Only implemented through torch: "bi - mamba2" , "vision- mamba2 -torch". support 1d/2d/3d/nd and support export by jit.script/onnx;
Nd-Mamba2 for any dimension by pytorch
BiMamba2, implemented solely using PyTorch, provides N-dimensional support, including support for 1d, 2d, and 3d data. With BiMamba2, you can easily stitch it into any model to improve accuracy.
The following paper provides a detailed introduction to the NdBimamba2 module of the project:
✅ Support for segmentation tasks (added vssd backbone based shuffle segmentation decoder)
✅ Support classification tasks (added vssd backbone based shuffle classification decoder)
✅ Support exporting to scientific and onnx formats
❌ Not supporting data formats such as 1d, 3d, and other dimensions
Reminder
If you want faster speed, you can replace Mamba2 in this project with Mamba2's official Cuda acceleration implementation and install various dependency packages according to official requirements. This will not affect the project's support for multi-dimensional data, but it will have an impact on the export of the model
Thanks
[Albert Gu], [Tri Dao] state-spaces/mamba - authors of the Mamba-2 architecture
[Yuheng Shi], [Minjing Dong] YuHengsss/VSSD - authors of the vssd architecture