GitLab

apply Whole Word Masking technique.
referred to [create_pretraining_data.py](https://github.com/google-research/bert/blob/master/create_pretraining_data.py)

fix issue of bert-base-multilingual and add support for uncased multilingual

Indentation change so that the output "nbest_predictions.json" is not empty.

Reason for issue was that optimzation steps where computed from example size, which is different from actual size of dataloader when an example is chunked into multiple instances.

Solution in this pull request is to compute num_optimization_steps directly from len(data_loader).

small fix to remove shifting of lm labels during pre process of roc stories, as this shifting happens interanlly in the model

error for users whose corpus is just one giant document.

Fix for

```> > > > 04/09/2019 21:39:38 - INFO - __main__ -   device: cuda n_gpu: 1, distributed training: False, 16-bits training: False
Traceback (most recent call last):
  File "/home/ubuntu/pytorch-pretrained-BERT/examples/lm_finetuning/simple_lm_finetuning.py", line 642, in <module>
    main()
  File "/home/ubuntu/pytorch-pretrained-BERT/examples/lm_finetuning/simple_lm_finetuning.py", line 502, in main
    raise ValueError("Training is currently the only implemented execution option. Please set `do_train`.")
ValueError: Training is currently the only implemented execution option. Please set `do_train`.
```

```while not args.unconditional:
   if not args.unconditional:
```
These lines have been updated

Modify 'unambigiously' to 'unambiguously'

If the value of rand_end is returned from the randint function, the value of sampled_doc_index that matches current_idx is returned from searchsorted.

example:
cumsum_max = {int64} 30
doc_cumsum = {ndarray} [ 5  7 11 19 30]
doc_lengths = {list} <class 'list'>: [5, 2, 4, 8, 11]
if current_idx  = 1,
rand_start = 7
rand_end = 35
sentence_index = randint(7, 35) % cumsum_max
if randint return 35, sentence_index becomes 5.
if sentence_index is 5, np.searchsorted returns 1 equal to current_index.