Any plan to integrate MCMC strategy into NeRFStudio?

[yt2639]

Hii, is there any plan to integrate MCMC strategy into NeRFStudio's splatfacto? I've tried myself but it seems not working. I essentially add MCMCStrategy.step_post_backward link in trainer.py between self.grad_scaler.scale(loss).backward() link and self.optimizers.optimizer_scaler_step_some(self.grad_scaler, needs_step) link. Additionally, I commented out the after_train and refinement_after operations.

However, the results are pretty weird. I understand MCMC is working well with gsplat code base so I am wondering if there is any plan to integrate MCMC into nerfstudio's code base?

Thanks!

[brentyi]

This PR from @jb-ye is relevant: nerfstudio-project/nerfstudio#3376
@kerrj also has a similar branch here for moving to the strategy API: https://github.com/nerfstudio-project/nerfstudio/tree/splatfacto-refactor

My impression of the likely outcome is that MCMC will be added as an option once we move over to gsplat's strategy API, but it won't be enabled by default.