSegment - math.floor(len(kernel)/2) + k]
segment_vector = np.zeros(100)
segTags = [segment[i] for i in range(0,len(segment),2)]
segProbs = ac.softmax([float(segment[i]) for i in range(1,len(segment),2)])
print(segTags)
print(segProbs)
normalize(segProbs)
for (tag, weight) in zip(segTags,segProbs):
if tag in tag_vectors.vocab:
#float(weight)
segment_vector = segment_vector + (tag_vectors[tag] * float(weight))
else:
print("unknown",tag)
# Pass if tag is unknown
if tag not in error_tags:
error_tags.append(tag)
kernel[k] = segment_vector
#여기에서 유사한 벡터들을 뽑아냄
#현재는 비디오id로 영상을 얻을 수 없으므로 반환값으로 비디오 아이디와 태그들, 확률 사용
video_tags = pd.read_csv('/mnt/e/khuhub/2015104192/web/backend/yt8m/esot3ria/segment_tags.csv', encoding='utf8',error_bad_lines=False)
videoVectorList = []
prevVideoId = ""
minimunVideoIds = [["",-1.0] for i in range(0,top_k)]
for i, row in video_tags.iterrows():
video_id = row[0]
if video_id == "vid_id":
continue
if prevVideoId == "":
prevVideoId = video_id
if video_id[0:4] != prevVideoId[0:4]:
#여기서 모다진걸로 컨볼루션 연산
#if('fIvl' == video_id[0:4]):
#print("====")
#for a in kernel:
# print(len(kernel),norm(a))
maxima, idx = convolution(videoVectorList,kernel,prevVideoId)
#print(video_id,maxima)
localMinima = 100
localMinimaIndex = -1
for seg in range(0,top_k):
if float(minimunVideoIds[seg][1]) < localMinima:
localMinima = float(minimunVideoIds[seg][1])
localMinimaIndex = seg
#print(maxima)
if localMinima < maxima:
print(prevVideoId[0:4] + "_" + str(idx),maxima)
minimunVideoIds[localMinimaIndex] = [prevVideoId[0:4] + "_" + str(idx),maxima]
videoVectorList.clear()
prevVideoId = video_id
if video_id == "finished":
break
videoVectorList.append(video_vectors[video_id])
similar_ids = []
for i in range(0,top_k):
similar_ids.append(minimunVideoIds[i][0])
#similar_ids = [x[0] for x in video_vectors.similar_by_vector(video_vector, top_k)]
print('results =' ,similar_ids)
return similar_ids
def cos_sim(A, B):
denom = norm(A)*norm(B)
if denom == 0:
#print("a = ",norm(A)," b = ",norm(B))
return 0
else:
return dot(A, B)/(norm(A)*norm(B))
def shiftKernel(kernel, newValue):
for i in range(0, len(kernel) - 1):
kernel[i] = kernel[i+1]
kernel[len(kernel) - 1] = newValue
def convolution(arrs, _kernel,vidId):
s = len(_kernel)
result = []
midpos = math.floor(s/2)
for i in range(0,midpos):
arrs.insert(0,np.zeros(100))
arrs.append(np.zeros(100))
for j in range(midpos,len(arrs) - midpos):
convResult = 0
for i in range(0, s):
convResult = convResult + cos_sim(arrs[j - midpos + i],_kernel[i])
result.append(convResult)
maxVal = max(result)
index = result.index(maxVal)
return maxVal,index
def normalize(arrs):
maximum = max(arrs)
minimum = min(arrs)
denom = maximum - minimum
for i in range(0,len(arrs)):
arrs[i] = (arrs[i] - minimum)/ denom