Skip to content

GitLab

Unverified Commit eb849f66 authored by Sylvain Gugger's avatar Sylvain Gugger Committed by GitHub
Browse files

Migrate doc files to Markdown. (#24376) 



* Rename index.mdx to index.md

* With saved modifs

* Address review comment

* Treat all files

* .mdx -> .md

* Remove special char

* Update utils/tests_fetcher.py
Co-authored-by: default avatarLysandre Debut <lysandre.debut@reseau.eseo.fr>

---------
Co-authored-by: default avatarLysandre Debut <lysandre.debut@reseau.eseo.fr>
parent b0513b01
Hide whitespace changes
Inline Side-by-side
Showing with 3481 additions and 3442 deletions
docs/source/en/community.md 0 → 100644
View file @ eb849f66
<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Community
This page regroups resources around 🤗 Transformers developed by the community.
## Community resources:
| Resource | Description | Author |
|:----------|:-------------|------:|
| [Hugging Face Transformers Glossary Flashcards](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | A set of flashcards based on the [Transformers Docs Glossary](glossary) that has been put into a form which can be easily learnt/revised using [Anki ](https://apps.ankiweb.net/) an open source, cross platform app specifically designed for long term knowledge retention. See this [Introductory video on how to use the flashcards](https://www.youtube.com/watch?v=Dji_h7PILrw). | [Darigov Research](https://www.darigovresearch.com/) |
## Community notebooks:
| Notebook | Description | Author | |
|:----------|:-------------|:-------------|------:|
| [Fine-tune a pre-trained Transformer to generate lyrics](https://github.com/AlekseyKorshuk/huggingartists) | How to generate lyrics in the style of your favorite artist by fine-tuning a GPT-2 model | [Aleksey Korshuk](https://github.com/AlekseyKorshuk) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb) |
| [Train T5 in Tensorflow 2 ](https://github.com/snapthat/TF-T5-text-to-text) | How to train T5 for any task using Tensorflow 2. This notebook demonstrates a Question & Answer task implemented in Tensorflow 2 using SQUAD | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
| [Train T5 on TPU](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb) | How to train T5 on SQUAD with Transformers and Nlp | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
| [Fine-tune T5 for Classification and Multiple Choice](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) | How to fine-tune T5 for classification and multiple choice tasks using a text-to-text format with PyTorch Lightning | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) |
| [Fine-tune DialoGPT on New Datasets and Languages](https://github.com/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) | How to fine-tune the DialoGPT model on a new dataset for open-dialog conversational chatbots | [Nathan Cooper](https://github.com/ncoop57) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) |
| [Long Sequence Modeling with Reformer](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb) | How to train on sequences as long as 500,000 tokens with Reformer | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb) |
| [Fine-tune BART for Summarization](https://github.com/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) | How to fine-tune BART for summarization with fastai using blurr | [Wayde Gilliam](https://ohmeow.com/) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) |
| [Fine-tune a pre-trained Transformer on anyone's tweets](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) | How to generate tweets in the style of your favorite Twitter account by fine-tuning a GPT-2 model | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) |
| [Optimize 🤗 Hugging Face models with Weights & Biases](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) | A complete tutorial showcasing W&B integration with Hugging Face | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) |
| [Pretrain Longformer](https://github.com/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) | How to build a "long" version of existing pretrained models | [Iz Beltagy](https://beltagy.net) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) |
| [Fine-tune Longformer for QA](https://github.com/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) | How to fine-tune longformer model for QA task | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) |
| [Evaluate Model with 🤗nlp](https://github.com/patrickvonplaten/notebooks/blob/master/How_to_evaluate_Longformer_on_TriviaQA_using_NLP.ipynb) | How to evaluate longformer on TriviaQA with `nlp` | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing) |
| [Fine-tune T5 for Sentiment Span Extraction](https://github.com/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) | How to fine-tune T5 for sentiment span extraction using a text-to-text format with PyTorch Lightning | [Lorenzo Ampil](https://github.com/enzoampil) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) |
| [Fine-tune DistilBert for Multiclass Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb) | How to fine-tune DistilBert for multiclass classification with PyTorch | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb)|
|[Fine-tune BERT for Multi-label Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|How to fine-tune BERT for multi-label classification using PyTorch|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|
|[Fine-tune T5 for Summarization](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|How to fine-tune T5 for summarization in PyTorch and track experiments with WandB|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|
|[Speed up Fine-Tuning in Transformers with Dynamic Padding / Bucketing](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)|How to speed up fine-tuning by a factor of 2 using dynamic padding / bucketing|[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|
|[Pretrain Reformer for Masked Language Modeling](https://github.com/patrickvonplaten/notebooks/blob/master/Reformer_For_Masked_LM.ipynb)| How to train a Reformer model with bi-directional self-attention layers | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tzzh0i8PgDQGV3SMFUGxM7_gGae3K-uW?usp=sharing)|
|[Expand and Fine Tune Sci-BERT](https://github.com/lordtt13/word-embeddings/blob/master/COVID-19%20Research%20Data/COVID-SciBERT.ipynb)| How to increase vocabulary of a pretrained SciBERT model from AllenAI on the CORD dataset and pipeline it. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1rqAR40goxbAfez1xvF3hBJphSCsvXmh8)|
|[Fine Tune BlenderBotSmall for Summarization using the Trainer API](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/fine-tune-blenderbot_small-for-summarization.ipynb)| How to fine tune BlenderBotSmall for summarization on a custom dataset, using the Trainer API. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/19Wmupuls7mykSGyRN_Qo6lPQhgp56ymq?usp=sharing)|
|[Fine-tune Electra and interpret with Integrated Gradients](https://github.com/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb) | How to fine-tune Electra for sentiment analysis and interpret predictions with Captum Integrated Gradients | [Eliza Szczechla](https://elsanns.github.io) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb)|
|[fine-tune a non-English GPT-2 Model with Trainer class](https://github.com/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb) | How to fine-tune a non-English GPT-2 Model with Trainer class | [Philipp Schmid](https://www.philschmid.de) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb)|
|[Fine-tune a DistilBERT Model for Multi Label Classification task](https://github.com/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb) | How to fine-tune a DistilBERT Model for Multi Label Classification task | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb)|
|[Fine-tune ALBERT for sentence-pair classification](https://github.com/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb) | How to fine-tune an ALBERT model or another BERT-based model for the sentence-pair classification task | [Nadir El Manouzi](https://github.com/NadirEM) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb)|
|[Fine-tune Roberta for sentiment analysis](https://github.com/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb) | How to fine-tune a Roberta model for sentiment analysis | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)|
|[Evaluating Question Generation Models](https://github.com/flexudy-pipe/qugeev) | How accurate are the answers to questions generated by your seq2seq transformer model? | [Pascal Zoleko](https://github.com/zolekode) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1bpsSqCQU-iw_5nNoRm_crPq6FRuJthq_?usp=sharing)|
|[Classify text with DistilBERT and Tensorflow](https://github.com/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb) | How to fine-tune DistilBERT for text classification in TensorFlow | [Peter Bayerle](https://github.com/peterbayerle) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb)|
|[Leverage BERT for Encoder-Decoder Summarization on CNN/Dailymail](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb) | How to warm-start a *EncoderDecoderModel* with a *bert-base-uncased* checkpoint for summarization on CNN/Dailymail | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb)|
|[Leverage RoBERTa for Encoder-Decoder Summarization on BBC XSum](https://github.com/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb) | How to warm-start a shared *EncoderDecoderModel* with a *roberta-base* checkpoint for summarization on BBC/XSum | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb)|
|[Fine-tune TAPAS on Sequential Question Answering (SQA)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb) | How to fine-tune *TapasForQuestionAnswering* with a *tapas-base* checkpoint on the Sequential Question Answering (SQA) dataset | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb)|
|[Evaluate TAPAS on Table Fact Checking (TabFact)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb) | How to evaluate a fine-tuned *TapasForSequenceClassification* with a *tapas-base-finetuned-tabfact* checkpoint using a combination of the 🤗 datasets and 🤗 transformers libraries | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb)|
|[Fine-tuning mBART for translation](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb) | How to fine-tune mBART using Seq2SeqTrainer for Hindi to English translation | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb)|
|[Fine-tune LayoutLM on FUNSD (a form understanding dataset)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb) | How to fine-tune *LayoutLMForTokenClassification* on the FUNSD dataset for information extraction from scanned documents | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb)|
|[Fine-Tune DistilGPT2 and Generate Text](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb) | How to fine-tune DistilGPT2 and generate text | [Aakash Tripathi](https://github.com/tripathiaakash) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb)|
|[Fine-Tune LED on up to 8K tokens](https://github.com/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb) | How to fine-tune LED on pubmed for long-range summarization | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb)|
|[Evaluate LED on Arxiv](https://github.com/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb) | How to effectively evaluate LED on long-range summarization | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb)|
|[Fine-tune LayoutLM on RVL-CDIP (a document image classification dataset)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb) | How to fine-tune *LayoutLMForSequenceClassification* on the RVL-CDIP dataset for scanned document classification | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb)|
|[Wav2Vec2 CTC decoding with GPT2 adjustment](https://github.com/voidful/huggingface_notebook/blob/main/xlsr_gpt.ipynb) | How to decode CTC sequence with language model adjustment | [Eric Lam](https://github.com/voidful) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1e_z5jQHYbO2YKEaUgzb1ww1WwiAyydAj?usp=sharing)|
|[Fine-tune BART for summarization in two languages with Trainer class](https://github.com/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb) | How to fine-tune BART for summarization in two languages with Trainer class | [Eliza Szczechla](https://github.com/elsanns) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb)|
|[Evaluate Big Bird on Trivia QA](https://github.com/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb) | How to evaluate BigBird on long document question answering on Trivia QA | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb)|
| [Create video captions using Wav2Vec2](https://github.com/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) | How to create YouTube captions from any video by transcribing the audio with Wav2Vec | [Niklas Muennighoff](https://github.com/Muennighoff) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) |
| [Fine-tune the Vision Transformer on CIFAR-10 using PyTorch Lightning](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and PyTorch Lightning | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) |
| [Fine-tune the Vision Transformer on CIFAR-10 using the 🤗 Trainer](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and the 🤗 Trainer | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) |
| [Evaluate LUKE on Open Entity, an entity typing dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) | How to evaluate *LukeForEntityClassification* on the Open Entity dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) |
| [Evaluate LUKE on TACRED, a relation extraction dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) | How to evaluate *LukeForEntityPairClassification* on the TACRED dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) |
| [Evaluate LUKE on CoNLL-2003, an important NER benchmark](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) | How to evaluate *LukeForEntitySpanClassification* on the CoNLL-2003 dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) |
| [Evaluate BigBird-Pegasus on PubMed dataset](https://github.com/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) | How to evaluate *BigBirdPegasusForConditionalGeneration* on PubMed dataset | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) |
| [Speech Emotion Classification with Wav2Vec2](https://github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) | How to leverage a pretrained Wav2Vec2 model for Emotion Classification on the MEGA dataset | [Mehrdad Farahani](https://github.com/m3hrdadfi) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) |
| [Detect objects in an image with DETR](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) | How to use a trained *DetrForObjectDetection* model to detect objects in an image and visualize attention | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) |
| [Fine-tune DETR on a custom object detection dataset](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) | How to fine-tune *DetrForObjectDetection* on a custom object detection dataset | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) |
| [Finetune T5 for Named Entity Recognition](https://github.com/ToluClassics/Notebooks/blob/main/T5_Ner_Finetuning.ipynb) | How to fine-tune *T5* on a Named Entity Recognition Task | [Ogundepo Odunayo](https://github.com/ToluClassics) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1obr78FY_cBmWY5ODViCmzdY6O1KB65Vc?usp=sharing) |
docs/source/en/community.mdx deleted 100644 → 0
View file @ b0513b01
# Community
This page regroups resources around 🤗 Transformers developed by the community.
## Community resources:
| Resource | Description | Author |
|:----------|:-------------|------:|
| [Hugging Face Transformers Glossary Flashcards](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | A set of flashcards based on the [Transformers Docs Glossary](glossary) that has been put into a form which can be easily learnt/revised using [Anki ](https://apps.ankiweb.net/) an open source, cross platform app specifically designed for long term knowledge retention. See this [Introductory video on how to use the flashcards](https://www.youtube.com/watch?v=Dji_h7PILrw). | [Darigov Research](https://www.darigovresearch.com/) |
## Community notebooks:
| Notebook | Description | Author | |
|:----------|:-------------|:-------------|------:|
| [Fine-tune a pre-trained Transformer to generate lyrics](https://github.com/AlekseyKorshuk/huggingartists) | How to generate lyrics in the style of your favorite artist by fine-tuning a GPT-2 model | [Aleksey Korshuk](https://github.com/AlekseyKorshuk) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb) |
| [Train T5 in Tensorflow 2 ](https://github.com/snapthat/TF-T5-text-to-text) | How to train T5 for any task using Tensorflow 2. This notebook demonstrates a Question & Answer task implemented in Tensorflow 2 using SQUAD | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
| [Train T5 on TPU](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb) | How to train T5 on SQUAD with Transformers and Nlp | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
| [Fine-tune T5 for Classification and Multiple Choice](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) | How to fine-tune T5 for classification and multiple choice tasks using a text-to-text format with PyTorch Lightning | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) |
| [Fine-tune DialoGPT on New Datasets and Languages](https://github.com/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) | How to fine-tune the DialoGPT model on a new dataset for open-dialog conversational chatbots | [Nathan Cooper](https://github.com/ncoop57) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) |
| [Long Sequence Modeling with Reformer](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb) | How to train on sequences as long as 500,000 tokens with Reformer | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb) |
| [Fine-tune BART for Summarization](https://github.com/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) | How to fine-tune BART for summarization with fastai using blurr | [Wayde Gilliam](https://ohmeow.com/) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) |
| [Fine-tune a pre-trained Transformer on anyone's tweets](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) | How to generate tweets in the style of your favorite Twitter account by fine-tuning a GPT-2 model | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) |
| [Optimize 🤗 Hugging Face models with Weights & Biases](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) | A complete tutorial showcasing W&B integration with Hugging Face | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) |
| [Pretrain Longformer](https://github.com/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) | How to build a "long" version of existing pretrained models | [Iz Beltagy](https://beltagy.net) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) |
| [Fine-tune Longformer for QA](https://github.com/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) | How to fine-tune longformer model for QA task | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) |
| [Evaluate Model with 🤗nlp](https://github.com/patrickvonplaten/notebooks/blob/master/How_to_evaluate_Longformer_on_TriviaQA_using_NLP.ipynb) | How to evaluate longformer on TriviaQA with `nlp` | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing) |
| [Fine-tune T5 for Sentiment Span Extraction](https://github.com/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) | How to fine-tune T5 for sentiment span extraction using a text-to-text format with PyTorch Lightning | [Lorenzo Ampil](https://github.com/enzoampil) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) |
| [Fine-tune DistilBert for Multiclass Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb) | How to fine-tune DistilBert for multiclass classification with PyTorch | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb)|
|[Fine-tune BERT for Multi-label Classification](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|How to fine-tune BERT for multi-label classification using PyTorch|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|
|[Fine-tune T5 for Summarization](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|How to fine-tune T5 for summarization in PyTorch and track experiments with WandB|[Abhishek Kumar Mishra](https://github.com/abhimishra91) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|
|[Speed up Fine-Tuning in Transformers with Dynamic Padding / Bucketing](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)|How to speed up fine-tuning by a factor of 2 using dynamic padding / bucketing|[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|
|[Pretrain Reformer for Masked Language Modeling](https://github.com/patrickvonplaten/notebooks/blob/master/Reformer_For_Masked_LM.ipynb)| How to train a Reformer model with bi-directional self-attention layers | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tzzh0i8PgDQGV3SMFUGxM7_gGae3K-uW?usp=sharing)|
|[Expand and Fine Tune Sci-BERT](https://github.com/lordtt13/word-embeddings/blob/master/COVID-19%20Research%20Data/COVID-SciBERT.ipynb)| How to increase vocabulary of a pretrained SciBERT model from AllenAI on the CORD dataset and pipeline it. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1rqAR40goxbAfez1xvF3hBJphSCsvXmh8)|
|[Fine Tune BlenderBotSmall for Summarization using the Trainer API](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/fine-tune-blenderbot_small-for-summarization.ipynb)| How to fine tune BlenderBotSmall for summarization on a custom dataset, using the Trainer API. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/19Wmupuls7mykSGyRN_Qo6lPQhgp56ymq?usp=sharing)|
|[Fine-tune Electra and interpret with Integrated Gradients](https://github.com/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb) | How to fine-tune Electra for sentiment analysis and interpret predictions with Captum Integrated Gradients | [Eliza Szczechla](https://elsanns.github.io) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb)|
|[fine-tune a non-English GPT-2 Model with Trainer class](https://github.com/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb) | How to fine-tune a non-English GPT-2 Model with Trainer class | [Philipp Schmid](https://www.philschmid.de) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb)|
|[Fine-tune a DistilBERT Model for Multi Label Classification task](https://github.com/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb) | How to fine-tune a DistilBERT Model for Multi Label Classification task | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb)|
|[Fine-tune ALBERT for sentence-pair classification](https://github.com/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb) | How to fine-tune an ALBERT model or another BERT-based model for the sentence-pair classification task | [Nadir El Manouzi](https://github.com/NadirEM) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb)|
|[Fine-tune Roberta for sentiment analysis](https://github.com/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb) | How to fine-tune a Roberta model for sentiment analysis | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)|
|[Evaluating Question Generation Models](https://github.com/flexudy-pipe/qugeev) | How accurate are the answers to questions generated by your seq2seq transformer model? | [Pascal Zoleko](https://github.com/zolekode) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1bpsSqCQU-iw_5nNoRm_crPq6FRuJthq_?usp=sharing)|
|[Classify text with DistilBERT and Tensorflow](https://github.com/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb) | How to fine-tune DistilBERT for text classification in TensorFlow | [Peter Bayerle](https://github.com/peterbayerle) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb)|
|[Leverage BERT for Encoder-Decoder Summarization on CNN/Dailymail](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb) | How to warm-start a *EncoderDecoderModel* with a *bert-base-uncased* checkpoint for summarization on CNN/Dailymail | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb)|
|[Leverage RoBERTa for Encoder-Decoder Summarization on BBC XSum](https://github.com/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb) | How to warm-start a shared *EncoderDecoderModel* with a *roberta-base* checkpoint for summarization on BBC/XSum | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb)|
|[Fine-tune TAPAS on Sequential Question Answering (SQA)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb) | How to fine-tune *TapasForQuestionAnswering* with a *tapas-base* checkpoint on the Sequential Question Answering (SQA) dataset | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb)|
|[Evaluate TAPAS on Table Fact Checking (TabFact)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb) | How to evaluate a fine-tuned *TapasForSequenceClassification* with a *tapas-base-finetuned-tabfact* checkpoint using a combination of the 🤗 datasets and 🤗 transformers libraries | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb)|
|[Fine-tuning mBART for translation](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb) | How to fine-tune mBART using Seq2SeqTrainer for Hindi to English translation | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb)|
|[Fine-tune LayoutLM on FUNSD (a form understanding dataset)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb) | How to fine-tune *LayoutLMForTokenClassification* on the FUNSD dataset for information extraction from scanned documents | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb)|
|[Fine-Tune DistilGPT2 and Generate Text](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb) | How to fine-tune DistilGPT2 and generate text | [Aakash Tripathi](https://github.com/tripathiaakash) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb)|
|[Fine-Tune LED on up to 8K tokens](https://github.com/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb) | How to fine-tune LED on pubmed for long-range summarization | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb)|
|[Evaluate LED on Arxiv](https://github.com/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb) | How to effectively evaluate LED on long-range summarization | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb)|
|[Fine-tune LayoutLM on RVL-CDIP (a document image classification dataset)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb) | How to fine-tune *LayoutLMForSequenceClassification* on the RVL-CDIP dataset for scanned document classification | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb)|
|[Wav2Vec2 CTC decoding with GPT2 adjustment](https://github.com/voidful/huggingface_notebook/blob/main/xlsr_gpt.ipynb) | How to decode CTC sequence with language model adjustment | [Eric Lam](https://github.com/voidful) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1e_z5jQHYbO2YKEaUgzb1ww1WwiAyydAj?usp=sharing)|
|[Fine-tune BART for summarization in two languages with Trainer class](https://github.com/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb) | How to fine-tune BART for summarization in two languages with Trainer class | [Eliza Szczechla](https://github.com/elsanns) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb)|
|[Evaluate Big Bird on Trivia QA](https://github.com/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb) | How to evaluate BigBird on long document question answering on Trivia QA | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb)|
| [Create video captions using Wav2Vec2](https://github.com/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) | How to create YouTube captions from any video by transcribing the audio with Wav2Vec | [Niklas Muennighoff](https://github.com/Muennighoff) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) |
| [Fine-tune the Vision Transformer on CIFAR-10 using PyTorch Lightning](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and PyTorch Lightning | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) |
| [Fine-tune the Vision Transformer on CIFAR-10 using the 🤗 Trainer](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) | How to fine-tune the Vision Transformer (ViT) on CIFAR-10 using HuggingFace Transformers, Datasets and the 🤗 Trainer | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) |
| [Evaluate LUKE on Open Entity, an entity typing dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) | How to evaluate *LukeForEntityClassification* on the Open Entity dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) |
| [Evaluate LUKE on TACRED, a relation extraction dataset](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) | How to evaluate *LukeForEntityPairClassification* on the TACRED dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) |
| [Evaluate LUKE on CoNLL-2003, an important NER benchmark](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) | How to evaluate *LukeForEntitySpanClassification* on the CoNLL-2003 dataset | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) |
| [Evaluate BigBird-Pegasus on PubMed dataset](https://github.com/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) | How to evaluate *BigBirdPegasusForConditionalGeneration* on PubMed dataset | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) |
| [Speech Emotion Classification with Wav2Vec2](https://github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) | How to leverage a pretrained Wav2Vec2 model for Emotion Classification on the MEGA dataset | [Mehrdad Farahani](https://github.com/m3hrdadfi) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) |
| [Detect objects in an image with DETR](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) | How to use a trained *DetrForObjectDetection* model to detect objects in an image and visualize attention | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) |
| [Fine-tune DETR on a custom object detection dataset](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) | How to fine-tune *DetrForObjectDetection* on a custom object detection dataset | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) |
| [Finetune T5 for Named Entity Recognition](https://github.com/ToluClassics/Notebooks/blob/main/T5_Ner_Finetuning.ipynb) | How to fine-tune *T5* on a Named Entity Recognition Task | [Ogundepo Odunayo](https://github.com/ToluClassics) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1obr78FY_cBmWY5ODViCmzdY6O1KB65Vc?usp=sharing) |
docs/source/en/create_a_model.md 0 → 100644
View file @ eb849f66
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Create a custom architecture
An [`AutoClass`](model_doc/auto) automatically infers the model architecture and downloads pretrained configuration and weights. Generally, we recommend using an `AutoClass` to produce checkpoint-agnostic code. But users who want more control over specific model parameters can create a custom 🤗 Transformers model from just a few base classes. This could be particularly useful for anyone who is interested in studying, training or experimenting with a 🤗 Transformers model. In this guide, dive deeper into creating a custom model without an `AutoClass`. Learn how to:
- Load and customize a model configuration.
- Create a model architecture.
- Create a slow and fast tokenizer for text.
- Create an image processor for vision tasks.
- Create a feature extractor for audio tasks.
- Create a processor for multimodal tasks.
## Configuration
A [configuration](main_classes/configuration) refers to a model's specific attributes. Each model configuration has different attributes; for instance, all NLP models have the `hidden_size`, `num_attention_heads`, `num_hidden_layers` and `vocab_size` attributes in common. These attributes specify the number of attention heads or hidden layers to construct a model with.
Get a closer look at [DistilBERT](model_doc/distilbert) by accessing [`DistilBertConfig`] to inspect it's attributes:
```py
>>> from transformers import DistilBertConfig
>>> config = DistilBertConfig()
>>> print(config)
DistilBertConfig {
"activation": "gelu",
"attention_dropout": 0.1,
"dim": 768,
"dropout": 0.1,
"hidden_dim": 3072,
"initializer_range": 0.02,
"max_position_embeddings": 512,
"model_type": "distilbert",
"n_heads": 12,
"n_layers": 6,
"pad_token_id": 0,
"qa_dropout": 0.1,
"seq_classif_dropout": 0.2,
"sinusoidal_pos_embds": false,
"transformers_version": "4.16.2",
"vocab_size": 30522
}
```
[`DistilBertConfig`] displays all the default attributes used to build a base [`DistilBertModel`]. All attributes are customizable, creating space for experimentation. For example, you can customize a default model to:
- Try a different activation function with the `activation` parameter.
- Use a higher dropout ratio for the attention probabilities with the `attention_dropout` parameter.
```py
>>> my_config = DistilBertConfig(activation="relu", attention_dropout=0.4)
>>> print(my_config)
DistilBertConfig {
"activation": "relu",
"attention_dropout": 0.4,
"dim": 768,
"dropout": 0.1,
"hidden_dim": 3072,
"initializer_range": 0.02,
"max_position_embeddings": 512,
"model_type": "distilbert",
"n_heads": 12,
"n_layers": 6,
"pad_token_id": 0,
"qa_dropout": 0.1,
"seq_classif_dropout": 0.2,
"sinusoidal_pos_embds": false,
"transformers_version": "4.16.2",
"vocab_size": 30522
}
```
Pretrained model attributes can be modified in the [`~PretrainedConfig.from_pretrained`] function:
```py
>>> my_config = DistilBertConfig.from_pretrained("distilbert-base-uncased", activation="relu", attention_dropout=0.4)
```
Once you are satisfied with your model configuration, you can save it with [`~PretrainedConfig.save_pretrained`]. Your configuration file is stored as a JSON file in the specified save directory:
```py
>>> my_config.save_pretrained(save_directory="./your_model_save_path")
```
To reuse the configuration file, load it with [`~PretrainedConfig.from_pretrained`]:
```py
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")
```
<Tip>
You can also save your configuration file as a dictionary or even just the difference between your custom configuration attributes and the default configuration attributes! See the [configuration](main_classes/configuration) documentation for more details.
</Tip>
## Model
The next step is to create a [model](main_classes/models). The model - also loosely referred to as the architecture - defines what each layer is doing and what operations are happening. Attributes like `num_hidden_layers` from the configuration are used to define the architecture. Every model shares the base class [`PreTrainedModel`] and a few common methods like resizing input embeddings and pruning self-attention heads. In addition, all models are also either a [`torch.nn.Module`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), [`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) or [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/flax.linen.html#module) subclass. This means models are compatible with each of their respective framework's usage.
<frameworkcontent>
<pt>
Load your custom configuration attributes into the model:
```py
>>> from transformers import DistilBertModel
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")
>>> model = DistilBertModel(my_config)
```
This creates a model with random values instead of pretrained weights. You won't be able to use this model for anything useful yet until you train it. Training is a costly and time-consuming process. It is generally better to use a pretrained model to obtain better results faster, while using only a fraction of the resources required for training.
Create a pretrained model with [`~PreTrainedModel.from_pretrained`]:
```py
>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased")
```
When you load pretrained weights, the default model configuration is automatically loaded if the model is provided by 🤗 Transformers. However, you can still replace - some or all of - the default model configuration attributes with your own if you'd like:
```py
>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
```
</pt>
<tf>
Load your custom configuration attributes into the model:
```py
>>> from transformers import TFDistilBertModel
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
>>> tf_model = TFDistilBertModel(my_config)
```
This creates a model with random values instead of pretrained weights. You won't be able to use this model for anything useful yet until you train it. Training is a costly and time-consuming process. It is generally better to use a pretrained model to obtain better results faster, while using only a fraction of the resources required for training.
Create a pretrained model with [`~TFPreTrainedModel.from_pretrained`]:
```py
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
```
When you load pretrained weights, the default model configuration is automatically loaded if the model is provided by 🤗 Transformers. However, you can still replace - some or all of - the default model configuration attributes with your own if you'd like:
```py
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
```
</tf>
</frameworkcontent>
### Model heads
At this point, you have a base DistilBERT model which outputs the *hidden states*. The hidden states are passed as inputs to a model head to produce the final output. 🤗 Transformers provides a different model head for each task as long as a model supports the task (i.e., you can't use DistilBERT for a sequence-to-sequence task like translation).
<frameworkcontent>
<pt>
For example, [`DistilBertForSequenceClassification`] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs.
```py
>>> from transformers import DistilBertForSequenceClassification
>>> model = DistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
```
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [`DistilBertForQuestionAnswering`] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output.
```py
>>> from transformers import DistilBertForQuestionAnswering
>>> model = DistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
```
</pt>
<tf>
For example, [`TFDistilBertForSequenceClassification`] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs.
```py
>>> from transformers import TFDistilBertForSequenceClassification
>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
```
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [`TFDistilBertForQuestionAnswering`] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output.
```py
>>> from transformers import TFDistilBertForQuestionAnswering
>>> tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
```
</tf>
</frameworkcontent>
## Tokenizer
The last base class you need before using a model for textual data is a [tokenizer](main_classes/tokenizer) to convert raw text to tensors. There are two types of tokenizers you can use with 🤗 Transformers:
- [`PreTrainedTokenizer`]: a Python implementation of a tokenizer.
- [`PreTrainedTokenizerFast`]: a tokenizer from our Rust-based [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) library. This tokenizer type is significantly faster - especially during batch tokenization - due to it's Rust implementation. The fast tokenizer also offers additional methods like *offset mapping* which maps tokens to their original words or characters.
Both tokenizers support common methods such as encoding and decoding, adding new tokens, and managing special tokens.
<Tip warning={true}>
Not every model supports a fast tokenizer. Take a look at this [table](index#supported-frameworks) to check if a model has fast tokenizer support.
</Tip>
If you trained your own tokenizer, you can create one from your *vocabulary* file:
```py
>>> from transformers import DistilBertTokenizer
>>> my_tokenizer = DistilBertTokenizer(vocab_file="my_vocab_file.txt", do_lower_case=False, padding_side="left")
```
It is important to remember the vocabulary from a custom tokenizer will be different from the vocabulary generated by a pretrained model's tokenizer. You need to use a pretrained model's vocabulary if you are using a pretrained model, otherwise the inputs won't make sense. Create a tokenizer with a pretrained model's vocabulary with the [`DistilBertTokenizer`] class:
```py
>>> from transformers import DistilBertTokenizer
>>> slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
```
Create a fast tokenizer with the [`DistilBertTokenizerFast`] class:
```py
>>> from transformers import DistilBertTokenizerFast
>>> fast_tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")
```
<Tip>
By default, [`AutoTokenizer`] will try to load a fast tokenizer. You can disable this behavior by setting `use_fast=False` in `from_pretrained`.
</Tip>
## Image Processor
An image processor processes vision inputs. It inherits from the base [`~image_processing_utils.ImageProcessingMixin`] class.
To use, create an image processor associated with the model you're using. For example, create a default [`ViTImageProcessor`] if you are using [ViT](model_doc/vit) for image classification:
```py
>>> from transformers import ViTImageProcessor
>>> vit_extractor = ViTImageProcessor()
>>> print(vit_extractor)
ViTImageProcessor {
"do_normalize": true,
"do_resize": true,
"feature_extractor_type": "ViTImageProcessor",
"image_mean": [
0.5,
0.5,
0.5
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": 2,
"size": 224
}
```
<Tip>
If you aren't looking for any customization, just use the `from_pretrained` method to load a model's default image processor parameters.
