LearnerPR

VPR student: DINOv2-Small + GeM, trained with RKD from Qwen3-VL-8B-Instruct (see PLAN.md).

• Course submission: root model.py implements the COMP 560 StudentModel API (device, encode, embedding_dim) for the provided evaluate.py. Place trained weights as submission_weights.pt or set LEARNERPR_WEIGHTS.
• Training / MSLS: src/train.py, Longleaf scripts under scripts/.
• Training / GSV-Cities: see below.
• COMP 560 alignment (metrics, datasets, integrity): PLAN.md §11.

Longleaf (UNC) Slurm

Batch jobs source scripts/slurm_longleaf_init.sh, which: loads the Anaconda module, runs conda activate "${CONDA_ENV_NAME:-learnerpr}", exports PYTHON as ${CONDA_PREFIX}/bin/python, and verifies import torch (skipped for the GSV download job via LEARNERPR_SKIP_TORCH_CHECK=1). All *.sl scripts invoke "${PYTHON}" src/... so Slurm never falls back to system /usr/bin/python.

Repo path: Each *.sl sets LEARNERPR_REPO (default /users/a/l/alshen/LearnerPR/LearnerPR) and sources scripts/slurm_longleaf_init.sh from there, then cds into that directory so python src/... works even when Slurm copies the batch script to /var/spool/... or you sbatch from a parent folder. Override if your clone lives elsewhere: LEARNERPR_REPO=/path/to/inner/LearnerPR sbatch scripts/train.sl.

Optional environment overrides: ANACONDA_MODULE (default anaconda/2024.02), CONDA_ENV_NAME (default learnerpr). Install deps in that env (pip install -r requirements.txt) so the torch check passes.

GSV-Cities setup

GSV-Cities (~530k images, ~62k places across 40+ cities) can be added as a second training source alongside MSLS.

Prerequisites:

1. A Kaggle account with an API token saved to ~/.kaggle/kaggle.json (chmod 600). Accept the dataset license at https://www.kaggle.com/datasets/amaralibey/gsv-cities.
2. ~60 GB free on your work filesystem.

Download (once):

sbatch scripts/download_gsv_cities.sl

Expected layout after unzip:

gsv-cities/
    Images/<City>/*.JPG        # ~530k images
    Dataframes/<City>.csv      # place_id metadata per city

Cache teacher embeddings for GSV-Cities (once):

sbatch scripts/cache_teacher_gsv.sl

Joint training (MSLS + GSV-Cities):

Edit configs/default.yaml:

training:
  train_datasets: [msls, gsv_cities]

Then run training as usual:

sbatch scripts/train.sl

GSV-Cities eval (held-out cities):

sbatch scripts/eval_gsv.sl