Skip to content

GitLab

"Dockerfile" did not exist on "86a26d1eb0ce5498e8b28b704a94f287e5df86e9"
Unverified Commit 126a7390 authored by DepuMeng's avatar DepuMeng Committed by GitHub
Browse files

Add support for conditional detr (#18948) 



* added conditional_detr files

* checked copies

* checked copies

* fixed style and copies

* fixed style and copies

* fixed hub

* fixed style

* Update README.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/_toctree.yml
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/index.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/conditional_detr.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixed some doc issue

* changed prefix to ConditionalDetr

* fixed docs

* Update README_ko.md

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fixed use_pretrained issue

* changed post-process

* added conditional_detr files

* checked copies

* checked copies

* fixed style and copies

* fixed style and copies

* fixed hub

* fixed style

* Update README.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/_toctree.yml
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/index.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixed some doc issue

* Update docs/source/en/model_doc/conditional_detr.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* changed prefix to ConditionalDetr

* fixed docs

* Update README_ko.md

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fixed use_pretrained issue

* changed post-process

* fix style quality and copies

* fix style quality and copies

* fix style quality and copies

* fix style quality and copies

* add more fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixed some variable names & added more fix-copies

* fixed some variable names & added more fix-copies

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added more copied from

* fixed quality

* changed pretrained config

* added more copied-from and fixed the issue in feature_extraction_auto

* added conditional_detr files

* checked copies

* checked copies

* fixed style and copies

* fixed style and copies

* fixed hub

* fixed style

* Update README.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/_toctree.yml
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/index.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixed some doc issue

* Update docs/source/en/model_doc/conditional_detr.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* changed prefix to ConditionalDetr

* fixed docs

* Update README_ko.md

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fixed use_pretrained issue

* changed post-process

* added conditional_detr files

* checked copies

* fixed style and copies

* fixed some doc issue

* changed prefix to ConditionalDetr

* fixed docs

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fix style quality and copies

* fix style quality and copies

* fix style quality and copies

* add more fix-copies

* fixed some variable names & added more fix-copies

* fixed some variable names & added more fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added more copied from

* fixed quality

* changed pretrained config

* added more copied-from and fixed the issue in feature_extraction_auto

* fixed style

* added conditional_detr files

* checked copies

* checked copies

* fixed style and copies

* fixed style and copies

* fixed hub

* fixed style

* Update README.md
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/_toctree.yml
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/index.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixed some doc issue

* Update docs/source/en/model_doc/conditional_detr.mdx
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* changed prefix to ConditionalDetr

* fixed docs

* Update README_ko.md

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fixed use_pretrained issue

* changed post-process

* added conditional_detr files

* checked copies

* fixed style and copies

* fixed some doc issue

* changed prefix to ConditionalDetr

* fixed docs

* added spatial_model_name

* fixed fix-copies

* Update src/transformers/models/conditional_detr/modeling_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added some copied from

* added some copied from

* added some copied from

* added some copied from

* fix style quality and copies

* fix style quality and copies

* fix style quality and copies

* add more fix-copies

* fixed some variable names & added more fix-copies

* fixed some variable names & added more fix-copies

* Update src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/conditional_detr/configuration_conditional_detr.py
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>

