Skip to content

GitLab

Unverified Commit d3eacbb8 authored by NielsRogge's avatar NielsRogge Committed by GitHub
Browse files

Add DETR (#11653) 



* Squash all commits of modeling_detr_v7 branch into one

* Improve docs

* Fix tests

* Style

* Improve docs some more and fix most tests

* Fix slow tests of ViT, DeiT and DETR

* Improve replacement of batch norm

* Restructure timm backbone forward

* Make DetrForSegmentation support any timm backbone

* Fix name of output

* Address most comments by @LysandreJik

* Give better names for variables

* Conditional imports + timm in setup.py

* Address additional comments by @sgugger

* Make style, add require_timm and require_vision to testsé

* Remove train_backbone attribute of DetrConfig, add methods to freeze/unfreeze backbone

* Add png files to fixtures

* Fix type hint

* Add timm to workflows

* Add `BatchNorm2d` to the weight initialization

* Fix retain_grad test

* Replace model checkpoints by Facebook namespace

* Fix name of checkpoint in test

* Add user-friendly message when scipy is not available

* Address most comments by @patrickvonplaten

* Remove return_intermediate_layers attribute of DetrConfig and simplify Joiner

* Better initialization

* Scipy is necessary to get sklearn metrics

* Rename TimmBackbone to DetrTimmConvEncoder and rename DetrJoiner to DetrConvModel

