Closed HighCWu closed 5 years ago
@HighCWu awesome man; have you tried that? I don't have a TPU and not an expert here (I don't like notebooks so I can't run in on colab) If you have tried it and it's working then I can merge it and thanks again for your contribution :+1:
@Separius I've tried it a lot of times on colab because of my needs. I found a bug when using load_bert. It seems it should be the same function as the load_google_bert, but has a wrong variable. I fix it. Now i modify the code to support theano again. The PR is updated. I think it's no problem to merge. And I also updated the colab code, so others can easily use.
cool cool, thanks again!
I have misunderstood some things before, so that only the prediction part is using tpu. Now the training can also use tpu.