Skip to main content

Repository of pre-trained NLP Transformer models: BERT & RoBERTa, GPT & GPT-2, Transformer-XL, XLNet and XLM

Project description

👾 PyTorch-Transformers


PyTorch-Transformers (formerly known as pytorch-pretrained-bert) is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).

The library currently contains PyTorch implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:

  1. BERT (from Google) released with the paper BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
  2. GPT (from OpenAI) released with the paper Improving Language Understanding by Generative Pre-Training by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
  3. GPT-2 (from OpenAI) released with the paper Language Models are Unsupervised Multitask Learners by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
  4. Transformer-XL (from Google/CMU) released with the paper Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
  5. XLNet (from Google/CMU) released with the paper ​XLNet: Generalized Autoregressive Pretraining for Language Understanding by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
  6. XLM (from Facebook) released together with the paper Cross-lingual Language Model Pretraining by Guillaume Lample and Alexis Conneau.
  7. RoBERTa (from Facebook), released together with the paper a Robustly Optimized BERT Pretraining Approach by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
  8. DistilBERT (from HuggingFace), released together with the blogpost Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT by Victor Sanh, Lysandre Debut and Thomas Wolf.

These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the documentation.

Section Description
Installation How to install the package
Quick tour: Usage Tokenizers & models usage: Bert and GPT-2
Quick tour: Fine-tuning/usage scripts Using provided scripts: GLUE, SQuAD and Text generation
Migrating from pytorch-pretrained-bert to pytorch-transformers Migrating your code from pytorch-pretrained-bert to pytorch-transformers
Documentation Full API documentation and more


This repo is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.0.0+

With pip

PyTorch-Transformers can be installed by pip as follows:

pip install pytorch-transformers

From source

Clone the repository and run:

pip install [--editable] .


A series of tests is included for the library and the example scripts. Library tests can be found in the tests folder and examples tests in the examples folder.

These tests can be run using pytest (install pytest if needed with pip install pytest).

You can run the tests from the root of the cloned repository with the commands:

python -m pytest -sv ./pytorch_transformers/tests/
python -m pytest -sv ./examples/

Do you want to run a Transformer model on a mobile device?

You should check out our swift-coreml-transformers repo.

It contains an example of a conversion script from a Pytorch trained Transformer model (here, GPT-2) to a CoreML model that runs on iOS devices.

At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML, or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!

Quick tour

Let's do a very quick overview of PyTorch-Transformers. Detailed examples for each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the full documentation.

import torch
from pytorch_transformers import *

# PyTorch-Transformers has a unified API
# for 7 transformer architectures and 30 pretrained weights.
#          Model          | Tokenizer          | Pretrained weights shortcut
MODELS = [(BertModel,       BertTokenizer,      'bert-base-uncased'),
          (OpenAIGPTModel,  OpenAIGPTTokenizer, 'openai-gpt'),
          (GPT2Model,       GPT2Tokenizer,      'gpt2'),
          (TransfoXLModel,  TransfoXLTokenizer, 'transfo-xl-wt103'),
          (XLNetModel,      XLNetTokenizer,     'xlnet-base-cased'),
          (XLMModel,        XLMTokenizer,       'xlm-mlm-enfr-1024'),
          (RobertaModel,    RobertaTokenizer,   'roberta-base')]

# Let's encode some text in a sequence of hidden-states using each model:
for model_class, tokenizer_class, pretrained_weights in MODELS:
    # Load pretrained model/tokenizer
    tokenizer = tokenizer_class.from_pretrained(pretrained_weights)
    model = model_class.from_pretrained(pretrained_weights)

    # Encode text
    input_ids = torch.tensor([tokenizer.encode("Here is some text to encode", add_special_tokens=True)])  # Add special tokens takes care of adding [CLS], [SEP], <s>... tokens in the right way for each model.
    with torch.no_grad():
        last_hidden_states = model(input_ids)[0]  # Models outputs are now tuples

# Each architecture is provided with several class for fine-tuning on down-stream tasks, e.g.
BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNextSentencePrediction,
                      BertForSequenceClassification, BertForMultipleChoice, BertForTokenClassification,

# All the classes for an architecture can be initiated from pretrained weights for this architecture
# Note that additional weights added for fine-tuning are only initialized
# and need to be trained on the down-stream task
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
for model_class in BERT_MODEL_CLASSES:
    # Load pretrained model/tokenizer
    model = model_class.from_pretrained('bert-base-uncased')

# Models can return full list of hidden-states & attentions weights at each layer
model = model_class.from_pretrained(pretrained_weights,
input_ids = torch.tensor([tokenizer.encode("Let's see all hidden-states and attentions on this text")])
all_hidden_states, all_attentions = model(input_ids)[-2:]

