Monocular SNN training

[CurryYH]

Hi, I've tried to reproduce the experimental results of SNN Monocular version on DENSE dataset. I have noticed that the label shape is different from the predicted tensor which is [N,1,260,346] and I have expanded the label dimension into 4 to fit the loss function. Besides, I have also change the model input channels into 2 to fit the single camera data.However, the network didn't perform well and the mean depth error was hundreds meters. The initial training loss in the first epoch was so big that it would take hundreds of epochs to converge. Therefore, I'd like to know the details of the monocular training procedures , the dataloading details regarding the label, xl,yl dimension and the monocular SNN model. It will be highly appreciated that you can offer some guidance regarding the issues mentioned above. Thanks!