license: apache-2.0
pipeline_tag: image-classification
library_name: pytorch
Vim Model Card
This repository contains the model based on the paper PTQ4VM: Post-Training Quantization for Visual Mamba.
Model Details
Vision Mamba (Vim) is a generic backbone trained on the ImageNet-1K dataset for vision tasks.
- Developed by: HUST, Horizon Robotics, BAAI
- Model type: A generic vision backbone based on the bidirectional state space model (SSM) architecture.
- License: Non-commercial license
Github repository:
https://github.com/YoungHyun197/ptq4vm
Model Sources
- Repository: https://github.com/hustvl/Vim
- Paper: https://arxiv.org/abs/2401.09417
Uses
The primary use of Vim is research on vision tasks, e.g., classification, segmentation, detection, and instance segmentation, with an SSM-based backbone. The primary intended users of the model are researchers and hobbyists in computer vision, machine learning, and artificial intelligence.
How to Get Started with the Model
- You can replace the backbone for vision tasks with the proposed Vim: https://github.com/hustvl/Vim/blob/main/vim/models_mamba.py
- Then you can load this checkpoint and start training.
Training Details
Vim is pretrained on ImageNet-1K with classification supervision. The training data is around 1.3M images from ImageNet-1K dataset. See more details in this paper.
Evaluation
Vim-tiny is evaluated on ImageNet-1K val set, and achieves 76.1% Top-1 Acc. By further finetuning at finer granularity, Vim-tiny achieves 78.3% Top-1 Acc. See more details in this paper.
Additional Information
Citation Information
@article{vim,
title={Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model},
author={Lianghui Zhu and Bencheng Liao and Qian Zhang and Xinlong Wang and Wenyu Liu and Xinggang Wang},
journal={arXiv preprint arXiv:2401.09417},
year={2024}
}
@article{cho2024ptq4vm,
title={PTQ4VM: Post-Training Quantization for Visual Mamba},
author={Cho, Younghyun and Lee, Changhun and Kim, Seonggon and Park, Eunhyeok},
journal={arXiv preprint arXiv:2412.20386},
year={2024}
}