* Make style

* Improve docs and add 2 community notebooks
Co-authored-by: default avatarLysandre <lysandre.debut@reseau.eseo.fr>
parent d14e0af2
Expand all Show whitespace changes
Inline Side-by-side
Showing with 4054 additions and 50 deletions
.circleci/config.yml
View file @ d3eacbb8
......@@ -139,7 +139,7 @@ jobs:
- v0.4-{{ checksum "setup.py" }}
- run: sudo apt-get -y update && sudo apt-get install -y libsndfile1-dev
- run: pip install --upgrade pip
- run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision]
- run: pip install .[sklearn,torch,testing,sentencepiece,speech,vision,timm]
- run: pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cpu.html
- save_cache:
key: v0.4-torch-{{ checksum "setup.py" }}
......
.github/workflows/self-push.yml
View file @ d3eacbb8
......@@ -37,7 +37,7 @@ jobs:
run: |
apt -y update && apt install -y libsndfile1-dev
pip install --upgrade pip
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
- name: Are GPUs recognized by our DL frameworks
run: |
......@@ -121,7 +121,7 @@ jobs:
run: |
apt -y update && apt install -y libsndfile1-dev
pip install --upgrade pip
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech]
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
- name: Are GPUs recognized by our DL frameworks
run: |
......
.github/workflows/self-scheduled.yml
View file @ d3eacbb8
......@@ -33,7 +33,7 @@ jobs:
run: |
apt -y update && apt install -y libsndfile1-dev
pip install --upgrade pip
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,integrations]
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
- name: Are GPUs recognized by our DL frameworks
run: |
......@@ -155,7 +155,7 @@ jobs:
run: |
apt -y update && apt install -y libsndfile1-dev
pip install --upgrade pip
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,integrations]
pip install .[sklearn,testing,onnxruntime,sentencepiece,speech,vision,timm]
- name: Are GPUs recognized by our DL frameworks
run: |
......
README.md
View file @ d3eacbb8
......@@ -215,6 +215,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
1. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (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](https://huggingface.co/transformers/model_doc/deberta_v2.html)** (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. **[DeiT](https://huggingface.co/transformers/model_doc/deit.html)** (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. **[DETR](https://huggingface.co/transformers/model_doc/detr.html)** (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](https://huggingface.co/transformers/model_doc/dialogpt.html)** (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. **[DistilBERT](https://huggingface.co/transformers/model_doc/distilbert.html)** (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/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
1. **[DPR](https://huggingface.co/transformers/model_doc/dpr.html)** (from Facebook) released with the paper [Dense Passage Retrieval
......
docs/source/community.md
View file @ d3eacbb8
......@@ -59,3 +59,5 @@ This page regroups resources around 🤗 Transformers developed by the community
| [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) |
docs/source/index.rst
View file @ d3eacbb8
......@@ -153,128 +153,131 @@ Supported models
19. :doc:`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.
20. :doc:`DialoGPT <model_doc/dialogpt>` (from Microsoft Research) released with the paper `DialoGPT: Large-Scale
20. :doc:`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.
21. :doc:`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.
21. :doc:`DistilBERT <model_doc/distilbert>` (from HuggingFace), released together with the paper `DistilBERT, a
22. :doc:`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/master/examples/distillation>`__, RoBERTa into `DistilRoBERTa
<https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, Multilingual BERT into
`DistilmBERT <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German
version of DistilBERT.
22. :doc:`DPR <model_doc/dpr>` (from Facebook) released with the paper `Dense Passage Retrieval for Open-Domain
23. :doc:`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.
23. :doc:`ELECTRA <model_doc/electra>` (from Google Research/Stanford University) released with the paper `ELECTRA:
24. :doc:`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.
24. :doc:`FlauBERT <model_doc/flaubert>` (from CNRS) released with the paper `FlauBERT: Unsupervised Language Model
25. :doc:`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.
25. :doc:`Funnel Transformer <model_doc/funnel>` (from CMU/Google Brain) released with the paper `Funnel-Transformer:
26. :doc:`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.
26. :doc:`GPT <model_doc/gpt>` (from OpenAI) released with the paper `Improving Language Understanding by Generative
27. :doc:`GPT <model_doc/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.
27. :doc:`GPT-2 <model_doc/gpt2>` (from OpenAI) released with the paper `Language Models are Unsupervised Multitask
28. :doc:`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**.
28. :doc:`GPT Neo <model_doc/gpt_neo>` (from EleutherAI) released in the repository `EleutherAI/gpt-neo
29. :doc:`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.
29. :doc:`I-BERT <model_doc/ibert>` (from Berkeley) released with the paper `I-BERT: Integer-only BERT Quantization
30. :doc:`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
30. :doc:`LayoutLM <model_doc/layoutlm>` (from Microsoft Research Asia) released with the paper `LayoutLM: Pre-training
31. :doc:`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.
31. :doc:`LED <model_doc/led>` (from AllenAI) released with the paper `Longformer: The Long-Document Transformer
32. :doc:`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.
32. :doc:`Longformer <model_doc/longformer>` (from AllenAI) released with the paper `Longformer: The Long-Document
33. :doc:`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.
33. :doc:`LUKE <model_doc/luke>` (from Studio Ousia) released with the paper `LUKE: Deep Contextualized Entity
34. :doc:`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.
34. :doc:`LXMERT <model_doc/lxmert>` (from UNC Chapel Hill) released with the paper `LXMERT: Learning Cross-Modality
35. :doc:`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.
35. :doc:`M2M100 <model_doc/m2m_100>` (from Facebook) released with the paper `Beyond English-Centric Multilingual
36. :doc:`M2M100 <model_doc/m2m_100>` (from Facebook) released with the paper `Beyond English-Centric Multilingual
Machine Translation <https://arxiv.org/abs/2010.11125>`__ by 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.
36. :doc:`MarianMT <model_doc/marian>` Machine translation models trained using `OPUS <http://opus.nlpl.eu/>`__ data by
37. :doc:`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.
37. :doc:`MBart <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Denoising Pre-training for
38. :doc:`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.
38. :doc:`MBart-50 <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Translation with Extensible
39. :doc:`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.
39. :doc:`Megatron-BERT <model_doc/megatron_bert>` (from NVIDIA) released with the paper `Megatron-LM: Training
40. :doc:`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.
40. :doc:`Megatron-GPT2 <model_doc/megatron_gpt2>` (from NVIDIA) released with the paper `Megatron-LM: Training
41. :doc:`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.
41. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
42. :doc:`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.
42. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
43. :doc:`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.
43. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
44. :doc:`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.
44. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
45. :doc:`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.
45. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
46. :doc:`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.
46. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
47. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper 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.
47. :doc:`RoFormer <model_doc/roformer>` (from ZhuiyiTechnology), released together with the paper a `RoFormer:
48. :doc:`RoFormer <model_doc/roformer>` (from ZhuiyiTechnology), released together with the paper a `RoFormer:
Enhanced Transformer with Rotary Position Embedding <https://arxiv.org/pdf/2104.09864v1.pdf>`__ by Jianlin Su and
Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
48. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
49. :doc:`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.
49. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
50. :doc:`SqueezeBert <model_doc/squeezebert>` 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.
50. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
51. :doc:`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.
51. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
52. :doc:`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.
52. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
53. :doc:`Transformer-XL <model_doc/transformerxl>` (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.
53. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
54. :doc:`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.
54. :doc:`VisualBERT <model_doc/visual_bert>` (from UCLA NLP) released with the paper `VisualBERT: A Simple and
55. :doc:`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.
55. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
56. :doc:`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.
56. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
57. :doc:`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.
57. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
58. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (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.
58. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
59. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (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.
59. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
60. :doc:`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.
60. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
61. :doc:`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.
......@@ -318,6 +321,8 @@ Flax), PyTorch, and/or TensorFlow.
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DPR | ✅ | ✅ | ✅ | ✅ | ❌ |
+-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
| DeBERTa | ✅ | ✅ | ✅ | ❌ | ❌ |
......@@ -502,6 +507,7 @@ Flax), PyTorch, and/or TensorFlow.
model_doc/deberta
model_doc/deberta_v2
model_doc/deit
model_doc/detr
model_doc/dialogpt
model_doc/distilbert
model_doc/dpr
......
docs/source/model_doc/detr.rst 0 → 100644
View file @ d3eacbb8
..
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.
DETR
-----------------------------------------------------------------------------------------------------------------------
Overview
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The DETR model was proposed in `End-to-End Object Detection with Transformers <https://arxiv.org/abs/2005.12872>`__ by
Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov and Sergey Zagoruyko. DETR
consists of a convolutional backbone followed by an encoder-decoder Transformer which can be trained end-to-end for
object detection. It greatly simplifies a lot of the complexity of models like Faster-R-CNN and Mask-R-CNN, which use
things like region proposals, non-maximum suppression procedure and anchor generation. Moreover, DETR can also be
naturally extended to perform panoptic segmentation, by simply adding a mask head on top of the decoder outputs.
The abstract from the paper is the following:
*We present a new method that views object detection as a direct set prediction problem. Our approach streamlines the
detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression
procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the
new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via
bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries,
DETR reasons about the relations of the objects and the global image context to directly output the final set of
predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many
other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and
highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily
generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive
baselines.*
This model was contributed by `nielsr <https://huggingface.co/nielsr>`__. The original code can be found `here
<https://github.com/facebookresearch/detr>`__.
Here's a TLDR explaining how :class:`~transformers.DetrForObjectDetection` works:
First, an image is sent through a pre-trained convolutional backbone (in the paper, the authors use
ResNet-50/ResNet-101). Let's assume we also add a batch dimension. This means that the input to the backbone is a
tensor of shape :obj:`(batch_size, 3, height, width)`, assuming the image has 3 color channels (RGB). The CNN backbone
outputs a new lower-resolution feature map, typically of shape :obj:`(batch_size, 2048, height/32, width/32)`. This is
then projected to match the hidden dimension of the Transformer of DETR, which is :obj:`256` by default, using a
:obj:`nn.Conv2D` layer. So now, we have a tensor of shape :obj:`(batch_size, 256, height/32, width/32).` Next, the
feature map is flattened and transposed to obtain a tensor of shape :obj:`(batch_size, seq_len, d_model)` =
:obj:`(batch_size, width/32*height/32, 256)`. So a difference with NLP models is that the sequence length is actually
longer than usual, but with a smaller :obj:`d_model` (which in NLP is typically 768 or higher).
Next, this is sent through the encoder, outputting :obj:`encoder_hidden_states` of the same shape (you can consider
these as image features). Next, so-called **object queries** are sent through the decoder. This is a tensor of shape
:obj:`(batch_size, num_queries, d_model)`, with :obj:`num_queries` typically set to 100 and initialized with zeros.
These input embeddings are learnt positional encodings that the authors refer to as object queries, and similarly to
the encoder, they are added to the input of each attention layer. Each object query will look for a particular object
in the image. The decoder updates these embeddings through multiple self-attention and encoder-decoder attention layers
to output :obj:`decoder_hidden_states` of the same shape: :obj:`(batch_size, num_queries, d_model)`. Next, two heads
are added on top for object detection: a linear layer for classifying each object query into one of the objects or "no
object", and a MLP to predict bounding boxes for each query.
The model is trained using a **bipartite matching loss**: so what we actually do is compare the predicted classes +
bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N
(so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as
bounding box). The `Hungarian matching algorithm <https://en.wikipedia.org/wiki/Hungarian_algorithm>`__ is used to find
an optimal one-to-one mapping of each of the N queries to each of the N annotations. Next, standard cross-entropy (for
the classes) and a linear combination of the L1 and `generalized IoU loss <https://giou.stanford.edu/>`__ (for the
bounding boxes) are used to optimize the parameters of the model.
DETR can be naturally extended to perform panoptic segmentation (which unifies semantic segmentation and instance
segmentation). :class:`~transformers.DetrForSegmentation` adds a segmentation mask head on top of
:class:`~transformers.DetrForObjectDetection`. The mask head can be trained either jointly, or in a two steps process,
where one first trains a :class:`~transformers.DetrForObjectDetection` model to detect bounding boxes around both
"things" (instances) and "stuff" (background things like trees, roads, sky), then freeze all the weights and train only
