Awesome

Vim-F: Visual State Space Model Benefiting from Learning in the Frequency Domain

Introduction

This project is based on Vim (paper, code) and we appreciate this excellent work. You can simply replace main.py and models_mamba.py with our versions to reimplement our work. Among them, main.py has no substantial modifications, and only the code related to position embedding in the original file has been removed to fit our work.

ImageNet classification

Pre-training

Long sequence fine-tuning

Unlike the pre-training stage, Vim's source code did not automatically use the linear scaling rule to adjust the learning rate during the fine-tuning stage. We did not manually adjust the learning rate, so the performance declined.