Finalize MCMC strategy and some tiny fix

[KevinXu02]

Finalize MCMC strategy #3436, fix bilagird lr rate for splatfacto-big #3383 and some small changes to colmap dataparser (auto iterates possible colmap paths).
Tested on bicycle with random/sfm initialization.
To use: ns-train splatfacto-mcmc

[pablovela5620]

tried to use this but got a torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.47 GiB. GPU error with mcmc, where I had no problems with vanilla splatfacto

[KevinXu02]

tried to use this but got a torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.47 GiB. GPU error with mcmc, where I had no problems with vanilla splatfacto

Yes this might happen since MCMC uses a fixed number of gs for the scene. Could you please try add--pipeline.model.max_cap [max num of gs] in your training cmd? And a common number would be 1000000 but you can make it smaller.

[pablovela5620]

tried to use this but got a torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.47 GiB. GPU error with mcmc, where I had no problems with vanilla splatfacto

Yes this might happen since MCMC uses a fixed number of gs for the scene. Could you please try add--pipeline.model.max_cap [max num of gs] in your training cmd? And a common number would be 1000000 but you can make it smaller.

Even setting to 500,000 leads to a OOM error sadly (I'm using a 3060), I had to reduce to 100,000 gaussian. That seems like too few(?) but I honestly don't know. I'll do some more testing

[gradeeterna]

I'm also getting OOM errors here with --pipeline.model.max-gs-num 1000000 on a 3090 with 24GB VRAM. Same dataset works fine with splatfacto and MCMC+bilagrid directly in gsplat.

[kerrj]

can we avoid redefining these here so that the defaults use these values?

[kerrj]

that said we also need to update cull_alpha_threshold to .005 here!

[kerrj]

I'm also getting OOM errors here with --pipeline.model.max-gs-num 1000000 on a 3090 with 24GB VRAM. Same dataset works fine with splatfacto and MCMC+bilagrid directly in gsplat.

It's possible there are some memory differences between gsplat and nerfstudio because of dataloader overhead. When you run gsplat is the memory usage very close to 24GB? it may help to make sure the images are cached on CPU inside splatfacto-mcmc (there's a FullImagesDataManager parameter for this). If splatfacto-mcmc is taking significantly more memory than gsplat's version that would be surprising, but I think a difference of ~1GB is expected. MCMC in general will take more memory than the default strategy since the re-sampling step is a little memory hungry.