mobile net 的 depthwise conv 和 origin conv 的对比

Posted 2022-06-11 血影雪梦

在inputchannel = k，outputchannel = q的情况下。

Mobilenet conv将 origin conv拆成了depthwise conv+pointwise conv，如下：

                Mobilenet conv:

                                depthwiseconv filter shape：[di,dj, k, channel_multiplier]，channel_multiplier默认=1

                                pointwiseconv filter shape：[1,1, k*channel_multiplier, q]

                因为pointwise conv就是origin conv，所以下面只对比depthwise conv和origin conv。

 

关于output channel数：

                origin conv         :              

                              output channel = q（output channel数）

                depthwise conv               :              

                              output channel = k（input channel数）

               PS：实际上outputchannel = k*channel_multiplier，但channel_multiplier默认=1

详情：                 origin conv:                                 outout: [b, i, j, q]                 depthwise conv:                                 for _k in 0..(k-1):                                                 for _q in 0..(channel_multiplier-1):                                                                 outout: [b, i, j, _k*channel_multiplier + _q]                                 因channel_multiplier默认= 1，所以可以写成：outout: [b, i, j, k]

 

 

关于算法实现：

                origin conv output channel 的每一层是：

                               in_channel_0*filter_0 + in_channel_1*filter_1+ ... + in_channel_k*filter_k

                depthwise conv output channel 的每一层是：

                                in_channel_i*filter_0

                                PS：实际上是：

                                             for i in0..(in_channel-1):

                                                                forj in 0..(channel_multiplier-1):

                                                                                in_channel_i*filter_j

                                因为channel_multiplier默认为=1，所以可以写成：in_channel_i*filter_0

详情：                 origin conv:                               output[b,i,j,0] = sum_di,dj,k input[b, strides[1]*i + di, strides[2] * j + dj,k] * filter[di, dj,k, 0]                               output[b,i,j,1] = sum_di,dj,k input[b, strides[1]*i + di, strides[2] * j + dj,k] * filter[di, dj,k, 1]                               ......                               output[b,i,j,q] = sum_di,dj,k input[b, strides[1]*i + di, strides[2] * j + dj,k] * filter[di, dj,k, q]                   depthwise conv:                               output[b,i,j, 0] = sum_di, dj input[b, strides[1] * i + di, strides[2] * j + dj,0] * filter[di, dj,0, 0]                               output[b,i,j, 1] = sum_di, dj input[b, strides[1] * i + di, strides[2] * j + dj,1] * filter[di, dj,1, 0]                               ......                               output[b,i, j, k] = sum_di, dj input[b, strides[1] * i + di, strides[2] * j + dj,k] * filter[di, dj,k, 0]                               PS：实际上是                                              for _k in 0..(k-1):                                                             for _q in 0..(channel_multiplier-1):                                                                            output[b, i, j, _k*channel_multiplier+_q] = sum_di, dj input[b, strides[1] * i + di, strides[2] * j + dj, _k] * filter[di, dj, _k, _q]                               因为 depthwise conv 的 filter 的 q 默认为 1，所以可以写成上面的形式

 

最后

               结合上面这些，论文中的这张图应该就很明了了：