File size: 12,012 Bytes
7c30e0a
 
 
 
c34ed4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
This code was mostly taken from backbone-unet by mkisantal:
https://github.com/mkisantal/backboned-unet/blob/master/backboned_unet/unet.py
"""
import torch
import torch.nn as nn
from torchvision import models
from torch.nn import functional as F

import torch.nn as nn
import torch
from torchvision import models


class AdaptiveConcatPool2d(nn.Module):
    """
    Layer that concats `AdaptiveAvgPool2d` and `AdaptiveMaxPool2d`.
    Source: Fastai. This code was taken from the fastai library at url
    https://github.com/fastai/fastai/blob/master/fastai/layers.py#L176
    """

    def __init__(self, sz=None):
        "Output will be 2*sz or 2 if sz is None"
        super().__init__()
        self.output_size = sz or 1
        self.ap = nn.AdaptiveAvgPool2d(self.output_size)
        self.mp = nn.AdaptiveMaxPool2d(self.output_size)

    def forward(self, x): return torch.cat([self.mp(x), self.ap(x)], 1)


class MyNorm(nn.Module):
    def __init__(self, num_channels):
        super(MyNorm, self).__init__()
        self.norm = nn.InstanceNorm2d(
            num_channels, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)

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


def resnet_fastai(model, pretrained,  url, replace_first_layer=None, replace_maxpool_layer=None, progress=True, map_location=None, **kwargs):
    cut = -2
    s = model(pretrained=False, **kwargs)
    if replace_maxpool_layer is not None:
        s.maxpool = replace_maxpool_layer
    if replace_first_layer is not None:
        body = nn.Sequential(replace_first_layer, *list(s.children())[1:cut])
    else:
        body = nn.Sequential(*list(s.children())[:cut])

    if pretrained:
        state = torch.hub.load_state_dict_from_url(url,
                                                   progress=progress, map_location=map_location)
        if replace_first_layer is not None:
            for each in list(state.keys()).copy():
                if each.find("0.0.") == 0:
                    del state[each]
        body_tail = nn.Sequential(body)
        ret = body_tail.load_state_dict(state, strict=False)
    return body


def get_backbone(name, pretrained=True, map_location=None):
    """ Loading backbone, defining names for skip-connections and encoder output. """

    first_layer_for_4chn = nn.Conv2d(
        4, 64, kernel_size=7, stride=2, padding=3, bias=False)
    max_pool_layer_replace = nn.Conv2d(
        64, 64, kernel_size=3, stride=2, padding=1, bias=False)
    # loading backbone model
    if name == 'resnet18':
        backbone = models.resnet18(pretrained=pretrained)
    if name == 'resnet18-4':
        backbone = models.resnet18(pretrained=pretrained)
        backbone.conv1 = first_layer_for_4chn
    elif name == 'resnet34':
        backbone = models.resnet34(pretrained=pretrained)
    elif name == 'resnet50':
        backbone = models.resnet50(pretrained=False, norm_layer=MyNorm)
        backbone.maxpool = max_pool_layer_replace
    elif name == 'resnet101':
        backbone = models.resnet101(pretrained=pretrained)
    elif name == 'resnet152':
        backbone = models.resnet152(pretrained=pretrained)
    elif name == 'vgg16':
        backbone = models.vgg16_bn(pretrained=pretrained).features
    elif name == 'vgg19':
        backbone = models.vgg19_bn(pretrained=pretrained).features
    elif name == 'resnet18_danbo-4':
        backbone = resnet_fastai(models.resnet18, url="https://github.com/RF5/danbooru-pretrained/releases/download/v0.1/resnet18-3f77756f.pth",
                                 pretrained=pretrained, map_location=map_location, norm_layer=MyNorm, replace_first_layer=first_layer_for_4chn)
    elif name == 'resnet50_danbo':
        backbone = resnet_fastai(models.resnet50, url="https://github.com/RF5/danbooru-pretrained/releases/download/v0.1/resnet50-13306192.pth",
                                 pretrained=pretrained, map_location=map_location, norm_layer=MyNorm, replace_maxpool_layer=max_pool_layer_replace)
    elif name == 'densenet121':
        backbone = models.densenet121(pretrained=True).features
    elif name == 'densenet161':
        backbone = models.densenet161(pretrained=True).features
    elif name == 'densenet169':
        backbone = models.densenet169(pretrained=True).features
    elif name == 'densenet201':
        backbone = models.densenet201(pretrained=True).features
    else:
        raise NotImplemented(
            '{} backbone model is not implemented so far.'.format(name))
    #print(backbone)
    # specifying skip feature and output names
    if name.startswith('resnet'):
        feature_names = [None, 'relu', 'layer1', 'layer2', 'layer3']
        backbone_output = 'layer4'
    elif name == 'vgg16':
        # TODO: consider using a 'bridge' for VGG models, there is just a MaxPool between last skip and backbone output
        feature_names = ['5', '12', '22', '32', '42']
        backbone_output = '43'
    elif name == 'vgg19':
        feature_names = ['5', '12', '25', '38', '51']
        backbone_output = '52'
    elif name.startswith('densenet'):
        feature_names = [None, 'relu0', 'denseblock1',
                         'denseblock2', 'denseblock3']
        backbone_output = 'denseblock4'
    elif name == 'unet_encoder':
        feature_names = ['module1', 'module2', 'module3', 'module4']
        backbone_output = 'module5'
    else:
        raise NotImplemented(
            '{} backbone model is not implemented so far.'.format(name))
    if name.find('_danbo') > 0:
        feature_names = [None, '2', '4', '5', '6']
        backbone_output = '7'
    return backbone, feature_names, backbone_output


class UpsampleBlock(nn.Module):

