File size: 11,208 Bytes
a3d6c18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
"""
Modified by Nikita Selin (OPHoperHPO)[https://github.com/OPHoperHPO].
Source url: https://github.com/MarcoForte/FBA_Matting
License: MIT License
"""
import torch
import torch.nn as nn
import carvekit.ml.arch.fba_matting.resnet_GN_WS as resnet_GN_WS
import carvekit.ml.arch.fba_matting.layers_WS as L
import carvekit.ml.arch.fba_matting.resnet_bn as resnet_bn
from functools import partial


class FBA(nn.Module):
    def __init__(self, encoder: str):
        super(FBA, self).__init__()
        self.encoder = build_encoder(arch=encoder)
        self.decoder = fba_decoder(batch_norm=True if "BN" in encoder else False)

    def forward(self, image, two_chan_trimap, image_n, trimap_transformed):
        resnet_input = torch.cat((image_n, trimap_transformed, two_chan_trimap), 1)
        conv_out, indices = self.encoder(resnet_input, return_feature_maps=True)
        return self.decoder(conv_out, image, indices, two_chan_trimap)


class ResnetDilatedBN(nn.Module):
    def __init__(self, orig_resnet, dilate_scale=8):
        super(ResnetDilatedBN, self).__init__()

        if dilate_scale == 8:
            orig_resnet.layer3.apply(partial(self._nostride_dilate, dilate=2))
            orig_resnet.layer4.apply(partial(self._nostride_dilate, dilate=4))
        elif dilate_scale == 16:
            orig_resnet.layer4.apply(partial(self._nostride_dilate, dilate=2))

        # take pretrained resnet, except AvgPool and FC
        self.conv1 = orig_resnet.conv1
        self.bn1 = orig_resnet.bn1
        self.relu1 = orig_resnet.relu1
        self.conv2 = orig_resnet.conv2
        self.bn2 = orig_resnet.bn2
        self.relu2 = orig_resnet.relu2
        self.conv3 = orig_resnet.conv3
        self.bn3 = orig_resnet.bn3
        self.relu3 = orig_resnet.relu3
        self.maxpool = orig_resnet.maxpool
        self.layer1 = orig_resnet.layer1
        self.layer2 = orig_resnet.layer2
        self.layer3 = orig_resnet.layer3
        self.layer4 = orig_resnet.layer4

    def _nostride_dilate(self, m, dilate):
        classname = m.__class__.__name__
        if classname.find("Conv") != -1:
            # the convolution with stride
            if m.stride == (2, 2):
                m.stride = (1, 1)
                if m.kernel_size == (3, 3):
                    m.dilation = (dilate // 2, dilate // 2)
                    m.padding = (dilate // 2, dilate // 2)
            # other convoluions
            else:
                if m.kernel_size == (3, 3):
                    m.dilation = (dilate, dilate)
                    m.padding = (dilate, dilate)

    def forward(self, x, return_feature_maps=False):
        conv_out = [x]
        x = self.relu1(self.bn1(self.conv1(x)))
        x = self.relu2(self.bn2(self.conv2(x)))
        x = self.relu3(self.bn3(self.conv3(x)))
        conv_out.append(x)
        x, indices = self.maxpool(x)
        x = self.layer1(x)
        conv_out.append(x)
        x = self.layer2(x)
        conv_out.append(x)
        x = self.layer3(x)
        conv_out.append(x)
        x = self.layer4(x)
        conv_out.append(x)

        if return_feature_maps:
            return conv_out, indices
        return [x]


class Resnet(nn.Module):
    def __init__(self, orig_resnet):
        super(Resnet, self).__init__()