* added more copied from

* fixed quality

* changed pretrained config

* added more copied-from and fixed the issue in feature_extraction_auto

* rebased
Co-authored-by: default avatarNielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: default avatarDepu Meng <depumeng@Depus-MacBook-Pro.local>
parent c7fd2899
Show whitespace changes
Inline Side-by-side
Showing with 4354 additions and 1 deletion
README.md
View file @ 126a7390
......@@ -278,6 +278,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
1. **[CANINE](https://huggingface.co/docs/transformers/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. **[CLIP](https://huggingface.co/docs/transformers/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. **[CodeGen](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/main/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](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/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. **[CPM](https://huggingface.co/docs/transformers/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.
......
README_ko.md
View file @ 126a7390
......@@ -228,6 +228,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
1. **[CANINE](https://huggingface.co/docs/transformers/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. **[CLIP](https://huggingface.co/docs/transformers/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. **[CodeGen](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/main/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](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/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. **[CPM](https://huggingface.co/docs/transformers/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.
......
README_zh-hans.md
View file @ 126a7390
......@@ -252,6 +252,7 @@ conda install -c huggingface transformers
1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (来自 Google Research) 伴随论文 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 由 Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 发布。
1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (来自 OpenAI) 伴随论文 [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) 由 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 发布。
1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (来自 Salesforce) 伴随论文 [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) 由 Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong 发布。
1. **[Conditional DETR](https://huggingface.co/docs/transformers/main/model_doc/conditional_detr)** (来自 Microsoft Research Asia) 伴随论文 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 由 Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 发布。
1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (来自 YituTech) 伴随论文 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 由 Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 发布。
1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。
1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (来自 Tsinghua University) 伴随论文 [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 由 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 发布。
......
README_zh-hant.md
View file @ 126a7390
......@@ -264,6 +264,7 @@ conda install -c huggingface transformers
1. **[CANINE](https://huggingface.co/docs/transformers/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. **[CLIP](https://huggingface.co/docs/transformers/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. **[CodeGen](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/main/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](https://huggingface.co/docs/transformers/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](https://huggingface.co/docs/transformers/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. **[CPM](https://huggingface.co/docs/transformers/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.
......
docs/source/en/_toctree.yml
View file @ 126a7390
......@@ -362,6 +362,8 @@
sections:
- local: model_doc/beit
title: BEiT
- local: model_doc/conditional_detr
title: Conditional DETR
- local: model_doc/convnext
title: ConvNeXT
- local: model_doc/cvt
......
docs/source/en/index.mdx
View file @ 126a7390
......@@ -68,6 +68,7 @@ The documentation is organized into five sections:
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. **[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. **[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. **[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.
......@@ -215,6 +216,7 @@ Flax), PyTorch, and/or TensorFlow.
| CANINE | ✅ | ❌ | ✅ | ❌ | ❌ |
| CLIP | ✅ | ✅ | ✅ | ✅ | ✅ |
| CodeGen | ✅ | ✅ | ✅ | ❌ | ❌ |
| Conditional DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ |
| ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ |
| CTRL | ✅ | ❌ | ✅ | ✅ | ❌ |
......
docs/source/en/model_doc/conditional_detr.mdx 0 → 100644
View file @ 126a7390
<!--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.
-->
# Conditional DETR
## Overview
The Conditional DETR model was proposed in [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. Conditional DETR presents a conditional cross-attention mechanism for fast DETR training. Conditional DETR converges 6.7× to 10× faster than DETR.
The abstract from the paper is the following:
*The recently-developed DETR approach applies the transformer encoder and decoder architecture to object detection and achieves promising performance. In this paper, we handle the critical issue, slow training convergence, and present a conditional cross-attention mechanism for fast DETR training. Our approach is motivated by that the cross-attention in DETR relies highly on the content embeddings for localizing the four extremities and predicting the box, which increases the need for high-quality content embeddings and thus the training difficulty. Our approach, named conditional DETR, learns a conditional spatial query from the decoder embedding for decoder multi-head cross-attention. The benefit is that through the conditional spatial query, each cross-attention head is able to attend to a band containing a distinct region, e.g., one object extremity or a region inside the object box. This narrows down the spatial range for localizing the distinct regions for object classification and box regression, thus relaxing the dependence on the content embeddings and easing the training. Empirical results show that conditional DETR converges 6.7× faster for the backbones R50 and R101 and 10× faster for stronger backbones DC5-R50 and DC5-R101. Code is available at https://github.com/Atten4Vis/ConditionalDETR.*
This model was contributed by [DepuMeng](https://huggingface.co/DepuMeng). The original code can be found [here](https://github.com/Atten4Vis/ConditionalDETR).
## ConditionalDetrConfig
[[autodoc]] ConditionalDetrConfig
## ConditionalDetrFeatureExtractor
[[autodoc]] ConditionalDetrFeatureExtractor
- __call__
- pad_and_create_pixel_mask
- post_process
- post_process_segmentation
- post_process_panoptic
## ConditionalDetrModel
[[autodoc]] ConditionalDetrModel
- forward
## ConditionalDetrForObjectDetection
[[autodoc]] ConditionalDetrForObjectDetection
- forward
## ConditionalDetrForSegmentation
[[autodoc]] ConditionalDetrForSegmentation
- forward
\ No newline at end of file
docs/source/en/serialization.mdx
View file @ 126a7390
......@@ -57,6 +57,7 @@ Ready-made configurations include the following architectures:
- CamemBERT
- CLIP
- CodeGen
- Conditional DETR
- ConvBERT
- ConvNeXT
- Data2VecText
......
src/transformers/__init__.py
View file @ 126a7390
......@@ -172,6 +172,7 @@ _import_structure = {
"CLIPVisionConfig",
],
"models.codegen": ["CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP", "CodeGenConfig", "CodeGenTokenizer"],
"models.conditional_detr": ["CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConditionalDetrConfig"],
"models.convbert": ["CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvBertConfig", "ConvBertTokenizer"],
"models.convnext": ["CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvNextConfig"],
"models.cpm": [],
......@@ -660,6 +661,7 @@ else:
_import_structure["models.convnext"].append("ConvNextFeatureExtractor")
_import_structure["models.deit"].append("DeiTFeatureExtractor")
_import_structure["models.detr"].append("DetrFeatureExtractor")
_import_structure["models.conditional_detr"].append("ConditionalDetrFeatureExtractor")
_import_structure["models.donut"].append("DonutFeatureExtractor")
_import_structure["models.dpt"].append("DPTFeatureExtractor")
_import_structure["models.flava"].extend(["FlavaFeatureExtractor", "FlavaProcessor"])
......@@ -708,6 +710,15 @@ else:
"DetrPreTrainedModel",
]
)
_import_structure["models.conditional_detr"].extend(
[
"CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
"ConditionalDetrForObjectDetection",
"ConditionalDetrForSegmentation",
"ConditionalDetrModel",
"ConditionalDetrPreTrainedModel",
]
)
try:
if not is_scatter_available():
......@@ -3075,6 +3086,7 @@ if TYPE_CHECKING:
CLIPVisionConfig,
)
from .models.codegen import CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP, CodeGenConfig, CodeGenTokenizer
from .models.conditional_detr import CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, ConditionalDetrConfig
from .models.convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig, ConvBertTokenizer
from .models.convnext import CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvNextConfig
from .models.ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig, CTRLTokenizer
......@@ -3498,6 +3510,7 @@ if TYPE_CHECKING:
from .image_utils import ImageFeatureExtractionMixin
from .models.beit import BeitFeatureExtractor
from .models.clip import CLIPFeatureExtractor
from .models.conditional_detr import ConditionalDetrFeatureExtractor
from .models.convnext import ConvNextFeatureExtractor
from .models.deit import DeiTFeatureExtractor
from .models.detr import DetrFeatureExtractor
......@@ -3527,6 +3540,13 @@ if TYPE_CHECKING:
except OptionalDependencyNotAvailable:
from .utils.dummy_timm_and_vision_objects import *
else:
from .models.conditional_detr import (
CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
ConditionalDetrForObjectDetection,
ConditionalDetrForSegmentation,
ConditionalDetrModel,
ConditionalDetrPreTrainedModel,
)
from .models.deformable_detr import (
DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
DeformableDetrForObjectDetection,
......
src/transformers/models/__init__.py
View file @ 126a7390
......@@ -38,6 +38,7 @@ from . import (
canine,
clip,
codegen,
conditional_detr,
convbert,
convnext,
cpm,
......
src/transformers/models/auto/configuration_auto.py
View file @ 126a7390
......@@ -43,6 +43,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
("canine", "CanineConfig"),
("clip", "CLIPConfig"),
("codegen", "CodeGenConfig"),
("conditional_detr", "ConditionalDetrConfig"),
("convbert", "ConvBertConfig"),
("convnext", "ConvNextConfig"),
("ctrl", "CTRLConfig"),
......@@ -175,6 +176,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
("canine", "CANINE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("clip", "CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("codegen", "CODEGEN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("conditional_detr", "CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("convbert", "CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("convnext", "CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
("ctrl", "CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
......@@ -300,6 +302,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
("canine", "CANINE"),
("clip", "CLIP"),
("codegen", "CodeGen"),
("conditional_detr", "Conditional DETR"),
("convbert", "ConvBERT"),
("convnext", "ConvNeXT"),
("cpm", "CPM"),
......
src/transformers/models/auto/feature_extraction_auto.py
View file @ 126a7390
......@@ -39,6 +39,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
[
("beit", "BeitFeatureExtractor"),
("clip", "CLIPFeatureExtractor"),
("conditional_detr", "ConditionalDetrFeatureExtractor"),
("convnext", "ConvNextFeatureExtractor"),
("cvt", "ConvNextFeatureExtractor"),
("data2vec-audio", "Wav2Vec2FeatureExtractor"),
......@@ -46,7 +47,6 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
("deformable_detr", "DetrFeatureExtractor"),
("deit", "DeiTFeatureExtractor"),
("detr", "DetrFeatureExtractor"),
("detr", "DetrFeatureExtractor"),
("donut", "DonutFeatureExtractor"),
("dpt", "DPTFeatureExtractor"),
("flava", "FlavaFeatureExtractor"),
......
src/transformers/models/auto/modeling_auto.py
View file @ 126a7390
......@@ -42,6 +42,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
("canine", "CanineModel"),
("clip", "CLIPModel"),
("codegen", "CodeGenModel"),
("conditional_detr", "ConditionalDetrModel"),
("convbert", "ConvBertModel"),
("convnext", "ConvNextModel"),
("ctrl", "CTRLModel"),
......@@ -455,6 +456,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
[
# Model for Object Detection mapping
("conditional_detr", "ConditionalDetrForObjectDetection"),
("deformable_detr", "DeformableDetrForObjectDetection"),
("detr", "DetrForObjectDetection"),
("yolos", "YolosForObjectDetection"),
......
src/transformers/models/conditional_detr/__init__.py 0 → 100644
View file @ 126a7390
# 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 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.
from typing import TYPE_CHECKING
from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_timm_available, is_vision_available
_import_structure = {
"configuration_conditional_detr": [
"CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP",
"ConditionalDetrConfig",
"ConditionalDetrOnnxConfig",
]
}
try:
if not is_vision_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["feature_extraction_conditional_detr"] = ["ConditionalDetrFeatureExtractor"]
try:
if not is_timm_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_conditional_detr"] = [
"CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
"ConditionalDetrForObjectDetection",
"ConditionalDetrForSegmentation",
"ConditionalDetrModel",
"ConditionalDetrPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_conditional_detr import (
CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP,
ConditionalDetrConfig,
ConditionalDetrOnnxConfig,
)
try:
if not is_vision_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .feature_extraction_conditional_detr import ConditionalDetrFeatureExtractor
try:
if not is_timm_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_conditional_detr import (
CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
ConditionalDetrForObjectDetection,
ConditionalDetrForSegmentation,
ConditionalDetrModel,
ConditionalDetrPreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
src/transformers/models/conditional_detr/configuration_conditional_detr.py 0 → 100644
View file @ 126a7390
# coding=utf-8
# Copyright 2022 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.
""" Conditional DETR model configuration"""
from collections import OrderedDict
from typing import Mapping
from packaging import version
from ...configuration_utils import PretrainedConfig
from ...onnx import OnnxConfig
from ...utils import logging
logger = logging.get_logger(__name__)
CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"microsoft/conditional-detr-resnet-50": (
"https://huggingface.co/microsoft/conditional-detr-resnet-50/resolve/main/config.json"
),
}
class ConditionalDetrConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`ConditionalDetrModel`]. It is used to instantiate
a Conditional 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 Conditional DETR
[microsoft/conditional-detr-resnet-50](https://huggingface.co/microsoft/conditional-detr-resnet-50) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
num_channels (`int`, *optional*, defaults to 3):
The number of input channels.
num_queries (`int`, *optional*, defaults to 100):
Number of object queries, i.e. detection slots. This is the maximal number of objects
[`ConditionalDetrModel`] can detect in a single image. For COCO, we recommend 100 queries.
d_model (`int`, *optional*, defaults to 256):
Dimension of the layers.
encoder_layers (`int`, *optional*, defaults to 6):
Number of encoder layers.
decoder_layers (`int`, *optional*, defaults to 6):
Number of decoder layers.
encoder_attention_heads (`int`, *optional*, defaults to 8):
Number of attention heads for each attention layer in the Transformer encoder.
decoder_attention_heads (`int`, *optional*, defaults to 8):
Number of attention heads for each attention layer in the Transformer decoder.
decoder_ffn_dim (`int`, *optional*, defaults to 2048):
Dimension of the "intermediate" (often named feed-forward) layer in decoder.
encoder_ffn_dim (`int`, *optional*, defaults to 2048):
Dimension of the "intermediate" (often named feed-forward) layer in decoder.
activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"silu"` and `"gelu_new"` are supported.
dropout (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
activation_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for activations inside the fully connected layer.
init_std (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
init_xavier_std (`float`, *optional*, defaults to 1):
The scaling factor used for the Xavier initialization gain in the HM Attention map module.
encoder_layerdrop (`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 (`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 (`bool`, *optional*, defaults to `False`):
Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
position_embedding_type (`str`, *optional*, defaults to `"sine"`):
Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.
backbone (`str`, *optional*, defaults to `"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).
use_pretrained_backbone (`bool`, *optional*, defaults to `True`):
Whether to use pretrained weights for the backbone.
dilation (`bool`, *optional*, defaults to `False`):
Whether to replace stride with dilation in the last convolutional block (DC5).
class_cost (`float`, *optional*, defaults to 1):
Relative weight of the classification error in the Hungarian matching cost.
bbox_cost (`float`, *optional*, defaults to 5):
Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.
giou_cost (`float`, *optional*, defaults to 2):
Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.
mask_loss_coefficient (`float`, *optional*, defaults to 1):
Relative weight of the Focal loss in the panoptic segmentation loss.
dice_loss_coefficient (`float`, *optional*, defaults to 1):
Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.
bbox_loss_coefficient (`float`, *optional*, defaults to 5):
Relative weight of the L1 bounding box loss in the object detection loss.
giou_loss_coefficient (`float`, *optional*, defaults to 2):
Relative weight of the generalized IoU loss in the object detection loss.
eos_coefficient (`float`, *optional*, defaults to 0.1):
Relative classification weight of the 'no-object' class in the object detection loss.
Examples:
```python
>>> from transformers import ConditionalDetrModel, ConditionalDetrConfig
>>> # Initializing a Conditional DETR microsoft/conditional-detr-resnet-50 style configuration
>>> configuration = ConditionalDetrConfig()
>>> # Initializing a model from the microsoft/conditional-detr-resnet-50 style configuration
>>> model = ConditionalDetrModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "conditional_detr"
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {
"hidden_size": "d_model",
"num_attention_heads": "encoder_attention_heads",
}
def __init__(
self,
num_channels=3,
num_queries=300,
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",
use_pretrained_backbone=True,
dilation=False,
class_cost=2,
bbox_cost=5,
giou_cost=2,
mask_loss_coefficient=1,
dice_loss_coefficient=1,
cls_loss_coefficient=2,
bbox_loss_coefficient=5,
giou_loss_coefficient=2,
focal_alpha=0.25,
**kwargs
):
self.num_channels = num_channels
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.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.use_pretrained_backbone = use_pretrained_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.cls_loss_coefficient = cls_loss_coefficient
self.bbox_loss_coefficient = bbox_loss_coefficient
self.giou_loss_coefficient = giou_loss_coefficient
self.focal_alpha = focal_alpha
super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
@property
def num_attention_heads(self) -> int:
return self.encoder_attention_heads
