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：
