sungguk's picture
Release visual answerability benchmark v1.0.0
e1ced61 verified
Raw
History Blame Contribute Delete
13.9 kB
"""Deterministic executor for official CLEVR functional programs."""
from __future__ import annotations
from collections.abc import Mapping, Sequence
from typing import Any
from ..dsl.ast import Arg, AstNode, Program
from ..sources.base import World
from .base import (
AnswerValue,
ExecutionResult,
ExecutionStatus,
ExecutorError,
_InvalidReferent,
_MissingInformation,
_Trace,
literal_value,
serialize_answer,
)
NAME = "clevr_executor"
VERSION = "clevr_dsl_v1/1"
_ATTRIBUTES = frozenset({"color", "material", "shape", "size"})
_RELATIONS = frozenset({"left", "right", "front", "behind"})
class ClevrExecutor:
"""Evaluate a CLEVR AST over a normalized ``clevr_world_v1`` scene."""
name = NAME
version = VERSION
def execute(self, program: Program, *, world: World) -> ExecutionResult:
trace = _Trace()
referent_cardinality: int | None = None
status: ExecutionStatus
try:
if program.dsl != "clevr_dsl_v1" or program.compile_status != "compiled":
raise ExecutorError("executor requires a compiled clevr_dsl_v1 program")
if world.get("world_schema") != "clevr_world_v1":
raise ExecutorError(
f"clevr executor cannot read world_schema {world.get('world_schema')!r}"
)
referent_cardinality = self._check_referent(program, world, trace)
value = self._eval(program.program, world, trace)
answer = self._project_answer(program.program, value)
_steps, dependencies, trace_hash = trace.finish()
return ExecutionResult(
status="UNIQUE",
answer_value=answer,
answer_canonical=serialize_answer(
answer,
return_type=program.program.return_type,
),
dependency_node_ids=dependencies,
referent_cardinality=referent_cardinality,
trace_sha256=trace_hash,
executor_name=self.name,
executor_version=self.version,
)
except _MissingInformation:
status = "MISSING_INFORMATION"
except _InvalidReferent:
status = "INVALID_REFERENT"
except ExecutorError:
status = "ERROR"
_steps, dependencies, trace_hash = trace.finish()
return ExecutionResult(
status=status,
answer_value=None,
answer_canonical=None,
dependency_node_ids=dependencies,
referent_cardinality=(
referent_cardinality
if referent_cardinality is not None
else self._safe_referent_cardinality(program, world)
),
trace_sha256=trace_hash,
executor_name=self.name,
executor_version=self.version,
)
def execute_selector(self, selector: AstNode, *, world: World) -> tuple[str, ...]:
"""Resolve an object/object-set selector without applying ``UNIQUE``."""
trace = _Trace()
if selector.op == "UNIQUE":
if len(selector.args) != 1 or not isinstance(selector.args[0], AstNode):
raise ExecutorError("UNIQUE selector is malformed")
value = self._eval(selector.args[0], world, trace)
else:
value = self._eval(selector, world, trace)
if isinstance(value, Mapping):
node_id = value.get("id")
return (str(node_id),) if node_id else ()
if isinstance(value, tuple):
return tuple(
str(item["id"]) for item in value if isinstance(item, Mapping) and item.get("id")
)
raise ExecutorError(f"selector {selector.op!r} did not produce objects")
def _check_referent(self, program: Program, world: World, trace: _Trace) -> int | None:
selector = program.referent_selector
required = program.required_referent_cardinality
if selector is None:
return None
refs = self.execute_selector(selector, world=world)
trace.record("REFERENT", refs, len(refs))
if required is not None and len(refs) != required:
raise _InvalidReferent(
f"scalar referent cardinality {len(refs)} != required {required}"
)
return len(refs)
def _safe_referent_cardinality(self, program: Program, world: World) -> int | None:
if program.referent_selector is None:
return None
try:
return len(self.execute_selector(program.referent_selector, world=world))