@property
def hidden_size(self) -> int:
return self.d_model
class ConditionalDetrOnnxConfig(OnnxConfig):
torch_onnx_minimum_version = version.parse("1.11")
@property
def inputs(self) -> Mapping[str, Mapping[int, str]]:
return OrderedDict(
[
("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),
("pixel_mask", {0: "batch"}),
]
)
@property
def atol_for_validation(self) -> float:
return 1e-5
@property
def default_onnx_opset(self) -> int:
return 12
src/transformers/models/conditional_detr/convert_conditional_detr_original_pytorch_checkpoint_to_pytorch.py 0 → 100644
View file @ 126a7390
# coding=utf-8
# Copyright 2022 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 Conditional DETR checkpoints."""
import argparse
import json
from collections import OrderedDict
from pathlib import Path
import torch
from PIL import Image
import requests
from huggingface_hub import hf_hub_download
from transformers import (
ConditionalDetrConfig,
ConditionalDetrFeatureExtractor,
ConditionalDetrForObjectDetection,
ConditionalDetrForSegmentation,
)
from transformers.utils import logging
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}.cross_attn.out_proj.weight",
f"decoder.layers.{i}.encoder_attn.out_proj.weight",
)
)
rename_keys.append(
(
f"transformer.decoder.layers.{i}.cross_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"))
# q, k, v projections in self/cross-attention in decoder for conditional DETR
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_qcontent_proj.weight", f"decoder.layers.{i}.sa_qcontent_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_kcontent_proj.weight", f"decoder.layers.{i}.sa_kcontent_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_qpos_proj.weight", f"decoder.layers.{i}.sa_qpos_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_kpos_proj.weight", f"decoder.layers.{i}.sa_kpos_proj.weight")
)
rename_keys.append((f"transformer.decoder.layers.{i}.sa_v_proj.weight", f"decoder.layers.{i}.sa_v_proj.weight"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_qcontent_proj.weight", f"decoder.layers.{i}.ca_qcontent_proj.weight")
)
# rename_keys.append((f"transformer.decoder.layers.{i}.ca_qpos_proj.weight", f"decoder.layers.{i}.ca_qpos_proj.weight"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_kcontent_proj.weight", f"decoder.layers.{i}.ca_kcontent_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_kpos_proj.weight", f"decoder.layers.{i}.ca_kpos_proj.weight")
)
rename_keys.append((f"transformer.decoder.layers.{i}.ca_v_proj.weight", f"decoder.layers.{i}.ca_v_proj.weight"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_qpos_sine_proj.weight", f"decoder.layers.{i}.ca_qpos_sine_proj.weight")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_qcontent_proj.bias", f"decoder.layers.{i}.sa_qcontent_proj.bias")
)
rename_keys.append(
(f"transformer.decoder.layers.{i}.sa_kcontent_proj.bias", f"decoder.layers.{i}.sa_kcontent_proj.bias")
)
rename_keys.append((f"transformer.decoder.layers.{i}.sa_qpos_proj.bias", f"decoder.layers.{i}.sa_qpos_proj.bias"))
rename_keys.append((f"transformer.decoder.layers.{i}.sa_kpos_proj.bias", f"decoder.layers.{i}.sa_kpos_proj.bias"))
rename_keys.append((f"transformer.decoder.layers.{i}.sa_v_proj.bias", f"decoder.layers.{i}.sa_v_proj.bias"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_qcontent_proj.bias", f"decoder.layers.{i}.ca_qcontent_proj.bias")
)
# rename_keys.append((f"transformer.decoder.layers.{i}.ca_qpos_proj.bias", f"decoder.layers.{i}.ca_qpos_proj.bias"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_kcontent_proj.bias", f"decoder.layers.{i}.ca_kcontent_proj.bias")
)
rename_keys.append((f"transformer.decoder.layers.{i}.ca_kpos_proj.bias", f"decoder.layers.{i}.ca_kpos_proj.bias"))
rename_keys.append((f"transformer.decoder.layers.{i}.ca_v_proj.bias", f"decoder.layers.{i}.ca_v_proj.bias"))
rename_keys.append(
(f"transformer.decoder.layers.{i}.ca_qpos_sine_proj.bias", f"decoder.layers.{i}.ca_qpos_sine_proj.bias")
)
# convolutional projection + query embeddings + layernorm of decoder + class and bounding box heads
# for conditional DETR, also convert reference point head and query scale MLP
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"),
("transformer.decoder.ref_point_head.layers.0.weight", "decoder.ref_point_head.layers.0.weight"),
("transformer.decoder.ref_point_head.layers.0.bias", "decoder.ref_point_head.layers.0.bias"),
("transformer.decoder.ref_point_head.layers.1.weight", "decoder.ref_point_head.layers.1.weight"),
("transformer.decoder.ref_point_head.layers.1.bias", "decoder.ref_point_head.layers.1.bias"),
("transformer.decoder.query_scale.layers.0.weight", "decoder.query_scale.layers.0.weight"),
("transformer.decoder.query_scale.layers.0.bias", "decoder.query_scale.layers.0.bias"),
("transformer.decoder.query_scale.layers.1.weight", "decoder.query_scale.layers.1.weight"),
("transformer.decoder.query_scale.layers.1.bias", "decoder.query_scale.layers.1.bias"),
("transformer.decoder.layers.0.ca_qpos_proj.weight", "decoder.layers.0.ca_qpos_proj.weight"),
("transformer.decoder.layers.0.ca_qpos_proj.bias", "decoder.layers.0.ca_qpos_proj.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 = "conditional_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:]
# 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_conditional_detr_checkpoint(model_name, pytorch_dump_folder_path):
"""
Copy/paste/tweak model's weights to our CONDITIONAL_DETR structure.
"""
# load default config
config = ConditionalDetrConfig()
# 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
repo_id = "datasets/huggingface/label-files"
filename = "coco-detection-id2label.json"
id2label = json.load(open(hf_hub_download(repo_id, filename), "r"))
id2label = {int(k): v for k, v in id2label.items()}
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 = ConditionalDetrFeatureExtractor(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
conditional_detr = torch.hub.load("DeppMeng/ConditionalDETR", model_name, pretrained=True).eval()
state_dict = conditional_detr.state_dict()
# rename keys
for src, dest in rename_keys:
if is_panoptic:
src = "conditional_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 = "conditional_detr.model." if is_panoptic else "model."
for key in state_dict.copy().keys():
if is_panoptic:
if (
key.startswith("conditional_detr")
and not key.startswith("class_labels_classifier")
and not key.startswith("bbox_predictor")
):
val = state_dict.pop(key)
state_dict["conditional_detr.model" + key[4:]] = val
elif "class_labels_classifier" in key or "bbox_predictor" in key:
val = state_dict.pop(key)
state_dict["conditional_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 = ConditionalDetrForSegmentation(config) if is_panoptic else ConditionalDetrForObjectDetection(config)
model.load_state_dict(state_dict)
model.eval()
model.push_to_hub(repo_id=model_name, organization="DepuMeng", commit_message="Add model")
# verify our conversion
original_outputs = conditional_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="conditional_detr_resnet50",
type=str,
help="Name of the CONDITIONAL_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_conditional_detr_checkpoint(args.model_name, args.pytorch_dump_folder_path)
src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py 0 → 100644
View file @ 126a7390
# coding=utf-8
# Copyright 2022 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.
"""Feature extractor class for Conditional DETR."""
import io
import pathlib
from collections import defaultdict
from typing import Dict, List, Optional, Union
import numpy as np
from PIL import Image
from ...feature_extraction_utils import BatchFeature, FeatureExtractionMixin
from ...image_utils import ImageFeatureExtractionMixin, is_torch_tensor
from ...utils import TensorType, is_torch_available, logging
if is_torch_available():
import torch
from torch import nn
logger = logging.get_logger(__name__)
ImageInput = Union[Image.Image, np.ndarray, "torch.Tensor", List[Image.Image], List[np.ndarray], List["torch.Tensor"]]
# Copied from transformers.models.detr.feature_extraction_detr.center_to_corners_format
def center_to_corners_format(x):
"""
Converts a PyTorch tensor of bounding boxes of center format (center_x, center_y, width, height) to corners format
(x_0, y_0, x_1, y_1).
"""
x_c, y_c, w, h = x.unbind(-1)
b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)]
return torch.stack(b, dim=-1)
# Copied from transformers.models.detr.feature_extraction_detr.corners_to_center_format
def corners_to_center_format(x):
"""
Converts a NumPy array of bounding boxes of shape (number of bounding boxes, 4) of corners format (x_0, y_0, x_1,
y_1) to center format (center_x, center_y, width, height).
"""
x_transposed = x.T
x0, y0, x1, y1 = x_transposed[0], x_transposed[1], x_transposed[2], x_transposed[3]
b = [(x0 + x1) / 2, (y0 + y1) / 2, (x1 - x0), (y1 - y0)]
return np.stack(b, axis=-1)
# Copied from transformers.models.detr.feature_extraction_detr.masks_to_boxes
def masks_to_boxes(masks):
"""
Compute the bounding boxes around the provided panoptic segmentation masks.
The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions.
Returns a [N, 4] tensor, with the boxes in corner (xyxy) format.
"""
if masks.size == 0:
return np.zeros((0, 4))
h, w = masks.shape[-2:]
y = np.arange(0, h, dtype=np.float32)
x = np.arange(0, w, dtype=np.float32)
# see https://github.com/pytorch/pytorch/issues/50276
y, x = np.meshgrid(y, x, indexing="ij")
x_mask = masks * np.expand_dims(x, axis=0)
x_max = x_mask.reshape(x_mask.shape[0], -1).max(-1)
x = np.ma.array(x_mask, mask=~(np.array(masks, dtype=bool)))
x_min = x.filled(fill_value=1e8)
x_min = x_min.reshape(x_min.shape[0], -1).min(-1)
y_mask = masks * np.expand_dims(y, axis=0)
y_max = y_mask.reshape(x_mask.shape[0], -1).max(-1)
y = np.ma.array(y_mask, mask=~(np.array(masks, dtype=bool)))
y_min = y.filled(fill_value=1e8)
y_min = y_min.reshape(y_min.shape[0], -1).min(-1)
return np.stack([x_min, y_min, x_max, y_max], 1)
# Copied from transformers.models.detr.feature_extraction_detr.rgb_to_id
def rgb_to_id(color):
if isinstance(color, np.ndarray) and len(color.shape) == 3:
if color.dtype == np.uint8:
color = color.astype(np.int32)
return color[:, :, 0] + 256 * color[:, :, 1] + 256 * 256 * color[:, :, 2]
return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
# Copied from transformers.models.detr.feature_extraction_detr.id_to_rgb
def id_to_rgb(id_map):
if isinstance(id_map, np.ndarray):
id_map_copy = id_map.copy()
rgb_shape = tuple(list(id_map.shape) + [3])
rgb_map = np.zeros(rgb_shape, dtype=np.uint8)
for i in range(3):
rgb_map[..., i] = id_map_copy % 256
id_map_copy //= 256
return rgb_map
color = []
for _ in range(3):
color.append(id_map % 256)
id_map //= 256
return color
class ConditionalDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
r"""
Constructs a Conditional DETR feature extractor.
This feature extractor inherits from [`FeatureExtractionMixin`] which contains most of the main methods. Users
should refer to this superclass for more information regarding those methods.
Args:
format (`str`, *optional*, defaults to `"coco_detection"`):
Data format of the annotations. One of "coco_detection" or "coco_panoptic".
do_resize (`bool`, *optional*, defaults to `True`):
Whether to resize the input to a certain `size`.
size (`int`, *optional*, defaults to 800):
Resize the input to the given size. Only has an effect if `do_resize` is set to `True`. If size is a
sequence like `(width, height)`, output size will be matched to this. If size is an int, smaller edge of
the image will be matched to this number. i.e, if `height > width`, then image will be rescaled to `(size *
height / width, size)`.
max_size (`int`, *optional*, defaults to `1333`):
The largest size an image dimension can have (otherwise it's capped). Only has an effect if `do_resize` is
set to `True`.
do_normalize (`bool`, *optional*, defaults to `True`):
Whether or not to normalize the input with mean and standard deviation.
image_mean (`int`, *optional*, defaults to `[0.485, 0.456, 0.406]`):
The sequence of means for each channel, to be used when normalizing images. Defaults to the ImageNet mean.
image_std (`int`, *optional*, defaults to `[0.229, 0.224, 0.225]`):
The sequence of standard deviations for each channel, to be used when normalizing images. Defaults to the
ImageNet std.
"""
model_input_names = ["pixel_values", "pixel_mask"]
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.__init__
def __init__(
self,
format="coco_detection",
do_resize=True,
size=800,
max_size=1333,
do_normalize=True,
image_mean=None,
image_std=None,
**kwargs
):
super().__init__(**kwargs)
self.format = self._is_valid_format(format)
self.do_resize = do_resize
self.size = size
self.max_size = max_size
self.do_normalize = do_normalize
self.image_mean = image_mean if image_mean is not None else [0.485, 0.456, 0.406] # ImageNet mean
self.image_std = image_std if image_std is not None else [0.229, 0.224, 0.225] # ImageNet std
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor._is_valid_format
def _is_valid_format(self, format):
if format not in ["coco_detection", "coco_panoptic"]:
raise ValueError(f"Format {format} not supported")
return format
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.prepare
def prepare(self, image, target, return_segmentation_masks=False, masks_path=None):
if self.format == "coco_detection":
image, target = self.prepare_coco_detection(image, target, return_segmentation_masks)
return image, target
elif self.format == "coco_panoptic":
image, target = self.prepare_coco_panoptic(image, target, masks_path)
return image, target
else:
raise ValueError(f"Format {self.format} not supported")
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.convert_coco_poly_to_mask
def convert_coco_poly_to_mask(self, segmentations, height, width):
try:
from pycocotools import mask as coco_mask
except ImportError:
raise ImportError("Pycocotools is not installed in your environment.")
masks = []
for polygons in segmentations:
rles = coco_mask.frPyObjects(polygons, height, width)
mask = coco_mask.decode(rles)
if len(mask.shape) < 3:
mask = mask[..., None]
mask = np.asarray(mask, dtype=np.uint8)
mask = np.any(mask, axis=2)
masks.append(mask)
if masks:
masks = np.stack(masks, axis=0)
else:
masks = np.zeros((0, height, width), dtype=np.uint8)
return masks
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.prepare_coco_detection with DETR->ConditionalDETR
def prepare_coco_detection(self, image, target, return_segmentation_masks=False):
"""
Convert the target in COCO format into the format expected by ConditionalDETR.
"""
w, h = image.size
image_id = target["image_id"]
image_id = np.asarray([image_id], dtype=np.int64)
# get all COCO annotations for the given image
anno = target["annotations"]
anno = [obj for obj in anno if "iscrowd" not in obj or obj["iscrowd"] == 0]
boxes = [obj["bbox"] for obj in anno]
# guard against no boxes via resizing
boxes = np.asarray(boxes, dtype=np.float32).reshape(-1, 4)
boxes[:, 2:] += boxes[:, :2]
boxes[:, 0::2] = boxes[:, 0::2].clip(min=0, max=w)
boxes[:, 1::2] = boxes[:, 1::2].clip(min=0, max=h)
classes = [obj["category_id"] for obj in anno]
classes = np.asarray(classes, dtype=np.int64)
if return_segmentation_masks:
segmentations = [obj["segmentation"] for obj in anno]
masks = self.convert_coco_poly_to_mask(segmentations, h, w)
keypoints = None
if anno and "keypoints" in anno[0]:
keypoints = [obj["keypoints"] for obj in anno]
keypoints = np.asarray(keypoints, dtype=np.float32)
num_keypoints = keypoints.shape[0]
if num_keypoints:
keypoints = keypoints.reshape((-1, 3))
keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0])
boxes = boxes[keep]
classes = classes[keep]
if return_segmentation_masks:
masks = masks[keep]
if keypoints is not None:
keypoints = keypoints[keep]
target = {}
target["boxes"] = boxes
target["class_labels"] = classes
if return_segmentation_masks:
target["masks"] = masks
target["image_id"] = image_id
if keypoints is not None:
target["keypoints"] = keypoints
# for conversion to coco api
area = np.asarray([obj["area"] for obj in anno], dtype=np.float32)
iscrowd = np.asarray([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in anno], dtype=np.int64)
target["area"] = area[keep]
target["iscrowd"] = iscrowd[keep]
target["orig_size"] = np.asarray([int(h), int(w)], dtype=np.int64)
target["size"] = np.asarray([int(h), int(w)], dtype=np.int64)
return image, target
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.prepare_coco_panoptic
def prepare_coco_panoptic(self, image, target, masks_path, return_masks=True):
w, h = image.size
ann_info = target.copy()
ann_path = pathlib.Path(masks_path) / ann_info["file_name"]
if "segments_info" in ann_info:
masks = np.asarray(Image.open(ann_path), dtype=np.uint32)
masks = rgb_to_id(masks)
ids = np.array([ann["id"] for ann in ann_info["segments_info"]])
masks = masks == ids[:, None, None]
masks = np.asarray(masks, dtype=np.uint8)
labels = np.asarray([ann["category_id"] for ann in ann_info["segments_info"]], dtype=np.int64)
target = {}
target["image_id"] = np.asarray(
[ann_info["image_id"] if "image_id" in ann_info else ann_info["id"]], dtype=np.int64
)
if return_masks:
target["masks"] = masks
target["class_labels"] = labels
target["boxes"] = masks_to_boxes(masks)
target["size"] = np.asarray([int(h), int(w)], dtype=np.int64)
target["orig_size"] = np.asarray([int(h), int(w)], dtype=np.int64)
if "segments_info" in ann_info:
target["iscrowd"] = np.asarray([ann["iscrowd"] for ann in ann_info["segments_info"]], dtype=np.int64)
target["area"] = np.asarray([ann["area"] for ann in ann_info["segments_info"]], dtype=np.float32)
return image, target
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor._resize
def _resize(self, image, size, target=None, max_size=None):
"""
Resize the image to the given size. Size can be min_size (scalar) or (w, h) tuple. If size is an int, smaller
edge of the image will be matched to this number.
If given, also resize the target accordingly.
"""
if not isinstance(image, Image.Image):
image = self.to_pil_image(image)
def get_size_with_aspect_ratio(image_size, size, max_size=None):
w, h = image_size
if max_size is not None:
min_original_size = float(min((w, h)))
max_original_size = float(max((w, h)))
if max_original_size / min_original_size * size > max_size:
size = int(round(max_size * min_original_size / max_original_size))
if (w <= h and w == size) or (h <= w and h == size):
return (h, w)
if w < h:
ow = size
oh = int(size * h / w)
else:
oh = size
ow = int(size * w / h)
return (oh, ow)
def get_size(image_size, size, max_size=None):
if isinstance(size, (list, tuple)):
return size
else:
# size returned must be (w, h) since we use PIL to resize images
# so we revert the tuple
return get_size_with_aspect_ratio(image_size, size, max_size)[::-1]
size = get_size(image.size, size, max_size)
rescaled_image = self.resize(image, size=size)
if target is None:
return rescaled_image, None
ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(rescaled_image.size, image.size))
ratio_width, ratio_height = ratios
target = target.copy()
if "boxes" in target:
boxes = target["boxes"]
scaled_boxes = boxes * np.asarray([ratio_width, ratio_height, ratio_width, ratio_height], dtype=np.float32)
target["boxes"] = scaled_boxes
if "area" in target:
area = target["area"]
scaled_area = area * (ratio_width * ratio_height)
target["area"] = scaled_area
w, h = size
target["size"] = np.asarray([h, w], dtype=np.int64)
if "masks" in target:
# use PyTorch as current workaround
# TODO replace by self.resize
masks = torch.from_numpy(target["masks"][:, None]).float()
interpolated_masks = nn.functional.interpolate(masks, size=(h, w), mode="nearest")[:, 0] > 0.5
target["masks"] = interpolated_masks.numpy()
return rescaled_image, target
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor._normalize
def _normalize(self, image, mean, std, target=None):
"""
Normalize the image with a certain mean and std.
If given, also normalize the target bounding boxes based on the size of the image.
"""
image = self.normalize(image, mean=mean, std=std)
if target is None:
return image, None
target = target.copy()
h, w = image.shape[-2:]
if "boxes" in target:
boxes = target["boxes"]
boxes = corners_to_center_format(boxes)
boxes = boxes / np.asarray([w, h, w, h], dtype=np.float32)
target["boxes"] = boxes
return image, target
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.__call__ with Detr->ConditionalDetr,DETR->ConditionalDETR
