Home

Awesome

Chainer implementation of Google AI's BERT model with a script to load Google's pre-trained models

This repository contains a Chainer reimplementation of Google's TensorFlow repository for the BERT model for the paper BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.

This implementation can load any pre-trained TensorFlow checkpoint for BERT (in particular Google's pre-trained models) and a conversion script is provided (see below).

In the current implementation, we can

Not implemented:

This README follows the great README of PyTorch's BERT repository by the huggingface team.

Loading a TensorFlow checkpoint (e.g. Google's pre-trained models)

You can convert any TensorFlow checkpoint for BERT (in particular the pre-trained models released by Google) in a Chainer save file by using the convert_tf_checkpoint_to_chainer.py script.

This script takes as input a TensorFlow checkpoint (three files starting with bert_model.ckpt) and creates a Chainer model (npz file) for this configuration, so that we can load the models using chainer.serializers.load_npz() by Chainer. (see examples in run_classifier.py or run_squad.py)

You only need to run this conversion script once to get a Chainer model. You can then disregard the TensorFlow checkpoint (the three files starting with bert_model.ckpt) but be sure to keep the configuration file (bert_config.json) and the vocabulary file (vocab.txt) as these are needed for the Chainer model too.

To run this specific conversion script you will need to have TensorFlow and Chainer installed (pip install tensorflow). The rest of the repository only requires Chainer.

Here is an example of the conversion process for a pre-trained BERT-Base Uncased model:

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12

python convert_tf_checkpoint_to_chainer.py \
  --tf_checkpoint_path $BERT_BASE_DIR/bert_model.ckpt \
  --npz_dump_path $BERT_BASE_DIR/arrays_bert_model.ckpt.npz

You can download Google's pre-trained models for the conversion here.

Chainer models for BERT

We included two Chainer models in this repository that you will find in modeling.py:

Here are some details on each class.

1. BertModel

BertModel is the basic BERT Transformer model with a layer of summed token, position and sequence embeddings followed by a series of identical self-attention blocks (12 for BERT-base, 24 for BERT-large).

The inputs and output are identical to the TensorFlow model inputs and outputs.

We detail them here. This model takes as inputs:

This model can return some kinds of outputs (by calling like .get_pooled_output()):

An example on how to use this class is given in the extract_features.py script which can be used to extract the hidden states of the model for a given input.

2. BertClassifier

BertClassifier is a fine-tuning model that includes BertModel and a sequence-level (sequence or pair of sequences) classifier on top of the BertModel.

The sequence-level classifier is a linear layer that takes as input the last hidden state of the first character in the input sequence (see Figures 3a and 3b in the BERT paper).

An example on how to use this class is given in the run_classifier.py script which can be used to fine-tune a single sequence (or pair of sequence) classifier using BERT, for example for the MRPC task.

3. BertForQuestionAnswering

BertSQuAD is a fine-tuning model that includes BertModel with a token-level classifiers on top of the full sequence of last hidden states.

The token-level classifier takes as input the full sequence of the last hidden state and compute several (e.g. two) scores for each tokens that can for example respectively be the score that a given token is a start_span and a end_span token (see Figures 3c and 3d in the BERT paper).

An example on how to use this class is given in the run_squad.py script which can be used to fine-tune a token classifier using BERT, for example for the SQuAD task.

Installation, requirements

This code was tested on Python 3.5+. The requirements are:

Fine-tuning with BERT: running the examples

We showcase the same examples as the original implementation: fine-tuning a sequence-level classifier on the MRPC classification corpus and a token-level classifier on the question answering dataset SQuAD.

Prepare the pretrained BERT model

First of all, please also download the BERT-Base checkpoint, unzip it to some directory $BERT_BASE_DIR, and convert it to its Chainer version as explained in the previous section.

wget https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip
unzip uncased_L-12_H-768_A-12.zip
export BERT_BASE_DIR=./uncased_L-12_H-768_A-12
python convert_tf_checkpoint_to_chainer.py \
  --tf_checkpoint_path $BERT_BASE_DIR/bert_model.ckpt \
  --npz_dump_path $BERT_BASE_DIR/arrays_bert_model.ckpt.npz

Sentence (or Pair) Classification with GLUE Dataset

Prepare GLUE dataset

Before running theses examples you should download the GLUE data by running this script and unpack it to some directory $GLUE_DIR.

wget https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py
python download_glue_data.py
export GLUE_DIR=./glue_data

Train and evaluate

This example code fine-tunes BERT-Base on the Microsoft Research Paraphrase Corpus (MRPC) corpus and runs in less than several minutes on a single Tesla P100.

python run_classifier.py \
  --task_name MRPC \
  --do_train True \
  --do_eval True \
  --do_lower_case True \
  --data_dir $GLUE_DIR/MRPC/ \
  --vocab_file $BERT_BASE_DIR/vocab.txt \
  --bert_config_file $BERT_BASE_DIR/bert_config.json \
  --init_checkpoint $BERT_BASE_DIR/arrays_bert_model.ckpt.npz \
  --max_seq_length 128 \
  --train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir=./mrpc_output

Our test run on a few seeds with the original implementation hyper-parameters gave evaluation results around 86.

Token-level Classification with SQuAD QA Dataset

Prepare SQuAD dataset

The data for SQuAD can be downloaded with the following links and should be saved in a $SQUAD_DIR directory. This runs in less than several hours on a single Tesla P100.

export SQUAD_DIR=/path/to/SQUAD

python run_squad.py \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/arrays_bert_model.ckpt.npz \
  --do_train=True \
  --do_predict=True \
  --train_file=$SQUAD_DIR/train-v1.1.json \
  --predict_file=$SQUAD_DIR/dev-v1.1.json \
  --train_batch_size=12 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=./squad_output

Training with the previous hyper-parameters gave us the following results:

{"exact_match": 79.81078524124882, "f1": 87.74743306449187}

The result was little worse than the original repository reported (e.g. f1=88.4).