This repository contains an op-for-op PyTorch reimplementation of [Google's TensorFlow repository for the BERT model](https://github.com/google-research/bert) that was released together with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
This implementation is provided with [Google's pre-trained models](https://github.com/google-research/bert), examples, notebooks and a command-line interface to load any pre-trained TensorFlow checkpoint for BERT is also provided.
A series of tests is included in the [tests folder](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/tests) and can be run using `pytest` (install pytest if needed: `pip install pytest`).
- Seven PyTorch models (`torch.nn.Module`) for Bert with pre-trained weights (in the [`modeling.py`](./pytorch_pretrained_bert/modeling.py) file):
- [`BertModel`](./pytorch_pretrained_bert/modeling.py#L537) - raw BERT Transformer model (**fully pre-trained**),
- [`BertForMaskedLM`](./pytorch_pretrained_bert/modeling.py#L691) - BERT Transformer with the pre-trained masked language modeling head on top (**fully pre-trained**),
- [`BertForNextSentencePrediction`](./pytorch_pretrained_bert/modeling.py#L752) - BERT Transformer with the pre-trained next sentence prediction classifier on top (**fully pre-trained**),
- [`BertForPreTraining`](./pytorch_pretrained_bert/modeling.py#L620) - BERT Transformer with masked language modeling head and next sentence prediction classifier on top (**fully pre-trained**),
- [`BertForSequenceClassification`](./pytorch_pretrained_bert/modeling.py#L814) - BERT Transformer with a sequence classification head on top (BERT Transformer is **pre-trained**, the sequence classification head **is only initialized and has to be trained**),
- [`BertForMultipleChoice`](./pytorch_pretrained_bert/modeling.py#L880) - BERT Transformer with a multiple choice head on top (used for task like Swag) (BERT Transformer is **pre-trained**, the sequence classification head **is only initialized and has to be trained**),
- [`BertForTokenClassification`](./pytorch_pretrained_bert/modeling.py#L949) - BERT Transformer with a token classification head on top (BERT Transformer is **pre-trained**, the token classification head **is only initialized and has to be trained**),
- [`BertForQuestionAnswering`](./pytorch_pretrained_bert/modeling.py#L1015) - BERT Transformer with a token classification head on top (BERT Transformer is **pre-trained**, the token classification head **is only initialized and has to be trained**).
- Three notebooks that were used to check that the TensorFlow and PyTorch models behave identically (in the [`notebooks` folder](./notebooks)):
- [`Comparing-TF-and-PT-models.ipynb`](./notebooks/Comparing-TF-and-PT-models.ipynb) - Compare the hidden states predicted by `BertModel`,
- [`Comparing-TF-and-PT-models-SQuAD.ipynb`](./notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb) - Compare the spans predicted by `BertForQuestionAnswering` instances,
- [`Comparing-TF-and-PT-models-MLM-NSP.ipynb`](./notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb) - Compare the predictions of the `BertForPretraining` instances.
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](#doc) below for all the details on these classes.
| [Loading Google AI's pre-trained weigths](#Loading-Google-AIs-pre-trained-weigths-and-PyTorch-dump) | How to load Google AI's pre-trained weight or a PyTorch saved instance |
| [PyTorch models](#PyTorch-models) | API of the seven PyTorch model classes: `BertModel`, `BertForMaskedLM`, `BertForNextSentencePrediction`, `BertForPreTraining`, `BertForSequenceClassification` or `BertForQuestionAnswering` |
To load one of Google AI's pre-trained models or a PyTorch saved model (an instance of `BertForPreTraining` saved with `torch.save()`), the PyTorch model classes and the tokenizer can be instantiated as
-`BERT_CLASS` is either the `BertTokenizer` class (to load the vocabulary) or one of the seven PyTorch model classes (to load the pre-trained weights): `BertModel`, `BertForMaskedLM`, `BertForNextSentencePrediction`, `BertForPreTraining`, `BertForSequenceClassification`, `BertForTokenClassification` or `BertForQuestionAnswering`, and
If `PRE_TRAINED_MODEL_NAME_OR_PATH` is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links [here](pytorch_pretrained_bert/modeling.py)) and stored in a cache folder to avoid future download (the cache folder can be found at `~/.pytorch_pretrained_bert/`).
