File size: 11,586 Bytes
f549064
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import torch
import torch.nn as nn
import torch.utils.checkpoint as cp
from mmcv.cnn import ConvModule, build_activation_layer
from mmengine.model import BaseModule
from mmengine.model.weight_init import constant_init, normal_init
from torch.nn.modules.batchnorm import _BatchNorm

from mmcls.models.utils import channel_shuffle, make_divisible
from mmcls.registry import MODELS
from .base_backbone import BaseBackbone


class ShuffleUnit(BaseModule):
    """ShuffleUnit block.

    ShuffleNet unit with pointwise group convolution (GConv) and channel
    shuffle.

    Args:
        in_channels (int): The input channels of the ShuffleUnit.
        out_channels (int): The output channels of the ShuffleUnit.
        groups (int): The number of groups to be used in grouped 1x1
            convolutions in each ShuffleUnit. Default: 3
        first_block (bool): Whether it is the first ShuffleUnit of a
            sequential ShuffleUnits. Default: True, which means not using the
            grouped 1x1 convolution.
        combine (str): The ways to combine the input and output
            branches. Default: 'add'.
        conv_cfg (dict, optional): Config dict for convolution layer.
            Default: None, which means using conv2d.
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='BN').
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='ReLU').
        with_cp (bool): Use checkpoint or not. Using checkpoint
            will save some memory while slowing down the training speed.
            Default: False.

    Returns:
        Tensor: The output tensor.
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 groups=3,
                 first_block=True,
                 combine='add',
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU'),
                 with_cp=False):
        super(ShuffleUnit, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.first_block = first_block
        self.combine = combine
        self.groups = groups
        self.bottleneck_channels = self.out_channels // 4
        self.with_cp = with_cp

        if self.combine == 'add':
            self.depthwise_stride = 1
            self._combine_func = self._add
            assert in_channels == out_channels, (
                'in_channels must be equal to out_channels when combine '
                'is add')
        elif self.combine == 'concat':
            self.depthwise_stride = 2
            self._combine_func = self._concat
            self.out_channels -= self.in_channels
            self.avgpool = nn.AvgPool2d(kernel_size=3, stride=2, padding=1)
        else:
            raise ValueError(f'Cannot combine tensors with {self.combine}. '
                             'Only "add" and "concat" are supported')

        self.first_1x1_groups = 1 if first_block else self.groups
        self.g_conv_1x1_compress = ConvModule(
            in_channels=self.in_channels,
            out_channels=self.bottleneck_channels,
            kernel_size=1,
            groups=self.first_1x1_groups,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)

        self.depthwise_conv3x3_bn = ConvModule(
            in_channels=self.bottleneck_channels,
            out_channels=self.bottleneck_channels,
            kernel_size=3,
            stride=self.depthwise_stride,
            padding=1,
            groups=self.bottleneck_channels,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=None)

        self.g_conv_1x1_expand = ConvModule(
            in_channels=self.bottleneck_channels,
            out_channels=self.out_channels,
            kernel_size=1,
            groups=self.groups,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=None)

        self.act = build_activation_layer(act_cfg)

    @staticmethod
    def _add(x, out):
        # residual connection
        return x + out

    @staticmethod
    def _concat(x, out):
        # concatenate along channel axis
        return torch.cat((x, out), 1)

    def forward(self, x):

        def _inner_forward(x):
            residual = x

            out = self.g_conv_1x1_compress(x)
            out = self.depthwise_conv3x3_bn(out)

            if self.groups > 1:
                out = channel_shuffle(out, self.groups)

            out = self.g_conv_1x1_expand(out)

            if self.combine == 'concat':
                residual = self.avgpool(residual)
                out = self.act(out)
                out = self._combine_func(residual, out)
            else:
                out = self._combine_func(residual, out)
                out = self.act(out)
            return out

        if self.with_cp and x.requires_grad:
            out = cp.checkpoint(_inner_forward, x)
        else:
            out = _inner_forward(x)

        return out


@MODELS.register_module()
class ShuffleNetV1(BaseBackbone):
    """ShuffleNetV1 backbone.

