Loss when training with my own data

[Sissi-lu]

Hi, thank you for sharing the code and data you used in Metric3D. It helps me a lot to understand your work. But when I trained with my own data with RGB-Depth-Normal label following by #105 ， I felt confused about what kind of loss I should apply.

In the picture, I saw you mainly use three losses, .i.e $L_d$ for supervising the depth, $L_n$ for supervising normal, and $L_{d-n}$ to align the predicted depth：

Specifically, the $L_{d}$ is composed by:

but it seems that $L_{PWN}$ and $L_{VNL}$ are the normal loss?

When I tried to understand this loss in your code, I found they are hard to correspond to. This brings me these main questions:
1、Can you tell me the specific meaning of the loss in your paper, as well as their corresponding sense? And what do they correspond to the code?
2、When I want to fine-tune the model with my own data with GT of depth and normal, which loss should I apply? I'm just too confused about the loss in code. Now I use the configuration loss of vit.raft5.giant2.nyu.py, .i.e,

losses=dict( 
    decoder_losses=[
        dict(type='VNLoss', sample_ratio=0.2, loss_weight=1.0),
        dict(type='GRUSequenceLoss', loss_weight=1.0, loss_gamma=0.9, stereo_sup=0),
        dict(type='NormalBranchLoss', loss_weight=1.5, loss_fn='NLL_ours_GRU'),
        dict(type='DeNoConsistencyLoss', loss_weight=0.001, loss_fn='CEL', scale=2), 
        dict(type='HDNRandomLoss', loss_weight=0.5, random_num=10), 
        dict(type='HDSNRandomLoss', loss_weight=0.5, random_num=20, batch_limit=4),  
        dict(type='PWNPlanesLoss', loss_weight=1),  
    ], 
)

I find some of them equal to zero or never reduce, just as follows:

Thanks a lot!