caffe源码-conv_layer

先来一下code，  基本思路就是把 input_shape( c x h x w ) ->  (out_h x out_w) x (i_c x k_h x k_w),  把filter（o_c x i_c x k_h x k_w）-> o_c x (i_c x k_h x k_w) 

然后就是矩阵乘法 IN * W.T， caffe里im2col :  input_shape( c x h x w ) -> (i_c x k_h x k_w) x (out_h x out_w) ,  W * IN

x is bottom.cpu_data

forward:    Y = W X

conv_layer: Foward_cpu() 

base_conv_layer: forward_cpu_gemm()   [ caffe_cpu_gemm() <- cblas] 

im2col

backword:

1. back_pro:   bottom_diff = W.T * top_diff_  

2. weight_gradient: this.diff = top_diff * X.T

-------------reference------------------

1. high performance convolution neural network for document processing

2.