How to avoid NaN in training

[frickyinn]

bts/pytorch/bts.py

Line 47 in dd62221

d = torch.log(depth_est[mask]) - torch.log(depth_gt[mask])

Nan may occur while training because the final depth estimation sometimes yields 0 values. However, the loss function never tries to avoid log(0).
The solution to the situation is simply changing to this:

d = torch.log(depth_est[mask] + 1e-4) - torch.log(depth_gt[mask])