File size: 14,436 Bytes
58f667f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
import torch
import torch.nn as nn
from annotator.uniformer.mmcv.cnn import (ConvModule, DepthwiseSeparableConvModule, constant_init,
                      kaiming_init)
from torch.nn.modules.batchnorm import _BatchNorm

from annotator.uniformer.mmseg.models.decode_heads.psp_head import PPM
from annotator.uniformer.mmseg.ops import resize
from ..builder import BACKBONES
from ..utils.inverted_residual import InvertedResidual


class LearningToDownsample(nn.Module):
    """Learning to downsample module.

    Args:
        in_channels (int): Number of input channels.
        dw_channels (tuple[int]): Number of output channels of the first and
            the second depthwise conv (dwconv) layers.
        out_channels (int): Number of output channels of the whole
            'learning to downsample' module.
        conv_cfg (dict | None): Config of conv layers. Default: None
        norm_cfg (dict | None): Config of norm layers. Default:
            dict(type='BN')
        act_cfg (dict): Config of activation layers. Default:
            dict(type='ReLU')
    """

    def __init__(self,
                 in_channels,
                 dw_channels,
                 out_channels,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU')):
        super(LearningToDownsample, self).__init__()
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        dw_channels1 = dw_channels[0]
        dw_channels2 = dw_channels[1]

        self.conv = ConvModule(
            in_channels,
            dw_channels1,
            3,
            stride=2,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)
        self.dsconv1 = DepthwiseSeparableConvModule(
            dw_channels1,
            dw_channels2,
            kernel_size=3,
            stride=2,
            padding=1,
            norm_cfg=self.norm_cfg)
        self.dsconv2 = DepthwiseSeparableConvModule(
            dw_channels2,
            out_channels,
            kernel_size=3,
            stride=2,
            padding=1,
            norm_cfg=self.norm_cfg)

    def forward(self, x):
        x = self.conv(x)
        x = self.dsconv1(x)
        x = self.dsconv2(x)
        return x


class GlobalFeatureExtractor(nn.Module):
    """Global feature extractor module.

    Args:
        in_channels (int): Number of input channels of the GFE module.
            Default: 64
        block_channels (tuple[int]): Tuple of ints. Each int specifies the
            number of output channels of each Inverted Residual module.
            Default: (64, 96, 128)
        out_channels(int): Number of output channels of the GFE module.
            Default: 128
        expand_ratio (int): Adjusts number of channels of the hidden layer
            in InvertedResidual by this amount.
            Default: 6
        num_blocks (tuple[int]): Tuple of ints. Each int specifies the
            number of times each Inverted Residual module is repeated.
            The repeated Inverted Residual modules are called a 'group'.
            Default: (3, 3, 3)
        strides (tuple[int]): Tuple of ints. Each int specifies
            the downsampling factor of each 'group'.
            Default: (2, 2, 1)
        pool_scales (tuple[int]): Tuple of ints. Each int specifies
            the parameter required in 'global average pooling' within PPM.
            Default: (1, 2, 3, 6)
        conv_cfg (dict | None): Config of conv layers. Default: None
        norm_cfg (dict | None): Config of norm layers. Default:
            dict(type='BN')
        act_cfg (dict): Config of activation layers. Default:
            dict(type='ReLU')
        align_corners (bool): align_corners argument of F.interpolate.
            Default: False
    """

    def __init__(self,
                 in_channels=64,
                 block_channels=(64, 96, 128),
                 out_channels=128,
                 expand_ratio=6,
                 num_blocks=(3, 3, 3),
                 strides=(2, 2, 1),
                 pool_scales=(1, 2, 3, 6),
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU'),
                 align_corners=False):
        super(GlobalFeatureExtractor, self).__init__()
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        assert len(block_channels) == len(num_blocks) == 3
        self.bottleneck1 = self._make_layer(in_channels, block_channels[0],
                                            num_blocks[0], strides[0],
                                            expand_ratio)
        self.bottleneck2 = self._make_layer(block_channels[0],
                                            block_channels[1], num_blocks[1],
                                            strides[1], expand_ratio)
        self.bottleneck3 = self._make_layer(block_channels[1],
                                            block_channels[2], num_blocks[2],
                                            strides[2], expand_ratio)
        self.ppm = PPM(
            pool_scales,
            block_channels[2],
            block_channels[2] // 4,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg,
            align_corners=align_corners)
        self.out = ConvModule(
            block_channels[2] * 2,
            out_channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

    def _make_layer(self,
                    in_channels,
                    out_channels,
                    blocks,
                    stride=1,
                    expand_ratio=6):
        layers = [
            InvertedResidual(
                in_channels,
                out_channels,
                stride,
                expand_ratio,
                norm_cfg=self.norm_cfg)
        ]
        for i in range(1, blocks):
            layers.append(
                InvertedResidual(
                    out_channels,
                    out_channels,
                    1,
                    expand_ratio,
                    norm_cfg=self.norm_cfg))
        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.bottleneck1(x)
        x = self.bottleneck2(x)
        x = self.bottleneck3(x)
        x = torch.cat([x, *self.ppm(x)], dim=1)
        x = self.out(x)
        return x


class FeatureFusionModule(nn.Module):
    """Feature fusion module.

