[CNN] recommended depth separable convolution alternative Conv2D-- quickly enhance image recognition CNN / visual recognition model

《AI工匠BOOK》持续更新AI算法与最新应用，如果您感兴趣，欢迎关注AI工匠（AI算法与最新应用前沿研究）。
Father in Keras depth study of the book, design a way for image recognition tasks to improve the performance of a few percentage points of the network layer, the network layer can not only replace Conv2D, and can make the model more lightweight, with less trainable weights parameter, faster, the network layer is separable convolution depth (depthwise separable convolution) layer (SeparableConv2D). ``

   ******深度可分离卷积原理介绍：******

This layer SeparableConv2D separately for each input channel performs spatial convolution, convolution and point by point (1 × 1 convolution) mixing the output channel, which is equivalent to learning and learning spatially separated wherein the channel characteristics, parameters required to a lot less, the amount of computation is also smaller, so you can get a smaller, faster model. Because it is a more efficient method of performing the convolution, it is often possible to use less data to learn a better representation of the model to obtain a better performance.
Separable convolution depth application scenario
assumed spatial position in the input image are highly correlated, different independent channels.

      ******深度可分离卷积构建模型代码实战：****** 

Here we use a separable convolution depth to build a complete model image recognition CNN,