-`cache_dir` can be an optional path to a specific directory to download and cache the pre-trained model weights. This option is useful in particular when you are using distributed training: to avoid concurrent access to the same weights you can set for example `cache_dir='./pretrained_model_{}'.format(args.local_rank)` (see the section on distributed training for more information)
`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](https://github.com/google-research/bert/blob/master/multilingual.md) or the original TensorFlow repository.
**When using an `uncased model`, make sure to pass `--do_lower_case` to the training scripts. (Or pass `do_lower_case=True` directly to FullTokenizer if you're using your own script.)**
`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**.
-`input_ids`: a torch.LongTensor of shape [batch_size, sequence_length] with the word token indices in the vocabulary (see the tokens preprocessing logic in the scripts [`extract_features.py`](./examples/extract_features.py), [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`](./examples/run_squad.py)), and
-`token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to a `sentence B` token (see BERT paper for more details).
-`attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [0, 1]. It's a mask to be used if some input sequence lengths are smaller than the max input sequence length of the current batch. It's the mask that we typically use for attention when a batch has varying length sentences.
-`output_all_encoded_layers=True`: outputs a list of the encoded-hidden-states at the end of each attention block (i.e. 12 full sequences for BERT-base, 24 for BERT-large), each encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
-`output_all_encoded_layers=False`: outputs only the encoded-hidden-states corresponding to the last attention block, i.e. a single torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
-`pooled_output`: a torch.FloatTensor of size [batch_size, hidden_size] which is the output of a classifier pretrained on top of the hidden state associated to the first character of the input (`CLF`) to train on the Next-Sentence task (see BERT's paper).
An example on how to use this class is given in the [`extract_features.py`](./examples/extract_features.py) script which can be used to extract the hidden states of the model for a given input.
-`masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss is only computed for the labels set in [0, ..., vocab_size]
-`next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size] with indices selected in [0, 1]. 0 => next sentence is the continuation, 1 => next sentence is a random sentence.
*Outputs*:
- if `masked_lm_labels` and `next_sentence_label` are not `None`: Outputs the total_loss which is the sum of the masked language modeling loss and the next sentence classification loss.
- if `masked_lm_labels` or `next_sentence_label` is `None`: Outputs a tuple comprising
-`masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss is only computed for the labels set in [0, ..., vocab_size]
*Outputs*:
- if `masked_lm_labels` is not `None`: Outputs the masked language modeling loss.
- if `masked_lm_labels` is `None`: Outputs the masked language modeling logits.
-`next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size] with indices selected in [0, 1]. 0 => next sentence is the continuation, 1 => next sentence is a random sentence.
*Outputs*:
- if `next_sentence_label` is not `None`: Outputs the next sentence classification loss.
- if `next_sentence_label` is `None`: Outputs the next sentence classification logits.
`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 [`run_classifier.py`](./examples/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.
`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 output 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](https://blog.openai.com/language-unsupervised/) and the answer of Jacob Devlin in the following [issue](https://github.com/google-research/bert/issues/38).
An example on how to use this class is given in the [`run_swag.py`](./examples/run_swag.py) script which can be used to fine-tune a multiple choice classifier using BERT, for example for the Swag task.
`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 [`run_squad.py`](./examples/run_squad.py) script which can be used to fine-tune a token classifier using BERT, for example for the SQuAD task.
Please refer to the doc strings and code in [`tokenization.py`](./pytorch_pretrained_bert/tokenization.py) for the details of the `BasicTokenizer` and `WordpieceTokenizer` classes. In general it is recommended to use `BertTokenizer` unless you know what you are doing.
`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:
| [Training large models: introduction, tools and examples](#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](#Fine-tuning-with-BERT-running-the-examples) | Running the examples in [`./examples`](./examples/): `extract_classif.py`, `run_classifier.py` and `run_squad.py` |
| [Fine-tuning BERT-large on GPUs](#Fine-tuning-BERT-large-on-GPUs) | How to fine tune `BERT large`|
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 five techniques that you can activate in the fine-tuning scripts [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`](./examples/run_squad.py): gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training . For more details on how to use these techniques you can read [the tips on training large batches in PyTorch](https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255) that I published earlier this month.
