File size: 9,753 Bytes
da406d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
import torch
import torch.nn as nn
import torch.nn.functional as F

from .backbones import SUPPORTED_BACKBONES


#------------------------------------------------------------------------------
#  MODNet Basic Modules
#------------------------------------------------------------------------------

class IBNorm(nn.Module):
    """ Combine Instance Norm and Batch Norm into One Layer
    """

    def __init__(self, in_channels):
        super(IBNorm, self).__init__()
        in_channels = in_channels
        self.bnorm_channels = int(in_channels / 2)
        self.inorm_channels = in_channels - self.bnorm_channels

        self.bnorm = nn.BatchNorm2d(self.bnorm_channels, affine=True)
        self.inorm = nn.InstanceNorm2d(self.inorm_channels, affine=False)
        
    def forward(self, x):
        bn_x = self.bnorm(x[:, :self.bnorm_channels, ...].contiguous())
        in_x = self.inorm(x[:, self.bnorm_channels:, ...].contiguous())

        return torch.cat((bn_x, in_x), 1)


class Conv2dIBNormRelu(nn.Module):
    """ Convolution + IBNorm + ReLu
    """

    def __init__(self, in_channels, out_channels, kernel_size, 
                 stride=1, padding=0, dilation=1, groups=1, bias=True, 
                 with_ibn=True, with_relu=True):
        super(Conv2dIBNormRelu, self).__init__()

        layers = [
            nn.Conv2d(in_channels, out_channels, kernel_size, 
                      stride=stride, padding=padding, dilation=dilation, 
                      groups=groups, bias=bias)
        ]

        if with_ibn:       
            layers.append(IBNorm(out_channels))
        if with_relu:
            layers.append(nn.ReLU(inplace=True))

        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x) 


class SEBlock(nn.Module):
    """ SE Block Proposed in https://arxiv.org/pdf/1709.01507.pdf 
    """

    def __init__(self, in_channels, out_channels, reduction=1):
        super(SEBlock, self).__init__()
        self.pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(in_channels, int(in_channels // reduction), bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(int(in_channels // reduction), out_channels, bias=False),
            nn.Sigmoid()
        )
    
    def forward(self, x):
        b, c, _, _ = x.size()
        w = self.pool(x).view(b, c)
        w = self.fc(w).view(b, c, 1, 1)

        return x * w.expand_as(x)


#------------------------------------------------------------------------------
#  MODNet Branches
#------------------------------------------------------------------------------

class LRBranch(nn.Module):
    """ Low Resolution Branch of MODNet
    """

    def __init__(self, backbone):
        super(LRBranch, self).__init__()

        enc_channels = backbone.enc_channels
        
        self.backbone = backbone
        self.se_block = SEBlock(enc_channels[4], enc_channels[4], reduction=4)
        self.conv_lr16x = Conv2dIBNormRelu(enc_channels[4], enc_channels[3], 5, stride=1, padding=2)
        self.conv_lr8x = Conv2dIBNormRelu(enc_channels[3], enc_channels[2], 5, stride=1, padding=2)
        self.conv_lr = Conv2dIBNormRelu(enc_channels[2], 1, kernel_size=3, stride=2, padding=1, with_ibn=False, with_relu=False)

    def forward(self, img, inference):
        enc_features = self.backbone.forward(img)
        enc2x, enc4x, enc32x = enc_features[0], enc_features[1], enc_features[4]

        enc32x = self.se_block(enc32x)
        lr16x = F.interpolate(enc32x, scale_factor=2, mode='bilinear', align_corners=False)
        lr16x = self.conv_lr16x(lr16x)
        lr8x = F.interpolate(lr16x, scale_factor=2, mode='bilinear', align_corners=False)
        lr8x = self.conv_lr8x(lr8x)

        pred_semantic = None
        if not inference:
            lr = self.conv_lr(lr8x)
            pred_semantic = torch.sigmoid(lr)

        return pred_semantic, lr8x, [enc2x, enc4x] 


class HRBranch(nn.Module):
    """ High Resolution Branch of MODNet
    """

    def __init__(self, hr_channels, enc_channels):
        super(HRBranch, self).__init__()

        self.tohr_enc2x = Conv2dIBNormRelu(enc_channels[0], hr_channels, 1, stride=1, padding=0)
        self.conv_enc2x = Conv2dIBNormRelu(hr_channels + 3, hr_channels, 3, stride=2, padding=1)

        self.tohr_enc4x = Conv2dIBNormRelu(enc_channels[1], hr_channels, 1, stride=1, padding=0)
        self.conv_enc4x = Conv2dIBNormRelu(2 * hr_channels, 2 * hr_channels, 3, stride=1, padding=1)

        self.conv_hr4x = nn.Sequential(
            Conv2dIBNormRelu(3 * hr_channels + 3, 2 * hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(2 * hr_channels, 2 * hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(2 * hr_channels, hr_channels, 3, stride=1, padding=1),
        )

        self.conv_hr2x = nn.Sequential(
            Conv2dIBNormRelu(2 * hr_channels, 2 * hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(2 * hr_channels, hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(hr_channels, hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(hr_channels, hr_channels, 3, stride=1, padding=1),
        )

