PyTorch Pretrained Bert


This repository contains an op-for-op PyTorch reimplementation of Google's TensorFlow repository for the BERT model that was released together with 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 is provided with Google's pre-trained models, examples, notebooks and a command-line interface to load any pre-trained TensorFlow checkpoint for BERT is also provided.


Section Description
Installation How to install the package
Overview Overview of the package
Usage Quickstart examples
Doc Detailed documentation
Examples Detailed examples on how to fine-tune Bert
Notebooks Introduction on the provided Jupyter Notebooks
TPU Notes on TPU support and pretraining scripts
Command-line interface Convert a TensorFlow checkpoint in a PyTorch dump


This repo was tested on Python 3.5+ and PyTorch 0.4.1/1.0.0

With pip

PyTorch pretrained bert can be installed by pip as follows:

pip install pytorch-pretrained-bert

From source

Clone the repository and run:

pip install [--editable] .

A series of tests is included in the tests folder and can be run using pytest (install pytest if needed: pip install pytest).

You can run the tests with the command:

python -m pytest -sv tests/


This package comprises the following classes that can be imported in Python and are detailed in the Doc section of this readme:

The repository further comprises:


Here is a quick-start example using BertTokenizer, BertModel and BertForMaskedLM class with Google AI's pre-trained Bert base uncased model. See the doc section below for all the details on these classes.

First let's prepare a tokenized input with BertTokenizer

import torch
from pytorch_pretrained_bert import BertTokenizer, BertModel, BertForMaskedLM

# Load pre-trained model tokenizer (vocabulary)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

# Tokenized input
text = "Who was Jim Henson ? Jim Henson was a puppeteer"
tokenized_text = tokenizer.tokenize(text)

# Mask a token that we will try to predict back with `BertForMaskedLM`
masked_index = 6
tokenized_text[masked_index] = '[MASK]'
assert tokenized_text == ['who', 'was', 'jim', 'henson', '?', 'jim', '[MASK]', 'was', 'a', 'puppet', '##eer']

# Convert token to vocabulary indices
indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
# Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
segments_ids = [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1]

# Convert inputs to PyTorch tensors
tokens_tensor = torch.tensor([indexed_tokens])
segments_tensors = torch.tensor([segments_ids])

Let's see how to use BertModel to get hidden states

# Load pre-trained model (weights)
model = BertModel.from_pretrained('bert-base-uncased')

# Predict hidden states features for each layer
encoded_layers, _ = model(tokens_tensor, segments_tensors)
# We have a hidden states for each of the 12 layers in model bert-base-uncased
assert len(encoded_layers) == 12

And how to use BertForMaskedLM

# Load pre-trained model (weights)
model = BertForMaskedLM.from_pretrained('bert-base-uncased')

# Predict all tokens
predictions = model(tokens_tensor, segments_tensors)

# confirm we were able to predict 'henson'
predicted_index = torch.argmax(predictions[0, masked_index]).item()
predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
assert predicted_token == 'henson'


Here is a detailed documentation of the classes in the package and how to use them:

Sub-section Description
Loading Google AI's pre-trained weigths How to load Google AI's pre-trained weight or a PyTorch saved instance
PyTorch models API of the eight PyTorch model classes: BertModel, BertForMaskedLM, BertForNextSentencePrediction, BertForPreTraining, BertForSequenceClassification, BertForMultipleChoice or BertForQuestionAnswering
Tokenizer: BertTokenizer API of the BertTokenizer class
Optimizer: BertAdam API of the BertAdam class

Loading Google AI's pre-trained weigths and PyTorch dump

To load one of Google AI's pre-trained models or a PyTorch saved model (an instance of BertForPreTraining saved with, the PyTorch model classes and the tokenizer can be instantiated as

model = BERT_CLASS.from_pretrained(PRE_TRAINED_MODEL_NAME_OR_PATH, cache_dir=None)


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). For information about the Multilingual and Chinese model, see the Multilingual README or the original TensorFlow repository.

When using an uncased model, make sure to pass --do_lower_case to the example training scripts (or pass do_lower_case=True to FullTokenizer if you're using your own script and loading the tokenizer your-self.).


tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
model = BertForSequenceClassification.from_pretrained('bert-base-uncased')

PyTorch models

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 outputs a tuple composed of:

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

2. BertForPreTraining

BertForPreTraining includes the BertModel Transformer followed by the two pre-training heads:

Inputs comprises the inputs of the BertModel class plus two optional labels:


An example on how to use this class is given in the script which can be used to fine-tune the BERT language model on your specific different text corpus. This should improve model performance, if the language style is different from the original BERT training corpus (Wiki + BookCorpus).

