Shuffle Net_ v1-shuffle_ v2

Catalog

1.shuffle operation

2. Computational complexity comparison ：

3.Shuffle Net_v1 Model architecture and parameter setting

4.Shuffle Net_v1 Accuracy rate

5.Shuffle Net_v2 Design criteria

6.Shuffle Net_v2 Network architecture improvement

7.Shuffle Net_v2 Parameters

---

shuffleNet combination ResNeXt Group convolution （GConv） And MobileNet Of DW Convolution .ResNeXt in 1*1 Convolution takes up 94.3% Of Mdds, Introduce group convolution GConv Reduce computational complexity . When performing group convolution , Each group of convolution is performed independently , So introduce shuffle operation bring Group And Group Information exchange between .

1.shuffle operation

2. Computational complexity comparison ：

  Compare with ResNet,ShuffleNet_v1 Yes 1*1 The convolution of , And in the middle 3*3 Convolution uses DW Convolution , Group number g= The channel number m.

3.Shuffle Net_v1 Model architecture and parameter setting

Next, let's look at the model architecture and the parameters of each layer ：

Every stage One of the first block Step length stride by 2, Next stage output channels Double （ Down sampling ） As shown in the structure diagram （c）.bottleneck（ middle 3*3DW Convolution ） The number of channels is... Of the output channel 1/4.

4.Shuffle Net_v1 Accuracy rate

 shufflenet_v2

5.Shuffle Net_v2 Design criteria

  shufflenet_v2 Four indicators for efficient network design are proposed ,

FLOPs Is an indirect indicator , And speed is the direct indicator of the speed of model calculation ,MAC（ Memory footprint ）、 platform 、 Parallelism is an index that affects the speed of the model , Put forward four guidelines for designing efficient networks ：

（1） When the number of channels of the input characteristic matrix and the output characteristic matrix of the convolution layer is equal ,MAC Minimum .

（2） When GConv Of group When it increases （FLOPs unchanged ）,MAC Also increase .

（3） The more fragmented the network is , The slower the network . Although the accuracy is improved , But it is not friendly to hardware devices with parallel computing .

 （4）Element—wise（ReLU、shortcut etc. ） The impact of the crisis cannot be ignored .

6.Shuffle Net_v2 Network architecture improvement

According to the above four principles , shufflenet_v2 Yes  shufflenet_v1 A series of improvements have been made , The network architecture is shown in the figure below ：

1 Division is adopted at channle_split The way , Reduce access to bneck The number of channels , In half .

2 Change the group convolution back to normal 1*1 Convolution , Follow the second 2, Fewer groups g.

3 Adopted at shuffle Location change of .

4 Adopted at concat Splicing ensures the whole input_c=output_c.

5 It's about ReLU Location change of .

 6bneck in 1*1—3*3—1*1 The number of input channels of three-layer convolution is equal to the number of output channels .

（d） The bottom sampling operation in the figure , Channel division channle_split, The number of output channels is doubled .

7.Shuffle Net_v2 Parameters

For each of these stage One of the first block,channel Double , such as stage2 Of block1, Each branch has 58 individual channels, Corresponding （d） chart . Others block use split Method will input_channels Divide equally .