***** New February 7th, 2019: TfHub Module *****

BERT has been uploaded to TensorFlow Hub. See run_classifier_with_tfhub.py for an example of how to use the TF Hub module, or run an example in the browser on Colab.

***** New November 23rd, 2018: Un-normalized multilingual model + Thai + Mongolian *****

We uploaded a new multilingual model which does not perform any normalization on the input (no lower casing, accent stripping, or Unicode normalization), and additionally inclues Thai and Mongolian.

It is recommended to use this version for developing multilingual models, especially on languages with non-Latin alphabets.

This does not require any code changes, and can be downloaded here:

• BERT-Base, Multilingual Cased: 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters

***** New November 15th, 2018: SOTA SQuAD 2.0 System *****

We released code changes to reproduce our 83% F1 SQuAD 2.0 system, which is currently 1st place on the leaderboard by 3%. See the SQuAD 2.0 section of the README for details.

***** New November 5th, 2018: Third-party PyTorch and Chainer versions of BERT available *****

NLP researchers from HuggingFace made a PyTorch version of BERT available which is compatible with our pre-trained checkpoints and is able to reproduce our results. Sosuke Kobayashi also made a Chainer version of BERT available (Thanks!) We were not involved in the creation or maintenance of the PyTorch implementation so please direct any questions towards the authors of that repository.

***** New November 3rd, 2018: Multilingual and Chinese models available *****

We have made two new BERT models available:

• BERT-Base, Multilingual (Not recommended, use Multilingual Cased instead): 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Chinese: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

We use character-based tokenization for Chinese, and WordPiece tokenization for all other languages. Both models should work out-of-the-box without any code changes. We did update the implementation of BasicTokenizer in tokenization.py to support Chinese character tokenization, so please update if you forked it. However, we did not change the tokenization API.

For more, see the Multilingual README.

***** End new information *****

BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks.

Our 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.

To give a few numbers, here are the results on the SQuAD v1.1 question answering task:

And several natural language inference tasks:

Plus many other tasks.

Moreover, these results were all obtained with almost no task-specific neural network architecture design.

If you already know what BERT is and you just want to get started, you can download the pre-trained models and run a state-of-the-art fine-tuning in only a few minutes.

BERT is a method of pre-training language representations, meaning that we train a general-purpose "language understanding" model on a large text corpus (like Wikipedia), and then use that model for downstream NLP tasks that we care about (like question answering). BERT outperforms previous methods because it is the first unsupervised, deeply bidirectional system for pre-training NLP.

Unsupervised means that BERT was trained using only a plain text corpus, which is important because an enormous amount of plain text data is publicly available on the web in many languages.

Pre-trained representations can also either be context-free or contextual, and contextual representations can further be unidirectional or bidirectional. Context-free models such as word2vec or GloVe generate a single "word embedding" representation for each word in the vocabulary, so bank would have the same representation in bank deposit and river bank. Contextual models instead generate a representation of each word that is based on the other words in the sentence.

BERT was built upon recent work in pre-training contextual representations — including Semi-supervised Sequence Learning, Generative Pre-Training, ELMo, and ULMFit — but crucially these models are all unidirectional or shallowly bidirectional. This means that each word is only contextualized using the words to its left (or right). For example, in the sentence I made a bank deposit the unidirectional representation of bank is only based on I made a but not deposit. Some previous work does combine the representations from separate left-context and right-context models, but only in a "shallow" manner. BERT represents "bank" using both its left and right context — I made a ... deposit — starting from the very bottom of a deep neural network, so it is deeply bidirectional.

BERT uses a simple approach for this: We mask out 15% of the words in the input, run the entire sequence through a deep bidirectional Transformer encoder, and then predict only the masked words. For example:

Input: the man went to the [MASK1] . he bought a [MASK2] of milk.
Labels: [MASK1] = store; [MASK2] = gallon

In order to learn relationships between sentences, we also train on a simple task which can be generated from any monolingual corpus: Given two sentences A and B, is B the actual next sentence that comes after A, or just a random sentence from the corpus?

Sentence A: the man went to the store .
Sentence B: he bought a gallon of milk .
Label: IsNextSentence

Sentence A: the man went to the store .
Sentence B: penguins are flightless .
Label: NotNextSentence

We then train a large model (12-layer to 24-layer Transformer) on a large corpus (Wikipedia + BookCorpus) for a long time (1M update steps), and that's BERT.

Using BERT has two stages: Pre-training and fine-tuning.

Pre-training is fairly expensive (four days on 4 to 16 Cloud TPUs), but is a one-time procedure for each language (current models are English-only, but multilingual models will be released in the near future). We are releasing a number of pre-trained models from the paper which were pre-trained at Google. Most NLP researchers will never need to pre-train their own model from scratch.

Fine-tuning is inexpensive. All of the results in the paper can be replicated in at most 1 hour on a single Cloud TPU, or a few hours on a GPU, starting from the exact same pre-trained model. SQuAD, for example, can be trained in around 30 minutes on a single Cloud TPU to achieve a Dev F1 score of 91.0%, which is the single system state-of-the-art.

