We include `three Jupyter Notebooks <https://github.com/huggingface/pytorch-transformers/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.
We include `three Jupyter Notebooks <https://github.com/huggingface/transformers/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 <https://github.com/huggingface/pytorch-transformers/blob/master/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 first NoteBook (\ `Comparing-TF-and-PT-models.ipynb <https://github.com/huggingface/transformers/blob/master/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 <https://github.com/huggingface/pytorch-transformers/blob/master/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 second NoteBook (\ `Comparing-TF-and-PT-models-SQuAD.ipynb <https://github.com/huggingface/transformers/blob/master/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 <https://github.com/huggingface/pytorch-transformers/blob/master/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.
The third NoteBook (\ `Comparing-TF-and-PT-models-MLM-NSP.ipynb <https://github.com/huggingface/transformers/blob/master/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.
Please follow the instructions given in the notebooks to run and modify them.
Please follow the instructions given in the notebooks to run and modify them.
PyTorch-Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
The library was designed with two strong goals in mind:
The library was designed with two strong goals in mind:
...
@@ -19,12 +19,12 @@ The library was designed with two strong goals in mind:
...
@@ -19,12 +19,12 @@ The library was designed with two strong goals in mind:
A few other goals:
A few other goals:
- expose the models internals as consistently as possible:
- expose the models' internals as consistently as possible:
- we give access, using a single API to the full hidden-states and attention weights,
- we give access, using a single API to the full hidden-states and attention weights,
- tokenizer and base model's API are standardized to easily switch between models.
- tokenizer and base model's API are standardized to easily switch between models.
- incorporate a subjective selection of promising tools for fine-tuning/investiguating these models:
- incorporate a subjective selection of promising tools for fine-tuning/investigating these models:
- a simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning,
- a simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning,
- simple ways to mask and prune transformer heads.
- simple ways to mask and prune transformer heads.
...
@@ -33,13 +33,13 @@ A few other goals:
...
@@ -33,13 +33,13 @@ A few other goals:
The library is build around three type of classes for each models:
The library is build around three type of classes for each models:
-**model classes** which are PyTorch models (`torch.nn.Modules`) of the 6 models architectures currently provided in the library, e.g. `BertModel`
-**model classes** which are PyTorch models (`torch.nn.Modules`) of the 8 models architectures currently provided in the library, e.g. `BertModel`
-**configuration classes** which store all the parameters required to build a model, e.g. `BertConfig`. You don't always need to instantiate these your-self, in particular if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model)
-**configuration classes** which store all the parameters required to build a model, e.g. `BertConfig`. You don't always need to instantiate these your-self, in particular if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model)
-**tokenizer classes** which store the vocabulary for each model and provide methods for encoding/decoding strings in list of token embeddings indices to be fed to a model, e.g. `BertTokenizer`
-**tokenizer classes** which store the vocabulary for each model and provide methods for encoding/decoding strings in list of token embeddings indices to be fed to a model, e.g. `BertTokenizer`
All these classes can be instantiated from pretrained instances and saved locally using two methods:
All these classes can be instantiated from pretrained instances and saved locally using two methods:
-`from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/pytorch-transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
-`from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
-`save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using `from_pretrained()`.
-`save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using `from_pretrained()`.
We'll finish this quickstart tour by going through a few simple quick-start examples to see how we can instantiate and use these classes. The rest of the documentation is organized in two parts:
We'll finish this quickstart tour by going through a few simple quick-start examples to see how we can instantiate and use these classes. The rest of the documentation is organized in two parts:
...
@@ -51,7 +51,7 @@ We'll finish this quickstart tour by going through a few simple quick-start exam
...
@@ -51,7 +51,7 @@ We'll finish this quickstart tour by going through a few simple quick-start exam
Here are two examples showcasing a few `Bert` and `GPT2` classes and pre-trained models.
Here are two examples showcasing a few `Bert` and `GPT2` classes and pre-trained models.
See full API reference for examples for each model classe.
See full API reference for examples for each model class.
### BERT example
### BERT example
...
@@ -59,7 +59,7 @@ Let's start by preparing a tokenized input (a list of token embeddings indices t
...
@@ -59,7 +59,7 @@ Let's start by preparing a tokenized input (a list of token embeddings indices t
| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks.
| [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
| [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
| [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet. |
| [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet. |
| [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision. |
| [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision. |
| [SQuAD](#squad) | Using BERT for question answering, examples with distributed training. |
| [SQuAD](#squad) | Using BERT/XLM/XLNet/RoBERTa for question answering, examples with distributed training. |
| [Multiple Choice](#multiple choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
| [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
## TensorFlow 2.0 Bert models on GLUE
Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
These options and the below benchmark are provided by @tlkh.
Quick benchmarks from the script (no other modifications):
| GPU | Mode | Time (2nd epoch) | Val Acc (3 runs) |
Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
Details and results for the fine-tuning provided by @stefan-it.
### Data (Download and pre-processing steps)
Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
### Evaluation
Evaluation on development dataset outputs the following for our example:
This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.
This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
**2019, September 19th - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
**October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
## What is DistilBERT
**October 3rd, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
**September 19th, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
For more information on DistilBERT, please refer to our [detailed blog post](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-distilbert-a-distilled-version-of-bert-8cf3380435b5
). *Please note that we will publish a formal write-up with updated and more complete results in the near future (September 19th).*
Here's the updated results on the dev sets of GLUE:
Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base` performance on GLUE while being twice faster and 35% smaller.
- and more to come! 🤗🤗🤗
For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
<sup>2</sup> Macro-score computed without WNLI.
<sup>3</sup> We compute this score ourselves for completeness.
## Setup
## Setup
This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0). It is important to note that there is a small internal bug in the current version of PyTorch available on pip that causes a memory leak in our training/distillation. It has been recently fixed and will likely be integrated into the next release. For the moment, we recommend to [compile PyTorch from source](https://github.com/pytorch/pytorch#from-source). Please refer to [issue 1179](https://github.com/huggingface/pytorch-transformers/issues/1179) for more details.
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
## How to use DistilBERT
## How to use DistilBERT
PyTorch-Transformers includes two pre-trained DistilBERT models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
Transformers includes two pre-trained Distil* models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
-`distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
-`distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
-`distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
-`distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
-`distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
-`distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
- and more to come! 🤗🤗🤗
Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
...
@@ -42,9 +64,14 @@ outputs = model(input_ids)
...
@@ -42,9 +64,14 @@ outputs = model(input_ids)
last_hidden_states=outputs[0]# The last hidden-state is the first element of the output tuple
last_hidden_states=outputs[0]# The last hidden-state is the first element of the output tuple
```
```
## How to train DistilBERT
Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
We highly encourage you to use distributed training for training DistilBert as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
We highly encourage you to use distributed training for training DistilBERT as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract_for_distil.py` to create a valid initialization checkpoint and use `--from_pretrained_weights`and `--from_pretrained_config`arguments to use this initialization for the distilled training!
**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
Happy distillation!
Happy distillation!
## Citation
If you find the ressource useful, you should cite the following paper:
```
@inproceedings{sanh2019distilbert,
title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
parser=argparse.ArgumentParser(description="Extraction some layers of the full RobertaForMaskedLM or GPT2LMHeadModel for Transfer Learned Distillation")
