File size: 11,943 Bytes
dae5c90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as model_zoo
from itertools import chain
import torch.utils.checkpoint as cp
from collections import defaultdict
from math import log2
from timm.models.registry import register_model

from ..utils.utils import _UpsampleBlend

__all__ = ['ResNet', 'resnet18', 'resnet34', 'BasicBlock']

model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}


def convkxk(in_planes, out_planes, stride=1, k=3):
    """kxk convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=k, stride=stride, padding=k // 2, bias=False)


def _bn_function_factory(conv, norm, relu=None):
    def bn_function(x):
        x = norm(conv(x))
        if relu is not None:
            x = relu(x)
        return x

    return bn_function


def do_efficient_fwd(block, x, efficient):
    # return block(x)
    if efficient and x.requires_grad:
        return cp.checkpoint(block, x)
    else:
        return block(x)


class Identity(nn.Module):
    def __init__(self, *args, **kwargs):
        super(Identity, self).__init__()

    def forward(self, input):
        return input


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None, efficient=True, bn_class=nn.BatchNorm2d, levels=3):
        super(BasicBlock, self).__init__()
        self.conv1 = convkxk(inplanes, planes, stride)
        self.bn1 = nn.ModuleList([bn_class(planes) for _ in range(levels)])
        self.relu_inp = nn.ReLU(inplace=True)
        self.relu = nn.ReLU(inplace=False)
        self.conv2 = convkxk(planes, planes)
        self.bn2 = nn.ModuleList([bn_class(planes) for _ in range(levels)])
        self.downsample = downsample
        self.stride = stride
        self.efficient = efficient
        self.num_levels = levels

    def forward(self, x, level):
        residual = x

        bn_1 = _bn_function_factory(self.conv1, self.bn1[level], self.relu_inp)
        bn_2 = _bn_function_factory(self.conv2, self.bn2[level])

        out = do_efficient_fwd(bn_1, x, self.efficient)
        out = do_efficient_fwd(bn_2, out, self.efficient)

        if self.downsample is not None:
            residual = self.downsample(x)

        out += residual
        relu = self.relu(out)

        return relu, out

    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,

                              missing_keys, unexpected_keys, error_msgs):
        super(BasicBlock, self)._load_from_state_dict(state_dict, prefix, local_metadata, False, missing_keys,
                                                      unexpected_keys, error_msgs)
        missing_keys = []
        unexpected_keys = []
        for bn in self.bn1:
            bn._load_from_state_dict(state_dict, prefix + 'bn1.', local_metadata, strict, missing_keys, unexpected_keys,
                                     error_msgs)
        for bn in self.bn2:
            bn._load_from_state_dict(state_dict, prefix + 'bn2.', local_metadata, strict, missing_keys, unexpected_keys,
                                     error_msgs)


class ResNet(nn.Module):
    def _make_layer(self, block, planes, blocks, stride=1, bn_class=nn.BatchNorm2d):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                bn_class(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, self.efficient, bn_class=bn_class,
                            levels=self.pyramid_levels))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, bn_class=bn_class, levels=self.pyramid_levels, efficient=self.efficient))

        return nn.Sequential(*layers)

    def __init__(self, block, layers, *, num_features=128, pyramid_levels=3, use_bn=True, k_bneck=1, k_upsample=3,

                 efficient=False, upsample_skip=True, mean=(73.1584, 82.9090, 72.3924),

                 std=(44.9149, 46.1529, 45.3192), scale=1, detach_upsample_skips=(), detach_upsample_in=False,

                 align_corners=None, pyramid_subsample='bicubic', target_size=None,

                 output_stride=4, **kwargs):
        self.inplanes = 64
        self.efficient = efficient
        super(ResNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
        bn_class = nn.BatchNorm2d if use_bn else Identity
        self.register_buffer('img_mean', torch.tensor(mean).view(1, -1, 1, 1))
        self.register_buffer('img_std', torch.tensor(std).view(1, -1, 1, 1))
        if scale != 1:
            self.register_buffer('img_scale', torch.tensor(scale).view(1, -1, 1, 1).float())

        self.pyramid_levels = pyramid_levels
        self.num_features = num_features
        self.replicated = False

        self.align_corners = align_corners
        self.pyramid_subsample = pyramid_subsample

        self.bn1 = nn.ModuleList([bn_class(64) for _ in range(pyramid_levels)])
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        bottlenecks = []
        self.layer1 = self._make_layer(block, 64, layers[0], bn_class=bn_class)
        bottlenecks += [convkxk(self.inplanes, num_features, k=k_bneck)]
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, bn_class=bn_class)
        bottlenecks += [convkxk(self.inplanes, num_features, k=k_bneck)]
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, bn_class=bn_class)
        bottlenecks += [convkxk(self.inplanes, num_features, k=k_bneck)]
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, bn_class=bn_class)
        bottlenecks += [convkxk(self.inplanes, num_features, k=k_bneck)]

