Why can the model correctly predict empty occupancies?

[lnSong]

Thank you for your work!
The code occ[..., 3][occ[..., 3] == 0] = 255 sets empty occupancies to 255,
and the loss criterion is defined as:
criterion = nn.CrossEntropyLoss(
ignore_index=255, reduction="mean"
)
Since the loss ignores the index 255, why is the model still able to correctly infer empty occupancies?

[lnSong]

non-occupied is trained as it is labeled as 0
def multiscale_supervision(gt_occ, ratio, gt_shape):
'''
change ground truth shape as (B, W, H, Z) for each level supervision
'''

gt = torch.zeros([gt_shape[0], gt_shape[2], gt_shape[3], gt_shape[4]]).to(gt_occ.device).type(torch.float) 
for i in range(gt.shape[0]):
    coords = gt_occ[i][:, :3].type(torch.long) // ratio
    gt[i, coords[:, 0], coords[:, 1], coords[:, 2]] =  gt_occ[i][:, 3]

return gt