scripts

scripts/

Helper scripts that wrap the build/run/Docker commands documented in the root ../README.md. They are conveniences — every step they run is also documented in ../docs/getting_started.md and ../docs/testing.md.

Script	Purpose
build.sh	Configure + build OpenSplat for a chosen backend (CPU/CUDA/HIP/MPS). Binaries → output/.
fetch_test_data.sh	Download a COLMAP test scene from the project's Hugging Face dataset into data/.
make_chunks.py	Build valid 2/4/8/16/32/64-image COLMAP sub-scenes (pycolmap) for the benchmark ladder.
smoke.sh	End-to-end sanity check: train a few iterations on a dataset and confirm output.
benchmark.sh	Run a scene and capture runtime / peak RAM / backend to a local profiles/ dir.
docker-build.sh	Build a Docker image, selecting a root Dockerfile variant (CUDA / ROCm*).

# Build (Apple Metal), fetch data, smoke-test, then benchmark the chunk ladder:
scripts/build.sh --libtorch ~/libtorch --backend MPS
scripts/fetch_test_data.sh db/drjohnson                 # -> data/db/drjohnson
scripts/smoke.sh data/db/drjohnson 50
scripts/make_chunks.py data/db/drjohnson                # -> data/chunks/drjohnson_{2..64}
for d in data/chunks/drjohnson_{2,4,8,16,32,64}; do scripts/benchmark.sh "$d" 2000; done

Conventions: binaries → output/, trained splats → splat_output/, datasets → data/ (all git-ignored). Pass -o splat_output/<name>.ply when running opensplat directly.

Notes:

• Dockerfiles intentionally live in the repo root (so docker build . works and matches CI .github/workflows/docker.yml); docker-build.sh just selects the -f variant.
• Scripts make no assumptions about installed backends — pass --backend/--libtorch explicitly.