Follow Tang Yudi pytorch On the third day , Roll up ！

Convolutional neural networks （ＣＮＮ）

        I've learned it several times , Here are only the points that should be paid attention to ：

       １、 The number of channels of convolution kernel should be the same as that of input image , And the parameter values of different channels of each convolution kernel are different . for example , The three channels of the convolution kernel convolute the three channels of the picture respectively and add them （ Then add ｂｉａｓ bias ） Get an eigenvalue ;

        2、 The convolution kernel size of different convolution layers can be different , The size of different convolution kernels in the same convolution layer needs to be the same , The number of characteristic graphs is the same as the number of convolution kernels ;

        3、padding Time filling 0 The reason is that , It will only expand the size without affecting the calculation ;

        4、 Convolution result calculation formula ： Feature map size ＝［（ Image size － Convolution kernel size 　+　2　*　padding）/　 step ］+　1;

       ５、 Only the layer with parameters can be counted as a layer , Such as convolution ;