the mask head for 25 epochs. Experimentally, these two approaches give similar results. Note that predicting boxes is
required for the training to be possible, since the Hungarian matching is computed using distances between boxes.
Tips:
- DETR uses so-called **object queries** to detect objects in an image. The number of queries determines the maximum
number of objects that can be detected in a single image, and is set to 100 by default (see parameter
:obj:`num_queries` of :class:`~transformers.DetrConfig`). Note that it's good to have some slack (in COCO, the
authors used 100, while the maximum number of objects in a COCO image is ~70).
- The decoder of DETR updates the query embeddings in parallel. This is different from language models like GPT-2,
which use autoregressive decoding instead of parallel. Hence, no causal attention mask is used.
- DETR adds position embeddings to the hidden states at each self-attention and cross-attention layer before projecting
to queries and keys. For the position embeddings of the image, one can choose between fixed sinusoidal or learned
absolute position embeddings. By default, the parameter :obj:`position_embedding_type` of
:class:`~transformers.DetrConfig` is set to :obj:`"sine"`.
- During training, the authors of DETR did find it helpful to use auxiliary losses in the decoder, especially to help
the model output the correct number of objects of each class. If you set the parameter :obj:`auxiliary_loss` of
:class:`~transformers.DetrConfig` to :obj:`True`, then prediction feedforward neural networks and Hungarian losses
are added after each decoder layer (with the FFNs sharing parameters).
- If you want to train the model in a distributed environment across multiple nodes, then one should update the
`num_boxes` variable in the `DetrLoss` class of `modeling_detr.py`. When training on multiple nodes, this should be
set to the average number of target boxes across all nodes, as can be seen in the original implementation `here
<https://github.com/facebookresearch/detr/blob/a54b77800eb8e64e3ad0d8237789fcbf2f8350c5/models/detr.py#L227-L232>`__.
- :class:`~transformers.DetrForObjectDetection` and :class:`~transformers.DetrForSegmentation` can be initialized with
any convolutional backbone available in the `timm library <https://github.com/rwightman/pytorch-image-models>`__.
Initializing with a MobileNet backbone for example can be done by setting the :obj:`backbone` attribute of
:class:`~transformers.DetrConfig` to :obj:`"tf_mobilenetv3_small_075"`, and then initializing the model with that
config.
- DETR resizes the input images such that the shortest side is at least a certain amount of pixels while the longest is
at most 1333 pixels. At training time, scale augmentation is used such that the shortest side is randomly set to at
least 480 and at most 800 pixels. At inference time, the shortest side is set to 800. One can use
:class:`~transformers.DetrFeatureExtractor` to prepare images (and optional annotations in COCO format) for the
model. Due to this resizing, images in a batch can have different sizes. DETR solves this by padding images up to the
largest size in a batch, and by creating a pixel mask that indicates which pixels are real/which are padding.
Alternatively, one can also define a custom :obj:`collate_fn` in order to batch images together, using
:meth:`~transformers.DetrFeatureExtractor.pad_and_create_pixel_mask`.
- The size of the images will determine the amount of memory being used, and will thus determine the :obj:`batch_size`.
It is advised to use a batch size of 2 per GPU. See `this Github thread
<https://github.com/facebookresearch/detr/issues/150>`__ for more info.
As a summary, consider the following table:
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Task** | **Object detection** | **Instance segmentation** | **Panoptic segmentation** |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Description** | Predicting bounding boxes and class labels around | Predicting masks around objects (i.e. instances) in an image | Predicting masks around both objects (i.e. instances) as well as |
| | objects in an image | | "stuff" (i.e. background things like trees and roads) in an image |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Model** | :class:`~transformers.DetrForObjectDetection` | :class:`~transformers.DetrForSegmentation` | :class:`~transformers.DetrForSegmentation` |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Example dataset** | COCO detection | COCO detection, | COCO panoptic |
| | | COCO panoptic | |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Format of annotations to provide to** | {‘image_id’: int, | {‘image_id’: int, | {‘file_name: str, |
| :class:`~transformers.DetrFeatureExtractor` | ‘annotations’: List[Dict]}, each Dict being a COCO | ‘annotations’: [List[Dict]] } (in case of COCO detection) | ‘image_id: int, |
| | object annotation (containing keys "image_id", | | ‘segments_info’: List[Dict] } |
| | | or | |
| | | | and masks_path (path to directory containing PNG files of the masks) |
| | | {‘file_name’: str, | |
| | | ‘image_id’: int, | |
| | | ‘segments_info’: List[Dict]} (in case of COCO panoptic) | |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **Postprocessing** (i.e. converting the | :meth:`~transformers.DetrFeatureExtractor.post_process` | :meth:`~transformers.DetrFeatureExtractor.post_process_segmentation` | :meth:`~transformers.DetrFeatureExtractor.post_process_segmentation`, |
| output of the model to COCO API) | | | :meth:`~transformers.DetrFeatureExtractor.post_process_panoptic` |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
| **evaluators** | :obj:`CocoEvaluator` with iou_types = “bbox” | :obj:`CocoEvaluator` with iou_types = “bbox”, “segm” | :obj:`CocoEvaluator` with iou_tupes = “bbox, “segm” |
| | | | |
| | | | :obj:`PanopticEvaluator` |
+---------------------------------------------+---------------------------------------------------------+----------------------------------------------------------------------+------------------------------------------------------------------------+
In short, one should prepare the data either in COCO detection or COCO panoptic format, then use
:class:`~transformers.DetrFeatureExtractor` to create :obj:`pixel_values`, :obj:`pixel_mask` and optional
:obj:`labels`, which can then be used to train (or fine-tune) a model. For evaluation, one should first convert the
outputs of the model using one of the postprocessing methods of :class:`~transformers.DetrFeatureExtractor`. These can
be be provided to either :obj:`CocoEvaluator` or :obj:`PanopticEvaluator`, which allow you to calculate metrics like
mean Average Precision (mAP) and Panoptic Quality (PQ). The latter objects are implemented in the `original repository
<https://github.com/facebookresearch/detr>`__. See the example notebooks for more info regarding evaluation.
DETR specific outputs
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.models.detr.modeling_detr.DetrModelOutput
:members:
.. autoclass:: transformers.models.detr.modeling_detr.DetrObjectDetectionOutput
:members:
.. autoclass:: transformers.models.detr.modeling_detr.DetrSegmentationOutput
:members:
DetrConfig
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DetrConfig
:members:
DetrFeatureExtractor
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DetrFeatureExtractor
:members: __call__, pad_and_create_pixel_mask, post_process, post_process_segmentation, post_process_panoptic
DetrModel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DetrModel
:members: forward
DetrForObjectDetection
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DetrForObjectDetection
:members: forward
DetrForSegmentation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
.. autoclass:: transformers.DetrForSegmentation
:members: forward
setup.py
View file @ d3eacbb8
......@@ -142,6 +142,7 @@ _deps = [
"tensorflow-cpu>=2.3",
"tensorflow>=2.3",
"timeout-decorator",
"timm",
"tokenizers>=0.10.1,<0.11",
"torch>=1.0",
"torchaudio",
