loss

[aicomplier]

为什么loss值一直浮动，在loss为多少的时候说明训练的差不多了，必须等到收敛吗

[Sebastianjhx]

Hi, the loss values at timestep from different stages are not on a similar scale, for example model on smaller timestep tends to have larger loss, vice versa. I'd recommend save checkpoints and evaluate on a regular basis to compare.