</Tip>
Modify any of the [`ViTImageProcessor`] parameters to create your custom image processor:
```py
>>> from transformers import ViTImageProcessor
>>> my_vit_extractor = ViTImageProcessor(resample="PIL.Image.BOX", do_normalize=False, image_mean=[0.3, 0.3, 0.3])
>>> print(my_vit_extractor)
ViTImageProcessor {
"do_normalize": false,
"do_resize": true,
"feature_extractor_type": "ViTImageProcessor",
"image_mean": [
0.3,
0.3,
0.3
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": "PIL.Image.BOX",
"size": 224
}
```
## Feature Extractor
A feature extractor processes audio inputs. It inherits from the base [`~feature_extraction_utils.FeatureExtractionMixin`] class, and may also inherit from the [`SequenceFeatureExtractor`] class for processing audio inputs.
To use, create a feature extractor associated with the model you're using. For example, create a default [`Wav2Vec2FeatureExtractor`] if you are using [Wav2Vec2](model_doc/wav2vec2) for audio classification:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> w2v2_extractor = Wav2Vec2FeatureExtractor()
>>> print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": true,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 16000
}
```
<Tip>
If you aren't looking for any customization, just use the `from_pretrained` method to load a model's default feature extractor parameters.
</Tip>
Modify any of the [`Wav2Vec2FeatureExtractor`] parameters to create your custom feature extractor:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> w2v2_extractor = Wav2Vec2FeatureExtractor(sampling_rate=8000, do_normalize=False)
>>> print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": false,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 8000
}
```
## Processor
For models that support multimodal tasks, 🤗 Transformers offers a processor class that conveniently wraps processing classes such as a feature extractor and a tokenizer into a single object. For example, let's use the [`Wav2Vec2Processor`] for an automatic speech recognition task (ASR). ASR transcribes audio to text, so you will need a feature extractor and a tokenizer.
Create a feature extractor to handle the audio inputs:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> feature_extractor = Wav2Vec2FeatureExtractor(padding_value=1.0, do_normalize=True)
```
Create a tokenizer to handle the text inputs:
```py
>>> from transformers import Wav2Vec2CTCTokenizer
>>> tokenizer = Wav2Vec2CTCTokenizer(vocab_file="my_vocab_file.txt")
```
Combine the feature extractor and tokenizer in [`Wav2Vec2Processor`]:
```py
>>> from transformers import Wav2Vec2Processor
>>> processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
```
With two basic classes - configuration and model - and an additional preprocessing class (tokenizer, image processor, feature extractor, or processor), you can create any of the models supported by 🤗 Transformers. Each of these base classes are configurable, allowing you to use the specific attributes you want. You can easily setup a model for training or modify an existing pretrained model to fine-tune.
docs/source/en/create_a_model.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Create a custom architecture
An [`AutoClass`](model_doc/auto) automatically infers the model architecture and downloads pretrained configuration and weights. Generally, we recommend using an `AutoClass` to produce checkpoint-agnostic code. But users who want more control over specific model parameters can create a custom 🤗 Transformers model from just a few base classes. This could be particularly useful for anyone who is interested in studying, training or experimenting with a 🤗 Transformers model. In this guide, dive deeper into creating a custom model without an `AutoClass`. Learn how to:
- Load and customize a model configuration.
- Create a model architecture.
- Create a slow and fast tokenizer for text.
- Create an image processor for vision tasks.
- Create a feature extractor for audio tasks.
- Create a processor for multimodal tasks.
## Configuration
A [configuration](main_classes/configuration) refers to a model's specific attributes. Each model configuration has different attributes; for instance, all NLP models have the `hidden_size`, `num_attention_heads`, `num_hidden_layers` and `vocab_size` attributes in common. These attributes specify the number of attention heads or hidden layers to construct a model with.
Get a closer look at [DistilBERT](model_doc/distilbert) by accessing [`DistilBertConfig`] to inspect it's attributes:
```py
>>> from transformers import DistilBertConfig
>>> config = DistilBertConfig()
>>> print(config)
DistilBertConfig {
"activation": "gelu",
"attention_dropout": 0.1,
"dim": 768,
"dropout": 0.1,
"hidden_dim": 3072,
"initializer_range": 0.02,
"max_position_embeddings": 512,
"model_type": "distilbert",
"n_heads": 12,
"n_layers": 6,
"pad_token_id": 0,
"qa_dropout": 0.1,
"seq_classif_dropout": 0.2,
"sinusoidal_pos_embds": false,
"transformers_version": "4.16.2",
"vocab_size": 30522
}
```
[`DistilBertConfig`] displays all the default attributes used to build a base [`DistilBertModel`]. All attributes are customizable, creating space for experimentation. For example, you can customize a default model to:
- Try a different activation function with the `activation` parameter.
- Use a higher dropout ratio for the attention probabilities with the `attention_dropout` parameter.
```py
>>> my_config = DistilBertConfig(activation="relu", attention_dropout=0.4)
>>> print(my_config)
DistilBertConfig {
"activation": "relu",
"attention_dropout": 0.4,
"dim": 768,
"dropout": 0.1,
"hidden_dim": 3072,
"initializer_range": 0.02,
"max_position_embeddings": 512,
"model_type": "distilbert",
"n_heads": 12,
"n_layers": 6,
"pad_token_id": 0,
"qa_dropout": 0.1,
"seq_classif_dropout": 0.2,
"sinusoidal_pos_embds": false,
"transformers_version": "4.16.2",
"vocab_size": 30522
}
```
Pretrained model attributes can be modified in the [`~PretrainedConfig.from_pretrained`] function:
```py
>>> my_config = DistilBertConfig.from_pretrained("distilbert-base-uncased", activation="relu", attention_dropout=0.4)
```
Once you are satisfied with your model configuration, you can save it with [`~PretrainedConfig.save_pretrained`]. Your configuration file is stored as a JSON file in the specified save directory:
```py
>>> my_config.save_pretrained(save_directory="./your_model_save_path")
```
To reuse the configuration file, load it with [`~PretrainedConfig.from_pretrained`]:
```py
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")
```
<Tip>
You can also save your configuration file as a dictionary or even just the difference between your custom configuration attributes and the default configuration attributes! See the [configuration](main_classes/configuration) documentation for more details.
</Tip>
## Model
The next step is to create a [model](main_classes/models). The model - also loosely referred to as the architecture - defines what each layer is doing and what operations are happening. Attributes like `num_hidden_layers` from the configuration are used to define the architecture. Every model shares the base class [`PreTrainedModel`] and a few common methods like resizing input embeddings and pruning self-attention heads. In addition, all models are also either a [`torch.nn.Module`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html), [`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) or [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/flax.linen.html#module) subclass. This means models are compatible with each of their respective framework's usage.
<frameworkcontent>
<pt>
Load your custom configuration attributes into the model:
```py
>>> from transformers import DistilBertModel
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")
>>> model = DistilBertModel(my_config)
```
This creates a model with random values instead of pretrained weights. You won't be able to use this model for anything useful yet until you train it. Training is a costly and time-consuming process. It is generally better to use a pretrained model to obtain better results faster, while using only a fraction of the resources required for training.
Create a pretrained model with [`~PreTrainedModel.from_pretrained`]:
```py
>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased")
```
When you load pretrained weights, the default model configuration is automatically loaded if the model is provided by 🤗 Transformers. However, you can still replace - some or all of - the default model configuration attributes with your own if you'd like:
```py
>>> model = DistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
```
</pt>
<tf>
Load your custom configuration attributes into the model:
```py
>>> from transformers import TFDistilBertModel
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
>>> tf_model = TFDistilBertModel(my_config)
```
This creates a model with random values instead of pretrained weights. You won't be able to use this model for anything useful yet until you train it. Training is a costly and time-consuming process. It is generally better to use a pretrained model to obtain better results faster, while using only a fraction of the resources required for training.
Create a pretrained model with [`~TFPreTrainedModel.from_pretrained`]:
```py
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased")
```
When you load pretrained weights, the default model configuration is automatically loaded if the model is provided by 🤗 Transformers. However, you can still replace - some or all of - the default model configuration attributes with your own if you'd like:
```py
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert-base-uncased", config=my_config)
```
</tf>
</frameworkcontent>
### Model heads
At this point, you have a base DistilBERT model which outputs the *hidden states*. The hidden states are passed as inputs to a model head to produce the final output. 🤗 Transformers provides a different model head for each task as long as a model supports the task (i.e., you can't use DistilBERT for a sequence-to-sequence task like translation).
<frameworkcontent>
<pt>
For example, [`DistilBertForSequenceClassification`] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs.
```py
>>> from transformers import DistilBertForSequenceClassification
>>> model = DistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
```
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [`DistilBertForQuestionAnswering`] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output.
```py
>>> from transformers import DistilBertForQuestionAnswering
>>> model = DistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
```
</pt>
<tf>
For example, [`TFDistilBertForSequenceClassification`] is a base DistilBERT model with a sequence classification head. The sequence classification head is a linear layer on top of the pooled outputs.
```py
>>> from transformers import TFDistilBertForSequenceClassification
>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert-base-uncased")
```
Easily reuse this checkpoint for another task by switching to a different model head. For a question answering task, you would use the [`TFDistilBertForQuestionAnswering`] model head. The question answering head is similar to the sequence classification head except it is a linear layer on top of the hidden states output.
```py
>>> from transformers import TFDistilBertForQuestionAnswering
>>> tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert-base-uncased")
```
</tf>
</frameworkcontent>
## Tokenizer
The last base class you need before using a model for textual data is a [tokenizer](main_classes/tokenizer) to convert raw text to tensors. There are two types of tokenizers you can use with 🤗 Transformers:
- [`PreTrainedTokenizer`]: a Python implementation of a tokenizer.
- [`PreTrainedTokenizerFast`]: a tokenizer from our Rust-based [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) library. This tokenizer type is significantly faster - especially during batch tokenization - due to it's Rust implementation. The fast tokenizer also offers additional methods like *offset mapping* which maps tokens to their original words or characters.
Both tokenizers support common methods such as encoding and decoding, adding new tokens, and managing special tokens.
<Tip warning={true}>
Not every model supports a fast tokenizer. Take a look at this [table](index#supported-frameworks) to check if a model has fast tokenizer support.
</Tip>
If you trained your own tokenizer, you can create one from your *vocabulary* file:
```py
>>> from transformers import DistilBertTokenizer
>>> my_tokenizer = DistilBertTokenizer(vocab_file="my_vocab_file.txt", do_lower_case=False, padding_side="left")
```
It is important to remember the vocabulary from a custom tokenizer will be different from the vocabulary generated by a pretrained model's tokenizer. You need to use a pretrained model's vocabulary if you are using a pretrained model, otherwise the inputs won't make sense. Create a tokenizer with a pretrained model's vocabulary with the [`DistilBertTokenizer`] class:
```py
>>> from transformers import DistilBertTokenizer
>>> slow_tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
```
Create a fast tokenizer with the [`DistilBertTokenizerFast`] class:
```py
>>> from transformers import DistilBertTokenizerFast
>>> fast_tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")
```
<Tip>
By default, [`AutoTokenizer`] will try to load a fast tokenizer. You can disable this behavior by setting `use_fast=False` in `from_pretrained`.
</Tip>
## Image Processor
An image processor processes vision inputs. It inherits from the base [`~image_processing_utils.ImageProcessingMixin`] class.
To use, create an image processor associated with the model you're using. For example, create a default [`ViTImageProcessor`] if you are using [ViT](model_doc/vit) for image classification:
```py
>>> from transformers import ViTImageProcessor
>>> vit_extractor = ViTImageProcessor()
>>> print(vit_extractor)
ViTImageProcessor {
"do_normalize": true,
"do_resize": true,
"feature_extractor_type": "ViTImageProcessor",
"image_mean": [
0.5,
0.5,
0.5
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": 2,
"size": 224
}
```
<Tip>
If you aren't looking for any customization, just use the `from_pretrained` method to load a model's default image processor parameters.
</Tip>
Modify any of the [`ViTImageProcessor`] parameters to create your custom image processor:
```py
>>> from transformers import ViTImageProcessor
>>> my_vit_extractor = ViTImageProcessor(resample="PIL.Image.BOX", do_normalize=False, image_mean=[0.3, 0.3, 0.3])
>>> print(my_vit_extractor)
ViTImageProcessor {
"do_normalize": false,
"do_resize": true,
"feature_extractor_type": "ViTImageProcessor",
"image_mean": [
0.3,
0.3,
0.3
],
"image_std": [
0.5,
0.5,
0.5
],
"resample": "PIL.Image.BOX",
"size": 224
}
```
## Feature Extractor
A feature extractor processes audio inputs. It inherits from the base [`~feature_extraction_utils.FeatureExtractionMixin`] class, and may also inherit from the [`SequenceFeatureExtractor`] class for processing audio inputs.
To use, create a feature extractor associated with the model you're using. For example, create a default [`Wav2Vec2FeatureExtractor`] if you are using [Wav2Vec2](model_doc/wav2vec2) for audio classification:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> w2v2_extractor = Wav2Vec2FeatureExtractor()
>>> print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": true,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 16000
}
```
<Tip>
If you aren't looking for any customization, just use the `from_pretrained` method to load a model's default feature extractor parameters.
</Tip>
Modify any of the [`Wav2Vec2FeatureExtractor`] parameters to create your custom feature extractor:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> w2v2_extractor = Wav2Vec2FeatureExtractor(sampling_rate=8000, do_normalize=False)
>>> print(w2v2_extractor)
Wav2Vec2FeatureExtractor {
"do_normalize": false,
"feature_extractor_type": "Wav2Vec2FeatureExtractor",
"feature_size": 1,
"padding_side": "right",
"padding_value": 0.0,
"return_attention_mask": false,
"sampling_rate": 8000
}
```
## Processor
For models that support multimodal tasks, 🤗 Transformers offers a processor class that conveniently wraps processing classes such as a feature extractor and a tokenizer into a single object. For example, let's use the [`Wav2Vec2Processor`] for an automatic speech recognition task (ASR). ASR transcribes audio to text, so you will need a feature extractor and a tokenizer.
Create a feature extractor to handle the audio inputs:
```py
>>> from transformers import Wav2Vec2FeatureExtractor
>>> feature_extractor = Wav2Vec2FeatureExtractor(padding_value=1.0, do_normalize=True)
```
Create a tokenizer to handle the text inputs:
```py
>>> from transformers import Wav2Vec2CTCTokenizer
>>> tokenizer = Wav2Vec2CTCTokenizer(vocab_file="my_vocab_file.txt")
```
Combine the feature extractor and tokenizer in [`Wav2Vec2Processor`]:
```py
>>> from transformers import Wav2Vec2Processor
>>> processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
```
With two basic classes - configuration and model - and an additional preprocessing class (tokenizer, image processor, feature extractor, or processor), you can create any of the models supported by 🤗 Transformers. Each of these base classes are configurable, allowing you to use the specific attributes you want. You can easily setup a model for training or modify an existing pretrained model to fine-tune.
docs/source/en/custom_models.md 0 → 100644
View file @ eb849f66
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Sharing custom models
The 🤗 Transformers library is designed to be easily extensible. Every model is fully coded in a given subfolder
of the repository with no abstraction, so you can easily copy a modeling file and tweak it to your needs.
If you are writing a brand new model, it might be easier to start from scratch. In this tutorial, we will show you
how to write a custom model and its configuration so it can be used inside Transformers, and how you can share it
with the community (with the code it relies on) so that anyone can use it, even if it's not present in the 🤗
Transformers library.
We will illustrate all of this on a ResNet model, by wrapping the ResNet class of the
[timm library](https://github.com/rwightman/pytorch-image-models) into a [`PreTrainedModel`].
## Writing a custom configuration
Before we dive into the model, let's first write its configuration. The configuration of a model is an object that
will contain all the necessary information to build the model. As we will see in the next section, the model can only
take a `config` to be initialized, so we really need that object to be as complete as possible.
In our example, we will take a couple of arguments of the ResNet class that we might want to tweak. Different
configurations will then give us the different types of ResNets that are possible. We then just store those arguments,
after checking the validity of a few of them.
```python
from transformers import PretrainedConfig
from typing import List
class ResnetConfig(PretrainedConfig):
model_type = "resnet"
def __init__(
self,
block_type="bottleneck",
layers: List[int] = [3, 4, 6, 3],
num_classes: int = 1000,
input_channels: int = 3,
cardinality: int = 1,
base_width: int = 64,
stem_width: int = 64,
stem_type: str = "",
avg_down: bool = False,
**kwargs,
):
if block_type not in ["basic", "bottleneck"]:
raise ValueError(f"`block_type` must be 'basic' or bottleneck', got {block_type}.")
if stem_type not in ["", "deep", "deep-tiered"]:
raise ValueError(f"`stem_type` must be '', 'deep' or 'deep-tiered', got {stem_type}.")
self.block_type = block_type
self.layers = layers
self.num_classes = num_classes
self.input_channels = input_channels
self.cardinality = cardinality
self.base_width = base_width
self.stem_width = stem_width
self.stem_type = stem_type
self.avg_down = avg_down
super().__init__(**kwargs)
```
The three important things to remember when writing you own configuration are the following:
- you have to inherit from `PretrainedConfig`,
- the `__init__` of your `PretrainedConfig` must accept any kwargs,
- those `kwargs` need to be passed to the superclass `__init__`.
The inheritance is to make sure you get all the functionality from the 🤗 Transformers library, while the two other
constraints come from the fact a `PretrainedConfig` has more fields than the ones you are setting. When reloading a
config with the `from_pretrained` method, those fields need to be accepted by your config and then sent to the
superclass.
Defining a `model_type` for your configuration (here `model_type="resnet"`) is not mandatory, unless you want to
register your model with the auto classes (see last section).
With this done, you can easily create and save your configuration like you would do with any other model config of the
library. Here is how we can create a resnet50d config and save it:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d_config.save_pretrained("custom-resnet")
```
This will save a file named `config.json` inside the folder `custom-resnet`. You can then reload your config with the
`from_pretrained` method:
```py
resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
```
You can also use any other method of the [`PretrainedConfig`] class, like [`~PretrainedConfig.push_to_hub`] to
directly upload your config to the Hub.
## Writing a custom model
Now that we have our ResNet configuration, we can go on writing the model. We will actually write two: one that
extracts the hidden features from a batch of images (like [`BertModel`]) and one that is suitable for image
classification (like [`BertForSequenceClassification`]).
As we mentioned before, we'll only write a loose wrapper of the model to keep it simple for this example. The only
thing we need to do before writing this class is a map between the block types and actual block classes. Then the
model is defined from the configuration by passing everything to the `ResNet` class:
```py
from transformers import PreTrainedModel
from timm.models.resnet import BasicBlock, Bottleneck, ResNet
from .configuration_resnet import ResnetConfig
BLOCK_MAPPING = {"basic": BasicBlock, "bottleneck": Bottleneck}
class ResnetModel(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor):
return self.model.forward_features(tensor)
```
For the model that will classify images, we just change the forward method:
```py
import torch
class ResnetModelForImageClassification(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor, labels=None):
logits = self.model(tensor)
if labels is not None:
loss = torch.nn.cross_entropy(logits, labels)
return {"loss": loss, "logits": logits}
return {"logits": logits}
```
In both cases, notice how we inherit from `PreTrainedModel` and call the superclass initialization with the `config`
(a bit like when you write a regular `torch.nn.Module`). The line that sets the `config_class` is not mandatory, unless
you want to register your model with the auto classes (see last section).
<Tip>
If your model is very similar to a model inside the library, you can re-use the same configuration as this model.
</Tip>
You can have your model return anything you want, but returning a dictionary like we did for
`ResnetModelForImageClassification`, with the loss included when labels are passed, will make your model directly
usable inside the [`Trainer`] class. Using another output format is fine as long as you are planning on using your own
training loop or another library for training.
Now that we have our model class, let's create one:
```py
resnet50d = ResnetModelForImageClassification(resnet50d_config)
```
Again, you can use any of the methods of [`PreTrainedModel`], like [`~PreTrainedModel.save_pretrained`] or
[`~PreTrainedModel.push_to_hub`]. We will use the second in the next section, and see how to push the model weights
with the code of our model. But first, let's load some pretrained weights inside our model.
In your own use case, you will probably be training your custom model on your own data. To go fast for this tutorial,
we will use the pretrained version of the resnet50d. Since our model is just a wrapper around it, it's going to be
easy to transfer those weights:
```py
import timm
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Now let's see how to make sure that when we do [`~PreTrainedModel.save_pretrained`] or [`~PreTrainedModel.push_to_hub`], the
code of the model is saved.
## Sending the code to the Hub
<Tip warning={true}>
This API is experimental and may have some slight breaking changes in the next releases.
</Tip>
First, make sure your model is fully defined in a `.py` file. It can rely on relative imports to some other files as
long as all the files are in the same directory (we don't support submodules for this feature yet). For our example,
we'll define a `modeling_resnet.py` file and a `configuration_resnet.py` file in a folder of the current working
directory named `resnet_model`. The configuration file contains the code for `ResnetConfig` and the modeling file
contains the code of `ResnetModel` and `ResnetModelForImageClassification`.
```
.
└── resnet_model
├── __init__.py
├── configuration_resnet.py
└── modeling_resnet.py
```
The `__init__.py` can be empty, it's just there so that Python detects `resnet_model` can be use as a module.
<Tip warning={true}>
If copying a modeling files from the library, you will need to replace all the relative imports at the top of the file
to import from the `transformers` package.
</Tip>
Note that you can re-use (or subclass) an existing configuration/model.
To share your model with the community, follow those steps: first import the ResNet model and config from the newly
created files:
```py
from resnet_model.configuration_resnet import ResnetConfig
from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
```
Then you have to tell the library you want to copy the code files of those objects when using the `save_pretrained`
method and properly register them with a given Auto class (especially for models), just run:
```py
ResnetConfig.register_for_auto_class()
ResnetModel.register_for_auto_class("AutoModel")
ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
```
Note that there is no need to specify an auto class for the configuration (there is only one auto class for them,
[`AutoConfig`]) but it's different for models. Your custom model could be suitable for many different tasks, so you
have to specify which one of the auto classes is the correct one for your model.
Next, let's create the config and models as we did before:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d = ResnetModelForImageClassification(resnet50d_config)
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Now to send the model to the Hub, make sure you are logged in. Either run in your terminal:
```bash
huggingface-cli login
```
or from a notebook:
```py
from huggingface_hub import notebook_login
notebook_login()
```
You can then push to your own namespace (or an organization you are a member of) like this:
```py
resnet50d.push_to_hub("custom-resnet50d")
```
On top of the modeling weights and the configuration in json format, this also copied the modeling and
configuration `.py` files in the folder `custom-resnet50d` and uploaded the result to the Hub. You can check the result
in this [model repo](https://huggingface.co/sgugger/custom-resnet50d).
See the [sharing tutorial](model_sharing) for more information on the push to Hub method.
## Using a model with custom code
You can use any configuration, model or tokenizer with custom code files in its repository with the auto-classes and
the `from_pretrained` method. All files and code uploaded to the Hub are scanned for malware (refer to the [Hub security](https://huggingface.co/docs/hub/security#malware-scanning) documentation for more information), but you should still
review the model code and author to avoid executing malicious code on your machine. Set `trust_remote_code=True` to use
a model with custom code:
```py
from transformers import AutoModelForImageClassification
model = AutoModelForImageClassification.from_pretrained("sgugger/custom-resnet50d", trust_remote_code=True)
```
It is also strongly encouraged to pass a commit hash as a `revision` to make sure the author of the models did not
update the code with some malicious new lines (unless you fully trust the authors of the models).
```py
commit_hash = "ed94a7c6247d8aedce4647f00f20de6875b5b292"
model = AutoModelForImageClassification.from_pretrained(
"sgugger/custom-resnet50d", trust_remote_code=True, revision=commit_hash
)
```
Note that when browsing the commit history of the model repo on the Hub, there is a button to easily copy the commit
hash of any commit.
## Registering a model with custom code to the auto classes
If you are writing a library that extends 🤗 Transformers, you may want to extend the auto classes to include your own
model. This is different from pushing the code to the Hub in the sense that users will need to import your library to
get the custom models (contrarily to automatically downloading the model code from the Hub).
As long as your config has a `model_type` attribute that is different from existing model types, and that your model
classes have the right `config_class` attributes, you can just add them to the auto classes likes this:
```py
from transformers import AutoConfig, AutoModel, AutoModelForImageClassification
AutoConfig.register("resnet", ResnetConfig)
AutoModel.register(ResnetConfig, ResnetModel)
AutoModelForImageClassification.register(ResnetConfig, ResnetModelForImageClassification)
```
Note that the first argument used when registering your custom config to [`AutoConfig`] needs to match the `model_type`
of your custom config, and the first argument used when registering your custom models to any auto model class needs
to match the `config_class` of those models.
docs/source/en/custom_models.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Sharing custom models
The 🤗 Transformers library is designed to be easily extensible. Every model is fully coded in a given subfolder
of the repository with no abstraction, so you can easily copy a modeling file and tweak it to your needs.
If you are writing a brand new model, it might be easier to start from scratch. In this tutorial, we will show you
how to write a custom model and its configuration so it can be used inside Transformers, and how you can share it
with the community (with the code it relies on) so that anyone can use it, even if it's not present in the 🤗
Transformers library.
We will illustrate all of this on a ResNet model, by wrapping the ResNet class of the
[timm library](https://github.com/rwightman/pytorch-image-models) into a [`PreTrainedModel`].
## Writing a custom configuration
Before we dive into the model, let's first write its configuration. The configuration of a model is an object that
will contain all the necessary information to build the model. As we will see in the next section, the model can only
take a `config` to be initialized, so we really need that object to be as complete as possible.
In our example, we will take a couple of arguments of the ResNet class that we might want to tweak. Different
configurations will then give us the different types of ResNets that are possible. We then just store those arguments,
after checking the validity of a few of them.
```python
from transformers import PretrainedConfig
from typing import List
class ResnetConfig(PretrainedConfig):
model_type = "resnet"
def __init__(
self,
block_type="bottleneck",
layers: List[int] = [3, 4, 6, 3],
num_classes: int = 1000,
input_channels: int = 3,
cardinality: int = 1,
base_width: int = 64,
stem_width: int = 64,
stem_type: str = "",
avg_down: bool = False,
**kwargs,
):
if block_type not in ["basic", "bottleneck"]:
raise ValueError(f"`block_type` must be 'basic' or bottleneck', got {block_type}.")
if stem_type not in ["", "deep", "deep-tiered"]:
raise ValueError(f"`stem_type` must be '', 'deep' or 'deep-tiered', got {stem_type}.")
self.block_type = block_type
self.layers = layers
self.num_classes = num_classes
self.input_channels = input_channels
self.cardinality = cardinality
self.base_width = base_width
self.stem_width = stem_width
self.stem_type = stem_type
self.avg_down = avg_down
super().__init__(**kwargs)
```
The three important things to remember when writing you own configuration are the following:
- you have to inherit from `PretrainedConfig`,
- the `__init__` of your `PretrainedConfig` must accept any kwargs,
- those `kwargs` need to be passed to the superclass `__init__`.
The inheritance is to make sure you get all the functionality from the 🤗 Transformers library, while the two other
constraints come from the fact a `PretrainedConfig` has more fields than the ones you are setting. When reloading a
config with the `from_pretrained` method, those fields need to be accepted by your config and then sent to the
superclass.
Defining a `model_type` for your configuration (here `model_type="resnet"`) is not mandatory, unless you want to
register your model with the auto classes (see last section).
With this done, you can easily create and save your configuration like you would do with any other model config of the
library. Here is how we can create a resnet50d config and save it:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d_config.save_pretrained("custom-resnet")
```
This will save a file named `config.json` inside the folder `custom-resnet`. You can then reload your config with the
`from_pretrained` method:
```py
resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
```
You can also use any other method of the [`PretrainedConfig`] class, like [`~PretrainedConfig.push_to_hub`] to
directly upload your config to the Hub.
## Writing a custom model
Now that we have our ResNet configuration, we can go on writing the model. We will actually write two: one that
extracts the hidden features from a batch of images (like [`BertModel`]) and one that is suitable for image
classification (like [`BertForSequenceClassification`]).
As we mentioned before, we'll only write a loose wrapper of the model to keep it simple for this example. The only
thing we need to do before writing this class is a map between the block types and actual block classes. Then the
model is defined from the configuration by passing everything to the `ResNet` class:
```py
from transformers import PreTrainedModel
from timm.models.resnet import BasicBlock, Bottleneck, ResNet
from .configuration_resnet import ResnetConfig
BLOCK_MAPPING = {"basic": BasicBlock, "bottleneck": Bottleneck}
class ResnetModel(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor):
return self.model.forward_features(tensor)
```
For the model that will classify images, we just change the forward method:
```py
import torch
class ResnetModelForImageClassification(PreTrainedModel):
config_class = ResnetConfig
def __init__(self, config):
super().__init__(config)
block_layer = BLOCK_MAPPING[config.block_type]
self.model = ResNet(
block_layer,
config.layers,
num_classes=config.num_classes,
in_chans=config.input_channels,
cardinality=config.cardinality,
base_width=config.base_width,
stem_width=config.stem_width,
stem_type=config.stem_type,
avg_down=config.avg_down,
)
def forward(self, tensor, labels=None):
logits = self.model(tensor)
if labels is not None:
loss = torch.nn.cross_entropy(logits, labels)
return {"loss": loss, "logits": logits}
return {"logits": logits}
```
In both cases, notice how we inherit from `PreTrainedModel` and call the superclass initialization with the `config`
(a bit like when you write a regular `torch.nn.Module`). The line that sets the `config_class` is not mandatory, unless
you want to register your model with the auto classes (see last section).
<Tip>
If your model is very similar to a model inside the library, you can re-use the same configuration as this model.
</Tip>
You can have your model return anything you want, but returning a dictionary like we did for
`ResnetModelForImageClassification`, with the loss included when labels are passed, will make your model directly
usable inside the [`Trainer`] class. Using another output format is fine as long as you are planning on using your own
training loop or another library for training.
Now that we have our model class, let's create one:
```py
resnet50d = ResnetModelForImageClassification(resnet50d_config)
```
Again, you can use any of the methods of [`PreTrainedModel`], like [`~PreTrainedModel.save_pretrained`] or
[`~PreTrainedModel.push_to_hub`]. We will use the second in the next section, and see how to push the model weights
with the code of our model. But first, let's load some pretrained weights inside our model.
In your own use case, you will probably be training your custom model on your own data. To go fast for this tutorial,
we will use the pretrained version of the resnet50d. Since our model is just a wrapper around it, it's going to be
easy to transfer those weights:
```py
import timm
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Now let's see how to make sure that when we do [`~PreTrainedModel.save_pretrained`] or [`~PreTrainedModel.push_to_hub`], the
code of the model is saved.
## Sending the code to the Hub
<Tip warning={true}>
This API is experimental and may have some slight breaking changes in the next releases.
</Tip>
First, make sure your model is fully defined in a `.py` file. It can rely on relative imports to some other files as
long as all the files are in the same directory (we don't support submodules for this feature yet). For our example,
we'll define a `modeling_resnet.py` file and a `configuration_resnet.py` file in a folder of the current working
directory named `resnet_model`. The configuration file contains the code for `ResnetConfig` and the modeling file
contains the code of `ResnetModel` and `ResnetModelForImageClassification`.
