The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
### Model2Model example
Encoder-decoder architectures require two tokenized inputs: one for the encoder and the other one for the decoder. Let's assume that we want to use `Model2Model` for generative question answering, and start by tokenizing the question and answer that will be fed to the model.
```python
importtorch
fromtransformersimportBertTokenizer,Model2Model
# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
# See the models docstrings for the detail of all the outputs
# In our case, the first element is the value of the LM loss
lm_loss=outputs[0]
```
This loss can be used to fine-tune `Model2Model` on the question answering task. Assuming that we fine-tuned the model, let us now see how to generate an answer:
```python
# Let's re-use the previous question
question="Who was Jim Henson?"
encoded_question=tokenizer.encode(question)
question_tensor=torch.tensor([encoded_question])
# This time we try to generate the answer, so we start with an empty sequence
| [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/RoBERTa/XLNet/XLM for question answering, examples with distributed training. |
| [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. |
| [Named Entity Recognition](https://github.com/huggingface/transformers/tree/master/examples/ner) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
| [Adversarial evaluation of model performances](#adversarial-evaluation-of-model-performances) | Testing a model with adversarial evaluation of natural language inference on the Heuristic Analysis for NLI Systems (HANS) dataset (McCoy et al., 2019.) |
@@ -10,14 +10,14 @@ This folder contains the original code used to train Distil* as well as examples
**October 23, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
**October 3, 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.**
**October 3, 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 supersedes 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.**
**September 19, 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 99% 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!
## What is Distil*
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 99% 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.
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.
We have applied the same method to other Transformer architectures and released the weights:
- 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 16.3 compared to 21.1 for **DistilGPT2** (after fine-tuning on the train set).
...
...
@@ -31,15 +31,15 @@ Here are the results on the dev sets of GLUE:
<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directly perform transfer learning on the pre-trained DistilRoBERTa.
<sup>2</sup> Macro-score computed without WNLI.
...
...
@@ -65,9 +65,9 @@ This part of the library has only be tested with Python3.6+. There are few speci
Transformers includes five pre-trained Distil* models, currently only provided for English and German (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-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 79.8 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 82.3 F1 score).
-`distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knowledge 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-cased`: 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-cased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 65M parameters.
-`distilbert-base-cased-distilled-squad`: A finetuned version of `distilbert-base-cased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 87.1 on the dev set (for comparison, Bert `bert-base-cased` version reaches a 88.7 F1 score).
-`distilbert-base-cased-distilled-squad`: A finetuned version of `distilbert-base-cased` finetuned using (a second step of) knowledge distillation on SQuAD 1.0. This model reaches a F1 score of 87.1 on the dev set (for comparison, Bert `bert-base-cased` version reaches a 88.7 F1 score).
-`distilbert-base-german-cased`: DistilBERT German language model pretrained on 1/2 of the data used to pretrain Bert using distillation with the supervision of the `bert-base-german-dbmdz-cased` version of German DBMDZ Bert. For NER tasks the model reaches a F1 score of 83.49 on the CoNLL-2003 test set (for comparison, `bert-base-german-dbmdz-cased` reaches a 84.52 F1 score), and a F1 score of 85.23 on the GermEval 2014 test set (`bert-base-german-dbmdz-cased` reaches a 86.89 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.
Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurences of each tokens in the data:
Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurrences of each tokens in the data:
