-
Notifications
You must be signed in to change notification settings - Fork 0
/
extract_BCF.py
167 lines (134 loc) · 5.16 KB
/
extract_BCF.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
import os, cv2
import torch,time
import clip
from PIL import Image
import numpy as np
from torchvision.transforms import Compose, ToTensor, Normalize
device = "cuda" if torch.cuda.is_available() else "cpu"
model, _ = clip.load("ViT-B/32", device=device)
preprocess = Compose([
Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
])
# prompts = [
# f'an underexposed photo',
# f'a well-exposed photo',
# f'a overexposed photo'
# ]
prompts = [
f'an underexposed photo',
f'a slightly underexposed photo',
f'a well-exposed photo',
f'a slightly overexposed photo',
f'a overexposed photo'
]
text = clip.tokenize(prompts).to(device)
to_tensor = ToTensor()
def GET_BC(videos_dir, video_name):
save_folder = 'BC_Features'
os.makedirs(save_folder, exist_ok=True)
filename = os.path.join(videos_dir, video_name)
video_name_str = video_name[:-4]
video_capture = cv2.VideoCapture()
video_capture.open(filename)
cap=cv2.VideoCapture(filename)
video_length = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
dim = (1120, 672)
frames = []
for i in range(video_length):
has_frames, frame = video_capture.read()
if has_frames:
read_frame = cv2.resize(frame, dim)
frames.append(read_frame)
final_res = []
res = []
length = len(frames)
print(length)
now = 0
interval = 10
while now + interval - 1 < length:
final = []
for i in range(interval):
read_frame = Image.fromarray(cv2.cvtColor(frames[i + now], cv2.COLOR_BGR2RGB))
final.append(to_tensor(read_frame).to(device))
image = torch.stack(final, dim = 0)
# 形状变为 [Batch, 3, 5, 224, 3, 224]
image = image.unfold(2, 224, 224).unfold(3, 224, 224)
# 接下来，调整维度的顺序
# 新的形状为 [Batch, 5, 3, 3, 224, 224]
image = image.permute(0, 3, 2, 1, 4, 5).contiguous()
# 最后，合并两个中间维度以形成最终的Tensor
# 形状变为 [Batch, 15, 3, 224, 224]
image = image.reshape(-1, 15, 3, 224, 224)
image = image.view(-1, 3, 224, 224)
image = preprocess(image)
with torch.no_grad():
logits_per_image, logits_per_text = model(image, text)
tmp = logits_per_image.softmax(dim=-1)
tmp = tmp.view(interval, -1)
tmp *= 10
# print(tmp.shape)
res.append(tmp)
now += interval
if length > now:
final = []
for i in range(now, length):
read_frame = Image.fromarray(cv2.cvtColor(frames[i], cv2.COLOR_BGR2RGB))
final.append(to_tensor(read_frame).to(device))
image = torch.stack(final, dim = 0)
# 形状变为 [Batch, 3, 5, 224, 3, 224]
image = image.unfold(2, 224, 224).unfold(3, 224, 224)
# 接下来，调整维度的顺序
# 新的形状为 [Batch, 5, 3, 3, 224, 224]
image = image.permute(0, 3, 2, 1, 4, 5).contiguous()
# 最后，合并两个中间维度以形成最终的Tensor
# 形状变为 [Batch, 15, 3, 224, 224]
image = image.reshape(-1, 15, 3, 224, 224)
image = image.view(-1, 3, 224, 224)
image = preprocess(image)
with torch.no_grad():
logits_per_image, logits_per_text = model(image, text)
tmp = logits_per_image.softmax(dim=-1)
tmp = tmp.view(length - now, -1)
tmp *= 10
# print(tmp)
res.append(tmp)
res = torch.cat(res, dim=0)
# print(res.shape)
final_res=[]
for step in [1, 2, 4, 8]:
chunk_number = 8 // step
chunk_size = length // chunk_number
chunks = []
for i in range(chunk_number):
if i < chunk_number - 1:
chunk = res[i * chunk_size : (i + 1) * chunk_size, :]
else:
chunk = res[(chunk_number - 1) * chunk_size : , :]
tmp = []
for j in range(step):
temp = chunk[j :: step, :]
tmp.append(torch.var(temp.float(), dim=0))
chunks.append(tmp)
final_res.append(chunks)
temp = []
for i in range(8):
temp.append(torch.cat(final_res[0][i]
+ [torch.mean(torch.stack(final_res[1][i // 2], dim=0), dim=0)]
+ [torch.mean(torch.stack(final_res[2][i // 4], dim=0), dim=0)]
+ [torch.mean(torch.stack(final_res[3][i // 8], dim=0), dim=0)]
, dim=0))
final_res = torch.stack(temp, dim=0)
print(final_res.shape)
if torch.isnan(final_res).any() or torch.isinf(final_res).any():
print(f'{video_name_str}: dead.')
else:
print(f'{video_name_str}: live.')
np.save(save_folder + f'/{video_name_str}_BC.npy', final_res.cpu())
def run(lb,ub):
cnt = 0
videos_dir = 'final'
for video_name in os.listdir(videos_dir):
if cnt in range(lb, ub):
GET_BC(videos_dir, video_name)
cnt += 1
run(0, 4518)