    Args:
        groups (int): The number of groups to be used in grouped 1x1
            convolutions in each ShuffleUnit. Default: 3.
        widen_factor (float): Width multiplier - adjusts the number
            of channels in each layer by this amount. Default: 1.0.
        out_indices (Sequence[int]): Output from which stages.
            Default: (2, )
        frozen_stages (int): Stages to be frozen (all param fixed).
            Default: -1, which means not freezing any parameters.
        conv_cfg (dict, optional): Config dict for convolution layer.
            Default: None, which means using conv2d.
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='BN').
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='ReLU').
        norm_eval (bool): Whether to set norm layers to eval mode, namely,
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only. Default: False.
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed. Default: False.
    """

    def __init__(self,
                 groups=3,
                 widen_factor=1.0,
                 out_indices=(2, ),
                 frozen_stages=-1,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 act_cfg=dict(type='ReLU'),
                 norm_eval=False,
                 with_cp=False,
                 init_cfg=None):
        super(ShuffleNetV1, self).__init__(init_cfg)
        self.init_cfg = init_cfg
        self.stage_blocks = [4, 8, 4]
        self.groups = groups

        for index in out_indices:
            if index not in range(0, 3):
                raise ValueError('the item in out_indices must in '
                                 f'range(0, 3). But received {index}')

        if frozen_stages not in range(-1, 3):
            raise ValueError('frozen_stages must be in range(-1, 3). '
                             f'But received {frozen_stages}')
        self.out_indices = out_indices
        self.frozen_stages = frozen_stages
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        self.norm_eval = norm_eval
        self.with_cp = with_cp

        if groups == 1:
            channels = (144, 288, 576)
        elif groups == 2:
            channels = (200, 400, 800)
        elif groups == 3:
            channels = (240, 480, 960)
        elif groups == 4:
            channels = (272, 544, 1088)
        elif groups == 8:
            channels = (384, 768, 1536)
        else:
            raise ValueError(f'{groups} groups is not supported for 1x1 '
                             'Grouped Convolutions')

        channels = [make_divisible(ch * widen_factor, 8) for ch in channels]

        self.in_channels = int(24 * widen_factor)

        self.conv1 = ConvModule(
            in_channels=3,
            out_channels=self.in_channels,
            kernel_size=3,
            stride=2,
            padding=1,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

        self.layers = nn.ModuleList()
        for i, num_blocks in enumerate(self.stage_blocks):
            first_block = True if i == 0 else False
            layer = self.make_layer(channels[i], num_blocks, first_block)
            self.layers.append(layer)

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            for param in self.conv1.parameters():
                param.requires_grad = False
        for i in range(self.frozen_stages):
            layer = self.layers[i]
            layer.eval()
            for param in layer.parameters():
                param.requires_grad = False

    def init_weights(self):
        super(ShuffleNetV1, self).init_weights()

        if (isinstance(self.init_cfg, dict)
                and self.init_cfg['type'] == 'Pretrained'):
            # Suppress default init if use pretrained model.
            return

        for name, m in self.named_modules():
            if isinstance(m, nn.Conv2d):
                if 'conv1' in name:
                    normal_init(m, mean=0, std=0.01)
                else:
                    normal_init(m, mean=0, std=1.0 / m.weight.shape[1])
            elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
                constant_init(m, val=1, bias=0.0001)
                if isinstance(m, _BatchNorm):
                    if m.running_mean is not None:
                        nn.init.constant_(m.running_mean, 0)

    def make_layer(self, out_channels, num_blocks, first_block=False):
        """Stack ShuffleUnit blocks to make a layer.

        Args:
            out_channels (int): out_channels of the block.
            num_blocks (int): Number of blocks.
            first_block (bool): Whether is the first ShuffleUnit of a
                sequential ShuffleUnits. Default: False, which means using
                the grouped 1x1 convolution.
        """
        layers = []
        for i in range(num_blocks):
            first_block = first_block if i == 0 else False
            combine_mode = 'concat' if i == 0 else 'add'
            layers.append(
                ShuffleUnit(
                    self.in_channels,
                    out_channels,
                    groups=self.groups,
                    first_block=first_block,
                    combine=combine_mode,
                    conv_cfg=self.conv_cfg,
                    norm_cfg=self.norm_cfg,
                    act_cfg=self.act_cfg,
                    with_cp=self.with_cp))
            self.in_channels = out_channels

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.maxpool(x)

        outs = []
        for i, layer in enumerate(self.layers):
            x = layer(x)
            if i in self.out_indices:
                outs.append(x)

        return tuple(outs)

    def train(self, mode=True):
        super(ShuffleNetV1, self).train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                if isinstance(m, _BatchNorm):
                    m.eval()