unexpected loss curves on my own driving datasets

[HaiyangPeng]

1.After training, I got the following unexpected loss curves where consistency and smooth losses are not descending. Fortunately, the predicted depth map seems right for my case. The photo loss dominates the total loss, so the consistency and smooth can be constrained as the paper mentions?

2.I set the following training params in your provided bash file:

TRAIN_SET=xxxx
python train.py $TRAIN_SET \
--resnet-layers 18 \
--num-scales 1 \
-b8 -s0.1 -c0.5 --epochs 500 --sequence-length 3 \
--with-ssim 1 \
--with-mask 1 \
--with-auto-mask 1 \
--with-pretrain 1 \
--log-output \
--name xxxx

The weights of photo-loss, smooth-loss and geometry-consistency are same as yours. I don't know why I got these unexpected losses, and if you have encountered this issue, I will appreciate your help. THX.