except (ExecutorError, _MissingInformation, _InvalidReferent):
return None
def _eval(self, node: Arg, world: World, trace: _Trace) -> Any:
if not isinstance(node, AstNode):
raise ExecutorError(f"expected AST node, got {type(node).__name__}")
op = node.op
if op == "LITERAL":
return literal_value(node)
if op == "SCENE":
objects = _objects(world)
trace.record("SCENE", (), len(objects))
return objects
if op.startswith("FILTER_"):
return self._filter(op.removeprefix("FILTER_").lower(), node, world, trace)
if op == "UNIQUE":
unique_values = _object_set(self._eval_arg(node, 0, world, trace), op)
trace.record(op, unique_values, len(unique_values))
if len(unique_values) != 1:
raise _InvalidReferent(f"UNIQUE received {len(unique_values)} objects")
return unique_values[0]
if op == "RELATE":
source = _object(self._eval_arg(node, 0, world, trace), op)
relation = _string(self._eval_arg(node, 1, world, trace), op)
related_values = _related(world, str(source["id"]), relation)
trace.record(op, (source, *related_values), len(related_values))
return related_values
if op.startswith("SAME_"):
attribute = op.removeprefix("SAME_").lower()
source = _object(self._eval_arg(node, 0, world, trace), op)
target = _attribute(source, attribute)
same_values: list[Mapping[str, Any]] = []
for candidate in _objects(world):
if candidate.get("id") == source.get("id"):
continue
if _attribute(candidate, attribute) == target:
same_values.append(candidate)
same_result = _ordered(same_values)
trace.record(op, (source, *same_result), len(same_result))
return same_result
if op in {"UNION", "INTERSECT"}:
left = _object_set(self._eval_arg(node, 0, world, trace), op)
right = _object_set(self._eval_arg(node, 1, world, trace), op)
left_by_id = {str(item["id"]): item for item in left}
right_by_id = {str(item["id"]): item for item in right}
ids = (
set(left_by_id).union(right_by_id)
if op == "UNION"
else set(left_by_id).intersection(right_by_id)
)
set_result = tuple((left_by_id | right_by_id)[node_id] for node_id in sorted(ids))
trace.record(op, set_result, len(set_result))
return set_result
if op == "COUNT":
count_values = _object_set(self._eval_arg(node, 0, world, trace), op)
trace.record(op, count_values, len(count_values))
return len(count_values)
if op == "EXIST":
exist_values = _object_set(self._eval_arg(node, 0, world, trace), op)
exists = bool(exist_values)
trace.record(op, exist_values, exists)
return exists
if op.startswith("QUERY_"):
attribute = op.removeprefix("QUERY_").lower()
obj = _object(self._eval_arg(node, 0, world, trace), op)
value = _attribute(obj, attribute)
trace.record(op, obj, value)
return value
if op.startswith("EQUAL_"):
left = self._eval_arg(node, 0, world, trace)
right = self._eval_arg(node, 1, world, trace)
comparison = _comparable(left) == _comparable(right)
trace.record(op, (), comparison)
return comparison
if op in {"LESS_THAN", "GREATER_THAN"}:
left_integer = _integer(self._eval_arg(node, 0, world, trace), op)
right_integer = _integer(self._eval_arg(node, 1, world, trace), op)
comparison = (
left_integer < right_integer if op == "LESS_THAN" else left_integer > right_integer
)
trace.record(op, (), comparison)
return comparison
raise ExecutorError(f"unsupported CLEVR op {op!r}")
def _eval_arg(
self,
node: AstNode,
position: int,
world: World,
trace: _Trace,
) -> Any:
try:
arg = node.args[position]
except IndexError as exc:
raise ExecutorError(f"{node.op} has no operand {position}") from exc
if not isinstance(arg, AstNode):
raise ExecutorError(f"{node.op} operand {position} is not an AST node")
return self._eval(arg, world, trace)
def _filter(
self,
attribute: str,
node: AstNode,
world: World,
trace: _Trace,
) -> tuple[Mapping[str, Any], ...]:
values = _object_set(self._eval_arg(node, 0, world, trace), node.op)
