Vision Mamba

Implementation of Vision Mamba from the paper: "Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model" It's 2.8x faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on high-res images.

PAPER LINK

Installation

pip install vision-mamba

Usage

import torch
from vision_mamba.model import Vim

# Create a random tensor
x = torch.randn(1, 3, 224, 224)

# Create an instance of the Vim model
model = Vim(
    dim=256,  # Dimension of the model
    heads=8,  # Number of attention heads
    dt_rank=32,  # Rank of the dynamic routing tensor
    dim_inner=256,  # Inner dimension of the model
    d_state=256,  # State dimension of the model
    num_classes=1000,  # Number of output classes
    image_size=224,  # Size of the input image
    patch_size=16,  # Size of the image patch
    channels=3,  # Number of input channels
    dropout=0.1,  # Dropout rate
    depth=12,  # Depth of the model
)

# Perform a forward pass through the model
out = model(x)

# Print the shape and output of the forward pass
print(out.shape)
print(out)

Citation

@misc{zhu2024vision,
    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},
    year={2024},
    eprint={2401.09417},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

License

MIT

Todo

• Fix the encoder block with the forward and backward convolutions
• Make a training script for imagenet