......@@ -249,6 +250,7 @@ extras["integrations"] = extras["optuna"] + extras["ray"]
extras["serving"] = deps_list("pydantic", "uvicorn", "fastapi", "starlette")
extras["speech"] = deps_list("soundfile", "torchaudio")
extras["vision"] = deps_list("Pillow")
extras["timm"] = deps_list("timm")
extras["sentencepiece"] = deps_list("sentencepiece", "protobuf")
extras["testing"] = (
......@@ -270,6 +272,7 @@ extras["all"] = (
+ extras["speech"]
+ extras["vision"]
+ extras["integrations"]
+ extras["timm"]
)
extras["docs_specific"] = deps_list(
......
src/transformers/__init__.py
View file @ d3eacbb8
......@@ -47,6 +47,7 @@ from .file_utils import (
is_sentencepiece_available,
is_speech_available,
is_tf_available,
is_timm_available,
is_tokenizers_available,
is_torch_available,
is_vision_available,
......@@ -101,10 +102,12 @@ _import_structure = {
"is_flax_available",
"is_psutil_available",
"is_py3nvml_available",
"is_scipy_available",
"is_sentencepiece_available",
"is_sklearn_available",
"is_speech_available",
"is_tf_available",
"is_timm_available",
"is_tokenizers_available",
"is_torch_available",
"is_torch_tpu_available",
......@@ -180,6 +183,7 @@ _import_structure = {
"models.deberta": ["DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaConfig", "DebertaTokenizer"],
"models.deberta_v2": ["DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaV2Config"],
"models.deit": ["DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeiTConfig"],
"models.detr": ["DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetrConfig"],
"models.distilbert": ["DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DistilBertConfig", "DistilBertTokenizer"],
"models.dpr": [
"DPR_PRETRAINED_CONFIG_ARCHIVE_MAP",
......@@ -405,6 +409,7 @@ if is_vision_available():
_import_structure["models.clip"].append("CLIPFeatureExtractor")
_import_structure["models.clip"].append("CLIPProcessor")
_import_structure["models.deit"].append("DeiTFeatureExtractor")
_import_structure["models.detr"].append("DetrFeatureExtractor")
_import_structure["models.vit"].append("ViTFeatureExtractor")
else:
from .utils import dummy_vision_objects
......@@ -413,6 +418,23 @@ else:
name for name in dir(dummy_vision_objects) if not name.startswith("_")
]
# Timm-backed objects
if is_timm_available() and is_vision_available():
_import_structure["models.detr"].extend(
[
"DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
"DetrForObjectDetection",
"DetrForSegmentation",
"DetrModel",
]
)
else:
from .utils import dummy_timm_objects
_import_structure["utils.dummy_timm_objects"] = [
name for name in dir(dummy_timm_objects) if not name.startswith("_")
]
# PyTorch-backed objects
if is_torch_available():
_import_structure["benchmark.benchmark"] = ["PyTorchBenchmark"]
......@@ -489,6 +511,7 @@ if is_torch_available():
"MODEL_FOR_MASKED_LM_MAPPING",
"MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
"MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
"MODEL_FOR_OBJECT_DETECTION_MAPPING",
"MODEL_FOR_PRETRAINING_MAPPING",
"MODEL_FOR_QUESTION_ANSWERING_MAPPING",
"MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
......@@ -1587,10 +1610,12 @@ if TYPE_CHECKING:
is_flax_available,
is_psutil_available,
is_py3nvml_available,
is_scipy_available,
is_sentencepiece_available,
is_sklearn_available,
is_speech_available,
is_tf_available,
is_timm_available,
is_tokenizers_available,
is_torch_available,
is_torch_tpu_available,
......@@ -1666,6 +1691,7 @@ if TYPE_CHECKING:
from .models.deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig, DebertaTokenizer
from .models.deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
from .models.deit import DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, DeiTConfig
from .models.detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
from .models.distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig, DistilBertTokenizer
from .models.dpr import (
DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
......@@ -1863,13 +1889,23 @@ if TYPE_CHECKING:
from .image_utils import ImageFeatureExtractionMixin
from .models.clip import CLIPFeatureExtractor, CLIPProcessor
from .models.deit import DeiTFeatureExtractor
from .models.detr import DetrFeatureExtractor
from .models.vit import ViTFeatureExtractor
else:
from .utils.dummy_vision_objects import *
# Modeling
if is_torch_available():
if is_timm_available() and is_vision_available():
from .models.detr import (
DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
DetrForObjectDetection,
DetrForSegmentation,
DetrModel,
)
else:
from .utils.dummy_timm_objects import *
if is_torch_available():
# Benchmarks
from .benchmark.benchmark import PyTorchBenchmark
from .benchmark.benchmark_args import PyTorchBenchmarkArguments
......@@ -1939,6 +1975,7 @@ if TYPE_CHECKING:
MODEL_FOR_MASKED_LM_MAPPING,
MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
MODEL_FOR_OBJECT_DETECTION_MAPPING,
MODEL_FOR_PRETRAINING_MAPPING,
MODEL_FOR_QUESTION_ANSWERING_MAPPING,
MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
......
src/transformers/dependency_versions_table.py
View file @ d3eacbb8
......@@ -59,6 +59,7 @@ deps = {
"tensorflow-cpu": "tensorflow-cpu>=2.3",
"tensorflow": "tensorflow>=2.3",
"timeout-decorator": "timeout-decorator",
"timm": "timm",
"tokenizers": "tokenizers>=0.10.1,<0.11",
"torch": "torch>=1.0",
"torchaudio": "torchaudio",
......
src/transformers/file_utils.py
View file @ d3eacbb8
......@@ -174,6 +174,14 @@ except importlib_metadata.PackageNotFoundError:
_soundfile_available = False
_timm_available = importlib.util.find_spec("timm") is not None
try:
_timm_version = importlib_metadata.version("timm")
logger.debug(f"Successfully imported timm version {_timm_version}")
except importlib_metadata.PackageNotFoundError:
_timm_available = False
_torchaudio_available = importlib.util.find_spec("torchaudio") is not None
try:
_torchaudio_version = importlib_metadata.version("torchaudio")
......@@ -317,12 +325,14 @@ def is_faiss_available():
return _faiss_available
def is_scipy_available():
return importlib.util.find_spec("scipy") is not None
def is_sklearn_available():
if importlib.util.find_spec("sklearn") is None:
return False
if importlib.util.find_spec("scipy") is None:
return False
return importlib.util.find_spec("sklearn.metrics") and importlib.util.find_spec("scipy.stats")
return is_scipy_available() and importlib.util.find_spec("sklearn.metrics")
def is_sentencepiece_available():
......@@ -411,6 +421,10 @@ def is_soundfile_availble():
return _soundfile_available
def is_timm_available():
return _timm_available
def is_torchaudio_available():
return _torchaudio_available
......@@ -536,12 +550,24 @@ explained here: https://pandas.pydata.org/pandas-docs/stable/getting_started/ins
"""
# docstyle-ignore
SCIPY_IMPORT_ERROR = """
{0} requires the scipy library but it was not found in your environment. You can install it with pip:
`pip install scipy`
"""
# docstyle-ignore
SPEECH_IMPORT_ERROR = """
{0} requires the torchaudio library but it was not found in your environment. You can install it with pip:
`pip install torchaudio`
"""
# docstyle-ignore
TIMM_IMPORT_ERROR = """
{0} requires the timm library but it was not found in your environment. You can install it with pip:
`pip install timm`
"""
# docstyle-ignore
VISION_IMPORT_ERROR = """
......@@ -562,9 +588,11 @@ BACKENDS_MAPPING = OrderedDict(
("sklearn", (is_sklearn_available, SKLEARN_IMPORT_ERROR)),
("speech", (is_speech_available, SPEECH_IMPORT_ERROR)),
("tf", (is_tf_available, TENSORFLOW_IMPORT_ERROR)),
("timm", (is_timm_available, TIMM_IMPORT_ERROR)),
("tokenizers", (is_tokenizers_available, TOKENIZERS_IMPORT_ERROR)),
("torch", (is_torch_available, PYTORCH_IMPORT_ERROR)),
("vision", (is_vision_available, VISION_IMPORT_ERROR)),
("scipy", (is_scipy_available, SCIPY_IMPORT_ERROR)),
]
)
......
src/transformers/models/__init__.py
View file @ d3eacbb8
......@@ -36,6 +36,7 @@ from . import (
ctrl,
deberta,
deit,
detr,
dialogpt,
distilbert,
dpr,
......
src/transformers/models/auto/__init__.py
View file @ d3eacbb8
......@@ -35,6 +35,7 @@ if is_torch_available():
"MODEL_FOR_MASKED_LM_MAPPING",
