search

MzeroMiko / VMamba

VMamba: Visual State Space Models，code is based on mamba
MIT License
2.06k stars 123 forks source link

SS2D Block #245

Open Xia-zx opened 3 months ago

Xia-zx commented 3 months ago

what is the diffences of SS2Dv0，SS2Dv1,SS2DV2,SS2Dm0 in vmamba.py

MzeroMiko commented 3 months ago

SS2Dv0 represent for the vanilla-VMamba in the arxiv paper, it just applied cross scan and cross merge into Mamba and added a Norm and the hieratical architecture into it.

SS2Dv2 represent for the VMamba, which adds sort of tricks into it to accelerate vanilla-VMamba while keeping its advantages.

There's no SS2Dv1 as I delete it.

SS2Dm0 represent for the support for Mamba2, but the code has not trained yet.

YCHYZW commented 1 month ago

about SS2Dv3？

MzeroMiko commented 1 month ago

SS2Dv3 (forwardtype=xv1a...) is a more simplified version of SS2Dv2 and is faster. But the training seems unstable when scaled to base model.

YCHYZW commented 1 month ago

thank you

xinpuliu commented 2 weeks ago

may i ask where to switch the version of ss2d?

MzeroMiko commented 1 week ago

with different forward_type, you can use different settings of SS2D forward, including the class type of SS2D (e.g SS2Dv0, SS2Dv2, SS2Dv3)