def __call__(
self,
images: ImageInput,
annotations: Union[List[Dict], List[List[Dict]]] = None,
return_segmentation_masks: Optional[bool] = False,
masks_path: Optional[pathlib.Path] = None,
pad_and_return_pixel_mask: Optional[bool] = True,
return_tensors: Optional[Union[str, TensorType]] = None,
**kwargs,
) -> BatchFeature:
"""
Main method to prepare for the model one or several image(s) and optional annotations. Images are by default
padded up to the largest image in a batch, and a pixel mask is created that indicates which pixels are
real/which are padding.
<Tip warning={true}>
NumPy arrays and PyTorch tensors are converted to PIL images when resizing, so the most efficient is to pass
PIL images.
</Tip>
Args:
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
number of channels, H and W are image height and width.
annotations (`Dict`, `List[Dict]`, *optional*):
The corresponding annotations in COCO format.
In case [`ConditionalDetrFeatureExtractor`] was initialized with `format = "coco_detection"`, the
annotations for each image should have the following format: {'image_id': int, 'annotations':
[annotation]}, with the annotations being a list of COCO object annotations.
In case [`ConditionalDetrFeatureExtractor`] was initialized with `format = "coco_panoptic"`, the
annotations for each image should have the following format: {'image_id': int, 'file_name': str,
'segments_info': [segment_info]} with segments_info being a list of COCO panoptic annotations.
return_segmentation_masks (`Dict`, `List[Dict]`, *optional*, defaults to `False`):
Whether to also include instance segmentation masks as part of the labels in case `format =
"coco_detection"`.
masks_path (`pathlib.Path`, *optional*):
Path to the directory containing the PNG files that store the class-agnostic image segmentations. Only
relevant in case [`ConditionalDetrFeatureExtractor`] was initialized with `format = "coco_panoptic"`.
pad_and_return_pixel_mask (`bool`, *optional*, defaults to `True`):
Whether or not to pad images up to the largest image in a batch and create a pixel mask.
If left to the default, will return a pixel mask that is:
- 1 for pixels that are real (i.e. **not masked**),
- 0 for pixels that are padding (i.e. **masked**).
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors instead of NumPy arrays. If set to `'pt'`, return PyTorch `torch.Tensor`
objects.
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **pixel_values** -- Pixel values to be fed to a model.
- **pixel_mask** -- Pixel mask to be fed to a model (when `pad_and_return_pixel_mask=True` or if
*"pixel_mask"* is in `self.model_input_names`).
- **labels** -- Optional labels to be fed to a model (when `annotations` are provided)
"""
# Input type checking for clearer error
valid_images = False
valid_annotations = False
valid_masks_path = False
# Check that images has a valid type
if isinstance(images, (Image.Image, np.ndarray)) or is_torch_tensor(images):
valid_images = True
elif isinstance(images, (list, tuple)):
if len(images) == 0 or isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]):
valid_images = True
if not valid_images:
raise ValueError(
"Images must of type `PIL.Image.Image`, `np.ndarray` or `torch.Tensor` (single example), "
"`List[PIL.Image.Image]`, `List[np.ndarray]` or `List[torch.Tensor]` (batch of examples)."
)
is_batched = bool(
isinstance(images, (list, tuple))
and (isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]))
)
# Check that annotations has a valid type
if annotations is not None:
if not is_batched:
if self.format == "coco_detection":
if isinstance(annotations, dict) and "image_id" in annotations and "annotations" in annotations:
if isinstance(annotations["annotations"], (list, tuple)):
# an image can have no annotations
if len(annotations["annotations"]) == 0 or isinstance(annotations["annotations"][0], dict):
valid_annotations = True
elif self.format == "coco_panoptic":
if isinstance(annotations, dict) and "image_id" in annotations and "segments_info" in annotations:
if isinstance(annotations["segments_info"], (list, tuple)):
# an image can have no segments (?)
if len(annotations["segments_info"]) == 0 or isinstance(
annotations["segments_info"][0], dict
):
valid_annotations = True
else:
if isinstance(annotations, (list, tuple)):
if len(images) != len(annotations):
raise ValueError("There must be as many annotations as there are images")
if isinstance(annotations[0], Dict):
if self.format == "coco_detection":
if isinstance(annotations[0]["annotations"], (list, tuple)):
valid_annotations = True
elif self.format == "coco_panoptic":
if isinstance(annotations[0]["segments_info"], (list, tuple)):
valid_annotations = True
if not valid_annotations:
raise ValueError(
"""
Annotations must of type `Dict` (single image) or `List[Dict]` (batch of images). In case of object
detection, each dictionary should contain the keys 'image_id' and 'annotations', with the latter
being a list of annotations in COCO format. In case of panoptic segmentation, each dictionary
should contain the keys 'file_name', 'image_id' and 'segments_info', with the latter being a list
of annotations in COCO format.
"""
)
# Check that masks_path has a valid type
if masks_path is not None:
if self.format == "coco_panoptic":
if isinstance(masks_path, pathlib.Path):
valid_masks_path = True
if not valid_masks_path:
raise ValueError(
"The path to the directory containing the mask PNG files should be provided as a"
" `pathlib.Path` object."
)
if not is_batched:
images = [images]
if annotations is not None:
annotations = [annotations]
# prepare (COCO annotations as a list of Dict -> ConditionalDETR target as a single Dict per image)
if annotations is not None:
for idx, (image, target) in enumerate(zip(images, annotations)):
if not isinstance(image, Image.Image):
image = self.to_pil_image(image)
image, target = self.prepare(image, target, return_segmentation_masks, masks_path)
images[idx] = image
annotations[idx] = target
# transformations (resizing + normalization)
if self.do_resize and self.size is not None:
if annotations is not None:
for idx, (image, target) in enumerate(zip(images, annotations)):
image, target = self._resize(image=image, target=target, size=self.size, max_size=self.max_size)
images[idx] = image
annotations[idx] = target
else:
for idx, image in enumerate(images):
images[idx] = self._resize(image=image, target=None, size=self.size, max_size=self.max_size)[0]
if self.do_normalize:
if annotations is not None:
for idx, (image, target) in enumerate(zip(images, annotations)):
image, target = self._normalize(
image=image, mean=self.image_mean, std=self.image_std, target=target
)
images[idx] = image
annotations[idx] = target
else:
images = [
self._normalize(image=image, mean=self.image_mean, std=self.image_std)[0] for image in images
]
if pad_and_return_pixel_mask:
# pad images up to largest image in batch and create pixel_mask
max_size = self._max_by_axis([list(image.shape) for image in images])
c, h, w = max_size
padded_images = []
pixel_mask = []
for image in images:
# create padded image
padded_image = np.zeros((c, h, w), dtype=np.float32)
padded_image[: image.shape[0], : image.shape[1], : image.shape[2]] = np.copy(image)
padded_images.append(padded_image)
# create pixel mask
mask = np.zeros((h, w), dtype=np.int64)
mask[: image.shape[1], : image.shape[2]] = True
pixel_mask.append(mask)
images = padded_images
# return as BatchFeature
data = {}
data["pixel_values"] = images
if pad_and_return_pixel_mask:
data["pixel_mask"] = pixel_mask
encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
if annotations is not None:
# Convert to TensorType
tensor_type = return_tensors
if not isinstance(tensor_type, TensorType):
tensor_type = TensorType(tensor_type)
if not tensor_type == TensorType.PYTORCH:
raise ValueError("Only PyTorch is supported for the moment.")
else:
if not is_torch_available():
raise ImportError("Unable to convert output to PyTorch tensors format, PyTorch is not installed.")
encoded_inputs["labels"] = [
{k: torch.from_numpy(v) for k, v in target.items()} for target in annotations
]
return encoded_inputs
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor._max_by_axis
def _max_by_axis(self, the_list):
# type: (List[List[int]]) -> List[int]
maxes = the_list[0]
for sublist in the_list[1:]:
for index, item in enumerate(sublist):
maxes[index] = max(maxes[index], item)
return maxes
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.pad_and_create_pixel_mask
def pad_and_create_pixel_mask(
self, pixel_values_list: List["torch.Tensor"], return_tensors: Optional[Union[str, TensorType]] = None
):
"""
Pad images up to the largest image in a batch and create a corresponding `pixel_mask`.
Args:
pixel_values_list (`List[torch.Tensor]`):
List of images (pixel values) to be padded. Each image should be a tensor of shape (C, H, W).
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors instead of NumPy arrays. If set to `'pt'`, return PyTorch `torch.Tensor`
objects.
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **pixel_values** -- Pixel values to be fed to a model.
- **pixel_mask** -- Pixel mask to be fed to a model (when `pad_and_return_pixel_mask=True` or if
*"pixel_mask"* is in `self.model_input_names`).
"""
max_size = self._max_by_axis([list(image.shape) for image in pixel_values_list])
c, h, w = max_size
padded_images = []
pixel_mask = []
for image in pixel_values_list:
# create padded image
padded_image = np.zeros((c, h, w), dtype=np.float32)
padded_image[: image.shape[0], : image.shape[1], : image.shape[2]] = np.copy(image)
padded_images.append(padded_image)
# create pixel mask
mask = np.zeros((h, w), dtype=np.int64)
mask[: image.shape[1], : image.shape[2]] = True
pixel_mask.append(mask)
# return as BatchFeature
data = {"pixel_values": padded_images, "pixel_mask": pixel_mask}
encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)
return encoded_inputs
# POSTPROCESSING METHODS
# inspired by https://github.com/Atten4Vis/conditionalDETR/blob/master/models/conditional_detr.py#L258
def post_process(self, outputs, target_sizes):
"""
Converts the output of [`ConditionalDetrForObjectDetection`] into the format expected by the COCO api. Only
supports PyTorch.
Args:
outputs ([`ConditionalDetrObjectDetectionOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
image size (before any data augmentation). For visualization, this should be the image size after data
augment, but before padding.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
in the batch as predicted by the model.
"""
out_logits, out_bbox = outputs.logits, outputs.pred_boxes
if len(out_logits) != len(target_sizes):
raise ValueError("Make sure that you pass in as many target sizes as the batch dimension of the logits")
if target_sizes.shape[1] != 2:
raise ValueError("Each element of target_sizes must contain the size (h, w) of each image of the batch")
prob = out_logits.sigmoid()
topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 300, dim=1)
scores = topk_values
topk_boxes = topk_indexes // out_logits.shape[2]
labels = topk_indexes % out_logits.shape[2]
boxes = center_to_corners_format(out_bbox)
boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
# and from relative [0, 1] to absolute [0, height] coordinates
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
results = [{"scores": s, "labels": l, "boxes": b} for s, l, b in zip(scores, labels, boxes)]
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_segmentation with Detr->ConditionalDetr
def post_process_segmentation(self, outputs, target_sizes, threshold=0.9, mask_threshold=0.5):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into image segmentation predictions. Only supports
PyTorch.
Parameters:
outputs ([`ConditionalDetrSegmentationOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction.
threshold (`float`, *optional*, defaults to 0.9):
Threshold to use to filter out queries.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, and masks for an image
in the batch as predicted by the model.
"""
out_logits, raw_masks = outputs.logits, outputs.pred_masks
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, size in zip(out_logits, raw_masks, target_sizes):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_masks = (cur_masks.sigmoid() > mask_threshold) * 1
predictions = {"scores": cur_scores, "labels": cur_classes, "masks": cur_masks}
preds.append(predictions)
return preds
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_instance with Detr->ConditionalDetr
def post_process_instance(self, results, outputs, orig_target_sizes, max_target_sizes, threshold=0.5):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into actual instance segmentation predictions. Only
supports PyTorch.
Args:
results (`List[Dict]`):
Results list obtained by [`~ConditionalDetrFeatureExtractor.post_process`], to which "masks" results
will be added.
outputs ([`ConditionalDetrSegmentationOutput`]):
Raw outputs of the model.
orig_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
image size (before any data augmentation).
max_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the maximum size (h, w) of each image of the batch. For evaluation, this must be the
original image size (before any data augmentation).
threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, boxes and masks for an
image in the batch as predicted by the model.
"""
if len(orig_target_sizes) != len(max_target_sizes):
raise ValueError("Make sure to pass in as many orig_target_sizes as max_target_sizes")
max_h, max_w = max_target_sizes.max(0)[0].tolist()
outputs_masks = outputs.pred_masks.squeeze(2)
outputs_masks = nn.functional.interpolate(
outputs_masks, size=(max_h, max_w), mode="bilinear", align_corners=False
)
outputs_masks = (outputs_masks.sigmoid() > threshold).cpu()
for i, (cur_mask, t, tt) in enumerate(zip(outputs_masks, max_target_sizes, orig_target_sizes)):
img_h, img_w = t[0], t[1]
results[i]["masks"] = cur_mask[:, :img_h, :img_w].unsqueeze(1)
results[i]["masks"] = nn.functional.interpolate(
results[i]["masks"].float(), size=tuple(tt.tolist()), mode="nearest"
).byte()
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_panoptic with Detr->ConditionalDetr
def post_process_panoptic(self, outputs, processed_sizes, target_sizes=None, is_thing_map=None, threshold=0.85):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into actual panoptic predictions. Only supports
PyTorch.
Parameters:
outputs ([`ConditionalDetrSegmentationOutput`]):
Raw outputs of the model.
processed_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) containing the size (h, w) of each image of the batch, i.e. the size after data
augmentation but before batching.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`, *optional*):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction. If left to
None, it will default to the `processed_sizes`.
is_thing_map (`torch.Tensor` of shape `(batch_size, 2)`, *optional*):
Dictionary mapping class indices to either True or False, depending on whether or not they are a thing.
If not set, defaults to the `is_thing_map` of COCO panoptic.
threshold (`float`, *optional*, defaults to 0.85):
Threshold to use to filter out queries.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing a PNG string and segments_info values for
an image in the batch as predicted by the model.
"""
if target_sizes is None:
target_sizes = processed_sizes
if len(processed_sizes) != len(target_sizes):
raise ValueError("Make sure to pass in as many processed_sizes as target_sizes")
if is_thing_map is None:
# default to is_thing_map of COCO panoptic
is_thing_map = {i: i <= 90 for i in range(201)}
out_logits, raw_masks, raw_boxes = outputs.logits, outputs.pred_masks, outputs.pred_boxes
if not len(out_logits) == len(raw_masks) == len(target_sizes):
raise ValueError(
"Make sure that you pass in as many target sizes as the batch dimension of the logits and masks"
)
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, cur_boxes, size, target_size in zip(
out_logits, raw_masks, raw_boxes, processed_sizes, target_sizes
):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_boxes = center_to_corners_format(cur_boxes[keep])
h, w = cur_masks.shape[-2:]
if len(cur_boxes) != len(cur_classes):
raise ValueError("Not as many boxes as there are classes")
# It may be that we have several predicted masks for the same stuff class.
# In the following, we track the list of masks ids for each stuff class (they are merged later on)
cur_masks = cur_masks.flatten(1)
stuff_equiv_classes = defaultdict(lambda: [])
for k, label in enumerate(cur_classes):
if not is_thing_map[label.item()]:
stuff_equiv_classes[label.item()].append(k)
def get_ids_area(masks, scores, dedup=False):
# This helper function creates the final panoptic segmentation image
# It also returns the area of the masks that appears on the image
m_id = masks.transpose(0, 1).softmax(-1)
if m_id.shape[-1] == 0:
# We didn't detect any mask :(
m_id = torch.zeros((h, w), dtype=torch.long, device=m_id.device)
else:
m_id = m_id.argmax(-1).view(h, w)
if dedup:
# Merge the masks corresponding to the same stuff class
for equiv in stuff_equiv_classes.values():
if len(equiv) > 1:
for eq_id in equiv:
m_id.masked_fill_(m_id.eq(eq_id), equiv[0])
final_h, final_w = to_tuple(target_size)
seg_img = Image.fromarray(id_to_rgb(m_id.view(h, w).cpu().numpy()))
seg_img = seg_img.resize(size=(final_w, final_h), resample=Image.NEAREST)
np_seg_img = torch.ByteTensor(torch.ByteStorage.from_buffer(seg_img.tobytes()))
np_seg_img = np_seg_img.view(final_h, final_w, 3)
np_seg_img = np_seg_img.numpy()
m_id = torch.from_numpy(rgb_to_id(np_seg_img))
area = []
for i in range(len(scores)):
area.append(m_id.eq(i).sum().item())
return area, seg_img
area, seg_img = get_ids_area(cur_masks, cur_scores, dedup=True)
if cur_classes.numel() > 0:
# We know filter empty masks as long as we find some
while True:
filtered_small = torch.as_tensor(
[area[i] <= 4 for i, c in enumerate(cur_classes)], dtype=torch.bool, device=keep.device
)
if filtered_small.any().item():
cur_scores = cur_scores[~filtered_small]
cur_classes = cur_classes[~filtered_small]
cur_masks = cur_masks[~filtered_small]
area, seg_img = get_ids_area(cur_masks, cur_scores)
else:
break
else:
cur_classes = torch.ones(1, dtype=torch.long, device=cur_classes.device)
segments_info = []
for i, a in enumerate(area):
cat = cur_classes[i].item()
segments_info.append({"id": i, "isthing": is_thing_map[cat], "category_id": cat, "area": a})
del cur_classes
with io.BytesIO() as out:
seg_img.save(out, format="PNG")
predictions = {"png_string": out.getvalue(), "segments_info": segments_info}
preds.append(predictions)
return preds
src/transformers/models/conditional_detr/modeling_conditional_detr.py 0 → 100644
View file @ 126a7390
# coding=utf-8
# Copyright 2022 Microsoft Research Asia 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.
""" PyTorch Conditional DETR model."""
import math
import random
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import torch
from torch import Tensor, nn
from ...activations import ACT2FN
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithCrossAttentions, Seq2SeqModelOutput
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import torch_int_div
from ...utils import (
ModelOutput,
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_scipy_available,
is_timm_available,
is_vision_available,
logging,
replace_return_docstrings,
requires_backends,
)
from .configuration_conditional_detr import ConditionalDetrConfig
if is_scipy_available():
from scipy.optimize import linear_sum_assignment
if is_vision_available():
from .feature_extraction_conditional_detr import center_to_corners_format
if is_timm_available():
from timm import create_model
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "ConditionalDetrConfig"
_CHECKPOINT_FOR_DOC = "microsoft/conditional-detr-resnet-50"
CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST = [
"microsoft/conditional-detr-resnet-50",
# See all Conditional DETR models at https://huggingface.co/models?filter=conditional_detr
]
@dataclass
class ConditionalDetrDecoderOutput(BaseModelOutputWithCrossAttentions):
"""
Base class for outputs of the Conditional DETR decoder. This class adds one attribute to
BaseModelOutputWithCrossAttentions, namely an optional stack of intermediate decoder activations, i.e. the output