```
.
└── resnet_model
├── __init__.py
├── configuration_resnet.py
└── modeling_resnet.py
```
The `__init__.py` can be empty, it's just there so that Python detects `resnet_model` can be use as a module.
<Tip warning={true}>
If copying a modeling files from the library, you will need to replace all the relative imports at the top of the file
to import from the `transformers` package.
</Tip>
Note that you can re-use (or subclass) an existing configuration/model.
To share your model with the community, follow those steps: first import the ResNet model and config from the newly
created files:
```py
from resnet_model.configuration_resnet import ResnetConfig
from resnet_model.modeling_resnet import ResnetModel, ResnetModelForImageClassification
```
Then you have to tell the library you want to copy the code files of those objects when using the `save_pretrained`
method and properly register them with a given Auto class (especially for models), just run:
```py
ResnetConfig.register_for_auto_class()
ResnetModel.register_for_auto_class("AutoModel")
ResnetModelForImageClassification.register_for_auto_class("AutoModelForImageClassification")
```
Note that there is no need to specify an auto class for the configuration (there is only one auto class for them,
[`AutoConfig`]) but it's different for models. Your custom model could be suitable for many different tasks, so you
have to specify which one of the auto classes is the correct one for your model.
Next, let's create the config and models as we did before:
```py
resnet50d_config = ResnetConfig(block_type="bottleneck", stem_width=32, stem_type="deep", avg_down=True)
resnet50d = ResnetModelForImageClassification(resnet50d_config)
pretrained_model = timm.create_model("resnet50d", pretrained=True)
resnet50d.model.load_state_dict(pretrained_model.state_dict())
```
Now to send the model to the Hub, make sure you are logged in. Either run in your terminal:
```bash
huggingface-cli login
```
or from a notebook:
```py
from huggingface_hub import notebook_login
notebook_login()
```
You can then push to your own namespace (or an organization you are a member of) like this:
```py
resnet50d.push_to_hub("custom-resnet50d")
```
On top of the modeling weights and the configuration in json format, this also copied the modeling and
configuration `.py` files in the folder `custom-resnet50d` and uploaded the result to the Hub. You can check the result
in this [model repo](https://huggingface.co/sgugger/custom-resnet50d).
See the [sharing tutorial](model_sharing) for more information on the push to Hub method.
## Using a model with custom code
You can use any configuration, model or tokenizer with custom code files in its repository with the auto-classes and
the `from_pretrained` method. All files and code uploaded to the Hub are scanned for malware (refer to the [Hub security](https://huggingface.co/docs/hub/security#malware-scanning) documentation for more information), but you should still
review the model code and author to avoid executing malicious code on your machine. Set `trust_remote_code=True` to use
a model with custom code:
```py
from transformers import AutoModelForImageClassification
model = AutoModelForImageClassification.from_pretrained("sgugger/custom-resnet50d", trust_remote_code=True)
```
It is also strongly encouraged to pass a commit hash as a `revision` to make sure the author of the models did not
update the code with some malicious new lines (unless you fully trust the authors of the models).
```py
commit_hash = "ed94a7c6247d8aedce4647f00f20de6875b5b292"
model = AutoModelForImageClassification.from_pretrained(
"sgugger/custom-resnet50d", trust_remote_code=True, revision=commit_hash
)
```
Note that when browsing the commit history of the model repo on the Hub, there is a button to easily copy the commit
hash of any commit.
## Registering a model with custom code to the auto classes
If you are writing a library that extends 🤗 Transformers, you may want to extend the auto classes to include your own
model. This is different from pushing the code to the Hub in the sense that users will need to import your library to
get the custom models (contrarily to automatically downloading the model code from the Hub).
As long as your config has a `model_type` attribute that is different from existing model types, and that your model
classes have the right `config_class` attributes, you can just add them to the auto classes likes this:
```py
from transformers import AutoConfig, AutoModel, AutoModelForImageClassification
AutoConfig.register("resnet", ResnetConfig)
AutoModel.register(ResnetConfig, ResnetModel)
AutoModelForImageClassification.register(ResnetConfig, ResnetModelForImageClassification)
```
Note that the first argument used when registering your custom config to [`AutoConfig`] needs to match the `model_type`
of your custom config, and the first argument used when registering your custom models to any auto model class needs
to match the `config_class` of those models.
docs/source/en/custom_tools.md 0 → 100644
View file @ eb849f66
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Custom Tools and Prompts
<Tip>
If you are not aware of what tools and agents are in the context of transformers, we recommend you read the
[Transformers Agents](transformers_agents) page first.
</Tip>
<Tip warning={true}>
Transformers Agent is an experimental API that is subject to change at any time. Results returned by the agents
can vary as the APIs or underlying models are prone to change.
</Tip>
Creating and using custom tools and prompts is paramount to empowering the agent and having it perform new tasks.
In this guide we'll take a look at:
- How to customize the prompt
- How to use custom tools
- How to create custom tools
## Customizing the prompt
As explained in [Transformers Agents](transformers_agents) agents can run in [`~Agent.run`] and [`~Agent.chat`] mode.
Both the `run` and `chat` modes underlie the same logic. The language model powering the agent is conditioned on a long
prompt and completes the prompt by generating the next tokens until the stop token is reached.
The only difference between the two modes is that during the `chat` mode the prompt is extended with
previous user inputs and model generations. This allows the agent to have access to past interactions,
seemingly giving the agent some kind of memory.
### Structure of the prompt
Let's take a closer look at how the prompt is structured to understand how it can be best customized.
The prompt is structured broadly into four parts.
- 1. Introduction: how the agent should behave, explanation of the concept of tools.
- 2. Description of all the tools. This is defined by a `<<all_tools>>` token that is dynamically replaced at runtime with the tools defined/chosen by the user.
- 3. A set of examples of tasks and their solution
- 4. Current example, and request for solution.
To better understand each part, let's look at a shortened version of how the `run` prompt can look like:
````text
I will ask you to perform a task, your job is to come up with a series of simple commands in Python that will perform the task.
[...]
You can print intermediate results if it makes sense to do so.
Tools:
- document_qa: This is a tool that answers a question about a document (pdf). It takes an input named `document` which should be the document containing the information, as well as a `question` that is the question about the document. It returns a text that contains the answer to the question.
- image_captioner: This is a tool that generates a description of an image. It takes an input named `image` which should be the image to the caption and returns a text that contains the description in English.
[...]
Task: "Answer the question in the variable `question` about the image stored in the variable `image`. The question is in French."
I will use the following tools: `translator` to translate the question into English and then `image_qa` to answer the question on the input image.
Answer:
```py
translated_question = translator(question=question, src_lang="French", tgt_lang="English")
print(f"The translated question is {translated_question}.")
answer = image_qa(image=image, question=translated_question)
print(f"The answer is {answer}")
```
Task: "Identify the oldest person in the `document` and create an image showcasing the result as a banner."
I will use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Answer:
```py
answer = document_qa(document, question="What is the oldest person?")
print(f"The answer is {answer}.")
image = image_generator("A banner showing " + answer)
```
[...]
Task: "Draw me a picture of rivers and lakes"
I will use the following
````
The introduction (the text before *"Tools:"*) explains precisely how the model shall behave and what it should do.
This part most likely does not need to be customized as the agent shall always behave the same way.
The second part (the bullet points below *"Tools"*) is dynamically added upon calling `run` or `chat`. There are
exactly as many bullet points as there are tools in `agent.toolbox` and each bullet point consists of the name
and description of the tool:
```text
- <tool.name>: <tool.description>
```
Let's verify this quickly by loading the document_qa tool and printing out the name and description.
```py
from transformers import load_tool
document_qa = load_tool("document-question-answering")
print(f"- {document_qa.name}: {document_qa.description}")
```
which gives:
```text
- document_qa: This is a tool that answers a question about a document (pdf). It takes an input named `document` which should be the document containing the information, as well as a `question` that is the question about the document. It returns a text that contains the answer to the question.
```
We can see that the tool name is short and precise. The description includes two parts, the first explaining
what the tool does and the second states what input arguments and return values are expected.
A good tool name and tool description are very important for the agent to correctly use it. Note that the only
information the agent has about the tool is its name and description, so one should make sure that both
are precisely written and match the style of the existing tools in the toolbox. In particular make sure the description
mentions all the arguments expected by name in code-style, along with the expected type and a description of what they
are.
<Tip>
Check the naming and description of the curated Transformers tools to better understand what name and
description a tool is expected to have. You can see all tools with the [`Agent.toolbox`] property.
</Tip>
The third part includes a set of curated examples that show the agent exactly what code it should produce
for what kind of user request. The large language models empowering the agent are extremely good at
recognizing patterns in a prompt and repeating the pattern with new data. Therefore, it is very important
that the examples are written in a way that maximizes the likelihood of the agent to generating correct,
executable code in practice.
Let's have a look at one example:
````text
Task: "Identify the oldest person in the `document` and create an image showcasing the result as a banner."
I will use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Answer:
```py
answer = document_qa(document, question="What is the oldest person?")
print(f"The answer is {answer}.")
image = image_generator("A banner showing " + answer)
```
````
The pattern the model is prompted to repeat has three parts: The task statement, the agent's explanation of
what it intends to do, and finally the generated code. Every example that is part of the prompt has this exact
pattern, thus making sure that the agent will reproduce exactly the same pattern when generating new tokens.
The prompt examples are curated by the Transformers team and rigorously evaluated on a set of
[problem statements](https://github.com/huggingface/transformers/blob/main/src/transformers/tools/evaluate_agent.py)
to ensure that the agent's prompt is as good as possible to solve real use cases of the agent.
The final part of the prompt corresponds to:
```text
Task: "Draw me a picture of rivers and lakes"
I will use the following
```
is a final and unfinished example that the agent is tasked to complete. The unfinished example
is dynamically created based on the actual user input. For the above example, the user ran:
```py
agent.run("Draw me a picture of rivers and lakes")
```
The user input - *a.k.a* the task: *"Draw me a picture of rivers and lakes"* is cast into the
prompt template: "Task: <task> \n\n I will use the following". This sentence makes up the final lines of the
prompt the agent is conditioned on, therefore strongly influencing the agent to finish the example
exactly in the same way it was previously done in the examples.
Without going into too much detail, the chat template has the same prompt structure with the
examples having a slightly different style, *e.g.*:
````text
[...]
=====
Human: Answer the question in the variable `question` about the image stored in the variable `image`.
Assistant: I will use the tool `image_qa` to answer the question on the input image.
```py
answer = image_qa(text=question, image=image)
print(f"The answer is {answer}")
```
Human: I tried this code, it worked but didn't give me a good result. The question is in French
Assistant: In this case, the question needs to be translated first. I will use the tool `translator` to do this.
```py
translated_question = translator(question=question, src_lang="French", tgt_lang="English")
print(f"The translated question is {translated_question}.")
answer = image_qa(text=translated_question, image=image)
print(f"The answer is {answer}")
```
=====
[...]
````
Contrary, to the examples of the `run` prompt, each `chat` prompt example has one or more exchanges between the
*Human* and the *Assistant*. Every exchange is structured similarly to the example of the `run` prompt.
The user's input is appended to behind *Human:* and the agent is prompted to first generate what needs to be done
before generating code. An exchange can be based on previous exchanges, therefore allowing the user to refer
to past exchanges as is done *e.g.* above by the user's input of "I tried **this** code" refers to the
previously generated code of the agent.
Upon running `.chat`, the user's input or *task* is cast into an unfinished example of the form:
```text
Human: <user-input>\n\nAssistant:
```
which the agent completes. Contrary to the `run` command, the `chat` command then appends the completed example
to the prompt, thus giving the agent more context for the next `chat` turn.
Great now that we know how the prompt is structured, let's see how we can customize it!
### Writing good user inputs
While large language models are getting better and better at understanding users' intentions, it helps
enormously to be as precise as possible to help the agent pick the correct task. What does it mean to be
as precise as possible?
The agent sees a list of tool names and their description in its prompt. The more tools are added the
more difficult it becomes for the agent to choose the correct tool and it's even more difficult to choose
the correct sequences of tools to run. Let's look at a common failure case, here we will only return
the code to analyze it.
```py
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
agent.run("Show me a tree", return_code=True)
```
gives:
```text
==Explanation from the agent==
I will use the following tool: `image_segmenter` to create a segmentation mask for the image.
==Code generated by the agent==
mask = image_segmenter(image, prompt="tree")
```
which is probably not what we wanted. Instead, it is more likely that we want an image of a tree to be generated.
To steer the agent more towards using a specific tool it can therefore be very helpful to use important keywords that
are present in the tool's name and description. Let's have a look.
```py
agent.toolbox["image_generator"].description
```
```text
'This is a tool that creates an image according to a prompt, which is a text description. It takes an input named `prompt` which contains the image description and outputs an image.
```
The name and description make use of the keywords "image", "prompt", "create" and "generate". Using these words will most likely work better here. Let's refine our prompt a bit.
```py
agent.run("Create an image of a tree", return_code=True)
```
gives:
```text
==Explanation from the agent==
I will use the following tool `image_generator` to generate an image of a tree.
==Code generated by the agent==
image = image_generator(prompt="tree")
```
Much better! That looks more like what we want. In short, when you notice that the agent struggles to
correctly map your task to the correct tools, try looking up the most pertinent keywords of the tool's name
and description and try refining your task request with it.
### Customizing the tool descriptions
As we've seen before the agent has access to each of the tools' names and descriptions. The base tools
should have very precise names and descriptions, however, you might find that it could help to change the
the description or name of a tool for your specific use case. This might become especially important
when you've added multiple tools that are very similar or if you want to use your agent only for a certain
domain, *e.g.* image generation and transformations.
A common problem is that the agent confuses image generation with image transformation/modification when
used a lot for image generation tasks, *e.g.*
```py
agent.run("Make an image of a house and a car", return_code=True)
```
returns
```text
==Explanation from the agent==
I will use the following tools `image_generator` to generate an image of a house and `image_transformer` to transform the image of a car into the image of a house.
==Code generated by the agent==
house_image = image_generator(prompt="A house")
car_image = image_generator(prompt="A car")
house_car_image = image_transformer(image=car_image, prompt="A house")
```
which is probably not exactly what we want here. It seems like the agent has a difficult time
to understand the difference between `image_generator` and `image_transformer` and often uses the two together.
We can help the agent here by changing the tool name and description of `image_transformer`. Let's instead call it `modifier`
to disassociate it a bit from "image" and "prompt":
```py
agent.toolbox["modifier"] = agent.toolbox.pop("image_transformer")
agent.toolbox["modifier"].description = agent.toolbox["modifier"].description.replace(
"transforms an image according to a prompt", "modifies an image"
)
```
Now "modify" is a strong cue to use the new image processor which should help with the above prompt. Let's run it again.
```py
agent.run("Make an image of a house and a car", return_code=True)
```
Now we're getting:
```text
==Explanation from the agent==
I will use the following tools: `image_generator` to generate an image of a house, then `image_generator` to generate an image of a car.
==Code generated by the agent==
house_image = image_generator(prompt="A house")
car_image = image_generator(prompt="A car")
```
which is definitely closer to what we had in mind! However, we want to have both the house and car in the same image. Steering the task more toward single image generation should help:
```py
agent.run("Create image: 'A house and car'", return_code=True)
```
```text
==Explanation from the agent==
I will use the following tool: `image_generator` to generate an image.
==Code generated by the agent==
image = image_generator(prompt="A house and car")
```
<Tip warning={true}>
Agents are still brittle for many use cases, especially when it comes to
slightly more complex use cases like generating an image of multiple objects.
Both the agent itself and the underlying prompt will be further improved in the coming
months making sure that agents become more robust to a variety of user inputs.
</Tip>
### Customizing the whole prompt
To give the user maximum flexibility, the whole prompt template as explained in [above](#structure-of-the-prompt)
can be overwritten by the user. In this case make sure that your custom prompt includes an introduction section,
a tool section, an example section, and an unfinished example section. If you want to overwrite the `run` prompt template,
you can do as follows:
```py
template = """ [...] """
agent = HfAgent(your_endpoint, run_prompt_template=template)
```
<Tip warning={true}>
Please make sure to have the `<<all_tools>>` string and the `<<prompt>>` defined somewhere in the `template` so that the agent can be aware
of the tools, it has available to it as well as correctly insert the user's prompt.
</Tip>
Similarly, one can overwrite the `chat` prompt template. Note that the `chat` mode always uses the following format for the exchanges:
```text
Human: <<task>>
Assistant:
```
Therefore it is important that the examples of the custom `chat` prompt template also make use of this format.
You can overwrite the `chat` template at instantiation as follows.
```
template = """ [...] """
agent = HfAgent(url_endpoint=your_endpoint, chat_prompt_template=template)
```
<Tip warning={true}>
Please make sure to have the `<<all_tools>>` string defined somewhere in the `template` so that the agent can be aware
of the tools, it has available to it.
</Tip>
In both cases, you can pass a repo ID instead of the prompt template if you would like to use a template hosted by someone in the community. The default prompts live in [this repo](https://huggingface.co/datasets/huggingface-tools/default-prompts) as an example.
To upload your custom prompt on a repo on the Hub and share it with the community just make sure:
- to use a dataset repository
- to put the prompt template for the `run` command in a file named `run_prompt_template.txt`
- to put the prompt template for the `chat` command in a file named `chat_prompt_template.txt`
## Using custom tools
In this section, we'll be leveraging two existing custom tools that are specific to image generation:
- We replace [huggingface-tools/image-transformation](https://huggingface.co/spaces/huggingface-tools/image-transformation),
with [diffusers/controlnet-canny-tool](https://huggingface.co/spaces/diffusers/controlnet-canny-tool)
to allow for more image modifications.
- We add a new tool for image upscaling to the default toolbox:
[diffusers/latent-upscaler-tool](https://huggingface.co/spaces/diffusers/latent-upscaler-tool) replace the existing image-transformation tool.
We'll start by loading the custom tools with the convenient [`load_tool`] function:
```py
from transformers import load_tool
controlnet_transformer = load_tool("diffusers/controlnet-canny-tool")
upscaler = load_tool("diffusers/latent-upscaler-tool")
```
Upon adding custom tools to an agent, the tools' descriptions and names are automatically
included in the agents' prompts. Thus, it is imperative that custom tools have
a well-written description and name in order for the agent to understand how to use them.
Let's take a look at the description and name of `controlnet_transformer`:
```py
print(f"Description: '{controlnet_transformer.description}'")
print(f"Name: '{controlnet_transformer.name}'")
```
gives
```text
Description: 'This is a tool that transforms an image with ControlNet according to a prompt.
It takes two inputs: `image`, which should be the image to transform, and `prompt`, which should be the prompt to use to change it. It returns the modified image.'
Name: 'image_transformer'
```
The name and description are accurate and fit the style of the [curated set of tools](./transformers_agents#a-curated-set-of-tools).
Next, let's instantiate an agent with `controlnet_transformer` and `upscaler`:
```py
tools = [controlnet_transformer, upscaler]
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=tools)
```
This command should give you the following info:
```text
image_transformer has been replaced by <transformers_modules.diffusers.controlnet-canny-tool.bd76182c7777eba9612fc03c0
8718a60c0aa6312.image_transformation.ControlNetTransformationTool object at 0x7f1d3bfa3a00> as provided in `additional_tools`
```
The set of curated tools already has an `image_transformer` tool which is hereby replaced with our custom tool.
<Tip>
Overwriting existing tools can be beneficial if we want to use a custom tool exactly for the same task as an existing tool
because the agent is well-versed in using the specific task. Beware that the custom tool should follow the exact same API
as the overwritten tool in this case, or you should adapt the prompt template to make sure all examples using that
tool are updated.
</Tip>
The upscaler tool was given the name `image_upscaler` which is not yet present in the default toolbox and is therefore simply added to the list of tools.
You can always have a look at the toolbox that is currently available to the agent via the `agent.toolbox` attribute:
```py
print("\n".join([f"- {a}" for a in agent.toolbox.keys()]))
```
```text
- document_qa
- image_captioner
- image_qa
- image_segmenter
- transcriber
- summarizer
- text_classifier
- text_qa
- text_reader
- translator
- image_transformer
- text_downloader
- image_generator
- video_generator
- image_upscaler
```
Note how `image_upscaler` is now part of the agents' toolbox.
Let's now try out the new tools! We will re-use the image we generated in [Transformers Agents Quickstart](./transformers_agents#single-execution-run).
```py
from diffusers.utils import load_image
image = load_image(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png"
)
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
Let's transform the image into a beautiful winter landscape:
```py
image = agent.run("Transform the image: 'A frozen lake and snowy forest'", image=image)
```
```text
==Explanation from the agent==
I will use the following tool: `image_transformer` to transform the image.
==Code generated by the agent==
image = image_transformer(image, prompt="A frozen lake and snowy forest")
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_winter.png" width=200>
The new image processing tool is based on ControlNet which can make very strong modifications to the image.
By default the image processing tool returns an image of size 512x512 pixels. Let's see if we can upscale it.
```py
image = agent.run("Upscale the image", image)
```
```text
==Explanation from the agent==
I will use the following tool: `image_upscaler` to upscale the image.
==Code generated by the agent==
upscaled_image = image_upscaler(image)
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_winter_upscale.png" width=400>
The agent automatically mapped our prompt "Upscale the image" to the just added upscaler tool purely based on the description and name of the upscaler tool
and was able to correctly run it.
Next, let's have a look at how you can create a new custom tool.
### Adding new tools
In this section, we show how to create a new tool that can be added to the agent.
#### Creating a new tool
We'll first start by creating a tool. We'll add the not-so-useful yet fun task of fetching the model on the Hugging Face
Hub with the most downloads for a given task.
We can do that with the following code:
```python
from huggingface_hub import list_models
task = "text-classification"
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
print(model.id)
```
For the task `text-classification`, this returns `'facebook/bart-large-mnli'`, for `translation` it returns `'t5-base`.
How do we convert this to a tool that the agent can leverage? All tools depend on the superclass `Tool` that holds the
main attributes necessary. We'll create a class that inherits from it:
```python
from transformers import Tool
class HFModelDownloadsTool(Tool):
pass
```
This class has a few needs:
- An attribute `name`, which corresponds to the name of the tool itself. To be in tune with other tools which have a
performative name, we'll name it `model_download_counter`.
- An attribute `description`, which will be used to populate the prompt of the agent.
- `inputs` and `outputs` attributes. Defining this will help the python interpreter make educated choices about types,
and will allow for a gradio-demo to be spawned when we push our tool to the Hub. They're both a list of expected
values, which can be `text`, `image`, or `audio`.
- A `__call__` method which contains the inference code. This is the code we've played with above!
Here's what our class looks like now:
```python
from transformers import Tool
from huggingface_hub import list_models
class HFModelDownloadsTool(Tool):
name = "model_download_counter"
description = (
"This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub. "
"It takes the name of the category (such as text-classification, depth-estimation, etc), and "
"returns the name of the checkpoint."
)
inputs = ["text"]
outputs = ["text"]
def __call__(self, task: str):
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
return model.id
```
We now have our tool handy. Save it in a file and import it from your main script. Let's name this file
`model_downloads.py`, so the resulting import code looks like this:
```python
from model_downloads import HFModelDownloadsTool
tool = HFModelDownloadsTool()
```
In order to let others benefit from it and for simpler initialization, we recommend pushing it to the Hub under your
namespace. To do so, just call `push_to_hub` on the `tool` variable:
```python
tool.push_to_hub("hf-model-downloads")
```
You now have your code on the Hub! Let's take a look at the final step, which is to have the agent use it.
#### Having the agent use the tool
We now have our tool that lives on the Hub which can be instantiated as such (change the user name for your tool):
```python
from transformers import load_tool
tool = load_tool("lysandre/hf-model-downloads")
```
In order to use it in the agent, simply pass it in the `additional_tools` parameter of the agent initialization method:
```python
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=[tool])
agent.run(
"Can you read out loud the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?"
)
```
which outputs the following:
```text
==Code generated by the agent==
model = model_download_counter(task="text-to-video")
print(f"The model with the most downloads is {model}.")
audio_model = text_reader(model)
==Result==
The model with the most downloads is damo-vilab/text-to-video-ms-1.7b.
```
and generates the following audio.
| **Audio** |
|------------------------------------------------------------------------------------------------------------------------------------------------------|
| <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/damo.wav" type="audio/wav"/> |
<Tip>
Depending on the LLM, some are quite brittle and require very exact prompts in order to work well. Having a well-defined
name and description of the tool is paramount to having it be leveraged by the agent.
</Tip>
### Replacing existing tools
Replacing existing tools can be done simply by assigning a new item to the agent's toolbox. Here's how one would do so:
```python
from transformers import HfAgent, load_tool
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
agent.toolbox["image-transformation"] = load_tool("diffusers/controlnet-canny-tool")
```
<Tip>
Beware when replacing tools with others! This will also adjust the agent's prompt. This can be good if you have a better
prompt suited for the task, but it can also result in your tool being selected way more than others or for other
tools to be selected instead of the one you have defined.
</Tip>
## Leveraging gradio-tools
[gradio-tools](https://github.com/freddyaboulton/gradio-tools) is a powerful library that allows using Hugging
Face Spaces as tools. It supports many existing Spaces as well as custom Spaces to be designed with it.
We offer support for `gradio_tools` by using the `Tool.from_gradio` method. For example, we want to take
advantage of the `StableDiffusionPromptGeneratorTool` tool offered in the `gradio-tools` toolkit so as to
improve our prompts and generate better images.
We first import the tool from `gradio_tools` and instantiate it:
```python
from gradio_tools import StableDiffusionPromptGeneratorTool
gradio_tool = StableDiffusionPromptGeneratorTool()
```
We pass that instance to the `Tool.from_gradio` method:
```python
from transformers import Tool
tool = Tool.from_gradio(gradio_tool)
```
Now we can manage it exactly as we would a usual custom tool. We leverage it to improve our prompt
` a rabbit wearing a space suit`:
```python
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=[tool])
agent.run("Generate an image of the `prompt` after improving it.", prompt="A rabbit wearing a space suit")
```
The model adequately leverages the tool:
```text
==Explanation from the agent==
I will use the following tools: `StableDiffusionPromptGenerator` to improve the prompt, then `image_generator` to generate an image according to the improved prompt.
==Code generated by the agent==
improved_prompt = StableDiffusionPromptGenerator(prompt)
print(f"The improved prompt is {improved_prompt}.")
image = image_generator(improved_prompt)
```
Before finally generating the image:
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png">
<Tip warning={true}>
gradio-tools requires *textual* inputs and outputs, even when working with different modalities. This implementation
works with image and audio objects. The two are currently incompatible, but will rapidly become compatible as we
work to improve the support.
</Tip>
## Future compatibility with Langchain
We love Langchain and think it has a very compelling suite of tools. In order to handle these tools,
Langchain requires *textual* inputs and outputs, even when working with different modalities.
This is often the serialized version (i.e., saved to disk) of the objects.
This difference means that multi-modality isn't handled between transformers-agents and langchain.
We aim for this limitation to be resolved in future versions, and welcome any help from avid langchain
users to help us achieve this compatibility.
We would love to have better support. If you would like to help, please
[open an issue](https://github.com/huggingface/transformers/issues/new) and share what you have in mind.
docs/source/en/custom_tools.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Custom Tools and Prompts
<Tip>
If you are not aware of what tools and agents are in the context of transformers, we recommend you read the
[Transformers Agents](transformers_agents) page first.
</Tip>
<Tip warning={true}>
Transformers Agent is an experimental API that is subject to change at any time. Results returned by the agents
can vary as the APIs or underlying models are prone to change.
</Tip>
Creating and using custom tools and prompts is paramount to empowering the agent and having it perform new tasks.
In this guide we'll take a look at:
- How to customize the prompt
- How to use custom tools
- How to create custom tools
## Customizing the prompt
As explained in [Transformers Agents](transformers_agents) agents can run in [`~Agent.run`] and [`~Agent.chat`] mode.
Both the `run` and `chat` modes underlie the same logic. The language model powering the agent is conditioned on a long
prompt and completes the prompt by generating the next tokens until the stop token is reached.
The only difference between the two modes is that during the `chat` mode the prompt is extended with
previous user inputs and model generations. This allows the agent to have access to past interactions,
seemingly giving the agent some kind of memory.
### Structure of the prompt
Let's take a closer look at how the prompt is structured to understand how it can be best customized.
The prompt is structured broadly into four parts.
- 1. Introduction: how the agent should behave, explanation of the concept of tools.
- 2. Description of all the tools. This is defined by a `<<all_tools>>` token that is dynamically replaced at runtime with the tools defined/chosen by the user.
- 3. A set of examples of tasks and their solution
- 4. Current example, and request for solution.
To better understand each part, let's look at a shortened version of how the `run` prompt can look like:
````text
I will ask you to perform a task, your job is to come up with a series of simple commands in Python that will perform the task.
[...]
You can print intermediate results if it makes sense to do so.
Tools:
- document_qa: This is a tool that answers a question about a document (pdf). It takes an input named `document` which should be the document containing the information, as well as a `question` that is the question about the document. It returns a text that contains the answer to the question.
- image_captioner: This is a tool that generates a description of an image. It takes an input named `image` which should be the image to the caption and returns a text that contains the description in English.
[...]
Task: "Answer the question in the variable `question` about the image stored in the variable `image`. The question is in French."
I will use the following tools: `translator` to translate the question into English and then `image_qa` to answer the question on the input image.
Answer:
```py
translated_question = translator(question=question, src_lang="French", tgt_lang="English")
print(f"The translated question is {translated_question}.")
answer = image_qa(image=image, question=translated_question)
print(f"The answer is {answer}")
```
Task: "Identify the oldest person in the `document` and create an image showcasing the result as a banner."
I will use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Answer:
```py
answer = document_qa(document, question="What is the oldest person?")
print(f"The answer is {answer}.")
image = image_generator("A banner showing " + answer)
```
[...]
Task: "Draw me a picture of rivers and lakes"
I will use the following
````
The introduction (the text before *"Tools:"*) explains precisely how the model shall behave and what it should do.
This part most likely does not need to be customized as the agent shall always behave the same way.
The second part (the bullet points below *"Tools"*) is dynamically added upon calling `run` or `chat`. There are
exactly as many bullet points as there are tools in `agent.toolbox` and each bullet point consists of the name
and description of the tool:
```text
- <tool.name>: <tool.description>
```
Let's verify this quickly by loading the document_qa tool and printing out the name and description.
```py
from transformers import load_tool
document_qa = load_tool("document-question-answering")
print(f"- {document_qa.name}: {document_qa.description}")
```
which gives:
```text
- document_qa: This is a tool that answers a question about a document (pdf). It takes an input named `document` which should be the document containing the information, as well as a `question` that is the question about the document. It returns a text that contains the answer to the question.
```
We can see that the tool name is short and precise. The description includes two parts, the first explaining
what the tool does and the second states what input arguments and return values are expected.
A good tool name and tool description are very important for the agent to correctly use it. Note that the only
information the agent has about the tool is its name and description, so one should make sure that both
are precisely written and match the style of the existing tools in the toolbox. In particular make sure the description
mentions all the arguments expected by name in code-style, along with the expected type and a description of what they
are.
<Tip>
Check the naming and description of the curated Transformers tools to better understand what name and
description a tool is expected to have. You can see all tools with the [`Agent.toolbox`] property.
