Hi, I've seen that for the squeeze co-excitation (SCE) part of the paper, a global average pooling operation is performed on the weighted non-local feature map. After I looked at the code implementation for SCE in faster_rcnn.py, I was a little confused because the global average pooling is not implemented in the forward pass in faster_rcnn.py, even though it's initialized in the match_block class, and I couldn't find any other implementations of it in the code base. I commented the GAP initialization and re-ran the test script on the COCO second group and found that it worked with similar accuracy.
Have I misinterpreted the implementation of GAP in SCE or am I looking in the wrong place?
Thanks!
Hi, I've seen that for the squeeze co-excitation (SCE) part of the paper, a global average pooling operation is performed on the weighted non-local feature map. After I looked at the code implementation for SCE in faster_rcnn.py, I was a little confused because the global average pooling is not implemented in the forward pass in faster_rcnn.py, even though it's initialized in the match_block class, and I couldn't find any other implementations of it in the code base. I commented the GAP initialization and re-ran the test script on the COCO second group and found that it worked with similar accuracy.
Have I misinterpreted the implementation of GAP in SCE or am I looking in the wrong place?
Thanks!