VMamba

VMamba: Visual State Space Model

Yue Liu¹,Yunjie Tian¹,Yuzhong Zhao¹, Hongtian Yu¹, Lingxi Xie², Yaowei Wang³, Qixiang Ye¹, Yunfan Liu¹

¹ University of Chinese Academy of Sciences, ² HUAWEI Inc., ³ PengCheng Lab.

Updates

Jan. 28th, 2024: we cloned main into a new branch called 20240128-achieve, the main branch has experienced a great update now. The code now are much easier to use in your own project, and the training speed is faster! This new version is totally compatible with original one, and you can use previous checkpoints without any modification. But if you want to use exactly the same models as original ones, just change forward_core = self.forward_corev1 into forward_core = self.forward_corev0 in classification/models/vmamba/vmamba.py#SS2D or you can change into the branch 20240128-archive instead.
Jan. 23th, 2024: we add an alternative for mamba_ssm and causal_conv1d. Typing pip install . in selective_scan and you can get rid of those two packages. ~~Just turn self.forward_core = self.forward_corev0 to self.forward_core = self.forward_corev1 in classification/models/vmamba/vmamba.py#SS2D.__init__ to enjoy that feature.~~ The training speed is expected to raise from 20min/epoch for tiny in 8x4090GPU to 17min/epoch, GPU memory cost reduces too. We have not trained our model with this feature, and we'll try in the future.
Jan. 22th, 2024: We have released VMamba-T/S pre-trained weights. The ema weights should be converted before transferring to downstream tasks to match the module names using get_ckpt.py.
Jan. 19th, 2024: The source code for classification, object detection, and semantic segmentation are provided.

Abstract

Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) stand as the two most popular foundation models for visual representation learning. While CNNs exhibit remarkable scalability with linear complexity w.r.t. image resolution, ViTs surpass them in fitting capabilities despite contending with quadratic complexity. A closer inspection reveals that ViTs achieve superior visual modeling performance through the incorporation of global receptive fields and dynamic weights. This observation motivates us to propose a novel architecture that inherits these components while enhancing computational efficiency. To this end, we draw inspiration from the recently introduced state space model and propose the Visual State Space Model (VMamba), which achieves linear complexity without sacrificing global receptive fields. To address the encountered direction-sensitive issue, we introduce the Cross-Scan Module (CSM) to traverse the spatial domain and convert any non-causal visual image into order patch sequences. Extensive experimental results substantiate that VMamba not only demonstrates promising capabilities across various visual perception tasks, but also exhibits more pronounced advantages over established benchmarks as the image resolution increases.

Overview

VMamba serves as a general-purpose backbone for computer vision with linear complexity and shows the advantages of global receptive fields and dynamic weights.

2D-Selective-Scan of VMamba

VMamba has global effective receptive field

Main Results

We will release all the pre-trained models/logs in few days!

Classification on ImageNet-1K

name	pretrain	resolution	acc@1	#params	FLOPs	checkpoints/logs
DeiT-S	ImageNet-1K	224x224	79.8	22M	4.6G	--
DeiT-B	ImageNet-1K	224x224	81.8	86M	17.5G	--
DeiT-B	ImageNet-1K	384x384	83.1	86M	55.4G	--
Swin-T	ImageNet-1K	224x224	81.2	28M	4.5G	--
Swin-S	ImageNet-1K	224x224	83.2	50M	8.7G	--
Swin-B	ImageNet-1K	224x224	83.5	88M	15.4G	--
VMamba-T	ImageNet-1K	224x224	82.2	22M	4.5G	ckpt/log
VMamba-S	ImageNet-1K	224x224	83.5	44M	9.1G	ckpt/log
VMamba-B	ImageNet-1K	224x224	84.0	75M	15.2G	waiting

Object Detection on COCO

Backbone	#params	FLOPs	Detector	box mAP	mask mAP	checkpoints/logs
Swin-T	48M	267G	MaskRCNN@1x	42.7	39.3	--
VMamba-T	42M	262G	MaskRCNN@1x	46.5	42.1	ckpt/log
Swin-S	69M	354G	MaskRCNN@1x	44.8	40.9	--
VMamba-S	64M	357G	MaskRCNN@1x	48.2	43.0	ckpt/log
Swin-B	107M	496G	MaskRCNN@1x	46.9	42.3	--
VMamba-B	96M	482G	MaskRCNN@1x	48.5	43.1	ckpt/log
Swin-T	48M	267G	MaskRCNN@3x	46.0	41.6	--
VMamba-T	42M	262G	MaskRCNN@3x	48.5	43.2	ckpt/log
Swin-S	69M	354G	MaskRCNN@3x	48.2	43.2	--
VMamba-S	64M	357G	MaskRCNN@3x	49.7	44.0	ckpt/log

Semantic Segmentation on ADE20K

Backbone	Input	#params	FLOPs	Segmentor	mIoU	checkpoints/logs
Swin-T	512x512	60M	945G	UperNet@160k	44.4	--
VMamba-T	512x512	55M	939G	UperNet@160k	47.3	ckpt/log
Swin-S	512x512	81M	1039G	UperNet@160k	47.6	--
VMamba-S	512x512	76M	1037G	UperNet@160k	49.5	ckpt/log
Swin-B	512x512	121M	1188G	UperNet@160k	48.1	--
VMamba-B	512x512	110M	1167G	UperNet@160k	50.0	ckpt/log
Swin-S	640x640	81M	1614G	UperNet@160k	47.9	--
VMamba-S	640x640	76M	1620G	UperNet@160k	50.8	ckpt/log

Getting Started

Install required packages:

conda_env="vmamba"
nvcc -V
conda create -n ${conda_env} --clone base
conda init bash && source ~/.bashrc && conda activate vmamba
python -VV
pip -V
pip install torch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu117
# We use py110 cu117 torch113
pip install packaging
pip install timm==0.4.12
pip install pytest chardet yacs termcolor
pip install submitit tensorboardX
pip install fvcore
pip install seaborn

cd selective_scan && pip install . && pytest
# you can also install packages below as an alternative...
# pip install triton==2.0.0
# pip install causal_conv1d==1.0.0  # causal_conv1d-1.0.0+cu118torch1.13cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
# pip install mamba_ssm==1.0.1  # mamba_ssm-1.0.1+cu118torch1.13cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

See more details at modelcard.sh.

Citation

@article{liu2024vmamba,
  title={VMamba: Visual State Space Model},
  author={Liu, Yue and Tian, Yunjie and Zhao, Yuzhong and Yu, Hongtian and Xie, Lingxi and Wang, Yaowei and Ye, Qixiang and Liu, Yunfan},
  journal={arXiv preprint arXiv:2401.10166},
  year={2024}
}

Acknowledgment

This project is based on Mamba (paper, code), Swin-Transformer (paper, code), ConvNeXt (paper, code), OpenMMLab, and the analyze/get_erf.py is adopted from replknet, thanks for their excellent works.

We release Fast-iTPN recently, which reports the best performance on ImageNet-1K at Tiny/Small/Base level models as far as we know. (Tiny-24M-86.5%, Small-40M-87.8%, Base-85M-88.75%)

xinyuliu-jeffrey/VMamba