Regarding the loss issue of training from scratch on one's own dataset

[DaIGaN2019]

Hello author, thank you for being willing to open source such outstanding code. I trained from scratch on my own dataset, and when using the original vit_L14 config, the loss became higher and higher. After adjusting the parameters, the loss gradually decreased, but after a small decrease, it almost stagnated (13.8-12.9, where dino_global-loss remained unchanged for a long time). Is this normal? Can you briefly tell me the extent to which your loss decreased during the training process? Also, the log shows that my keleo-loss has decreased from a positive number to a negative number. Is this reasonable? Thank you again for your contribution. Looking forward to your reply.