solve problem of normalization

[DianaDI]

• local (per input) normalization -> good reconstruction of relief, bigger loss. Depth is not learned due to local depth values only
• global (whole dataset) normalization -> small l1 loss, bad reconstruction. Only predicts constant like depths
• no normalization -> since values range between 600-700, it might be bad for NN

Solutions:

• use another loss, e.g. loss assuming the distribution, maybe smoothing loss
• learn to predict relief and min, max depths