Hello, I found some strange thing in alternating-training code.
From stage1 fast-rcnn training, it return trained weight that is un-normalized (by snapshot function).
But, in stage2 fast-rcnn training, it copy the un-normalized weight(from stage1 fast-rcnn) and
train by 'normalized targets'.
If it's true, stage2 fast-rcnn training will be hard to converge, because scale is different.
Hello, I found some strange thing in alternating-training code.
From stage1 fast-rcnn training, it return trained weight that is un-normalized (by snapshot function). But, in stage2 fast-rcnn training, it copy the un-normalized weight(from stage1 fast-rcnn) and train by 'normalized targets'.
If it's true, stage2 fast-rcnn training will be hard to converge, because scale is different.
Am i missing something??