def shufflenet_v2_x1_0(pretrained=False, progress=True, **kwargs):
    """
    Constructs a ShuffleNetV2 with 1.0x output channels, as described in
    `"ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design"
    <https://arxiv.org/abs/1807.11164>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    return _shufflenetv2('shufflenetv2_x1.0', pretrained, progress,
                         [4, 8, 4], [24, 116, 232, 464, 1024], **kwargs)
1
2
3
4
5
6
7
8
9
10
11
12

return _shufflenetv2('shufflenetv2_x0.5', pretrained, progress,
                         [4, 8, 4], [24, 48, 96, 192, 1024], **kwargs)
return _shufflenetv2('shufflenetv2_x1.5', pretrained, progress,
                         [4, 8, 4], [24, 176, 352, 704, 1024], **kwargs)
return _shufflenetv2('shufflenetv2_x2.0', pretrained, progress,
                         [4, 8, 4], [24, 244, 488, 976, 2048], **kwargs)
1
2
3
4
5
6

def _shufflenetv2(arch, pretrained, progress, *args, **kwargs):
    model = ShuffleNetV2(*args, **kwargs)

    if pretrained:
        model_url = model_urls[arch]
        if model_url is None:
            raise NotImplementedError('pretrained {} is not supported as of now'.format(arch))
        else:
            state_dict = load_state_dict_from_url(model_url, progress=progress)
            model.load_state_dict(state_dict)

    return model
1
2
3
4
5
6
7
8
9
10
11
12

class ShuffleNetV2(nn.Module):
    def __init__(self, stages_repeats, stages_out_channels, num_classes=1000, inverted_residual=InvertedResidual):
    #  重复次数stages_repeats=[4, 8, 4], 输出通道数stages_out_channels=[24, 116, 232, 464, 1024]
        super(ShuffleNetV2, self).__init__()

        if len(stages_repeats) != 3:
            raise ValueError('expected stages_repeats as list of 3 positive ints')
        if len(stages_out_channels) != 5:
            raise ValueError('expected stages_out_channels as list of 5 positive ints')
        self._stage_out_channels = stages_out_channels  # [24, 116, 232, 464, 1024]

        input_channels = 3
        output_channels = self._stage_out_channels[0]  # 24
        self.conv1 = nn.Sequential(
            nn.Conv2d(input_channels, output_channels, 3, 2, 1, bias=False),  # 3 24 3x3卷积
            nn.BatchNorm2d(output_channels),
            nn.ReLU(inplace=True),
        )
        input_channels = output_channels  # 更新input_channels值24

        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)  # 最大池化

        stage_names = ['stage{}'.format(i) for i in [2, 3, 4]]  # 阶段名字list ['stage2', 'stage3', 'stage4']
        for name, repeats, output_channels in zip(
                stage_names, stages_repeats, self._stage_out_channels[1:]):
            seq = [inverted_residual(input_channels, output_channels, 2)]  # 24 116 2(步长)
            for i in range(repeats - 1):  # 重复数减1 ，上面的是0，现在是从1开始
                seq.append(inverted_residual(output_channels, output_channels, 1))  # 输入与输出相等，步长为1
            setattr(self, name, nn.Sequential(*seq))  # setattr() 函数对应函数 getattr()，用于设置属性值
            input_channels = output_channels  # 更新input_channels值

        output_channels = self._stage_out_channels[-1]  # 1024
        self.conv5 = nn.Sequential(
            nn.Conv2d(input_channels, output_channels, 1, 1, 0, bias=False),  # 1x1卷积
            nn.BatchNorm2d(output_channels),
            nn.ReLU(inplace=True),
        )

        self.fc = nn.Linear(output_channels, num_classes)  # 分类器层

    def _forward_impl(self, x):
        # See note [TorchScript super()]
        x = self.conv1(x)
        x = self.maxpool(x)
        x = self.stage2(x)
        x = self.stage3(x)
        x = self.stage4(x)
        x = self.conv5(x)
        x = x.mean([2, 3])  # globalpool
        x = self.fc(x)
        return x

    def forward(self, x):
        return self._forward_impl(x)
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

class InvertedResidual(nn.Module):
    def __init__(self, inp, oup, stride):
    	# 输入，输出，步长
        super(InvertedResidual, self).__init__()

        if not (1 <= stride <= 3):
            raise ValueError('illegal stride value')
        self.stride = stride

        branch_features = oup // 2  # 分支通道数，输出通道数减半
        assert (self.stride != 1) or (inp == branch_features << 1)

        if self.stride > 1:  # 步长大于1
            self.branch1 = nn.Sequential(
                self.depthwise_conv(inp, inp, kernel_size=3, stride=self.stride, padding=1),  # 3x3可分离卷积
                nn.BatchNorm2d(inp),
                nn.Conv2d(inp, branch_features, kernel_size=1, stride=1, padding=0, bias=False),  # 1x1卷积
                nn.BatchNorm2d(branch_features),
                nn.ReLU(inplace=True),
            )
        else:  # 否则为list
            self.branch1 = nn.Sequential()

        self.branch2 = nn.Sequential(
            nn.Conv2d(inp if (self.stride > 1) else branch_features,
                      branch_features, kernel_size=1, stride=1, padding=0, bias=False),  # 1x1卷积
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
            self.depthwise_conv(branch_features, branch_features, kernel_size=3, stride=self.stride, padding=1),  # 分支的3x3可分离卷积
            nn.BatchNorm2d(branch_features),
            nn.Conv2d(branch_features, branch_features, kernel_size=1, stride=1, padding=0, bias=False),  # 1x1卷积
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
        )

    @staticmethod
    def depthwise_conv(i, o, kernel_size, stride=1, padding=0, bias=False):
    # i = o 可分离卷积函数
        return nn.Conv2d(i, o, kernel_size, stride, padding, bias=bias, groups=i)   # groups 等于输入通道数

    def forward(self, x):
        if self.stride == 1:  # 如果步长等于1
            x1, x2 = x.chunk(2, dim=1)   # 沿1轴分为2部分
            out = torch.cat((x1, self.branch2(x2)), dim=1)  # 合并 按通道数
        else:  # 否则 两个分支合并
            out = torch.cat((self.branch1(x), self.branch2(x)), dim=1)

        out = channel_shuffle(out, 2)  # 

        return out
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

def channel_shuffle(x, groups):
	# groups = 2
    # type: (torch.Tensor, int) -> torch.Tensor
    batchsize, num_channels, height, width = x.data.size()  # 
    channels_per_group = num_channels // groups  # 分成两部分

    # reshape
    x = x.view(batchsize, groups,
               channels_per_group, height, width)

    x = torch.transpose(x, 1, 2).contiguous()   # 交换1,2通道，转为连续类型

    # flatten
    x = x.view(batchsize, -1, height, width)   # 变换维度

    return x
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16