SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Add Neighborhood Attention Transformer (NAT) and Dilated NAT (DiNAT) models (#20219)

Browse Source 
* Add DiNAT

* Adds DiNAT + tests

* Minor fixes

* Added HF model

* Add natten to dependencies.

* Cleanup

* Minor fixup

* Reformat

* Optional NATTEN import.

* Reformat & add doc to _toctree

* Reformat (finally)

* Dummy objects for DiNAT

* Add NAT + minor changes

Adds NAT as its own independent model + docs, tests
Adds NATTEN to ext deps to ensure ci picks it up.

* Remove natten from `all` and `dev-torch` deps, add manual pip install to ci tests

* Minor fixes.

* Fix READMEs.

* Requested changes to docs + minor fixes.

* Requested changes.

* Add NAT/DiNAT tests to layoutlm_job

* Correction to Dinat doc.

* Requested changes.
This commit is contained in:
Ali Hassani
2022-11-18 10:08:26 -08:00
committed by GitHub
parent 8d6de0b9cf
commit fc4a993e1b
34 changed files with 3067 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
docs/source/en/_toctree.yml
View File

@@ -396,6 +396,8 @@
title: DeiT
- local: model_doc/detr
title: DETR
- local: model_doc/dinat
title: DiNAT
- local: model_doc/dit
title: DiT
- local: model_doc/dpt
@@ -412,6 +414,8 @@
title: MobileNetV2
- local: model_doc/mobilevit
title: MobileViT
- local: model_doc/nat
title: NAT
- local: model_doc/poolformer
title: PoolFormer
- local: model_doc/regnet

4
docs/source/en/index.mdx
View File

@@ -83,6 +83,7 @@ The documentation is organized into five sections:
1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
1. **[DiNAT](model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
1. **[Donut](model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
@@ -136,6 +137,7 @@ The documentation is organized into five sections:
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
1. **[NAT](model_doc/nat)** (from SHI Labs) released with the paper [Neighborhood Attention Transformer](https://arxiv.org/abs/2204.07143) by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
1. **[Nezha](model_doc/nezha)** (from Huawei Noah’s Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
@@ -244,6 +246,7 @@ Flax), PyTorch, and/or TensorFlow.
| Deformable DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DeiT | ❌ | ❌ | ✅ | ✅ | ❌ |
| DETR | ❌ | ❌ | ✅ | ❌ | ❌ |
| DiNAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| DistilBERT | ✅ | ✅ | ✅ | ✅ | ✅ |
| DonutSwin | ❌ | ❌ | ✅ | ❌ | ❌ |
| DPR | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -290,6 +293,7 @@ Flax), PyTorch, and/or TensorFlow.
| MPNet | ✅ | ✅ | ✅ | ✅ | ❌ |
| MT5 | ✅ | ✅ | ✅ | ✅ | ✅ |
| MVP | ✅ | ✅ | ✅ | ❌ | ❌ |
| NAT | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nezha | ❌ | ❌ | ✅ | ❌ | ❌ |
| Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ |
| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ |

78
docs/source/en/model_doc/dinat.mdx Normal file
View File

@@ -0,0 +1,78 @@
<!--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.
-->
# Dilated Neighborhood Attention Transformer
## Overview
DiNAT was proposed in [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001)
by Ali Hassani and Humphrey Shi.
It extends [NAT](nat) by adding a Dilated Neighborhood Attention pattern to capture global context,
and shows significant performance improvements over it.
The abstract from the paper is the following:
*Transformers are quickly becoming one of the most heavily applied deep learning architectures across modalities,
domains, and tasks. In vision, on top of ongoing efforts into plain transformers, hierarchical transformers have
also gained significant attention, thanks to their performance and easy integration into existing frameworks.
These models typically employ localized attention mechanisms, such as the sliding-window Neighborhood Attention (NA)
or Swin Transformer's Shifted Window Self Attention. While effective at reducing self attention's quadratic complexity,
local attention weakens two of the most desirable properties of self attention: long range inter-dependency modeling,
and global receptive field. In this paper, we introduce Dilated Neighborhood Attention (DiNA), a natural, flexible and
efficient extension to NA that can capture more global context and expand receptive fields exponentially at no
additional cost. NA's local attention and DiNA's sparse global attention complement each other, and therefore we
introduce Dilated Neighborhood Attention Transformer (DiNAT), a new hierarchical vision transformer built upon both.
DiNAT variants enjoy significant improvements over strong baselines such as NAT, Swin, and ConvNeXt.
Our large model is faster and ahead of its Swin counterpart by 1.5% box AP in COCO object detection,
1.3% mask AP in COCO instance segmentation, and 1.1% mIoU in ADE20K semantic segmentation.
Paired with new frameworks, our large variant is the new state of the art panoptic segmentation model on COCO (58.2 PQ)
and ADE20K (48.5 PQ), and instance segmentation model on Cityscapes (44.5 AP) and ADE20K (35.4 AP) (no extra data).
It also matches the state of the art specialized semantic segmentation models on ADE20K (58.2 mIoU),
and ranks second on Cityscapes (84.5 mIoU) (no extra data). *
Tips:
- One can use the [`AutoImageProcessor`] API to prepare images for the model.
- DiNAT can be used as a *backbone*. When `output_hidden_states = True`,
it will output both `hidden_states` and `reshaped_hidden_states`. The `reshaped_hidden_states` have a shape of `(batch, num_channels, height, width)` rather than `(batch_size, height, width, num_channels)`.
Notes:
- DiNAT depends on [NATTEN](https://github.com/SHI-Labs/NATTEN/)'s implementation of Neighborhood Attention and Dilated Neighborhood Attention.
You can install it with pre-built wheels for Linux by referring to [shi-labs.com/natten](https://shi-labs.com/natten), or build on your system by running `pip install natten`.
Note that the latter will likely take time to compile. NATTEN does not support Windows devices yet.
- Patch size of 4 is only supported at the moment.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/dilated-neighborhood-attention-pattern.jpg"
alt="drawing" width="600"/>
<small> Neighborhood Attention with different dilation values.
Taken from the <a href="https://arxiv.org/abs/2209.15001">original paper</a>.</small>
This model was contributed by [Ali Hassani](https://huggingface.co/alihassanijr).
The original code can be found [here](https://github.com/SHI-Labs/Neighborhood-Attention-Transformer).
## DinatConfig
[[autodoc]] DinatConfig
## DinatModel
[[autodoc]] DinatModel
- forward
## DinatForImageClassification
[[autodoc]] DinatForImageClassification
- forward

73
docs/source/en/model_doc/nat.mdx Normal file
View File

@@ -0,0 +1,73 @@
<!--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.
-->
# Neighborhood Attention Transformer
## Overview
NAT was proposed in [Neighborhood Attention Transformer](https://arxiv.org/abs/2204.07143)
by Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi.
It is a hierarchical vision transformer based on Neighborhood Attention, a sliding-window self attention pattern.
The abstract from the paper is the following:
*We present Neighborhood Attention (NA), the first efficient and scalable sliding-window attention mechanism for vision.
NA is a pixel-wise operation, localizing self attention (SA) to the nearest neighboring pixels, and therefore enjoys a
linear time and space complexity compared to the quadratic complexity of SA. The sliding-window pattern allows NA's
receptive field to grow without needing extra pixel shifts, and preserves translational equivariance, unlike
Swin Transformer's Window Self Attention (WSA). We develop NATTEN (Neighborhood Attention Extension), a Python package
with efficient C++ and CUDA kernels, which allows NA to run up to 40% faster than Swin's WSA while using up to 25% less
memory. We further present Neighborhood Attention Transformer (NAT), a new hierarchical transformer design based on NA
that boosts image classification and downstream vision performance. Experimental results on NAT are competitive;
NAT-Tiny reaches 83.2% top-1 accuracy on ImageNet, 51.4% mAP on MS-COCO and 48.4% mIoU on ADE20K, which is 1.9%
ImageNet accuracy, 1.0% COCO mAP, and 2.6% ADE20K mIoU improvement over a Swin model with similar size. *
Tips:
- One can use the [`AutoImageProcessor`] API to prepare images for the model.
- NAT can be used as a *backbone*. When `output_hidden_states = True`,
it will output both `hidden_states` and `reshaped_hidden_states`.
The `reshaped_hidden_states` have a shape of `(batch, num_channels, height, width)` rather than
`(batch_size, height, width, num_channels)`.
Notes:
- NAT depends on [NATTEN](https://github.com/SHI-Labs/NATTEN/)'s implementation of Neighborhood Attention.
You can install it with pre-built wheels for Linux by referring to [shi-labs.com/natten](https://shi-labs.com/natten),
or build on your system by running `pip install natten`.
Note that the latter will likely take time to compile. NATTEN does not support Windows devices yet.
- Patch size of 4 is only supported at the moment.
<img
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/neighborhood-attention-pattern.jpg"
alt="drawing" width="600"/>
<small> Neighborhood Attention compared to other attention patterns.
Taken from the <a href="https://arxiv.org/abs/2204.07143">original paper</a>.</small>
This model was contributed by [Ali Hassani](https://huggingface.co/alihassanijr).
The original code can be found [here](https://github.com/SHI-Labs/Neighborhood-Attention-Transformer).
## NatConfig
[[autodoc]] NatConfig
## NatModel
[[autodoc]] NatModel
- forward
## NatForImageClassification
[[autodoc]] NatForImageClassification
- forward