The other important aspect of BERT is that it can be adapted to many types of NLP tasks very easily. In the paper, we demonstrate state-of-the-art results on sentence-level (e.g., SST-2), sentence-pair-level (e.g., MultiNLI), word-level (e.g., NER), and span-level (e.g., SQuAD) tasks with almost no task-specific modifications.

We are releasing the following:

• TensorFlow code for the BERT model architecture (which is mostly a standard Transformer architecture).
• Pre-trained checkpoints for both the lowercase and cased version of BERT-Base and BERT-Large from the paper.
• TensorFlow code for push-button replication of the most important fine-tuning experiments from the paper, including SQuAD, MultiNLI, and MRPC.

All of the code in this repository works out-of-the-box with CPU, GPU, and Cloud TPU.

We are releasing the BERT-Base and BERT-Large models from the paper. Uncased means that the text has been lowercased before WordPiece tokenization, e.g., John Smith becomes john smith. The Uncased model also strips out any accent markers. Cased means that the true case and accent markers are preserved. Typically, the Uncased model is better unless you know that case information is important for your task (e.g., Named Entity Recognition or Part-of-Speech tagging).

These models are all released under the same license as the source code (Apache 2.0).

For information about the Multilingual and Chinese model, see the Multilingual README.

When using a cased model, make sure to pass --do_lower=False to the training scripts. (Or pass do_lower_case=False directly to FullTokenizer if you're using your own script.)

The links to the models are here (right-click, 'Save link as...' on the name):

• BERT-Base, Uncased: 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Large, Uncased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Cased: 12-layer, 768-hidden, 12-heads , 110M parameters
• BERT-Large, Cased: 24-layer, 1024-hidden, 16-heads, 340M parameters
• BERT-Base, Multilingual Cased (New, recommended): 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Multilingual Uncased (Orig, not recommended) (Not recommended, use Multilingual Cased instead): 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
• BERT-Base, Chinese: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

Each .zip file contains three items:

• A TensorFlow checkpoint (bert_model.ckpt) containing the pre-trained weights (which is actually 3 files).
• A vocab file (vocab.txt) to map WordPiece to word id.
• A config file (bert_config.json) which specifies the hyperparameters of the model.

Important: All results on the paper were fine-tuned on a single Cloud TPU, which has 64GB of RAM. It is currently not possible to re-produce most of the BERT-Large results on the paper using a GPU with 12GB - 16GB of RAM, because the maximum batch size that can fit in memory is too small. We are working on adding code to this repository which allows for much larger effective batch size on the GPU. See the section on out-of-memory issues for more details.

This code was tested with TensorFlow 1.11.0. It was tested with Python2 and Python3 (but more thoroughly with Python2, since this is what's used internally in Google).

The fine-tuning examples which use BERT-Base should be able to run on a GPU that has at least 12GB of RAM using the hyperparameters given.

Most of the examples below assumes that you will be running training/evaluation on your local machine, using a GPU like a Titan X or GTX 1080.

However, if you have access to a Cloud TPU that you want to train on, just add the following flags to run_classifier.py or run_squad.py:

  --use_tpu=True \
  --tpu_name=$TPU_NAME

Please see the Google Cloud TPU tutorial for how to use Cloud TPUs. Alternatively, you can use the Google Colab notebook "BERT FineTuning with Cloud TPUs".

On Cloud TPUs, the pretrained model and the output directory will need to be on Google Cloud Storage. For example, if you have a bucket named some_bucket, you might use the following flags instead:

  --output_dir=gs:https://some_bucket/my_output_dir/

The unzipped pre-trained model files can also be found in the Google Cloud Storage folder gs:https://bert_models/2018_10_18. For example:

export BERT_BASE_DIR=gs:https://bert_models/2018_10_18/uncased_L-12_H-768_A-12

Before running this example you must download the GLUE data by running this script and unpack it to some directory $GLUE_DIR. Next, download the BERT-Base checkpoint and unzip it to some directory $BERT_BASE_DIR.

This example code fine-tunes BERT-Base 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.

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

python run_classifier.py \
  --task_name=MRPC \
  --do_train=true \
  --do_eval=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/bert_model.ckpt \
  --max_seq_length=128 \
  --train_batch_size=32 \
  --learning_rate=2e-5 \
  --num_train_epochs=3.0 \
  --output_dir=/tmp/mrpc_output/

You should see output like this:

***** Eval results *****
  eval_accuracy = 0.845588
  eval_loss = 0.505248
  global_step = 343
  loss = 0.505248

This means that the Dev set accuracy was 84.55%. Small sets like MRPC have a high variance in the Dev set accuracy, even when starting from the same pre-training checkpoint. If you re-run multiple times (making sure to point to different output_dir), you should see results between 84% and 88%.

A few other pre-trained models are implemented off-the-shelf in run_classifier.py, so it should be straightforward to follow those examples to use BERT for any single-sentence or sentence-pair classification task.

Note: You might see a message Running train on CPU. This really just means that it's running on something other than a Cloud TPU, which includes a GPU.

Once you have trained your classifier you can use it in inference mode by using the --do_predict=true command. You need to have a file named test.tsv in the input folder. Output will be created in file called test_results.tsv in the output folder. Each line will contain output for each sample, columns are the class probabilities.

