Awesome

exBERT

The details of the model is in paper.

Pre-train an exBERT model (only the extension part)

In command line:

python Pretraining.py -e 1 
  -b 256 
  -sp path_to_storage
  -dv 0 1 2 3 -lr 1e-04 
  -str exBERT    
  -config path_to_config_file_of_the_OFF_THE_SHELF_MODEL ./config_and_vocab/exBERT/bert_config_ex_s3.json  
  -vocab ./config_and_vocab/exBERT/exBERT_vocab.txt 
  -pm_p path_to_state_dict_of_the_OFF_THE_SHELF_MODEL
  -dp path_to_your_training_data
  -ls 128 
  -p 1

You can replace the path_to_config_file_of_the_OFF_THE_SHELF_MODEL and path_to_state_dict_of_the_OFF_THE_SHELF_MODEL to any weel pre-trained model in BERT archietecture. ./config_and_vocab/exBERT/bert_config_ex_s3.json defines the size of extension module.

Pre-train an exBERT model (whole model)

python Pretraining.py -e 1 
  -b 256 
  -sp path_to_storage
  -dv 0 1 2 3 -lr 1e-04 
  -str exBERT    
  -config path_to_config_file_of_the_OFF_THE_SHELF_MODEL ./config_and_vocab/exBERT/bert_config_ex_s3.json  
  -vocab ./config_and_vocab/exBERT/exBERT_vocab.txt 
  -pm_p path_to_state_dict_of_the_OFF_THE_SHELF_MODEL
  -dp path_to_your_training_data
  -ls 128 
  -p 1
  -t_ex_only ""

-t_ex_only "" enable training the whole model

Pre-train an exBERT with no extension of vocab

python Pretraining.py -e 1 
  -b 256 
  -sp path_to_storage
  -dv 0 1 2 3 -lr 1e-04 
  -str exBERT    
  -config path_to_config_file_of_the_OFF_THE_SHELF_MODEL config_and_vocab/exBERT_no_ex_vocab/bert_config_ex_s3.json
  -vocab path_to_vocab_file_of_the_OFF_THE_SHELF_MODEL
  -pm_p path_to_state_dict_of_the_OFF_THE_SHELF_MODEL
  -dp path_to_your_training_data
  -ls 128 
  -p 1
  -t_ex_only ""

Data preparation

Input data for pre-training script should be a .pkl file which contains two a list with two elements, e.g. [list1, list2]. list1 and list2 should contains the sentences like [CLS] sentence A [SEP] sentence B [SEP]. The only differnece between list1 and list2 is the relationship between sentence A and sentence B is IsNext or NotNext. Please check example_data.pkl

We also provide a simple script to generate the data from raw text file. python data_preprocess.py -voc path_to_vocab_file -ls 128 -dp path_to_txt_file -n_c 5 -rd 1 -sp ./your_data.pkl replace 128 to the max length limit you want try python data_preprocess.py -voc ./exBERT_vocab.txt -ls 128 -dp ./example_raw_text.txt -n_c 5 -rd 1 -sp ./example_data.pkl

Or you can do your own data preparation and organize the data with the format metioned above.