1 ShuffleNet的思想

ShuffleNet使用Group convolution和Channel shuffle改进ResNet，可以看作是ResNet的压缩版本。

• Group convolution
• Channel shuffle

图1 ShuffleNet的微结构

ShuffleNet的本质是将卷积运算限制在每个Group内，这样模型的计算量取得了显著的下降。然而导致模型的信息流限制在各个Group内，组与组之间没有信息交换，这会影响模型的表示能力。因此，需要引入组间信息交换的机制，即Channel Shuffle操作。同时Channel Shuffle是可导的，可以实现end-to-end一次性训练网络。

2 核心代码

分组shuffle通道：

def shuffle_channels(x, groups):
    """shuffle channels of a 4-D Tensor"""
    batch_size, channels, height, width = x.size()
    assert channels % groups == 0
    channels_per_group = channels // groups
    # split into groups
    x = x.view(batch_size, groups, channels_per_group,
               height, width)
    # transpose 1, 2 axis
    x = x.transpose(1, 2).contiguous()
    # reshape into orignal
    x = x.view(batch_size, channels, height, width)
    return x

ShuffleNet的A单元：

class ShuffleNetUnitA(nn.Module):
    """ShuffleNet unit for stride=1"""
    def __init__(self, in_channels, out_channels, groups=3):
        super(ShuffleNetUnitA, self).__init__()
        assert in_channels == out_channels
        assert out_channels % 4 == 0
        bottleneck_channels = out_channels // 4
        self.groups = groups
        self.group_conv1 = nn.Conv2d(in_channels, bottleneck_channels,
                                        1, groups=groups, stride=1)
        self.bn2 = nn.BatchNorm2d(bottleneck_channels)
        self.depthwise_conv3 = nn.Conv2d(bottleneck_channels,
                                         bottleneck_channels,
                                         3, padding=1, stride=1,
                                         groups=bottleneck_channels)
        self.bn4 = nn.BatchNorm2d(bottleneck_channels)
        self.group_conv5 = nn.Conv2d(bottleneck_channels, out_channels,
                                     1, stride=1, groups=groups)
        self.bn6 = nn.BatchNorm2d(out_channels)

    def forward(self, x):
        out = self.group_conv1(x)
        out = F.relu(self.bn2(out))
        out = shuffle_channels(out, groups=self.groups)
        out = self.depthwise_conv3(out)
        out = self.bn4(out)
        out = self.group_conv5(out)
        out = self.bn6(out)
        out = F.relu(x + out)
        return out

ShuffleNet的B单元：

class ShuffleNetUnitB(nn.Module):
    """ShuffleNet unit for stride=2"""
    def __init__(self, in_channels, out_channels, groups=3):
        super(ShuffleNetUnitB, self).__init__()
        out_channels -= in_channels
        assert out_channels % 4 == 0
        bottleneck_channels = out_channels // 4
        self.groups = groups
        self.group_conv1 = nn.Conv2d(in_channels, bottleneck_channels,
                                     1, groups=groups, stride=1)
        self.bn2 = nn.BatchNorm2d(bottleneck_channels)
        self.depthwise_conv3 = nn.Conv2d(bottleneck_channels,
                                         bottleneck_channels,
                                         3, padding=1, stride=2,
                                         groups=bottleneck_channels)
        self.bn4 = nn.BatchNorm2d(bottleneck_channels)
        self.group_conv5 = nn.Conv2d(bottleneck_channels, out_channels,
                                     1, stride=1, groups=groups)
        self.bn6 = nn.BatchNorm2d(out_channels)

    def forward(self, x):
        out = self.group_conv1(x)
        out = F.relu(self.bn2(out))
        out = shuffle_channels(out, groups=self.groups)
        out = self.depthwise_conv3(out)
        out = self.bn4(out)
        out = self.group_conv5(out)
        out = self.bn6(out)
        x = F.avg_pool2d(x, 3, stride=2, padding=1)
        out = F.relu(torch.cat([x, out], dim=1))
        return out

3 优缺点分析

缺点：

• Shuffle channel在实现的时候需要大量的指针跳转和Memory set，这本身就是极其耗时的；同时又特别依赖实现细节，导致实际运行速度不会那么理想。
• Shuffle channel规则是人工设计出来的，不是网络自己学出来的。这不符合网络通过负反馈自动学习特征的基本原则，又陷入人工设计特征的老路（如sift/HOG等）。

4 总结

首先介绍了ShuffleNet的基本思想，然后介绍了ShuffleNet的核心代码实现，最后分析了ShuffleNet V1的缺点，指明了改进方向。
最后插一句，想得到未必能做到，能做到未必能做好，能做好未必能产生效益。
告诉自己，想到了就要做到，做到了就要尽力做好，做好了就要想办法变现。