of each decoder layer, each of them gone through a layernorm. This is useful when training the model with auxiliary
decoding losses.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the model.
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer
plus the initial embedding outputs.
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
the self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax,
used to compute the weighted average in the cross-attention heads.
intermediate_hidden_states (`torch.FloatTensor` of shape `(config.decoder_layers, batch_size, num_queries, hidden_size)`, *optional*, returned when `config.auxiliary_loss=True`):
Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a
layernorm.
"""
intermediate_hidden_states: Optional[torch.FloatTensor] = None
reference_points: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
class ConditionalDetrModelOutput(Seq2SeqModelOutput):
"""
Base class for outputs of the Conditional DETR encoder-decoder model. This class adds one attribute to
Seq2SeqModelOutput, namely an optional stack of intermediate decoder activations, i.e. the output of each decoder
layer, each of them gone through a layernorm. This is useful when training the model with auxiliary decoding
losses.
Args:
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the decoder at the output of each
layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax,
used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each
layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the encoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
intermediate_hidden_states (`torch.FloatTensor` of shape `(config.decoder_layers, batch_size, sequence_length, hidden_size)`, *optional*, returned when `config.auxiliary_loss=True`):
Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a
layernorm.
"""
intermediate_hidden_states: Optional[torch.FloatTensor] = None
reference_points: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
# Copied from transformers.models.detr.modeling_detr.DetrObjectDetectionOutput with Detr->ConditionalDetr
class ConditionalDetrObjectDetectionOutput(ModelOutput):
"""
Output type of [`ConditionalDetrForObjectDetection`].
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` are provided)):
Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a
bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized
scale-invariant IoU loss.
loss_dict (`Dict`, *optional*):
A dictionary containing the individual losses. Useful for logging.
logits (`torch.FloatTensor` of shape `(batch_size, num_queries, num_classes + 1)`):
Classification logits (including no-object) for all queries.
pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
possible padding). You can use [`~ConditionalDetrFeatureExtractor.post_process`] to retrieve the
unnormalized bounding boxes.
auxiliary_outputs (`list[Dict]`, *optional*):
Optional, only returned when auxilary losses are activated (i.e. `config.auxiliary_loss` is set to `True`)
and labels are provided. It is a list of dictionaries containing the two above keys (`logits` and
`pred_boxes`) for each decoder layer.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the decoder at the output of each
layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax,
used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each
layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the encoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
loss_dict: Optional[Dict] = None
logits: torch.FloatTensor = None
pred_boxes: torch.FloatTensor = None
auxiliary_outputs: Optional[List[Dict]] = None
last_hidden_state: Optional[torch.FloatTensor] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
@dataclass
# Copied from transformers.models.detr.modeling_detr.DetrSegmentationOutput with Detr->ConditionalDetr
class ConditionalDetrSegmentationOutput(ModelOutput):
"""
Output type of [`ConditionalDetrForSegmentation`].
Args:
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` are provided)):
Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a
bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized
scale-invariant IoU loss.
loss_dict (`Dict`, *optional*):
A dictionary containing the individual losses. Useful for logging.
logits (`torch.FloatTensor` of shape `(batch_size, num_queries, num_classes + 1)`):
Classification logits (including no-object) for all queries.
pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
possible padding). You can use [`~ConditionalDetrFeatureExtractor.post_process`] to retrieve the
unnormalized bounding boxes.
pred_masks (`torch.FloatTensor` of shape `(batch_size, num_queries, height/4, width/4)`):
Segmentation masks logits for all queries. See also
[`~ConditionalDetrFeatureExtractor.post_process_segmentation`] or
[`~ConditionalDetrFeatureExtractor.post_process_panoptic`] to evaluate instance and panoptic segmentation
masks respectively.
auxiliary_outputs (`list[Dict]`, *optional*):
Optional, only returned when auxiliary losses are activated (i.e. `config.auxiliary_loss` is set to `True`)
and labels are provided. It is a list of dictionaries containing the two above keys (`logits` and
`pred_boxes`) for each decoder layer.
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the decoder of the model.
decoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the decoder at the output of each
layer plus the initial embedding outputs.
decoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
cross_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax,
used to compute the weighted average in the cross-attention heads.
encoder_last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder of the model.
encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each
layer plus the initial embedding outputs.
encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`. Attentions weights of the encoder, after the attention softmax, used to compute the
weighted average in the self-attention heads.
"""
loss: Optional[torch.FloatTensor] = None
loss_dict: Optional[Dict] = None
logits: torch.FloatTensor = None
pred_boxes: torch.FloatTensor = None
pred_masks: torch.FloatTensor = None
auxiliary_outputs: Optional[List[Dict]] = None
last_hidden_state: Optional[torch.FloatTensor] = None
decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
encoder_last_hidden_state: Optional[torch.FloatTensor] = None
encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
# Copied from transformers.models.detr.modeling_detr.DetrFrozenBatchNorm2d with Detr->ConditionalDetr
class ConditionalDetrFrozenBatchNorm2d(nn.Module):
"""
BatchNorm2d where the batch statistics and the affine parameters are fixed.
Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than
torchvision.models.resnet[18,34,50,101] produce nans.
"""
def __init__(self, n):
super().__init__()
self.register_buffer("weight", torch.ones(n))
self.register_buffer("bias", torch.zeros(n))
self.register_buffer("running_mean", torch.zeros(n))
self.register_buffer("running_var", torch.ones(n))
def _load_from_state_dict(
self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
):
num_batches_tracked_key = prefix + "num_batches_tracked"
if num_batches_tracked_key in state_dict:
del state_dict[num_batches_tracked_key]
super()._load_from_state_dict(
state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
)
def forward(self, x):
# move reshapes to the beginning
# to make it user-friendly
weight = self.weight.reshape(1, -1, 1, 1)
bias = self.bias.reshape(1, -1, 1, 1)
running_var = self.running_var.reshape(1, -1, 1, 1)
running_mean = self.running_mean.reshape(1, -1, 1, 1)
epsilon = 1e-5
scale = weight * (running_var + epsilon).rsqrt()
bias = bias - running_mean * scale
return x * scale + bias
# Copied from transformers.models.detr.modeling_detr.replace_batch_norm with Detr->ConditionalDetr
def replace_batch_norm(m, name=""):
for attr_str in dir(m):
target_attr = getattr(m, attr_str)
if isinstance(target_attr, nn.BatchNorm2d):
frozen = ConditionalDetrFrozenBatchNorm2d(target_attr.num_features)
bn = getattr(m, attr_str)
frozen.weight.data.copy_(bn.weight)
frozen.bias.data.copy_(bn.bias)
frozen.running_mean.data.copy_(bn.running_mean)
frozen.running_var.data.copy_(bn.running_var)
setattr(m, attr_str, frozen)
for n, ch in m.named_children():
replace_batch_norm(ch, n)
# Copied from transformers.models.detr.modeling_detr.DetrTimmConvEncoder
class ConditionalDetrTimmConvEncoder(nn.Module):
"""
Convolutional encoder (backbone) from the timm library.
nn.BatchNorm2d layers are replaced by DetrFrozenBatchNorm2d as defined above.
"""
def __init__(self, name: str, dilation: bool, use_pretrained_backbone: bool, num_channels: int = 3):
super().__init__()
kwargs = {}
if dilation:
kwargs["output_stride"] = 16
requires_backends(self, ["timm"])
backbone = create_model(
name,
pretrained=use_pretrained_backbone,
features_only=True,
out_indices=(1, 2, 3, 4),
in_chans=num_channels,
**kwargs,
)
# replace batch norm by frozen batch norm
with torch.no_grad():
replace_batch_norm(backbone)
self.model = backbone
self.intermediate_channel_sizes = self.model.feature_info.channels()
if "resnet" in name:
for name, parameter in self.model.named_parameters():
if "layer2" not in name and "layer3" not in name and "layer4" not in name:
parameter.requires_grad_(False)
def forward(self, pixel_values: torch.Tensor, pixel_mask: torch.Tensor):
# send pixel_values through the model to get list of feature maps
features = self.model(pixel_values)
out = []
for feature_map in features:
# downsample pixel_mask to match shape of corresponding feature_map
mask = nn.functional.interpolate(pixel_mask[None].float(), size=feature_map.shape[-2:]).to(torch.bool)[0]
out.append((feature_map, mask))
return out
# Copied from transformers.models.detr.modeling_detr.DetrConvModel with Detr->ConditionalDetr
class ConditionalDetrConvModel(nn.Module):
"""
This module adds 2D position embeddings to all intermediate feature maps of the convolutional encoder.
"""
def __init__(self, conv_encoder, position_embedding):
super().__init__()
self.conv_encoder = conv_encoder
self.position_embedding = position_embedding
def forward(self, pixel_values, pixel_mask):
# send pixel_values and pixel_mask through backbone to get list of (feature_map, pixel_mask) tuples
out = self.conv_encoder(pixel_values, pixel_mask)
pos = []
for feature_map, mask in out:
# position encoding
pos.append(self.position_embedding(feature_map, mask).to(feature_map.dtype))
return out, pos
# Copied from transformers.models.detr.modeling_detr._expand_mask with Detr->ConditionalDetr
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, target_len: Optional[int] = None):
"""
Expands attention_mask from `[batch_size, seq_len]` to `[batch_size, 1, target_seq_len, source_seq_len]`.
"""
batch_size, source_len = mask.size()
target_len = target_len if target_len is not None else source_len
expanded_mask = mask[:, None, None, :].expand(batch_size, 1, target_len, source_len).to(dtype)
inverted_mask = 1.0 - expanded_mask
return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
# Copied from transformers.models.detr.modeling_detr.DetrSinePositionEmbedding with Detr->ConditionalDetr
class ConditionalDetrSinePositionEmbedding(nn.Module):
"""
This is a more standard version of the position embedding, very similar to the one used by the Attention is all you
need paper, generalized to work on images.
"""
def __init__(self, embedding_dim=64, temperature=10000, normalize=False, scale=None):
super().__init__()
self.embedding_dim = embedding_dim
self.temperature = temperature
self.normalize = normalize
if scale is not None and normalize is False:
raise ValueError("normalize should be True if scale is passed")
if scale is None:
scale = 2 * math.pi
self.scale = scale
def forward(self, pixel_values, pixel_mask):
if pixel_mask is None:
raise ValueError("No pixel mask provided")
y_embed = pixel_mask.cumsum(1, dtype=torch.float32)
x_embed = pixel_mask.cumsum(2, dtype=torch.float32)
if self.normalize:
y_embed = y_embed / (y_embed[:, -1:, :] + 1e-6) * self.scale
x_embed = x_embed / (x_embed[:, :, -1:] + 1e-6) * self.scale
dim_t = torch.arange(self.embedding_dim, dtype=torch.float32, device=pixel_values.device)
dim_t = self.temperature ** (2 * torch_int_div(dim_t, 2) / self.embedding_dim)
pos_x = x_embed[:, :, :, None] / dim_t
pos_y = y_embed[:, :, :, None] / dim_t
pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
return pos
# Copied from transformers.models.detr.modeling_detr.DetrLearnedPositionEmbedding with Detr->ConditionalDetr
class ConditionalDetrLearnedPositionEmbedding(nn.Module):
"""
This module learns positional embeddings up to a fixed maximum size.
"""
def __init__(self, embedding_dim=256):
super().__init__()
self.row_embeddings = nn.Embedding(50, embedding_dim)
self.column_embeddings = nn.Embedding(50, embedding_dim)
def forward(self, pixel_values, pixel_mask=None):
height, width = pixel_values.shape[-2:]
width_values = torch.arange(width, device=pixel_values.device)
height_values = torch.arange(height, device=pixel_values.device)
x_emb = self.column_embeddings(width_values)
y_emb = self.row_embeddings(height_values)
pos = torch.cat([x_emb.unsqueeze(0).repeat(height, 1, 1), y_emb.unsqueeze(1).repeat(1, width, 1)], dim=-1)
pos = pos.permute(2, 0, 1)
pos = pos.unsqueeze(0)
pos = pos.repeat(pixel_values.shape[0], 1, 1, 1)
return pos
# Copied from transformers.models.detr.modeling_detr.build_position_encoding with Detr->ConditionalDetr
def build_position_encoding(config):
n_steps = config.d_model // 2
if config.position_embedding_type == "sine":
# TODO find a better way of exposing other arguments
position_embedding = ConditionalDetrSinePositionEmbedding(n_steps, normalize=True)
elif config.position_embedding_type == "learned":
position_embedding = ConditionalDetrLearnedPositionEmbedding(n_steps)
else:
raise ValueError(f"Not supported {config.position_embedding_type}")
return position_embedding
# function to generate sine positional embedding for 2d coordinates
def gen_sine_position_embeddings(pos_tensor):
scale = 2 * math.pi
dim_t = torch.arange(128, dtype=torch.float32, device=pos_tensor.device)
dim_t = 10000 ** (2 * (dim_t // 2) / 128)
x_embed = pos_tensor[:, :, 0] * scale
y_embed = pos_tensor[:, :, 1] * scale
pos_x = x_embed[:, :, None] / dim_t
pos_y = y_embed[:, :, None] / dim_t
pos_x = torch.stack((pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3).flatten(2)
pos_y = torch.stack((pos_y[:, :, 0::2].sin(), pos_y[:, :, 1::2].cos()), dim=3).flatten(2)
pos = torch.cat((pos_y, pos_x), dim=2)
return pos
def inverse_sigmoid(x, eps=1e-5):
x = x.clamp(min=0, max=1)
x1 = x.clamp(min=eps)
x2 = (1 - x).clamp(min=eps)
return torch.log(x1 / x2)
# Copied from transformers.models.detr.modeling_detr.DetrAttention
class DetrAttention(nn.Module):
"""
Multi-headed attention from 'Attention Is All You Need' paper.
Here, we add position embeddings to the queries and keys (as explained in the DETR paper).
"""
def __init__(
self,
embed_dim: int,
num_heads: int,
dropout: float = 0.0,
is_decoder: bool = False,
bias: bool = True,
):
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
if self.head_dim * num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
def _shape(self, tensor: torch.Tensor, seq_len: int, batch_size: int):
return tensor.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def with_pos_embed(self, tensor: torch.Tensor, position_embeddings: Optional[Tensor]):
return tensor if position_embeddings is None else tensor + position_embeddings
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
key_value_states: Optional[torch.Tensor] = None,
key_value_position_embeddings: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
# if key_value_states are provided this layer is used as a cross-attention layer
# for the decoder
is_cross_attention = key_value_states is not None
batch_size, target_len, embed_dim = hidden_states.size()
# add position embeddings to the hidden states before projecting to queries and keys
if position_embeddings is not None:
hidden_states_original = hidden_states
hidden_states = self.with_pos_embed(hidden_states, position_embeddings)
# add key-value position embeddings to the key value states
if key_value_position_embeddings is not None:
key_value_states_original = key_value_states
key_value_states = self.with_pos_embed(key_value_states, key_value_position_embeddings)
# get query proj
query_states = self.q_proj(hidden_states) * self.scaling
# get key, value proj
if is_cross_attention:
# cross_attentions
key_states = self._shape(self.k_proj(key_value_states), -1, batch_size)
value_states = self._shape(self.v_proj(key_value_states_original), -1, batch_size)
else:
# self_attention
key_states = self._shape(self.k_proj(hidden_states), -1, batch_size)
value_states = self._shape(self.v_proj(hidden_states_original), -1, batch_size)
proj_shape = (batch_size * self.num_heads, -1, self.head_dim)
query_states = self._shape(query_states, target_len, batch_size).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*proj_shape)
source_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (batch_size * self.num_heads, target_len, source_len):
raise ValueError(
f"Attention weights should be of size {(batch_size * self.num_heads, target_len, source_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (batch_size, 1, target_len, source_len):
raise ValueError(
f"Attention mask should be of size {(batch_size, 1, target_len, source_len)}, but is"
f" {attention_mask.size()}"
)
attn_weights = attn_weights.view(batch_size, self.num_heads, target_len, source_len) + attention_mask
attn_weights = attn_weights.view(batch_size * self.num_heads, target_len, source_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(batch_size, self.num_heads, target_len, source_len)
attn_weights = attn_weights_reshaped.view(batch_size * self.num_heads, target_len, source_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (batch_size * self.num_heads, target_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(batch_size, self.num_heads, target_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(batch_size, self.num_heads, target_len, self.head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(batch_size, target_len, embed_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped
class ConditionalDetrAttention(nn.Module):
"""
Cross-Attention used in Conditional DETR 'Conditional DETR for Fast Training Convergence' paper.
The key q_proj, k_proj, v_proj are defined outside the attention. This attention allows the dim of q, k to be
different to v.
"""
def __init__(
self,
embed_dim: int,
out_dim: int,
num_heads: int,
dropout: float = 0.0,
bias: bool = True,
):
super().__init__()
self.embed_dim = embed_dim
self.out_dim = out_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
if self.head_dim * num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {num_heads})."
)
# head dimension of values
self.v_head_dim = out_dim // num_heads
if self.v_head_dim * num_heads != self.out_dim:
raise ValueError(
f"out_dim must be divisible by num_heads (got `out_dim`: {self.out_dim} and `num_heads`: {num_heads})."
)
self.scaling = self.head_dim**-0.5
self.out_proj = nn.Linear(out_dim, out_dim, bias=bias)
def _qk_shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def _v_shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.v_head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
key_states: Optional[torch.Tensor] = None,
value_states: Optional[torch.Tensor] = None,
output_attentions: bool = False,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, _ = hidden_states.size()
# get query proj
query_states = hidden_states * self.scaling
# get key, value proj
key_states = self._qk_shape(key_states, -1, bsz)
value_states = self._v_shape(value_states, -1, bsz)
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
v_proj_shape = (bsz * self.num_heads, -1, self.v_head_dim)
query_states = self._qk_shape(query_states, tgt_len, bsz).view(*proj_shape)
key_states = key_states.view(*proj_shape)
value_states = value_states.view(*v_proj_shape)
src_len = key_states.size(1)
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
raise ValueError(
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
if output_attentions:
# this operation is a bit awkward, but it's required to
# make sure that attn_weights keeps its gradient.
# In order to do so, attn_weights have to reshaped
# twice and have to be reused in the following
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
else:
attn_weights_reshaped = None