"MODEL_FOR_MULTIPLE_CHOICE_MAPPING",
"MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING",
"MODEL_FOR_OBJECT_DETECTION_MAPPING",
"MODEL_FOR_PRETRAINING_MAPPING",
"MODEL_FOR_QUESTION_ANSWERING_MAPPING",
"MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING",
......@@ -119,6 +120,7 @@ if TYPE_CHECKING:
MODEL_FOR_MASKED_LM_MAPPING,
MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
MODEL_FOR_OBJECT_DETECTION_MAPPING,
MODEL_FOR_PRETRAINING_MAPPING,
MODEL_FOR_QUESTION_ANSWERING_MAPPING,
MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
......
src/transformers/models/auto/configuration_auto.py
View file @ d3eacbb8
......@@ -39,6 +39,7 @@ from ..ctrl.configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLCo
from ..deberta.configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
from ..deberta_v2.configuration_deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
from ..deit.configuration_deit import DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, DeiTConfig
from ..detr.configuration_detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
from ..distilbert.configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
from ..dpr.configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
from ..electra.configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
......@@ -99,6 +100,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP,
DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP,
LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP,
DETR_PRETRAINED_CONFIG_ARCHIVE_MAP,
GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP,
BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP,
MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
......@@ -156,6 +158,7 @@ CONFIG_MAPPING = OrderedDict(
("bigbird_pegasus", BigBirdPegasusConfig),
("deit", DeiTConfig),
("luke", LukeConfig),
("detr", DetrConfig),
("gpt_neo", GPTNeoConfig),
("big_bird", BigBirdConfig),
("speech_to_text", Speech2TextConfig),
......@@ -219,6 +222,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
("bigbird_pegasus", "BigBirdPegasus"),
("deit", "DeiT"),
("luke", "LUKE"),
("detr", "DETR"),
("gpt_neo", "GPT Neo"),
("big_bird", "BigBird"),
("speech_to_text", "Speech2Text"),
......
src/transformers/models/auto/modeling_auto.py
View file @ d3eacbb8
......@@ -106,6 +106,7 @@ from ..deberta_v2.modeling_deberta_v2 import (
DebertaV2Model,
)
from ..deit.modeling_deit import DeiTForImageClassification, DeiTForImageClassificationWithTeacher, DeiTModel
from ..detr.modeling_detr import DetrForObjectDetection, DetrModel
from ..distilbert.modeling_distilbert import (
DistilBertForMaskedLM,
DistilBertForMultipleChoice,
......@@ -316,6 +317,7 @@ from .configuration_auto import (
DebertaConfig,
DebertaV2Config,
DeiTConfig,
DetrConfig,
DistilBertConfig,
DPRConfig,
ElectraConfig,
......@@ -372,6 +374,7 @@ MODEL_MAPPING = OrderedDict(
(BigBirdPegasusConfig, BigBirdPegasusModel),
(DeiTConfig, DeiTModel),
(LukeConfig, LukeModel),
(DetrConfig, DetrModel),
(GPTNeoConfig, GPTNeoModel),
(BigBirdConfig, BigBirdModel),
(Speech2TextConfig, Speech2TextModel),
......@@ -586,6 +589,13 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
]
)
MODEL_FOR_OBJECT_DETECTION_MAPPING = OrderedDict(
[
# Model for Object Detection mapping
(DetrConfig, DetrForObjectDetection),
]
)
MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
[
# Model for Seq2Seq Causal LM mapping
......
src/transformers/models/detr/__init__.py 0 → 100644
View file @ d3eacbb8
# flake8: noqa
# There's no way to ignore "F401 '...' imported but unused" warnings in this
# module, but to preserve other warnings. So, don't check this module at all.
# 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.
from typing import TYPE_CHECKING
from ...file_utils import _BaseLazyModule, is_timm_available, is_vision_available
_import_structure = {
"configuration_detr": ["DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetrConfig"],
}
if is_vision_available():
_import_structure["feature_extraction_detr"] = ["DetrFeatureExtractor"]
if is_timm_available():
_import_structure["modeling_detr"] = [
"DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
"DetrForObjectDetection",
"DetrForSegmentation",
"DetrModel",
"DetrPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
if is_vision_available():
from .feature_extraction_detr import DetrFeatureExtractor
if is_timm_available():
from .modeling_detr import (
DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
DetrForObjectDetection,
DetrForSegmentation,
DetrModel,
DetrPreTrainedModel,
)
else:
import importlib
import os
import sys
class _LazyModule(_BaseLazyModule):
"""
Module class that surfaces all objects but only performs associated imports when the objects are requested.
"""
__file__ = globals()["__file__"]
__path__ = [os.path.dirname(__file__)]
def _get_module(self, module_name: str):
return importlib.import_module("." + module_name, self.__name__)
sys.modules[__name__] = _LazyModule(__name__, _import_structure)
src/transformers/models/detr/configuration_detr.py 0 → 100644
View file @ d3eacbb8
# coding=utf-8
# Copyright 2021 Facebook AI Research and The HuggingFace Inc. 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.
""" DETR model configuration """
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
DETR_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"facebook/detr-resnet-50": "https://huggingface.co/facebook/detr-resnet-50/resolve/main/config.json",
# See all DETR models at https://huggingface.co/models?filter=detr
}
class DetrConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a :class:`~transformers.DetrModel`. It is used to
instantiate a DETR model according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the DETR `facebook/detr-resnet-50
<https://huggingface.co/facebook/detr-resnet-50>`__ architecture.
Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
Args:
num_queries (:obj:`int`, `optional`, defaults to 100):
Number of object queries, i.e. detection slots. This is the maximal number of objects
:class:`~transformers.DetrModel` can detect in a single image. For COCO, we recommend 100 queries.
d_model (:obj:`int`, `optional`, defaults to 256):
Dimension of the layers.
encoder_layers (:obj:`int`, `optional`, defaults to 6):
Number of encoder layers.
decoder_layers (:obj:`int`, `optional`, defaults to 6):
Number of decoder layers.
encoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
Number of attention heads for each attention layer in the Transformer encoder.
decoder_attention_heads (:obj:`int`, `optional`, defaults to 8):
Number of attention heads for each attention layer in the Transformer decoder.
decoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
Dimension of the "intermediate" (often named feed-forward) layer in decoder.
encoder_ffn_dim (:obj:`int`, `optional`, defaults to 2048):
Dimension of the "intermediate" (often named feed-forward) layer in decoder.
activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"relu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string,
:obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
dropout (:obj:`float`, `optional`, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_dropout (:obj:`float`, `optional`, defaults to 0.0):
The dropout ratio for the attention probabilities.
activation_dropout (:obj:`float`, `optional`, defaults to 0.0):
The dropout ratio for activations inside the fully connected layer.
classifier_dropout (:obj:`float`, `optional`, defaults to 0.0):
The dropout ratio for classifier.
max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
init_std (:obj:`float`, `optional`, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
init_xavier_std (:obj:`float`, `optional`, defaults to 1):
The scaling factor used for the Xavier initialization gain in the HM Attention map module.
encoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
The LayerDrop probability for the encoder. See the `LayerDrop paper <see
https://arxiv.org/abs/1909.11556>`__ for more details.
decoder_layerdrop: (:obj:`float`, `optional`, defaults to 0.0):
