Fine-tuning with my own dataset

[XYQ0121]

Hello, I noticed that the fine-tuning of depth-anything-v1 was performed using the KITTI dataset (with sparse depth maps as ground truth). However, depth-anything-v2 was fine-tuned using Virtual KITTI 2 (with dense depth maps as ground truth). When I fine-tune depth-anything-v2 using my self-collected dataset (with sparse depth maps as ground truth), there are issues with the loss. Is there any good solution for this?

[LiheYoung]

Hi, did you correctly ignore the regions without groundtruth values during loss computation?

[XYQ0121]

Hi, did you correctly ignore the regions without groundtruth values during loss computation?

It seems that the issue with the loss function occurred because I did not set the learning rate correctly during training. However, even when I correctly set the training parameters and used the same data to fine-tune both v1 and v2, v2 performed worse than v1. Could this be because I trained using sparse depth maps?

[XiaoLuoLYG]

@XYQ0121 Hi, I met the same Nan loss problem, how did you solve this. I set LR as default 0.000005