        # take pretrained resnet, except AvgPool and FC
        self.conv1 = orig_resnet.conv1
        self.bn1 = orig_resnet.bn1
        self.relu1 = orig_resnet.relu1
        self.conv2 = orig_resnet.conv2
        self.bn2 = orig_resnet.bn2
        self.relu2 = orig_resnet.relu2
        self.conv3 = orig_resnet.conv3
        self.bn3 = orig_resnet.bn3
        self.relu3 = orig_resnet.relu3
        self.maxpool = orig_resnet.maxpool
        self.layer1 = orig_resnet.layer1
        self.layer2 = orig_resnet.layer2
        self.layer3 = orig_resnet.layer3
        self.layer4 = orig_resnet.layer4

    def forward(self, x, return_feature_maps=False):
        conv_out = []

        x = self.relu1(self.bn1(self.conv1(x)))
        x = self.relu2(self.bn2(self.conv2(x)))
        x = self.relu3(self.bn3(self.conv3(x)))
        conv_out.append(x)
        x, indices = self.maxpool(x)

        x = self.layer1(x)
        conv_out.append(x)
        x = self.layer2(x)
        conv_out.append(x)
        x = self.layer3(x)
        conv_out.append(x)
        x = self.layer4(x)
        conv_out.append(x)

        if return_feature_maps:
            return conv_out
        return [x]


class ResnetDilated(nn.Module):
    def __init__(self, orig_resnet, dilate_scale=8):
        super(ResnetDilated, self).__init__()

        if dilate_scale == 8:
            orig_resnet.layer3.apply(partial(self._nostride_dilate, dilate=2))
            orig_resnet.layer4.apply(partial(self._nostride_dilate, dilate=4))
        elif dilate_scale == 16:
            orig_resnet.layer4.apply(partial(self._nostride_dilate, dilate=2))

        # take pretrained resnet, except AvgPool and FC
        self.conv1 = orig_resnet.conv1
        self.bn1 = orig_resnet.bn1
        self.relu = orig_resnet.relu
        self.maxpool = orig_resnet.maxpool
        self.layer1 = orig_resnet.layer1
        self.layer2 = orig_resnet.layer2
        self.layer3 = orig_resnet.layer3
        self.layer4 = orig_resnet.layer4

    def _nostride_dilate(self, m, dilate):
        classname = m.__class__.__name__
        if classname.find("Conv") != -1:
            # the convolution with stride
            if m.stride == (2, 2):
                m.stride = (1, 1)
                if m.kernel_size == (3, 3):
                    m.dilation = (dilate // 2, dilate // 2)
                    m.padding = (dilate // 2, dilate // 2)
            # other convoluions
            else:
                if m.kernel_size == (3, 3):
                    m.dilation = (dilate, dilate)
                    m.padding = (dilate, dilate)

    def forward(self, x, return_feature_maps=False):
        conv_out = [x]
        x = self.relu(self.bn1(self.conv1(x)))
        conv_out.append(x)
        x, indices = self.maxpool(x)
        x = self.layer1(x)
        conv_out.append(x)
        x = self.layer2(x)
        conv_out.append(x)
        x = self.layer3(x)
        conv_out.append(x)
        x = self.layer4(x)
        conv_out.append(x)

        if return_feature_maps:
            return conv_out, indices
        return [x]


def norm(dim, bn=False):
    if bn is False:
        return nn.GroupNorm(32, dim)
    else:
        return nn.BatchNorm2d(dim)


def fba_fusion(alpha, img, F, B):
    F = alpha * img + (1 - alpha**2) * F - alpha * (1 - alpha) * B
    B = (1 - alpha) * img + (2 * alpha - alpha**2) * B - alpha * (1 - alpha) * F

    F = torch.clamp(F, 0, 1)
    B = torch.clamp(B, 0, 1)
    la = 0.1
    alpha = (alpha * la + torch.sum((img - B) * (F - B), 1, keepdim=True)) / (
        torch.sum((F - B) * (F - B), 1, keepdim=True) + la
    )
    alpha = torch.clamp(alpha, 0, 1)
    return alpha, F, B


class fba_decoder(nn.Module):
    def __init__(self, batch_norm=False):
        super(fba_decoder, self).__init__()
        pool_scales = (1, 2, 3, 6)
        self.batch_norm = batch_norm

        self.ppm = []

        for scale in pool_scales:
            self.ppm.append(
                nn.Sequential(
                    nn.AdaptiveAvgPool2d(scale),
                    L.Conv2d(2048, 256, kernel_size=1, bias=True),
                    norm(256, self.batch_norm),
                    nn.LeakyReLU(),
                )
            )
        self.ppm = nn.ModuleList(self.ppm)

