Closed Wuziyi616 closed 1 year ago
ok so I double-check the code, this line is the only place affected by this bug, because you're reading the correct step value here. All the others (lr, resolution, which are after this correct line) are not affected.
After I change the previous line to the correct one, the performance on my custom data improves a lot (I think this is because I'm using lr_fg_begin_step>0
). As shown below, the orange curve is after correcting that line.
There is also another places affected by this, since curr_step
is always 0, the ResampleCallBack
condition here will never be True, so it will never upsample the Plenoxel voxel grids
Sorry for being late. This issue is often found when using a different pytorch_lightning version. Could you share your environment version?
In the case of the last comment, we'll double check it and change the config if some arguments are not used.
Feel free to reopen the issue if you suffer from any problems regarding this issue.
Hi, thanks for releasing this great work! I think there is an error in the implementation of
LitPlenoxel
. You defineself.curr_step
here and use it to determine lr decay, voxel upsampling, etc. However, it seems that you never increase this value? I add aprint(gstep)
after this line, and it always prints0
.