Few-shot views

Extension, not part of the paper

This page describes tooling for studying how fine-tuning behaves with less labelled data. The paper reports no few-shot experiments, and this page shows no results.

A few-shot view is a copy of a dataset's training split that keeps only a fraction of the training images; the test split stays unchanged. Pair mining, fine-tuning and evaluation then run on that view exactly as on the full one (mining, fine-tuning, evaluation), with one extra option each.

How a view is built

Rule	Effect
Exact budget	the training split keeps round(fraction × N_train) images in total
Proportional	the budget is shared across individuals in proportion to their image counts
Minimum of two	an individual never drops below two training images, so every remaining image still has a positive; the other individuals share what is left of the budget
Nested	the images kept for an individual are a prefix of one seeded permutation, so for one seed the 1/8 view lies inside the 1/4 view, which lies inside the 1/2 view
Same frame names	images keep their names from the full view, so the full view's feature caches serve every few-shot view

Individuals with a single image keep it: they never yield a training pair, but they stay in the gallery as in the full view. When the two-image minimum alone exceeds the budget, the view keeps more than the requested fraction. Datasets with many small individuals reach this limit early; SalamanderID2025, for example, keeps 70 % of its training images for any fraction up to 1/2. The view's fewshot.json records the effective fraction and budget_feasible.

CzechLynx few-shot views exist for the time-closed split only.

Mining

The full view and its feature cache must exist first (see Pair mining). --fraction and --seed then select the few-shot view in every wildmatch mine step:

wildmatch mine plan   --dataset hyenaid2022 --backend loma --fraction 0.25 --seed 0
wildmatch mine view   --dataset hyenaid2022 --backend loma --fraction 0.25 --seed 0
wildmatch mine submit --dataset hyenaid2022 --backend loma --fraction 0.25 --seed 0

Views, indices, checkpoints and metadata go to their own folders under the path profile's fewshot_root (by default <data_root>/fewshot, or WILDMATCH_FEWSHOT_ROOT): {views,indices,checkpoints,metadata}/<dataset>/<protocol>/frac<F>-seed<S>/.

Fine-tuning

sbatch slurm/finetune_matcher.sbatch dataset=hyenaid2022 matcher_finetune=loma \
    matcher_finetune.fewshot.fraction=0.25 matcher_finetune.fewshot.seed=0

The run reads the few-shot view and its mined index and writes to <fewshot_root>/checkpoints/<dataset>/<protocol>/frac<F>-seed<S>/<matcher>-finetuned/. The recipe is the paper's (see Matcher fine-tuning).

Evaluation

Building a view also writes a copy of the dataset's metadata with one extra column, split_frac<F>_seed<S>: train for the kept training images, unused for the dropped ones, and test unchanged. Evaluating with that column puts only the kept images in the gallery, which models a smaller labelled collection:

wildmatch evaluate dataset=hyenaid2022 \
  dataset.metadata_file=<fewshot_root>/metadata/HyenaID2022/legacy/frac0.25-seed0/metadata.csv \
  dataset.split_col=split_frac0.25_seed0 \
  dataset.database_split_value=train dataset.query_split_value=test \
  benchmark.method=vismatch benchmark.methods.vismatch.matcher=loma \
  benchmark.methods.vismatch.checkpoint_source=custom \
  benchmark.methods.vismatch.checkpoint_path=<fewshot_root>/checkpoints/HyenaID2022/legacy/frac0.25-seed0/loma-finetuned/epoch_299/model.safetensors \
  benchmark.methods.vismatch.checkpoint_components=matcher_only \
  benchmark.candidate_k=250

The metadata path makes the run's cache and run identity separate from the full dataset's.