Jihun Choi

> @xyfZzz > Can you test if below workaround is useful? > > Workaround: in `modeling.py/get_assignment_map_from_checkpoint` > > ``` > - assignment_map[name] = name > + assignment_map[name] = name_to_variable[name] >...

I found the hang arises due to the call of `BertServer.close`, regardless of the version (at least the problem also persists in v1.8.6, v1.8.0, v1.7.0). Also, `Ctrl-C`-ing the server terminates...

Then, can this issue be closed? If modification is needed, please let me know more about the details! 😃

I tried to use the initialization scheme stated in the paper, however it seems there’s a problem in that. I will see the problem in days. Thanks!

Hi, as per my understanding, the most common usage of the packed sequence is to give it as input to pre-defined RNN modules (e.g. torch.nn.LSTM, torch.nn.GRU, ...). However the batch-normalized...

It seems like an oversight. Thanks! I have a plan to refactor the entire code when I have some time, including what you've pointed out and other suggestions. (Or it...

Sorry for the late reply. As far as I understood, the paper states that an element whose time index is larger than `T_max` just uses the population statistics of time...