wanted = _string(self._eval_arg(node, 1, world, trace), node.op)
selected = tuple(item for item in values if _attribute(item, attribute) == wanted)
trace.record(node.op, selected, len(selected))
return selected
def _project_answer(self, root: AstNode, value: Any) -> AnswerValue:
if root.return_type == "integer":
return _integer(value, root.op)
if root.return_type == "boolean":
return bool(value)
if root.return_type == "string":
return _string(value, root.op)
raise ExecutorError(
f"CLEVR root {root.op!r} has non-answer return type {root.return_type!r}"
)
def _objects(world: World) -> tuple[Mapping[str, Any], ...]:
raw = world.get("objects")
if not isinstance(raw, list):
raise ExecutorError("CLEVR world objects is not an array")
values: list[Mapping[str, Any]] = []
seen: set[str] = set()
for item in raw:
if not isinstance(item, Mapping) or not item.get("id"):
raise ExecutorError("CLEVR world contains an object without an id")
node_id = str(item["id"])
if node_id in seen:
raise ExecutorError(f"CLEVR world has duplicate object id {node_id!r}")
seen.add(node_id)
values.append(item)
return _ordered(values)
def _ordered(values: Sequence[Mapping[str, Any]]) -> tuple[Mapping[str, Any], ...]:
return tuple(sorted(values, key=lambda item: str(item.get("id", ""))))
def _object_set(value: Any, op: str) -> tuple[Mapping[str, Any], ...]:
if not isinstance(value, tuple) or any(not isinstance(item, Mapping) for item in value):
raise ExecutorError(f"{op} expected an object set")
return _ordered(value)
def _object(value: Any, op: str) -> Mapping[str, Any]:
if not isinstance(value, Mapping) or not value.get("id"):
raise ExecutorError(f"{op} expected one object")
return value
def _attribute(obj: Mapping[str, Any], attribute: str) -> str:
if attribute not in _ATTRIBUTES:
raise ExecutorError(f"unknown CLEVR attribute {attribute!r}")
value = obj.get(attribute)
if value is None:
raise _MissingInformation(f"object {obj.get('id')} {attribute} is hidden")
if not isinstance(value, str) or not value:
raise ExecutorError(f"object {obj.get('id')} has invalid {attribute}")
return value
def _related(
world: World,
source_id: str,
relation: str,
) -> tuple[Mapping[str, Any], ...]:
if relation not in _RELATIONS:
raise ExecutorError(f"unknown CLEVR relation {relation!r}")
relationships = world.get("relationships")
if not isinstance(relationships, Mapping):
raise ExecutorError("CLEVR relationships is not an object")
relation_map = relationships.get(relation)
if not isinstance(relation_map, Mapping):
raise ExecutorError(f"CLEVR relationship {relation!r} is missing")
raw_targets = relation_map.get(source_id, [])
if not isinstance(raw_targets, list):
raise ExecutorError(f"CLEVR relationship {relation!r}/{source_id!r} is malformed")
by_id = {str(item["id"]): item for item in _objects(world)}
targets: list[Mapping[str, Any]] = []
for node_id in raw_targets:
target = by_id.get(str(node_id))
if target is None:
raise ExecutorError(
f"CLEVR relationship {relation!r} references absent object {node_id!r}"
)
targets.append(target)
return _ordered(targets)
def _string(value: Any, op: str) -> str:
if not isinstance(value, str):
raise ExecutorError(f"{op} expected a string, got {type(value).__name__}")
return value
def _integer(value: Any, op: str) -> int:
if isinstance(value, bool) or not isinstance(value, int):
raise ExecutorError(f"{op} expected an integer, got {type(value).__name__}")
return int(value)
def _comparable(value: Any) -> Any:
if isinstance(value, Mapping):
node_id = value.get("id")
if not node_id:
raise ExecutorError("object comparison received an object without an id")
return ("object", str(node_id))
if isinstance(value, (str, int, bool)):
return value
raise ExecutorError(f"value {value!r} is not comparable")
__all__ = ["ClevrExecutor"]