cmsc714-group-prj-Spring25

Clone original implementation

git clone https://github.com/graphdeco-inria/gaussian-splatting --recursive

Replace files with versions from the "modified" folder

How to set up environment

Check ./setup_env/README.md

How to set up batching and running code

Request a 20 minute bash session with 2 gpus:

srun --partition=dpart --qos=medium --gres=gpu:2 --time 0:20:0 --pty bash
# OR
srun --partition=dpart --qos=normal --gres=gpu:a100:2 --time 0:20:0 --pty bash 

Note:

# zaratan (/.../scratch.cmsc714/) has --qos = normal, scavenger, high, gpu

run the following to train with two gpus for resolution 1/8 files, for 200 iterations:

torchrun --nproc_per_node=2 train.py -s ./bicycle/ --test_iterations 200 --iterations 200 -r 8

How to run profiling:

Files are on Zaratan /scratch/zt1/project/cmsc714/shared/3dgs_proj/gaussian-splatting

Python

Step 1

python -m cProfile -o profiling_results/profile_train.prof train.py -s data/bicycle/ --iterations 200

Step 2:

python pyProfile.py

Original dataset used in paper:

https://storage.googleapis.com/gresearch/refraw360/360_v2.zip

We use images from 360_v2/bicycle/images_8 (resolution 1/8 files)

Cuda

Integrate in train_cudaProfile.py code. Run CMDs as above to train, just replace train.py with train_cudaProfile.py.