- **Gradient Accumulation**: Gradient accumulation can be used by supplying a integer greater than 1 to the `--gradient_accumulation_steps` argument. The batch at each step will be divided by this integer and gradient will be accumulated over `gradient_accumulation_steps` steps.
- **Multi-GPU**: Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs.
- **Distributed training**: Distributed training can be activated by supplying an integer greater or equal to 0 to the `--local_rank` argument (see below).
- **Optimize on CPU**: The Adam optimizer stores 2 moving average of the weights of the model. If you keep them on GPU 1 (typical behavior), your first GPU will have to store 3-times the size of the model. This is not optimal for large models like `BERT-large` and means your batch size is a lot lower than it could be. This option will perform the optimization and store the averages on the CPU/RAM to free more room on the GPU(s). As the most computational intensive operation is usually the backward pass, this doesn't have a significant impact on the training time. Activate this option with `--optimize_on_cpu` on the [`run_squad.py`](./examples/run_squad.py) script.
- **16-bits training**: 16-bits training, also called mixed-precision training, can reduce the memory requirement of your model on the GPU by using half-precision training, basically allowing to double the batch size. If you have a recent GPU (starting from NVIDIA Volta architecture) you should see no decrease in speed. A good introduction to Mixed precision training can be found [here](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) and a full documentation is [here](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html). In our scripts, this option can be activated by setting the `--fp16` flag and you can play with loss scaling using the `--loss_scaling` flag (see the previously linked documentation for details on loss scaling). If the loss scaling is too high (`Nan` in the gradients) it will be automatically scaled down until the value is acceptable. The default loss scaling is 128 which behaved nicely in our tests.
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]((https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255)) for more details):
python -m torch.distributed.launch --nproc_per_node=4 --nnodes=2 --node_rank=$THIS_MACHINE_INDEX --master_addr="192.168.1.1" --master_port=1234 run_classifier.py (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)
Where `$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 `192.168.1.1` and an open port `1234`.
We showcase the same examples as [the original implementation](https://github.com/google-research/bert/): fine-tuning a sequence-level classifier on the MRPC classification corpus and a token-level classifier on the question answering dataset SQuAD.
Our test ran on a few seeds with [the original implementation hyper-parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation results between 84% and 88%.
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):
We include [three Jupyter Notebooks](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/notebooks) that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
- The first NoteBook ([Comparing-TF-and-PT-models.ipynb](./notebooks/Comparing-TF-and-PT-models.ipynb)) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
- The second NoteBook ([Comparing-TF-and-PT-models-SQuAD.ipynb](./notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb)) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the `BertForQuestionAnswering` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
- The third NoteBook ([Comparing-TF-and-PT-models-MLM-NSP.ipynb](./notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb)) compares the predictions computed by the TensorFlow and the PyTorch models for masked token language modeling using the pre-trained masked language modeling model.
You can convert any TensorFlow checkpoint for BERT (in particular [the pre-trained models released by Google](https://github.com/google-research/bert#pre-trained-models)) in a PyTorch save file by using the [`./pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py`](convert_tf_checkpoint_to_pytorch.py) 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 [`extract_features.py`](./examples/extract_features.py), [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`]((./examples/run_squad.py))).
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:
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](https://cloud.google.com/blog/products/ai-machine-learning/introducing-pytorch-across-google-cloud)).
We will add TPU support when this next release is published.
The original TensorFlow code further comprises two scripts for pre-training BERT: [create_pretraining_data.py](https://github.com/google-research/bert/blob/master/create_pretraining_data.py) and [run_pretraining.py](https://github.com/google-research/bert/blob/master/run_pretraining.py).
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](https://github.com/google-research/bert#pre-training-with-bert)) we have decided to wait for the inclusion of TPU support in PyTorch to convert these pre-training scripts.