# Models are compatible with Torchscript
model = model_class.from_pretrained(pretrained_weights, torchscript=True)
traced_model = torch.jit.trace(model, (input_ids,))

# Simple serialization for models and tokenizers
model.save_pretrained('./directory/to/save/')  # save
model = model_class.from_pretrained('./directory/to/save/')  # re-load
tokenizer.save_pretrained('./directory/to/save/')  # save
tokenizer = tokenizer_class.from_pretrained('./directory/to/save/')  # re-load

# SOTA examples for GLUE, SQUAD, text generation...

Quick tour of the fine-tuning/usage scripts

The library comprises several example scripts with SOTA performances for NLU and NLG tasks:

  • an example fine-tuning Bert, XLNet and XLM on nine different GLUE tasks (sequence-level classification)
  • an example fine-tuning Bert, XLNet and XLM on the question answering dataset SQuAD 2.0 (token-level classification)
  • an example using GPT, GPT-2, Transformer-XL and XLNet for conditional language generation
  • other model-specific examples (see the documentation).

Here are three quick usage examples for these scripts: Fine-tuning on GLUE tasks for sequence classification

The General Language Understanding Evaluation (GLUE) benchmark is a collection of nine sentence- or sentence-pair language understanding tasks for evaluating and analyzing natural language understanding systems.

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

You should also install the additional packages required by the examples:

pip install -r ./examples/requirements.txt
export GLUE_DIR=/path/to/glue

python ./examples/ \
    --model_type bert \
    --model_name_or_path bert-base-uncased \
    --task_name $TASK_NAME \
    --do_train \
    --do_eval \
    --do_lower_case \
    --data_dir $GLUE_DIR/$TASK_NAME \
    --max_seq_length 128 \
    --per_gpu_eval_batch_size=8   \
    --per_gpu_train_batch_size=8   \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir /tmp/$TASK_NAME/

where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.

The dev set results will be present within the text file 'eval_results.txt' in the specified output_dir. In case of MNLI, since there are two separate dev sets, matched and mismatched, there will be a separate output folder called '/tmp/MNLI-MM/' in addition to '/tmp/MNLI/'.

Fine-tuning XLNet model on the STS-B regression task

This example code fine-tunes XLNet on the STS-B corpus using parallel training on a server with 4 V100 GPUs. Parallel training is a simple way to use several GPUs (but is slower and less flexible than distributed training, see below).

export GLUE_DIR=/path/to/glue

python ./examples/ \
    --model_type xlnet \
    --model_name_or_path xlnet-large-cased \
    --do_train  \
    --do_eval   \
    --task_name=sts-b     \
    --data_dir=${GLUE_DIR}/STS-B  \
    --output_dir=./proc_data/sts-b-110   \
    --max_seq_length=128   \
    --per_gpu_eval_batch_size=8   \
    --per_gpu_train_batch_size=8   \
    --gradient_accumulation_steps=1 \
    --max_steps=1200  \
    --model_name=xlnet-large-cased   \
    --overwrite_output_dir   \
    --overwrite_cache \

On this machine we thus have a batch size of 32, please increase gradient_accumulation_steps to reach the same batch size if you have a smaller machine. These hyper-parameters should result in a Pearson correlation coefficient of +0.917 on the development set.

Fine-tuning Bert model on the MRPC classification task

This example code fine-tunes the Bert Whole Word Masking model on the Microsoft Research Paraphrase Corpus (MRPC) corpus using distributed training on 8 V100 GPUs to reach a F1 > 92.

python -m torch.distributed.launch --nproc_per_node 8 ./examples/   \
    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking \
    --task_name MRPC \
    --do_train   \
    --do_eval   \
    --do_lower_case   \
    --data_dir $GLUE_DIR/MRPC/   \
    --max_seq_length 128   \
    --per_gpu_eval_batch_size=8   \
    --per_gpu_train_batch_size=8   \
    --learning_rate 2e-5   \
    --num_train_epochs 3.0  \
    --output_dir /tmp/mrpc_output/ \
    --overwrite_output_dir   \
    --overwrite_cache \

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

  acc = 0.8823529411764706
  acc_and_f1 = 0.901702786377709
  eval_loss = 0.3418912578906332
  f1 = 0.9210526315789473
  global_step = 174
  loss = 0.07231863956341798 Fine-tuning on SQuAD for question-answering

