CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped

Overview

CSWin-Transformer

PWC PWC

This repo is the official implementation of "CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows". The code and models for downstream tasks are coming soon.

Introduction

CSWin Transformer (the name CSWin stands for Cross-Shaped Window) is introduced in arxiv, which is a new general-purpose backbone for computer vision. It is a hierarchical Transformer and replaces the traditional full attention with our newly proposed cross-shaped window self-attention. The cross-shaped window self-attention mechanism computes self-attention in the horizontal and vertical stripes in parallel that from a cross-shaped window, with each stripe obtained by splitting the input feature into stripes of equal width. With CSWin, we could realize global attention with a limited computation cost.

CSWin Transformer achieves strong performance on ImageNet classification (87.5 on val with only 97G flops) and ADE20K semantic segmentation (55.7 mIoU on val), surpassing previous models by a large margin.

teaser

Main Results on ImageNet

model pretrain resolution [email protected] #params FLOPs 22K model 1K model
CSWin-T ImageNet-1K 224x224 82.8 23M 4.3G - model
CSWin-S ImageNet-1k 224x224 83.6 35M 6.9G - model
CSWin-B ImageNet-1k 224x224 84.2 78M 15.0G - model
CSWin-B ImageNet-1k 384x384 85.5 78M 47.0G - model
CSWin-L ImageNet-22k 224x224 86.5 173M 31.5G model model
CSWin-L ImageNet-22k 384x384 87.5 173M 96.8G - model

Main Results on Downstream Tasks

COCO Object Detection

backbone Method pretrain lr Schd box mAP mask mAP #params FLOPS
CSwin-T Mask R-CNN ImageNet-1K 3x 49.0 43.6 42M 279G
CSwin-S Mask R-CNN ImageNet-1K 3x 50.0 44.5 54M 342G
CSwin-B Mask R-CNN ImageNet-1K 3x 50.8 44.9 97M 526G
CSwin-T Cascade Mask R-CNN ImageNet-1K 3x 52.5 45.3 80M 757G
CSwin-S Cascade Mask R-CNN ImageNet-1K 3x 53.7 46.4 92M 820G
CSwin-B Cascade Mask R-CNN ImageNet-1K 3x 53.9 46.4 135M 1004G

ADE20K Semantic Segmentation (val)

Backbone Method pretrain Crop Size Lr Schd mIoU mIoU (ms+flip) #params FLOPs
CSwin-T Semantic FPN ImageNet-1K 512x512 80K 48.2 - 26M 202G
CSwin-S Semantic FPN ImageNet-1K 512x512 80K 49.2 - 39M 271G
CSwin-B Semantic FPN ImageNet-1K 512x512 80K 49.9 - 81M 464G
CSwin-T UPerNet ImageNet-1K 512x512 160K 49.3 50.4 60M 959G
CSwin-S UperNet ImageNet-1K 512x512 160K 50.0 50.8 65M 1027G
CSwin-B UperNet ImageNet-1K 512x512 160K 50.8 51.7 109M 1222G
CSwin-B UPerNet ImageNet-22K 640x640 160K 51.8 52.6 109M 1941G
CSwin-L UperNet ImageNet-22K 640x640 160K 53.4 55.7 208M 2745G

Requirements

timm==0.3.4, pytorch>=1.4, opencv, ... , run:

bash install_req.sh

Apex for mixed precision training is used for finetuning. To install apex, run:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Data prepare: ImageNet with the following folder structure, you can extract imagenet by this script.

│imagenet/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Train

Train the three lite variants: CSWin-Tiny, CSWin-Small and CSWin-Base:

bash train.sh 8 --data <data path> --model CSWin_64_12211_tiny_224 -b 256 --lr 2e-3 --weight-decay .05 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99984 --drop-path 0.2
bash train.sh 8 --data <data path> --model CSWin_64_24322_small_224 -b 256 --lr 2e-3 --weight-decay .05 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99984 --drop-path 0.4
bash train.sh 8 --data <data path> --model CSWin_96_24322_base_224 -b 128 --lr 1e-3 --weight-decay .1 --amp --img-size 224 --warmup-epochs 20 --model-ema-decay 0.99992 --drop-path 0.5

If you want to train our CSWin on images with 384x384 resolution, please use '--img-size 384'.