        self.conv_hr = nn.Sequential(
            Conv2dIBNormRelu(hr_channels + 3, hr_channels, 3, stride=1, padding=1),
            Conv2dIBNormRelu(hr_channels, 1, kernel_size=1, stride=1, padding=0, with_ibn=False, with_relu=False),
        )

    def forward(self, img, enc2x, enc4x, lr8x, inference):
        img2x = F.interpolate(img, scale_factor=1/2, mode='bilinear', align_corners=False)
        img4x = F.interpolate(img, scale_factor=1/4, mode='bilinear', align_corners=False)

        enc2x = self.tohr_enc2x(enc2x)
        hr4x = self.conv_enc2x(torch.cat((img2x, enc2x), dim=1))

        enc4x = self.tohr_enc4x(enc4x)
        hr4x = self.conv_enc4x(torch.cat((hr4x, enc4x), dim=1))

        lr4x = F.interpolate(lr8x, scale_factor=2, mode='bilinear', align_corners=False)
        hr4x = self.conv_hr4x(torch.cat((hr4x, lr4x, img4x), dim=1))

        hr2x = F.interpolate(hr4x, scale_factor=2, mode='bilinear', align_corners=False)
        hr2x = self.conv_hr2x(torch.cat((hr2x, enc2x), dim=1))

        pred_detail = None
        if not inference:
            hr = F.interpolate(hr2x, scale_factor=2, mode='bilinear', align_corners=False)
            hr = self.conv_hr(torch.cat((hr, img), dim=1))
            pred_detail = torch.sigmoid(hr)

        return pred_detail, hr2x


class FusionBranch(nn.Module):
    """ Fusion Branch of MODNet
    """

    def __init__(self, hr_channels, enc_channels):
        super(FusionBranch, self).__init__()
        self.conv_lr4x = Conv2dIBNormRelu(enc_channels[2], hr_channels, 5, stride=1, padding=2)
        
        self.conv_f2x = Conv2dIBNormRelu(2 * hr_channels, hr_channels, 3, stride=1, padding=1)
        self.conv_f = nn.Sequential(
            Conv2dIBNormRelu(hr_channels + 3, int(hr_channels / 2), 3, stride=1, padding=1),
            Conv2dIBNormRelu(int(hr_channels / 2), 1, 1, stride=1, padding=0, with_ibn=False, with_relu=False),
        )

    def forward(self, img, lr8x, hr2x):
        lr4x = F.interpolate(lr8x, scale_factor=2, mode='bilinear', align_corners=False)
        lr4x = self.conv_lr4x(lr4x)
        lr2x = F.interpolate(lr4x, scale_factor=2, mode='bilinear', align_corners=False)

        f2x = self.conv_f2x(torch.cat((lr2x, hr2x), dim=1))
        f = F.interpolate(f2x, scale_factor=2, mode='bilinear', align_corners=False)
        f = self.conv_f(torch.cat((f, img), dim=1))
        pred_matte = torch.sigmoid(f)

        return pred_matte


#------------------------------------------------------------------------------
#  MODNet
#------------------------------------------------------------------------------

class MODNet(nn.Module):
    """ Architecture of MODNet
    """

    def __init__(self, in_channels=3, hr_channels=32, backbone_arch='mobilenetv2', backbone_pretrained=True):
        super(MODNet, self).__init__()

        self.in_channels = in_channels
        self.hr_channels = hr_channels
        self.backbone_arch = backbone_arch
        self.backbone_pretrained = backbone_pretrained

        self.backbone = SUPPORTED_BACKBONES[self.backbone_arch](self.in_channels)

        self.lr_branch = LRBranch(self.backbone)
        self.hr_branch = HRBranch(self.hr_channels, self.backbone.enc_channels)
        self.f_branch = FusionBranch(self.hr_channels, self.backbone.enc_channels)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                self._init_conv(m)
            elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.InstanceNorm2d):
                self._init_norm(m)

        if self.backbone_pretrained:
            self.backbone.load_pretrained_ckpt()                

    def forward(self, img, inference):
        pred_semantic, lr8x, [enc2x, enc4x] = self.lr_branch(img, inference)
        pred_detail, hr2x = self.hr_branch(img, enc2x, enc4x, lr8x, inference)
        pred_matte = self.f_branch(img, lr8x, hr2x)

        return pred_semantic, pred_detail, pred_matte
    
    def freeze_norm(self):
        norm_types = [nn.BatchNorm2d, nn.InstanceNorm2d]
        for m in self.modules():
            for n in norm_types:
                if isinstance(m, n):
                    m.eval()
                    continue

    def _init_conv(self, conv):
        nn.init.kaiming_uniform_(
            conv.weight, a=0, mode='fan_in', nonlinearity='relu')
        if conv.bias is not None:
            nn.init.constant_(conv.bias, 0)

    def _init_norm(self, norm):
        if norm.weight is not None:
            nn.init.constant_(norm.weight, 1)
            nn.init.constant_(norm.bias, 0)