</Tip>
The third part includes a set of curated examples that show the agent exactly what code it should produce
for what kind of user request. The large language models empowering the agent are extremely good at
recognizing patterns in a prompt and repeating the pattern with new data. Therefore, it is very important
that the examples are written in a way that maximizes the likelihood of the agent to generating correct,
executable code in practice.
Let's have a look at one example:
````text
Task: "Identify the oldest person in the `document` and create an image showcasing the result as a banner."
I will use the following tools: `document_qa` to find the oldest person in the document, then `image_generator` to generate an image according to the answer.
Answer:
```py
answer = document_qa(document, question="What is the oldest person?")
print(f"The answer is {answer}.")
image = image_generator("A banner showing " + answer)
```
````
The pattern the model is prompted to repeat has three parts: The task statement, the agent's explanation of
what it intends to do, and finally the generated code. Every example that is part of the prompt has this exact
pattern, thus making sure that the agent will reproduce exactly the same pattern when generating new tokens.
The prompt examples are curated by the Transformers team and rigorously evaluated on a set of
[problem statements](https://github.com/huggingface/transformers/blob/main/src/transformers/tools/evaluate_agent.py)
to ensure that the agent's prompt is as good as possible to solve real use cases of the agent.
The final part of the prompt corresponds to:
```text
Task: "Draw me a picture of rivers and lakes"
I will use the following
```
is a final and unfinished example that the agent is tasked to complete. The unfinished example
is dynamically created based on the actual user input. For the above example, the user ran:
```py
agent.run("Draw me a picture of rivers and lakes")
```
The user input - *a.k.a* the task: *"Draw me a picture of rivers and lakes"* is cast into the
prompt template: "Task: <task> \n\n I will use the following". This sentence makes up the final lines of the
prompt the agent is conditioned on, therefore strongly influencing the agent to finish the example
exactly in the same way it was previously done in the examples.
Without going into too much detail, the chat template has the same prompt structure with the
examples having a slightly different style, *e.g.*:
````text
[...]
=====
Human: Answer the question in the variable `question` about the image stored in the variable `image`.
Assistant: I will use the tool `image_qa` to answer the question on the input image.
```py
answer = image_qa(text=question, image=image)
print(f"The answer is {answer}")
```
Human: I tried this code, it worked but didn't give me a good result. The question is in French
Assistant: In this case, the question needs to be translated first. I will use the tool `translator` to do this.
```py
translated_question = translator(question=question, src_lang="French", tgt_lang="English")
print(f"The translated question is {translated_question}.")
answer = image_qa(text=translated_question, image=image)
print(f"The answer is {answer}")
```
=====
[...]
````
Contrary, to the examples of the `run` prompt, each `chat` prompt example has one or more exchanges between the
*Human* and the *Assistant*. Every exchange is structured similarly to the example of the `run` prompt.
The user's input is appended to behind *Human:* and the agent is prompted to first generate what needs to be done
before generating code. An exchange can be based on previous exchanges, therefore allowing the user to refer
to past exchanges as is done *e.g.* above by the user's input of "I tried **this** code" refers to the
previously generated code of the agent.
Upon running `.chat`, the user's input or *task* is cast into an unfinished example of the form:
```text
Human: <user-input>\n\nAssistant:
```
which the agent completes. Contrary to the `run` command, the `chat` command then appends the completed example
to the prompt, thus giving the agent more context for the next `chat` turn.
Great now that we know how the prompt is structured, let's see how we can customize it!
### Writing good user inputs
While large language models are getting better and better at understanding users' intentions, it helps
enormously to be as precise as possible to help the agent pick the correct task. What does it mean to be
as precise as possible?
The agent sees a list of tool names and their description in its prompt. The more tools are added the
more difficult it becomes for the agent to choose the correct tool and it's even more difficult to choose
the correct sequences of tools to run. Let's look at a common failure case, here we will only return
the code to analyze it.
```py
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
agent.run("Show me a tree", return_code=True)
```
gives:
```text
==Explanation from the agent==
I will use the following tool: `image_segmenter` to create a segmentation mask for the image.
==Code generated by the agent==
mask = image_segmenter(image, prompt="tree")
```
which is probably not what we wanted. Instead, it is more likely that we want an image of a tree to be generated.
To steer the agent more towards using a specific tool it can therefore be very helpful to use important keywords that
are present in the tool's name and description. Let's have a look.
```py
agent.toolbox["image_generator"].description
```
```text
'This is a tool that creates an image according to a prompt, which is a text description. It takes an input named `prompt` which contains the image description and outputs an image.
```
The name and description make use of the keywords "image", "prompt", "create" and "generate". Using these words will most likely work better here. Let's refine our prompt a bit.
```py
agent.run("Create an image of a tree", return_code=True)
```
gives:
```text
==Explanation from the agent==
I will use the following tool `image_generator` to generate an image of a tree.
==Code generated by the agent==
image = image_generator(prompt="tree")
```
Much better! That looks more like what we want. In short, when you notice that the agent struggles to
correctly map your task to the correct tools, try looking up the most pertinent keywords of the tool's name
and description and try refining your task request with it.
### Customizing the tool descriptions
As we've seen before the agent has access to each of the tools' names and descriptions. The base tools
should have very precise names and descriptions, however, you might find that it could help to change the
the description or name of a tool for your specific use case. This might become especially important
when you've added multiple tools that are very similar or if you want to use your agent only for a certain
domain, *e.g.* image generation and transformations.
A common problem is that the agent confuses image generation with image transformation/modification when
used a lot for image generation tasks, *e.g.*
```py
agent.run("Make an image of a house and a car", return_code=True)
```
returns
```text
==Explanation from the agent==
I will use the following tools `image_generator` to generate an image of a house and `image_transformer` to transform the image of a car into the image of a house.
==Code generated by the agent==
house_image = image_generator(prompt="A house")
car_image = image_generator(prompt="A car")
house_car_image = image_transformer(image=car_image, prompt="A house")
```
which is probably not exactly what we want here. It seems like the agent has a difficult time
to understand the difference between `image_generator` and `image_transformer` and often uses the two together.
We can help the agent here by changing the tool name and description of `image_transformer`. Let's instead call it `modifier`
to disassociate it a bit from "image" and "prompt":
```py
agent.toolbox["modifier"] = agent.toolbox.pop("image_transformer")
agent.toolbox["modifier"].description = agent.toolbox["modifier"].description.replace(
"transforms an image according to a prompt", "modifies an image"
)
```
Now "modify" is a strong cue to use the new image processor which should help with the above prompt. Let's run it again.
```py
agent.run("Make an image of a house and a car", return_code=True)
```
Now we're getting:
```text
==Explanation from the agent==
I will use the following tools: `image_generator` to generate an image of a house, then `image_generator` to generate an image of a car.
==Code generated by the agent==
house_image = image_generator(prompt="A house")
car_image = image_generator(prompt="A car")
```
which is definitely closer to what we had in mind! However, we want to have both the house and car in the same image. Steering the task more toward single image generation should help:
```py
agent.run("Create image: 'A house and car'", return_code=True)
```
```text
==Explanation from the agent==
I will use the following tool: `image_generator` to generate an image.
==Code generated by the agent==
image = image_generator(prompt="A house and car")
```
<Tip warning={true}>
Agents are still brittle for many use cases, especially when it comes to
slightly more complex use cases like generating an image of multiple objects.
Both the agent itself and the underlying prompt will be further improved in the coming
months making sure that agents become more robust to a variety of user inputs.
</Tip>
### Customizing the whole prompt
To give the user maximum flexibility, the whole prompt template as explained in [above](#structure-of-the-prompt)
can be overwritten by the user. In this case make sure that your custom prompt includes an introduction section,
a tool section, an example section, and an unfinished example section. If you want to overwrite the `run` prompt template,
you can do as follows:
```py
template = """ [...] """
agent = HfAgent(your_endpoint, run_prompt_template=template)
```
<Tip warning={true}>
Please make sure to have the `<<all_tools>>` string and the `<<prompt>>` defined somewhere in the `template` so that the agent can be aware
of the tools, it has available to it as well as correctly insert the user's prompt.
</Tip>
Similarly, one can overwrite the `chat` prompt template. Note that the `chat` mode always uses the following format for the exchanges:
```text
Human: <<task>>
Assistant:
```
Therefore it is important that the examples of the custom `chat` prompt template also make use of this format.
You can overwrite the `chat` template at instantiation as follows.
```
template = """ [...] """
agent = HfAgent(url_endpoint=your_endpoint, chat_prompt_template=template)
```
<Tip warning={true}>
Please make sure to have the `<<all_tools>>` string defined somewhere in the `template` so that the agent can be aware
of the tools, it has available to it.
</Tip>
In both cases, you can pass a repo ID instead of the prompt template if you would like to use a template hosted by someone in the community. The default prompts live in [this repo](https://huggingface.co/datasets/huggingface-tools/default-prompts) as an example.
To upload your custom prompt on a repo on the Hub and share it with the community just make sure:
- to use a dataset repository
- to put the prompt template for the `run` command in a file named `run_prompt_template.txt`
- to put the prompt template for the `chat` command in a file named `chat_prompt_template.txt`
## Using custom tools
In this section, we'll be leveraging two existing custom tools that are specific to image generation:
- We replace [huggingface-tools/image-transformation](https://huggingface.co/spaces/huggingface-tools/image-transformation),
with [diffusers/controlnet-canny-tool](https://huggingface.co/spaces/diffusers/controlnet-canny-tool)
to allow for more image modifications.
- We add a new tool for image upscaling to the default toolbox:
[diffusers/latent-upscaler-tool](https://huggingface.co/spaces/diffusers/latent-upscaler-tool) replace the existing image-transformation tool.
We'll start by loading the custom tools with the convenient [`load_tool`] function:
```py
from transformers import load_tool
controlnet_transformer = load_tool("diffusers/controlnet-canny-tool")
upscaler = load_tool("diffusers/latent-upscaler-tool")
```
Upon adding custom tools to an agent, the tools' descriptions and names are automatically
included in the agents' prompts. Thus, it is imperative that custom tools have
a well-written description and name in order for the agent to understand how to use them.
Let's take a look at the description and name of `controlnet_transformer`:
```py
print(f"Description: '{controlnet_transformer.description}'")
print(f"Name: '{controlnet_transformer.name}'")
```
gives
```text
Description: 'This is a tool that transforms an image with ControlNet according to a prompt.
It takes two inputs: `image`, which should be the image to transform, and `prompt`, which should be the prompt to use to change it. It returns the modified image.'
Name: 'image_transformer'
```
The name and description are accurate and fit the style of the [curated set of tools](./transformers_agents#a-curated-set-of-tools).
Next, let's instantiate an agent with `controlnet_transformer` and `upscaler`:
```py
tools = [controlnet_transformer, upscaler]
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=tools)
```
This command should give you the following info:
```text
image_transformer has been replaced by <transformers_modules.diffusers.controlnet-canny-tool.bd76182c7777eba9612fc03c0
8718a60c0aa6312.image_transformation.ControlNetTransformationTool object at 0x7f1d3bfa3a00> as provided in `additional_tools`
```
The set of curated tools already has an `image_transformer` tool which is hereby replaced with our custom tool.
<Tip>
Overwriting existing tools can be beneficial if we want to use a custom tool exactly for the same task as an existing tool
because the agent is well-versed in using the specific task. Beware that the custom tool should follow the exact same API
as the overwritten tool in this case, or you should adapt the prompt template to make sure all examples using that
tool are updated.
</Tip>
The upscaler tool was given the name `image_upscaler` which is not yet present in the default toolbox and is therefore simply added to the list of tools.
You can always have a look at the toolbox that is currently available to the agent via the `agent.toolbox` attribute:
```py
print("\n".join([f"- {a}" for a in agent.toolbox.keys()]))
```
```text
- document_qa
- image_captioner
- image_qa
- image_segmenter
- transcriber
- summarizer
- text_classifier
- text_qa
- text_reader
- translator
- image_transformer
- text_downloader
- image_generator
- video_generator
- image_upscaler
```
Note how `image_upscaler` is now part of the agents' toolbox.
Let's now try out the new tools! We will re-use the image we generated in [Transformers Agents Quickstart](./transformers_agents#single-execution-run).
```py
from diffusers.utils import load_image
image = load_image(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png"
)
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
Let's transform the image into a beautiful winter landscape:
```py
image = agent.run("Transform the image: 'A frozen lake and snowy forest'", image=image)
```
```text
==Explanation from the agent==
I will use the following tool: `image_transformer` to transform the image.
==Code generated by the agent==
image = image_transformer(image, prompt="A frozen lake and snowy forest")
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_winter.png" width=200>
The new image processing tool is based on ControlNet which can make very strong modifications to the image.
By default the image processing tool returns an image of size 512x512 pixels. Let's see if we can upscale it.
```py
image = agent.run("Upscale the image", image)
```
```text
==Explanation from the agent==
I will use the following tool: `image_upscaler` to upscale the image.
==Code generated by the agent==
upscaled_image = image_upscaler(image)
```
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_winter_upscale.png" width=400>
The agent automatically mapped our prompt "Upscale the image" to the just added upscaler tool purely based on the description and name of the upscaler tool
and was able to correctly run it.
Next, let's have a look at how you can create a new custom tool.
### Adding new tools
In this section, we show how to create a new tool that can be added to the agent.
#### Creating a new tool
We'll first start by creating a tool. We'll add the not-so-useful yet fun task of fetching the model on the Hugging Face
Hub with the most downloads for a given task.
We can do that with the following code:
```python
from huggingface_hub import list_models
task = "text-classification"
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
print(model.id)
```
For the task `text-classification`, this returns `'facebook/bart-large-mnli'`, for `translation` it returns `'t5-base`.
How do we convert this to a tool that the agent can leverage? All tools depend on the superclass `Tool` that holds the
main attributes necessary. We'll create a class that inherits from it:
```python
from transformers import Tool
class HFModelDownloadsTool(Tool):
pass
```
This class has a few needs:
- An attribute `name`, which corresponds to the name of the tool itself. To be in tune with other tools which have a
performative name, we'll name it `model_download_counter`.
- An attribute `description`, which will be used to populate the prompt of the agent.
- `inputs` and `outputs` attributes. Defining this will help the python interpreter make educated choices about types,
and will allow for a gradio-demo to be spawned when we push our tool to the Hub. They're both a list of expected
values, which can be `text`, `image`, or `audio`.
- A `__call__` method which contains the inference code. This is the code we've played with above!
Here's what our class looks like now:
```python
from transformers import Tool
from huggingface_hub import list_models
class HFModelDownloadsTool(Tool):
name = "model_download_counter"
description = (
"This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub. "
"It takes the name of the category (such as text-classification, depth-estimation, etc), and "
"returns the name of the checkpoint."
)
inputs = ["text"]
outputs = ["text"]
def __call__(self, task: str):
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
return model.id
```
We now have our tool handy. Save it in a file and import it from your main script. Let's name this file
`model_downloads.py`, so the resulting import code looks like this:
```python
from model_downloads import HFModelDownloadsTool
tool = HFModelDownloadsTool()
```
In order to let others benefit from it and for simpler initialization, we recommend pushing it to the Hub under your
namespace. To do so, just call `push_to_hub` on the `tool` variable:
```python
tool.push_to_hub("hf-model-downloads")
```
You now have your code on the Hub! Let's take a look at the final step, which is to have the agent use it.
#### Having the agent use the tool
We now have our tool that lives on the Hub which can be instantiated as such (change the user name for your tool):
```python
from transformers import load_tool
tool = load_tool("lysandre/hf-model-downloads")
```
In order to use it in the agent, simply pass it in the `additional_tools` parameter of the agent initialization method:
```python
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=[tool])
agent.run(
"Can you read out loud the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?"
)
```
which outputs the following:
```text
==Code generated by the agent==
model = model_download_counter(task="text-to-video")
print(f"The model with the most downloads is {model}.")
audio_model = text_reader(model)
==Result==
The model with the most downloads is damo-vilab/text-to-video-ms-1.7b.
```
and generates the following audio.
| **Audio** |
|------------------------------------------------------------------------------------------------------------------------------------------------------|
| <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/damo.wav" type="audio/wav"/> |
<Tip>
Depending on the LLM, some are quite brittle and require very exact prompts in order to work well. Having a well-defined
name and description of the tool is paramount to having it be leveraged by the agent.
</Tip>
### Replacing existing tools
Replacing existing tools can be done simply by assigning a new item to the agent's toolbox. Here's how one would do so:
```python
from transformers import HfAgent, load_tool
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
agent.toolbox["image-transformation"] = load_tool("diffusers/controlnet-canny-tool")
```
<Tip>
Beware when replacing tools with others! This will also adjust the agent's prompt. This can be good if you have a better
prompt suited for the task, but it can also result in your tool being selected way more than others or for other
tools to be selected instead of the one you have defined.
</Tip>
## Leveraging gradio-tools
[gradio-tools](https://github.com/freddyaboulton/gradio-tools) is a powerful library that allows using Hugging
Face Spaces as tools. It supports many existing Spaces as well as custom Spaces to be designed with it.
We offer support for `gradio_tools` by using the `Tool.from_gradio` method. For example, we want to take
advantage of the `StableDiffusionPromptGeneratorTool` tool offered in the `gradio-tools` toolkit so as to
improve our prompts and generate better images.
We first import the tool from `gradio_tools` and instantiate it:
```python
from gradio_tools import StableDiffusionPromptGeneratorTool
gradio_tool = StableDiffusionPromptGeneratorTool()
```
We pass that instance to the `Tool.from_gradio` method:
```python
from transformers import Tool
tool = Tool.from_gradio(gradio_tool)
```
Now we can manage it exactly as we would a usual custom tool. We leverage it to improve our prompt
` a rabbit wearing a space suit`:
```python
from transformers import HfAgent
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder", additional_tools=[tool])
agent.run("Generate an image of the `prompt` after improving it.", prompt="A rabbit wearing a space suit")
```
The model adequately leverages the tool:
```text
==Explanation from the agent==
I will use the following tools: `StableDiffusionPromptGenerator` to improve the prompt, then `image_generator` to generate an image according to the improved prompt.
==Code generated by the agent==
improved_prompt = StableDiffusionPromptGenerator(prompt)
print(f"The improved prompt is {improved_prompt}.")
image = image_generator(improved_prompt)
```
Before finally generating the image:
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit.png">
<Tip warning={true}>
gradio-tools requires *textual* inputs and outputs, even when working with different modalities. This implementation
works with image and audio objects. The two are currently incompatible, but will rapidly become compatible as we
work to improve the support.
</Tip>
## Future compatibility with Langchain
We love Langchain and think it has a very compelling suite of tools. In order to handle these tools,
Langchain requires *textual* inputs and outputs, even when working with different modalities.
This is often the serialized version (i.e., saved to disk) of the objects.
This difference means that multi-modality isn't handled between transformers-agents and langchain.
We aim for this limitation to be resolved in future versions, and welcome any help from avid langchain
users to help us achieve this compatibility.
We would love to have better support. If you would like to help, please
[open an issue](https://github.com/huggingface/transformers/issues/new) and share what you have in mind.
docs/source/en/debugging.md 0 → 100644
View file @ eb849f66
<!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Debugging
## Multi-GPU Network Issues Debug
When training or inferencing with `DistributedDataParallel` and multiple GPU, if you run into issue of inter-communication between processes and/or nodes, you can use the following script to diagnose network issues.
```bash
wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py
```
For example to test how 2 GPUs interact do:
```bash
python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
```
If both processes can talk to each and allocate GPU memory each will print an OK status.
For more GPUs or nodes adjust the arguments in the script.
You will find a lot more details inside the diagnostics script and even a recipe to how you could run it in a SLURM environment.
An additional level of debug is to add `NCCL_DEBUG=INFO` environment variable as follows:
```bash
NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
```
This will dump a lot of NCCL-related debug information, which you can then search online if you find that some problems are reported. Or if you're not sure how to interpret the output you can share the log file in an Issue.
## Underflow and Overflow Detection
<Tip>
This feature is currently available for PyTorch-only.
</Tip>
<Tip>
For multi-GPU training it requires DDP (`torch.distributed.launch`).
</Tip>
<Tip>
This feature can be used with any `nn.Module`-based model.
</Tip>
If you start getting `loss=NaN` or the model inhibits some other abnormal behavior due to `inf` or `nan` in
activations or weights one needs to discover where the first underflow or overflow happens and what led to it. Luckily
you can accomplish that easily by activating a special module that will do the detection automatically.
If you're using [`Trainer`], you just need to add:
```bash
--debug underflow_overflow
```
to the normal command line arguments, or pass `debug="underflow_overflow"` when creating the
[`TrainingArguments`] object.
If you're using your own training loop or another Trainer you can accomplish the same with:
```python
from transformers.debug_utils import DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model)
```
[`~debug_utils.DebugUnderflowOverflow`] inserts hooks into the model that immediately after each
forward call will test input and output variables and also the corresponding module's weights. As soon as `inf` or
`nan` is detected in at least one element of the activations or weights, the program will assert and print a report
like this (this was caught with `google/mt5-small` under fp16 mixed precision):
```
Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min abs max metadata
encoder.block.1.layer.1.DenseReluDense.dropout Dropout
0.00e+00 2.57e+02 input[0]
0.00e+00 2.85e+02 output
[...]
encoder.block.2.layer.0 T5LayerSelfAttention
6.78e-04 3.15e+03 input[0]
2.65e-04 3.42e+03 output[0]
None output[1]
2.25e-01 1.00e+04 output[2]
encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output
encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
encoder.block.2.layer.1.DenseReluDense.dropout Dropout
0.00e+00 8.76e+03 input[0]
0.00e+00 9.74e+03 output
encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00 inf output
```
The example output has been trimmed in the middle for brevity.
The second column shows the value of the absolute largest element, so if you have a closer look at the last few frames,
the inputs and outputs were in the range of `1e4`. So when this training was done under fp16 mixed precision the very
last step overflowed (since under `fp16` the largest number before `inf` is `64e3`). To avoid overflows under
`fp16` the activations must remain way below `1e4`, because `1e4 * 1e4 = 1e8` so any matrix multiplication with
large activations is going to lead to a numerical overflow condition.
At the very start of the trace you can discover at which batch number the problem occurred (here `Detected inf/nan during batch_number=0` means the problem occurred on the first batch).
Each reported frame starts by declaring the fully qualified entry for the corresponding module this frame is reporting
for. If we look just at this frame:
```
encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output
```
Here, `encoder.block.2.layer.1.layer_norm` indicates that it was a layer norm for the first layer, of the second
block of the encoder. And the specific calls of the `forward` is `T5LayerNorm`.
Let's look at the last few frames of that report:
```
Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min abs max metadata
[...]
encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00 inf output
```
The last frame reports for `Dropout.forward` function with the first entry for the only input and the second for the
only output. You can see that it was called from an attribute `dropout` inside `DenseReluDense` class. We can see
that it happened during the first layer, of the 2nd block, during the very first batch. Finally, the absolute largest
input elements was `6.27e+04` and same for the output was `inf`.
You can see here, that `T5DenseGatedGeluDense.forward` resulted in output activations, whose absolute max value was
around 62.7K, which is very close to fp16's top limit of 64K. In the next frame we have `Dropout` which renormalizes
the weights, after it zeroed some of the elements, which pushes the absolute max value to more than 64K, and we get an
overflow (`inf`).
As you can see it's the previous frames that we need to look into when the numbers start going into very large for fp16
numbers.
Let's match the report to the code from `models/t5/modeling_t5.py`:
```python
class T5DenseGatedGeluDense(nn.Module):
def __init__(self, config):
super().__init__()
self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
self.dropout = nn.Dropout(config.dropout_rate)
self.gelu_act = ACT2FN["gelu_new"]
def forward(self, hidden_states):
hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
hidden_linear = self.wi_1(hidden_states)
hidden_states = hidden_gelu * hidden_linear
hidden_states = self.dropout(hidden_states)
hidden_states = self.wo(hidden_states)
return hidden_states
```
Now it's easy to see the `dropout` call, and all the previous calls as well.
Since the detection is happening in a forward hook, these reports are printed immediately after each `forward`
returns.
Going back to the full report, to act on it and to fix the problem, we need to go a few frames up where the numbers
started to go up and most likely switch to the `fp32` mode here, so that the numbers don't overflow when multiplied
or summed up. Of course, there might be other solutions. For example, we could turn off `amp` temporarily if it's
enabled, after moving the original `forward` into a helper wrapper, like so:
```python
def _forward(self, hidden_states):
hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
hidden_linear = self.wi_1(hidden_states)
hidden_states = hidden_gelu * hidden_linear
hidden_states = self.dropout(hidden_states)
hidden_states = self.wo(hidden_states)
return hidden_states
import torch
def forward(self, hidden_states):
if torch.is_autocast_enabled():
with torch.cuda.amp.autocast(enabled=False):
return self._forward(hidden_states)
else:
return self._forward(hidden_states)
```
Since the automatic detector only reports on inputs and outputs of full frames, once you know where to look, you may
want to analyse the intermediary stages of any specific `forward` function as well. In such a case you can use the
`detect_overflow` helper function to inject the detector where you want it, for example:
```python
from debug_utils import detect_overflow
class T5LayerFF(nn.Module):
[...]
def forward(self, hidden_states):
forwarded_states = self.layer_norm(hidden_states)
detect_overflow(forwarded_states, "after layer_norm")
forwarded_states = self.DenseReluDense(forwarded_states)
detect_overflow(forwarded_states, "after DenseReluDense")
return hidden_states + self.dropout(forwarded_states)
```
You can see that we added 2 of these and now we track if `inf` or `nan` for `forwarded_states` was detected
somewhere in between.
Actually, the detector already reports these because each of the calls in the example above is a `nn.Module`, but
let's say if you had some local direct calculations this is how you'd do that.
Additionally, if you're instantiating the debugger in your own code, you can adjust the number of frames printed from
its default, e.g.:
```python
from transformers.debug_utils import DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)
```
### Specific batch absolute min and max value tracing
The same debugging class can be used for per-batch tracing with the underflow/overflow detection feature turned off.
Let's say you want to watch the absolute min and max values for all the ingredients of each `forward` call of a given
batch, and only do that for batches 1 and 3. Then you instantiate this class as:
```python
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3])
```
And now full batches 1 and 3 will be traced using the same format as the underflow/overflow detector does.
Batches are 0-indexed.
This is helpful if you know that the program starts misbehaving after a certain batch number, so you can fast-forward
right to that area. Here is a sample truncated output for such configuration:
```
*** Starting batch number=1 ***
abs min abs max metadata
shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.47e+04 input[0]
5.36e-05 7.92e+02 output
[...]
decoder.dropout Dropout
1.60e-07 2.27e+01 input[0]
0.00e+00 2.52e+01 output
decoder T5Stack
not a tensor output
lm_head Linear
1.01e-06 7.92e+02 weight
0.00e+00 1.11e+00 input[0]
6.06e-02 8.39e+01 output
T5ForConditionalGeneration
not a tensor output
*** Starting batch number=3 ***
abs min abs max metadata
shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.78e+04 input[0]
5.36e-05 7.92e+02 output
[...]
```
Here you will get a huge number of frames dumped - as many as there were forward calls in your model, so it may or may
not what you want, but sometimes it can be easier to use for debugging purposes than a normal debugger. For example, if
a problem starts happening at batch number 150. So you can dump traces for batches 149 and 150 and compare where
numbers started to diverge.
You can also specify the batch number after which to stop the training, with:
```python
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3], abort_after_batch_num=3)
```
docs/source/en/debugging.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Debugging
## Multi-GPU Network Issues Debug
When training or inferencing with `DistributedDataParallel` and multiple GPU, if you run into issue of inter-communication between processes and/or nodes, you can use the following script to diagnose network issues.
```bash
wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py
```
For example to test how 2 GPUs interact do:
```bash
python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
```
If both processes can talk to each and allocate GPU memory each will print an OK status.
For more GPUs or nodes adjust the arguments in the script.
You will find a lot more details inside the diagnostics script and even a recipe to how you could run it in a SLURM environment.
An additional level of debug is to add `NCCL_DEBUG=INFO` environment variable as follows:
```bash
NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py
```
This will dump a lot of NCCL-related debug information, which you can then search online if you find that some problems are reported. Or if you're not sure how to interpret the output you can share the log file in an Issue.
## Underflow and Overflow Detection
<Tip>
This feature is currently available for PyTorch-only.
</Tip>
<Tip>
For multi-GPU training it requires DDP (`torch.distributed.launch`).
</Tip>
<Tip>
This feature can be used with any `nn.Module`-based model.
</Tip>
If you start getting `loss=NaN` or the model inhibits some other abnormal behavior due to `inf` or `nan` in
activations or weights one needs to discover where the first underflow or overflow happens and what led to it. Luckily
you can accomplish that easily by activating a special module that will do the detection automatically.
If you're using [`Trainer`], you just need to add:
```bash
--debug underflow_overflow
```
to the normal command line arguments, or pass `debug="underflow_overflow"` when creating the
[`TrainingArguments`] object.
If you're using your own training loop or another Trainer you can accomplish the same with:
```python
from transformers.debug_utils import DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model)
```
[`~debug_utils.DebugUnderflowOverflow`] inserts hooks into the model that immediately after each
forward call will test input and output variables and also the corresponding module's weights. As soon as `inf` or
`nan` is detected in at least one element of the activations or weights, the program will assert and print a report
like this (this was caught with `google/mt5-small` under fp16 mixed precision):
```
Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min abs max metadata
encoder.block.1.layer.1.DenseReluDense.dropout Dropout
0.00e+00 2.57e+02 input[0]
0.00e+00 2.85e+02 output
[...]
encoder.block.2.layer.0 T5LayerSelfAttention
6.78e-04 3.15e+03 input[0]
2.65e-04 3.42e+03 output[0]
None output[1]
2.25e-01 1.00e+04 output[2]
encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output
encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
encoder.block.2.layer.1.DenseReluDense.dropout Dropout
0.00e+00 8.76e+03 input[0]
0.00e+00 9.74e+03 output
encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00 inf output
```
The example output has been trimmed in the middle for brevity.
The second column shows the value of the absolute largest element, so if you have a closer look at the last few frames,
the inputs and outputs were in the range of `1e4`. So when this training was done under fp16 mixed precision the very
last step overflowed (since under `fp16` the largest number before `inf` is `64e3`). To avoid overflows under
`fp16` the activations must remain way below `1e4`, because `1e4 * 1e4 = 1e8` so any matrix multiplication with
large activations is going to lead to a numerical overflow condition.
At the very start of the trace you can discover at which batch number the problem occurred (here `Detected inf/nan during batch_number=0` means the problem occurred on the first batch).
Each reported frame starts by declaring the fully qualified entry for the corresponding module this frame is reporting
for. If we look just at this frame:
```
encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output
```
Here, `encoder.block.2.layer.1.layer_norm` indicates that it was a layer norm for the first layer, of the second
block of the encoder. And the specific calls of the `forward` is `T5LayerNorm`.
Let's look at the last few frames of that report:
```
Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min abs max metadata
[...]
encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00 inf output
```
The last frame reports for `Dropout.forward` function with the first entry for the only input and the second for the
only output. You can see that it was called from an attribute `dropout` inside `DenseReluDense` class. We can see
that it happened during the first layer, of the 2nd block, during the very first batch. Finally, the absolute largest
input elements was `6.27e+04` and same for the output was `inf`.
You can see here, that `T5DenseGatedGeluDense.forward` resulted in output activations, whose absolute max value was
around 62.7K, which is very close to fp16's top limit of 64K. In the next frame we have `Dropout` which renormalizes
the weights, after it zeroed some of the elements, which pushes the absolute max value to more than 64K, and we get an
overflow (`inf`).