If the GPU memory is not enough, please use '-b 128 --lr 1e-3 --model-ema-decay 0.99992' or use checkpoint '--use-chk'.

Finetune

Finetune CSWin-Base with 384x384 resolution:

bash finetune.sh 8 --data <data path> --model CSWin_96_24322_base_384 -b 32 --lr 5e-6 --min-lr 5e-7 --weight-decay 1e-8 --amp --img-size 384 --warmup-epochs 0 --model-ema-decay 0.9998 --finetune <pretrained 224 model> --epochs 20 --mixup 0.1 --cooldown-epochs 10 --drop-path 0.7 --ema-finetune --lr-scale 1 --cutmix 0.1

Finetune ImageNet-22K pretrained CSWin-Large with 224x224 resolution:

bash finetune.sh 8 --data <data path> --model CSWin_144_24322_large_224 -b 64 --lr 2.5e-4 --min-lr 5e-7 --weight-decay 1e-8 --amp --img-size 224 --warmup-epochs 0 --model-ema-decay 0.9996 --finetune <22k-pretrained model> --epochs 30 --mixup 0.01 --cooldown-epochs 10 --interpolation bicubic  --lr-scale 0.05 --drop-path 0.2 --cutmix 0.3 --use-chk --fine-22k --ema-finetune

If the GPU memory is not enough, please use checkpoint '--use-chk'.

Cite CSWin Transformer

@misc{dong2021cswin,
      title={CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows}, 
        author={Xiaoyi Dong and Jianmin Bao and Dongdong Chen and Weiming Zhang and Nenghai Yu and Lu Yuan and Dong Chen and Baining Guo},
        year={2021},
        eprint={2107.00652},
        archivePrefix={arXiv},
        primaryClass={cs.CV}
}

Acknowledgement

This repository is built using the timm library and the DeiT repository.

License

This project is licensed under the license found in the LICENSE file in the root directory of this source tree.

Microsoft Open Source Code of Conduct

Contact Information

For help or issues using CSWin Transformer, please submit a GitHub issue.

For other communications related to CSWin Transformer, please contact Jianmin Bao ([email protected]), Dong Chen ([email protected]).

Comments
  • About the patches_resolution of the segmentation model

    About the patches_resolution of the segmentation model

    Hello, this work is interesting but I have some questions about the 'patches_resolution' of the segmentation model. I notice that the long side of the cross-shaped windows is the 'patches_resolution' rather than the real feature resoulution. For example, in the stage-3, the long side is 224 / 16 = 14. Do I understand it correctly? Does that make it impossible to exchannge information outside the 'patches_resolution' ?

    opened by danczs 2
  • The results of downstream task by my realization are poor

    The results of downstream task by my realization are poor

    There are some questions:

    1. the split size is still [1 2 7 7]?
    2. last stage branch_num is 2 or 1 ? The downstream task image resolution in last stages cannot equal to 7(split size). If not 1, the pretrained weights size is not matched
    3. pading is right in my realization ? pad_l = pad_t = 0 pad_r = (W_sp - W % W_sp) % W_sp pad_b = (H_sp - H % H_sp) % H_sp q = q.transpose(-2,-1).contiguous().view(B, H, W, C) k = q.transpose(-2,-1).contiguous().view(B, H, W, C) v = q.transpose(-2,-1).contiguous().view(B, H, W, C) if pad_r > 0 or pad_b > 0: q = F.pad(q, (0, 0, pad_l, pad_r, pad_t, pad_b)) k = F.pad(k, (0, 0, pad_l, pad_r, pad_t, pad_b)) v = F.pad(v, (0, 0, pad_l, pad_r, pad_t, pad_b)) _, Hp, Wp, _ = q.shape
    opened by Sunting78 2
  • Experiment setting for semantic segmentation

    Experiment setting for semantic segmentation

    Hi, thank you for the code. I implemented CSwin-T with FPN for semantic segmentation in ADE20K but couldn't reach the mIoU value of 48.2 as mentioned by you in the table. The maximum I could get was 39.9 mIoU, it will be great if you could share the exact experiment settings you used? Thanks

    opened by AnukritiSinghh 2
  • CSWin significantly slower than Swin?

