Jiejing Zhang
Jiejing Zhang
> @paupereira, I have this in my spacemacs/user-config. The problem form your approach is that git-gutter+ is lazy loaded, and may be loaded after your redefinition. > > ```emacs-lisp >...
I found same error in paper. @taolei87
Hi Tao, I think TF cannot do this computation automatically, because I pass W and X as two input, and do the mm() operation inside the OP's Copmute() function(like forward()...
Hi, Actually TF cannot do this computation automatically, because the new defined OP take "W" as input, and do the `mm` inside the OP 's compute function(like forward() function in...
Thanks for your detail explain.
@taolei87 I did some gradient check in my implementation of SRU in Tensorflow, I found some gradient error between grad function and numeric gradient, here is some data: When k...
Hi @taolei87, Thanks for you reply. I don't very clear what's you mean highway sub-layer compute, I don't see the concate(h_f[t], h_b][t]) in cuda_functional.py code in sru_forward() and sru_backward() and...
Thanks for your detail explanation, I still have some question about how grad x was produced. if U's tranpose dot with W can produce X, do you mean the U...
是在什么虚拟机？ 或者看下这个命令： `lscpu` 的输出？
这个CPU 太旧了，不支持任何fp16的指令，demo默认的推理是bf16的， 你可以换新一点的CPU试一下， 如果可能得话，最好用ECS上的第八代ECS，那个CPU不仅支持bf16精度，还可以用amx指令加速计算。