驗證resneXt，densenet，mobilenet和SENet的特點結構

簡介

圖像分類對網絡結構的要求，一個是精度，另外一個是速度。這兩個需求推進了網絡結構的發展。

• resneXt：分組卷積，下降了網絡參數個數。
• densenet：密集的跳鏈接。
• mobilenet：標準卷積分解成深度卷積和逐點卷積，即深度分離卷積。
• SENet：注意力機制。

簡單起見，使用了[1]的代碼，註釋掉 layer4，做爲基本框架resnet14。而後改變局部結構，驗證分類效果。

實驗結果

GPU：gtx1070
超參數：epochs=80,lr=0.001,optim=Adam
數據集：cifar10，batch_size=100

分組卷積

# 3x3 convolution with grouping
def conv3x3(in_channels, out_channels, stride=1, groups=1):
    return nn.Conv2d(in_channels, out_channels, kernel_size=3,
                     stride=stride, padding=1, bias=False,groups=groups)

_	參數個數(k)	GPU內存(M)	訓練時間(s)	測試時間(s)	精度(%)
resnet14	195	617	665	0.34	87
分組=2	99	615	727	0.40	85
分組=4	50	615	834	0.50	81

結論：卷積分組下降了參數個數，同時也下降了速度和精度。

密集鏈接

def forward(self, x): # basic block
        residual = x
        if self.downsample:
            residual = self.downsample(x)
        out = self.layer1(x)
        out = self.relu(out)
        out2 = self.layer2(out)
        out2 = self.relu(out2)
        out3 = torch.cat([out,out2],1)
        out = self.layer3(out3)
        out4 = self.relu(out)
        out5 = torch.cat([out3,out4],1)
        out = self.layer4(out5) # back to the specified channels
        return out

_	參數個數(k)	GPU內存(M)	訓練時間(s)	測試時間(s)	精度(%)
resnet14	195	617	665	0.34	87
密集鏈接	341	679	703	0.43	88

結論：參數個數和精度有所增長，速度降低一點點。

深度分離卷積

def Conv2d(in_channels, out_channels,kernel_size=1,padding=0,stride=1):
    return nn.Sequential(*[
            nn.Conv2d(in_channels, in_channels,kernel_size,stride=stride,padding=padding,groups=in_channels,bias=False),
            nn.Conv2d(in_channels, out_channels,1,bias=False),
        ])

_	參數個數(k)	GPU內存(M)	訓練時間(s)	測試時間(s)	精度(%)
resnet14	195	617	665	0.34	87
分組=2	99	615	727	0.40	85
分組=4	50	615	834	0.50	81
深度分離卷積	27	665	788	0.40	84

結論：深度分離卷積下降了參數個數，同時也下降了速度和精度。與分組卷積（分組=4）相比，精度要高一點。

注意力機制

利用[2]的代碼，修正通道個數

def forward(self, x): # BasicBlock
        residual = x
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.bn2(out)
        if self.downsample:
            residual = self.downsample(x)
        # attention
        original_out = out
        out = F.avg_pool2d(out,out.size()[2:])
        out = out.view(out.size(0), -1)
        out = self.fc1(out)
        out = self.relu(out)
        out = self.fc2(out)
        out = self.sigmoid(out)
        out = out.view(out.size(0), out.size(1), 1, 1)
        out = out * original_out
        out += residual
        out = self.relu(out)
        return out

_	參數個數(k)	GPU內存(M)	訓練時間(s)	測試時間(s)	精度(%)
resnet14	195	617	665	0.34	87
注意力	201	641	838	0.51	87

結論：參數個數和精度變更不大，速度下降比較明顯。

引用

[1] https://github.com/yunjey/pytorch-tutorial/tree/master/tutorials/02-intermediate/deep_residual_network/main.py
[2] https://github.com/miraclewkf/SENet-PyTorch/blob/master/se_resnet.py

參考文獻

• Chollet, François. Xception: Deep Learning with Depthwise Separable Convolutions[J]. 2016.
• Xie S , Girshick R , Dollár, Piotr, et al. Aggregated Residual Transformations for Deep Neural Networks[J]. 2016.
• Huang G, Liu Z, Laurens V D M, et al. Densely Connected Convolutional Networks[J]. 2016.
• Howard A G , Zhu M , Chen B , et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[J]. 2017.
• Hu J , Shen L , Albanie S , et al. Squeeze-and-Excitation Networks[J]. 2017.
• http://www.javashuo.com/article/p-geacblrr-hr.html