Hi, Happynear,Depth-wise separable convolution is widely adopted in recent state-of-art neural networks, such as Xception, mobilenet, shufflenet, but there is no such layer in the original caffe implementation (using convolution layer with group parameter is excruciatingly slow! ), I am wondering if you have any plan to implement a high efficiency depth-wise separable convolution layer in this repo in the near future?
Hi, Happynear,Depth-wise separable convolution is widely adopted in recent state-of-art neural networks, such as Xception, mobilenet, shufflenet, but there is no such layer in the original caffe implementation (using convolution layer with group parameter is excruciatingly slow! ), I am wondering if you have any plan to implement a high efficiency depth-wise separable convolution layer in this repo in the near future?