the last depthwise_conv_block(b18)

[stainless-steel-rat]

I noticed that the last depthwise_conv_block(b18) is annotated. You used a standard conv layer to replace it. This standard conv layer cost about 20% computation. So why not use depthwise_conv_block as above layers?

[akirasosa]

@stainless-steel-rat
If the conv layer is put in the middle of U-Net, the computational cost will be high. But they are on the edge. It does not have so much differences.

[stainless-steel-rat]

@akirasosa
emm, thanks for the response. there is something wrong in my calculations indeedly, the percentage is lower than 20%. But the considering the larger size of feature maps on the edge, the cost may also high. Does the reason of using standard conv for a better performance?

[akirasosa]

@stainless-steel-rat
Not so much reason. The original MobileNets uses conv block at 1st block. So, I made it symmetric.