The LayerDrop probability for the decoder. See the `LayerDrop paper <see
https://arxiv.org/abs/1909.11556>`__ for more details.
auxiliary_loss (:obj:`bool`, `optional`, defaults to :obj:`False`):
Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
position_embedding_type (:obj:`str`, `optional`, defaults to :obj:`"sine"`):
Type of position embeddings to be used on top of the image features. One of :obj:`"sine"` or
:obj:`"learned"`.
backbone (:obj:`str`, `optional`, defaults to :obj:`"resnet50"`):
Name of convolutional backbone to use. Supports any convolutional backbone from the timm package. For a
list of all available models, see `this page
<https://rwightman.github.io/pytorch-image-models/#load-a-pretrained-model>`__.
dilation (:obj:`bool`, `optional`, defaults to :obj:`False`):
Whether to replace stride with dilation in the last convolutional block (DC5).
class_cost (:obj:`float`, `optional`, defaults to 1):
Relative weight of the classification error in the Hungarian matching cost.
bbox_cost (:obj:`float`, `optional`, defaults to 5):
Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.
giou_cost (:obj:`float`, `optional`, defaults to 2):
Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.
mask_loss_coefficient (:obj:`float`, `optional`, defaults to 1):
Relative weight of the Focal loss in the panoptic segmentation loss.
dice_loss_coefficient (:obj:`float`, `optional`, defaults to 1):
Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.
bbox_loss_coefficient (:obj:`float`, `optional`, defaults to 5):
Relative weight of the L1 bounding box loss in the object detection loss.
giou_loss_coefficient (:obj:`float`, `optional`, defaults to 2):
Relative weight of the generalized IoU loss in the object detection loss.
eos_coefficient (:obj:`float`, `optional`, defaults to 0.1):
Relative classification weight of the 'no-object' class in the object detection loss.
Examples::
>>> from transformers import DetrModel, DetrConfig
>>> # Initializing a DETR facebook/detr-resnet-50 style configuration
>>> configuration = DetrConfig()
>>> # Initializing a model from the facebook/detr-resnet-50 style configuration
>>> model = DetrModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
"""
model_type = "detr"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
num_queries=100,
max_position_embeddings=1024,
encoder_layers=6,
encoder_ffn_dim=2048,
encoder_attention_heads=8,
decoder_layers=6,
decoder_ffn_dim=2048,
decoder_attention_heads=8,
encoder_layerdrop=0.0,
decoder_layerdrop=0.0,
is_encoder_decoder=True,
activation_function="relu",
d_model=256,
dropout=0.1,
attention_dropout=0.0,
activation_dropout=0.0,
init_std=0.02,
init_xavier_std=1.0,
classifier_dropout=0.0,
scale_embedding=False,
auxiliary_loss=False,
position_embedding_type="sine",
backbone="resnet50",
dilation=False,
class_cost=1,
bbox_cost=5,
giou_cost=2,
mask_loss_coefficient=1,
dice_loss_coefficient=1,
bbox_loss_coefficient=5,
giou_loss_coefficient=2,
eos_coefficient=0.1,
**kwargs
):
super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
self.num_queries = num_queries
self.max_position_embeddings = max_position_embeddings
self.d_model = d_model
self.encoder_ffn_dim = encoder_ffn_dim
self.encoder_layers = encoder_layers
self.encoder_attention_heads = encoder_attention_heads
self.decoder_ffn_dim = decoder_ffn_dim
self.decoder_layers = decoder_layers
self.decoder_attention_heads = decoder_attention_heads
self.dropout = dropout
self.attention_dropout = attention_dropout
self.activation_dropout = activation_dropout
self.activation_function = activation_function
self.init_std = init_std
self.init_xavier_std = init_xavier_std
self.encoder_layerdrop = encoder_layerdrop
self.decoder_layerdrop = decoder_layerdrop
self.classifier_dropout = classifier_dropout
self.num_hidden_layers = encoder_layers
self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
self.auxiliary_loss = auxiliary_loss
self.position_embedding_type = position_embedding_type
self.backbone = backbone
self.dilation = dilation
# Hungarian matcher
self.class_cost = class_cost
self.bbox_cost = bbox_cost
self.giou_cost = giou_cost
# Loss coefficients
self.mask_loss_coefficient = mask_loss_coefficient
self.dice_loss_coefficient = dice_loss_coefficient
self.bbox_loss_coefficient = bbox_loss_coefficient
self.giou_loss_coefficient = giou_loss_coefficient
self.eos_coefficient = eos_coefficient
@property
def num_attention_heads(self) -> int:
return self.encoder_attention_heads
@property
def hidden_size(self) -> int:
return self.d_model
src/transformers/models/detr/convert_detr_original_pytorch_checkpoint_to_pytorch.py 0 → 100644
View file @ d3eacbb8
# coding=utf-8
# Copyright 2020 The HuggingFace Inc. team.
#
# 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.
"""Convert DETR checkpoints."""
import argparse
from collections import OrderedDict
from pathlib import Path
import torch
from PIL import Image
import requests
from transformers import DetrConfig, DetrFeatureExtractor, DetrForObjectDetection, DetrForSegmentation
from transformers.utils import logging
from transformers.utils.coco_classes import id2label
logging.set_verbosity_info()
logger = logging.get_logger(__name__)
# here we list all keys to be renamed (original name on the left, our name on the right)
rename_keys = []
for i in range(6):
# encoder layers: output projection, 2 feedforward neural networks and 2 layernorms
rename_keys.append(
(f"transformer.encoder.layers.{i}.self_attn.out_proj.weight", f"encoder.layers.{i}.self_attn.out_proj.weight")
)
rename_keys.append(
(f"transformer.encoder.layers.{i}.self_attn.out_proj.bias", f"encoder.layers.{i}.self_attn.out_proj.bias")
)
rename_keys.append((f"transformer.encoder.layers.{i}.linear1.weight", f"encoder.layers.{i}.fc1.weight"))
rename_keys.append((f"transformer.encoder.layers.{i}.linear1.bias", f"encoder.layers.{i}.fc1.bias"))
rename_keys.append((f"transformer.encoder.layers.{i}.linear2.weight", f"encoder.layers.{i}.fc2.weight"))
rename_keys.append((f"transformer.encoder.layers.{i}.linear2.bias", f"encoder.layers.{i}.fc2.bias"))
rename_keys.append(
(f"transformer.encoder.layers.{i}.norm1.weight", f"encoder.layers.{i}.self_attn_layer_norm.weight")
)
rename_keys.append((f"transformer.encoder.layers.{i}.norm1.bias", f"encoder.layers.{i}.self_attn_layer_norm.bias"))
rename_keys.append((f"transformer.encoder.layers.{i}.norm2.weight", f"encoder.layers.{i}.final_layer_norm.weight"))
rename_keys.append((f"transformer.encoder.layers.{i}.norm2.bias", f"encoder.layers.{i}.final_layer_norm.bias"))
# decoder layers: 2 times output projection, 2 feedforward neural networks and 3 layernorms
rename_keys.append(
(f"transformer.decoder.layers.{i}.self_attn.out_proj.weight", f"decoder.layers.{i}.self_attn.out_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.self_attn.out_proj.bias", f"decoder.layers.{i}.self_attn.out_proj.bias")
)
rename_keys.append(
(
f"transformer.decoder.layers.{i}.multihead_attn.out_proj.weight",
f"decoder.layers.{i}.encoder_attn.out_proj.weight",
)
)
rename_keys.append(
(
f"transformer.decoder.layers.{i}.multihead_attn.out_proj.bias",
f"decoder.layers.{i}.encoder_attn.out_proj.bias",
)
)
rename_keys.append((f"transformer.decoder.layers.{i}.linear1.weight", f"decoder.layers.{i}.fc1.weight"))
rename_keys.append((f"transformer.decoder.layers.{i}.linear1.bias", f"decoder.layers.{i}.fc1.bias"))
rename_keys.append((f"transformer.decoder.layers.{i}.linear2.weight", f"decoder.layers.{i}.fc2.weight"))
rename_keys.append((f"transformer.decoder.layers.{i}.linear2.bias", f"decoder.layers.{i}.fc2.bias"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.norm1.weight", f"decoder.layers.{i}.self_attn_layer_norm.weight")
)
rename_keys.append((f"transformer.decoder.layers.{i}.norm1.bias", f"decoder.layers.{i}.self_attn_layer_norm.bias"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.norm2.weight", f"decoder.layers.{i}.encoder_attn_layer_norm.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.norm2.bias", f"decoder.layers.{i}.encoder_attn_layer_norm.bias")
)
rename_keys.append((f"transformer.decoder.layers.{i}.norm3.weight", f"decoder.layers.{i}.final_layer_norm.weight"))
