Qibin Hou
Qibin Hou
Channel attention is based on the capability of capturing global information. We do this by using two different types of pooling operations.
According to my knowledge, the answer is yes.
In mobile network training, it would be better to use ReLU6 or Swich, which is smooth. MobileNetV3 has demonstrated this.
You may try to put it after the layer merge.
Do you mean point cloud data?
I think it is correct.
It needs a try.
This is for better optimization for mobile networks.
Yes, you are highly suggested to pretrain your model on ImageNet first.
If you used adaptive_avg_pool, that would not be a problem.