In README.md, you wrote that the training steps are 40000 at most.
However, when I tried to reproduce the fine-tuned model you kindly provide, it seems the training steps are too many if I run the script at the default setting (100.0 epochs are set as default for run_union.py).
The initial checkpoint of BERT can be downloaded from bert. We use the uncased base version of BERT (about 110M parameters). We train the model for 40000 steps at most. The training process will task about 1~2 days.
After the 100-epoch training, I got model.ckpt-1414000.
If you don't mind, could you please tell me the appropriate args to be used for reproducing your fine-tuned UNION models used in the paper?
Is it enough to change the training epochs?
Hello,
In README.md, you wrote that the training steps are 40000 at most. However, when I tried to reproduce the fine-tuned model you kindly provide, it seems the training steps are too many if I run the script at the default setting (100.0 epochs are set as default for
run_union.py).After the 100-epoch training, I got
model.ckpt-1414000.If you don't mind, could you please tell me the appropriate args to be used for reproducing your fine-tuned UNION models used in the paper? Is it enough to change the training epochs?
I'm sorry if there are already details somewhere.
Thank you in advance.