ShuffleNet V2 神经网络简介与代码实战

1.介绍
    ShuffleNet V2也是由旷视科技提出的，它是ShuffleNet v1的升级版（可参见我的一篇博客ShuffleNet 神经网络简介与代码实战），旷视科技是根据四个规则来改进ShuffleNet v1，更加详细的介绍可以参见：ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design。

 
2.模型结构
   下图中分析了ShuffleNet v1和MobileNets V2在GPU和ARM上的耗时分布，从图中可看出Conv操作占了其时间的绝大多数，但Elemwise/Data IO等内存读写密集型操作也占了相当比例的时间，因此，我们在设计网络的时候，不仅要考虑Conv的个数也要考虑其他操作。

文中总结了四个准则来指导优化网络：

1）卷积层的输入和输出特征通道数相等时MAC(内存访问消耗时间)最小，此时模型速度最快。

2）过多的group操作会增大MAC，从而使模型速度变慢。

3）模型中的网络分支数量越少，模型速度越快。

4 ）element-wise操作所带来的时间消耗远比在FLOPs上的体现的数值要多，因此要尽可能减少element-wise操作。

接下里，我具体来看下图中网络结构的改变，是怎样对应着这四个准则，其中(a)和(b)是ShuffleNet v1文章中两种结构，(c)和(d)则是根据(a)和(b)改进而来。(c)结构相对于(a)结构，增加了channel split操作，就是为了保证concat操作之后满足前面第1)点。(c)中取消了GCconv是为了满足前面第2)点，同时把channel shuffle的操作移到了concat后面，和前面第3)点对应。将将element-wise add操作替换成concat，和和前面第4)点对应。

3.模型特点

  通过四个上面所提到的四个准则设计网络，而不是盲目的设计网络，这样做的好处，可以减少计算量和运算时间。

4.代码实现 pytorch

def conv_bn(inp, oup, stride):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU(inplace=True)
    )
 
 
def conv_1x1_bn(inp, oup):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU(inplace=True)
    )
 
def channel_shuffle(x, groups):
    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, benchmodel):
        super(InvertedResidual, self).__init__()
        self.benchmodel = benchmodel
        self.stride = stride
        assert stride in [1, 2]
 
        oup_inc = oup//2
        
        if self.benchmodel == 1:
            #assert inp == oup_inc
        	self.banch2 = nn.Sequential(
                # pw
                nn.Conv2d(oup_inc, oup_inc, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup_inc),
                nn.ReLU(inplace=True),
                # dw
                nn.Conv2d(oup_inc, oup_inc, 3, stride, 1, groups=oup_inc, bias=False),
                nn.BatchNorm2d(oup_inc),
                # pw-linear
                nn.Conv2d(oup_inc, oup_inc, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup_inc),
                nn.ReLU(inplace=True),
            )                
        else:                  
            self.banch1 = nn.Sequential(
                # dw
                nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
                nn.BatchNorm2d(inp),
                # pw-linear
                nn.Conv2d(inp, oup_inc, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup_inc),
                nn.ReLU(inplace=True),
            )        
    
            self.banch2 = nn.Sequential(
                # pw
                nn.Conv2d(inp, oup_inc, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup_inc),
                nn.ReLU(inplace=True),
                # dw
                nn.Conv2d(oup_inc, oup_inc, 3, stride, 1, groups=oup_inc, bias=False),
                nn.BatchNorm2d(oup_inc),
                # pw-linear
                nn.Conv2d(oup_inc, oup_inc, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup_inc),
                nn.ReLU(inplace=True),
            )
          
    @staticmethod
    def _concat(x, out):
        # concatenate along channel axis
        return torch.cat((x, out), 1)        
 
    def forward(self, x):
        if 1==self.benchmodel:
            x1 = x[:, :(x.shape[1]//2), :, :]
            x2 = x[:, (x.shape[1]//2):, :, :]
            out = self._concat(x1, self.banch2(x2))
        elif 2==self.benchmodel:
            out = self._concat(self.banch1(x), self.banch2(x))
 
        return channel_shuffle(out, 2)
 
 
class ShuffleNetV2(nn.Module):
    def __init__(self, n_class=1000, input_size=224, width_mult=1.):
        super(ShuffleNetV2, self).__init__()
        
        assert input_size % 32 == 0
        
        self.stage_repeats = [4, 8, 4]
        # index 0 is invalid and should never be called.
        # only used for indexing convenience.
        if width_mult == 0.5:
            self.stage_out_channels = [-1, 24,  48,  96, 192, 1024]
        elif width_mult == 1.0:
            self.stage_out_channels = [-1, 24, 116, 232, 464, 1024]
        elif width_mult == 1.5:
            self.stage_out_channels = [-1, 24, 176, 352, 704, 1024]
        elif width_mult == 2.0:
            self.stage_out_channels = [-1, 24, 224, 488, 976, 2048]
        else:
            raise ValueError(
                """{} groups is not supported for
                       1x1 Grouped Convolutions""".format(num_groups))
 
        # building first layer
        input_channel = self.stage_out_channels[1]
        self.conv1 = conv_bn(3, input_channel, 2)    
	self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        
        self.features = []
        # building inverted residual blocks
        for idxstage in range(len(self.stage_repeats)):
            numrepeat = self.stage_repeats[idxstage]
            output_channel = self.stage_out_channels[idxstage+2]
            for i in range(numrepeat):
                if i == 0:
	            #inp, oup, stride, benchmodel):
                    self.features.append(InvertedResidual(input_channel, output_channel, 2, 2))
                else:
                    self.features.append(InvertedResidual(input_channel, output_channel, 1, 1))
                input_channel = output_channel
                
                
        # make it nn.Sequential
        self.features = nn.Sequential(*self.features)
 
        # building last several layers
        self.conv_last      = conv_1x1_bn(input_channel, self.stage_out_channels[-1])
	self.globalpool = nn.Sequential(nn.AvgPool2d(int(input_size/32)))              
    
	# building classifier
	self.classifier = nn.Sequential(nn.Linear(self.stage_out_channels[-1], n_class))
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.maxpool(x)
        x = self.features(x)
        x = self.conv_last(x)
        x = self.globalpool(x)
        x = x.view(-1, self.stage_out_channels[-1])
        x = self.classifier(x)
        return x