-
Notifications
You must be signed in to change notification settings - Fork 7
/
compression_module.py
123 lines (79 loc) · 3.77 KB
/
compression_module.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import transforms
from torchvision.utils import save_image
import numpy as np
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
class BEC(torch.autograd.Function):
@staticmethod
def forward(ctx, x, p=0.2):
x_tmp = torch.round(x * 256)
#x_tmp = x_tmp.int()
x_tmp = x_tmp.byte()
p_complement = 1-p
std = x
binomial_noise = np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 1 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 2 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 4 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 8 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 16 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 32 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 64 + \
np.random.binomial(1,p_complement,(std.size())).astype(np.uint8) * 128
binomial_noise = torch.ByteTensor(binomial_noise).to(device)
x_tmp_filter = x_tmp & binomial_noise
x_tmp_filter = x_tmp_filter.float()
x_tmp_filter = x_tmp_filter + (255.0 - binomial_noise.float()) / 2.0
x_tmp_filter /= 255.0
return x_tmp_filter
@staticmethod
def backward(ctx, grad_output):
grad_input = grad_output.clone()
return grad_input, None
class compression_module(nn.Module):
def __init__(self, input_channel=256, hidden_channel=128, noise=10, channel = 1,spatial = 0):
super(compression_module, self).__init__()
self.conv1 = nn.Conv2d(input_channel+1,hidden_channel,kernel_size = 3,stride=1,padding=1)
self.conv2 = nn.Conv2d(hidden_channel,input_channel,kernel_size = 3,stride=1,padding=1)
self.batchnorm1 = nn.BatchNorm2d(hidden_channel)
self.batchnorm2 = nn.BatchNorm2d(input_channel)
self.conv3 = nn.Conv2d(input_channel+1,hidden_channel,kernel_size=2,stride=2)
self.conv4 = nn.ConvTranspose2d(hidden_channel,input_channel,kernel_size=2,stride=2)
self.noise = noise
self.channel =channel
self.spatial = spatial
def forward(self, x):
H = x.size()[2]
C = x.size()[1]
B = x.size()[0]
noise_factor = torch.rand(1) * self.noise
#noise_factor = torch.FloatTenspr([1]) * self.noise
p = noise_factor.numpy()
noise_factor = noise_factor.to(device)
noise_matrix = torch.FloatTensor(np.ones((B,1,H,H))).to(device) * noise_factor
x = torch.cat((x,noise_matrix),dim = 1)
if self.spatial == 0:
x = torch.sigmoid(self.batchnorm1(self.conv1(x)))
elif self.spatial == 1:
x = torch.sigmoid(self.batchnorm1(self.conv3(x)))
x_tmp = x
if self.channel == 'a':
x = awgn_noise(x,noise_factor)
elif self.channel == 'e':
bec = BEC.apply
x = bec(x,p)
elif self.channel == 'w':
x = x
else:
print('error')
if self.spatial == 1:
x = F.relu(self.batchnorm2(self.conv2(x)))
else:
x = F.relu(self.batchnorm2(self.conv4(x)))
return x
def awgn_noise(x, noise_factor):
return x + torch.randn_like(x) * noise_factor