        self.conv_up1 = nn.Sequential(
            L.Conv2d(
                2048 + len(pool_scales) * 256, 256, kernel_size=3, padding=1, bias=True
            ),
            norm(256, self.batch_norm),
            nn.LeakyReLU(),
            L.Conv2d(256, 256, kernel_size=3, padding=1),
            norm(256, self.batch_norm),
            nn.LeakyReLU(),
        )

        self.conv_up2 = nn.Sequential(
            L.Conv2d(256 + 256, 256, kernel_size=3, padding=1, bias=True),
            norm(256, self.batch_norm),
            nn.LeakyReLU(),
        )
        if self.batch_norm:
            d_up3 = 128
        else:
            d_up3 = 64
        self.conv_up3 = nn.Sequential(
            L.Conv2d(256 + d_up3, 64, kernel_size=3, padding=1, bias=True),
            norm(64, self.batch_norm),
            nn.LeakyReLU(),
        )

        self.unpool = nn.MaxUnpool2d(2, stride=2)

        self.conv_up4 = nn.Sequential(
            nn.Conv2d(64 + 3 + 3 + 2, 32, kernel_size=3, padding=1, bias=True),
            nn.LeakyReLU(),
            nn.Conv2d(32, 16, kernel_size=3, padding=1, bias=True),
            nn.LeakyReLU(),
            nn.Conv2d(16, 7, kernel_size=1, padding=0, bias=True),
        )

    def forward(self, conv_out, img, indices, two_chan_trimap):
        conv5 = conv_out[-1]

        input_size = conv5.size()
        ppm_out = [conv5]
        for pool_scale in self.ppm:
            ppm_out.append(
                nn.functional.interpolate(
                    pool_scale(conv5),
                    (input_size[2], input_size[3]),
                    mode="bilinear",
                    align_corners=False,
                )
            )
        ppm_out = torch.cat(ppm_out, 1)
        x = self.conv_up1(ppm_out)

        x = torch.nn.functional.interpolate(
            x, scale_factor=2, mode="bilinear", align_corners=False
        )

        x = torch.cat((x, conv_out[-4]), 1)

        x = self.conv_up2(x)
        x = torch.nn.functional.interpolate(
            x, scale_factor=2, mode="bilinear", align_corners=False
        )

        x = torch.cat((x, conv_out[-5]), 1)
        x = self.conv_up3(x)

        x = torch.nn.functional.interpolate(
            x, scale_factor=2, mode="bilinear", align_corners=False
        )
        x = torch.cat((x, conv_out[-6][:, :3], img, two_chan_trimap), 1)

        output = self.conv_up4(x)

        alpha = torch.clamp(output[:, 0][:, None], 0, 1)
        F = torch.sigmoid(output[:, 1:4])
        B = torch.sigmoid(output[:, 4:7])

        # FBA Fusion
        alpha, F, B = fba_fusion(alpha, img, F, B)

        output = torch.cat((alpha, F, B), 1)

        return output


def build_encoder(arch="resnet50_GN"):
    if arch == "resnet50_GN_WS":
        orig_resnet = resnet_GN_WS.__dict__["l_resnet50"]()
        net_encoder = ResnetDilated(orig_resnet, dilate_scale=8)
    elif arch == "resnet50_BN":
        orig_resnet = resnet_bn.__dict__["l_resnet50"]()
        net_encoder = ResnetDilatedBN(orig_resnet, dilate_scale=8)

    else:
        raise ValueError("Architecture undefined!")

    num_channels = 3 + 6 + 2

    if num_channels > 3:
        net_encoder_sd = net_encoder.state_dict()
        conv1_weights = net_encoder_sd["conv1.weight"]

        c_out, c_in, h, w = conv1_weights.size()
        conv1_mod = torch.zeros(c_out, num_channels, h, w)
        conv1_mod[:, :3, :, :] = conv1_weights

        conv1 = net_encoder.conv1
        conv1.in_channels = num_channels
        conv1.weight = torch.nn.Parameter(conv1_mod)

        net_encoder.conv1 = conv1

        net_encoder_sd["conv1.weight"] = conv1_mod

        net_encoder.load_state_dict(net_encoder_sd)
    return net_encoder