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
attn_output = torch.bmm(attn_probs, value_states)
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.v_head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.v_head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.v_head_dim)
attn_output = attn_output.transpose(1, 2)
attn_output = attn_output.reshape(bsz, tgt_len, self.out_dim)
attn_output = self.out_proj(attn_output)
return attn_output, attn_weights_reshaped
# Copied from transformers.models.detr.modeling_detr.DetrEncoderLayer with DetrEncoderLayer->ConditionalDetrEncoderLayer,DetrConfig->ConditionalDetrConfig
class ConditionalDetrEncoderLayer(nn.Module):
def __init__(self, config: ConditionalDetrConfig):
super().__init__()
self.embed_dim = config.d_model
self.self_attn = DetrAttention(
embed_dim=self.embed_dim,
num_heads=config.encoder_attention_heads,
dropout=config.attention_dropout,
)
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor,
position_embeddings: torch.Tensor = None,
output_attentions: bool = False,
):
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, target_len, source_len)` where padding elements are indicated by very large negative
values.
position_embeddings (`torch.FloatTensor`, *optional*): position embeddings, to be added to hidden_states.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
hidden_states, attn_weights = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_embeddings=position_embeddings,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
if self.training:
if torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any():
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
outputs = (hidden_states,)
if output_attentions:
outputs += (attn_weights,)
return outputs
class ConditionalDetrDecoderLayer(nn.Module):
def __init__(self, config: ConditionalDetrConfig):
super().__init__()
self.embed_dim = config.d_model
d_model = config.d_model
# Decoder Self-Attention projections
self.sa_qcontent_proj = nn.Linear(d_model, d_model)
self.sa_qpos_proj = nn.Linear(d_model, d_model)
self.sa_kcontent_proj = nn.Linear(d_model, d_model)
self.sa_kpos_proj = nn.Linear(d_model, d_model)
self.sa_v_proj = nn.Linear(d_model, d_model)
self.self_attn = ConditionalDetrAttention(
embed_dim=self.embed_dim,
out_dim=self.embed_dim,
num_heads=config.decoder_attention_heads,
dropout=config.attention_dropout,
)
self.dropout = config.dropout
self.activation_fn = ACT2FN[config.activation_function]
self.activation_dropout = config.activation_dropout
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
# Decoder Cross-Attention projections
self.ca_qcontent_proj = nn.Linear(d_model, d_model)
self.ca_qpos_proj = nn.Linear(d_model, d_model)
self.ca_kcontent_proj = nn.Linear(d_model, d_model)
self.ca_kpos_proj = nn.Linear(d_model, d_model)
self.ca_v_proj = nn.Linear(d_model, d_model)
self.ca_qpos_sine_proj = nn.Linear(d_model, d_model)
self.encoder_attn = ConditionalDetrAttention(
self.embed_dim * 2, self.embed_dim, config.decoder_attention_heads, dropout=config.attention_dropout
)
self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
self.final_layer_norm = nn.LayerNorm(self.embed_dim)
self.nhead = config.decoder_attention_heads
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[torch.Tensor] = None,
query_position_embeddings: Optional[torch.Tensor] = None,
query_sine_embed: Optional[torch.Tensor] = None,
encoder_hidden_states: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = False,
is_first: Optional[bool] = False,
):
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
attention_mask (`torch.FloatTensor`): attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
position_embeddings (`torch.FloatTensor`, *optional*):
position embeddings that are added to the queries and keys
in the cross-attention layer.
query_position_embeddings (`torch.FloatTensor`, *optional*):
position embeddings that are added to the queries and keys
in the self-attention layer.
encoder_hidden_states (`torch.FloatTensor`):
cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
"""
residual = hidden_states
# ========== Begin of Self-Attention =============
# Apply projections here
# shape: num_queries x batch_size x 256
q_content = self.sa_qcontent_proj(
hidden_states
) # target is the input of the first decoder layer. zero by default.
q_pos = self.sa_qpos_proj(query_position_embeddings)
k_content = self.sa_kcontent_proj(hidden_states)
k_pos = self.sa_kpos_proj(query_position_embeddings)
v = self.sa_v_proj(hidden_states)
_, num_queries, n_model = q_content.shape
q = q_content + q_pos
k = k_content + k_pos
hidden_states, self_attn_weights = self.self_attn(
hidden_states=q,
attention_mask=attention_mask,
key_states=k,
value_states=v,
output_attentions=output_attentions,
)
# ============ End of Self-Attention =============
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.self_attn_layer_norm(hidden_states)
# ========== Begin of Cross-Attention =============
# Apply projections here
# shape: num_queries x batch_size x 256
q_content = self.ca_qcontent_proj(hidden_states)
k_content = self.ca_kcontent_proj(encoder_hidden_states)
v = self.ca_v_proj(encoder_hidden_states)
batch_size, num_queries, n_model = q_content.shape
_, src_len, _ = k_content.shape
k_pos = self.ca_kpos_proj(position_embeddings)
# For the first decoder layer, we concatenate the positional embedding predicted from
# the object query (the positional embedding) into the original query (key) in DETR.
if is_first:
q_pos = self.ca_qpos_proj(query_position_embeddings)
q = q_content + q_pos
k = k_content + k_pos
else:
q = q_content
k = k_content
q = q.view(batch_size, num_queries, self.nhead, n_model // self.nhead)
query_sine_embed = self.ca_qpos_sine_proj(query_sine_embed)
query_sine_embed = query_sine_embed.view(batch_size, num_queries, self.nhead, n_model // self.nhead)
q = torch.cat([q, query_sine_embed], dim=3).view(batch_size, num_queries, n_model * 2)
k = k.view(batch_size, src_len, self.nhead, n_model // self.nhead)
k_pos = k_pos.view(batch_size, src_len, self.nhead, n_model // self.nhead)
k = torch.cat([k, k_pos], dim=3).view(batch_size, src_len, n_model * 2)
# Cross-Attention Block
cross_attn_weights = None
if encoder_hidden_states is not None:
residual = hidden_states
hidden_states, cross_attn_weights = self.encoder_attn(
hidden_states=q,
attention_mask=encoder_attention_mask,
key_states=k,
value_states=v,
output_attentions=output_attentions,
)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.encoder_attn_layer_norm(hidden_states)
# ============ End of Cross-Attention =============
# Fully Connected
residual = hidden_states
hidden_states = self.activation_fn(self.fc1(hidden_states))
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
hidden_states = self.fc2(hidden_states)
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
hidden_states = residual + hidden_states
hidden_states = self.final_layer_norm(hidden_states)
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights, cross_attn_weights)
return outputs
# Copied from transformers.models.detr.modeling_detr.DetrClassificationHead with Detr->ConditionalDetr
class ConditionalDetrClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, input_dim: int, inner_dim: int, num_classes: int, pooler_dropout: float):
super().__init__()
self.dense = nn.Linear(input_dim, inner_dim)
self.dropout = nn.Dropout(p=pooler_dropout)
self.out_proj = nn.Linear(inner_dim, num_classes)
def forward(self, hidden_states: torch.Tensor):
hidden_states = self.dropout(hidden_states)
hidden_states = self.dense(hidden_states)
hidden_states = torch.tanh(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.out_proj(hidden_states)
return hidden_states
# Copied from transformers.models.detr.modeling_detr.DetrMLPPredictionHead with DetrMLPPredictionHead->MLP
class MLP(nn.Module):
"""
Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates,
height and width of a bounding box w.r.t. an image.
Copied from https://github.com/facebookresearch/detr/blob/master/models/detr.py
"""
def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
super().__init__()
self.num_layers = num_layers
h = [hidden_dim] * (num_layers - 1)
self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
def forward(self, x):
for i, layer in enumerate(self.layers):
x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
return x
# Copied from transformers.models.detr.modeling_detr.DetrPreTrainedModel with Detr->ConditionalDetr
class ConditionalDetrPreTrainedModel(PreTrainedModel):
config_class = ConditionalDetrConfig
base_model_prefix = "model"
main_input_name = "pixel_values"
def _init_weights(self, module):
std = self.config.init_std
xavier_std = self.config.init_xavier_std
if isinstance(module, ConditionalDetrMHAttentionMap):
nn.init.zeros_(module.k_linear.bias)
nn.init.zeros_(module.q_linear.bias)
nn.init.xavier_uniform_(module.k_linear.weight, gain=xavier_std)
nn.init.xavier_uniform_(module.q_linear.weight, gain=xavier_std)
elif isinstance(module, ConditionalDetrLearnedPositionEmbedding):
nn.init.uniform_(module.row_embeddings.weight)
nn.init.uniform_(module.column_embeddings.weight)
if isinstance(module, (nn.Linear, nn.Conv2d, nn.BatchNorm2d)):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, ConditionalDetrDecoder):
module.gradient_checkpointing = value
CONDITIONAL_DETR_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`ConditionalDetrConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
CONDITIONAL_DETR_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it.
Pixel values can be obtained using [`ConditionalDetrFeatureExtractor`]. See
[`ConditionalDetrFeatureExtractor.__call__`] for details.
pixel_mask (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*):
Mask to avoid performing attention on padding pixel values. Mask values selected in `[0, 1]`:
- 1 for pixels that are real (i.e. **not masked**),
- 0 for pixels that are padding (i.e. **masked**).
[What are attention masks?](../glossary#attention-mask)
decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, num_queries)`, *optional*):
Not used by default. Can be used to mask object queries.
encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
`last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you
can choose to directly pass a flattened representation of an image.
decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`, *optional*):
Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an
embedded representation.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
# Copied from transformers.models.detr.modeling_detr.DetrEncoder with Detr->ConditionalDetr,DETR->ConditionalDETR
class ConditionalDetrEncoder(ConditionalDetrPreTrainedModel):
"""
Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
[`ConditionalDetrEncoderLayer`].
The encoder updates the flattened feature map through multiple self-attention layers.
Small tweak for ConditionalDETR:
- position_embeddings are added to the forward pass.
Args:
config: ConditionalDetrConfig
"""
def __init__(self, config: ConditionalDetrConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.encoder_layerdrop
self.layers = nn.ModuleList([ConditionalDetrEncoderLayer(config) for _ in range(config.encoder_layers)])
# in the original ConditionalDETR, no layernorm is used at the end of the encoder, as "normalize_before" is set to False by default
# Initialize weights and apply final processing
self.post_init()
def forward(
self,
inputs_embeds=None,
attention_mask=None,
position_embeddings=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Flattened feature map (output of the backbone + projection layer) that is passed to the encoder.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding pixel features. Mask values selected in `[0, 1]`:
- 1 for pixel features that are real (i.e. **not masked**),
- 0 for pixel features that are padding (i.e. **masked**).
[What are attention masks?](../glossary#attention-mask)
position_embeddings (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
Position embeddings that are added to the queries and keys in each self-attention layer.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
hidden_states = inputs_embeds
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
# expand attention_mask
if attention_mask is not None:
# [batch_size, seq_len] -> [batch_size, 1, target_seq_len, source_seq_len]
attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
for i, encoder_layer in enumerate(self.layers):
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
dropout_probability = random.uniform(0, 1)
if self.training and (dropout_probability < self.layerdrop): # skip the layer
layer_outputs = (None, None)
else:
# we add position_embeddings as extra input to the encoder_layer
layer_outputs = encoder_layer(
hidden_states,
attention_mask,
position_embeddings=position_embeddings,
output_attentions=output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if output_hidden_states:
encoder_states = encoder_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
)
class ConditionalDetrDecoder(ConditionalDetrPreTrainedModel):
"""
Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`ConditionalDetrDecoderLayer`].
The decoder updates the query embeddings through multiple self-attention and cross-attention layers.
Some small tweaks for Conditional DETR:
- position_embeddings and query_position_embeddings are added to the forward pass.
- if self.config.auxiliary_loss is set to True, also returns a stack of activations from all decoding layers.
Args:
config: ConditionalDetrConfig
"""
def __init__(self, config: ConditionalDetrConfig):
super().__init__(config)
self.dropout = config.dropout
self.layerdrop = config.decoder_layerdrop
self.layers = nn.ModuleList([ConditionalDetrDecoderLayer(config) for _ in range(config.decoder_layers)])
# in Conditional DETR, the decoder uses layernorm after the last decoder layer output
self.layernorm = nn.LayerNorm(config.d_model)
d_model = config.d_model
self.gradient_checkpointing = False
# query_scale is the FFN applied on f to generate transformation T
self.query_scale = MLP(d_model, d_model, d_model, 2)
self.ref_point_head = MLP(d_model, d_model, 2, 2)
for layer_id in range(config.decoder_layers - 1):
self.layers[layer_id + 1].ca_qpos_proj = None
# Initialize weights and apply final processing
self.post_init()
def forward(
self,
inputs_embeds=None,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
position_embeddings=None,
query_position_embeddings=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Args:
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
The query embeddings that are passed into the decoder.
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on certain queries. Mask values selected in `[0, 1]`:
- 1 for queries that are **not masked**,
- 0 for queries that are **masked**.
[What are attention masks?](../glossary#attention-mask)
encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
of the decoder.
encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
Mask to avoid performing cross-attention on padding pixel_values of the encoder. Mask values selected
in `[0, 1]`:
- 1 for pixels that are real (i.e. **not masked**),
- 0 for pixels that are padding (i.e. **masked**).
position_embeddings (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Position embeddings that are added to the queries and keys in each cross-attention layer.
query_position_embeddings (`torch.FloatTensor` of shape `(batch_size, num_queries, hidden_size)`):
, *optional*): Position embeddings that are added to the queries and keys in each self-attention layer.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
for more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if inputs_embeds is not None:
hidden_states = inputs_embeds
input_shape = inputs_embeds.size()[:-1]
combined_attention_mask = None
if attention_mask is not None and combined_attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
combined_attention_mask = combined_attention_mask + _expand_mask(
attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
)
# expand encoder attention mask
if encoder_hidden_states is not None and encoder_attention_mask is not None:
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
encoder_attention_mask = _expand_mask(
encoder_attention_mask, inputs_embeds.dtype, target_len=input_shape[-1]
)
# optional intermediate hidden states
intermediate = () if self.config.auxiliary_loss else None
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
reference_points_before_sigmoid = self.ref_point_head(
query_position_embeddings
) # [num_queries, batch_size, 2]
reference_points = reference_points_before_sigmoid.sigmoid().transpose(0, 1)
obj_center = reference_points[..., :2].transpose(0, 1)
# get sine embedding for the query vector
query_sine_embed_before_transformation = gen_sine_position_embeddings(obj_center)
for idx, decoder_layer in enumerate(self.layers):
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
if output_hidden_states:
all_hidden_states += (hidden_states,)
dropout_probability = random.uniform(0, 1)
if self.training and (dropout_probability < self.layerdrop):
continue
if idx == 0:
pos_transformation = 1
else:
pos_transformation = self.query_scale(hidden_states)
# apply transformation
query_sine_embed = query_sine_embed_before_transformation * pos_transformation
if self.gradient_checkpointing and self.training:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(decoder_layer),
hidden_states,
combined_attention_mask,
position_embeddings,
query_position_embeddings,
query_sine_embed,
encoder_hidden_states,
encoder_attention_mask,
None,
None,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=combined_attention_mask,
position_embeddings=position_embeddings,
query_position_embeddings=query_position_embeddings,
query_sine_embed=query_sine_embed,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
is_first=(idx == 0),
)
hidden_states = layer_outputs[0]
if self.config.auxiliary_loss:
hidden_states = self.layernorm(hidden_states)
intermediate += (hidden_states,)
if output_attentions:
all_self_attns += (layer_outputs[1],)
if encoder_hidden_states is not None:
all_cross_attentions += (layer_outputs[2],)
# finally, apply layernorm
hidden_states = self.layernorm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
# stack intermediate decoder activations
if self.config.auxiliary_loss:
intermediate = torch.stack(intermediate)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
all_hidden_states,
all_self_attns,
all_cross_attentions,
intermediate,
reference_points,
]
if v is not None
)
return ConditionalDetrDecoderOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=all_self_attns,
cross_attentions=all_cross_attentions,
intermediate_hidden_states=intermediate,
reference_points=reference_points,
)
@add_start_docstrings(
"""
The bare Conditional DETR Model (consisting of a backbone and encoder-decoder Transformer) outputting raw
hidden-states without any specific head on top.
""",
CONDITIONAL_DETR_START_DOCSTRING,
)
class ConditionalDetrModel(ConditionalDetrPreTrainedModel):
def __init__(self, config: ConditionalDetrConfig):
super().__init__(config)
# Create backbone + positional encoding
backbone = ConditionalDetrTimmConvEncoder(
config.backbone, config.dilation, config.use_pretrained_backbone, config.num_channels
)
position_embeddings = build_position_encoding(config)
self.backbone = ConditionalDetrConvModel(backbone, position_embeddings)
# Create projection layer
self.input_projection = nn.Conv2d(backbone.intermediate_channel_sizes[-1], config.d_model, kernel_size=1)
self.query_position_embeddings = nn.Embedding(config.num_queries, config.d_model)
self.encoder = ConditionalDetrEncoder(config)
self.decoder = ConditionalDetrDecoder(config)
# Initialize weights and apply final processing
self.post_init()
def get_encoder(self):
return self.encoder
def get_decoder(self):
return self.decoder
def freeze_backbone(self):
for name, param in self.backbone.conv_encoder.model.named_parameters():
param.requires_grad_(False)
def unfreeze_backbone(self):
for name, param in self.backbone.conv_encoder.model.named_parameters():
param.requires_grad_(True)
@add_start_docstrings_to_model_forward(CONDITIONAL_DETR_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=ConditionalDetrModelOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values,
pixel_mask=None,
decoder_attention_mask=None,
encoder_outputs=None,
inputs_embeds=None,
