Datasets:
Download src/explicit_learning/interventions/control.py from sungguk/visual-answerability: direct link, hf CLI and curl.
- Browser
- Download file 7.49 kB
-
https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/interventions/control.py
- Command line
-
hf download hf://datasets/sungguk/visual-answerability/src/explicit_learning/interventions/control.py
-
curl -L -o control.py https://huggingface.co/datasets/sungguk/visual-answerability/resolve/main/src/explicit_learning/interventions/control.py
7.49 kB
| """``CONTROL_MATCHED_V1`` — a matched non-dependency control node (§10.5). | |
| A control view edits a node the program does **not** depend on, so the formal | |
| answer is unchanged while the image changes in a region matched to the target. | |
| The matched node shares the target's visual type, pixel-area quantile, and | |
| center-position bin, and is edited with the same operator (§10.5). | |
| Pixels do not exist until P6, so area/position are proxied from the semantic | |
| world: *visual type* = :func:`~explicit_learning.sources.world_ops.node_kind`, | |
| *area quantile* = the node's value-magnitude rank among same-kind nodes, and | |
| *position bin* = its index rank in world order. P6/P9 replace these proxies | |
| with real pixel audits; the selection contract is unchanged. | |
| When no matched non-dependency node exists, the spec lets the caller either form | |
| a control-less group or reject. :func:`find_control_match` returns ``None`` so | |
| the group builder (P11) can choose; :meth:`ControlMatchedOperator.apply` | |
| rejects via :class:`InterventionError` for callers that want a hard fail. | |
| """ | |
| from __future__ import annotations | |
| import hashlib | |
| from typing import Any | |
| from ..dsl.ast import Dsl | |
| from ..sources.base import World | |
| from ..sources.world_ops import ( | |
| find_point, | |
| find_row, | |
| find_series, | |
| is_geometry, | |
| is_plot, | |
| is_table, | |
| node_kind, | |
| world_sha256, | |
| ) | |
| from .base import ( | |
| InterventionError, | |
| OperatorName, | |
| TransformedWorld, | |
| transformed_hash, | |
| ) | |
| _NAME: OperatorName = "CONTROL_MATCHED_V1" | |
| _BINS = 4 | |
| class ControlMatchedOperator: | |
| """Apply the matched edit to a non-dependency control node.""" | |
| name = _NAME | |
| def apply( | |
| self, | |
| world: World, | |
| target_node: str, | |
| *, | |
| seed: int, | |
| dsl: Dsl, | |
| dependency_node_ids: tuple[str, ...] | frozenset[str] = (), | |
| edit_operator: OperatorName = "REDACT_SOLID_V1", | |
| **kwargs: Any, | |
| ) -> TransformedWorld: | |
| control = find_control_match( | |
| world, | |
| target_node, | |
| frozenset(dependency_node_ids), | |
| edit_operator=edit_operator, | |
| seed=seed, | |
| ) | |
| if control is None: | |
| raise InterventionError( | |
| f"CONTROL_MATCHED_V1: no matched non-dependency node for " | |
| f"{target_node!r} ({edit_operator})" | |
| ) | |
| # Apply the matched edit to the control node. The transformation is | |
| # labelled CONTROL_MATCHED_V1; the underlying edit is in the description. | |
| from . import get_operator # local: the registry is finalized at import | |
| inner = get_operator(edit_operator).apply(world, control, seed=seed, dsl=dsl, **kwargs) | |
| return TransformedWorld( | |
| world=inner.world, | |
| operator=_NAME, | |
| target_node=control, | |
| seed=seed, | |
| transformed_world_sha256=transformed_hash( | |
| operator=_NAME, | |
| target_node=control, | |
| seed=seed, | |
| source_world_sha256=world_sha256(world), | |
| transformed_world_sha256=world_sha256(inner.world), | |
| hidden_nodes=inner.hidden_nodes, | |
| substitute_value=inner.substitute_value, | |
| ), | |
| hidden_nodes=inner.hidden_nodes, | |
| substitute_value=inner.substitute_value, | |
| description=f"control-matched {edit_operator} on {control} (for {target_node})", | |
| ) | |
| def find_control_match( | |
