Why use padding = [0, 0] in downsampling layers of FBNet V3 ?

[changlin31]

In FBNet V3, the padding of downsampling layer is set to [1, 1] for kernel_size 5 and [0, 0] for kernel_size 3, resulting in smaller resolution in later stages (i.e. 13x13 and 6x6 in the last two stages). This is not a common setting in other network, including MBNet and previous version of FBNet.

Why do you make such adjustment in FBNet V3? How does it affect the accuracy? Thanks in advance.

[zhanghang1989]

The architecture is searched automatically.

[changlin31]

I know, but I believe padding is not included in the search space.

[laisimiao]

Does this repo provide model or codes of v3 version? @changlin31

[changlin31]

Yes, @laisimiao here: https://github.com/facebookresearch/mobile-vision/blob/b2964a4476167589b0f7fad207bf51f614af5fd7/mobile_cv/arch/fbnet_v2/fbnet_modeldef_cls_fbnetv3.py

[laisimiao]

@changlin31 Thanks. But it just provide the pretrained model with v1 and v2?

[sbl1996]

@changlin31 Do you know the reason of the uncommon padding now? Thanks very much.

[changlin31]

@sbl1996 I have no clue. Maybe it's just a trick to reduce the FLOPS.