This example code fine-tunes BERT on the SQuAD dataset using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:

python -m torch.distributed.launch --nproc_per_node=8 ./examples/ \
    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking \
    --do_train \
    --do_eval \
    --do_lower_case \
    --train_file $SQUAD_DIR/train-v1.1.json \
    --predict_file $SQUAD_DIR/dev-v1.1.json \
    --learning_rate 3e-5 \
    --num_train_epochs 2 \
    --max_seq_length 384 \
    --doc_stride 128 \
    --output_dir ../models/wwm_uncased_finetuned_squad/ \
    --per_gpu_eval_batch_size=3   \
    --per_gpu_train_batch_size=3   \

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

python $SQUAD_DIR/ $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
{"exact_match": 86.91579943235573, "f1": 93.1532499015869}

This is the model provided as bert-large-uncased-whole-word-masking-finetuned-squad. Text generation with GPT, GPT-2, Transformer-XL and XLNet

A conditional generation script is also included to generate text from a prompt. The generation script includes the tricks proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).

Here is how to run the script with the small version of OpenAI GPT-2 model:

python ./examples/ \
    --model_type=gpt2 \
    --length=20 \
    --model_name_or_path=gpt2 \

Migrating from pytorch-pretrained-bert to pytorch-transformers

Here is a quick summary of what you should take care of when migrating from pytorch-pretrained-bert to pytorch-transformers

Models always output tuples

The main breaking change when migrating from pytorch-pretrained-bert to pytorch-transformers is that the models forward method always outputs a tuple with various elements depending on the model and the configuration parameters.

The exact content of the tuples for each model are detailed in the models' docstrings and the documentation.

In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in pytorch-pretrained-bert.

Here is a pytorch-pretrained-bert to pytorch-transformers conversion example for a BertForSequenceClassification classification model:

# Let's load our model
model = BertForSequenceClassification.from_pretrained('bert-base-uncased')

# If you used to have this line in pytorch-pretrained-bert:
loss = model(input_ids, labels=labels)

# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
outputs = model(input_ids, labels=labels)
loss = outputs[0]

# In pytorch-transformers you can also have access to the logits:
loss, logits = outputs[:2]

# And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
model = BertForSequenceClassification.from_pretrained('bert-base-uncased', output_attentions=True)
outputs = model(input_ids, labels=labels)
loss, logits, attentions = outputs


Breaking change in the from_pretrained()method:

  1. Models are now set in evaluation mode by default when instantiated with the from_pretrained() method. To train them don't forget to set them back in training mode (model.train()) to activate the dropout modules.

  2. The additional *input and **kwargs arguments supplied to the from_pretrained() method used to be directly passed to the underlying model's class __init__() method. They are now used to update the model configuration attribute instead which can break derived model classes build based on the previous BertForSequenceClassification examples. We are working on a way to mitigate this breaking change in #866 by forwarding the the model __init__() method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.

Also, while not a breaking change, the serialization methods have been standardized and you probably should switch to the new method save_pretrained(save_directory) if you were using any other serialization method before.

Here is an example:

### Let's load a model and tokenizer
model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

### Do some stuff to our model and tokenizer
# Ex: add new tokens to the vocabulary and embeddings of our model
tokenizer.add_tokens(['[SPECIAL_TOKEN_1]', '[SPECIAL_TOKEN_2]'])
# Train our model

### Now let's save our model and tokenizer to a directory

### Reload the model and the tokenizer
model = BertForSequenceClassification.from_pretrained('./my_saved_model_directory/')
tokenizer = BertTokenizer.from_pretrained('./my_saved_model_directory/')

Optimizers: BertAdam & OpenAIAdam are now AdamW, schedules are standard PyTorch schedules

The two optimizers previously included, BertAdam and OpenAIAdam, have been replaced by a single AdamW optimizer which has a few differences:

  • it only implements weights decay correction,
  • schedules are now externals (see below),
  • gradient clipping is now also external (see below).

The new optimizer AdamW matches PyTorch Adam optimizer API and let you use standard PyTorch or apex methods for the schedule and clipping.

The schedules are now standard PyTorch learning rate schedulers and not part of the optimizer anymore.

Here is a conversion examples from BertAdam with a linear warmup and decay schedule to AdamW and the same schedule:

# Parameters:
lr = 1e-3
max_grad_norm = 1.0
num_total_steps = 1000
num_warmup_steps = 100
warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1

### Previously BertAdam optimizer was instantiated like this:
optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
### and used like this:
for batch in train_data:
    loss = model(batch)

### In PyTorch-Transformers, optimizer and schedules are splitted and instantiated like this:
optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
### and used like this:
for batch in train_data:
    loss = model(batch)
    torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)


At the moment, there is no paper associated to PyTorch-Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pytorch_transformers-1.2.0.tar.gz (152.3 kB view hashes)

Uploaded source

Built Distributions

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page