Three questions of the results in your paper

[Gaffey]

Hi, I've read your paper but there are three questions remain.

1. Why the number of parameter of your GSANet is about 1/40 of the original GhostNet? (table 1, gamma=2) In my opinion, if gamma=2, the network should be just the same as the original GhostNet. The only difference is that convolution is substituted by shift and adder, but this would not influence the number of parameters.
2. In table 1, when gamma=6, the accuracy is comparable with original GhostNet. But you said that gamma=2 and gamma=4 are used for cpu and gpu. Why the final accuracy is even better than GhostNet?
3. I've tried your code, but I find that GhostSANet is much slower than the original GhostNet in my Nvidia V100 GPUs. Why the speed on Jetson Nano is 3 times faster than original GhostNet?

I would appreciate it if you can take the time to answer these.