    CSWin significantly slower than Swin?

    Greetings,

    From my benchmarks I have noticed that CSwin seems to be significantly slower than Swin when it comes to inference times, is this the expected behavior? While I can get predictions as fast as 20 miliseconds on Swin Large 384 it takes above 900 milisecond on CSWin_144_24322_large_384.

    I performed tests using FP16, torchscript, optimize_for_inference and torch.inference_mode

    opened by ErenBalatkan 2
  • Pretrained settings for object detection

    Pretrained settings for object detection

    Hi, I'm impressed by your excellent work.

    I have a question.

    I wonder which type of the pre-trained weights (224x224 or 384x384 finetuned) is used for object detection.

    I know both 224x224 and 384x384 are pre-trained on ImageNet-1k.

    opened by youngwanLEE 2
  • Error about building 384 models

    Error about building 384 models

    The code: https://github.com/microsoft/CSWin-Transformer/blob/d8be74a7833898f7bd9c77eb8c051d1b8bd5d753/models/cswin.py#L407 shoud be:

    model = CSWinTransformer(img_size=384, patch_size=4, embed_dim=96, depth=[2,4,32,2],
    

    And as the same: https://github.com/microsoft/CSWin-Transformer/blob/d8be74a7833898f7bd9c77eb8c051d1b8bd5d753/models/cswin.py#L414

    opened by TingquanGao 2
  • The problem is shown in the figure

    The problem is shown in the figure

    model = CSWinTransformer(patch_size=4, embed_dim=96, depth=[2,4,32,2], split_size=[1,2,12,12], num_heads=[4,8,16,32], mlp_ratio=4.).cuda().eval() inp = torch.rand(1, 3, 224, 224).cuda() outs = model(inp) for out in outs: print(out.shape)

    RuntimeError: shape '[1, 192, 1, 14, 1, 12]' is invalid for input of size 37632

    why? image

    opened by rui-cf 2
  • about the test result on imagenet-2012

    about the test result on imagenet-2012

    Hi! I've test the CSwin-Tiny-224 released pretrained weight, this is my data transforms during testing:

    DEFAULT_CROP_SIZE = 0.9
    scale_size = int(math.floor(image_size / DEFAULT_CROP_SIZE))
    transform = transforms.Compose(
            [
                transforms.Resize(scale_size, interpolation=3)  # 3: bibubic
                if image_size == 224
                else transforms.Resize(image_size, interpolation=3),
                transforms.CenterCrop(image_size),
                transforms.ToTensor(),
                transforms.Normalize(IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD),
            ]
        )
    

    I can only get 80.5% on imagenet2012 dataset which is inconsistent with the results as you mentioned in this repo, did I miss some details about the data-augmentation during testing?

    opened by rentainhe 1
  • CSwin Code for Segmentation with MMSegmentation

    CSwin Code for Segmentation with MMSegmentation

    Hi, Thank you for your work! I wonder if you plan to release the mmsegmentation code you used for the downstream segmentation task, just like in Swin-Transformer repo.

    Best,

    opened by WalBouss 1
  • How do you produce Table 9  (ablation on different attention mecahnisms) in the  paper?

    How do you produce Table 9 (ablation on different attention mecahnisms) in the paper?

    Hi, thanks for your nice work. I'm doing some comparison on different attention mechanisms, and want to follow your experimental settings. I meet two problems:

    1. Why the reported mIoU is 41.9 for Swin-T in Table 9, while it is 46.1 in Swin Paper?
    2. Can you provide detailed experimental settings for semantic segmentation and object detection in table 9 ?
    opened by rayleizhu 0
  • about the setting of --use_chk

    about the setting of --use_chk

    give the parameter of --use_chk can launch the torch.utils.checkpoint to save the GPU memory, and I wonder if this could hurt the final performance, thanks a lot!

    opened by go-ahead-maker 0
  • Input image normalization parameters for semantic segmentation.

