forward-backward ssm

[swarajnanda2021]

It appears, per algorithm 1 of the vision mamba paper, the state-space model runs bi-directionally along the sequence. But in this implementation I see that both forward and backward convolutions are standard 1D convolutions, rather than going backward and forward in space. Can you explain your rationale behind this?

Because otherwise, there seems to be no difference between the backward and forward operations, which would otherwise be providing the mamba block a bi-directional way of applying the state-space selection operation.

Upvote & Fund

• We're using Polar.sh so you can upvote and help fund this issue.
• We receive the funding once the issue is completed & confirmed by you.
• Thank you in advance for helping prioritize & fund our backlog.

[github-actions[bot]]

Hello there, thank you for opening an Issue ! 🙏🏻 The team was notified and they will get back to you asap.

[kyegomez]

@swarajnanda2021 yes I did this for now until I can understand what forward and backward into space means

[swarajnanda2021]

Thanks for the fast response. Per the stencil operation used by most practitioners, it seems to me this is just a forward and backward for-looping in the scan process. Concur?

[github-actions[bot]]

Stale issue message

[thucz]

So the forward and backward SSM have no difference in fact?