Hardcoded user-specific absolute paths as dataset defaults

[c-mulliken]

Problem

Several dataset/aggregation defaults are hardcoded to one user's scratch space, so they're wrong for everyone else (and, as of now, wrong even for that user — the paths no longer exist on Oscar).

Confirmed on Oscar: the default root_dir below does not exist (/users/jamullik/scratch/tvsd/data/TVSD is absent), so constructing a dataset with defaults fails.

Offending defaults

• utils/dataset.py — TVSD_BaseDataset.__init__, TVSD_Dataset.__init__, TVSD_TestDataset.__init__:
  root_dir="/users/jamullik/scratch/tvsd/data/TVSD"
• utils/dataset.py — get_things:
  things_path="/users/jamullik/scratch/TVSD-real/data/object_images"
• utils/aggregate.py — __main__:
  results_dir="/users/jamullik/scratch/tvsd-copy/outputs/results/"

Inconsistency

download_tvsd.sh / download_things.sh populate ./data/TVSD and ./data/THINGS (cwd-relative), and benchmark.py / generate_activations.py already default --root_dir/--things_path to f"{os.getcwd()}/data/...". So the dataset-class defaults contradict the rest of the pipeline and the documented download layout.

Proposed fix

• Default root_dir / things_path to the repo-relative ./data/... layout the download scripts create (or None + require explicit paths), matching the argparse defaults.
• Optionally honor an env var (e.g. TVSD_DATA_ROOT) for cluster overrides.
• Remove all /users/jamullik/... absolute paths from committed defaults.

Low risk; no scientific impact. Purely portability/reproducibility.

🤖 Generated with Claude Code