    Input image normalization parameters for semantic segmentation.

    Hey, guys, cool work!

    Unfortunately, for me it is not quite obvious, which normalization parameters (mean, std) did you use, when trained semantic segmentation model on ADE20K. Are they still IMAGENET_DEFAULT_MEAN or you used another values?

    opened by NikAleksFed 0
  • Using transfer to train over food101

    Using transfer to train over food101

    Hi all! I'm trying to train a model for food101 using the using the CSWin_64_12211_tiny_224 model with its pretrained values. The thing is, during execution it looks like its training from 0 rather than reusing the pretrained weights. By this I mean the initial top5 accuracy is around 5% but my initial thoughts is that it should be higher than this.

    For this I loaded the pretrained model and changed it's classification layer in a separate script and saved it for use as follows

    model = create_model( 'CSWin_64_12211_tiny_224', pretrained=True, num_classes=1000, drop_rate=0.0, drop_connect_rate=None, # DEPRECATED, use drop_path drop_path_rate=0.2, drop_block_rate=None, global_pool=None, bn_tf=False, bn_momentum=None, bn_eps=None, checkpoint_path='', img_size=224, use_chk=True) chk_path = './pretrained/cswin_tiny_224.pth' load_checkpoint(model, chk_path) model.reset_classifier(101, 'max')

    These are some of the runs I tried ` bash finetune.sh 1 --data ../food-101 --model CSWin_64_12211_tiny_224 -b 32 --lr 5e-6 --min-lr 5e-7 --weight-decay 1e-8 --amp --img-size 224 --warmup-epochs 0 --model-ema-decay 0.9998 --epochs 20 --mixup 0.1 --cooldown-epochs 10 --drop-path 0.7 --ema-finetune --lr-scale 1 --cutmix 0.1 --use-chk --num-classes 101 --pretrained --finetune ./pretrained/CSWin_64_12211_tiny_224101.pth

    bash finetune.sh 1 --data ../food-101 --model CSWin_64_12211_tiny_224 -b 32 --lr 2e-3 --weight-decay .05 --amp --img-size 224 --warmup-epochs 0 --model-ema-decay 0.9998 --epochs 20 --cooldown-epochs 10 --drop-path 0.2 --ema-finetune --cutmix 0.1 --use-chk --num-classes 101 --initial-checkpoint ./pretrained/CSWin_64_12211_tiny_224101.pth --lr-scale 1.0 --output ./full_base `

    Is there something I'm missing or a proper way I should try this?

    Thanks in advance for any help! :)

    opened by andreynz691 0
Owner
Microsoft
Open source projects and samples from Microsoft
Microsoft
The coda and data for "Measuring Fine-Grained Domain Relevance of Terms: A Hierarchical Core-Fringe Approach" (ACL '21)

We propose a hierarchical core-fringe learning framework to measure fine-grained domain relevance of terms – the degree that a term is relevant to a broad (e.g., computer science) or narrow (e.g., de

Jie Huang 14 Oct 21, 2022
[CVPR 2021] NormalFusion: Real-Time Acquisition of Surface Normals for High-Resolution RGB-D Scanning

NormalFusion: Real-Time Acquisition of Surface Normals for High-Resolution RGB-D Scanning Project Page | Paper | Supplemental material #1 | Supplement

KAIST VCLAB 49 Nov 24, 2022
Reverse engineering Rosetta 2 in M1 Mac

Project Champollion About this project Rosetta 2 is an emulation mechanism to run the x86_64 applications on Arm-based Apple Silicon with Ahead-Of-Tim

FFRI Security, Inc. 258 Jan 07, 2023
CIFAR-10_train-test - training and testing codes for dataset CIFAR-10

CIFAR-10_train-test - training and testing codes for dataset CIFAR-10

Frederick Wang 3 Apr 26, 2022
AAAI 2022 paper - Unifying Model Explainability and Robustness for Joint Text Classification and Rationale Extraction

AT-BMC Unifying Model Explainability and Robustness for Joint Text Classification and Rationale Extraction (AAAI 2022) Paper Prerequisites Install pac