As you can see it's the previous frames that we need to look into when the numbers start going into very large for fp16
numbers.
Let's match the report to the code from `models/t5/modeling_t5.py`:
```python
class T5DenseGatedGeluDense(nn.Module):
def __init__(self, config):
super().__init__()
self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
self.dropout = nn.Dropout(config.dropout_rate)
self.gelu_act = ACT2FN["gelu_new"]
def forward(self, hidden_states):
hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
hidden_linear = self.wi_1(hidden_states)
hidden_states = hidden_gelu * hidden_linear
hidden_states = self.dropout(hidden_states)
hidden_states = self.wo(hidden_states)
return hidden_states
```
Now it's easy to see the `dropout` call, and all the previous calls as well.
Since the detection is happening in a forward hook, these reports are printed immediately after each `forward`
returns.
Going back to the full report, to act on it and to fix the problem, we need to go a few frames up where the numbers
started to go up and most likely switch to the `fp32` mode here, so that the numbers don't overflow when multiplied
or summed up. Of course, there might be other solutions. For example, we could turn off `amp` temporarily if it's
enabled, after moving the original `forward` into a helper wrapper, like so:
```python
def _forward(self, hidden_states):
hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
hidden_linear = self.wi_1(hidden_states)
hidden_states = hidden_gelu * hidden_linear
hidden_states = self.dropout(hidden_states)
hidden_states = self.wo(hidden_states)
return hidden_states
import torch
def forward(self, hidden_states):
if torch.is_autocast_enabled():
with torch.cuda.amp.autocast(enabled=False):
return self._forward(hidden_states)
else:
return self._forward(hidden_states)
```
Since the automatic detector only reports on inputs and outputs of full frames, once you know where to look, you may
want to analyse the intermediary stages of any specific `forward` function as well. In such a case you can use the
`detect_overflow` helper function to inject the detector where you want it, for example:
```python
from debug_utils import detect_overflow
class T5LayerFF(nn.Module):
[...]
def forward(self, hidden_states):
forwarded_states = self.layer_norm(hidden_states)
detect_overflow(forwarded_states, "after layer_norm")
forwarded_states = self.DenseReluDense(forwarded_states)
detect_overflow(forwarded_states, "after DenseReluDense")
return hidden_states + self.dropout(forwarded_states)
```
You can see that we added 2 of these and now we track if `inf` or `nan` for `forwarded_states` was detected
somewhere in between.
Actually, the detector already reports these because each of the calls in the example above is a `nn.Module`, but
let's say if you had some local direct calculations this is how you'd do that.
Additionally, if you're instantiating the debugger in your own code, you can adjust the number of frames printed from
its default, e.g.:
```python
from transformers.debug_utils import DebugUnderflowOverflow
debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)
```
### Specific batch absolute min and max value tracing
The same debugging class can be used for per-batch tracing with the underflow/overflow detection feature turned off.
Let's say you want to watch the absolute min and max values for all the ingredients of each `forward` call of a given
batch, and only do that for batches 1 and 3. Then you instantiate this class as:
```python
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3])
```
And now full batches 1 and 3 will be traced using the same format as the underflow/overflow detector does.
Batches are 0-indexed.
This is helpful if you know that the program starts misbehaving after a certain batch number, so you can fast-forward
right to that area. Here is a sample truncated output for such configuration:
```
*** Starting batch number=1 ***
abs min abs max metadata
shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.47e+04 input[0]
5.36e-05 7.92e+02 output
[...]
decoder.dropout Dropout
1.60e-07 2.27e+01 input[0]
0.00e+00 2.52e+01 output
decoder T5Stack
not a tensor output
lm_head Linear
1.01e-06 7.92e+02 weight
0.00e+00 1.11e+00 input[0]
6.06e-02 8.39e+01 output
T5ForConditionalGeneration
not a tensor output
*** Starting batch number=3 ***
abs min abs max metadata
shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.78e+04 input[0]
5.36e-05 7.92e+02 output
[...]
```
Here you will get a huge number of frames dumped - as many as there were forward calls in your model, so it may or may
not what you want, but sometimes it can be easier to use for debugging purposes than a normal debugger. For example, if
a problem starts happening at batch number 150. So you can dump traces for batches 149 and 150 and compare where
numbers started to diverge.
You can also specify the batch number after which to stop the training, with:
```python
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3], abort_after_batch_num=3)
```
docs/source/en/fast_tokenizers.md 0 → 100644
View file @ eb849f66
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Use tokenizers from 🤗 Tokenizers
The [`PreTrainedTokenizerFast`] depends on the [🤗 Tokenizers](https://huggingface.co/docs/tokenizers) library. The tokenizers obtained from the 🤗 Tokenizers library can be
loaded very simply into 🤗 Transformers.
Before getting in the specifics, let's first start by creating a dummy tokenizer in a few lines:
```python
>>> from tokenizers import Tokenizer
>>> from tokenizers.models import BPE
>>> from tokenizers.trainers import BpeTrainer
>>> from tokenizers.pre_tokenizers import Whitespace
>>> tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
>>> trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
>>> tokenizer.pre_tokenizer = Whitespace()
>>> files = [...]
>>> tokenizer.train(files, trainer)
```
We now have a tokenizer trained on the files we defined. We can either continue using it in that runtime, or save it to
a JSON file for future re-use.
## Loading directly from the tokenizer object
Let's see how to leverage this tokenizer object in the 🤗 Transformers library. The
[`PreTrainedTokenizerFast`] class allows for easy instantiation, by accepting the instantiated
*tokenizer* object as an argument:
```python
>>> from transformers import PreTrainedTokenizerFast
>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
```
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to [the tokenizer
page](main_classes/tokenizer) for more information.
## Loading from a JSON file
In order to load a tokenizer from a JSON file, let's first start by saving our tokenizer:
```python
>>> tokenizer.save("tokenizer.json")
```
The path to which we saved this file can be passed to the [`PreTrainedTokenizerFast`] initialization
method using the `tokenizer_file` parameter:
```python
>>> from transformers import PreTrainedTokenizerFast
>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_file="tokenizer.json")
```
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to [the tokenizer
page](main_classes/tokenizer) for more information.
docs/source/en/fast_tokenizers.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Use tokenizers from 🤗 Tokenizers
The [`PreTrainedTokenizerFast`] depends on the [🤗 Tokenizers](https://huggingface.co/docs/tokenizers) library. The tokenizers obtained from the 🤗 Tokenizers library can be
loaded very simply into 🤗 Transformers.
Before getting in the specifics, let's first start by creating a dummy tokenizer in a few lines:
```python
>>> from tokenizers import Tokenizer
>>> from tokenizers.models import BPE
>>> from tokenizers.trainers import BpeTrainer
>>> from tokenizers.pre_tokenizers import Whitespace
>>> tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
>>> trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
>>> tokenizer.pre_tokenizer = Whitespace()
>>> files = [...]
>>> tokenizer.train(files, trainer)
```
We now have a tokenizer trained on the files we defined. We can either continue using it in that runtime, or save it to
a JSON file for future re-use.
## Loading directly from the tokenizer object
Let's see how to leverage this tokenizer object in the 🤗 Transformers library. The
[`PreTrainedTokenizerFast`] class allows for easy instantiation, by accepting the instantiated
*tokenizer* object as an argument:
```python
>>> from transformers import PreTrainedTokenizerFast
>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
```
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to [the tokenizer
page](main_classes/tokenizer) for more information.
## Loading from a JSON file
In order to load a tokenizer from a JSON file, let's first start by saving our tokenizer:
```python
>>> tokenizer.save("tokenizer.json")
```
The path to which we saved this file can be passed to the [`PreTrainedTokenizerFast`] initialization
method using the `tokenizer_file` parameter:
```python
>>> from transformers import PreTrainedTokenizerFast
>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_file="tokenizer.json")
```
This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to [the tokenizer
page](main_classes/tokenizer) for more information.
docs/source/en/generation_strategies.md 0 → 100644
View file @ eb849f66
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Text generation strategies
Text generation is essential to many NLP tasks, such as open-ended text generation, summarization, translation, and
more. It also plays a role in a variety of mixed-modality applications that have text as an output like speech-to-text
and vision-to-text. Some of the models that can generate text include
GPT2, XLNet, OpenAI GPT, CTRL, TransformerXL, XLM, Bart, T5, GIT, Whisper.
Check out a few examples that use [`~transformers.generation_utils.GenerationMixin.generate`] method to produce
text outputs for different tasks:
* [Text summarization](./tasks/summarization#inference)
* [Image captioning](./model_doc/git#transformers.GitForCausalLM.forward.example)
* [Audio transcription](./model_doc/whisper#transformers.WhisperForConditionalGeneration.forward.example)
Note that the inputs to the generate method depend on the model's modality. They are returned by the model's preprocessor
class, such as AutoTokenizer or AutoProcessor. If a model's preprocessor creates more than one kind of input, pass all
the inputs to generate(). You can learn more about the individual model's preprocessor in the corresponding model's documentation.
The process of selecting output tokens to generate text is known as decoding, and you can customize the decoding strategy
that the `generate()` method will use. Modifying a decoding strategy does not change the values of any trainable parameters.
However, it can have a noticeable impact on the quality of the generated output. It can help reduce repetition in the text
and make it more coherent.
This guide describes:
* default generation configuration
* common decoding strategies and their main parameters
* saving and sharing custom generation configurations with your fine-tuned model on 🤗 Hub
## Default text generation configuration
A decoding strategy for a model is defined in its generation configuration. When using pre-trained models for inference
within a [`pipeline`], the models call the `PreTrainedModel.generate()` method that applies a default generation
configuration under the hood. The default configuration is also used when no custom configuration has been saved with
the model.
When you load a model explicitly, you can inspect the generation configuration that comes with it through
`model.generation_config`:
```python
>>> from transformers import AutoModelForCausalLM
>>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
>>> model.generation_config
GenerationConfig {
"_from_model_config": true,
"bos_token_id": 50256,
"eos_token_id": 50256,
"transformers_version": "4.26.0.dev0"
}
```
Printing out the `model.generation_config` reveals only the values that are different from the default generation
configuration, and does not list any of the default values.
The default generation configuration limits the size of the output combined with the input prompt to a maximum of 20
tokens to avoid running into resource limitations. The default decoding strategy is greedy search, which is the simplest decoding strategy that picks a token with the highest probability as the next token. For many tasks
and small output sizes this works well. However, when used to generate longer outputs, greedy search can start
producing highly repetitive results.
## Customize text generation
You can override any `generation_config` by passing the parameters and their values directly to the [`generate`] method:
```python
>>> my_model.generate(**inputs, num_beams=4, do_sample=True)
```
Even if the default decoding strategy mostly works for your task, you can still tweak a few things. Some of the
commonly adjusted parameters include:
- `max_new_tokens`: the maximum number of tokens to generate. In other words, the size of the output sequence, not
including the tokens in the prompt.
- `num_beams`: by specifying a number of beams higher than 1, you are effectively switching from greedy search to
beam search. This strategy evaluates several hypotheses at each time step and eventually chooses the hypothesis that
has the overall highest probability for the entire sequence. This has the advantage of identifying high-probability
sequences that start with a lower probability initial tokens and would've been ignored by the greedy search.
- `do_sample`: if set to `True`, this parameter enables decoding strategies such as multinomial sampling, beam-search
multinomial sampling, Top-K sampling and Top-p sampling. All these strategies select the next token from the probability
distribution over the entire vocabulary with various strategy-specific adjustments.
- `num_return_sequences`: the number of sequence candidates to return for each input. This options is only available for
the decoding strategies that support multiple sequence candidates, e.g. variations of beam search and sampling. Decoding
strategies like greedy search and contrastive search return a single output sequence.
## Save a custom decoding strategy with your model
If you would like to share your fine-tuned model with a specific generation configuration, you can:
* Create a [`GenerationConfig`] class instance
* Specify the decoding strategy parameters
* Save your generation configuration with [`GenerationConfig.save_pretrained`], making sure to leave its `config_file_name` argument empty
* Set `push_to_hub` to `True` to upload your config to the model's repo
```python
>>> from transformers import AutoModelForCausalLM, GenerationConfig
>>> model = AutoModelForCausalLM.from_pretrained("my_account/my_model")
>>> generation_config = GenerationConfig(
... max_new_tokens=50, do_sample=True, top_k=50, eos_token_id=model.config.eos_token_id
... )
>>> generation_config.save_pretrained("my_account/my_model", push_to_hub=True)
```
You can also store several generation configurations in a single directory, making use of the `config_file_name`
argument in [`GenerationConfig.save_pretrained`]. You can later instantiate them with [`GenerationConfig.from_pretrained`]. This is useful if you want to
store several generation configurations for a single model (e.g. one for creative text generation with sampling, and
one for summarization with beam search). You must have the right Hub permissions to add configuration files to a model.
```python
>>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig
>>> tokenizer = AutoTokenizer.from_pretrained("t5-small")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-small")
>>> translation_generation_config = GenerationConfig(
... num_beams=4,
... early_stopping=True,
... decoder_start_token_id=0,
... eos_token_id=model.config.eos_token_id,
... pad_token=model.config.pad_token_id,
... )
>>> translation_generation_config.save_pretrained("t5-small", "translation_generation_config.json", push_to_hub=True)
>>> # You could then use the named generation config file to parameterize generation
>>> generation_config = GenerationConfig.from_pretrained("t5-small", "translation_generation_config.json")
>>> inputs = tokenizer("translate English to French: Configuration files are easy to use!", return_tensors="pt")
>>> outputs = model.generate(**inputs, generation_config=generation_config)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
['Les fichiers de configuration sont faciles à utiliser !']
```
## Streaming
The `generate()` supports streaming, through its `streamer` input. The `streamer` input is compatible any instance
from a class that has the following methods: `put()` and `end()`. Internally, `put()` is used to push new tokens and
`end()` is used to flag the end of text generation.
<Tip warning={true}>
The API for the streamer classes is still under development and may change in the future.
</Tip>
In practice, you can craft your own streaming class for all sorts of purposes! We also have basic streaming classes
ready for you to use. For example, you can use the [`TextStreamer`] class to stream the output of `generate()` into
your screen, one word at a time:
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
>>> tok = AutoTokenizer.from_pretrained("gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("gpt2")
>>> inputs = tok(["An increasing sequence: one,"], return_tensors="pt")
>>> streamer = TextStreamer(tok)
>>> # Despite returning the usual output, the streamer will also print the generated text to stdout.
>>> _ = model.generate(**inputs, streamer=streamer, max_new_tokens=20)
An increasing sequence: one, two, three, four, five, six, seven, eight, nine, ten, eleven,
```
## Decoding strategies
Certain combinations of the `generate()` parameters, and ultimately `generation_config`, can be used to enable specific
decoding strategies. If you are new to this concept, we recommend reading [this blog post that illustrates how common decoding strategies work](https://huggingface.co/blog/how-to-generate).
Here, we'll show some of the parameters that control the decoding strategies and illustrate how you can use them.
### Greedy Search
[`generate`] uses greedy search decoding by default so you don't have to pass any parameters to enable it. This means the parameters `num_beams` is set to 1 and `do_sample=False`.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "I look forward to"
>>> checkpoint = "distilgpt2"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['I look forward to seeing you all again!\n\n\n\n\n\n\n\n\n\n\n']
```
### Contrastive search
The contrastive search decoding strategy was proposed in the 2022 paper [A Contrastive Framework for Neural Text Generation](https://arxiv.org/abs/2202.06417).
It demonstrates superior results for generating non-repetitive yet coherent long outputs. To learn how contrastive search
works, check out [this blog post](https://huggingface.co/blog/introducing-csearch).
The two main parameters that enable and control the behavior of contrastive search are `penalty_alpha` and `top_k`:
```python
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> checkpoint = "gpt2-large"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> prompt = "Hugging Face Company is"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> outputs = model.generate(**inputs, penalty_alpha=0.6, top_k=4, max_new_tokens=100)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Hugging Face Company is a family owned and operated business. \
We pride ourselves on being the best in the business and our customer service is second to none.\
\n\nIf you have any questions about our products or services, feel free to contact us at any time.\
We look forward to hearing from you!']
```
### Multinomial sampling
As opposed to greedy search that always chooses a token with the highest probability as the
next token, multinomial sampling (also called ancestral sampling) randomly selects the next token based on the probability distribution over the entire
vocabulary given by the model. Every token with a non-zero probability has a chance of being selected, thus reducing the
risk of repetition.
To enable multinomial sampling set `do_sample=True` and `num_beams=1`.
```python
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> checkpoint = "gpt2-large"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> prompt = "Today was an amazing day because"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> outputs = model.generate(**inputs, do_sample=True, num_beams=1, max_new_tokens=100)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Today was an amazing day because we are now in the final stages of our trip to New York City which was very tough. \
It is a difficult schedule and a challenging part of the year but still worth it. I have been taking things easier and \
I feel stronger and more motivated to be out there on their tour. Hopefully, that experience is going to help them with \
their upcoming events which are currently scheduled in Australia.\n\nWe love that they are here. They want to make a \
name for themselves and become famous for what they']
```
### Beam-search decoding
Unlike greedy search, beam-search decoding keeps several hypotheses at each time step and eventually chooses
the hypothesis that has the overall highest probability for the entire sequence. This has the advantage of identifying high-probability
sequences that start with lower probability initial tokens and would've been ignored by the greedy search.
To enable this decoding strategy, specify the `num_beams` (aka number of hypotheses to keep track of) that is greater than 1.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "It is astonishing how one can"
>>> checkpoint = "gpt2-medium"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, max_new_tokens=50)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['It is astonishing how one can have such a profound impact on the lives of so many people in such a short period of \
time."\n\nHe added: "I am very proud of the work I have been able to do in the last few years.\n\n"I have']
```
### Beam-search multinomial sampling
As the name implies, this decoding strategy combines beam search with multinomial sampling. You need to specify
the `num_beams` greater than 1, and set `do_sample=True` to use this decoding strategy.
```python
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> prompt = "translate English to German: The house is wonderful."
>>> checkpoint = "t5-small"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, do_sample=True)
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'Das Haus ist wunderbar.'
```
### Diverse beam search decoding
The diverse beam search decoding strategy is an extension of the beam search strategy that allows for generating a more diverse
set of beam sequences to choose from. To learn how it works, refer to [Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models](https://arxiv.org/pdf/1610.02424.pdf).
This approach has two main parameters: `num_beams` and `num_beam_groups`.
The groups are selected to ensure they are distinct enough compared to the others, and regular beam search is used within each group.
```python
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> checkpoint = "google/pegasus-xsum"
>>> prompt = "The Permaculture Design Principles are a set of universal design principles \
>>> that can be applied to any location, climate and culture, and they allow us to design \
>>> the most efficient and sustainable human habitation and food production systems. \
>>> Permaculture is a design system that encompasses a wide variety of disciplines, such \
>>> as ecology, landscape design, environmental science and energy conservation, and the \
>>> Permaculture design principles are drawn from these various disciplines. Each individual \
>>> design principle itself embodies a complete conceptual framework based on sound \
>>> scientific principles. When we bring all these separate principles together, we can \
>>> create a design system that both looks at whole systems, the parts that these systems \
>>> consist of, and how those parts interact with each other to create a complex, dynamic, \
>>> living system. Each design principle serves as a tool that allows us to integrate all \
>>> the separate parts of a design, referred to as elements, into a functional, synergistic, \
>>> whole system, where the elements harmoniously interact and work together in the most \
>>> efficient way possible."
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, num_beam_groups=5, max_new_tokens=30)
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'The Design Principles are a set of universal design principles that can be applied to any location, climate and culture, and they allow us to design the most efficient and sustainable human habitation and food production systems.'
```
This guide illustrates the main parameters that enable various decoding strategies. More advanced parameters exist for the
[`generate`] method, which gives you even further control over the [`generate`] method's behavior.
For the complete list of the available parameters, refer to the [API documentation](./main_classes/text_generation.md).
### Assisted Decoding
Assisted decoding is a modification of the decoding strategies above that uses an assistant model with the same
tokenizer (ideally a much smaller model) to greedily generate a few candidate tokens. The main model then validates
the candidate tokens in a single forward pass, which speeds up the decoding process. Currently, only greedy search
and sampling are supported with assisted decoding, and doesn't support batched inputs. To learn more about assisted
decoding, check [this blog post](https://huggingface.co/blog/assisted-generation).
To enable assisted decoding, set the `assistant_model` argument with a model.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "Alice and Bob"
>>> checkpoint = "EleutherAI/pythia-1.4b-deduped"
>>> assistant_checkpoint = "EleutherAI/pythia-160m-deduped"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> assistant_model = AutoModelForCausalLM.from_pretrained(assistant_checkpoint)
>>> outputs = model.generate(**inputs, assistant_model=assistant_model)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Alice and Bob are sitting in a bar. Alice is drinking a beer and Bob is drinking a']
```
When using assisted decoding with sampling methods, you can use the `temperarure` argument to control the randomness
just like in multinomial sampling. However, in assisted decoding, reducing the temperature will help improving latency.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "Alice and Bob"
>>> checkpoint = "EleutherAI/pythia-1.4b-deduped"
>>> assistant_checkpoint = "EleutherAI/pythia-160m-deduped"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> assistant_model = AutoModelForCausalLM.from_pretrained(assistant_checkpoint)
>>> outputs = model.generate(**inputs, assistant_model=assistant_model, do_sample=True, temperature=0.5)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
["Alice and Bob are sitting on the sofa. Alice says, 'I'm going to my room"]
```
docs/source/en/generation_strategies.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Text generation strategies
Text generation is essential to many NLP tasks, such as open-ended text generation, summarization, translation, and
more. It also plays a role in a variety of mixed-modality applications that have text as an output like speech-to-text
and vision-to-text. Some of the models that can generate text include
GPT2, XLNet, OpenAI GPT, CTRL, TransformerXL, XLM, Bart, T5, GIT, Whisper.
Check out a few examples that use [`~transformers.generation_utils.GenerationMixin.generate`] method to produce
text outputs for different tasks:
* [Text summarization](./tasks/summarization#inference)
* [Image captioning](./model_doc/git#transformers.GitForCausalLM.forward.example)
* [Audio transcription](./model_doc/whisper#transformers.WhisperForConditionalGeneration.forward.example)
Note that the inputs to the generate method depend on the model's modality. They are returned by the model's preprocessor
class, such as AutoTokenizer or AutoProcessor. If a model's preprocessor creates more than one kind of input, pass all
the inputs to generate(). You can learn more about the individual model's preprocessor in the corresponding model's documentation.
The process of selecting output tokens to generate text is known as decoding, and you can customize the decoding strategy
that the `generate()` method will use. Modifying a decoding strategy does not change the values of any trainable parameters.
However, it can have a noticeable impact on the quality of the generated output. It can help reduce repetition in the text
and make it more coherent.
This guide describes:
* default generation configuration
* common decoding strategies and their main parameters
* saving and sharing custom generation configurations with your fine-tuned model on 🤗 Hub
## Default text generation configuration
A decoding strategy for a model is defined in its generation configuration. When using pre-trained models for inference
within a [`pipeline`], the models call the `PreTrainedModel.generate()` method that applies a default generation
configuration under the hood. The default configuration is also used when no custom configuration has been saved with
the model.
When you load a model explicitly, you can inspect the generation configuration that comes with it through
`model.generation_config`:
```python
>>> from transformers import AutoModelForCausalLM
>>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
>>> model.generation_config
GenerationConfig {
"_from_model_config": true,
"bos_token_id": 50256,
"eos_token_id": 50256,
"transformers_version": "4.26.0.dev0"
}
```
Printing out the `model.generation_config` reveals only the values that are different from the default generation
configuration, and does not list any of the default values.
The default generation configuration limits the size of the output combined with the input prompt to a maximum of 20
tokens to avoid running into resource limitations. The default decoding strategy is greedy search, which is the simplest decoding strategy that picks a token with the highest probability as the next token. For many tasks
and small output sizes this works well. However, when used to generate longer outputs, greedy search can start
producing highly repetitive results.
## Customize text generation
You can override any `generation_config` by passing the parameters and their values directly to the [`generate`] method:
```python
>>> my_model.generate(**inputs, num_beams=4, do_sample=True)
```
Even if the default decoding strategy mostly works for your task, you can still tweak a few things. Some of the
commonly adjusted parameters include:
- `max_new_tokens`: the maximum number of tokens to generate. In other words, the size of the output sequence, not
including the tokens in the prompt.
- `num_beams`: by specifying a number of beams higher than 1, you are effectively switching from greedy search to
beam search. This strategy evaluates several hypotheses at each time step and eventually chooses the hypothesis that
has the overall highest probability for the entire sequence. This has the advantage of identifying high-probability
sequences that start with a lower probability initial tokens and would've been ignored by the greedy search.
- `do_sample`: if set to `True`, this parameter enables decoding strategies such as multinomial sampling, beam-search
multinomial sampling, Top-K sampling and Top-p sampling. All these strategies select the next token from the probability
distribution over the entire vocabulary with various strategy-specific adjustments.
- `num_return_sequences`: the number of sequence candidates to return for each input. This options is only available for
the decoding strategies that support multiple sequence candidates, e.g. variations of beam search and sampling. Decoding
strategies like greedy search and contrastive search return a single output sequence.
## Save a custom decoding strategy with your model
If you would like to share your fine-tuned model with a specific generation configuration, you can:
* Create a [`GenerationConfig`] class instance
* Specify the decoding strategy parameters
* Save your generation configuration with [`GenerationConfig.save_pretrained`], making sure to leave its `config_file_name` argument empty
* Set `push_to_hub` to `True` to upload your config to the model's repo
```python
>>> from transformers import AutoModelForCausalLM, GenerationConfig
>>> model = AutoModelForCausalLM.from_pretrained("my_account/my_model")
>>> generation_config = GenerationConfig(
... max_new_tokens=50, do_sample=True, top_k=50, eos_token_id=model.config.eos_token_id
... )
>>> generation_config.save_pretrained("my_account/my_model", push_to_hub=True)
```
You can also store several generation configurations in a single directory, making use of the `config_file_name`
argument in [`GenerationConfig.save_pretrained`]. You can later instantiate them with [`GenerationConfig.from_pretrained`]. This is useful if you want to
store several generation configurations for a single model (e.g. one for creative text generation with sampling, and
one for summarization with beam search). You must have the right Hub permissions to add configuration files to a model.
```python
>>> from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig
>>> tokenizer = AutoTokenizer.from_pretrained("t5-small")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-small")
>>> translation_generation_config = GenerationConfig(
... num_beams=4,
... early_stopping=True,
... decoder_start_token_id=0,
... eos_token_id=model.config.eos_token_id,
... pad_token=model.config.pad_token_id,
... )
>>> translation_generation_config.save_pretrained("t5-small", "translation_generation_config.json", push_to_hub=True)
>>> # You could then use the named generation config file to parameterize generation
>>> generation_config = GenerationConfig.from_pretrained("t5-small", "translation_generation_config.json")
>>> inputs = tokenizer("translate English to French: Configuration files are easy to use!", return_tensors="pt")
>>> outputs = model.generate(**inputs, generation_config=generation_config)
>>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
['Les fichiers de configuration sont faciles à utiliser !']
```
## Streaming
The `generate()` supports streaming, through its `streamer` input. The `streamer` input is compatible any instance
from a class that has the following methods: `put()` and `end()`. Internally, `put()` is used to push new tokens and
`end()` is used to flag the end of text generation.
<Tip warning={true}>
The API for the streamer classes is still under development and may change in the future.
</Tip>
In practice, you can craft your own streaming class for all sorts of purposes! We also have basic streaming classes
ready for you to use. For example, you can use the [`TextStreamer`] class to stream the output of `generate()` into
your screen, one word at a time:
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
>>> tok = AutoTokenizer.from_pretrained("gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("gpt2")
>>> inputs = tok(["An increasing sequence: one,"], return_tensors="pt")
>>> streamer = TextStreamer(tok)
>>> # Despite returning the usual output, the streamer will also print the generated text to stdout.
>>> _ = model.generate(**inputs, streamer=streamer, max_new_tokens=20)
An increasing sequence: one, two, three, four, five, six, seven, eight, nine, ten, eleven,
```
## Decoding strategies
Certain combinations of the `generate()` parameters, and ultimately `generation_config`, can be used to enable specific
decoding strategies. If you are new to this concept, we recommend reading [this blog post that illustrates how common decoding strategies work](https://huggingface.co/blog/how-to-generate).
Here, we'll show some of the parameters that control the decoding strategies and illustrate how you can use them.
### Greedy Search
[`generate`] uses greedy search decoding by default so you don't have to pass any parameters to enable it. This means the parameters `num_beams` is set to 1 and `do_sample=False`.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "I look forward to"
>>> checkpoint = "distilgpt2"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['I look forward to seeing you all again!\n\n\n\n\n\n\n\n\n\n\n']
```
### Contrastive search
The contrastive search decoding strategy was proposed in the 2022 paper [A Contrastive Framework for Neural Text Generation](https://arxiv.org/abs/2202.06417).
It demonstrates superior results for generating non-repetitive yet coherent long outputs. To learn how contrastive search
works, check out [this blog post](https://huggingface.co/blog/introducing-csearch).
The two main parameters that enable and control the behavior of contrastive search are `penalty_alpha` and `top_k`:
```python
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> checkpoint = "gpt2-large"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> prompt = "Hugging Face Company is"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> outputs = model.generate(**inputs, penalty_alpha=0.6, top_k=4, max_new_tokens=100)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Hugging Face Company is a family owned and operated business. \
We pride ourselves on being the best in the business and our customer service is second to none.\
\n\nIf you have any questions about our products or services, feel free to contact us at any time.\
We look forward to hearing from you!']
```
### Multinomial sampling
As opposed to greedy search that always chooses a token with the highest probability as the
next token, multinomial sampling (also called ancestral sampling) randomly selects the next token based on the probability distribution over the entire
vocabulary given by the model. Every token with a non-zero probability has a chance of being selected, thus reducing the
risk of repetition.
To enable multinomial sampling set `do_sample=True` and `num_beams=1`.
```python
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> checkpoint = "gpt2-large"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> prompt = "Today was an amazing day because"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> outputs = model.generate(**inputs, do_sample=True, num_beams=1, max_new_tokens=100)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Today was an amazing day because we are now in the final stages of our trip to New York City which was very tough. \
It is a difficult schedule and a challenging part of the year but still worth it. I have been taking things easier and \
I feel stronger and more motivated to be out there on their tour. Hopefully, that experience is going to help them with \
their upcoming events which are currently scheduled in Australia.\n\nWe love that they are here. They want to make a \
name for themselves and become famous for what they']
```
### Beam-search decoding
Unlike greedy search, beam-search decoding keeps several hypotheses at each time step and eventually chooses
the hypothesis that has the overall highest probability for the entire sequence. This has the advantage of identifying high-probability
sequences that start with lower probability initial tokens and would've been ignored by the greedy search.
To enable this decoding strategy, specify the `num_beams` (aka number of hypotheses to keep track of) that is greater than 1.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "It is astonishing how one can"
>>> checkpoint = "gpt2-medium"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, max_new_tokens=50)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['It is astonishing how one can have such a profound impact on the lives of so many people in such a short period of \
time."\n\nHe added: "I am very proud of the work I have been able to do in the last few years.\n\n"I have']
```
### Beam-search multinomial sampling
As the name implies, this decoding strategy combines beam search with multinomial sampling. You need to specify
the `num_beams` greater than 1, and set `do_sample=True` to use this decoding strategy.
