Machine Learning

Recently I’ve been experimenting again with quite a few computer vision models (including Vision Mamba). The fact that researchers share their code make experimenting with models a much more pleasant experience! Sometimes, though, the released code is very sensitive to CUDA versions. Luckily, switching between CUDA versions on Ubuntu has now become a relatively easy task using update-alternatives. Install the first CUDA version (e.g. CUDA 11.8) This can be achieved by following the documentation by NVIDIA, found in the CUDA Toolkit Download page....

At the beginning of 2024, an article came out that made the whole computer vision community excited: “Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model”. It’s a very interesting approach that promises to mitigate the cost of self-attention in transformers using a state space model (SSM) based vision backbone. The authors of Vision Mamba (Vim) shared both the source code and some of the pre-trained model weights. It took me a bit to figure out how to put all the things together, so I thought to share this to help the community....