loss not decrease

[hdmjdp]

hi, in my training process, the loss do not decrease. How about your super parms setting?

[veqtor]

Can you show a loss graph? Where does your loss get stuck? Batch size? What GPU?

[hdmjdp]

121k step, batchsize=8