decoder_inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
Returns:
Examples:
```python
>>> from transformers import ConditionalDetrFeatureExtractor, ConditionalDetrModel
>>> from PIL import Image
>>> import requests
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> feature_extractor = ConditionalDetrFeatureExtractor.from_pretrained("microsoft/conditional-detr-resnet-50")
>>> model = ConditionalDetrModel.from_pretrained("microsoft/conditional-detr-resnet-50")
>>> # prepare image for the model
>>> inputs = feature_extractor(images=image, return_tensors="pt")
>>> # forward pass
>>> outputs = model(**inputs)
>>> # the last hidden states are the final query embeddings of the Transformer decoder
>>> # these are of shape (batch_size, num_queries, hidden_size)
>>> last_hidden_states = outputs.last_hidden_state
>>> list(last_hidden_states.shape)
[1, 300, 256]
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
batch_size, num_channels, height, width = pixel_values.shape
device = pixel_values.device
if pixel_mask is None:
pixel_mask = torch.ones(((batch_size, height, width)), device=device)
# First, sent pixel_values + pixel_mask through Backbone to obtain the features
# pixel_values should be of shape (batch_size, num_channels, height, width)
# pixel_mask should be of shape (batch_size, height, width)
features, position_embeddings_list = self.backbone(pixel_values, pixel_mask)
# get final feature map and downsampled mask
feature_map, mask = features[-1]
if mask is None:
raise ValueError("Backbone does not return downsampled pixel mask")
# Second, apply 1x1 convolution to reduce the channel dimension to d_model (256 by default)
projected_feature_map = self.input_projection(feature_map)
# Third, flatten the feature map + position embeddings of shape NxCxHxW to NxCxHW, and permute it to NxHWxC
# In other words, turn their shape into (batch_size, sequence_length, hidden_size)
flattened_features = projected_feature_map.flatten(2).permute(0, 2, 1)
position_embeddings = position_embeddings_list[-1].flatten(2).permute(0, 2, 1)
flattened_mask = mask.flatten(1)
# Fourth, sent flattened_features + flattened_mask + position embeddings through encoder
# flattened_features is a Tensor of shape (batch_size, heigth*width, hidden_size)
# flattened_mask is a Tensor of shape (batch_size, heigth*width)
if encoder_outputs is None:
encoder_outputs = self.encoder(
inputs_embeds=flattened_features,
attention_mask=flattened_mask,
position_embeddings=position_embeddings,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
# If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
encoder_outputs = BaseModelOutput(
last_hidden_state=encoder_outputs[0],
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
)
# Fifth, sent query embeddings + position embeddings through the decoder (which is conditioned on the encoder output)
query_position_embeddings = self.query_position_embeddings.weight.unsqueeze(0).repeat(batch_size, 1, 1)
queries = torch.zeros_like(query_position_embeddings)
# decoder outputs consists of (dec_features, dec_hidden, dec_attn)
decoder_outputs = self.decoder(
inputs_embeds=queries,
attention_mask=None,
position_embeddings=position_embeddings,
query_position_embeddings=query_position_embeddings,
encoder_hidden_states=encoder_outputs[0],
encoder_attention_mask=flattened_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
if not return_dict:
return decoder_outputs + encoder_outputs
return ConditionalDetrModelOutput(
last_hidden_state=decoder_outputs.last_hidden_state,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
intermediate_hidden_states=decoder_outputs.intermediate_hidden_states,
reference_points=decoder_outputs.reference_points,
)
@add_start_docstrings(
"""
CONDITIONAL_DETR Model (consisting of a backbone and encoder-decoder Transformer) with object detection heads on
top, for tasks such as COCO detection.
""",
CONDITIONAL_DETR_START_DOCSTRING,
)
class ConditionalDetrForObjectDetection(ConditionalDetrPreTrainedModel):
def __init__(self, config: ConditionalDetrConfig):
super().__init__(config)
# CONDITIONAL DETR encoder-decoder model
self.model = ConditionalDetrModel(config)
# Object detection heads
self.class_labels_classifier = nn.Linear(
config.d_model, config.num_labels
) # We add one for the "no object" class
self.bbox_predictor = ConditionalDetrMLPPredictionHead(
input_dim=config.d_model, hidden_dim=config.d_model, output_dim=4, num_layers=3
)
# Initialize weights and apply final processing
self.post_init()
# taken from https://github.com/Atten4Vis/conditionalDETR/blob/master/models/conditional_detr.py
@torch.jit.unused
def _set_aux_loss(self, outputs_class, outputs_coord):
# this is a workaround to make torchscript happy, as torchscript
# doesn't support dictionary with non-homogeneous values, such
# as a dict having both a Tensor and a list.
return [{"logits": a, "pred_boxes": b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1])]
@add_start_docstrings_to_model_forward(CONDITIONAL_DETR_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=ConditionalDetrObjectDetectionOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values,
pixel_mask=None,
decoder_attention_mask=None,
encoder_outputs=None,
inputs_embeds=None,
decoder_inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (`List[Dict]` of len `(batch_size,)`, *optional*):
Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the
following 2 keys: 'class_labels' and 'boxes' (the class labels and bounding boxes of an image in the batch
respectively). The class labels themselves should be a `torch.LongTensor` of len `(number of bounding boxes
in the image,)` and the boxes a `torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)`.
Returns:
Examples:
```python
>>> from transformers import ConditionalDetrFeatureExtractor, ConditionalDetrForObjectDetection
>>> from PIL import Image
>>> import requests
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> feature_extractor = ConditionalDetrFeatureExtractor.from_pretrained("microsoft/conditional-detr-resnet-50")
>>> model = ConditionalDetrForObjectDetection.from_pretrained("microsoft/conditional-detr-resnet-50")
>>> inputs = feature_extractor(images=image, return_tensors="pt")
>>> outputs = model(**inputs)
>>> # model predicts bounding boxes and corresponding COCO classes
>>> logits = outputs.logits
>>> bboxes = outputs.pred_boxes
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# First, sent images through CONDITIONAL_DETR base model to obtain encoder + decoder outputs
outputs = self.model(
pixel_values,
pixel_mask=pixel_mask,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_outputs,
inputs_embeds=inputs_embeds,
decoder_inputs_embeds=decoder_inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
# class logits + predicted bounding boxes
logits = self.class_labels_classifier(sequence_output)
reference = outputs.reference_points if return_dict else outputs[-1]
reference_before_sigmoid = inverse_sigmoid(reference).transpose(0, 1)
outputs_coords = []
hs = sequence_output
tmp = self.bbox_predictor(hs)
tmp[..., :2] += reference_before_sigmoid
pred_boxes = tmp.sigmoid()
# pred_boxes = self.bbox_predictor(sequence_output).sigmoid()
loss, loss_dict, auxiliary_outputs = None, None, None
if labels is not None:
# First: create the matcher
matcher = ConditionalDetrHungarianMatcher(
class_cost=self.config.class_cost, bbox_cost=self.config.bbox_cost, giou_cost=self.config.giou_cost
)
# Second: create the criterion
losses = ["labels", "boxes", "cardinality"]
criterion = ConditionalDetrLoss(
matcher=matcher,
num_classes=self.config.num_labels,
focal_alpha=self.config.focal_alpha,
losses=losses,
)
criterion.to(self.device)
# Third: compute the losses, based on outputs and labels
outputs_loss = {}
outputs_loss["logits"] = logits
outputs_loss["pred_boxes"] = pred_boxes
if self.config.auxiliary_loss:
intermediate = outputs.intermediate_hidden_states if return_dict else outputs[4]
outputs_class = self.class_labels_classifier(intermediate)
for lvl in range(hs.shape[0]):
tmp = self.bbox_predictor(hs[lvl])
tmp[..., :2] += reference_before_sigmoid
outputs_coord = tmp.sigmoid()
outputs_coords.append(outputs_coord)
outputs_coord = torch.stack(outputs_coords)
auxiliary_outputs = self._set_aux_loss(outputs_class, outputs_coord)
outputs_loss["auxiliary_outputs"] = auxiliary_outputs
loss_dict = criterion(outputs_loss, labels)
# Fourth: compute total loss, as a weighted sum of the various losses
weight_dict = {"loss_ce": self.config.cls_loss_coefficient, "loss_bbox": self.config.bbox_loss_coefficient}
weight_dict["loss_giou"] = self.config.giou_loss_coefficient
if self.config.auxiliary_loss:
aux_weight_dict = {}
for i in range(self.config.decoder_layers - 1):
aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
weight_dict.update(aux_weight_dict)
loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
if not return_dict:
if auxiliary_outputs is not None:
output = (logits, pred_boxes) + auxiliary_outputs + outputs
else:
output = (logits, pred_boxes) + outputs
return ((loss, loss_dict) + output) if loss is not None else output
return ConditionalDetrObjectDetectionOutput(
loss=loss,
loss_dict=loss_dict,
logits=logits,
pred_boxes=pred_boxes,
auxiliary_outputs=auxiliary_outputs,
last_hidden_state=outputs.last_hidden_state,
decoder_hidden_states=outputs.decoder_hidden_states,
decoder_attentions=outputs.decoder_attentions,
cross_attentions=outputs.cross_attentions,
encoder_last_hidden_state=outputs.encoder_last_hidden_state,
encoder_hidden_states=outputs.encoder_hidden_states,
encoder_attentions=outputs.encoder_attentions,
)
@add_start_docstrings(
"""
CONDITIONAL_DETR Model (consisting of a backbone and encoder-decoder Transformer) with a segmentation head on top,
for tasks such as COCO panoptic.
""",
CONDITIONAL_DETR_START_DOCSTRING,
)
class ConditionalDetrForSegmentation(ConditionalDetrPreTrainedModel):
def __init__(self, config: ConditionalDetrConfig):
super().__init__(config)
# object detection model
self.conditional_detr = ConditionalDetrForObjectDetection(config)
# segmentation head
hidden_size, number_of_heads = config.d_model, config.encoder_attention_heads
intermediate_channel_sizes = self.conditional_detr.model.backbone.conv_encoder.intermediate_channel_sizes
self.mask_head = ConditionalDetrMaskHeadSmallConv(
hidden_size + number_of_heads, intermediate_channel_sizes[::-1][-3:], hidden_size
)
self.bbox_attention = ConditionalDetrMHAttentionMap(
hidden_size, hidden_size, number_of_heads, dropout=0.0, std=config.init_xavier_std
)
# Initialize weights and apply final processing
self.post_init()
@add_start_docstrings_to_model_forward(CONDITIONAL_DETR_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=ConditionalDetrSegmentationOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
pixel_values,
pixel_mask=None,
decoder_attention_mask=None,
encoder_outputs=None,
inputs_embeds=None,
decoder_inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (`List[Dict]` of len `(batch_size,)`, *optional*):
Labels for computing the bipartite matching loss, DICE/F-1 loss and Focal loss. List of dicts, each
dictionary containing at least the following 3 keys: 'class_labels', 'boxes' and 'masks' (the class labels,
bounding boxes and segmentation masks of an image in the batch respectively). The class labels themselves
should be a `torch.LongTensor` of len `(number of bounding boxes in the image,)`, the boxes a
`torch.FloatTensor` of shape `(number of bounding boxes in the image, 4)` and the masks a
`torch.FloatTensor` of shape `(number of bounding boxes in the image, height, width)`.
Returns:
Examples:
```python
>>> import io
>>> import requests
>>> from PIL import Image
>>> import torch
>>> import numpy
>>> from transformers import ConditionalDetrFeatureExtractor, ConditionalDetrForSegmentation
>>> from transformers.models.conditional_detr.feature_extraction_conditional_detr import rgb_to_id
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)
>>> feature_extractor = ConditionalDetrFeatureExtractor.from_pretrained(
... "facebook/conditional_detr-resnet-50-panoptic"
... )
>>> model = ConditionalDetrForSegmentation.from_pretrained("facebook/conditional_detr-resnet-50-panoptic")
>>> # prepare image for the model
>>> inputs = feature_extractor(images=image, return_tensors="pt")
>>> # forward pass
>>> outputs = model(**inputs)
>>> # use the `post_process_panoptic` method of `ConditionalDetrFeatureExtractor` to convert to COCO format
>>> processed_sizes = torch.as_tensor(inputs["pixel_values"].shape[-2:]).unsqueeze(0)
>>> result = feature_extractor.post_process_panoptic(outputs, processed_sizes)[0]
>>> # the segmentation is stored in a special-format png
>>> panoptic_seg = Image.open(io.BytesIO(result["png_string"]))
>>> panoptic_seg = numpy.array(panoptic_seg, dtype=numpy.uint8)
>>> # retrieve the ids corresponding to each mask
>>> panoptic_seg_id = rgb_to_id(panoptic_seg)
>>> panoptic_seg_id.shape
(800, 1066)
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
batch_size, num_channels, height, width = pixel_values.shape
device = pixel_values.device
if pixel_mask is None:
pixel_mask = torch.ones((batch_size, height, width), device=device)
# First, get list of feature maps and position embeddings
features, position_embeddings_list = self.conditional_detr.model.backbone(pixel_values, pixel_mask=pixel_mask)
# Second, apply 1x1 convolution to reduce the channel dimension to d_model (256 by default)
feature_map, mask = features[-1]
batch_size, num_channels, height, width = feature_map.shape
projected_feature_map = self.conditional_detr.model.input_projection(feature_map)
# Third, flatten the feature map + position embeddings of shape NxCxHxW to NxCxHW, and permute it to NxHWxC
# In other words, turn their shape into (batch_size, sequence_length, hidden_size)
flattened_features = projected_feature_map.flatten(2).permute(0, 2, 1)
position_embeddings = position_embeddings_list[-1].flatten(2).permute(0, 2, 1)
flattened_mask = mask.flatten(1)
# Fourth, sent flattened_features + flattened_mask + position embeddings through encoder
# flattened_features is a Tensor of shape (batch_size, heigth*width, hidden_size)
# flattened_mask is a Tensor of shape (batch_size, heigth*width)
if encoder_outputs is None:
encoder_outputs = self.conditional_detr.model.encoder(
inputs_embeds=flattened_features,
attention_mask=flattened_mask,
position_embeddings=position_embeddings,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
# If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
encoder_outputs = BaseModelOutput(
last_hidden_state=encoder_outputs[0],
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
)
# Fifth, sent query embeddings + position embeddings through the decoder (which is conditioned on the encoder output)
query_position_embeddings = self.conditional_detr.model.query_position_embeddings.weight.unsqueeze(0).repeat(
batch_size, 1, 1
)
queries = torch.zeros_like(query_position_embeddings)
# decoder outputs consists of (dec_features, dec_hidden, dec_attn)
decoder_outputs = self.conditional_detr.model.decoder(
inputs_embeds=queries,
attention_mask=None,
position_embeddings=position_embeddings,
query_position_embeddings=query_position_embeddings,
encoder_hidden_states=encoder_outputs[0],
encoder_attention_mask=flattened_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = decoder_outputs[0]
# Sixth, compute logits, pred_boxes and pred_masks
logits = self.conditional_detr.class_labels_classifier(sequence_output)
pred_boxes = self.conditional_detr.bbox_predictor(sequence_output).sigmoid()
memory = encoder_outputs[0].permute(0, 2, 1).view(batch_size, self.config.d_model, height, width)
mask = flattened_mask.view(batch_size, height, width)
# FIXME h_boxes takes the last one computed, keep this in mind
# important: we need to reverse the mask, since in the original implementation the mask works reversed
# bbox_mask is of shape (batch_size, num_queries, number_of_attention_heads in bbox_attention, height/32, width/32)
bbox_mask = self.bbox_attention(sequence_output, memory, mask=~mask)
seg_masks = self.mask_head(projected_feature_map, bbox_mask, [features[2][0], features[1][0], features[0][0]])
pred_masks = seg_masks.view(
batch_size, self.conditional_detr.config.num_queries, seg_masks.shape[-2], seg_masks.shape[-1]
)
loss, loss_dict, auxiliary_outputs = None, None, None
if labels is not None:
# First: create the matcher
matcher = ConditionalDetrHungarianMatcher(
class_cost=self.config.class_cost, bbox_cost=self.config.bbox_cost, giou_cost=self.config.giou_cost
)
# Second: create the criterion
losses = ["labels", "boxes", "cardinality", "masks"]
criterion = ConditionalDetrLoss(
matcher=matcher,
num_classes=self.config.num_labels,
focal_alpha=self.config.focal_alpha,
losses=losses,
)
criterion.to(self.device)
# Third: compute the losses, based on outputs and labels
outputs_loss = {}
outputs_loss["logits"] = logits
outputs_loss["pred_boxes"] = pred_boxes
outputs_loss["pred_masks"] = pred_masks
if self.config.auxiliary_loss:
intermediate = decoder_outputs.intermediate_hidden_states if return_dict else decoder_outputs[-1]
outputs_class = self.class_labels_classifier(intermediate)
outputs_coord = self.bbox_predictor(intermediate).sigmoid()
auxiliary_outputs = self._set_aux_loss(outputs_class, outputs_coord)
outputs_loss["auxiliary_outputs"] = auxiliary_outputs
loss_dict = criterion(outputs_loss, labels)
# Fourth: compute total loss, as a weighted sum of the various losses
weight_dict = {"loss_ce": 1, "loss_bbox": self.config.bbox_loss_coefficient}
weight_dict["loss_giou"] = self.config.giou_loss_coefficient
weight_dict["loss_mask"] = self.config.mask_loss_coefficient
weight_dict["loss_dice"] = self.config.dice_loss_coefficient
if self.config.auxiliary_loss:
aux_weight_dict = {}
for i in range(self.config.decoder_layers - 1):
aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
weight_dict.update(aux_weight_dict)
loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict)
if not return_dict:
if auxiliary_outputs is not None:
output = (logits, pred_boxes, pred_masks) + auxiliary_outputs + decoder_outputs + encoder_outputs
else:
output = (logits, pred_boxes, pred_masks) + decoder_outputs + encoder_outputs
return ((loss, loss_dict) + output) if loss is not None else output
return ConditionalDetrSegmentationOutput(
loss=loss,
loss_dict=loss_dict,
logits=logits,
pred_boxes=pred_boxes,
pred_masks=pred_masks,
auxiliary_outputs=auxiliary_outputs,
last_hidden_state=decoder_outputs.last_hidden_state,
decoder_hidden_states=decoder_outputs.hidden_states,
decoder_attentions=decoder_outputs.attentions,
cross_attentions=decoder_outputs.cross_attentions,
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
encoder_hidden_states=encoder_outputs.hidden_states,
encoder_attentions=encoder_outputs.attentions,
)
def _expand(tensor, length: int):
return tensor.unsqueeze(1).repeat(1, int(length), 1, 1, 1).flatten(0, 1)
# taken from https://github.com/facebookresearch/detr/blob/master/models/segmentation.py
# Copied from transformers.models.detr.modeling_detr.DetrMaskHeadSmallConv with Detr->ConditionalDetr
class ConditionalDetrMaskHeadSmallConv(nn.Module):
"""
Simple convolutional head, using group norm. Upsampling is done using a FPN approach
"""
def __init__(self, dim, fpn_dims, context_dim):
super().__init__()
if dim % 8 != 0:
raise ValueError(
"The hidden_size + number of attention heads must be divisible by 8 as the number of groups in"
" GroupNorm is set to 8"
)
inter_dims = [dim, context_dim // 2, context_dim // 4, context_dim // 8, context_dim // 16, context_dim // 64]
self.lay1 = nn.Conv2d(dim, dim, 3, padding=1)
self.gn1 = nn.GroupNorm(8, dim)
self.lay2 = nn.Conv2d(dim, inter_dims[1], 3, padding=1)
self.gn2 = nn.GroupNorm(8, inter_dims[1])
self.lay3 = nn.Conv2d(inter_dims[1], inter_dims[2], 3, padding=1)
self.gn3 = nn.GroupNorm(8, inter_dims[2])
self.lay4 = nn.Conv2d(inter_dims[2], inter_dims[3], 3, padding=1)
self.gn4 = nn.GroupNorm(8, inter_dims[3])
self.lay5 = nn.Conv2d(inter_dims[3], inter_dims[4], 3, padding=1)
self.gn5 = nn.GroupNorm(8, inter_dims[4])
self.out_lay = nn.Conv2d(inter_dims[4], 1, 3, padding=1)
self.dim = dim
self.adapter1 = nn.Conv2d(fpn_dims[0], inter_dims[1], 1)
self.adapter2 = nn.Conv2d(fpn_dims[1], inter_dims[2], 1)
self.adapter3 = nn.Conv2d(fpn_dims[2], inter_dims[3], 1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_uniform_(m.weight, a=1)
nn.init.constant_(m.bias, 0)