```python
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> prompt = "translate English to German: The house is wonderful."
>>> checkpoint = "t5-small"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, do_sample=True)
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'Das Haus ist wunderbar.'
```
### Diverse beam search decoding
The diverse beam search decoding strategy is an extension of the beam search strategy that allows for generating a more diverse
set of beam sequences to choose from. To learn how it works, refer to [Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models](https://arxiv.org/pdf/1610.02424.pdf).
This approach has two main parameters: `num_beams` and `num_beam_groups`.
The groups are selected to ensure they are distinct enough compared to the others, and regular beam search is used within each group.
```python
>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
>>> checkpoint = "google/pegasus-xsum"
>>> prompt = "The Permaculture Design Principles are a set of universal design principles \
>>> that can be applied to any location, climate and culture, and they allow us to design \
>>> the most efficient and sustainable human habitation and food production systems. \
>>> Permaculture is a design system that encompasses a wide variety of disciplines, such \
>>> as ecology, landscape design, environmental science and energy conservation, and the \
>>> Permaculture design principles are drawn from these various disciplines. Each individual \
>>> design principle itself embodies a complete conceptual framework based on sound \
>>> scientific principles. When we bring all these separate principles together, we can \
>>> create a design system that both looks at whole systems, the parts that these systems \
>>> consist of, and how those parts interact with each other to create a complex, dynamic, \
>>> living system. Each design principle serves as a tool that allows us to integrate all \
>>> the separate parts of a design, referred to as elements, into a functional, synergistic, \
>>> whole system, where the elements harmoniously interact and work together in the most \
>>> efficient way possible."
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint)
>>> outputs = model.generate(**inputs, num_beams=5, num_beam_groups=5, max_new_tokens=30)
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
'The Design Principles are a set of universal design principles that can be applied to any location, climate and culture, and they allow us to design the most efficient and sustainable human habitation and food production systems.'
```
This guide illustrates the main parameters that enable various decoding strategies. More advanced parameters exist for the
[`generate`] method, which gives you even further control over the [`generate`] method's behavior.
For the complete list of the available parameters, refer to the [API documentation](./main_classes/text_generation.mdx).
### Assisted Decoding
Assisted decoding is a modification of the decoding strategies above that uses an assistant model with the same
tokenizer (ideally a much smaller model) to greedily generate a few candidate tokens. The main model then validates
the candidate tokens in a single forward pass, which speeds up the decoding process. Currently, only greedy search
and sampling are supported with assisted decoding, and doesn't support batched inputs. To learn more about assisted
decoding, check [this blog post](https://huggingface.co/blog/assisted-generation).
To enable assisted decoding, set the `assistant_model` argument with a model.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "Alice and Bob"
>>> checkpoint = "EleutherAI/pythia-1.4b-deduped"
>>> assistant_checkpoint = "EleutherAI/pythia-160m-deduped"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> assistant_model = AutoModelForCausalLM.from_pretrained(assistant_checkpoint)
>>> outputs = model.generate(**inputs, assistant_model=assistant_model)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['Alice and Bob are sitting in a bar. Alice is drinking a beer and Bob is drinking a']
```
When using assisted decoding with sampling methods, you can use the `temperarure` argument to control the randomness
just like in multinomial sampling. However, in assisted decoding, reducing the temperature will help improving latency.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> prompt = "Alice and Bob"
>>> checkpoint = "EleutherAI/pythia-1.4b-deduped"
>>> assistant_checkpoint = "EleutherAI/pythia-160m-deduped"
>>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> model = AutoModelForCausalLM.from_pretrained(checkpoint)
>>> assistant_model = AutoModelForCausalLM.from_pretrained(assistant_checkpoint)
>>> outputs = model.generate(**inputs, assistant_model=assistant_model, do_sample=True, temperature=0.5)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
["Alice and Bob are sitting on the sofa. Alice says, 'I'm going to my room"]
```
docs/source/en/glossary.md 0 → 100644
View file @ eb849f66
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Glossary
This glossary defines general machine learning and 🤗 Transformers terms to help you better understand the
documentation.
## A
### attention mask
The attention mask is an optional argument used when batching sequences together.
<Youtube id="M6adb1j2jPI"/>
This argument indicates to the model which tokens should be attended to, and which should not.
For example, consider these two sequences:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence_a = "This is a short sequence."
>>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
>>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
>>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
```
The encoded versions have different lengths:
```python
>>> len(encoded_sequence_a), len(encoded_sequence_b)
(8, 19)
```
Therefore, we can't put them together in the same tensor as-is. The first sequence needs to be padded up to the length
of the second one, or the second one needs to be truncated down to the length of the first one.
In the first case, the list of IDs will be extended by the padding indices. We can pass a list to the tokenizer and ask
it to pad like this:
```python
>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
```
We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:
```python
>>> padded_sequences["input_ids"]
[[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
```
This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating the
position of the padded indices so that the model does not attend to them. For the [`BertTokenizer`], `1` indicates a
value that should be attended to, while `0` indicates a padded value. This attention mask is in the dictionary returned
by the tokenizer under the key "attention_mask":
```python
>>> padded_sequences["attention_mask"]
[[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
```
### autoencoding models
See [encoder models](#encoder-models) and [masked language modeling](#masked-language-modeling-mlm)
### autoregressive models
See [causal language modeling](#causal-language-modeling) and [decoder models](#decoder-models)
## B
### backbone
The backbone is the network (embeddings and layers) that outputs the raw hidden states or features. It is usually connected to a [head](#head) which accepts the features as its input to make a prediction. For example, [`ViTModel`] is a backbone without a specific head on top. Other models can also use [`VitModel`] as a backbone such as [DPT](model_doc/dpt).
## C
### causal language modeling
A pretraining task where the model reads the texts in order and has to predict the next word. It's usually done by
reading the whole sentence but using a mask inside the model to hide the future tokens at a certain timestep.
### channel
Color images are made up of some combination of values in three channels - red, green, and blue (RGB) - and grayscale images only have one channel. In 🤗 Transformers, the channel can be the first or last dimension of an image's tensor: [`n_channels`, `height`, `width`] or [`height`, `width`, `n_channels`].
### connectionist temporal classification (CTC)
An algorithm which allows a model to learn without knowing exactly how the input and output are aligned; CTC calculates the distribution of all possible outputs for a given input and chooses the most likely output from it. CTC is commonly used in speech recognition tasks because speech doesn't always cleanly align with the transcript for a variety of reasons such as a speaker's different speech rates.
### convolution
A type of layer in a neural network where the input matrix is multiplied element-wise by a smaller matrix (kernel or filter) and the values are summed up in a new matrix. This is known as a convolutional operation which is repeated over the entire input matrix. Each operation is applied to a different segment of the input matrix. Convolutional neural networks (CNNs) are commonly used in computer vision.
## D
### decoder input IDs
This input is specific to encoder-decoder models, and contains the input IDs that will be fed to the decoder. These
inputs should be used for sequence to sequence tasks, such as translation or summarization, and are usually built in a
way specific to each model.
Most encoder-decoder models (BART, T5) create their `decoder_input_ids` on their own from the `labels`. In such models,
passing the `labels` is the preferred way to handle training.
Please check each model's docs to see how they handle these input IDs for sequence to sequence training.
### decoder models
Also referred to as autoregressive models, decoder models involve a pretraining task (called causal language modeling) where the model reads the texts in order and has to predict the next word. It's usually done by
reading the whole sentence with a mask to hide future tokens at a certain timestep.
<Youtube id="d_ixlCubqQw"/>
### deep learning (DL)
Machine learning algorithms which uses neural networks with several layers.
## E
### encoder models
Also known as autoencoding models, encoder models take an input (such as text or images) and transform them into a condensed numerical representation called an embedding. Oftentimes, encoder models are pretrained using techniques like [masked language modeling](#masked-language-modeling-mlm), which masks parts of the input sequence and forces the model to create more meaningful representations.
<Youtube id="H39Z_720T5s"/>
## F
### feature extraction
The process of selecting and transforming raw data into a set of features that are more informative and useful for machine learning algorithms. Some examples of feature extraction include transforming raw text into word embeddings and extracting important features such as edges or shapes from image/video data.
### feed forward chunking
In each residual attention block in transformers the self-attention layer is usually followed by 2 feed forward layers.
The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (e.g., for
`bert-base-uncased`).
For an input of size `[batch_size, sequence_length]`, the memory required to store the intermediate feed forward
embeddings `[batch_size, sequence_length, config.intermediate_size]` can account for a large fraction of the memory
use. The authors of [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) noticed that since the
computation is independent of the `sequence_length` dimension, it is mathematically equivalent to compute the output
embeddings of both feed forward layers `[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n`
individually and concat them afterward to `[batch_size, sequence_length, config.hidden_size]` with `n =
sequence_length`, which trades increased computation time against reduced memory use, but yields a mathematically
**equivalent** result.
For models employing the function [`apply_chunking_to_forward`], the `chunk_size` defines the number of output
embeddings that are computed in parallel and thus defines the trade-off between memory and time complexity. If
`chunk_size` is set to 0, no feed forward chunking is done.
### finetuned models
Finetuning is a form of transfer learning which involves taking a pretrained model, freezing its weights, and replacing the output layer with a newly added [model head](#head). The model head is trained on your target dataset.
See the [Fine-tune a pretrained model](https://huggingface.co/docs/transformers/training) tutorial for more details, and learn how to fine-tune models with 🤗 Transformers.
## H
### head
The model head refers to the last layer of a neural network that accepts the raw hidden states and projects them onto a different dimension. There is a different model head for each task. For example:
* [`GPT2ForSequenceClassification`] is a sequence classification head - a linear layer - on top of the base [`GPT2Model`].
* [`ViTForImageClassification`] is an image classification head - a linear layer on top of the final hidden state of the `CLS` token - on top of the base [`ViTModel`].
* [`Wav2Vec2ForCTC`] ia a language modeling head with [CTC](#connectionist-temporal-classification-(CTC)) on top of the base [`Wav2Vec2Model`].
## I
### image patch
Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the `patch_size` - or resolution - of the model in it's configuration.
### inference
Inference is the process of evaluating a model on new data after training is complete. See the [Pipeline for inference](https://huggingface.co/docs/transformers/pipeline_tutorial) tutorial to learn how to perform inference with 🤗 Transformers.
### input IDs
The input ids are often the only required parameters to be passed to the model as input. They are token indices,
numerical representations of tokens building the sequences that will be used as input by the model.
<Youtube id="VFp38yj8h3A"/>
Each tokenizer works differently but the underlying mechanism remains the same. Here's an example using the BERT
tokenizer, which is a [WordPiece](https://arxiv.org/pdf/1609.08144.pdf) tokenizer:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence = "A Titan RTX has 24GB of VRAM"
```
The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.
```python
>>> tokenized_sequence = tokenizer.tokenize(sequence)
```
The tokens are either words or subwords. Here for instance, "VRAM" wasn't in the model vocabulary, so it's been split
in "V", "RA" and "M". To indicate those tokens are not separate words but parts of the same word, a double-hash prefix
is added for "RA" and "M":
```python
>>> print(tokenized_sequence)
['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
```
These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding
the sentence to the tokenizer, which leverages the Rust implementation of [🤗
Tokenizers](https://github.com/huggingface/tokenizers) for peak performance.
```python
>>> inputs = tokenizer(sequence)
```
The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
token indices are under the key `input_ids`:
```python
>>> encoded_sequence = inputs["input_ids"]
>>> print(encoded_sequence)
[101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
```
Note that the tokenizer automatically adds "special tokens" (if the associated model relies on them) which are special
IDs the model sometimes uses.
If we decode the previous sequence of ids,
```python
>>> decoded_sequence = tokenizer.decode(encoded_sequence)
```
we will see
```python
>>> print(decoded_sequence)
[CLS] A Titan RTX has 24GB of VRAM [SEP]
```
because this is the way a [`BertModel`] is going to expect its inputs.
## L
### labels
The labels are an optional argument which can be passed in order for the model to compute the loss itself. These labels
should be the expected prediction of the model: it will use the standard loss in order to compute the loss between its
predictions and the expected value (the label).
These labels are different according to the model head, for example:
- For sequence classification models, ([`BertForSequenceClassification`]), the model expects a tensor of dimension
`(batch_size)` with each value of the batch corresponding to the expected label of the entire sequence.
- For token classification models, ([`BertForTokenClassification`]), the model expects a tensor of dimension
`(batch_size, seq_length)` with each value corresponding to the expected label of each individual token.
- For masked language modeling, ([`BertForMaskedLM`]), the model expects a tensor of dimension `(batch_size,
seq_length)` with each value corresponding to the expected label of each individual token: the labels being the token
ID for the masked token, and values to be ignored for the rest (usually -100).
- For sequence to sequence tasks, ([`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]), the model
expects a tensor of dimension `(batch_size, tgt_seq_length)` with each value corresponding to the target sequences
associated with each input sequence. During training, both BART and T5 will make the appropriate
`decoder_input_ids` and decoder attention masks internally. They usually do not need to be supplied. This does not
apply to models leveraging the Encoder-Decoder framework.
- For image classification models, ([`ViTForImageClassification`]), the model expects a tensor of dimension
`(batch_size)` with each value of the batch corresponding to the expected label of each individual image.
- For semantic segmentation models, ([`SegformerForSemanticSegmentation`]), the model expects a tensor of dimension
`(batch_size, height, width)` with each value of the batch corresponding to the expected label of each individual pixel.
- For object detection models, ([`DetrForObjectDetection`]), the model expects a list of dictionaries with a
`class_labels` and `boxes` key where each value of the batch corresponds to the expected label and number of bounding boxes of each individual image.
- For automatic speech recognition models, ([`Wav2Vec2ForCTC`]), the model expects a tensor of dimension `(batch_size,
target_length)` with each value corresponding to the expected label of each individual token.
<Tip>
Each model's labels may be different, so be sure to always check the documentation of each model for more information
about their specific labels!
</Tip>
The base models ([`BertModel`]) do not accept labels, as these are the base transformer models, simply outputting
features.
### large language models (LLM)
A generic term that refers to transformer language models (GPT-3, BLOOM, OPT) that were trained on a large quantity of data. These models also tend to have a large number of learnable parameters (e.g. 175 billion for GPT-3).
## M
### masked language modeling (MLM)
A pretraining task where the model sees a corrupted version of the texts, usually done by
masking some tokens randomly, and has to predict the original text.
### multimodal
A task that combines texts with another kind of inputs (for instance images).
## N
### Natural language generation (NLG)
All tasks related to generating text (for instance, [Write With Transformers](https://transformer.huggingface.co/), translation).
### Natural language processing (NLP)
A generic way to say "deal with texts".
### Natural language understanding (NLU)
All tasks related to understanding what is in a text (for instance classifying the
whole text, individual words).
## P
### pipeline
A pipeline in 🤗 Transformers is an abstraction referring to a series of steps that are executed in a specific order to preprocess and transform data and return a prediction from a model. Some example stages found in a pipeline might be data preprocessing, feature extraction, and normalization.
For more details, see [Pipelines for inference](https://huggingface.co/docs/transformers/pipeline_tutorial).
### pixel values
A tensor of the numerical representations of an image that is passed to a model. The pixel values have a shape of [`batch_size`, `num_channels`, `height`, `width`], and are generated from an image processor.
### pooling
An operation that reduces a matrix into a smaller matrix, either by taking the maximum or average of the pooled dimension(s). Pooling layers are commonly found between convolutional layers to downsample the feature representation.
### position IDs
Contrary to RNNs that have the position of each token embedded within them, transformers are unaware of the position of
each token. Therefore, the position IDs (`position_ids`) are used by the model to identify each token's position in the
list of tokens.
They are an optional parameter. If no `position_ids` are passed to the model, the IDs are automatically created as
absolute positional embeddings.
Absolute positional embeddings are selected in the range `[0, config.max_position_embeddings - 1]`. Some models use
other types of positional embeddings, such as sinusoidal position embeddings or relative position embeddings.
### preprocessing
The task of preparing raw data into a format that can be easily consumed by machine learning models. For example, text is typically preprocessed by tokenization. To gain a better idea of what preprocessing looks like for other input types, check out the [Preprocess](https://huggingface.co/docs/transformers/preprocessing) tutorial.
### pretrained model
A model that has been pretrained on some data (for instance all of Wikipedia). Pretraining methods involve a
self-supervised objective, which can be reading the text and trying to predict the next word (see [causal language
modeling](#causal-language-modeling)) or masking some words and trying to predict them (see [masked language
modeling](#masked-language-modeling-mlm)).
Speech and vision models have their own pretraining objectives. For example, Wav2Vec2 is a speech model pretrained on a contrastive task which requires the model to identify the "true" speech representation from a set of "false" speech representations. On the other hand, BEiT is a vision model pretrained on a masked image modeling task which masks some of the image patches and requires the model to predict the masked patches (similar to the masked language modeling objective).
## R
### recurrent neural network (RNN)
A type of model that uses a loop over a layer to process texts.
### representation learning
A subfield of machine learning which focuses on learning meaningful representations of raw data. Some examples of representation learning techniques include word embeddings, autoencoders, and Generative Adversarial Networks (GANs).
## S
### sampling rate
A measurement in hertz of the number of samples (the audio signal) taken per second. The sampling rate is a result of discretizing a continuous signal such as speech.
### self-attention
Each element of the input finds out which other elements of the input they should attend to.
### self-supervised learning
A category of machine learning techniques in which a model creates its own learning objective from unlabeled data. It differs from [unsupervised learning](#unsupervised-learning) and [supervised learning](#supervised-learning) in that the learning process is supervised, but not explicitly from the user.
One example of self-supervised learning is [masked language modeling](#masked-language-modeling-mlm), where a model is passed sentences with a proportion of its tokens removed and learns to predict the missing tokens.
### semi-supervised learning
A broad category of machine learning training techniques that leverages a small amount of labeled data with a larger quantity of unlabeled data to improve the accuracy of a model, unlike [supervised learning](#supervised-learning) and [unsupervised learning](#unsupervised-learning).
An example of a semi-supervised learning approach is "self-training", in which a model is trained on labeled data, and then used to make predictions on the unlabeled data. The portion of the unlabeled data that the model predicts with the most confidence gets added to the labeled dataset and used to retrain the model.
### sequence-to-sequence (seq2seq)
Models that generate a new sequence from an input, like translation models, or summarization models (such as
[Bart](model_doc/bart) or [T5](model_doc/t5)).
### stride
In [convolution](#convolution) or [pooling](#pooling), the stride refers to the distance the kernel is moved over a matrix. A stride of 1 means the kernel is moved one pixel over at a time, and a stride of 2 means the kernel is moved two pixels over at a time.
### supervised learning
A form of model training that directly uses labeled data to correct and instruct model performance. Data is fed into the model being trained, and its predictions are compared to the known labels. The model updates its weights based on how incorrect its predictions were, and the process is repeated to optimize model performance.
## T
### token
A part of a sentence, usually a word, but can also be a subword (non-common words are often split in subwords) or a
punctuation symbol.
### token Type IDs
Some models' purpose is to do classification on pairs of sentences or question answering.
<Youtube id="0u3ioSwev3s"/>
These require two different sequences to be joined in a single "input_ids" entry, which usually is performed with the
help of special tokens, such as the classifier (`[CLS]`) and separator (`[SEP]`) tokens. For example, the BERT model
builds its two sequence input as such:
```python
>>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
```
We can use our tokenizer to automatically generate such a sentence by passing the two sequences to `tokenizer` as two
arguments (and not a list, like before) like this:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence_a = "HuggingFace is based in NYC"
>>> sequence_b = "Where is HuggingFace based?"
>>> encoded_dict = tokenizer(sequence_a, sequence_b)
>>> decoded = tokenizer.decode(encoded_dict["input_ids"])
```
which will return:
```python
>>> print(decoded)
[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
```
This is enough for some models to understand where one sequence ends and where another begins. However, other models,
such as BERT, also deploy token type IDs (also called segment IDs). They are represented as a binary mask identifying
the two types of sequence in the model.
The tokenizer returns this mask as the "token_type_ids" entry:
```python
>>> encoded_dict["token_type_ids"]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
```
The first sequence, the "context" used for the question, has all its tokens represented by a `0`, whereas the second
sequence, corresponding to the "question", has all its tokens represented by a `1`.
Some models, like [`XLNetModel`] use an additional token represented by a `2`.
### transfer learning
A technique that involves taking a pretrained model and adapting it to a dataset specific to your task. Instead of training a model from scratch, you can leverage knowledge obtained from an existing model as a starting point. This speeds up the learning process and reduces the amount of training data needed.
### transformer
Self-attention based deep learning model architecture.
## U
### unsupervised learning
A form of model training in which data provided to the model is not labeled. Unsupervised learning techniques leverage statistical information of the data distribution to find patterns useful for the task at hand.
docs/source/en/glossary.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Glossary
This glossary defines general machine learning and 🤗 Transformers terms to help you better understand the
documentation.
## A
### attention mask
The attention mask is an optional argument used when batching sequences together.
<Youtube id="M6adb1j2jPI"/>
This argument indicates to the model which tokens should be attended to, and which should not.
For example, consider these two sequences:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence_a = "This is a short sequence."
>>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
>>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
>>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
```
The encoded versions have different lengths:
```python
>>> len(encoded_sequence_a), len(encoded_sequence_b)
(8, 19)
```
Therefore, we can't put them together in the same tensor as-is. The first sequence needs to be padded up to the length
of the second one, or the second one needs to be truncated down to the length of the first one.
In the first case, the list of IDs will be extended by the padding indices. We can pass a list to the tokenizer and ask
it to pad like this:
```python
>>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
```
We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:
```python
>>> padded_sequences["input_ids"]
[[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
```
This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating the
position of the padded indices so that the model does not attend to them. For the [`BertTokenizer`], `1` indicates a
value that should be attended to, while `0` indicates a padded value. This attention mask is in the dictionary returned
by the tokenizer under the key "attention_mask":
```python
>>> padded_sequences["attention_mask"]
[[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
```
### autoencoding models
See [encoder models](#encoder-models) and [masked language modeling](#masked-language-modeling-mlm)
### autoregressive models
See [causal language modeling](#causal-language-modeling) and [decoder models](#decoder-models)
## B
### backbone
The backbone is the network (embeddings and layers) that outputs the raw hidden states or features. It is usually connected to a [head](#head) which accepts the features as its input to make a prediction. For example, [`ViTModel`] is a backbone without a specific head on top. Other models can also use [`VitModel`] as a backbone such as [DPT](model_doc/dpt).
## C
### causal language modeling
A pretraining task where the model reads the texts in order and has to predict the next word. It's usually done by
reading the whole sentence but using a mask inside the model to hide the future tokens at a certain timestep.
### channel
Color images are made up of some combination of values in three channels - red, green, and blue (RGB) - and grayscale images only have one channel. In 🤗 Transformers, the channel can be the first or last dimension of an image's tensor: [`n_channels`, `height`, `width`] or [`height`, `width`, `n_channels`].
### connectionist temporal classification (CTC)
An algorithm which allows a model to learn without knowing exactly how the input and output are aligned; CTC calculates the distribution of all possible outputs for a given input and chooses the most likely output from it. CTC is commonly used in speech recognition tasks because speech doesn't always cleanly align with the transcript for a variety of reasons such as a speaker's different speech rates.
### convolution
A type of layer in a neural network where the input matrix is multiplied element-wise by a smaller matrix (kernel or filter) and the values are summed up in a new matrix. This is known as a convolutional operation which is repeated over the entire input matrix. Each operation is applied to a different segment of the input matrix. Convolutional neural networks (CNNs) are commonly used in computer vision.
## D
### decoder input IDs
This input is specific to encoder-decoder models, and contains the input IDs that will be fed to the decoder. These
inputs should be used for sequence to sequence tasks, such as translation or summarization, and are usually built in a
way specific to each model.
Most encoder-decoder models (BART, T5) create their `decoder_input_ids` on their own from the `labels`. In such models,
passing the `labels` is the preferred way to handle training.
Please check each model's docs to see how they handle these input IDs for sequence to sequence training.
### decoder models
Also referred to as autoregressive models, decoder models involve a pretraining task (called causal language modeling) where the model reads the texts in order and has to predict the next word. It's usually done by
reading the whole sentence with a mask to hide future tokens at a certain timestep.
<Youtube id="d_ixlCubqQw"/>
### deep learning (DL)
Machine learning algorithms which uses neural networks with several layers.
## E
### encoder models
Also known as autoencoding models, encoder models take an input (such as text or images) and transform them into a condensed numerical representation called an embedding. Oftentimes, encoder models are pretrained using techniques like [masked language modeling](#masked-language-modeling-mlm), which masks parts of the input sequence and forces the model to create more meaningful representations.
<Youtube id="H39Z_720T5s"/>
## F
### feature extraction
The process of selecting and transforming raw data into a set of features that are more informative and useful for machine learning algorithms. Some examples of feature extraction include transforming raw text into word embeddings and extracting important features such as edges or shapes from image/video data.
### feed forward chunking
In each residual attention block in transformers the self-attention layer is usually followed by 2 feed forward layers.
The intermediate embedding size of the feed forward layers is often bigger than the hidden size of the model (e.g., for
`bert-base-uncased`).
For an input of size `[batch_size, sequence_length]`, the memory required to store the intermediate feed forward
embeddings `[batch_size, sequence_length, config.intermediate_size]` can account for a large fraction of the memory
use. The authors of [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) noticed that since the
computation is independent of the `sequence_length` dimension, it is mathematically equivalent to compute the output
embeddings of both feed forward layers `[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n`
individually and concat them afterward to `[batch_size, sequence_length, config.hidden_size]` with `n =
sequence_length`, which trades increased computation time against reduced memory use, but yields a mathematically
**equivalent** result.
For models employing the function [`apply_chunking_to_forward`], the `chunk_size` defines the number of output
embeddings that are computed in parallel and thus defines the trade-off between memory and time complexity. If
`chunk_size` is set to 0, no feed forward chunking is done.
### finetuned models
Finetuning is a form of transfer learning which involves taking a pretrained model, freezing its weights, and replacing the output layer with a newly added [model head](#head). The model head is trained on your target dataset.
See the [Fine-tune a pretrained model](https://huggingface.co/docs/transformers/training) tutorial for more details, and learn how to fine-tune models with 🤗 Transformers.
## H
### head
The model head refers to the last layer of a neural network that accepts the raw hidden states and projects them onto a different dimension. There is a different model head for each task. For example:
* [`GPT2ForSequenceClassification`] is a sequence classification head - a linear layer - on top of the base [`GPT2Model`].
* [`ViTForImageClassification`] is an image classification head - a linear layer on top of the final hidden state of the `CLS` token - on top of the base [`ViTModel`].
* [`Wav2Vec2ForCTC`] ia a language modeling head with [CTC](#connectionist-temporal-classification-(CTC)) on top of the base [`Wav2Vec2Model`].
## I
### image patch
Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the `patch_size` - or resolution - of the model in it's configuration.
### inference
Inference is the process of evaluating a model on new data after training is complete. See the [Pipeline for inference](https://huggingface.co/docs/transformers/pipeline_tutorial) tutorial to learn how to perform inference with 🤗 Transformers.
### input IDs
The input ids are often the only required parameters to be passed to the model as input. They are token indices,
numerical representations of tokens building the sequences that will be used as input by the model.
<Youtube id="VFp38yj8h3A"/>
Each tokenizer works differently but the underlying mechanism remains the same. Here's an example using the BERT
tokenizer, which is a [WordPiece](https://arxiv.org/pdf/1609.08144.pdf) tokenizer:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence = "A Titan RTX has 24GB of VRAM"
```
The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.
```python
>>> tokenized_sequence = tokenizer.tokenize(sequence)
```
The tokens are either words or subwords. Here for instance, "VRAM" wasn't in the model vocabulary, so it's been split
in "V", "RA" and "M". To indicate those tokens are not separate words but parts of the same word, a double-hash prefix
is added for "RA" and "M":
```python
>>> print(tokenized_sequence)
['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
```
These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding
the sentence to the tokenizer, which leverages the Rust implementation of [🤗
Tokenizers](https://github.com/huggingface/tokenizers) for peak performance.
```python
>>> inputs = tokenizer(sequence)
```
The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
token indices are under the key `input_ids`:
```python
>>> encoded_sequence = inputs["input_ids"]
>>> print(encoded_sequence)
[101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
```
Note that the tokenizer automatically adds "special tokens" (if the associated model relies on them) which are special
IDs the model sometimes uses.
If we decode the previous sequence of ids,
```python
>>> decoded_sequence = tokenizer.decode(encoded_sequence)
```
we will see
```python
>>> print(decoded_sequence)
[CLS] A Titan RTX has 24GB of VRAM [SEP]
```
because this is the way a [`BertModel`] is going to expect its inputs.
## L
### labels
The labels are an optional argument which can be passed in order for the model to compute the loss itself. These labels
should be the expected prediction of the model: it will use the standard loss in order to compute the loss between its
predictions and the expected value (the label).
These labels are different according to the model head, for example:
- For sequence classification models, ([`BertForSequenceClassification`]), the model expects a tensor of dimension
`(batch_size)` with each value of the batch corresponding to the expected label of the entire sequence.
- For token classification models, ([`BertForTokenClassification`]), the model expects a tensor of dimension
`(batch_size, seq_length)` with each value corresponding to the expected label of each individual token.
- For masked language modeling, ([`BertForMaskedLM`]), the model expects a tensor of dimension `(batch_size,
seq_length)` with each value corresponding to the expected label of each individual token: the labels being the token
ID for the masked token, and values to be ignored for the rest (usually -100).
- For sequence to sequence tasks, ([`BartForConditionalGeneration`], [`MBartForConditionalGeneration`]), the model
expects a tensor of dimension `(batch_size, tgt_seq_length)` with each value corresponding to the target sequences
associated with each input sequence. During training, both BART and T5 will make the appropriate
`decoder_input_ids` and decoder attention masks internally. They usually do not need to be supplied. This does not
apply to models leveraging the Encoder-Decoder framework.
- For image classification models, ([`ViTForImageClassification`]), the model expects a tensor of dimension
`(batch_size)` with each value of the batch corresponding to the expected label of each individual image.
- For semantic segmentation models, ([`SegformerForSemanticSegmentation`]), the model expects a tensor of dimension
`(batch_size, height, width)` with each value of the batch corresponding to the expected label of each individual pixel.
- For object detection models, ([`DetrForObjectDetection`]), the model expects a list of dictionaries with a
`class_labels` and `boxes` key where each value of the batch corresponds to the expected label and number of bounding boxes of each individual image.
- For automatic speech recognition models, ([`Wav2Vec2ForCTC`]), the model expects a tensor of dimension `(batch_size,
target_length)` with each value corresponding to the expected label of each individual token.
<Tip>
Each model's labels may be different, so be sure to always check the documentation of each model for more information
about their specific labels!
</Tip>
The base models ([`BertModel`]) do not accept labels, as these are the base transformer models, simply outputting
features.
### large language models (LLM)
A generic term that refers to transformer language models (GPT-3, BLOOM, OPT) that were trained on a large quantity of data. These models also tend to have a large number of learnable parameters (e.g. 175 billion for GPT-3).
## M
### masked language modeling (MLM)
A pretraining task where the model sees a corrupted version of the texts, usually done by
masking some tokens randomly, and has to predict the original text.
### multimodal
A task that combines texts with another kind of inputs (for instance images).
