I finished the installation, and wanted to try Training a CLIP-Fields (https://github.com/notmahi/clip-fields) directly by doing:
python train.py dataset_path=nyu.r3d
I amb monitoring both RAM and GPUram. I see that when the code starts, data starts to load on RAM:
Loading data: 100%|████████████████████████████████████████████████████████| 757/757 [00:05<00:00, 135.05it/s]
Upscaling depth and conf: 100%|████████████████████████████████████████████| 757/757 [00:04<00:00, 157.72it/s]
Calculating global XYZs: 100%|██████████████████████████████████████████████| 757/757 [00:14<00:00, 51.72it/s]
The previous code ocupies about 30GB of RAM. Then models such as Detic load on GPU, but when I arrive to line 177, /dataloaders/real_dataset.py my PC kills the process because of RAM OOM:
# First, setup detic with the combined classes.
self._setup_detic_all_classes(view_data)
Why is data load on ram and not on GPU? Is there any way to lower the GB of memory used?
