coredd-bench / src /coredd /benchmark.py
witrin's picture
Pool metrics by default and derive score and interval from the MC draws
7deba50 verified
Raw
History Blame Contribute Delete
8.61 kB
"""The public benchmark API.
``Benchmark.load`` reads the released labels and the noise model; ``evaluate`` scores a
candidate's predictions and returns a point score with a label-uncertainty interval.
A candidate predicts **lines** (``Line(path, line, side, score)``). CoReDD lifts each
line to the syntax node that carries it -- the same projection that built the labels --
and scores on **nodes**: the released defective nodes are the truth, the predicted-onto
nodes are the prediction. The uncertainty comes from perturbing the truth (see
``perturbation``).
"""
from __future__ import annotations
import json
from collections.abc import Iterable, Mapping
from dataclasses import dataclass
from pathlib import Path
import numpy as np
from sklearn.metrics import f1_score
from coredd.noise import NoiseModel
from coredd.perturbation import Change, interval
from coredd.history import RepositorySet
@dataclass(frozen=True)
class Line:
"""A predicted line: a file, a line number, its diff side, and a score.
``side`` is ``"after"`` for an added/context line or ``"base"`` for a deleted line.
``score`` is the prediction's confidence, used by rank metrics; it defaults to 1.0
so a bare hit still ranks above the unpredicted nodes (score 0).
"""
path: str
line: int
side: str
score: float = 1.0
@dataclass(frozen=True)
class Result:
"""A benchmark score with its label-uncertainty band."""
score: float
interval: tuple[float, float]
draws: int
level: float = 0.95
def __repr__(self) -> str:
pct = int(round(self.level * 100))
return (
f"Score: {self.score:.2f}\n"
f"{pct}% label-uncertainty interval: "
f"[{self.interval[0]:.2f}, {self.interval[1]:.2f}]\n"
f"Monte Carlo draws: {self.draws:,}"
)
class Benchmark:
"""The released benchmark: labels, noise model, and repository clones."""
def __init__(self, records: list[dict], repositories: RepositorySet,
noise_path: Path) -> None:
self._records = records
self._repositories = repositories
self._noise_path = noise_path
self._by_key = {
self.key(record["repository"], record["pr"]): record for record in records
}
@classmethod
def load(cls, repositories: str | Path, *, labels: str | Path,
noise: str | Path) -> "Benchmark":
"""Load the benchmark against local clones under *repositories*.
``labels`` (the released ``labels.jsonl``) and ``noise`` (the ``noise.json``
noise model) are the benchmark's released data files and are supplied by the
caller; both are required.
"""
records = _read_labels(labels)
return cls(records, RepositorySet(repositories), Path(noise))
@staticmethod
def key(repository: str, pr: int, base: str | None = None,
after: str | None = None) -> str:
"""The prediction key for a change.
``owner/repo#pr`` maps the prediction to a change; the optional
``@base..after`` suffix names the diff the prediction is made against, and both
SHAs flow through to the node match. Omit them to build the mapping key used to
index a change; supply them (a released node's ``base``/``after``) to predict.
"""
change = f"{repository}#{pr}"
if base is not None and after is not None:
return f"{change}@{base}..{after}"
return change
def evaluate(self, predictions: Mapping[str, Iterable[Line]], *,
metric=f1_score, pooled: bool = True,
draws: int = 10_000, seed: int | None = 0) -> Result:
"""Score *predictions* and return the point score with its uncertainty band.
``predictions`` maps a change key (:meth:`key`, ``owner/repo#pr``) to its
predicted lines; a change absent from the map is scored with no prediction.
Score and interval are the mean and the [2.5%, 97.5%] quantile of the metric
over the Monte Carlo draws that perturb the labels at the audited rates.
Scoring is over the candidate syntax nodes N(c) of each change. With ``pooled``
(the default) the metric runs once over the nodes of every change pooled into
one population; with ``pooled=False`` it runs per change and is averaged
equally across changes, dropping changes where it is undefined (NaN) -- the
form per-case metrics such as :func:`coredd.metrics.mrr` are reported in.
"""
prepared = _normalise(predictions)
empty = _Prediction(base=None, after=None, lines=[])
noise = NoiseModel.load(self._noise_path)
changes = [
self._change(record, prepared.get(key, empty))
for key, record in self._by_key.items()
]
rng = np.random.default_rng(seed)
score, band = interval(changes, metric, noise, draws=draws, pooled=pooled,
rng=rng)
return Result(score=score, interval=band, draws=draws)
def _change(self, record: dict, prediction: "_Prediction") -> Change:
nodes = record["nodes"]
projection = self._repositories.projection(record["repository"])
predicted = self._predict(projection, prediction)
return self._node_change(nodes, predicted)
def _predict(self, projection, prediction: "_Prediction") -> dict:
# A prediction carries the (base, after) diff it is made against, in the key
# ``owner/repo#pr@base..after``. Lines are lifted against that after commit, and
# the resulting node keeps the (base, after) so it matches a label node only when
# both the span and the diff agree -- a prediction against a different diff hits
# nothing. Lines that lift to no node (blank/comment/non-Python) drop out.
scores: dict[tuple, float] = {}
if prediction.after is None:
return scores
for line in prediction.lines:
# A prediction may name a commit the local clone does not have (a wrong or
# unknown diff); that lifts to nothing rather than raising -- it simply hits
# no label node.
try:
node = projection.project(
prediction.after, line.path, line.line, line.side
)
except (ValueError, KeyError):
continue
if node is None:
continue
key = (node.path, node.start, node.end, prediction.base, prediction.after)
scores[key] = max(scores.get(key, float("-inf")), line.score)
return scores
def _node_change(self, nodes: list[dict], predicted: dict) -> Change:
y_true, y_pred, scores, bins = [], [], [], []
for node in nodes:
key = (
node["path"], tuple(node["start"]), tuple(node["end"]),
node["base"], node["after"],
)
hit = key in predicted
y_true.append("provenance" in node)
y_pred.append(hit)
scores.append(predicted[key] if hit else 0.0)
bins.append(int(node["bin"]))
return Change(
y_true=np.asarray(y_true, dtype=bool),
y_pred=np.asarray(y_pred, dtype=bool),
scores=np.asarray(scores, dtype=float),
bin=np.asarray(bins, dtype=int),
)
@dataclass(frozen=True)
class _Prediction:
"""A change's predicted lines together with the diff they are made against."""
base: str | None
after: str | None
lines: list[Line]
def _normalise(
predictions: Mapping[str, Iterable[Line]],
) -> dict[str, "_Prediction"]:
# The key is ``owner/repo#pr@base..after``: the ``owner/repo#pr`` part maps the
# prediction to a change, and the ``@base..after`` part names the diff it is made
# against. Both SHAs flow through to the node match, so a prediction against a
# different diff than the released one lifts to nodes that match nothing.
prepared: dict[str, "_Prediction"] = {}
for key, lines in predictions.items():
change, _, diff = key.partition("@")
base, after = (None, None)
if diff:
base, _, after = diff.partition("..")
base, after = base or None, after or None
prepared[change] = _Prediction(base=base, after=after, lines=list(lines))
return prepared
def _read_labels(labels: str | Path) -> list[dict]:
text = Path(labels).read_text(encoding="utf-8")
return [json.loads(line) for line in text.splitlines() if line.strip()]