loss

[xzhdxe]

Hello, I would like to ask: What is the value that the loss approaches approximately at the end, or to what extent does the loss need to reach, to be considered that the model has converged?

[xzhdxe]

Could you please tell me why the loss doesn't converge and always fluctuates? I would really appreciate your guidance on the possible reasons for this issue. Thank you very much.

[Eono701]

Have you solved this problem?

[aicomplier]

Hi！Have you solved this problem?

[Sebastianjhx]

Hi, the loss at smaller timestep tends to be of greater value. I'd recommend save checkpoints and evaluate on a regular basis to compare.