## N
### Natural language generation (NLG)
All tasks related to generating text (for instance, [Write With Transformers](https://transformer.huggingface.co/), translation).
### Natural language processing (NLP)
A generic way to say "deal with texts".
### Natural language understanding (NLU)
All tasks related to understanding what is in a text (for instance classifying the
whole text, individual words).
## P
### pipeline
A pipeline in 🤗 Transformers is an abstraction referring to a series of steps that are executed in a specific order to preprocess and transform data and return a prediction from a model. Some example stages found in a pipeline might be data preprocessing, feature extraction, and normalization.
For more details, see [Pipelines for inference](https://huggingface.co/docs/transformers/pipeline_tutorial).
### pixel values
A tensor of the numerical representations of an image that is passed to a model. The pixel values have a shape of [`batch_size`, `num_channels`, `height`, `width`], and are generated from an image processor.
### pooling
An operation that reduces a matrix into a smaller matrix, either by taking the maximum or average of the pooled dimension(s). Pooling layers are commonly found between convolutional layers to downsample the feature representation.
### position IDs
Contrary to RNNs that have the position of each token embedded within them, transformers are unaware of the position of
each token. Therefore, the position IDs (`position_ids`) are used by the model to identify each token's position in the
list of tokens.
They are an optional parameter. If no `position_ids` are passed to the model, the IDs are automatically created as
absolute positional embeddings.
Absolute positional embeddings are selected in the range `[0, config.max_position_embeddings - 1]`. Some models use
other types of positional embeddings, such as sinusoidal position embeddings or relative position embeddings.
### preprocessing
The task of preparing raw data into a format that can be easily consumed by machine learning models. For example, text is typically preprocessed by tokenization. To gain a better idea of what preprocessing looks like for other input types, check out the [Preprocess](https://huggingface.co/docs/transformers/preprocessing) tutorial.
### pretrained model
A model that has been pretrained on some data (for instance all of Wikipedia). Pretraining methods involve a
self-supervised objective, which can be reading the text and trying to predict the next word (see [causal language
modeling](#causal-language-modeling)) or masking some words and trying to predict them (see [masked language
modeling](#masked-language-modeling-mlm)).
Speech and vision models have their own pretraining objectives. For example, Wav2Vec2 is a speech model pretrained on a contrastive task which requires the model to identify the "true" speech representation from a set of "false" speech representations. On the other hand, BEiT is a vision model pretrained on a masked image modeling task which masks some of the image patches and requires the model to predict the masked patches (similar to the masked language modeling objective).
## R
### recurrent neural network (RNN)
A type of model that uses a loop over a layer to process texts.
### representation learning
A subfield of machine learning which focuses on learning meaningful representations of raw data. Some examples of representation learning techniques include word embeddings, autoencoders, and Generative Adversarial Networks (GANs).
## S
### sampling rate
A measurement in hertz of the number of samples (the audio signal) taken per second. The sampling rate is a result of discretizing a continuous signal such as speech.
### self-attention
Each element of the input finds out which other elements of the input they should attend to.
### self-supervised learning
A category of machine learning techniques in which a model creates its own learning objective from unlabeled data. It differs from [unsupervised learning](#unsupervised-learning) and [supervised learning](#supervised-learning) in that the learning process is supervised, but not explicitly from the user.
One example of self-supervised learning is [masked language modeling](#masked-language-modeling-mlm), where a model is passed sentences with a proportion of its tokens removed and learns to predict the missing tokens.
### semi-supervised learning
A broad category of machine learning training techniques that leverages a small amount of labeled data with a larger quantity of unlabeled data to improve the accuracy of a model, unlike [supervised learning](#supervised-learning) and [unsupervised learning](#unsupervised-learning).
An example of a semi-supervised learning approach is "self-training", in which a model is trained on labeled data, and then used to make predictions on the unlabeled data. The portion of the unlabeled data that the model predicts with the most confidence gets added to the labeled dataset and used to retrain the model.
### sequence-to-sequence (seq2seq)
Models that generate a new sequence from an input, like translation models, or summarization models (such as
[Bart](model_doc/bart) or [T5](model_doc/t5)).
### stride
In [convolution](#convolution) or [pooling](#pooling), the stride refers to the distance the kernel is moved over a matrix. A stride of 1 means the kernel is moved one pixel over at a time, and a stride of 2 means the kernel is moved two pixels over at a time.
### supervised learning
A form of model training that directly uses labeled data to correct and instruct model performance. Data is fed into the model being trained, and its predictions are compared to the known labels. The model updates its weights based on how incorrect its predictions were, and the process is repeated to optimize model performance.
## T
### token
A part of a sentence, usually a word, but can also be a subword (non-common words are often split in subwords) or a
punctuation symbol.
### token Type IDs
Some models' purpose is to do classification on pairs of sentences or question answering.
<Youtube id="0u3ioSwev3s"/>
These require two different sequences to be joined in a single "input_ids" entry, which usually is performed with the
help of special tokens, such as the classifier (`[CLS]`) and separator (`[SEP]`) tokens. For example, the BERT model
builds its two sequence input as such:
```python
>>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
```
We can use our tokenizer to automatically generate such a sentence by passing the two sequences to `tokenizer` as two
arguments (and not a list, like before) like this:
```python
>>> from transformers import BertTokenizer
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
>>> sequence_a = "HuggingFace is based in NYC"
>>> sequence_b = "Where is HuggingFace based?"
>>> encoded_dict = tokenizer(sequence_a, sequence_b)
>>> decoded = tokenizer.decode(encoded_dict["input_ids"])
```
which will return:
```python
>>> print(decoded)
[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
```
This is enough for some models to understand where one sequence ends and where another begins. However, other models,
such as BERT, also deploy token type IDs (also called segment IDs). They are represented as a binary mask identifying
the two types of sequence in the model.
The tokenizer returns this mask as the "token_type_ids" entry:
```python
>>> encoded_dict["token_type_ids"]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
```
The first sequence, the "context" used for the question, has all its tokens represented by a `0`, whereas the second
sequence, corresponding to the "question", has all its tokens represented by a `1`.
Some models, like [`XLNetModel`] use an additional token represented by a `2`.
### transfer learning
A technique that involves taking a pretrained model and adapting it to a dataset specific to your task. Instead of training a model from scratch, you can leverage knowledge obtained from an existing model as a starting point. This speeds up the learning process and reduces the amount of training data needed.
### transformer
Self-attention based deep learning model architecture.
## U
### unsupervised learning
A form of model training in which data provided to the model is not labeled. Unsupervised learning techniques leverage statistical information of the data distribution to find patterns useful for the task at hand.
docs/source/en/hpo_train.md 0 → 100644
View file @ eb849f66
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Hyperparameter Search using Trainer API
🤗 Transformers provides a [`Trainer`] class optimized for training 🤗 Transformers models, making it easier to start training without manually writing your own training loop. The [`Trainer`] provides API for hyperparameter search. This doc shows how to enable it in example.
## Hyperparameter Search backend
[`Trainer`] supports four hyperparameter search backends currently:
[optuna](https://optuna.org/), [sigopt](https://sigopt.com/), [raytune](https://docs.ray.io/en/latest/tune/index.html) and [wandb](https://wandb.ai/site/sweeps).
you should install them before using them as the hyperparameter search backend
```bash
pip install optuna/sigopt/wandb/ray[tune]
```
## How to enable Hyperparameter search in example
Define the hyperparameter search space, different backends need different format.
For sigopt, see sigopt [object_parameter](https://docs.sigopt.com/ai-module-api-references/api_reference/objects/object_parameter), it's like following:
```py
>>> def sigopt_hp_space(trial):
... return [
... {"bounds": {"min": 1e-6, "max": 1e-4}, "name": "learning_rate", "type": "double"},
... {
... "categorical_values": ["16", "32", "64", "128"],
... "name": "per_device_train_batch_size",
... "type": "categorical",
... },
... ]
```
For optuna, see optuna [object_parameter](https://optuna.readthedocs.io/en/stable/tutorial/10_key_features/002_configurations.html#sphx-glr-tutorial-10-key-features-002-configurations-py), it's like following:
```py
>>> def optuna_hp_space(trial):
... return {
... "learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
... "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [16, 32, 64, 128]),
... }
```
For raytune, see raytune [object_parameter](https://docs.ray.io/en/latest/tune/api_docs/search_space.html), it's like following:
```py
>>> def ray_hp_space(trial):
... return {
... "learning_rate": tune.loguniform(1e-6, 1e-4),
... "per_device_train_batch_size": tune.choice([16, 32, 64, 128]),
... }
```
For wandb, see wandb [object_parameter](https://docs.wandb.ai/guides/sweeps/configuration), it's like following:
```py
>>> def wandb_hp_space(trial):
... return {
... "method": "random",
... "metric": {"name": "objective", "goal": "minimize"},
... "parameters": {
... "learning_rate": {"distribution": "uniform", "min": 1e-6, "max": 1e-4},
... "per_device_train_batch_size": {"values": [16, 32, 64, 128]},
... },
... }
```
Define a `model_init` function and pass it to the [`Trainer`], as an example:
```py
>>> def model_init(trial):
... return AutoModelForSequenceClassification.from_pretrained(
... model_args.model_name_or_path,
... from_tf=bool(".ckpt" in model_args.model_name_or_path),
... config=config,
... cache_dir=model_args.cache_dir,
... revision=model_args.model_revision,
... use_auth_token=True if model_args.use_auth_token else None,
... )
```
Create a [`Trainer`] with your `model_init` function, training arguments, training and test datasets, and evaluation function:
```py
>>> trainer = Trainer(
... model=None,
... args=training_args,
... train_dataset=small_train_dataset,
... eval_dataset=small_eval_dataset,
... compute_metrics=compute_metrics,
... tokenizer=tokenizer,
... model_init=model_init,
... data_collator=data_collator,
... )
```
Call hyperparameter search, get the best trial parameters, backend could be `"optuna"`/`"sigopt"`/`"wandb"`/`"ray"`. direction can be`"minimize"` or `"maximize"`, which indicates whether to optimize greater or lower objective.
You could define your own compute_objective function, if not defined, the default compute_objective will be called, and the sum of eval metric like f1 is returned as objective value.
```py
>>> best_trial = trainer.hyperparameter_search(
... direction="maximize",
... backend="optuna",
... hp_space=optuna_hp_space,
... n_trials=20,
... compute_objective=compute_objective,
... )
```
## Hyperparameter search For DDP finetune
Currently, Hyperparameter search for DDP is enabled for optuna and sigopt. Only the rank-zero process will generate the search trial and pass the argument to other ranks.
\ No newline at end of file
docs/source/en/hpo_train.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-->
# Hyperparameter Search using Trainer API
🤗 Transformers provides a [`Trainer`] class optimized for training 🤗 Transformers models, making it easier to start training without manually writing your own training loop. The [`Trainer`] provides API for hyperparameter search. This doc shows how to enable it in example.
## Hyperparameter Search backend
[`Trainer`] supports four hyperparameter search backends currently:
[optuna](https://optuna.org/), [sigopt](https://sigopt.com/), [raytune](https://docs.ray.io/en/latest/tune/index.html) and [wandb](https://wandb.ai/site/sweeps).
you should install them before using them as the hyperparameter search backend
```bash
pip install optuna/sigopt/wandb/ray[tune]
```
## How to enable Hyperparameter search in example
Define the hyperparameter search space, different backends need different format.
For sigopt, see sigopt [object_parameter](https://docs.sigopt.com/ai-module-api-references/api_reference/objects/object_parameter), it's like following:
```py
>>> def sigopt_hp_space(trial):
... return [
... {"bounds": {"min": 1e-6, "max": 1e-4}, "name": "learning_rate", "type": "double"},
... {
... "categorical_values": ["16", "32", "64", "128"],
... "name": "per_device_train_batch_size",
... "type": "categorical",
... },
... ]
```
For optuna, see optuna [object_parameter](https://optuna.readthedocs.io/en/stable/tutorial/10_key_features/002_configurations.html#sphx-glr-tutorial-10-key-features-002-configurations-py), it's like following:
```py
>>> def optuna_hp_space(trial):
... return {
... "learning_rate": trial.suggest_float("learning_rate", 1e-6, 1e-4, log=True),
... "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [16, 32, 64, 128]),
... }
```
For raytune, see raytune [object_parameter](https://docs.ray.io/en/latest/tune/api_docs/search_space.html), it's like following:
```py
>>> def ray_hp_space(trial):
... return {
... "learning_rate": tune.loguniform(1e-6, 1e-4),
... "per_device_train_batch_size": tune.choice([16, 32, 64, 128]),
... }
```
For wandb, see wandb [object_parameter](https://docs.wandb.ai/guides/sweeps/configuration), it's like following:
```py
>>> def wandb_hp_space(trial):
... return {
... "method": "random",
... "metric": {"name": "objective", "goal": "minimize"},
... "parameters": {
... "learning_rate": {"distribution": "uniform", "min": 1e-6, "max": 1e-4},
... "per_device_train_batch_size": {"values": [16, 32, 64, 128]},
... },
... }
```
Define a `model_init` function and pass it to the [`Trainer`], as an example:
```py
>>> def model_init(trial):
... return AutoModelForSequenceClassification.from_pretrained(
... model_args.model_name_or_path,
... from_tf=bool(".ckpt" in model_args.model_name_or_path),
... config=config,
... cache_dir=model_args.cache_dir,
... revision=model_args.model_revision,
... use_auth_token=True if model_args.use_auth_token else None,
... )
```
Create a [`Trainer`] with your `model_init` function, training arguments, training and test datasets, and evaluation function:
```py
>>> trainer = Trainer(
... model=None,
... args=training_args,
... train_dataset=small_train_dataset,
... eval_dataset=small_eval_dataset,
... compute_metrics=compute_metrics,
... tokenizer=tokenizer,
... model_init=model_init,
... data_collator=data_collator,
... )
```
Call hyperparameter search, get the best trial parameters, backend could be `"optuna"`/`"sigopt"`/`"wandb"`/`"ray"`. direction can be`"minimize"` or `"maximize"`, which indicates whether to optimize greater or lower objective.
You could define your own compute_objective function, if not defined, the default compute_objective will be called, and the sum of eval metric like f1 is returned as objective value.
```py
>>> best_trial = trainer.hyperparameter_search(
... direction="maximize",
... backend="optuna",
... hp_space=optuna_hp_space,
... n_trials=20,
... compute_objective=compute_objective,
... )
```
## Hyperparameter search For DDP finetune
Currently, Hyperparameter search for DDP is enabled for optuna and sigopt. Only the rank-zero process will generate the search trial and pass the argument to other ranks.
\ No newline at end of file
docs/source/en/index.md 0 → 100644
View file @ eb849f66
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# 🤗 Transformers
State-of-the-art Machine Learning for [PyTorch](https://pytorch.org/), [TensorFlow](https://www.tensorflow.org/), and [JAX](https://jax.readthedocs.io/en/latest/).
🤗 Transformers provides APIs and tools to easily download and train state-of-the-art pretrained models. Using pretrained models can reduce your compute costs, carbon footprint, and save you the time and resources required to train a model from scratch. These models support common tasks in different modalities, such as:
📝 **Natural Language Processing**: text classification, named entity recognition, question answering, language modeling, summarization, translation, multiple choice, and text generation.<br>
🖼️ **Computer Vision**: image classification, object detection, and segmentation.<br>
🗣️ **Audio**: automatic speech recognition and audio classification.<br>
🐙 **Multimodal**: table question answering, optical character recognition, information extraction from scanned documents, video classification, and visual question answering.
🤗 Transformers support framework interoperability between PyTorch, TensorFlow, and JAX. This provides the flexibility to use a different framework at each stage of a model's life; train a model in three lines of code in one framework, and load it for inference in another. Models can also be exported to a format like ONNX and TorchScript for deployment in production environments.
Join the growing community on the [Hub](https://huggingface.co/models), [forum](https://discuss.huggingface.co/), or [Discord](https://discord.com/invite/JfAtkvEtRb) today!
## If you are looking for custom support from the Hugging Face team
<a target="_blank" href="https://huggingface.co/support">
<img alt="HuggingFace Expert Acceleration Program" src="https://cdn-media.huggingface.co/marketing/transformers/new-support-improved.png" style="width: 100%; max-width: 600px; border: 1px solid #eee; border-radius: 4px; box-shadow: 0 1px 2px 0 rgba(0, 0, 0, 0.05);">
</a>
## Contents
The documentation is organized into five sections:
- **GET STARTED** provides a quick tour of the library and installation instructions to get up and running.
- **TUTORIALS** are a great place to start if you're a beginner. This section will help you gain the basic skills you need to start using the library.
- **HOW-TO GUIDES** show you how to achieve a specific goal, like finetuning a pretrained model for language modeling or how to write and share a custom model.
- **CONCEPTUAL GUIDES** offers more discussion and explanation of the underlying concepts and ideas behind models, tasks, and the design philosophy of 🤗 Transformers.
- **API** describes all classes and functions:
- **MAIN CLASSES** details the most important classes like configuration, model, tokenizer, and pipeline.
- **MODELS** details the classes and functions related to each model implemented in the library.
- **INTERNAL HELPERS** details utility classes and functions used internally.
### Supported models
<!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
1. **[Autoformer](model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
1. **[BEiT](model_doc/beit)** (from Microsoft) released with the paper [BEiT: BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254) by Hangbo Bao, Li Dong, Furu Wei.
1. **[BERT](model_doc/bert)** (from Google) released 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.
1. **[BERT For Sequence Generation](model_doc/bert-generation)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[BERTweet](model_doc/bertweet)** (from VinAI Research) released with the paper [BERTweet: A pre-trained language model for English Tweets](https://aclanthology.org/2020.emnlp-demos.2/) by Dat Quoc Nguyen, Thanh Vu and Anh Tuan Nguyen.
1. **[BigBird-Pegasus](model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BigBird-RoBERTa](model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BioGpt](model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
1. **[BiT](model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BLIP](model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
1. **[BLIP-2](model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
1. **[Chinese-CLIP](model_doc/chinese_clip)** (from OFA-Sys) released with the paper [Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese](https://arxiv.org/abs/2211.01335) by An Yang, Junshu Pan, Junyang Lin, Rui Men, Yichang Zhang, Jingren Zhou, Chang Zhou.
1. **[CLAP](model_doc/clap)** (from LAION-AI) released with the paper [Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation](https://arxiv.org/abs/2211.06687) by Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, Shlomo Dubnov.
1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
1. **[CLIPSeg](model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie.
1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
1. **[CPM-Ant](model_doc/cpmant)** (from OpenBMB) released by the [OpenBMB](https://www.openbmb.org/).
1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang.
1. **[Data2Vec](model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec: A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[Decision Transformer](model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
1. **[Deformable DETR](model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
1. **[DePlot](model_doc/deplot)** (from Google AI) released with the paper [DePlot: One-shot visual language reasoning by plot-to-table translation](https://arxiv.org/abs/2212.10505) by Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, Yasemin Altun.
1. **[DETA](model_doc/deta)** (from The University of Texas at Austin) released with the paper [NMS Strikes Back](https://arxiv.org/abs/2212.06137) by Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, Philipp Krähenbühl.
1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
1. **[DiNAT](model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
1. **[Donut](model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
1. **[DPR](model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
1. **[DPT](master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
1. **[EfficientFormer](model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
1. **[EfficientNet](model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
1. **[EnCodec](model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[ERNIE](model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
1. **[ErnieM](model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.
1. **[ESM](model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2 and ESMFold** were released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives.
1. **[FLAN-T5](model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
1. **[FLAN-UL2](model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
1. **[FocalNet](model_doc/focalnet)** (from Microsoft Research) released with the paper [Focal Modulation Networks](https://arxiv.org/abs/2203.11926) by Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, Jianfeng Gao.
1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
1. **[GIT](model_doc/git)** (from Microsoft Research) released with the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) by Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang.
1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
1. **[GPT NeoX Japanese](model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori.
1. **[GPT-2](model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
1. **[GPT-J](model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
1. **[GPT-Sw3](model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
1. **[GPTBigCode](model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
1. **[Graphormer](model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
1. **[ImageGPT](model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[Jukebox](model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[LayoutLM](model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
1. **[LayoutLMv2](model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
1. **[LayoutLMv3](model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.
1. **[LayoutXLM](model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
1. **[LED](model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LeViT](model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
1. **[LiLT](model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
1. **[LLaMA](model_doc/llama)** (from The FAIR team of Meta AI) released with the paper [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) by Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample.
1. **[Longformer](model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LongT5](model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
1. **[LUKE](model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
1. **[LXMERT](model_doc/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
1. **[M-CTC-T](model_doc/mctct)** (from Facebook) released with the paper [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) by Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert.
1. **[M2M100](model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
1. **[MarianMT](model_doc/marian)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
1. **[MarkupLM](model_doc/markuplm)** (from Microsoft Research Asia) released with the paper [MarkupLM: Pre-training of Text and Markup Language for Visually-rich Document Understanding](https://arxiv.org/abs/2110.08518) by Junlong Li, Yiheng Xu, Lei Cui, Furu Wei.
1. **[Mask2Former](model_doc/mask2former)** (from FAIR and UIUC) released with the paper [Masked-attention Mask Transformer for Universal Image Segmentation](https://arxiv.org/abs/2112.01527) by Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar.
1. **[MaskFormer](model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov.
1. **[MatCha](model_doc/matcha)** (from Google AI) released with the paper [MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering](https://arxiv.org/abs/2212.09662) by Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, Julian Martin Eisenschlos.
1. **[mBART](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
1. **[mBART-50](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
1. **[MEGA](model_doc/mega)** (from Meta/USC/CMU/SJTU) released with the paper [Mega: Moving Average Equipped Gated Attention](https://arxiv.org/abs/2209.10655) by Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer.
1. **[Megatron-BERT](model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[Megatron-GPT2](model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[MGP-STR](model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
1. **[mLUKE](model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
1. **[MMS](model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
1. **[MobileNetV1](model_doc/mobilenet_v1)** (from Google Inc.) released with the paper [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861) by Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam.
1. **[MobileNetV2](model_doc/mobilenet_v2)** (from Google Inc.) released with the paper [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/abs/1801.04381) by Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen.
1. **[MobileViT](model_doc/mobilevit)** (from Apple) released with the paper [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) by Sachin Mehta and Mohammad Rastegari.
1. **[MobileViTV2](model_doc/mobilevitv2)** (from Apple) released with the paper [Separable Self-attention for Mobile Vision Transformers](https://arxiv.org/abs/2206.02680) by Sachin Mehta and Mohammad Rastegari.
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
1. **[NAT](model_doc/nat)** (from SHI Labs) released with the paper [Neighborhood Attention Transformer](https://arxiv.org/abs/2204.07143) by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
1. **[Nezha](model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[NLLB-MOE](model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
1. **[OneFormer](model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
1. **[OpenLlama](model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
1. **[OWL-ViT](model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
1. **[PEGASUS-X](model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
1. **[Pix2Struct](model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
1. **[PoolFormer](model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
1. **[ProphetNet](model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[QDQBert](model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
1. **[RAG](model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
1. **[REALM](model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
1. **[Reformer](model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
1. **[RegNet](model_doc/regnet)** (from META Platforms) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
1. **[RemBERT](model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
1. **[ResNet](model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
1. **[RoBERTa](model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
1. **[RoBERTa-PreLayerNorm](model_doc/roberta-prelayernorm)** (from Facebook) released with the paper [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://arxiv.org/abs/1904.01038) by Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, Michael Auli.
1. **[RoCBert](model_doc/roc_bert)** (from WeChatAI) released with the paper [RoCBert: Robust Chinese Bert with Multimodal Contrastive Pretraining](https://aclanthology.org/2022.acl-long.65.pdf) by HuiSu, WeiweiShi, XiaoyuShen, XiaoZhou, TuoJi, JiaruiFang, JieZhou.
1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[RWKV](model_doc/rwkv)** (from Bo Peng), released on [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
1. **[Segment Anything](model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SpeechT5](model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei.
1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau.
1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy.
1. **[SqueezeBERT](model_doc/squeezebert)** (from Berkeley) released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
1. **[SwiftFormer](model_doc/swiftformer)** (from MBZUAI) released with the paper [SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications](https://arxiv.org/abs/2303.15446) by Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan.
1. **[Swin Transformer](model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo.
1. **[Swin Transformer V2](model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo.
1. **[Swin2SR](model_doc/swin2sr)** (from University of Würzburg) released with the paper [Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration](https://arxiv.org/abs/2209.11345) by Marcos V. Conde, Ui-Jin Choi, Maxime Burchi, Radu Timofte.
1. **[SwitchTransformers](model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer.
1. **[T5](model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[T5v1.1](model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[Table Transformer](model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
1. **[Time Series Transformer](model_doc/time_series_transformer)** (from HuggingFace).
1. **[TimeSformer](model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
1. **[TVLT](model_doc/tvlt)** (from UNC Chapel Hill) released with the paper [TVLT: Textless Vision-Language Transformer](https://arxiv.org/abs/2209.14156) by Zineng Tang, Jaemin Cho, Yixin Nie, Mohit Bansal.
1. **[UL2](model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
1. **[UniSpeech](model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
1. **[UniSpeechSat](model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
1. **[UPerNet](model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
1. **[VAN](model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
1. **[VideoMAE](model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
1. **[ViLT](model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
1. **[ViT Hybrid](model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
1. **[ViTMSN](model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
1. **[Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
1. **[Wav2Vec2Phoneme](model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
1. **[WavLM](model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
1. **[Whisper](model_doc/whisper)** (from OpenAI) released with the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
1. **[X-CLIP](model_doc/xclip)** (from Microsoft Research) released with the paper [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) by Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling.
1. **[X-MOD](model_doc/xmod)** (from Meta AI) released with the paper [Lifting the Curse of Multilinguality by Pre-training Modular Transformers](http://dx.doi.org/10.18653/v1/2022.naacl-main.255) by Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, Mikel Artetxe.
1. **[XGLM](model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
1. **[XLM](model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
1. **[XLM-ProphetNet](model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[XLM-RoBERTa](model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
1. **[XLM-RoBERTa-XL](model_doc/xlm-roberta-xl)** (from Facebook AI), released together with the paper [Larger-Scale Transformers for Multilingual Masked Language Modeling](https://arxiv.org/abs/2105.00572) by Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, Alexis Conneau.
1. **[XLM-V](model_doc/xlm-v)** (from Meta AI) released with the paper [XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models](https://arxiv.org/abs/2301.10472) by Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, Madian Khabsa.
1. **[XLNet](model_doc/xlnet)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
1. **[XLS-R](model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
1. **[XLSR-Wav2Vec2](model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
1. **[YOLOS](model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
1. **[YOSO](model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
### Supported frameworks
The table below represents the current support in the library for each of those models, whether they have a Python
tokenizer (called "slow"). A "fast" tokenizer backed by the 🤗 Tokenizers library, whether they have support in Jax (via
Flax), PyTorch, and/or TensorFlow.