    Args:
        higher_in_channels (int): Number of input channels of the
            higher-resolution branch.
        lower_in_channels (int): Number of input channels of the
            lower-resolution branch.
        out_channels (int): Number of output channels.
        conv_cfg (dict | None): Config of conv layers. Default: None
        norm_cfg (dict | None): Config of norm layers. Default:
            dict(type='BN')
        act_cfg (dict): Config of activation layers. Default:
            dict(type='ReLU')
        align_corners (bool): align_corners argument of F.interpolate.
            Default: False
    """

    def __init__(self,
                 higher_in_channels,
                 lower_in_channels,
                 out_channels,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU'),
                 align_corners=False):
        super(FeatureFusionModule, self).__init__()
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        self.align_corners = align_corners
        self.dwconv = ConvModule(
            lower_in_channels,
            out_channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)
        self.conv_lower_res = ConvModule(
            out_channels,
            out_channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=None)
        self.conv_higher_res = ConvModule(
            higher_in_channels,
            out_channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=None)
        self.relu = nn.ReLU(True)

    def forward(self, higher_res_feature, lower_res_feature):
        lower_res_feature = resize(
            lower_res_feature,
            size=higher_res_feature.size()[2:],
            mode='bilinear',
            align_corners=self.align_corners)
        lower_res_feature = self.dwconv(lower_res_feature)
        lower_res_feature = self.conv_lower_res(lower_res_feature)

        higher_res_feature = self.conv_higher_res(higher_res_feature)
        out = higher_res_feature + lower_res_feature
        return self.relu(out)


@BACKBONES.register_module()
class FastSCNN(nn.Module):
    """Fast-SCNN Backbone.

    Args:
        in_channels (int): Number of input image channels. Default: 3.
        downsample_dw_channels (tuple[int]): Number of output channels after
            the first conv layer & the second conv layer in
            Learning-To-Downsample (LTD) module.
            Default: (32, 48).
        global_in_channels (int): Number of input channels of
            Global Feature Extractor(GFE).
            Equal to number of output channels of LTD.
            Default: 64.
        global_block_channels (tuple[int]): Tuple of integers that describe
            the output channels for each of the MobileNet-v2 bottleneck
            residual blocks in GFE.
            Default: (64, 96, 128).
        global_block_strides (tuple[int]): Tuple of integers
            that describe the strides (downsampling factors) for each of the
            MobileNet-v2 bottleneck residual blocks in GFE.
            Default: (2, 2, 1).
        global_out_channels (int): Number of output channels of GFE.
            Default: 128.
        higher_in_channels (int): Number of input channels of the higher
            resolution branch in FFM.
            Equal to global_in_channels.
            Default: 64.
        lower_in_channels (int): Number of input channels of  the lower
            resolution branch in FFM.
            Equal to global_out_channels.
            Default: 128.
        fusion_out_channels (int): Number of output channels of FFM.
            Default: 128.
        out_indices (tuple): Tuple of indices of list
            [higher_res_features, lower_res_features, fusion_output].
            Often set to (0,1,2) to enable aux. heads.
            Default: (0, 1, 2).
        conv_cfg (dict | None): Config of conv layers. Default: None
        norm_cfg (dict | None): Config of norm layers. Default:
            dict(type='BN')
        act_cfg (dict): Config of activation layers. Default:
            dict(type='ReLU')
        align_corners (bool): align_corners argument of F.interpolate.
            Default: False
    """

    def __init__(self,
                 in_channels=3,
                 downsample_dw_channels=(32, 48),
                 global_in_channels=64,
                 global_block_channels=(64, 96, 128),
                 global_block_strides=(2, 2, 1),
                 global_out_channels=128,
                 higher_in_channels=64,
                 lower_in_channels=128,
                 fusion_out_channels=128,
                 out_indices=(0, 1, 2),
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU'),
                 align_corners=False):

        super(FastSCNN, self).__init__()
        if global_in_channels != higher_in_channels:
            raise AssertionError('Global Input Channels must be the same \
                                 with Higher Input Channels!')
        elif global_out_channels != lower_in_channels:
            raise AssertionError('Global Output Channels must be the same \
                                with Lower Input Channels!')

        self.in_channels = in_channels
        self.downsample_dw_channels1 = downsample_dw_channels[0]
        self.downsample_dw_channels2 = downsample_dw_channels[1]
        self.global_in_channels = global_in_channels
        self.global_block_channels = global_block_channels
        self.global_block_strides = global_block_strides
        self.global_out_channels = global_out_channels
        self.higher_in_channels = higher_in_channels
        self.lower_in_channels = lower_in_channels
        self.fusion_out_channels = fusion_out_channels
        self.out_indices = out_indices
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        self.align_corners = align_corners
        self.learning_to_downsample = LearningToDownsample(
            in_channels,
            downsample_dw_channels,
            global_in_channels,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)
        self.global_feature_extractor = GlobalFeatureExtractor(
            global_in_channels,
            global_block_channels,
            global_out_channels,
            strides=self.global_block_strides,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg,
            align_corners=self.align_corners)
        self.feature_fusion = FeatureFusionModule(
            higher_in_channels,
            lower_in_channels,
            fusion_out_channels,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg,
            align_corners=self.align_corners)

    def init_weights(self, pretrained=None):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                kaiming_init(m)
            elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
                constant_init(m, 1)

    def forward(self, x):
        higher_res_features = self.learning_to_downsample(x)
        lower_res_features = self.global_feature_extractor(higher_res_features)
        fusion_output = self.feature_fusion(higher_res_features,
                                            lower_res_features)

        outs = [higher_res_features, lower_res_features, fusion_output]
        outs = [outs[i] for i in self.out_indices]
        return tuple(outs)