git clone https://github.com/noahmeurer/monocular-depth-estimation.git ~/monocular-depth-estimation
cd ~/monocular-depth-estimation

Depth Anything 3 is installed automatically by uv sync (git dependency in pyproject.toml).

Uses uv with CUDA 13.0 PyTorch wheels from pyproject.toml.

Copy .env.example to .env and fill in your username under the /work/scratch/<your_username>/... paths and your HF_TOKEN:

cp .env.example .env
# edit HF_TOKEN, HF_HOME, XDG_CACHE_HOME
source .env

On x86 GPU nodes, load CUDA before uv sync:

module add cuda/13.0

Keep uv’s cache on scratch (20GB home quota):

export UV_CACHE_DIR=/work/scratch/<your_username>/uv-cache

On the login node or any x86 GPU node:

cd ~/monocular-depth-estimation
module add cuda/13.0
uv sync
source .venv/bin/activate

GB10 is aarch64 — use .venv-gb10, not .venv. Interactive setup and drop-caches: CLUSTER.md → GB10 (interactive).

python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

Train: /cluster/courses/cil/monocular-depth-estimation/train

Test: /cluster/courses/cil/monocular-depth-estimation/test

Interactive session (x86 GPU):

srun --pty --gpus 5060ti:1 -A cil_jobs -t 120 bash --login

Batch job:

sbatch scripts/baseline1.slurm

Slurm accounts, storage, Jupyter: CLUSTER.md.

Scratch dirs are world-readable on the cluster, so the simplest way to share between teammates is to rsync/cp straight from another user's /work/scratch/<user>/.... See SCRATCH.md for the index of known artifacts (baseline outputs, pseudo-labels, etc.).

For sharing off-cluster (e.g. to a teammate without cluster access, or as a durable backup), scripts/hf_scratch_sync.py pushes/pulls scratch paths to a private Hub repo (HF_SCRATCH_REPO_ID in .env):

# upload outputs/baseline1 from your scratch to the team's private HF repo
python scripts/hf_scratch_sync.py push outputs/baseline1

# pull it back on another machine into the same scratch-relative path
python scripts/hf_scratch_sync.py pull outputs/baseline1

# list what's on the Hub
python scripts/hf_scratch_sync.py list outputs/

For folders with thousands of files (e.g. pseudo-labels), use --chunk-size 1000 --sleep-between-chunks 30 to stay under HF's commit-rate limits. --scratch-root /work/scratch/<other_user> lets you push someone else's tree without touching your own.