About global average pooling layer about senet HOT 7 CLOSED

We have conducted experiments on Cityscapes Dataset using FCN to validate effectiveness of SE upon segmentation task. With same training configurations, SE outperforms its non-SE counterpart. Additionally, we also assessed SE on COCO detection task. Similarly, SE achieves great improvement against non-SE.

from senet.

Thanks.

from senet.

@hujie-frank could you please give more detail about how to apply SE on FCN, like in which layer to add the SE modules? Thanks.

from senet.

We apply SE in front half of FCN (down-sample stage) to enhance feature expression.

from senet.

Thanks for replying, I will try that. If you don't mind, could you please give me more information about your experiments on COCO detection task, such as how much improvement you get on the mAP compared with non-SE, and which network you use in the detection task ( still FCN?) and how you apply the SE modules? Thanks soooo much!!

from senet.

hi @hujie-frank,
i have a question, why you put the global_ave backward op in PoolingParameter_PoolMethod_STOCHASTIC case?

from senet.

@zimenglan-sysu-512 Sorry for that, I called the function in wrong case and I fixed it just now. Thank you for your attention.

from senet.