Loss doesn't go down.

[Emibobo]

Take 5s video segments form hundreds of videos, each 5s video segment takes 10 frames of images to train DINOv2 from the beginning, the input tensor shape of the model is [B,3,10,H,W], the batchsize is set to 12, run on 4 A100 (80GB) GPUs, the training parameter defaults to /configs/train/ vitl14.yaml.

[Shiyao-Xu]

Same here, did you finally solve that issue?