16 Nov 26, 2022
PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation (TPAMI).

PFENet This is the implementation of our paper PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation that has been accepted to IEE

DV Lab 230 Dec 31, 2022
Implement the Pareto Optimizer and pcgrad to make a self-adaptive loss for multi-task

multi-task_losses_optimizer Implement the Pareto Optimizer and pcgrad to make a self-adaptive loss for multi-task 已经实验过了,不会有cuda out of memory情况 ##Par

14 Dec 25, 2022
Julia and Matlab codes to simulated all problems in El-Hachem, McCue and Simpson (2021)

Substrate_Mediated_Invasion Julia and Matlab codes to simulated all problems in El-Hachem, McCue and Simpson (2021) 2DSolver.jl reproduces the simulat

Matthew Simpson 0 Nov 09, 2021
Tools to create pixel-wise object masks, bounding box labels (2D and 3D) and 3D object model (PLY triangle mesh) for object sequences filmed with an RGB-D camera.

Tools to create pixel-wise object masks, bounding box labels (2D and 3D) and 3D object model (PLY triangle mesh) for object sequences filmed with an RGB-D camera. This project prepares training and t

305 Dec 16, 2022
Generative Autoregressive, Normalized Flows, VAEs, Score-based models (GANVAS)

GANVAS-models This is an implementation of various generative models. It contains implementations of the following: Autoregressive Models: PixelCNN, G

MRSAIL (Mini Robotics, Software & AI Lab) 6 Nov 26, 2022
TorchX: A PyTorch Extension Library for More Efficient Deep Learning

TorchX TorchX: A PyTorch Extension Library for More Efficient Deep Learning. @misc{torchx, author = {Ansheng You and Changxu Wang}, title = {T

Donny You 8 May 28, 2022
Constraint-based geometry sketcher for blender

Constraint-based sketcher addon for Blender that allows to create precise 2d shapes by defining a set of geometric constraints like tangent, distance,

1.7k Dec 31, 2022
MoCoPnet - Deformable 3D Convolution for Video Super-Resolution

MoCoPnet: Exploring Local Motion and Contrast Priors for Infrared Small Target Super-Resolution Pytorch implementation of local motion and contrast pr

Xinyi Ying 28 Dec 15, 2022
Code release of paper Improving neural implicit surfaces geometry with patch warping

NeuralWarp: Improving neural implicit surfaces geometry with patch warping Project page | Paper Code release of paper Improving neural implicit surfac

François Darmon 167 Dec 30, 2022
Example of semantic segmentation in Keras

keras-semantic-segmentation-example Example of semantic segmentation in Keras Single class example: Generated data: random ellipse with random color o

53 Mar 23, 2022
A Number Recognition algorithm

Paddle-VisualAttention Results_Compared SVHN Dataset Methods Steps GPU Batch Size Learning Rate Patience Decay Step Decay Rate Training Speed (FPS) Ac

1 Nov 12, 2021
PyTorch implementation of SMODICE: Versatile Offline Imitation Learning via State Occupancy Matching

SMODICE: Versatile Offline Imitation Learning via State Occupancy Matching This is the official PyTorch implementation of SMODICE: Versatile Offline I

Jason Ma 14 Aug 30, 2022
Public repository created to store my custom-made tools for Just Dance (UbiArt Engine)

Woody's Just Dance Tools Public repository created to store my custom-made tools for Just Dance (UbiArt Engine) Development and updates Almost all of

Wodson de Andrade 8 Dec 24, 2022
A sequence of Jupyter notebooks featuring the 12 Steps to Navier-Stokes

CFD Python Please cite as: Barba, Lorena A., and Forsyth, Gilbert F. (2018). CFD Python: the 12 steps to Navier-Stokes equations. Journal of Open Sour

Barba group 2.6k Dec 30, 2022
Neural Fixed-Point Acceleration for Convex Optimization

Licensing The majority of neural-scs is licensed under the CC BY-NC 4.0 License, however, portions of the project are available under separate license

Facebook Research 27 Oct 06, 2022