Simple sorting of selective kernel networks methods

Link to the original text ：Selective Kernel Networks | IEEE Conference Publication | IEEE Xplore

         For different sizes of input information, adaptive learning convolution kernel size .

It is divided into 3 A link ：

        split： Transformation of two different convolution kernels （ Successive grouped/depth-wise Convolution +BN+ReLU） Act on respectively On , obtain and .

        fuse： Add the outputs of the above two transformations to get , Then, the average value is pooled to get . Pass a layer FC、BN and ReLU, Reduce the dimension to the original （ At the same time, in order to avoid the dimension being too small, set the minimum value ）, obtain Dimension vector .

        select： Using multiple linear layers will After transformation softmax, obtain soft attention vector （ In the figure ）. namely

among Is the linear layer weight Of the That's ok , yes Of the Elements ; Empathy .

         Finally, press the output ：

The above is the case of two branches . It is easy to expand to multiple branches .