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.

Paper: (arXiv 2401.10166)

• updates
• abstract
• overview
• main results
• getting started
• star history
• citation
• acknowledgment

✅ Updates

• May. 7th, 2024: Update: Important! using torch.backends.cudnn.enabled=True in downstream tasks is quite slow, disable it in vmamba.py!

• April. 10th, 2024: Update: we have released arXiv 2401.10166v2, which contains lots of updates we made related to VMambav2!

• March. 20th, 2024: Update: we have released the configs/logs/checkpoints for classification/detection/segmentation of VMambav2. We'are still working on VMambav3!

• March. 16th, 2024: Improvement: we implemented models with channel_first data layout, which GREATLY raises the throughput of the model on A100 (On V100, due to the slow implementation of F.conv2d compared to F.linear, it would not speed up.), Try using norm_layer="ln2d" (when inferencing or training) rather than norm_layer="ln" to unlock this feature with almost no performance cost!

• March. 8th, 2024: Update + Improvement: we update the performance of VMamba-T, Vmamba-S, VMamba-B with nightly build, checkpoints and logs are coming soon. (Note that these models are trained without CrossScanTriton or forwardtype=v4, you can modify those configs yourself to raise the speed with almost no cost!)

• March. 8th, 2024: Improvement: we implemented CrossScan and CrossMerge in triton, which speed the training up again. CrossScan and CrossMerge implemented in triton is ~2x faster than implemented in pytorch. Meanwhile, use v4 rather than v3 or v2 in forwardtype also raise the speed GREATLY!.

• Feb. 26th, 2024: Improvement: we now support flexible output of selective scan. That means whatever type the input is, the output can always be float32. The feature is useful as when training with float16, the loss often get nan due to the overflow over float16. In the meantime, training with float32 costs more time. Input with float16 and output with float32 can be fast, but in the meantime, the loss is less likely to be NaN. Try SelectiveScanOflex with float16 input and float32 output to enjoy that feature!

• Feb. 22th, 2024: Pre-Release: we set a pre-release to share nightly-build checkpoints in classificaion. Feel free to enjoy those new features with faster code and higher performance!

• Feb. 18th, 2024: Release: all the checkpoints and logs of VMamba (VSSM version 0) in classification have been released. These checkpoints correspond to the experiments done before date #20240119, if there is any mismatch to the latest code in main, please let me know, and I'll fix that. This is related to issue#1 and issue#37.

• Feb. 16th, 2024: Fix bug + Improvement: SS2D.forward_corev1 is deprecated. Fixed some bugs related to issue#30 (in test_selective scan.py, we now compare ours with mamba_ssm rather than selective_scan_ref), issue#32, issue#31. backward nrow has been added and tested in selective_scan.

• Feb. 4th, 2024: Fix bug + Improvement: Do not use SS2D.forward_corev1 with float32=False for training (testing is ok), as it's unstable training in float16 for selective scan. We released SS2D.forward_corev2, which is in float32, and is faster than SS2D.forward_corev1.

• Feb. 1st, 2024: Fix bug: we now calculate FLOPs with the algrithm @albertgu provides, which will be bigger than previous calculation (which is based on the selective_scan_ref function, and ignores the hardware-aware algrithm).

• Jan. 31st, 2024: Add feature: selective_scan now supports an extra argument nrow in [1, 2, 4]. If you find your device is strong and the time consumption keeps as d_state rises, try this feature to speed up nrows x without any cost ! Note this feature is actually a bug fix for mamba.

• Jan. 28th, 2024: Add feature: 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: Add feature: 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.

• 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 & Derivations

• 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

📖 Attention: The configs/logs/checkpoints of Classification on ImageNet-1K, Object Detection on COCO, Semantic Segmentation on ADE20K listed below corresponds to VMambav2arXiv 2401.10166v2, which is also named V9 in section Accelerating VMamba.

📖 Attention: The configs/logs/checkpoints of Classification on ImageNet-1K, Object Detection on COCO, Semantic Segmentation on ADE20K corresponding to arXiv 2401.10166v1 has been moved here.

Classification on ImageNet-1K with VMambav2

name	pretrain	resolution	acc@1	#params	FLOPs	configs/logs/ckpts	best epoch	use ema	GPU Mem	time/epoch
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(0230)	ImageNet-1K	224x224	82.5	30M	4.8G	config/log/ckpt	262	true	18234M	8.12min
VMamba-S	ImageNet-1K	224x224	83.6	50M	8.7G	config/log/ckpt	222	true	27634M	11.86min
VMamba-B	ImageNet-1K	224x224	83.9	89M	15.4G	config/log/ckpt	237	true	37122M	15.08min

• Models in this subsection is trained from scratch with random or manual initialization.

• We use ema because our model is still under development.

• we now calculate FLOPs with the algrithm @albertgu provides, which will be bigger than previous calculation (which is based on the selective_scan_ref function, and ignores the hardware-aware algrithm).

