Open
Description
Hi, I am confused about the calculation of bg_loss. here:
https://github.com/qiuyu96/CoDeF/blob/137f16c5423d484846857327597bf65c06b92994/train.py#L329
first of all, what does self_bg mean? Next, when self_bg=False, it seems that we are computing the mse loss between grid (coordinate values) and rgb values. I suppose this is wrong.
if self.hparams.bg_loss:
mk1 = torch.logical_not(mk_t)
if self.hparams.self_bg:
grid_flattened = rgbs_flattend
else:
grid_flattened = rearrange(grid, 'b n c -> (b n) c')
grid_flattened = torch.cat(
[grid_flattened, grid_flattened[:, :1]], -1)
if self.hparams.bg_loss and self.hparams.mask_dir:
loss = loss + self.hparams.bg_loss * self.color_loss(
results[mk1], grid_flattened[mk1])
although all experiments use self_bg=True according to the config. I'm just curious about the config self_bg, what does it mean? and what are we doing when self_bg=False ?
Metadata
Metadata
Assignees
Labels
No labels