Skip to content

Latest commit

 

History

History
 
 

bert

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
__init__.py
__init__.py
 
 
classifier_data_lib.py
classifier_data_lib.py
 
 
common_flags.py
common_flags.py
 
 
create_finetuning_data.py
create_finetuning_data.py
 
 
create_pretraining_data.py
create_pretraining_data.py
 
 
export_tfhub.py
export_tfhub.py
 
 
export_tfhub_test.py
export_tfhub_test.py
 
 
input_pipeline.py
input_pipeline.py
 
 
model_saving_utils.py
model_saving_utils.py
 
 
run_classifier.py
run_classifier.py
 
 
run_pretraining.py
run_pretraining.py
 
 
run_squad.py
run_squad.py
 
 
squad_lib.py
squad_lib.py
 
 
tf1_checkpoint_converter_lib.py
tf1_checkpoint_converter_lib.py
 
 
tf1_to_keras_checkpoint_converter.py
tf1_to_keras_checkpoint_converter.py
 
 
tf2_checkpoint_converter.py
tf2_checkpoint_converter.py
 
 
tokenization.py
tokenization.py
 
 
tokenization_test.py
tokenization_test.py
 
 

README.md

BERT (Bidirectional Encoder Representations from Transformers)

The academic paper which describes BERT in detail and provides full results on a number of tasks can be found here: https://arxiv.org/abs/1810.04805.

This repository contains TensorFlow 2 implementation for BERT.

N.B. This repository is under active development. Though we intend to keep the top-level BERT Keras model interface stable, expect continued changes to the training code, utility function interface and flags.

Contents

Pre-trained Models

Our current released checkpoints are exactly the same as TF 1.x official BERT repository, thus inside BertConfig, there is backward_compatible=True. We are going to release new pre-trained checkpoints soon.

Access to Pretrained Checkpoints

We provide checkpoints that are converted from google-research/bert, in order to keep consistent with BERT paper.

The stable model checkpoints work with v2.0 release.

Note: these checkpoints are not compatible with the current master run_classifier.py example.

Note: We are in the middle of a transition stage to switch BERT implementation to use Keras functional-style networks in nlp/modeling. The checkpoint above will be deleted once transition is done.

The new checkpoints work with run_classifier.py example are:

We recommend to host checkpoints on Google Cloud storage buckets when you use Cloud GPU/TPU.

Restoring from Checkpoints

tf.train.Checkpoint is used to manage model checkpoints in TF 2. To restore weights from provided pre-trained checkpoints, you can use the following code:

init_checkpoint='the pretrained model checkpoint path.'
model=tf.keras.Model() # Bert pre-trained model as feature extractor.
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(init_checkpoint)

Checkpoints featuring native serialized Keras models (i.e. model.load()/load_weights()) will be available soon.

Set Up

export PYTHONPATH="$PYTHONPATH:/path/to/models"

Install tf-nightly to get latest updates:

pip install tf-nightly-gpu

With TPU, GPU support is not necessary. First, you need to create a tf-nigthly TPU with cptu tool:

ctpu up -name <instance name> --tf-version=”nightly”

Second, you need to install TF 2 tf-night on your VM:

pip install tf-nightly

Warning: More details TPU-specific set-up instructions and tutorial should come along with official TF 2.x release for TPU. Note that this repo is not officially supported by Google Cloud TPU team yet.

Process Datasets

Pre-training

There is no change to generate pre-training data. Please use the script create_pretraining_data.py which is essentially branched from BERT research repo to get processed pre-training data and it adapts to TF2 symbols and python3 compatibility.

Fine-tuning

To prepare the fine-tuning data for final model training, use the create_finetuning_data.py script. Resulting datasets in tf_record format and training meta data should be later passed to training or evaluation scripts. The task-specific arguments are described in following sections:

  • GLUE

Users can download the GLUE data by running this script and unpack it to some directory $GLUE_DIR.

export GLUE_DIR=~/glue
export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16

export TASK_NAME=MNLI
export OUTPUT_DIR=gs://some_bucket/datasets
python create_finetuning_data.py \
 --input_data_dir=${GLUE_DIR}/${TASK_NAME}/ \
 --vocab_file=${BERT_BASE_DIR}/vocab.txt \
 --train_data_output_path=${OUTPUT_DIR}/${TASK_NAME}_train.tf_record \
 --eval_data_output_path=${OUTPUT_DIR}/${TASK_NAME}_eval.tf_record \
 --meta_data_file_path=${OUTPUT_DIR}/${TASK_NAME}_meta_data \
 --fine_tuning_task_type=classification --max_seq_length=128 \
 --classification_task_name=${TASK_NAME}
  • SQUAD

The SQuAD website contains detailed information about the SQuAD datasets and evaluation.