Object Detection on COCO with VMambav2

Backbone	#params	FLOPs	Detector	box mAP	mask mAP	configs/logs/ckpts	best epoch
Swin-T	48M	267G	MaskRCNN@1x	42.7	39.3	--	--
VMamba-T	50M	270G	MaskRCNN@1x	47.4	42.7	config/log/ckpt	12
Swin-S	69M	354G	MaskRCNN@1x	44.8	40.9	--	--
VMamba-S	70M	384G	MaskRCNN@1x	48.7	43.7	config/log/ckpt	11
Swin-B	107M	496G	MaskRCNN@1x	46.9	42.3	--	--
VMamba-B*	108M	485G	MaskRCNN@1x	49.2	43.9	config/log/ckpt	12
Swin-T	48M	267G	MaskRCNN@3x	46.0	41.6	--	--
VMamba-T	50M	270G	MaskRCNN@3x	48.9	43.7	config/log/ckpt	36
Swin-S	69M	354G	MaskRCNN@3x	48.2	43.2	--	--
VMamba-S	70M	384G	MaskRCNN@3x	49.9	44.2	config/log/ckpt	32

• Models in this subsection is initialized from the models trained in classfication.

• The total batch size of VMamba-B in COCO is 8, which is supposed to be 16 as in other experiments. This is a mistake, not feature. We may fix that later.

• we now calculate FLOPs with the algrithm @albertgu provides, which will be bigger than previous calculation (which is based on the selective_scan_ref function, and ignores the hardware-aware algrithm).

Semantic Segmentation on ADE20K with VMambav2

Backbone	Input	#params	FLOPs	Segmentor	mIoU(SS)	mIoU(MS)	configs/logs/logs(ms)/ckpts	best iter
Swin-T	512x512	60M	945G	UperNet@160k	44.4	45.8	--	--
VMamba-T	512x512	62M	948G	UperNet@160k	48.3	48.6	config/log/log(ms)/ckpt	160k
Swin-S	512x512	81M	1039G	UperNet@160k	47.6	49.5	--	--
VMamba-S	512x512	82M	1028G	UperNet@160k	50.6	51.2	config/log/log(ms)/ckpt	144k
Swin-B	512x512	121M	1188G	UperNet@160k	48.1	49.7	--
VMamba-B	512x512	122M	1170G	UperNet@160k	51.0	51.6	config/log/log(ms)/ckpt	160k

• Models in this subsection is initialized from the models trained in classfication.

• we now calculate FLOPs with the algrithm @albertgu provides, which will be bigger than previous calculation (which is based on the selective_scan_ref function, and ignores the hardware-aware algrithm).

Getting Started

Installation

Step 1: Clone the VMamba repository:

To get started, first clone the VMamba repository and navigate to the project directory:

git clone https://github.com/MzeroMiko/VMamba.git
cd VMamba

Step 2: Environment Setup:

VMamba recommends setting up a conda environment and installing dependencies via pip. Use the following commands to set up your environment:

Create and activate a new conda environment

conda create -n vmamba
conda activate vmamba

Install Dependencies

pip install -r requirements.txt
cd kernels/selective_scan && pip install .

Check Selective Scan (optional)

• If you want to check the modules compared with mamba_ssm, install mamba_ssm first!

• If you want to check if the implementation of selective scan of ours is the same with mamba_ssm, selective_scan/test_selective_scan.py is here for you. Change to MODE = "mamba_ssm_sscore" in selective_scan/test_selective_scan.py, and run pytest selective_scan/test_selective_scan.py.

• If you want to check if the implementation of selective scan of ours is the same with reference code (selective_scan_ref), change to MODE = "sscore" in selective_scan/test_selective_scan.py, and run pytest selective_scan/test_selective_scan.py.

• MODE = "mamba_ssm" stands for checking whether the results of mamba_ssm is close to selective_scan_ref, and "sstest" is preserved for development.

• If you find mamba_ssm (selective_scan_cuda) or selective_scan ( selctive_scan_cuda_core) is not close enough to selective_scan_ref, and the test failed, do not worry. Check if mamba_ssm and selective_scan are close enough instead.

• If you are interested in selective scan, you can check mamba, mamba-mini, mamba.py mamba-minimal for more information.

Dependencies for Detection and Segmentation (optional)

pip install mmengine==0.10.1 mmcv==2.1.0 opencv-python-headless ftfy regex
pip install mmdet==3.3.0 mmsegmentation==1.2.2 mmpretrain==1.2.0

Model Training and Inference

Classification

To train VMamba models for classification on ImageNet, use the following commands for different configurations:

python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=8 --master_addr="127.0.0.1" --master_port=29501 main.py --cfg </path/to/config> --batch-size 128 --data-path </path/of/dataset> --output /tmp

If you only want to test the performance (together with params and flops):

python -m torch.distributed.launch --nnodes=1 --node_rank=0 --nproc_per_node=1 --master_addr="127.0.0.1" --master_port=29501 main.py --cfg </path/to/config> --batch-size 128 --data-path </path/of/dataset> --output /tmp --pretrained </path/of/checkpoint>

Detection and Segmentation

To evaluate with mmdetection or mmsegmentation:

bash ./tools/dist_test.sh </path/to/config> </path/to/checkpoint> 1

use --tta to get the mIoU(ms) in segmentation

To train with mmdetection or mmsegmentation:

bash ./tools/dist_train.sh </path/to/config> 8

For more information about detection and segmentation tasks, please refer to the manual of mmdetection and mmsegmentation. Remember to use the appropriate backbone configurations in the configs directory.

Analysis Tools

VMamba includes tools for visualizing mamba "attention" and effective receptive field, analysing throughput and train-throughput. Use the following commands to perform analysis:

# Visualize Mamba "Attention"
CUDA_VISIBLE_DEVICES=0 python analyze/attnmap.py

# Analyze the effective receptive field
CUDA_VISIBLE_DEVICES=0 python analyze/erf.py

# Analyze the throughput and train throughput
CUDA_VISIBLE_DEVICES=0 python analyze/tp.py

Star History

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%)