I have recently been reading both the paper and the current training code carefully, and I have a question about how RefineNet is intended to be trained in practice.
In my own experiments and reproduction attempts, I observed the following behavior: RefineNet seems to take over most of the prediction/correction work, while the coarse branch itself does not appear to learn very strong semantic or coarse estimation ability. In other words, the refinement stage seems to dominate the optimization, while the coarse stage behaves more like a relatively weak initializer.
So I wanted to ask whether this is the expected behavior in the original WiLoR training setup, or whether I may be missing some important detail in my understanding or reproduction.
In addition, I noticed that in the current code, get_parameters() seems to return only self.backbone.parameters(). I am not sure whether this is intentional, or whether self.refine_net.parameters() should also be included in the optimizer parameter groups.
I would really appreciate any clarification, as I am trying to better understand the optimization behavior of this coarse-to-fine design.
I have recently been reading both the paper and the current training code carefully, and I have a question about how RefineNet is intended to be trained in practice.
In my own experiments and reproduction attempts, I observed the following behavior: RefineNet seems to take over most of the prediction/correction work, while the coarse branch itself does not appear to learn very strong semantic or coarse estimation ability. In other words, the refinement stage seems to dominate the optimization, while the coarse stage behaves more like a relatively weak initializer.
So I wanted to ask whether this is the expected behavior in the original WiLoR training setup, or whether I may be missing some important detail in my understanding or reproduction.
In addition, I noticed that in the current code, get_parameters() seems to return only self.backbone.parameters(). I am not sure whether this is intentional, or whether self.refine_net.parameters() should also be included in the optimizer parameter groups.
I would really appreciate any clarification, as I am trying to better understand the optimization behavior of this coarse-to-fine design.