作者从内存访问代价（Memory Access Cost，MAC）和GPU并行性的方向分析了网络应该怎么设计才能进一步减少运行时间，直接的提高模型的效率

设计原则

• Equal channel width minimizes memory access cost (MAC)
• Excessive group convolution increases MAC
• Network fragmentation reduces degree of parallelism
• Element-wise operations are non-negligible

Equal channel width minimizes memory access cost

当输入通道和输出通道数量相同的时候,mac最小.

假如一个卷积操作的输入特征图尺寸为:,输出特征图大小为: . 则该卷积操作的FLOPS为.
在这个计算过程中,输入特征图占用内存为:,输出特征图占用内存为:,卷积核占用内存为:

则总计内存访问为:

令,则得:

因为：

当C1 等于C2的时候,mac最小

Excessive group convolution increases MAC.

group convolution,分组卷积,通过将所有通道之间的密集卷积改为稀疏(仅在通道组内),可以显著降低计算复杂度.

其MAC计算公式为:

当

所以当分组数g增大时,内存访问增大;而论文中也用一组实验证明了该观点:

image

显然，使用较大的组数会大大降低运行速度。 例如，在GPU上使用8组的速度比在1组（标准密集卷积）上的速度慢两倍以上，而在ARM上则要慢30％。 这主要是由于MAC增加。因此要谨慎地设计变量g;

Network fragmentation reduces degree of parallelism

“multi- path” structure,多路径网络(如googlenet中的 四个分支)使用了许多小型操作符,尽管这种零散的结构已经显示了对准确性的提高,但是它对于GPU并行计算不太友好,因此可能会降低效率.作者设计了以下不同网络来计算其效率:

image

以下结果证明并行操作会降低运行速度,如 e的结果比c慢了三倍;

image

Element-wise operations are non-negligible

element-wise 操作非常耗时.

在计算FlOPs时,我们通常只考虑卷积中的乘法操作,但是一些逐个元素(element-wise)的操作会占用相当长的时间,尤其是在GPU.它们的FLOP较小，但MAC相对较重。各个操作占时如下图所示:

image

guide conclusion

在设计高性能网络时,尽量做到:

• 使用输入通道和输出通道相同的卷积操作
• 谨慎使用分组卷积
• 减少网络并行分支
• 减少元素级(element-wise)操作

网络基础结构

作者在论文中提到,回顾于shuffleNet v1,其改进经验应该为:

Therefore, in order to achieve high model capacity and efficiency, the key issue is how to maintain a large number and equally wide channels with neither dense convolution nor too many groups

如何在既不密集卷积(使用分组卷积)又不分太多组的情况下保持大量且同样宽的通道数量.

Channel split

如下图所示:
在每个单元的开头,将c个特征通道输入分为两个分支和,依据之前的网络设计原则G3(不宜有太多并行分支),其中一个分支将作为高速残差通道直接与另一分支结果直接concat;另一个分支中,由三个具有相同的输入通道数和输出通道数的卷积组成(以满足设计原则G1).两个1 * 1的卷积不再按组进行,原因:channel split 已经产生了两个组.concat后按照shuffleNet v1的方式,将两个分支的通道内容进行"channel shuffle"的操作来启用两个分支之间的信息通信.
像ReLU和深度卷积之类的元素操作仅存在于一个分支中。同样，三个连续的元素方式操作（“ Concat”，“ Channel Shuffle”和“ Channel Split”）合并为单个元素方式操作。

Channel split

import torch
import torch.nn as nn

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

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

    x = torch.transpose(x, 1, 2).contiguous()

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

    return x

class InvertedResidual(nn.Module):
    def __init__(self,inp,oup,stride):
        super(InvertedResidual,self).__init__()
        branch_features = oup // 2
        self.stride = stride

        self.branch_1 = nn.Sequential()

        self.branch_2 = nn.Sequential(
            nn.Conv2d(inp if (self.stride > 1) else branch_features,
                      branch_features,kernel_size=1,stride=1,padding=0,bias=False),
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
            ##depth_wise_conv
            nn.Conv2d(branch_features,branch_features,kernel_size=3,stride=stride,padding=0,bias=False,groups=branch_features),
            nn.BatchNorm2d(branch_features),
            nn.Conv2d(branch_features,branch_features,kernel_size=1,stride=1,padding=0,bias=False),
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
        )

    def forward(self, x):
        out = torch.cat((self.branch_1(x),self.branch_2(x)),dim=1)
        out = channel_shuffle(out,2)
        return out

down sample

channel split 用来保持图片分辨率,而更多情况下,我们需要对图片进行采样信息抽取,所以stride=2,改为如下结构:通道拆分运算符已删除。因此，输出通道的数量增加了一倍.branch_1 不再是直接高速通道直接连接,而是变成一个3 * 3 的深度分离卷积和一个1 * 1的卷积进行信息抽取.