    # TODO: separate parametric and non-parametric classes?
    # TODO: skip connection concatenated OR added

    def __init__(self, ch_in, ch_out=None, skip_in=0, use_bn=True, parametric=False):
        super(UpsampleBlock, self).__init__()

        self.parametric = parametric
        ch_out = ch_in/2 if ch_out is None else ch_out

        # first convolution: either transposed conv, or conv following the skip connection
        if parametric:
            # versions: kernel=4 padding=1, kernel=2 padding=0
            self.up = nn.ConvTranspose2d(in_channels=ch_in, out_channels=ch_out, kernel_size=(4, 4),
                                         stride=2, padding=1, output_padding=0, bias=(not use_bn))
            self.bn1 = MyNorm(ch_out) if use_bn else None
        else:
            self.up = None
            ch_in = ch_in + skip_in
            self.conv1 = nn.Conv2d(in_channels=ch_in, out_channels=ch_out, kernel_size=(3, 3),
                                   stride=1, padding=1, bias=(not use_bn))
            self.bn1 = MyNorm(ch_out) if use_bn else None

        self.relu = nn.ReLU(inplace=True)

        # second convolution
        conv2_in = ch_out if not parametric else ch_out + skip_in
        self.conv2 = nn.Conv2d(in_channels=conv2_in, out_channels=ch_out, kernel_size=(3, 3),
                               stride=1, padding=1, bias=(not use_bn))
        self.bn2 = MyNorm(ch_out) if use_bn else None

    def forward(self, x, skip_connection=None):

        x = self.up(x) if self.parametric else F.interpolate(x, size=None, scale_factor=2, mode='bilinear',
                                                             align_corners=None)
        if self.parametric:
            x = self.bn1(x) if self.bn1 is not None else x
            x = self.relu(x)

        if skip_connection is not None:
            x = torch.cat([x, skip_connection], dim=1)

        if not self.parametric:
            x = self.conv1(x)
            x = self.bn1(x) if self.bn1 is not None else x
            x = self.relu(x)
        x = self.conv2(x)
        x = self.bn2(x) if self.bn2 is not None else x
        x = self.relu(x)

        return x


class ResEncUnet(nn.Module):

    """ U-Net (https://arxiv.org/pdf/1505.04597.pdf) implementation with pre-trained torchvision backbones."""

    def __init__(self,
                 backbone_name,
                 pretrained=True,
                 encoder_freeze=False,
                 classes=21,
                 decoder_filters=(512, 256, 128, 64, 32),
                 parametric_upsampling=True,
                 shortcut_features='default',
                 decoder_use_instancenorm=True,
                 map_location=None
                 ):
        super(ResEncUnet, self).__init__()

        self.backbone_name = backbone_name

        self.backbone, self.shortcut_features, self.bb_out_name = get_backbone(
            backbone_name, pretrained=pretrained, map_location=map_location)
        shortcut_chs, bb_out_chs = self.infer_skip_channels()
        if shortcut_features != 'default':
            self.shortcut_features = shortcut_features

        # build decoder part
        self.upsample_blocks = nn.ModuleList()
        # avoiding having more blocks than skip connections
        decoder_filters = decoder_filters[:len(self.shortcut_features)]
        decoder_filters_in = [bb_out_chs] + list(decoder_filters[:-1])
        num_blocks = len(self.shortcut_features)
        for i, [filters_in, filters_out] in enumerate(zip(decoder_filters_in, decoder_filters)):
            self.upsample_blocks.append(UpsampleBlock(filters_in, filters_out,
                                                      skip_in=shortcut_chs[num_blocks-i-1],
                                                      parametric=parametric_upsampling,
                                                      use_bn=decoder_use_instancenorm))
        self.final_conv = nn.Conv2d(
            decoder_filters[-1], classes, kernel_size=(1, 1))

        if encoder_freeze:
            self.freeze_encoder()

    def freeze_encoder(self):
        """ Freezing encoder parameters, the newly initialized decoder parameters are remaining trainable. """

        for param in self.backbone.parameters():
            param.requires_grad = False

    def forward(self, *input, ret_parser_out=True):
        """ Forward propagation in U-Net. """

        x, features = self.forward_backbone(*input)
        output_feature = [x]
        for skip_name, upsample_block in zip(self.shortcut_features[::-1], self.upsample_blocks):
            skip_features = features[skip_name]
            if skip_features is not None:
                output_feature.append(skip_features)
            if ret_parser_out:
                x = upsample_block(x, skip_features)
        if ret_parser_out:
            x = self.final_conv(x)
            # apply sigmoid later
        else:
            x = None

        return x, output_feature

    def forward_backbone(self, x):
        """ Forward propagation in backbone encoder network.  """

        features = {None: None} if None in self.shortcut_features else dict()
        for name, child in self.backbone.named_children():
            x = child(x)
            if name in self.shortcut_features:
                features[name] = x
            if name == self.bb_out_name:
                break

        return x, features

    def infer_skip_channels(self):
        """ Getting the number of channels at skip connections and at the output of the encoder. """
        if self.backbone_name.find("-4") > 0:
            x = torch.zeros(1, 4, 224, 224)
        else:
            x = torch.zeros(1, 3, 224, 224)
        has_fullres_features = self.backbone_name.startswith(
            'vgg') or self.backbone_name == 'unet_encoder'
        # only VGG has features at full resolution
        channels = [] if has_fullres_features else [0]

        # forward run in backbone to count channels (dirty solution but works for *any* Module)
        for name, child in self.backbone.named_children():
            x = child(x)
            if name in self.shortcut_features:
                channels.append(x.shape[1])
            if name == self.bb_out_name:
                out_channels = x.shape[1]
                break
        return channels, out_channels