cac — Active Hypothesis Exploration for Crowd Counting

HypoExplore (arXiv:2604.12999) applied to FSC147 class-agnostic counting. The trajectory tree IS a directory tree: every branch in tree/ (physical nesting) is the lineage of one architecture family, rooted at the certified champion N0001_champion (MAE 19.647 / RMSE 74.05 / 31.32M).

tree/ (topology + info.json) and memory/hypotheses.jsonl (append-only ledger) are the only ground truth; src/cac/expt/ is the paper's decision rules implemented as code. Read AGENTS.md first.

Quickstart (local, no GPU needed)

pip install -e .           # python 3.13
python scripts/conformance.py          # must print CONFORMANCE OK
python scripts/discovery.py tree       # ASCII tree
python scripts/discovery.py validate   # format-gate hypothesis ledger
python tests/test_all.py 2>/dev/null || python -m pytest tests

One evolution step (each new branch is one iteration)

python scripts/discovery.py parent                 # Eq.2–4 parent
python scripts/discovery.py hypo <parent>          # Eq.5–6 child + hypothesis set
python scripts/novelty_check.py "<new hypothesis>" # novelty gate
... edit tree/<child>/config.toml + model.py (delta from parent only) ...
python scripts/conformance.py                      # green before commit
# on cac-server:
python scripts/run_node.py <child>                 # smoke → budget (τ_max 1800s)
python scripts/discovery.py calibration            # bin table

Layout

• src/cac/expt/ — constants, node (tree), hypothesis (ledger+Eq.1), select (Eq.2–6), gates (calibration + conformance).
• src/cac/engine|models|data|calls/ — runner (smoke→budget, checksums), champion counter, Lightning pl_module, FSC147 datamodule, best/EMA callbacks.
• scripts/ — discovery, conformance, novelty_check, run_node, install_key.
• tree/ — trajectory tree; memory/ — ledger + index; journal/ — events; docs/research_direction.md — mission (changes only there, journaled).

Reproducibility

A node reruns bit-identical under its config.toml seed (training seeds torch/numpy/random, seeded loader + worker_init). Every result.json records config_sha256/model_sha256.