BERT¶

BERT is an autoencoding language model with a final loss composed of 1. masked language model loss and 2. next sentence prediction. The model architecture is published in BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding [NLP-BERT1]. The model is originally trained on English Wikipedia and BookCorpus. BERT is often used an language model encoder for downstream tasks, e.g. Token Classification (Named Entity Recognition) Model, Text Classification Model, Question Answering Model, etc. Domain-specific BERT models can be advantageous for a wide range of applications. One notable is domain-specific BERT in a biomedical setting, e.g. BioBERT [NLP-BERT2] or its improved derivative BioMegatron [NLP-BERT3]. For the latter refer to Megatron-LM for Downstream Tasks.

Quick Start¶

from nemo.collections.nlp.models import BERTLMModel

# to get the list of pre-trained models
BERTLMModel.list_available_models()

# Download and load the pre-trained BERT-based model
model = BERTLMModel.from_pretrained("bertbaseuncased")

Available Models¶

*Pretrained Models*¶
Model	Pretrained Checkpoint
bertbaseuncased	https://ngc.nvidia.com/catalog/models/nvidia:nemo:bertbaseuncased
bertlargeuncased	https://ngc.nvidia.com/catalog/models/nvidia:nemo:bertlargeuncased

Data Input for BERT Model¶

Data preprocessing can be either done on the fly during training or offline before training. The latter is optimized and recommended for large text corpora. This was also used in original paper to train the model on Wikipedia and BookCorpus. For on the fly data processing, provide text files with sentences for training and validation, where words are separated by spaces, i.e.: [WORD] [SPACE] [WORD] [SPACE] [WORD]. To use this pipeline in training use the dedicated configuration file NeMo/examples/nlp/language_modeling/conf/bert_pretraining_from_preprocessed_config.yaml.

To process data offline in advance, go to https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling/BERT#quick-start-guide. To recreate the original Wikipedia and BookCorpus datasets, follow steps 1-5 in the Quick-Start-Guide and run the script ./data/create_datasets_from_start.sh inside the docker container. The downloaded folder should include two sub folders lower_case_[0,1]_seq_len_128_max_pred_20_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5 and lower_case_[0,1]_seq_len_512_max_pred_80_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5, containing sequences of length 128 with a maximum of 20 masked tokens and sequences of length 512 with a maximum of 80 masked tokens respectively. To use this pipeline in training use the dedicated configuration file NeMo/examples/nlp/language_modeling/conf/bert_pretraining_from_text_config.yaml and specify the path to the created hd5f files.

Training BERT Model¶

Example of model configuration for on the fly data preprocessing: NeMo/examples/nlp/language_modeling/conf/bert_pretraining_from_text_config.yaml. Example of model configuration for offline data preprocessing: NeMo/examples/nlp/language_modeling/conf/bert_pretraining_from_preprocessed_config.yaml.

The specification can be roughly grouped into three categories:

Parameters that describe the training process: trainer
Parameters that describe the datasets: model.train_ds, model.validation_ds
Parameters that describe the model: model, model.tokenizer, model.language_model

More details about parameters in the config file:

Parameter	Data Type	Description
model.only_mlm_loss	bool	Only uses masked language model without next sentence prediction
train_ds.data_file	string	Name of the text file or hdf5 data directory
train_ds.num_samples	integer	Number of samples to use from the training dataset, -1 mean all

More details about parameters for offline data preprocessing:

Parameter	Data Type	Description
train_ds.max_predictions_per_seq	integer	Maximum number of masked tokens in a sequence in the preprocessed data

More details about parameters for online data preprocessing:

Parameter	Data Type	Description
model.max_seq_length	integer	The maximum total input sequence length after tokenization
model.mask_prob	float	Probability of masking a token in the input text during data processing
model.short_seq_prob	float	Probability of having a sequence shorter than the maximum sequence length

Note

For offline data preprocessing model.tokenizer is null. For downstream task use the same tokenizer that was used for offline preprocessing.

For online data preprocessing model.tokenizer needs to be specified. See also Model NLP for details.

Example of the command for training the model:

python bert_pretraining.py \
       model.train_ds.data_file=<PATH_TO_DATA>  \
       trainer.max_epochs=<NUM_EPOCHS> \
       trainer.gpus=[<CHANGE_TO_GPU(s)_YOU_WANT_TO_USE>]

Finetuning on Downstream Tasks¶

To use trained BERT model checkpoint on an NeMo NLP downstream task, e.g. Question Answering Model, specify model.language_model.lm_checkpoint=<PATH_TO_CHECKPOINT>.

References¶

NLP-BERT1: Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
NLP-BERT2: Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. Biobert: a pre-trained biomedical language representation model for biomedical text mining. 2019. arXiv:1901.08746.
NLP-BERT3: Hoo-Chang Shin, Yang Zhang, Evelina Bakhturina, Raul Puri, Mostofa Patwary, Mohammad Shoeybi, and Raghav Mani. Biomegatron: larger biomedical domain language model. 2020. arXiv:2010.06060.