the problem of lr

[ffxz]

I find that the lr of your code is 0.0001, but the paper's lr is 0.0015, which one is better? I train the model, the loss can not decrease, i don't know why.

[zhr1201]

I didn't spend too much effort fine tune the model when I wrote this piece of code. I just found that this lr work for the data set and the model, but there may be better choices.