search

BaowenZ / RaDe-GS

RaDe-GS: Rasterizing Depth in Gaussian Splatting
Other
460 stars 25 forks source link

Splitting Data to RAM instead of GPU #23

Open Fanie-Visagie opened 3 months ago

Fanie-Visagie commented 3 months ago

As per 2DGS, I get an error saying the data is not all loaded onto then GPU when running the Command:

python train.py -s Processing --data_device cpu

Any reason why this is ??

It is of utmost importance for optimal processing...

Mark

BaowenZ commented 3 months ago

Hi! I treat all variables as if they were on GPU when writing train.py. With '--data_device cpu', the variables are saved on device CPU. And you could transform variables to GPU in train.py L130~131:

gt_image = viewpoint_cam.original_image.cuda()
edge = viewpoint_cam.edge.cuda()

and L261

gt_image = torch.clamp(viewpoint.original_image.cuda(), 0.0, 1.0)