        num_bn_remove = max(0, int(log2(output_stride) - 2))
        self.num_skip_levels = self.pyramid_levels + 3 - num_bn_remove
        bottlenecks = bottlenecks[num_bn_remove:]

        self.fine_tune = [self.conv1, self.maxpool, self.layer1, self.layer2, self.layer3, self.layer4, self.bn1]

        self.upsample_bottlenecks = nn.ModuleList(bottlenecks[::-1])
        num_pyr_modules = 2 + pyramid_levels - num_bn_remove
        self.target_size = target_size
        if self.target_size is not None:
            h, w = target_size
            target_sizes = [(h // 2 ** i, w // 2 ** i) for i in range(2, 2 + num_pyr_modules)][::-1]
        else:
            target_sizes = [None] * num_pyr_modules
        self.upsample_blends = nn.ModuleList(
            [_UpsampleBlend(num_features,
                            use_bn=use_bn,
                            use_skip=upsample_skip,
                            detach_skip=i in detach_upsample_skips,
                            fixed_size=ts,
                            k=k_upsample)
             for i, ts in enumerate(target_sizes)])
        self.detach_upsample_in = detach_upsample_in

        self.random_init = [self.upsample_bottlenecks, self.upsample_blends]

        self.features = num_features

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def random_init_params(self):
        return chain(*[f.parameters() for f in self.random_init])

    def fine_tune_params(self):
        return chain(*[f.parameters() for f in self.fine_tune])

    def forward_resblock(self, x, layers, idx):
        skip = None
        for l in layers:
            x = l(x) if not isinstance(l, BasicBlock) else l(x, idx)
            if isinstance(x, tuple):
                x, skip = x
        return x, skip

    def forward_down(self, image, skips, idx=-1):
        x = self.conv1(image)
        x = self.bn1[idx](x)
        x = self.relu(x)
        x = self.maxpool(x)

        features = []
        x, skip = self.forward_resblock(x, self.layer1, idx)
        features += [skip]
        x, skip = self.forward_resblock(x, self.layer2, idx)
        features += [skip]
        x, skip = self.forward_resblock(x, self.layer3, idx)
        features += [skip]
        x, skip = self.forward_resblock(x, self.layer4, idx)
        features += [skip]

        skip_feats = [b(f) for b, f in zip(self.upsample_bottlenecks, reversed(features))]

        for i, s in enumerate(reversed(skip_feats)):
            skips[idx + i] += [s]

        return skips

    def forward(self, image):
        if isinstance(self.bn1[0], nn.BatchNorm2d):
            if hasattr(self, 'img_scale'):
                image /= self.img_scale
            image -= self.img_mean
            image /= self.img_std
        pyramid = [image]
        for l in range(1, self.pyramid_levels):
            if self.target_size is not None:
                ts = list([si // 2 ** l for si in self.target_size])
                pyramid += [
                    F.interpolate(image, size=ts, mode=self.pyramid_subsample, align_corners=self.align_corners)]
            else:
                pyramid += [F.interpolate(image, scale_factor=1 / 2 ** l, mode=self.pyramid_subsample,
                                          align_corners=self.align_corners)]
        skips = [[] for _ in range(self.num_skip_levels)]
        additional = {'pyramid': pyramid}
        for idx, p in enumerate(pyramid):
            skips = self.forward_down(p, skips, idx=idx)
        skips = skips[::-1]
        x = skips[0][0]
        if self.detach_upsample_in:
            x = x.detach()
        for i, (sk, blend) in enumerate(zip(skips[1:], self.upsample_blends)):
            x = blend(x, sum(sk))
        return x, additional

    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,

                              missing_keys, unexpected_keys, error_msgs):
        super(ResNet, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys,
                                                  unexpected_keys, error_msgs)
        for bn in self.bn1:
            bn._load_from_state_dict(state_dict, prefix + 'bn1.', local_metadata, strict, missing_keys, unexpected_keys,
                                     error_msgs)

@register_model
def resnet18(pretrained=True, **kwargs):
    """Constructs a ResNet-18 model.

    Args:

        pretrained (bool): If True, returns a model pre-trained on ImageNet

    """
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
    return model

@register_model
def resnet34(pretrained=True, **kwargs):
    """Constructs a ResNet-34 model.

    Args:

        pretrained (bool): If True, returns a model pre-trained on ImageNet

    """
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet34']), strict=False)
    return model