-
Notifications
You must be signed in to change notification settings - Fork 1
/
models.py
145 lines (107 loc) · 4.49 KB
/
models.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
import torch
import torch.nn as nn
import timm
from botnet import BoTStack
class Identity(nn.Module):
def __init__(self):
super(Identity, self).__init__()
def forward(self, x):
return x
class Swin_b_384_in22k(torch.nn.Module):
def __init__(self, global_pool):
super(Swin_b_384_in22k, self).__init__()
swin_b = timm.create_model('swin_base_patch4_window12_384_in22k', pretrained=True, global_pool=global_pool)
swin_b.head = Identity()
self.feature_extraction = swin_b
self.quality = self.quality_regression(1024+256, 128,1)
def quality_regression(self,in_channels, middle_channels, out_channels):
regression_block = nn.Sequential(
nn.Linear(in_channels, middle_channels),
nn.Linear(middle_channels, out_channels),
)
return regression_block
def forward(self, x):
x = self.feature_extraction(x)
x = self.quality(x)
return x
class RQ_VQA(torch.nn.Module):
def __init__(self, pretrained_path):
super(RQ_VQA, self).__init__()
model = Swin_b_384_in22k(global_pool='')
if pretrained_path!= None:
model.load_state_dict(torch.load(pretrained_path))
model.quality = Identity()
swin_b = model
self.feature_extraction = swin_b
self.bot4 = BoTStack(dim=1024, dim_out=1024, num_layers=3, fmap_size=(12, 12), stride=1, rel_pos_emb=True)
self.quality = self.quality_regression(1024+256+4096+495+768, 128,1)
def quality_regression(self,in_channels, middle_channels, out_channels):
regression_block = nn.Sequential(
nn.Linear(in_channels, middle_channels),
nn.Linear(middle_channels, out_channels),
)
return regression_block
def forward(self, x, x_3D_features, x_LLM, x_LIQE, x_SlowFast):
# input dimension: batch x frames x 3 x height x width
x_size = x.shape
# x_3D: batch x frames x 2048
x_3D_features_size = x_3D_features.shape
x_LLM_size = x_LLM.shape
x_LIQE_size = x_LIQE.shape
x_SlowFast_size = x_SlowFast.shape
# x: batch * frames x 3 x height x width
x = x.view(-1, x_size[2], x_size[3], x_size[4])
# x_3D: batch * frames x 2048
x_3D_features = x_3D_features.view(-1, x_3D_features_size[2])
x_LLM = x_LLM.view(-1, x_LLM_size[2])
x_LIQE = x_LIQE.view(-1, x_LIQE_size[2])
x_SlowFast = x_SlowFast.view(-1, x_SlowFast_size[2])
x = self.feature_extraction(x)
x = x.permute(0, 2, 1)
x = x.view(-1, 1024, 12, 12)
x = self.bot4(x)
x = x.view(-1, 1024, 144)
x = x.permute(0, 1, 2)
x = x.mean(dim=2)
x = torch.cat((x, x_3D_features, x_LLM, x_LIQE, x_SlowFast), dim = 1)
# print(x.shape)
x = self.quality(x)
# x: batch x frames
x = x.view(x_size[0],x_size[1])
# x: batch x 1
x = torch.mean(x, dim = 1)
return x
class RQ_VQA_base_model(torch.nn.Module):
def __init__(self, pretrained_path):
super(RQ_VQA_base_model, self).__init__()
model = Swin_b_384_in22k(global_pool='avg')
if pretrained_path!= None:
model.load_state_dict(torch.load(pretrained_path))
model.quality = Identity()
swin_b = model
self.feature_extraction = swin_b
self.quality = self.quality_regression(1024+256, 128,1)
def quality_regression(self,in_channels, middle_channels, out_channels):
regression_block = nn.Sequential(
nn.Linear(in_channels, middle_channels),
nn.Linear(middle_channels, out_channels),
)
return regression_block
def forward(self, x, x_3D_features):
# input dimension: batch x frames x 3 x height x width
x_size = x.shape
# x_3D: batch x frames x 2048
x_3D_features_size = x_3D_features.shape
# x: batch * frames x 3 x height x width
x = x.view(-1, x_size[2], x_size[3], x_size[4])
# x_3D: batch * frames x 2048
x_3D_features = x_3D_features.view(-1, x_3D_features_size[2])
x = self.feature_extraction(x)
x = torch.cat((x, x_3D_features), dim = 1)
# print(x.shape)
x = self.quality(x)
# x: batch x frames
x = x.view(x_size[0],x_size[1])
# x: batch x 1
x = torch.mean(x, dim = 1)
return x