| world: World, | |
| target_node: str, | |
| dependency_node_ids: frozenset[str], | |
| *, | |
| edit_operator: OperatorName, | |
| seed: int, | |
| ) -> str | None: | |
| """The best same-kind, non-dependency match for ``target_node`` (or ``None``). | |
| Ranks every same-kind node by area (value-magnitude) and position (world | |
| index) into ``_BINS`` quantile buckets and prefers a candidate sharing both | |
| of the target's buckets; ties break by seeded hash for determinism. The | |
| target itself and every dependency node are excluded. | |
| """ | |
| target_kind = node_kind(world, target_node) | |
| if target_kind is None: | |
| return None | |
| same_kind = [nid for nid in _all_nodes(world) if node_kind(world, nid) == target_kind] | |
| if not same_kind: | |
| return None | |
| area_rank = _ranks(same_kind, lambda nid: _area_metric(world, nid, target_kind)) | |
| pos_rank = _ranks(same_kind, lambda nid: float(same_kind.index(nid))) | |
| t_area = _bucket(area_rank[target_node]) | |
| t_pos = _bucket(pos_rank[target_node]) | |
| candidates = [nid for nid in same_kind if nid != target_node and nid not in dependency_node_ids] | |
| if not candidates: | |
| return None | |
| def _score(nid: str) -> tuple[int, int, int]: | |
| # (area-bucket match, position-bucket match, seeded tie-break) — lower is | |
| # better; the seeded hash keeps selection deterministic. | |
| a = 0 if _bucket(area_rank[nid]) == t_area else 1 | |
| p = 0 if _bucket(pos_rank[nid]) == t_pos else 1 | |
| h = int.from_bytes(hashlib.sha256(f"{seed}|control|{nid}".encode()).digest()[:4], "big") | |
| return (a, p, h) | |
| candidates.sort(key=_score) | |
| return candidates[0] | |
| # --- node enumeration + metrics ------------------------------------------- | |
| def _all_nodes(world: World) -> list[str]: | |
| out: list[str] = [] | |
| if is_plot(world): | |
| for s in world.get("series", []) or []: | |
| if s.get("id"): | |
| out.append(str(s["id"])) | |
| for p in s.get("points", []) or []: | |
| if p.get("id"): | |
| out.append(str(p["id"])) | |
| if is_table(world): | |
| for r in world.get("rows", []) or []: | |
| if r.get("id"): | |
| out.append(str(r["id"])) | |
| if is_geometry(world): | |
| for e in world.get("entities", []) or []: | |
| if e.get("id"): | |
| out.append(str(e["id"])) | |
| for c in world.get("constraints", []) or []: | |
| for a in c.get("args") or []: | |
| if isinstance(a, str) and not _is_numeric(a): | |
| out.append(str(a)) | |
| # de-dup, preserve order | |
| seen: set[str] = set() | |
| uniq: list[str] = [] | |
| for nid in out: | |
| if nid not in seen: | |
| seen.add(nid) | |
| uniq.append(nid) | |
| return uniq | |
| def _area_metric(world: World, nid: str, kind: str) -> float: | |
| if kind == "point": | |
| point = find_point(world, nid) | |
| if point is not None: | |
| _s, _p, p = point | |
| try: | |
| return abs(float(p.get("y", 0))) | |
| except (TypeError, ValueError): | |
| return 0.0 | |
| if kind == "series": | |
| series = find_series(world, nid) | |
| if series is not None: | |
| _i, s = series | |
| return float(len(s.get("points", []) or [])) | |
| if kind == "row": | |
| row = find_row(world, nid) | |
| if row is not None: | |
| _i, r = row | |
| return float(len(r.get("cells", {}) or {})) | |
| return 0.0 | |
| def _ranks(items: list[str], metric: Any) -> dict[str, float]: | |
| """Normalized ``[0, 1]`` rank of each item by ``metric`` (ties average).""" | |
| if len(items) <= 1: | |
| return {items[0]: 0.0} if items else {} | |
| keyed = sorted(items, key=metric) | |
| return {nid: idx / (len(keyed) - 1) for idx, nid in enumerate(keyed)} | |
| def _bucket(rank: float) -> int: | |
| return min(_BINS - 1, int(rank * _BINS)) | |
| def _is_numeric(text: str) -> bool: | |
| try: | |
| float(text) | |
| except (TypeError, ValueError): | |
| return False | |
| return True | |
| __all__ = ["ControlMatchedOperator", "find_control_match"] | |