kernel size in primary convolution of Ghost module

[YUHAN666]

Hi,
It is said in your paper that the primary convolution in Ghost module can have customized kernel
size, which is a major difference from existing efficient convolution schemes. However, it seems that in this code all the kernel size of primary convolution in Ghost module are set to [1, 1], and the kernel set in _CONV_DEFS_0 are only used in blocks of stride=2. Is it set intentionally?

[iamhankai]

The code is not clean enough. kernel set in _CONV_DEFS_0 for bottleneck of stride=1 is useless.

[YUHAN666]

So the kernel size of primary convolution in Ghost module is fixed to [1, 1] by default?

[YUHAN666]

https://github.com/huawei-noah/ghostnet/blob/4ccf0993cec68a0b04faeea3271a14b453a96d9b/ghost_net.py#L288-L289
https://github.com/huawei-noah/ghostnet/blob/4ccf0993cec68a0b04faeea3271a14b453a96d9b/ghost_net.py#L270-L271

[iamhankai]

Yes, the kernel size of primary convolution in Ghost module is fixed to [1, 1] by default.

[YUHAN666]

Have you tried with larger size?

[KnightWin123]

Have you tried with larger size?

I have the same question

[iamhankai]

@YUHAN666 @dyj1234567890 Not yet.

[KnightWin123]

@YUHAN666 Have you tried with larger size?

[YUHAN666]

@dyj1234567890 Not yet, but when applied on my own project, the GhostNet trained with default setting of this code has the best performance among all the backbone I have tried.