<!--This table is updated automatically from the auto modules with _make fix-copies_. Do not update manually!-->
| Model | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support |
|:-----------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:|
| ALBERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| ALIGN | ❌ | ❌ | ✅ | ❌ | ❌ |
| AltCLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| Audio Spectrogram Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Autoformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| BART | ✅ | ✅ | ✅ | ✅ | ✅ |
| BEiT | ❌ | ❌ | ✅ | ❌ | ✅ |
| BERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| Bert Generation | ✅ | ❌ | ✅ | ❌ | ❌ |
| BigBird | ✅ | ✅ | ✅ | ❌ | ✅ |
| BigBird-Pegasus | ❌ | ❌ | ✅ | ❌ | ❌ |
| BioGpt | ✅ | ❌ | ✅ | ❌ | ❌ |
| BiT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Blenderbot | ✅ | ✅ | ✅ | ✅ | ✅ |
| BlenderbotSmall | ✅ | ✅ | ✅ | ✅ | ✅ |
| BLIP | ❌ | ❌ | ✅ | ✅ | ❌ |
| BLIP-2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| BLOOM | ❌ | ✅ | ✅ | ❌ | ❌ |
| BridgeTower | ❌ | ❌ | ✅ | ❌ | ❌ |
| CamemBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| CANINE | ✅ | ❌ | ✅ | ❌ | ❌ |
| Chinese-CLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| CLAP | ❌ | ❌ | ✅ | ❌ | ❌ |
| CLIP | ✅ | ✅ | ✅ | ✅ | ✅ |
| CLIPSeg | ❌ | ❌ | ✅ | ❌ | ❌ |
| CodeGen | ✅ | ✅ | ✅ | ❌ | ❌ |
| Conditional DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ |
| ConvNeXTV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| CPM-Ant | ✅ | ❌ | ✅ | ❌ | ❌ |
| CTRL | ✅ | ❌ | ✅ | ✅ | ❌ |
| CvT | ❌ | ❌ | ✅ | ✅ | ❌ |
| Data2VecAudio | ❌ | ❌ | ✅ | ❌ | ❌ |
| Data2VecText | ❌ | ❌ | ✅ | ❌ | ❌ |
| Data2VecVision | ❌ | ❌ | ✅ | ✅ | ❌ |
| DeBERTa | ✅ | ✅ | ✅ | ✅ | ❌ |
| DeBERTa-v2 | ✅ | ✅ | ✅ | ✅ | ❌ |
| Decision Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Deformable DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DeiT | ❌ | ❌ | ✅ | ✅ | ❌ |
| DETA | ❌ | ❌ | ✅ | ❌ | ❌ |
| DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DiNAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| DistilBERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| DonutSwin | ❌ | ❌ | ✅ | ❌ | ❌ |
| DPR | ✅ | ✅ | ✅ | ✅ | ❌ |
| DPT | ❌ | ❌ | ✅ | ❌ | ❌ |
| EfficientFormer | ❌ | ❌ | ✅ | ✅ | ❌ |
| EfficientNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| ELECTRA | ✅ | ✅ | ✅ | ✅ | ✅ |
| EnCodec | ❌ | ❌ | ✅ | ❌ | ❌ |
| Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| ERNIE | ❌ | ❌ | ✅ | ❌ | ❌ |
| ErnieM | ✅ | ❌ | ✅ | ❌ | ❌ |
| ESM | ✅ | ❌ | ✅ | ✅ | ❌ |
| FairSeq Machine-Translation | ✅ | ❌ | ✅ | ❌ | ❌ |
| FlauBERT | ✅ | ❌ | ✅ | ✅ | ❌ |
| FLAVA | ❌ | ❌ | ✅ | ❌ | ❌ |
| FNet | ✅ | ✅ | ✅ | ❌ | ❌ |
| FocalNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ |
| GIT | ❌ | ❌ | ✅ | ❌ | ❌ |
| GLPN | ❌ | ❌ | ✅ | ❌ | ❌ |
| GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ |
| GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ |
| GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
| GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ |
| GPT-Sw3 | ✅ | ✅ | ✅ | ✅ | ✅ |
| GPTBigCode | ❌ | ❌ | ✅ | ❌ | ❌ |
| GPTSAN-japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
| Graphormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| GroupViT | ❌ | ❌ | ✅ | ✅ | ❌ |
| Hubert | ❌ | ❌ | ✅ | ✅ | ❌ |
| I-BERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| ImageGPT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Informer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Jukebox | ✅ | ❌ | ✅ | ❌ | ❌ |
| LayoutLM | ✅ | ✅ | ✅ | ✅ | ❌ |
| LayoutLMv2 | ✅ | ✅ | ✅ | ❌ | ❌ |
| LayoutLMv3 | ✅ | ✅ | ✅ | ✅ | ❌ |
| LED | ✅ | ✅ | ✅ | ✅ | ❌ |
| LeViT | ❌ | ❌ | ✅ | ❌ | ❌ |
| LiLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| LLaMA | ✅ | ✅ | ✅ | ❌ | ❌ |
| Longformer | ✅ | ✅ | ✅ | ✅ | ❌ |
| LongT5 | ❌ | ❌ | ✅ | ❌ | ✅ |
| LUKE | ✅ | ❌ | ✅ | ❌ | ❌ |
| LXMERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| M-CTC-T | ❌ | ❌ | ✅ | ❌ | ❌ |
| M2M100 | ✅ | ❌ | ✅ | ❌ | ❌ |
| Marian | ✅ | ❌ | ✅ | ✅ | ✅ |
| MarkupLM | ✅ | ✅ | ✅ | ❌ | ❌ |
| Mask2Former | ❌ | ❌ | ✅ | ❌ | ❌ |
| MaskFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| MaskFormerSwin | ❌ | ❌ | ❌ | ❌ | ❌ |
| mBART | ✅ | ✅ | ✅ | ✅ | ✅ |
| MEGA | ❌ | ❌ | ✅ | ❌ | ❌ |
| Megatron-BERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| MGP-STR | ✅ | ❌ | ✅ | ❌ | ❌ |
| MobileBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| MobileNetV1 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MobileNetV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MobileViT | ❌ | ❌ | ✅ | ✅ | ❌ |
| MobileViTV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MPNet | ✅ | ✅ | ✅ | ✅ | ❌ |
| MT5 | ✅ | ✅ | ✅ | ✅ | ✅ |
| MVP | ✅ | ✅ | ✅ | ❌ | ❌ |
| NAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nezha | ❌ | ❌ | ✅ | ❌ | ❌ |
| NLLB-MOE | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| OneFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ |
| OpenAI GPT-2 | ✅ | ✅ | ✅ | ✅ | ✅ |
| OpenLlama | ❌ | ❌ | ✅ | ❌ | ❌ |
| OPT | ❌ | ❌ | ✅ | ✅ | ✅ |
| OWL-ViT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Pegasus | ✅ | ✅ | ✅ | ✅ | ✅ |
| PEGASUS-X | ❌ | ❌ | ✅ | ❌ | ❌ |
| Perceiver | ✅ | ❌ | ✅ | ❌ | ❌ |
| Pix2Struct | ❌ | ❌ | ✅ | ❌ | ❌ |
| PLBart | ✅ | ❌ | ✅ | ❌ | ❌ |
| PoolFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ |
| QDQBert | ❌ | ❌ | ✅ | ❌ | ❌ |
| RAG | ✅ | ❌ | ✅ | ✅ | ❌ |
| REALM | ✅ | ✅ | ✅ | ❌ | ❌ |
| Reformer | ✅ | ✅ | ✅ | ❌ | ❌ |
| RegNet | ❌ | ❌ | ✅ | ✅ | ✅ |
| RemBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| ResNet | ❌ | ❌ | ✅ | ✅ | ✅ |
| RetriBERT | ✅ | ✅ | ✅ | ❌ | ❌ |
| RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ |
| RoBERTa-PreLayerNorm | ❌ | ❌ | ✅ | ✅ | ✅ |
| RoCBert | ✅ | ❌ | ✅ | ❌ | ❌ |
| RoFormer | ✅ | ✅ | ✅ | ✅ | ✅ |
| RWKV | ❌ | ❌ | ✅ | ❌ | ❌ |
| SAM | ❌ | ❌ | ✅ | ✅ | ❌ |
| SegFormer | ❌ | ❌ | ✅ | ✅ | ❌ |
| SEW | ❌ | ❌ | ✅ | ❌ | ❌ |
| SEW-D | ❌ | ❌ | ✅ | ❌ | ❌ |
| Speech Encoder decoder | ❌ | ❌ | ✅ | ❌ | ✅ |
| Speech2Text | ✅ | ❌ | ✅ | ✅ | ❌ |
| Speech2Text2 | ✅ | ❌ | ❌ | ❌ | ❌ |
| SpeechT5 | ✅ | ❌ | ✅ | ❌ | ❌ |
| Splinter | ✅ | ✅ | ✅ | ❌ | ❌ |
| SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ |
| SwiftFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Swin Transformer | ❌ | ❌ | ✅ | ✅ | ❌ |
| Swin Transformer V2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| Swin2SR | ❌ | ❌ | ✅ | ❌ | ❌ |
| SwitchTransformers | ❌ | ❌ | ✅ | ❌ | ❌ |
| T5 | ✅ | ✅ | ✅ | ✅ | ✅ |
| Table Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TAPAS | ✅ | ❌ | ✅ | ✅ | ❌ |
| Time Series Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TimeSformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TimmBackbone | ❌ | ❌ | ❌ | ❌ | ❌ |
| Trajectory Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Transformer-XL | ✅ | ❌ | ✅ | ✅ | ❌ |
| TrOCR | ❌ | ❌ | ✅ | ❌ | ❌ |
| TVLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| UniSpeech | ❌ | ❌ | ✅ | ❌ | ❌ |
| UniSpeechSat | ❌ | ❌ | ✅ | ❌ | ❌ |
| UPerNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| VAN | ❌ | ❌ | ✅ | ❌ | ❌ |
| VideoMAE | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Vision Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| VisionTextDualEncoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| VisualBERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViT | ❌ | ❌ | ✅ | ✅ | ✅ |
| ViT Hybrid | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViTMAE | ❌ | ❌ | ✅ | ✅ | ❌ |
| ViTMSN | ❌ | ❌ | ✅ | ❌ | ❌ |
| Wav2Vec2 | ✅ | ❌ | ✅ | ✅ | ✅ |
| Wav2Vec2-Conformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| WavLM | ❌ | ❌ | ✅ | ❌ | ❌ |
| Whisper | ✅ | ✅ | ✅ | ✅ | ✅ |
| X-CLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| X-MOD | ❌ | ❌ | ✅ | ❌ | ❌ |
| XGLM | ✅ | ✅ | ✅ | ✅ | ✅ |
| XLM | ✅ | ❌ | ✅ | ✅ | ❌ |
| XLM-ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ |
| XLM-RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ |
| XLM-RoBERTa-XL | ❌ | ❌ | ✅ | ❌ | ❌ |
| XLNet | ✅ | ✅ | ✅ | ✅ | ❌ |
| YOLOS | ❌ | ❌ | ✅ | ❌ | ❌ |
| YOSO | ❌ | ❌ | ✅ | ❌ | ❌ |
<!-- End table-->
docs/source/en/index.mdx deleted 100644 → 0
View file @ b0513b01
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# 🤗 Transformers
State-of-the-art Machine Learning for [PyTorch](https://pytorch.org/), [TensorFlow](https://www.tensorflow.org/), and [JAX](https://jax.readthedocs.io/en/latest/).
🤗 Transformers provides APIs and tools to easily download and train state-of-the-art pretrained models. Using pretrained models can reduce your compute costs, carbon footprint, and save you the time and resources required to train a model from scratch. These models support common tasks in different modalities, such as:
📝 **Natural Language Processing**: text classification, named entity recognition, question answering, language modeling, summarization, translation, multiple choice, and text generation.<br>
🖼️ **Computer Vision**: image classification, object detection, and segmentation.<br>
🗣️ **Audio**: automatic speech recognition and audio classification.<br>
🐙 **Multimodal**: table question answering, optical character recognition, information extraction from scanned documents, video classification, and visual question answering.
🤗 Transformers support framework interoperability between PyTorch, TensorFlow, and JAX. This provides the flexibility to use a different framework at each stage of a model's life; train a model in three lines of code in one framework, and load it for inference in another. Models can also be exported to a format like ONNX and TorchScript for deployment in production environments.
Join the growing community on the [Hub](https://huggingface.co/models), [forum](https://discuss.huggingface.co/), or [Discord](https://discord.com/invite/JfAtkvEtRb) today!
## If you are looking for custom support from the Hugging Face team
<a target="_blank" href="https://huggingface.co/support">
<img alt="HuggingFace Expert Acceleration Program" src="https://cdn-media.huggingface.co/marketing/transformers/new-support-improved.png" style="width: 100%; max-width: 600px; border: 1px solid #eee; border-radius: 4px; box-shadow: 0 1px 2px 0 rgba(0, 0, 0, 0.05);">
</a>
## Contents
The documentation is organized into five sections:
- **GET STARTED** provides a quick tour of the library and installation instructions to get up and running.
- **TUTORIALS** are a great place to start if you're a beginner. This section will help you gain the basic skills you need to start using the library.
- **HOW-TO GUIDES** show you how to achieve a specific goal, like finetuning a pretrained model for language modeling or how to write and share a custom model.
- **CONCEPTUAL GUIDES** offers more discussion and explanation of the underlying concepts and ideas behind models, tasks, and the design philosophy of 🤗 Transformers.
- **API** describes all classes and functions:
- **MAIN CLASSES** details the most important classes like configuration, model, tokenizer, and pipeline.
- **MODELS** details the classes and functions related to each model implemented in the library.
- **INTERNAL HELPERS** details utility classes and functions used internally.
### Supported models
<!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
1. **[Autoformer](model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
1. **[BEiT](model_doc/beit)** (from Microsoft) released with the paper [BEiT: BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254) by Hangbo Bao, Li Dong, Furu Wei.
1. **[BERT](model_doc/bert)** (from Google) released 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.
1. **[BERT For Sequence Generation](model_doc/bert-generation)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[BERTweet](model_doc/bertweet)** (from VinAI Research) released with the paper [BERTweet: A pre-trained language model for English Tweets](https://aclanthology.org/2020.emnlp-demos.2/) by Dat Quoc Nguyen, Thanh Vu and Anh Tuan Nguyen.
1. **[BigBird-Pegasus](model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BigBird-RoBERTa](model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BioGpt](model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
1. **[BiT](model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BLIP](model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
1. **[BLIP-2](model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
1. **[Chinese-CLIP](model_doc/chinese_clip)** (from OFA-Sys) released with the paper [Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese](https://arxiv.org/abs/2211.01335) by An Yang, Junshu Pan, Junyang Lin, Rui Men, Yichang Zhang, Jingren Zhou, Chang Zhou.
1. **[CLAP](model_doc/clap)** (from LAION-AI) released with the paper [Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation](https://arxiv.org/abs/2211.06687) by Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, Shlomo Dubnov.
1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
1. **[CLIPSeg](model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie.
1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
1. **[CPM-Ant](model_doc/cpmant)** (from OpenBMB) released by the [OpenBMB](https://www.openbmb.org/).
1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang.
1. **[Data2Vec](model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec: A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[Decision Transformer](model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
1. **[Deformable DETR](model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
1. **[DePlot](model_doc/deplot)** (from Google AI) released with the paper [DePlot: One-shot visual language reasoning by plot-to-table translation](https://arxiv.org/abs/2212.10505) by Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, Yasemin Altun.
1. **[DETA](model_doc/deta)** (from The University of Texas at Austin) released with the paper [NMS Strikes Back](https://arxiv.org/abs/2212.06137) by Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, Philipp Krähenbühl.
1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
1. **[DiNAT](model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
1. **[Donut](model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
1. **[DPR](model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
1. **[DPT](master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
1. **[EfficientFormer](model_doc/efficientformer)** (from Snap Research) released with the paper [EfficientFormer: Vision Transformers at MobileNetSpeed](https://arxiv.org/abs/2206.01191) by Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, Jian Ren.
1. **[EfficientNet](model_doc/efficientnet)** (from Google Brain) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan, Quoc V. Le.
1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
1. **[EnCodec](model_doc/encodec)** (from Meta AI) released with the paper [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[ERNIE](model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu.
1. **[ErnieM](model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.
1. **[ESM](model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2 and ESMFold** were released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives.
1. **[FLAN-T5](model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
1. **[FLAN-UL2](model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei
1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
1. **[FocalNet](model_doc/focalnet)** (from Microsoft Research) released with the paper [Focal Modulation Networks](https://arxiv.org/abs/2203.11926) by Jianwei Yang, Chunyuan Li, Xiyang Dai, Lu Yuan, Jianfeng Gao.
1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
1. **[GIT](model_doc/git)** (from Microsoft Research) released with the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) by Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang.
1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
1. **[GPT NeoX Japanese](model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori.
1. **[GPT-2](model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
1. **[GPT-J](model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
1. **[GPT-Sw3](model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
1. **[GPTBigCode](model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
1. **[Graphormer](model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
1. **[ImageGPT](model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[Jukebox](model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[LayoutLM](model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
1. **[LayoutLMv2](model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
1. **[LayoutLMv3](model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.
1. **[LayoutXLM](model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
1. **[LED](model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LeViT](model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
1. **[LiLT](model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
1. **[LLaMA](model_doc/llama)** (from The FAIR team of Meta AI) released with the paper [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) by Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample.
1. **[Longformer](model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LongT5](model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
1. **[LUKE](model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
1. **[LXMERT](model_doc/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
1. **[M-CTC-T](model_doc/mctct)** (from Facebook) released with the paper [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) by Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert.
1. **[M2M100](model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
1. **[MarianMT](model_doc/marian)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
1. **[MarkupLM](model_doc/markuplm)** (from Microsoft Research Asia) released with the paper [MarkupLM: Pre-training of Text and Markup Language for Visually-rich Document Understanding](https://arxiv.org/abs/2110.08518) by Junlong Li, Yiheng Xu, Lei Cui, Furu Wei.
1. **[Mask2Former](model_doc/mask2former)** (from FAIR and UIUC) released with the paper [Masked-attention Mask Transformer for Universal Image Segmentation](https://arxiv.org/abs/2112.01527) by Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar.
1. **[MaskFormer](model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov.
1. **[MatCha](model_doc/matcha)** (from Google AI) released with the paper [MatCha: Enhancing Visual Language Pretraining with Math Reasoning and Chart Derendering](https://arxiv.org/abs/2212.09662) by Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, Julian Martin Eisenschlos.
1. **[mBART](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
1. **[mBART-50](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
1. **[MEGA](model_doc/mega)** (from Meta/USC/CMU/SJTU) released with the paper [Mega: Moving Average Equipped Gated Attention](https://arxiv.org/abs/2209.10655) by Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer.
1. **[Megatron-BERT](model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[Megatron-GPT2](model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[MGP-STR](model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://arxiv.org/abs/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
1. **[mLUKE](model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
1. **[MMS](model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://arxiv.org/abs/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
1. **[MobileNetV1](model_doc/mobilenet_v1)** (from Google Inc.) released with the paper [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861) by Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam.
1. **[MobileNetV2](model_doc/mobilenet_v2)** (from Google Inc.) released with the paper [MobileNetV2: Inverted Residuals and Linear Bottlenecks](https://arxiv.org/abs/1801.04381) by Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen.
1. **[MobileViT](model_doc/mobilevit)** (from Apple) released with the paper [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) by Sachin Mehta and Mohammad Rastegari.
1. **[MobileViTV2](model_doc/mobilevitv2)** (from Apple) released with the paper [Separable Self-attention for Mobile Vision Transformers](https://arxiv.org/abs/2206.02680) by Sachin Mehta and Mohammad Rastegari.
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
1. **[NAT](model_doc/nat)** (from SHI Labs) released with the paper [Neighborhood Attention Transformer](https://arxiv.org/abs/2204.07143) by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
1. **[Nezha](model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[NLLB-MOE](model_doc/nllb-moe)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
1. **[OneFormer](model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
1. **[OpenLlama](model_doc/open-llama)** (from [s-JoL](https://huggingface.co/s-JoL)) released in [Open-Llama](https://github.com/s-JoL/Open-Llama).
1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
1. **[OWL-ViT](model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
1. **[PEGASUS-X](model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
1. **[Pix2Struct](model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
1. **[PoolFormer](model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
1. **[ProphetNet](model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[QDQBert](model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
1. **[RAG](model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
1. **[REALM](model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
1. **[Reformer](model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
1. **[RegNet](model_doc/regnet)** (from META Platforms) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
1. **[RemBERT](model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
1. **[ResNet](model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
1. **[RoBERTa](model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
1. **[RoBERTa-PreLayerNorm](model_doc/roberta-prelayernorm)** (from Facebook) released with the paper [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://arxiv.org/abs/1904.01038) by Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, Michael Auli.
1. **[RoCBert](model_doc/roc_bert)** (from WeChatAI) released with the paper [RoCBert: Robust Chinese Bert with Multimodal Contrastive Pretraining](https://aclanthology.org/2022.acl-long.65.pdf) by HuiSu, WeiweiShi, XiaoyuShen, XiaoZhou, TuoJi, JiaruiFang, JieZhou.
1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[RWKV](model_doc/rwkv)** (from Bo Peng), released on [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
1. **[Segment Anything](model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SpeechT5](model_doc/speecht5)** (from Microsoft Research) released with the paper [SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing](https://arxiv.org/abs/2110.07205) by Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, Furu Wei.
1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau.
1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy.
1. **[SqueezeBERT](model_doc/squeezebert)** (from Berkeley) released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
1. **[SwiftFormer](model_doc/swiftformer)** (from MBZUAI) released with the paper [SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications](https://arxiv.org/abs/2303.15446) by Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, Fahad Shahbaz Khan.
1. **[Swin Transformer](model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo.
1. **[Swin Transformer V2](model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo.
1. **[Swin2SR](model_doc/swin2sr)** (from University of Würzburg) released with the paper [Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration](https://arxiv.org/abs/2209.11345) by Marcos V. Conde, Ui-Jin Choi, Maxime Burchi, Radu Timofte.
1. **[SwitchTransformers](model_doc/switch_transformers)** (from Google) released with the paper [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://arxiv.org/abs/2101.03961) by William Fedus, Barret Zoph, Noam Shazeer.
1. **[T5](model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[T5v1.1](model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[Table Transformer](model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
1. **[Time Series Transformer](model_doc/time_series_transformer)** (from HuggingFace).
1. **[TimeSformer](model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
1. **[TVLT](model_doc/tvlt)** (from UNC Chapel Hill) released with the paper [TVLT: Textless Vision-Language Transformer](https://arxiv.org/abs/2209.14156) by Zineng Tang, Jaemin Cho, Yixin Nie, Mohit Bansal.
1. **[UL2](model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
1. **[UniSpeech](model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
1. **[UniSpeechSat](model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
1. **[UPerNet](model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
1. **[VAN](model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
1. **[VideoMAE](model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
1. **[ViLT](model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
1. **[ViT Hybrid](model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
1. **[ViTMSN](model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
1. **[Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
1. **[Wav2Vec2Phoneme](model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
1. **[WavLM](model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
1. **[Whisper](model_doc/whisper)** (from OpenAI) released with the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
1. **[X-CLIP](model_doc/xclip)** (from Microsoft Research) released with the paper [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) by Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling.
1. **[X-MOD](model_doc/xmod)** (from Meta AI) released with the paper [Lifting the Curse of Multilinguality by Pre-training Modular Transformers](http://dx.doi.org/10.18653/v1/2022.naacl-main.255) by Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, Mikel Artetxe.
1. **[XGLM](model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
1. **[XLM](model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
1. **[XLM-ProphetNet](model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[XLM-RoBERTa](model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
1. **[XLM-RoBERTa-XL](model_doc/xlm-roberta-xl)** (from Facebook AI), released together with the paper [Larger-Scale Transformers for Multilingual Masked Language Modeling](https://arxiv.org/abs/2105.00572) by Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, Alexis Conneau.
1. **[XLM-V](model_doc/xlm-v)** (from Meta AI) released with the paper [XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models](https://arxiv.org/abs/2301.10472) by Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, Madian Khabsa.
1. **[XLNet](model_doc/xlnet)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
1. **[XLS-R](model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
1. **[XLSR-Wav2Vec2](model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
1. **[YOLOS](model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
1. **[YOSO](model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
### Supported frameworks
The table below represents the current support in the library for each of those models, whether they have a Python
tokenizer (called "slow"). A "fast" tokenizer backed by the 🤗 Tokenizers library, whether they have support in Jax (via
Flax), PyTorch, and/or TensorFlow.
<!--This table is updated automatically from the auto modules with _make fix-copies_. Do not update manually!-->
| Model | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support |
|:-----------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:|
| ALBERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| ALIGN | ❌ | ❌ | ✅ | ❌ | ❌ |
| AltCLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| Audio Spectrogram Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Autoformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| BART | ✅ | ✅ | ✅ | ✅ | ✅ |
| BEiT | ❌ | ❌ | ✅ | ❌ | ✅ |
| BERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| Bert Generation | ✅ | ❌ | ✅ | ❌ | ❌ |
| BigBird | ✅ | ✅ | ✅ | ❌ | ✅ |
| BigBird-Pegasus | ❌ | ❌ | ✅ | ❌ | ❌ |
| BioGpt | ✅ | ❌ | ✅ | ❌ | ❌ |
| BiT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Blenderbot | ✅ | ✅ | ✅ | ✅ | ✅ |
| BlenderbotSmall | ✅ | ✅ | ✅ | ✅ | ✅ |
| BLIP | ❌ | ❌ | ✅ | ✅ | ❌ |
| BLIP-2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| BLOOM | ❌ | ✅ | ✅ | ❌ | ❌ |
| BridgeTower | ❌ | ❌ | ✅ | ❌ | ❌ |
| CamemBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| CANINE | ✅ | ❌ | ✅ | ❌ | ❌ |
| Chinese-CLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| CLAP | ❌ | ❌ | ✅ | ❌ | ❌ |
| CLIP | ✅ | ✅ | ✅ | ✅ | ✅ |
| CLIPSeg | ❌ | ❌ | ✅ | ❌ | ❌ |
| CodeGen | ✅ | ✅ | ✅ | ❌ | ❌ |
| Conditional DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ |
| ConvNeXTV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| CPM-Ant | ✅ | ❌ | ✅ | ❌ | ❌ |
| CTRL | ✅ | ❌ | ✅ | ✅ | ❌ |
| CvT | ❌ | ❌ | ✅ | ✅ | ❌ |
| Data2VecAudio | ❌ | ❌ | ✅ | ❌ | ❌ |
| Data2VecText | ❌ | ❌ | ✅ | ❌ | ❌ |
| Data2VecVision | ❌ | ❌ | ✅ | ✅ | ❌ |
| DeBERTa | ✅ | ✅ | ✅ | ✅ | ❌ |
| DeBERTa-v2 | ✅ | ✅ | ✅ | ✅ | ❌ |
| Decision Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Deformable DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DeiT | ❌ | ❌ | ✅ | ✅ | ❌ |
| DETA | ❌ | ❌ | ✅ | ❌ | ❌ |
| DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DiNAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| DistilBERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| DonutSwin | ❌ | ❌ | ✅ | ❌ | ❌ |
| DPR | ✅ | ✅ | ✅ | ✅ | ❌ |
| DPT | ❌ | ❌ | ✅ | ❌ | ❌ |
| EfficientFormer | ❌ | ❌ | ✅ | ✅ | ❌ |
| EfficientNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| ELECTRA | ✅ | ✅ | ✅ | ✅ | ✅ |
| EnCodec | ❌ | ❌ | ✅ | ❌ | ❌ |
| Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| ERNIE | ❌ | ❌ | ✅ | ❌ | ❌ |
| ErnieM | ✅ | ❌ | ✅ | ❌ | ❌ |
| ESM | ✅ | ❌ | ✅ | ✅ | ❌ |
| FairSeq Machine-Translation | ✅ | ❌ | ✅ | ❌ | ❌ |
| FlauBERT | ✅ | ❌ | ✅ | ✅ | ❌ |
| FLAVA | ❌ | ❌ | ✅ | ❌ | ❌ |
| FNet | ✅ | ✅ | ✅ | ❌ | ❌ |
| FocalNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ |
| GIT | ❌ | ❌ | ✅ | ❌ | ❌ |
| GLPN | ❌ | ❌ | ✅ | ❌ | ❌ |
| GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ |
| GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ |
| GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
| GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ |
| GPT-Sw3 | ✅ | ✅ | ✅ | ✅ | ✅ |
| GPTBigCode | ❌ | ❌ | ✅ | ❌ | ❌ |
| GPTSAN-japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
| Graphormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| GroupViT | ❌ | ❌ | ✅ | ✅ | ❌ |
| Hubert | ❌ | ❌ | ✅ | ✅ | ❌ |
| I-BERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| ImageGPT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Informer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Jukebox | ✅ | ❌ | ✅ | ❌ | ❌ |
| LayoutLM | ✅ | ✅ | ✅ | ✅ | ❌ |
| LayoutLMv2 | ✅ | ✅ | ✅ | ❌ | ❌ |
| LayoutLMv3 | ✅ | ✅ | ✅ | ✅ | ❌ |
| LED | ✅ | ✅ | ✅ | ✅ | ❌ |
| LeViT | ❌ | ❌ | ✅ | ❌ | ❌ |
| LiLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| LLaMA | ✅ | ✅ | ✅ | ❌ | ❌ |
| Longformer | ✅ | ✅ | ✅ | ✅ | ❌ |
| LongT5 | ❌ | ❌ | ✅ | ❌ | ✅ |
| LUKE | ✅ | ❌ | ✅ | ❌ | ❌ |
| LXMERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| M-CTC-T | ❌ | ❌ | ✅ | ❌ | ❌ |
| M2M100 | ✅ | ❌ | ✅ | ❌ | ❌ |
| Marian | ✅ | ❌ | ✅ | ✅ | ✅ |
| MarkupLM | ✅ | ✅ | ✅ | ❌ | ❌ |
| Mask2Former | ❌ | ❌ | ✅ | ❌ | ❌ |
| MaskFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| MaskFormerSwin | ❌ | ❌ | ❌ | ❌ | ❌ |
| mBART | ✅ | ✅ | ✅ | ✅ | ✅ |
| MEGA | ❌ | ❌ | ✅ | ❌ | ❌ |
| Megatron-BERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| MGP-STR | ✅ | ❌ | ✅ | ❌ | ❌ |
| MobileBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| MobileNetV1 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MobileNetV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MobileViT | ❌ | ❌ | ✅ | ✅ | ❌ |
| MobileViTV2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| MPNet | ✅ | ✅ | ✅ | ✅ | ❌ |
| MT5 | ✅ | ✅ | ✅ | ✅ | ✅ |
| MVP | ✅ | ✅ | ✅ | ❌ | ❌ |
| NAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nezha | ❌ | ❌ | ✅ | ❌ | ❌ |
| NLLB-MOE | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| OneFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ |
| OpenAI GPT-2 | ✅ | ✅ | ✅ | ✅ | ✅ |
| OpenLlama | ❌ | ❌ | ✅ | ❌ | ❌ |
| OPT | ❌ | ❌ | ✅ | ✅ | ✅ |
| OWL-ViT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Pegasus | ✅ | ✅ | ✅ | ✅ | ✅ |
| PEGASUS-X | ❌ | ❌ | ✅ | ❌ | ❌ |
| Perceiver | ✅ | ❌ | ✅ | ❌ | ❌ |
| Pix2Struct | ❌ | ❌ | ✅ | ❌ | ❌ |
| PLBart | ✅ | ❌ | ✅ | ❌ | ❌ |
| PoolFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ |
| QDQBert | ❌ | ❌ | ✅ | ❌ | ❌ |
| RAG | ✅ | ❌ | ✅ | ✅ | ❌ |
| REALM | ✅ | ✅ | ✅ | ❌ | ❌ |
| Reformer | ✅ | ✅ | ✅ | ❌ | ❌ |
| RegNet | ❌ | ❌ | ✅ | ✅ | ✅ |
| RemBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| ResNet | ❌ | ❌ | ✅ | ✅ | ✅ |
| RetriBERT | ✅ | ✅ | ✅ | ❌ | ❌ |
| RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ |
| RoBERTa-PreLayerNorm | ❌ | ❌ | ✅ | ✅ | ✅ |
| RoCBert | ✅ | ❌ | ✅ | ❌ | ❌ |
| RoFormer | ✅ | ✅ | ✅ | ✅ | ✅ |
| RWKV | ❌ | ❌ | ✅ | ❌ | ❌ |
| SAM | ❌ | ❌ | ✅ | ✅ | ❌ |
| SegFormer | ❌ | ❌ | ✅ | ✅ | ❌ |
| SEW | ❌ | ❌ | ✅ | ❌ | ❌ |
| SEW-D | ❌ | ❌ | ✅ | ❌ | ❌ |
| Speech Encoder decoder | ❌ | ❌ | ✅ | ❌ | ✅ |
| Speech2Text | ✅ | ❌ | ✅ | ✅ | ❌ |
| Speech2Text2 | ✅ | ❌ | ❌ | ❌ | ❌ |
| SpeechT5 | ✅ | ❌ | ✅ | ❌ | ❌ |
| Splinter | ✅ | ✅ | ✅ | ❌ | ❌ |
| SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ |
| SwiftFormer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Swin Transformer | ❌ | ❌ | ✅ | ✅ | ❌ |
| Swin Transformer V2 | ❌ | ❌ | ✅ | ❌ | ❌ |
| Swin2SR | ❌ | ❌ | ✅ | ❌ | ❌ |
| SwitchTransformers | ❌ | ❌ | ✅ | ❌ | ❌ |
| T5 | ✅ | ✅ | ✅ | ✅ | ✅ |
| Table Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TAPAS | ✅ | ❌ | ✅ | ✅ | ❌ |
| Time Series Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TimeSformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| TimmBackbone | ❌ | ❌ | ❌ | ❌ | ❌ |
| Trajectory Transformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| Transformer-XL | ✅ | ❌ | ✅ | ✅ | ❌ |
| TrOCR | ❌ | ❌ | ✅ | ❌ | ❌ |
| TVLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| UniSpeech | ❌ | ❌ | ✅ | ❌ | ❌ |
| UniSpeechSat | ❌ | ❌ | ✅ | ❌ | ❌ |
| UPerNet | ❌ | ❌ | ✅ | ❌ | ❌ |
| VAN | ❌ | ❌ | ✅ | ❌ | ❌ |
| VideoMAE | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViLT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Vision Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| VisionTextDualEncoder | ❌ | ❌ | ✅ | ✅ | ✅ |
| VisualBERT | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViT | ❌ | ❌ | ✅ | ✅ | ✅ |
| ViT Hybrid | ❌ | ❌ | ✅ | ❌ | ❌ |
| ViTMAE | ❌ | ❌ | ✅ | ✅ | ❌ |
| ViTMSN | ❌ | ❌ | ✅ | ❌ | ❌ |
| Wav2Vec2 | ✅ | ❌ | ✅ | ✅ | ✅ |
| Wav2Vec2-Conformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| WavLM | ❌ | ❌ | ✅ | ❌ | ❌ |
| Whisper | ✅ | ✅ | ✅ | ✅ | ✅ |
| X-CLIP | ❌ | ❌ | ✅ | ❌ | ❌ |
| X-MOD | ❌ | ❌ | ✅ | ❌ | ❌ |
| XGLM | ✅ | ✅ | ✅ | ✅ | ✅ |
| XLM | ✅ | ❌ | ✅ | ✅ | ❌ |
| XLM-ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ |
| XLM-RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ |
| XLM-RoBERTa-XL | ❌ | ❌ | ✅ | ❌ | ❌ |
| XLNet | ✅ | ✅ | ✅ | ✅ | ❌ |
| YOLOS | ❌ | ❌ | ✅ | ❌ | ❌ |
| YOSO | ❌ | ❌ | ✅ | ❌ | ❌ |
<!-- End table-->
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment