File size: 6,718 Bytes
2514fb4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import torch.nn as nn
import models.basicblock as B
import functools
import torch.nn.functional as F
import torch.nn.init as init


"""
# --------------------------------------------
# modified SRResNet
#   -- MSRResNet0 (v0.0)
#   -- MSRResNet1 (v0.1)
# --------------------------------------------
References:
@inproceedings{wang2018esrgan,
  title={Esrgan: Enhanced super-resolution generative adversarial networks},
  author={Wang, Xintao and Yu, Ke and Wu, Shixiang and Gu, Jinjin and Liu, Yihao and Dong, Chao and Qiao, Yu and Change Loy, Chen},
  booktitle={European Concerence on Computer Vision (ECCV)},
  pages={0--0},
  year={2018}
}
@inproceedings{ledig2017photo,
  title={Photo-realistic single image super-resolution using a generative adversarial network},
  author={Ledig, Christian and Theis, Lucas and Husz{\'a}r, Ferenc and Caballero, Jose and Cunningham, Andrew and Acosta, Alejandro and Aitken, Andrew and Tejani, Alykhan and Totz, Johannes and Wang, Zehan and others},
  booktitle={IEEE concerence on computer vision and pattern recognition},
  pages={4681--4690},
  year={2017}
}
# --------------------------------------------
"""


# --------------------------------------------
# modified SRResNet v0.0
# https://github.com/xinntao/ESRGAN
# --------------------------------------------
class MSRResNet0(nn.Module):
    def __init__(self, in_nc=3, out_nc=3, nc=64, nb=16, upscale=4, act_mode='R', upsample_mode='upconv'):
        """
        in_nc: channel number of input
        out_nc: channel number of output
        nc: channel number
        nb: number of residual blocks
        upscale: up-scale factor
        act_mode: activation function
        upsample_mode: 'upconv' | 'pixelshuffle' | 'convtranspose'
        """
        super(MSRResNet0, self).__init__()
        assert 'R' in act_mode or 'L' in act_mode, 'Examples of activation function: R, L, BR, BL, IR, IL'

        n_upscale = int(math.log(upscale, 2))
        if upscale == 3:
            n_upscale = 1

        m_head = B.conv(in_nc, nc, mode='C')

        m_body = [B.ResBlock(nc, nc, mode='C'+act_mode+'C') for _ in range(nb)]
        m_body.append(B.conv(nc, nc, mode='C'))

        if upsample_mode == 'upconv':
            upsample_block = B.upsample_upconv
        elif upsample_mode == 'pixelshuffle':
            upsample_block = B.upsample_pixelshuffle
        elif upsample_mode == 'convtranspose':
            upsample_block = B.upsample_convtranspose
        else:
            raise NotImplementedError('upsample mode [{:s}] is not found'.format(upsample_mode))
        if upscale == 3:
            m_uper = upsample_block(nc, nc, mode='3'+act_mode)
        else:
            m_uper = [upsample_block(nc, nc, mode='2'+act_mode) for _ in range(n_upscale)]

        H_conv0 = B.conv(nc, nc, mode='C'+act_mode)
        H_conv1 = B.conv(nc, out_nc, bias=False, mode='C')
        m_tail = B.sequential(H_conv0, H_conv1)

        self.model = B.sequential(m_head, B.ShortcutBlock(B.sequential(*m_body)), *m_uper, m_tail)

    def forward(self, x):
        x = self.model(x)
        return x


# --------------------------------------------
# modified SRResNet v0.1
# https://github.com/xinntao/ESRGAN
# --------------------------------------------
class MSRResNet1(nn.Module):
    def __init__(self, in_nc=3, out_nc=3, nc=64, nb=16, upscale=4, act_mode='R', upsample_mode='upconv'):
        super(MSRResNet1, self).__init__()
        self.upscale = upscale

        self.conv_first = nn.Conv2d(in_nc, nc, 3, 1, 1, bias=True)
        basic_block = functools.partial(ResidualBlock_noBN, nc=nc)
        self.recon_trunk = make_layer(basic_block, nb)

        # upsampling
        if self.upscale == 2:
            self.upconv1 = nn.Conv2d(nc, nc * 4, 3, 1, 1, bias=True)
            self.pixel_shuffle = nn.PixelShuffle(2)
        elif self.upscale == 3:
            self.upconv1 = nn.Conv2d(nc, nc * 9, 3, 1, 1, bias=True)
            self.pixel_shuffle = nn.PixelShuffle(3)
        elif self.upscale == 4:
            self.upconv1 = nn.Conv2d(nc, nc * 4, 3, 1, 1, bias=True)
            self.upconv2 = nn.Conv2d(nc, nc * 4, 3, 1, 1, bias=True)
            self.pixel_shuffle = nn.PixelShuffle(2)

        self.HRconv = nn.Conv2d(nc, nc, 3, 1, 1, bias=True)
        self.conv_last = nn.Conv2d(nc, out_nc, 3, 1, 1, bias=True)

        # activation function
        self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)

        # initialization
        initialize_weights([self.conv_first, self.upconv1, self.HRconv, self.conv_last], 0.1)
        if self.upscale == 4:
            initialize_weights(self.upconv2, 0.1)

    def forward(self, x):
        fea = self.lrelu(self.conv_first(x))
        out = self.recon_trunk(fea)

        if self.upscale == 4:
            out = self.lrelu(self.pixel_shuffle(self.upconv1(out)))
            out = self.lrelu(self.pixel_shuffle(self.upconv2(out)))
        elif self.upscale == 3 or self.upscale == 2:
            out = self.lrelu(self.pixel_shuffle(self.upconv1(out)))

        out = self.conv_last(self.lrelu(self.HRconv(out)))
        base = F.interpolate(x, scale_factor=self.upscale, mode='bilinear', align_corners=False)
        out += base
        return out


def initialize_weights(net_l, scale=1):
    if not isinstance(net_l, list):
        net_l = [net_l]
    for net in net_l:
        for m in net.modules():
            if isinstance(m, nn.Conv2d):
                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
                m.weight.data *= scale  # for residual block
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                init.kaiming_normal_(m.weight, a=0, mode='fan_in')
                m.weight.data *= scale
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                init.constant_(m.weight, 1)
                init.constant_(m.bias.data, 0.0)


def make_layer(block, n_layers):
    layers = []
    for _ in range(n_layers):
        layers.append(block())
    return nn.Sequential(*layers)


class ResidualBlock_noBN(nn.Module):
    '''Residual block w/o BN
    ---Conv-ReLU-Conv-+-
     |________________|
    '''

    def __init__(self, nc=64):
        super(ResidualBlock_noBN, self).__init__()
        self.conv1 = nn.Conv2d(nc, nc, 3, 1, 1, bias=True)
        self.conv2 = nn.Conv2d(nc, nc, 3, 1, 1, bias=True)

        # initialization
        initialize_weights([self.conv1, self.conv2], 0.1)

    def forward(self, x):
        identity = x
        out = F.relu(self.conv1(x), inplace=True)
        out = self.conv2(out)
        return identity + out