-
Notifications
You must be signed in to change notification settings - Fork 41
/
Copy pathCNNfeatures.py
173 lines (144 loc) · 6.82 KB
/
CNNfeatures.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
"""Extracting Content-Aware Perceptual Features using Pre-Trained ResNet-50"""
# Author: Dingquan Li
# Email: dingquanli AT pku DOT edu DOT cn
# Date: 2018/3/27
#
# CUDA_VISIBLE_DEVICES=0 python CNNfeatures.py --database=KoNViD-1k --frame_batch_size=64
# CUDA_VISIBLE_DEVICES=1 python CNNfeatures.py --database=CVD2014 --frame_batch_size=32
# CUDA_VISIBLE_DEVICES=0 python CNNfeatures.py --database=LIVE-Qualcomm --frame_batch_size=8
import torch
from torchvision import transforms, models
import torch.nn as nn
from torch.utils.data import Dataset
import skvideo.io
from PIL import Image
import os
import h5py
import numpy as np
import random
from argparse import ArgumentParser
class VideoDataset(Dataset):
"""Read data from the original dataset for feature extraction"""
def __init__(self, videos_dir, video_names, score, video_format='RGB', width=None, height=None):
super(VideoDataset, self).__init__()
self.videos_dir = videos_dir
self.video_names = video_names
self.score = score
self.format = video_format
self.width = width
self.height = height
def __len__(self):
return len(self.video_names)
def __getitem__(self, idx):
video_name = self.video_names[idx]
assert self.format == 'YUV420' or self.format == 'RGB'
if self.format == 'YUV420':
video_data = skvideo.io.vread(os.path.join(self.videos_dir, video_name), self.height, self.width, inputdict={'-pix_fmt':'yuvj420p'})
else:
video_data = skvideo.io.vread(os.path.join(self.videos_dir, video_name))
video_score = self.score[idx]
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
video_length = video_data.shape[0]
video_channel = video_data.shape[3]
video_height = video_data.shape[1]
video_width = video_data.shape[2]
transformed_video = torch.zeros([video_length, video_channel, video_height, video_width])
for frame_idx in range(video_length):
frame = video_data[frame_idx]
frame = Image.fromarray(frame)
frame = transform(frame)
transformed_video[frame_idx] = frame
sample = {'video': transformed_video,
'score': video_score}
return sample
class ResNet50(torch.nn.Module):
"""Modified ResNet50 for feature extraction"""
def __init__(self):
super(ResNet50, self).__init__()
self.features = nn.Sequential(*list(models.resnet50(pretrained=True).children())[:-2])
for p in self.features.parameters():
p.requires_grad = False
def forward(self, x):
# features@: 7->res5c
for ii, model in enumerate(self.features):
x = model(x)
if ii == 7:
features_mean = nn.functional.adaptive_avg_pool2d(x, 1)
features_std = global_std_pool2d(x)
return features_mean, features_std
def global_std_pool2d(x):
"""2D global standard variation pooling"""
return torch.std(x.view(x.size()[0], x.size()[1], -1, 1),
dim=2, keepdim=True)
def get_features(video_data, frame_batch_size=64, device='cuda'):
"""feature extraction"""
extractor = ResNet50().to(device)
video_length = video_data.shape[0]
frame_start = 0
frame_end = frame_start + frame_batch_size
output1 = torch.Tensor().to(device)
output2 = torch.Tensor().to(device)
extractor.eval()
with torch.no_grad():
while frame_end < video_length:
batch = video_data[frame_start:frame_end].to(device)
features_mean, features_std = extractor(batch)
output1 = torch.cat((output1, features_mean), 0)
output2 = torch.cat((output2, features_std), 0)
frame_end += frame_batch_size
frame_start += frame_batch_size
last_batch = video_data[frame_start:video_length].to(device)
features_mean, features_std = extractor(last_batch)
output1 = torch.cat((output1, features_mean), 0)
output2 = torch.cat((output2, features_std), 0)
output = torch.cat((output1, output2), 1).squeeze()
return output
if __name__ == "__main__":
parser = ArgumentParser(description='"Extracting Content-Aware Perceptual Features using Pre-Trained ResNet-50')
parser.add_argument("--seed", type=int, default=19920517)
parser.add_argument('--database', default='KoNViD-1k', type=str,
help='database name (default: KoNViD-1k)')
parser.add_argument('--frame_batch_size', type=int, default=64,
help='frame batch size for feature extraction (default: 64)')
parser.add_argument('--disable_gpu', action='store_true',
help='flag whether to disable GPU')
args = parser.parse_args()
torch.manual_seed(args.seed) #
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(args.seed)
random.seed(args.seed)
torch.utils.backcompat.broadcast_warning.enabled = True
if args.database == 'KoNViD-1k':
videos_dir = '/home/ldq/Downloads/KoNViD-1k/' # videos dir
features_dir = 'CNN_features_KoNViD-1k/' # features dir
datainfo = 'data/KoNViD-1kinfo.mat' # database info: video_names, scores; video format, width, height, index, ref_ids, max_len, etc.
if args.database == 'CVD2014':
videos_dir = '/media/ldq/Research/Data/CVD2014/'
features_dir = 'CNN_features_CVD2014/'
datainfo = 'data/CVD2014info.mat'
if args.database == 'LIVE-Qualcomm':
videos_dir = '/media/ldq/Others/Data/12.LIVE-Qualcomm Mobile In-Capture Video Quality Database/'
features_dir = 'CNN_features_LIVE-Qualcomm/'
datainfo = 'data/LIVE-Qualcomminfo.mat'
if not os.path.exists(features_dir):
os.makedirs(features_dir)
device = torch.device("cuda" if not args.disable_gpu and torch.cuda.is_available() else "cpu")
Info = h5py.File(datainfo, 'r')
video_names = [Info[Info['video_names'][0, :][i]][()].tobytes()[::2].decode() for i in range(len(Info['video_names'][0, :]))]
scores = Info['scores'][0, :]
video_format = Info['video_format'][()].tobytes()[::2].decode()
width = int(Info['width'][0])
height = int(Info['height'][0])
dataset = VideoDataset(videos_dir, video_names, scores, video_format, width, height)
for i in range(len(dataset)):
current_data = dataset[i]
current_video = current_data['video']
current_score = current_data['score']
print('Video {}: length {}'.format(i, current_video.shape[0]))
features = get_features(current_video, args.frame_batch_size, device)
np.save(features_dir + str(i) + '_resnet-50_res5c', features.to('cpu').numpy())
np.save(features_dir + str(i) + '_score', current_score)