3. BertForMaskedLM

BertForMaskedLM includes the BertModel Transformer followed by the (possibly) pre-trained masked language modeling head.

Inputs comprises the inputs of the BertModel class plus optional label:


4. BertForNextSentencePrediction

BertForNextSentencePrediction includes the BertModel Transformer followed by the next sentence classification head.

Inputs comprises the inputs of the BertModel class plus an optional label:


5. BertForSequenceClassification

BertForSequenceClassification 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 script which can be used to fine-tune a single sequence (or pair of sequence) classifier using BERT, for example for the MRPC task.

6. BertForMultipleChoice

BertForMultipleChoice is a fine-tuning model that includes BertModel and a linear layer on top of the BertModel.

The linear layer outputs a single value for each choice of a multiple choice problem, then all the outputs corresponding to an instance are passed through a softmax to get the model choice.

This implementation is largely inspired by the work of OpenAI in Improving Language Understanding by Generative Pre-Training and the answer of Jacob Devlin in the following issue.

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

7. BertForTokenClassification

BertForTokenClassification is a fine-tuning model that includes BertModel and a token-level classifier on top of the BertModel.

The token-level classifier is a linear layer that takes as input the last hidden state of the sequence.

8. BertForQuestionAnswering

BertForQuestionAnswering 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 script which can be used to fine-tune a token classifier using BERT, for example for the SQuAD task.

Tokenizer: BertTokenizer

BertTokenizer perform end-to-end tokenization, i.e. basic tokenization followed by WordPiece tokenization.

This class has two arguments:

and three methods:

Please refer to the doc strings and code in for the details of the BasicTokenizer and WordpieceTokenizer classes. In general it is recommended to use BertTokenizer unless you know what you are doing.

Optimizer: BertAdam

BertAdam is a torch.optimizer adapted to be closer to the optimizer used in the TensorFlow implementation of Bert. The differences with PyTorch Adam optimizer are the following:

The optimizer accepts the following arguments:


Sub-section Description
Training large models: introduction, tools and examples How to use gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training to train Bert models
Fine-tuning with BERT: running the examples Running the examples in ./examples:,, and
Fine-tuning BERT-large on GPUs How to fine tune BERT large

Training large models: introduction, tools and examples

BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32).

To help with fine-tuning these models, we have included several techniques that you can activate in the fine-tuning scripts and gradient-accumulation, multi-gpu training, distributed training and 16-bits training . For more details on how to use these techniques you can read the tips on training large batches in PyTorch that I published earlier this month.

Here is how to use these techniques in our scripts:

To use 16-bits training and distributed training, you need to install NVIDIA's apex extension as detailed here. You will find more information regarding the internals of apex and how to use apex in the doc and the associated repository. The results of the tests performed on pytorch-BERT by the NVIDIA team (and my trials at reproducing them) can be consulted in the relevant PR of the present repository.

Note: To use Distributed Training, you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see the above mentioned blog post for more details):

python -m torch.distributed.launch --nproc_per_node=4 --nnodes=2 --node_rank=$THIS_MACHINE_INDEX --master_addr="" --master_port=1234 (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)

Where $THIS_MACHINE_INDEX is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP address and an open port 1234.

Fine-tuning with BERT: running the examples

We showcase several fine-tuning examples based on (and extended from) the original implementation:


This example code fine-tunes BERT on the Microsoft Research Paraphrase Corpus (MRPC) corpus and runs in less than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.

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

export GLUE_DIR=/path/to/glue

python \
  --task_name MRPC \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir $GLUE_DIR/MRPC/ \
  --bert_model bert-base-uncased \
  --max_seq_length 128 \
  --train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir /tmp/mrpc_output/

Our test ran on a few seeds with the original implementation hyper-parameters gave evaluation results between 84% and 88%.

Fast run with apex and 16 bit precision: fine-tuning on MRPC in 27 seconds! First install apex as indicated here. Then run

export GLUE_DIR=/path/to/glue

python \
  --task_name MRPC \
  --do_train \
  --do_eval \
  --do_lower_case \
  --data_dir $GLUE_DIR/MRPC/ \
  --bert_model bert-base-uncased \
  --max_seq_length 128 \
  --train_batch_size 32 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --output_dir /tmp/mrpc_output/


This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB.

