从头训练MASK BERT
In most cases, the Google pretrained BERT model or a further fine-tuning base on it is enough. Howerver, sometimes maybe you want to train a BERT or a small BERT with your specific dataset. This case is few but it does exist. This repo provides a mask BERT pretraining code, which has not been implemented in Pytorch at some popular repos of BERT, such as huggingface. Note that in this repo, the next sentence prediction is not jointed to pretraining.
- Preparing your dataset, train and dev, and place them into the data directionary
- Building a vocab as bert_vocab.txt. Note that the vocab should include [PAD] [UNK] [CLS] [SEP] [MASK]
- Change the bert_config.json. Alternatively, you can change the model size to fit your memory. The current num_layer is set to 4, same as the MASS model (k=1) by microsoft. Besides, you also should change the vocab_size in the config file based on your vocab
- python run_pretraining.py