cc
cc
是这边的后半部分，您看一下
哦哦，原来是这样啊，怪我代码没仔细推敲，多谢您的指点！
I have the question too. I searched on the stackoverflow and know that there is something with the shape of tensor. However ,when I change the 'self.lstm.state_size' to [256,256],another question...
I have met the problem too,converting midi to mp3 really solved my problem,but I still don't know why this happened. Finnally,I solved the problem.It maybe depands on what Interpreter you...
> > > 这是正常的 因为我微调后也这样了 > > > 这显然不正常。 > > 如果你理解ptuning在干什么，这显然正常。 本来一个对话模型，现在把输入变成非流畅文本，每层还额外加入prompt，怎么可能保留之前的能力。 如果你需要保留对话能力，肯定要把输入指令全部改写成人话，并且用lora的方式仅微调q、v来减少影响。我理解ptuning这种引入额外结构的根本不适用在这种场景下。 没错，直观理解起来就是你通过p-tuning之后，等于在基座模型之上加了一层全连接，现在它只支持当前的任务了。
I am facing the same problem. The grad_norm is na after 50 steps.