The data for SQuAD can be downloaded with the following links and should be saved in a $SQUAD_DIR directory.

export SQUAD_DIR=/path/to/SQUAD

python \
  --bert_model bert-base-uncased \
  --do_train \
  --do_predict \
  --do_lower_case \
  --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 /tmp/debug_squad/

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

{"f1": 88.52381567990474, "exact_match": 81.22043519394512}


The data for SWAG can be downloaded by cloning the following repository

export SWAG_DIR=/path/to/SWAG

python \
  --bert_model bert-base-uncased \
  --do_train \
  --do_lower_case \
  --do_eval \
  --data_dir $SWAG_DIR/data \
  --train_batch_size 16 \
  --learning_rate 2e-5 \
  --num_train_epochs 3.0 \
  --max_seq_length 80 \
  --output_dir /tmp/swag_output/ \
  --gradient_accumulation_steps 4

Training with the previous hyper-parameters on a single GPU gave us the following results:

eval_accuracy = 0.8062081375587323
eval_loss = 0.5966546792367169
global_step = 13788
loss = 0.06423990014260186

LM Fine-tuning

The data should be a text file in the same format as sample_text.txt (one sentence per line, docs separated by empty line). You can download an exemplary training corpus generated from wikipedia articles and splitted into ~500k sentences with spaCy. Training one epoch on this corpus takes about 1:20h on 4 x NVIDIA Tesla P100 with train_batch_size=200 and max_seq_length=128:

python \
  --bert_model bert-base-cased \
  --do_train \
  --train_file samples/sample_text.txt \
  --output_dir models \
  --num_train_epochs 5.0 \
  --learning_rate 3e-5 \
  --train_batch_size 32 \
  --max_seq_length 128 

Fine-tuning BERT-large on GPUs

The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.

For example, fine-tuning BERT-large on SQuAD can be done on a server with 4 k-80 (these are pretty old now) in 18 hours. Our results are similar to the TensorFlow implementation results (actually slightly higher):

{"exact_match": 84.56953642384106, "f1": 91.04028647786927}

To get these results we used a combination of:

Here is the full list of hyper-parameters for this run:

python ./ \
  --bert_model bert-large-uncased \
  --do_train \
  --do_predict \
  --do_lower_case \
  --train_file $SQUAD_TRAIN \
  --predict_file $SQUAD_EVAL \
  --learning_rate 3e-5 \
  --num_train_epochs 2 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir $OUTPUT_DIR \
  --train_batch_size 24 \
  --gradient_accumulation_steps 2 

If you have a recent GPU (starting from NVIDIA Volta series), you should try 16-bit fine-tuning (FP16).

Here is an example of hyper-parameters for a FP16 run we tried:

python ./ \
  --bert_model bert-large-uncased \
  --do_train \
  --do_predict \
  --do_lower_case \
  --train_file $SQUAD_TRAIN \
  --predict_file $SQUAD_EVAL \
  --learning_rate 3e-5 \
  --num_train_epochs 2 \
  --max_seq_length 384 \
  --doc_stride 128 \
  --output_dir $OUTPUT_DIR \
  --train_batch_size 24 \
  --fp16 \
  --loss_scale 128

The results were similar to the above FP32 results (actually slightly higher):

{"exact_match": 84.65468306527909, "f1": 91.238669287002}


We include three Jupyter Notebooks that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.

Please follow the instructions given in the notebooks to run and modify them.

Command-line interface

A command-line interface is provided to convert a TensorFlow checkpoint in a PyTorch dump of the BertForPreTraining class (see above).

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

This CLI takes as input a TensorFlow checkpoint (three files starting with bert_model.ckpt) and the associated configuration file (bert_config.json), and creates a PyTorch model for this configuration, loads the weights from the TensorFlow checkpoint in the PyTorch model and saves the resulting model in a standard PyTorch save file that can be imported using torch.load() (see examples in, and

You only need to run this conversion script once to get a PyTorch 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 PyTorch model too.

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

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

pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch \
  $BERT_BASE_DIR/bert_model.ckpt \
  $BERT_BASE_DIR/bert_config.json \

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


TPU support and pretraining scripts

TPU are not supported by the current stable release of PyTorch (0.4.1). However, the next version of PyTorch (v1.0) should support training on TPU and is expected to be released soon (see the recent official announcement).

We will add TPU support when this next release is published.

The original TensorFlow code further comprises two scripts for pre-training BERT: and

Since, pre-training BERT is a particularly expensive operation that basically requires one or several TPUs to be completed in a reasonable amout of time (see details here) we have decided to wait for the inclusion of TPU support in PyTorch to convert these pre-training scripts.