image

#所以我们修改branch_1的定义即可
self.branch_1 = nn.Sequential(
            nn.Conv2d(inp,inp,kernel_size=3,stride=2,padding=1,groups=inp),
            nn.BatchNorm2d(inp),
            nn.Conv2d(inp,branch_features,kernel_size=1,stride=1,padding=0,bias=False),
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True)
        )

在shuffle net中,按照bottlenet的通道数按照比例缩放以生成复杂度不同的网络.

shuffle net v2整体结构

按照论文中提供网络结构图,

image

我们来用pytorch代码搭建shufflenet v2;

首先是改写bottlenet结构,按照stride是否为2,来判断另一个分支上是否需要用33深度分离卷积跟11卷积进行信息抽取;

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

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

        branch_features = oup // 2
        assert (self.stride != 1) or (inp == branch_features << 1)

        if self.stride > 1:
            self.branch1 = nn.Sequential(
                self.depthwise_conv(inp, inp, kernel_size=3, stride=self.stride, padding=1),
                nn.BatchNorm2d(inp),
                nn.Conv2d(inp, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
                nn.BatchNorm2d(branch_features),
                nn.ReLU(inplace=True),
            )
        else:
            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),
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
            self.depthwise_conv(branch_features, branch_features, kernel_size=3, stride=self.stride, padding=1),
            nn.BatchNorm2d(branch_features),
            nn.Conv2d(branch_features, branch_features, kernel_size=1, stride=1, padding=0, bias=False),
            nn.BatchNorm2d(branch_features),
            nn.ReLU(inplace=True),
        )

    @staticmethod
    def depthwise_conv(i, o, kernel_size, stride=1, padding=0, bias=False):
        return nn.Conv2d(i, o, kernel_size, stride, padding, bias=bias, groups=i)

    def forward(self, x):
        if self.stride == 1:
            x1, x2 = x.chunk(2, dim=1)
            out = torch.cat((x1, self.branch2(x2)), dim=1)
        else:
            out = torch.cat((self.branch1(x), self.branch2(x)), dim=1)

        out = channel_shuffle(out, 2)

        return out

然后我们按照stage不同,搭建起整体shufflenet 用于imagenet 1000分类的整体结构:

class ShuffleNetV2(nn.Module):
    def __init__(self, stages_repeats, stages_out_channels, num_classes=1000, inverted_residual=InvertedResidual):
        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

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

        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

        stage_names = ['stage{}'.format(i) for i in [2, 3, 4]]
        for name, repeats, output_channels in zip(
                stage_names, stages_repeats, self._stage_out_channels[1:]):
            seq = [inverted_residual(input_channels, output_channels, 2)]
            for i in range(repeats - 1):
                seq.append(inverted_residual(output_channels, output_channels, 1))
            setattr(self, name, nn.Sequential(*seq))
            input_channels = output_channels

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

        self.fc = nn.Linear(output_channels, num_classes)

    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)


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

如果按照通道数不同来进行对网络的缩放,则:
0.5:

return _shufflenetv2('shufflenetv2_x0.5', pretrained, progress,
                         [4, 8, 4], [24, 48, 96, 192, 1024], **kwargs)

1:

return _shufflenetv2('shufflenetv2_x1.0', pretrained, progress,
                         [4, 8, 4], [24, 116, 232, 464, 1024], **kwargs)

总结与分析

shuffleNet v2效率高且准确的原因主要有两个:

• 每个构建模块(building block)中高效地利用了feature channel和network capacity;
• 每个模块中,一半的特征通道信息直接参与下一个block中(高速残差信息通道).被看做是另一种方式的特征复用.

在densenet中,分析得到相邻层之间的连接要强于其他层之间的信息连接,也就是意味着所有层之间的密集连接可能带来冗余,特征重用量随两个块之间的距离呈指数衰减。在远距离的块之间，功能重用变得很弱。shuffleNet 用 channel split的方式进行特征复用.

Reference

• [1] Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 116-131.

• [2] https://github.com/pytorch/vision/blob/master/torchvision/models/shufflenetv2.py