def forward(self, x: Tensor, bbox_mask: Tensor, fpns: List[Tensor]):
# here we concatenate x, the projected feature map, of shape (batch_size, d_model, heigth/32, width/32) with
# the bbox_mask = the attention maps of shape (batch_size, n_queries, n_heads, height/32, width/32).
# We expand the projected feature map to match the number of heads.
x = torch.cat([_expand(x, bbox_mask.shape[1]), bbox_mask.flatten(0, 1)], 1)
x = self.lay1(x)
x = self.gn1(x)
x = nn.functional.relu(x)
x = self.lay2(x)
x = self.gn2(x)
x = nn.functional.relu(x)
cur_fpn = self.adapter1(fpns[0])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + nn.functional.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay3(x)
x = self.gn3(x)
x = nn.functional.relu(x)
cur_fpn = self.adapter2(fpns[1])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + nn.functional.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay4(x)
x = self.gn4(x)
x = nn.functional.relu(x)
cur_fpn = self.adapter3(fpns[2])
if cur_fpn.size(0) != x.size(0):
cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0))
x = cur_fpn + nn.functional.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest")
x = self.lay5(x)
x = self.gn5(x)
x = nn.functional.relu(x)
x = self.out_lay(x)
return x
# Copied from transformers.models.detr.modeling_detr.DetrMHAttentionMap with Detr->ConditionalDetr
class ConditionalDetrMHAttentionMap(nn.Module):
"""This is a 2D attention module, which only returns the attention softmax (no multiplication by value)"""
def __init__(self, query_dim, hidden_dim, num_heads, dropout=0.0, bias=True, std=None):
super().__init__()
self.num_heads = num_heads
self.hidden_dim = hidden_dim
self.dropout = nn.Dropout(dropout)
self.q_linear = nn.Linear(query_dim, hidden_dim, bias=bias)
self.k_linear = nn.Linear(query_dim, hidden_dim, bias=bias)
self.normalize_fact = float(hidden_dim / self.num_heads) ** -0.5
def forward(self, q, k, mask: Optional[Tensor] = None):
q = self.q_linear(q)
k = nn.functional.conv2d(k, self.k_linear.weight.unsqueeze(-1).unsqueeze(-1), self.k_linear.bias)
queries_per_head = q.view(q.shape[0], q.shape[1], self.num_heads, self.hidden_dim // self.num_heads)
keys_per_head = k.view(k.shape[0], self.num_heads, self.hidden_dim // self.num_heads, k.shape[-2], k.shape[-1])
weights = torch.einsum("bqnc,bnchw->bqnhw", queries_per_head * self.normalize_fact, keys_per_head)
if mask is not None:
weights.masked_fill_(mask.unsqueeze(1).unsqueeze(1), torch.finfo(weights.dtype).min)
weights = nn.functional.softmax(weights.flatten(2), dim=-1).view(weights.size())
weights = self.dropout(weights)
return weights
def dice_loss(inputs, targets, num_boxes):
"""
Compute the DICE loss, similar to generalized IOU for masks
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs (0 for the negative class and 1 for the positive
class).
"""
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
numerator = 2 * (inputs * targets).sum(1)
denominator = inputs.sum(-1) + targets.sum(-1)
loss = 1 - (numerator + 1) / (denominator + 1)
return loss.sum() / num_boxes
def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2):
"""
Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs (0 for the negative class and 1 for the positive
class).
alpha: (optional) Weighting factor in range (0,1) to balance
positive vs negative examples. Default = -1 (no weighting).
gamma: Exponent of the modulating factor (1 - p_t) to
balance easy vs hard examples.
Returns:
Loss tensor
"""
prob = inputs.sigmoid()
ce_loss = nn.functional.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
p_t = prob * targets + (1 - prob) * (1 - targets)
loss = ce_loss * ((1 - p_t) ** gamma)
if alpha >= 0:
alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
loss = alpha_t * loss
return loss.mean(1).sum() / num_boxes
# taken from https://github.com/Atten4Vis/conditionalDETR/blob/master/models/conditional_detr.py
class ConditionalDetrLoss(nn.Module):
"""
This class computes the losses for ConditionalDetrForObjectDetection/ConditionalDetrForSegmentation. The process
happens in two steps: 1) we compute hungarian assignment between ground truth boxes and the outputs of the model 2)
we supervise each pair of matched ground-truth / prediction (supervise class and box).
Args:
matcher (`ConditionalDetrHungarianMatcher`):
Module able to compute a matching between targets and proposals.
num_classes (`int`):
Number of object categories, omitting the special no-object category.
focal_alpha (`float`):
Alpha parmeter in focal loss.
losses (`List[str]`):
List of all the losses to be applied. See `get_loss` for a list of all available losses.
"""
def __init__(self, matcher, num_classes, focal_alpha, losses):
super().__init__()
self.matcher = matcher
self.num_classes = num_classes
self.focal_alpha = focal_alpha
self.losses = losses
# removed logging parameter, which was part of the original implementation
def loss_labels(self, outputs, targets, indices, num_boxes):
"""
Classification loss (Binary focal loss) targets dicts must contain the key "class_labels" containing a tensor
of dim [nb_target_boxes]
"""
if "logits" not in outputs:
raise KeyError("No logits were found in the outputs")
src_logits = outputs["logits"]
idx = self._get_src_permutation_idx(indices)
target_classes_o = torch.cat([t["class_labels"][J] for t, (_, J) in zip(targets, indices)])
target_classes = torch.full(
src_logits.shape[:2], self.num_classes, dtype=torch.int64, device=src_logits.device
)
target_classes[idx] = target_classes_o
target_classes_onehot = torch.zeros(
[src_logits.shape[0], src_logits.shape[1], src_logits.shape[2] + 1],
dtype=src_logits.dtype,
layout=src_logits.layout,
device=src_logits.device,
)
target_classes_onehot.scatter_(2, target_classes.unsqueeze(-1), 1)
target_classes_onehot = target_classes_onehot[:, :, :-1]
loss_ce = (
sigmoid_focal_loss(src_logits, target_classes_onehot, num_boxes, alpha=self.focal_alpha, gamma=2)
* src_logits.shape[1]
)
losses = {"loss_ce": loss_ce}
return losses
@torch.no_grad()
def loss_cardinality(self, outputs, targets, indices, num_boxes):
"""
Compute the cardinality error, i.e. the absolute error in the number of predicted non-empty boxes.
This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients.
"""
logits = outputs["logits"]
device = logits.device
tgt_lengths = torch.as_tensor([len(v["class_labels"]) for v in targets], device=device)
# Count the number of predictions that are NOT "no-object" (which is the last class)
card_pred = (logits.argmax(-1) != logits.shape[-1] - 1).sum(1)
card_err = nn.functional.l1_loss(card_pred.float(), tgt_lengths.float())
losses = {"cardinality_error": card_err}
return losses
def loss_boxes(self, outputs, targets, indices, num_boxes):
"""
Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss.
Targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]. The target boxes
are expected in format (center_x, center_y, w, h), normalized by the image size.
"""
if "pred_boxes" not in outputs:
raise KeyError("No predicted boxes found in outputs")
idx = self._get_src_permutation_idx(indices)
src_boxes = outputs["pred_boxes"][idx]
target_boxes = torch.cat([t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0)
loss_bbox = nn.functional.l1_loss(src_boxes, target_boxes, reduction="none")
losses = {}
losses["loss_bbox"] = loss_bbox.sum() / num_boxes
loss_giou = 1 - torch.diag(
generalized_box_iou(center_to_corners_format(src_boxes), center_to_corners_format(target_boxes))
)
losses["loss_giou"] = loss_giou.sum() / num_boxes
return losses
def loss_masks(self, outputs, targets, indices, num_boxes):
"""
Compute the losses related to the masks: the focal loss and the dice loss.
Targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w].
"""
if "pred_masks" not in outputs:
raise KeyError("No predicted masks found in outputs")
src_idx = self._get_src_permutation_idx(indices)
tgt_idx = self._get_tgt_permutation_idx(indices)
src_masks = outputs["pred_masks"]
src_masks = src_masks[src_idx]
masks = [t["masks"] for t in targets]
# TODO use valid to mask invalid areas due to padding in loss
target_masks, valid = nested_tensor_from_tensor_list(masks).decompose()
target_masks = target_masks.to(src_masks)
target_masks = target_masks[tgt_idx]
# upsample predictions to the target size
src_masks = nn.functional.interpolate(
src_masks[:, None], size=target_masks.shape[-2:], mode="bilinear", align_corners=False
)
src_masks = src_masks[:, 0].flatten(1)
target_masks = target_masks.flatten(1)
target_masks = target_masks.view(src_masks.shape)
losses = {
"loss_mask": sigmoid_focal_loss(src_masks, target_masks, num_boxes),
"loss_dice": dice_loss(src_masks, target_masks, num_boxes),
}
return losses
def _get_src_permutation_idx(self, indices):
# permute predictions following indices
batch_idx = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])
src_idx = torch.cat([src for (src, _) in indices])
return batch_idx, src_idx
def _get_tgt_permutation_idx(self, indices):
# permute targets following indices
batch_idx = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)])
tgt_idx = torch.cat([tgt for (_, tgt) in indices])
return batch_idx, tgt_idx
def get_loss(self, loss, outputs, targets, indices, num_boxes):
loss_map = {
"labels": self.loss_labels,
"cardinality": self.loss_cardinality,
"boxes": self.loss_boxes,
"masks": self.loss_masks,
}
if loss not in loss_map:
raise ValueError(f"Loss {loss} not supported")
return loss_map[loss](outputs, targets, indices, num_boxes)
def forward(self, outputs, targets):
"""
This performs the loss computation.
Args:
outputs (`dict`, *optional*):
Dictionary of tensors, see the output specification of the model for the format.
targets (`List[dict]`, *optional*):
List of dicts, such that len(targets) == batch_size. The expected keys in each dict depends on the
losses applied, see each loss' doc.
"""
outputs_without_aux = {k: v for k, v in outputs.items() if k != "auxiliary_outputs"}
# Retrieve the matching between the outputs of the last layer and the targets
indices = self.matcher(outputs_without_aux, targets)
# Compute the average number of target boxes accross all nodes, for normalization purposes
num_boxes = sum(len(t["class_labels"]) for t in targets)
num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device)
# (Niels): comment out function below, distributed training to be added
# if is_dist_avail_and_initialized():
# torch.distributed.all_reduce(num_boxes)
# (Niels) in original implementation, num_boxes is divided by get_world_size()
num_boxes = torch.clamp(num_boxes, min=1).item()
# Compute all the requested losses
losses = {}
for loss in self.losses:
losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes))
# In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
if "auxiliary_outputs" in outputs:
for i, auxiliary_outputs in enumerate(outputs["auxiliary_outputs"]):
indices = self.matcher(auxiliary_outputs, targets)
for loss in self.losses:
if loss == "masks":
# Intermediate masks losses are too costly to compute, we ignore them.
continue
l_dict = self.get_loss(loss, auxiliary_outputs, targets, indices, num_boxes)
l_dict = {k + f"_{i}": v for k, v in l_dict.items()}
losses.update(l_dict)
return losses
# Copied from transformers.models.detr.modeling_detr.DetrMLPPredictionHead with Detr->ConditionalDetr
class ConditionalDetrMLPPredictionHead(nn.Module):
"""
Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates,
height and width of a bounding box w.r.t. an image.
Copied from https://github.com/facebookresearch/detr/blob/master/models/detr.py
"""
def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
super().__init__()
self.num_layers = num_layers
h = [hidden_dim] * (num_layers - 1)
self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
def forward(self, x):
for i, layer in enumerate(self.layers):
x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
return x
class ConditionalDetrHungarianMatcher(nn.Module):
"""
This class computes an assignment between the targets and the predictions of the network.
For efficiency reasons, the targets don't include the no_object. Because of this, in general, there are more
predictions than targets. In this case, we do a 1-to-1 matching of the best predictions, while the others are
un-matched (and thus treated as non-objects).
Args:
class_cost:
The relative weight of the classification error in the matching cost.
bbox_cost:
The relative weight of the L1 error of the bounding box coordinates in the matching cost.
giou_cost:
The relative weight of the giou loss of the bounding box in the matching cost.
"""
def __init__(self, class_cost: float = 1, bbox_cost: float = 1, giou_cost: float = 1):
super().__init__()
requires_backends(self, ["scipy"])
self.class_cost = class_cost
self.bbox_cost = bbox_cost
self.giou_cost = giou_cost
if class_cost == 0 and bbox_cost == 0 and giou_cost == 0:
raise ValueError("All costs of the Matcher can't be 0")
@torch.no_grad()
def forward(self, outputs, targets):
"""
Args:
outputs (`dict`):
A dictionary that contains at least these entries:
* "logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
* "pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates.
targets (`List[dict]`):
A list of targets (len(targets) = batch_size), where each target is a dict containing:
* "class_labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of
ground-truth
objects in the target) containing the class labels
* "boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates.
Returns:
`List[Tuple]`: A list of size `batch_size`, containing tuples of (index_i, index_j) where:
- index_i is the indices of the selected predictions (in order)
- index_j is the indices of the corresponding selected targets (in order)
For each batch element, it holds: len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
"""
batch_size, num_queries = outputs["logits"].shape[:2]
# We flatten to compute the cost matrices in a batch
out_prob = outputs["logits"].flatten(0, 1).sigmoid() # [batch_size * num_queries, num_classes]
out_bbox = outputs["pred_boxes"].flatten(0, 1) # [batch_size * num_queries, 4]
# Also concat the target labels and boxes
tgt_ids = torch.cat([v["class_labels"] for v in targets])
tgt_bbox = torch.cat([v["boxes"] for v in targets])
# Compute the classification cost.
alpha = 0.25
gamma = 2.0
neg_cost_class = (1 - alpha) * (out_prob**gamma) * (-(1 - out_prob + 1e-8).log())
pos_cost_class = alpha * ((1 - out_prob) ** gamma) * (-(out_prob + 1e-8).log())
class_cost = pos_cost_class[:, tgt_ids] - neg_cost_class[:, tgt_ids]
# Compute the L1 cost between boxes
bbox_cost = torch.cdist(out_bbox, tgt_bbox, p=1)
# Compute the giou cost between boxes
giou_cost = -generalized_box_iou(center_to_corners_format(out_bbox), center_to_corners_format(tgt_bbox))
# Final cost matrix
cost_matrix = self.bbox_cost * bbox_cost + self.class_cost * class_cost + self.giou_cost * giou_cost
cost_matrix = cost_matrix.view(batch_size, num_queries, -1).cpu()
sizes = [len(v["boxes"]) for v in targets]
indices = [linear_sum_assignment(c[i]) for i, c in enumerate(cost_matrix.split(sizes, -1))]
return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
# below: bounding box utilities taken from https://github.com/facebookresearch/detr/blob/master/util/box_ops.py
def _upcast(t: Tensor) -> Tensor:
# Protects from numerical overflows in multiplications by upcasting to the equivalent higher type
if t.is_floating_point():
return t if t.dtype in (torch.float32, torch.float64) else t.float()
else:
return t if t.dtype in (torch.int32, torch.int64) else t.int()
# Copied from transformers.models.detr.modeling_detr.box_area
def box_area(boxes: Tensor) -> Tensor:
"""
Computes the area of a set of bounding boxes, which are specified by its (x1, y1, x2, y2) coordinates.
Args:
boxes (`torch.FloatTensor` of shape `(number_of_boxes, 4)`):
Boxes for which the area will be computed. They are expected to be in (x1, y1, x2, y2) format with `0 <= x1
< x2` and `0 <= y1 < y2`.
Returns:
`torch.FloatTensor`: a tensor containing the area for each box.
"""
boxes = _upcast(boxes)
return (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
# Copied from transformers.models.detr.modeling_detr.box_iou
def box_iou(boxes1, boxes2):
area1 = box_area(boxes1)
area2 = box_area(boxes2)
left_top = torch.max(boxes1[:, None, :2], boxes2[:, :2]) # [N,M,2]
right_bottom = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) # [N,M,2]
width_height = (right_bottom - left_top).clamp(min=0) # [N,M,2]
inter = width_height[:, :, 0] * width_height[:, :, 1] # [N,M]
union = area1[:, None] + area2 - inter
iou = inter / union
return iou, union
# Copied from transformers.models.detr.modeling_detr.generalized_box_iou
def generalized_box_iou(boxes1, boxes2):
"""
Generalized IoU from https://giou.stanford.edu/. The boxes should be in [x0, y0, x1, y1] (corner) format.
Returns:
`torch.FloatTensor`: a [N, M] pairwise matrix, where N = len(boxes1) and M = len(boxes2)
"""
# degenerate boxes gives inf / nan results
# so do an early check
if not (boxes1[:, 2:] >= boxes1[:, :2]).all():
raise ValueError(f"boxes1 must be in [x0, y0, x1, y1] (corner) format, but got {boxes1}")
if not (boxes2[:, 2:] >= boxes2[:, :2]).all():
raise ValueError(f"boxes2 must be in [x0, y0, x1, y1] (corner) format, but got {boxes2}")
iou, union = box_iou(boxes1, boxes2)
top_left = torch.min(boxes1[:, None, :2], boxes2[:, :2])
bottom_right = torch.max(boxes1[:, None, 2:], boxes2[:, 2:])
width_height = (bottom_right - top_left).clamp(min=0) # [N,M,2]
area = width_height[:, :, 0] * width_height[:, :, 1]
return iou - (area - union) / area
# below: taken from https://github.com/facebookresearch/detr/blob/master/util/misc.py#L306
def _max_by_axis(the_list):
# type: (List[List[int]]) -> List[int]
maxes = the_list[0]
for sublist in the_list[1:]:
for index, item in enumerate(sublist):
maxes[index] = max(maxes[index], item)
return maxes
# Copied from transformers.models.detr.modeling_detr.NestedTensor
class NestedTensor(object):
def __init__(self, tensors, mask: Optional[Tensor]):
self.tensors = tensors
self.mask = mask
def to(self, device):
cast_tensor = self.tensors.to(device)
mask = self.mask
if mask is not None:
cast_mask = mask.to(device)
else:
cast_mask = None
return NestedTensor(cast_tensor, cast_mask)
def decompose(self):
return self.tensors, self.mask
def __repr__(self):
return str(self.tensors)
# Copied from transformers.models.detr.modeling_detr.nested_tensor_from_tensor_list
def nested_tensor_from_tensor_list(tensor_list: List[Tensor]):
if tensor_list[0].ndim == 3:
max_size = _max_by_axis([list(img.shape) for img in tensor_list])
batch_shape = [len(tensor_list)] + max_size
batch_size, num_channels, height, width = batch_shape
dtype = tensor_list[0].dtype
device = tensor_list[0].device
tensor = torch.zeros(batch_shape, dtype=dtype, device=device)
mask = torch.ones((batch_size, height, width), dtype=torch.bool, device=device)
for img, pad_img, m in zip(tensor_list, tensor, mask):
pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
m[: img.shape[1], : img.shape[2]] = False
else:
raise ValueError("Only 3-dimensional tensors are supported")
return NestedTensor(tensor, mask)
src/transformers/utils/dummy_timm_and_vision_objects.py
View file @ 126a7390
......@@ -3,6 +3,37 @@
from ..utils import DummyObject, requires_backends
CONDITIONAL_DETR_PRETRAINED_MODEL_ARCHIVE_LIST = None
class ConditionalDetrForObjectDetection(metaclass=DummyObject):
_backends = ["timm", "vision"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["timm", "vision"])
class ConditionalDetrForSegmentation(metaclass=DummyObject):
_backends = ["timm", "vision"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["timm", "vision"])
class ConditionalDetrModel(metaclass=DummyObject):
_backends = ["timm", "vision"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["timm", "vision"])
class ConditionalDetrPreTrainedModel(metaclass=DummyObject):
_backends = ["timm", "vision"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["timm", "vision"])
DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST = None
......
src/transformers/utils/dummy_vision_objects.py
View file @ 126a7390
......@@ -24,6 +24,13 @@ class CLIPFeatureExtractor(metaclass=DummyObject):
requires_backends(self, ["vision"])
class ConditionalDetrFeatureExtractor(metaclass=DummyObject):
_backends = ["vision"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["vision"])
class ConvNextFeatureExtractor(metaclass=DummyObject):
_backends = ["vision"]
......
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