The necessary files can be found here:

export SQUAD_DIR=~/squad
export SQUAD_VERSION=v1.1
export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16
export OUTPUT_DIR=gs://some_bucket/datasets

python create_finetuning_data.py \
 --squad_data_file=${SQUAD_DIR}/train-${SQUAD_VERSION}.json \
 --vocab_file=${BERT_BASE_DIR}/vocab.txt \
 --train_data_output_path=${OUTPUT_DIR}/squad_${SQUAD_VERSION}_train.tf_record \
 --meta_data_file_path=${OUTPUT_DIR}/squad_${SQUAD_VERSION}_meta_data \
 --fine_tuning_task_type=squad --max_seq_length=384

Fine-tuning with BERT

Cloud GPUs and TPUs

  • Cloud Storage

The unzipped pre-trained model files can also be found in the Google Cloud Storage folder gs://cloud-tpu-checkpoints/bert/tf_20. For example:

export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16
export MODEL_DIR=gs://some_bucket/my_output_dir

Currently, users are able to access to tf-nightly TPUs and the following TPU script should run with tf-nightly.

  • GPU -> TPU

Just add the following flags to run_classifier.py or run_squad.py:

  --distribution_strategy=tpu
  --tpu=grpc://${TPU_IP_ADDRESS}:8470

Sentence and Sentence-pair Classification Tasks

This example code fine-tunes BERT-Large on the Microsoft Research Paraphrase Corpus (MRPC) corpus, which only contains 3,600 examples and can fine-tune in a few minutes on most GPUs.

We use the BERT-Large (uncased_L-24_H-1024_A-16) as an example throughout the workflow. For GPU memory of 16GB or smaller, you may try to use BERT-Base (uncased_L-12_H-768_A-12).

export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16
export MODEL_DIR=gs://some_bucket/my_output_dir
export GLUE_DIR=gs://some_bucket/datasets
export TASK=MRPC

python run_classifier.py \
  --mode='train_and_eval' \
  --input_meta_data_path=${GLUE_DIR}/${TASK}_meta_data \
  --train_data_path=${GLUE_DIR}/${TASK}_train.tf_record \
  --eval_data_path=${GLUE_DIR}/${TASK}_eval.tf_record \
  --bert_config_file=${BERT_BASE_DIR}/bert_config.json \
  --init_checkpoint=${BERT_BASE_DIR}/bert_model.ckpt \
  --train_batch_size=4 \
  --eval_batch_size=4 \
  --steps_per_loop=1 \
  --learning_rate=2e-5 \
  --num_train_epochs=3 \
  --model_dir=${MODEL_DIR} \
  --distribution_strategy=mirror

To use TPU, you only need to switch distribution strategy type to tpu with TPU information and use remote storage for model checkpoints.

export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16
export TPU_IP_ADDRESS='???'
export MODEL_DIR=gs://some_bucket/my_output_dir
export GLUE_DIR=gs://some_bucket/datasets

python run_classifier.py \
  --mode='train_and_eval' \
  --input_meta_data_path=${GLUE_DIR}/${TASK}_meta_data \
  --train_data_path=${GLUE_DIR}/${TASK}_train.tf_record \
  --eval_data_path=${GLUE_DIR}/${TASK}_eval.tf_record \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --train_batch_size=32 \
  --eval_batch_size=32 \
  --learning_rate=2e-5 \
  --num_train_epochs=3 \
  --model_dir=${MODEL_DIR} \
  --distribution_strategy=tpu \
  --tpu=grpc://${TPU_IP_ADDRESS}:8470

SQuAD 1.1

The Stanford Question Answering Dataset (SQuAD) is a popular question answering benchmark dataset. See more in SQuAD website.

We use the BERT-Large (uncased_L-24_H-1024_A-16) as an example throughout the workflow. For GPU memory of 16GB or smaller, you may try to use BERT-Base (uncased_L-12_H-768_A-12).

export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16
export SQUAD_DIR=gs://some_bucket/datasets
export MODEL_DIR=gs://some_bucket/my_output_dir
export SQUAD_VERSION=v1.1

python run_squad.py \
  --input_meta_data_path=${SQUAD_DIR}/squad_${SQUAD_VERSION}_meta_data \
  --train_data_path=${SQUAD_DIR}/squad_${SQUAD_VERSION}_train.tf_record \
  --predict_file=${SQUAD_DIR}/dev-v1.1.json \
  --vocab_file=${BERT_BASE_DIR}/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --train_batch_size=4 \
  --predict_batch_size=4 \
  --learning_rate=8e-5 \
  --num_train_epochs=2 \
  --model_dir=${MODEL_DIR} \
  --distribution_strategy=mirror

To use TPU, you need switch distribution strategy type to tpu with TPU information.

export BERT_BASE_DIR=gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16
export TPU_IP_ADDRESS='???'
export MODEL_DIR=gs://some_bucket/my_output_dir
export SQUAD_DIR=gs://some_bucket/datasets
export SQUAD_VERSION=v1.1

python run_squad.py \
  --input_meta_data_path=${SQUAD_DIR}/squad_${SQUAD_VERSION}_meta_data \
  --train_data_path=${SQUAD_DIR}/squad_${SQUAD_VERSION}_train.tf_record \
  --predict_file=${SQUAD_DIR}/dev-v1.1.json \
  --vocab_file=${BERT_BASE_DIR}/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --train_batch_size=32 \
  --learning_rate=8e-5 \
  --num_train_epochs=2 \
  --model_dir=${MODEL_DIR} \
  --distribution_strategy=tpu \
  --tpu=grpc://${TPU_IP_ADDRESS}:8470

The dev set predictions will be saved into a file called predictions.json in the model_dir:

python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ./squad/predictions.json