rename_keys.append((f"transformer.decoder.layers.{i}.norm3.bias", f"decoder.layers.{i}.final_layer_norm.bias"))
# convolutional projection + query embeddings + layernorm of decoder + class and bounding box heads
rename_keys.extend(
[
("input_proj.weight", "input_projection.weight"),
("input_proj.bias", "input_projection.bias"),
("query_embed.weight", "query_position_embeddings.weight"),
("transformer.decoder.norm.weight", "decoder.layernorm.weight"),
("transformer.decoder.norm.bias", "decoder.layernorm.bias"),
("class_embed.weight", "class_labels_classifier.weight"),
("class_embed.bias", "class_labels_classifier.bias"),
("bbox_embed.layers.0.weight", "bbox_predictor.layers.0.weight"),
("bbox_embed.layers.0.bias", "bbox_predictor.layers.0.bias"),
("bbox_embed.layers.1.weight", "bbox_predictor.layers.1.weight"),
("bbox_embed.layers.1.bias", "bbox_predictor.layers.1.bias"),
("bbox_embed.layers.2.weight", "bbox_predictor.layers.2.weight"),
("bbox_embed.layers.2.bias", "bbox_predictor.layers.2.bias"),
]
)
def rename_key(state_dict, old, new):
val = state_dict.pop(old)
state_dict[new] = val
def rename_backbone_keys(state_dict):
new_state_dict = OrderedDict()
for key, value in state_dict.items():
if "backbone.0.body" in key:
new_key = key.replace("backbone.0.body", "backbone.conv_encoder.model")
new_state_dict[new_key] = value
else:
new_state_dict[key] = value
return new_state_dict
def read_in_q_k_v(state_dict, is_panoptic=False):
prefix = ""
if is_panoptic:
prefix = "detr."
# first: transformer encoder
for i in range(6):
# read in weights + bias of input projection layer (in PyTorch's MultiHeadAttention, this is a single matrix + bias)
in_proj_weight = state_dict.pop(f"{prefix}transformer.encoder.layers.{i}.self_attn.in_proj_weight")
in_proj_bias = state_dict.pop(f"{prefix}transformer.encoder.layers.{i}.self_attn.in_proj_bias")
# next, add query, keys and values (in that order) to the state dict
state_dict[f"encoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:256, :]
state_dict[f"encoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:256]
state_dict[f"encoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[256:512, :]
state_dict[f"encoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[256:512]
state_dict[f"encoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-256:, :]
state_dict[f"encoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-256:]
# next: transformer decoder (which is a bit more complex because it also includes cross-attention)
for i in range(6):
# read in weights + bias of input projection layer of self-attention
in_proj_weight = state_dict.pop(f"{prefix}transformer.decoder.layers.{i}.self_attn.in_proj_weight")
in_proj_bias = state_dict.pop(f"{prefix}transformer.decoder.layers.{i}.self_attn.in_proj_bias")
# next, add query, keys and values (in that order) to the state dict
state_dict[f"decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:256, :]
state_dict[f"decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:256]
state_dict[f"decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[256:512, :]
state_dict[f"decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[256:512]
state_dict[f"decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-256:, :]
state_dict[f"decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-256:]
# read in weights + bias of input projection layer of cross-attention
in_proj_weight_cross_attn = state_dict.pop(
f"{prefix}transformer.decoder.layers.{i}.multihead_attn.in_proj_weight"
)
in_proj_bias_cross_attn = state_dict.pop(f"{prefix}transformer.decoder.layers.{i}.multihead_attn.in_proj_bias")
# next, add query, keys and values (in that order) of cross-attention to the state dict
state_dict[f"decoder.layers.{i}.encoder_attn.q_proj.weight"] = in_proj_weight_cross_attn[:256, :]
state_dict[f"decoder.layers.{i}.encoder_attn.q_proj.bias"] = in_proj_bias_cross_attn[:256]
state_dict[f"decoder.layers.{i}.encoder_attn.k_proj.weight"] = in_proj_weight_cross_attn[256:512, :]
state_dict[f"decoder.layers.{i}.encoder_attn.k_proj.bias"] = in_proj_bias_cross_attn[256:512]
state_dict[f"decoder.layers.{i}.encoder_attn.v_proj.weight"] = in_proj_weight_cross_attn[-256:, :]
state_dict[f"decoder.layers.{i}.encoder_attn.v_proj.bias"] = in_proj_bias_cross_attn[-256:]
# We will verify our results on an image of cute cats
def prepare_img():
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
im = Image.open(requests.get(url, stream=True).raw)
return im
@torch.no_grad()
def convert_detr_checkpoint(model_name, pytorch_dump_folder_path):
"""
Copy/paste/tweak model's weights to our DETR structure.
"""
# load default config
config = DetrConfig()
# set backbone and dilation attributes
if "resnet101" in model_name:
config.backbone = "resnet101"
if "dc5" in model_name:
config.dilation = True
is_panoptic = "panoptic" in model_name
if is_panoptic:
config.num_labels = 250
else:
config.num_labels = 91
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
# load feature extractor
format = "coco_panoptic" if is_panoptic else "coco_detection"
feature_extractor = DetrFeatureExtractor(format=format)
# prepare image
img = prepare_img()
encoding = feature_extractor(images=img, return_tensors="pt")
pixel_values = encoding["pixel_values"]
logger.info(f"Converting model {model_name}...")
# load original model from torch hub
detr = torch.hub.load("facebookresearch/detr", model_name, pretrained=True).eval()
state_dict = detr.state_dict()
# rename keys
for src, dest in rename_keys:
if is_panoptic:
src = "detr." + src
rename_key(state_dict, src, dest)
state_dict = rename_backbone_keys(state_dict)
# query, key and value matrices need special treatment
read_in_q_k_v(state_dict, is_panoptic=is_panoptic)
# important: we need to prepend a prefix to each of the base model keys as the head models use different attributes for them
prefix = "detr.model." if is_panoptic else "model."
for key in state_dict.copy().keys():
if is_panoptic:
if (
key.startswith("detr")
and not key.startswith("class_labels_classifier")
and not key.startswith("bbox_predictor")
):
val = state_dict.pop(key)
state_dict["detr.model" + key[4:]] = val
elif "class_labels_classifier" in key or "bbox_predictor" in key:
val = state_dict.pop(key)
state_dict["detr." + key] = val
elif key.startswith("bbox_attention") or key.startswith("mask_head"):
continue
else:
val = state_dict.pop(key)
state_dict[prefix + key] = val
else:
if not key.startswith("class_labels_classifier") and not key.startswith("bbox_predictor"):
val = state_dict.pop(key)
state_dict[prefix + key] = val
# finally, create HuggingFace model and load state dict
model = DetrForSegmentation(config) if is_panoptic else DetrForObjectDetection(config)
model.load_state_dict(state_dict)
model.eval()
# verify our conversion
original_outputs = detr(pixel_values)
outputs = model(pixel_values)
assert torch.allclose(outputs.logits, original_outputs["pred_logits"], atol=1e-4)
assert torch.allclose(outputs.pred_boxes, original_outputs["pred_boxes"], atol=1e-4)
if is_panoptic:
assert torch.allclose(outputs.pred_masks, original_outputs["pred_masks"], atol=1e-4)
# Save model and feature extractor
logger.info(f"Saving PyTorch model and feature extractor to {pytorch_dump_folder_path}...")
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
model.save_pretrained(pytorch_dump_folder_path)
feature_extractor.save_pretrained(pytorch_dump_folder_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--model_name", default="detr_resnet50", type=str, help="Name of the DETR model you'd like to convert."
)
parser.add_argument(
"--pytorch_dump_folder_path", default=None, type=str, help="Path to the folder to output PyTorch model."
)
args = parser.parse_args()
convert_detr_checkpoint(args.model_name, args.pytorch_dump_folder_path)
src/transformers/models/detr/feature_extraction_detr.py 0 → 100644
View file @ d3eacbb8
This diff is collapsed.
src/transformers/models/detr/modeling_detr.py 0 → 100644
View file @ d3eacbb8
This diff is collapsed.
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