Resolve #218: Support soft ids as inputs for BERT/GPT2/RoBERTa/XLNet

[gpengzhi]

Support soft_ids as inputs for encoders, classifiers, and regressors.

[gpengzhi]

Resolve #218

[huzecong]

LGTM. What worries me is that our implementation of modules based on pre-trained stuff is a bit too repetitive, so that a seemingly small change would require modifying a bunch of files. This is also true for a lot of tests (not limited to pre-trained ones). Let's keep this in mind so we can improve this in the future.

[gpengzhi]

LGTM. What worries me is that our implementation of modules based on pre-trained stuff is a bit too repetitive, so that a seemingly small change would require modifying a bunch of files. This is also true for a lot of tests (not limited to pre-trained ones). Let's keep this in mind so we can improve this in the future.

This is a very good suggestion. For the code redundancy in the models, I feel that we can extract some task-specific headers for all or most of the pre-trained models. For the testing issue, some of the tests can be shared across different models. I think we can write a common test script that can be used by a group of classes for unit testing.

[huzecong]

That sounds good. Let's open a separate issue to keep track of this, so we can merge this PR.

[huzecong]

Feel free to merge when you're done with the changes.