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
export GLUE_DIR=/path/to/glue
export TRAINED_CLASSIFIER=/path/to/fine/tuned/classifier

python run_classifier.py \
  --task_name=MRPC \
  --do_predict=true \
  --data_dir=$GLUE_DIR/MRPC \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$TRAINED_CLASSIFIER \
  --max_seq_length=128 \
  --output_dir=/tmp/mrpc_output/

The Stanford Question Answering Dataset (SQuAD) is a popular question answering benchmark dataset. BERT (at the time of the release) obtains state-of-the-art results on SQuAD with almost no task-specific network architecture modifications or data augmentation. However, it does require semi-complex data pre-processing and post-processing to deal with (a) the variable-length nature of SQuAD context paragraphs, and (b) the character-level answer annotations which are used for SQuAD training. This processing is implemented and documented in run_squad.py.

To run on SQuAD, you will first need to download the dataset. The SQuAD website does not seem to link to the v1.1 datasets any longer, but the necessary files can be found here:

• train-v1.1.json
• dev-v1.1.json
• evaluate-v1.1.py

Download these to some directory $SQUAD_DIR.

The state-of-the-art SQuAD results from the paper currently cannot be reproduced on a 12GB-16GB GPU due to memory constraints (in fact, even batch size 1 does not seem to fit on a 12GB GPU using BERT-Large). However, a reasonably strong BERT-Base model can be trained on the GPU with these hyperparameters:

python run_squad.py \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v1.1.json \
  --do_predict=True \
  --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=/tmp/squad_base/

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

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

Which should produce an output like this:

{"f1": 88.41249612335034, "exact_match": 81.2488174077578}

You should see a result similar to the 88.5% reported in the paper for BERT-Base.

If you have access to a Cloud TPU, you can train with BERT-Large. Here is a set of hyperparameters (slightly different than the paper) which consistently obtain around 90.5%-91.0% F1 single-system trained only on SQuAD:

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v1.1.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v1.1.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs:https://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME

For example, one random run with these parameters produces the following Dev scores:

{"f1": 90.87081895814865, "exact_match": 84.38978240302744}

If you fine-tune for one epoch on TriviaQA before this the results will be even better, but you will need to convert TriviaQA into the SQuAD json format.

This model is also implemented and documented in run_squad.py.

To run on SQuAD 2.0, you will first need to download the dataset. The necessary files can be found here:

• train-v2.0.json
• dev-v2.0.json
• evaluate-v2.0.py

Download these to some directory $SQUAD_DIR.

On Cloud TPU you can run with BERT-Large as follows:

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=True \
  --train_file=$SQUAD_DIR/train-v2.0.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v2.0.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs:https://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME \
  --version_2_with_negative=True

We assume you have copied everything from the output directory to a local directory called ./squad/. The initial dev set predictions will be at ./squad/predictions.json and the differences between the score of no answer ("") and the best non-null answer for each question will be in the file ./squad/null_odds.json

Run this script to tune a threshold for predicting null versus non-null answers:

python $SQUAD_DIR/evaluate-v2.0.py $SQUAD_DIR/dev-v2.0.json ./squad/predictions.json --na-prob-file ./squad/null_odds.json

Assume the script outputs "best_f1_thresh" THRESH. (Typical values are between -1.0 and -5.0). You can now re-run the model to generate predictions with the derived threshold or alternatively you can extract the appropriate answers from ./squad/nbest_predictions.json.

python run_squad.py \
  --vocab_file=$BERT_LARGE_DIR/vocab.txt \
  --bert_config_file=$BERT_LARGE_DIR/bert_config.json \
  --init_checkpoint=$BERT_LARGE_DIR/bert_model.ckpt \
  --do_train=False \
  --train_file=$SQUAD_DIR/train-v2.0.json \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-v2.0.json \
  --train_batch_size=24 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=gs:https://some_bucket/squad_large/ \
  --use_tpu=True \
  --tpu_name=$TPU_NAME \
  --version_2_with_negative=True \
  --null_score_diff_threshold=$THRESH

All experiments in the paper were fine-tuned on a Cloud TPU, which has 64GB of device RAM. Therefore, when using a GPU with 12GB - 16GB of RAM, you are likely to encounter out-of-memory issues if you use the same hyperparameters described in the paper.

The factors that affect memory usage are:

• max_seq_length: The released models were trained with sequence lengths up to 512, but you can fine-tune with a shorter max sequence length to save substantial memory. This is controlled by the max_seq_length flag in our example code.

• train_batch_size: The memory usage is also directly proportional to the batch size.

• Model type, BERT-Base vs. BERT-Large: The BERT-Large model requires significantly more memory than BERT-Base.

• Optimizer: The default optimizer for BERT is Adam, which requires a lot of extra memory to store the m and v vectors. Switching to a more memory efficient optimizer can reduce memory usage, but can also affect the results. We have not experimented with other optimizers for fine-tuning.

Using the default training scripts (run_classifier.py and run_squad.py), we benchmarked the maximum batch size on single Titan X GPU (12GB RAM) with TensorFlow 1.11.0: