Train loss about bsp-net-pytorch HOT 5 OPEN

I am not sure whether the range of the loss is normal or not. But the loss will be higher when training on a higher resolution. So in your case, it is normal for the loss to jump from 0.002 to 0.02 to 0.05 when the training resolution switches from 16 to 32 to 64.

To confirm whether the training is successful, check the quality of the sample output shapes during training.

from bsp-net-pytorch.

Thank you for your reply. I found some problems with my ground truth data.

Its interior is hollow. I think some problems happened in data processing using https://github.com/czq142857/IM-NET/tree/master/point_sampling

from bsp-net-pytorch.

It is likely due to improper normalization. Please normalize your shapes (voxels) so that the object is centered and the diagonal of the object bounding box is 1.

from bsp-net-pytorch.

Thank you for your reply. I'm sorry I don't know how to normalize. Is it using the binvox?

from bsp-net-pytorch.

You can refer to 1_simplify_obj.py and 2_voxelize.py in this data preparation code for normalizing shapes in obj formats and voxelizing them, respectively.

from bsp-net-pytorch.