File size: 10,140 Bytes
ecaa1ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7deba50
ecaa1ff
 
 
 
7deba50
 
ecaa1ff
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11d2b0b
ecaa1ff
 
 
 
7deba50
 
 
ecaa1ff
7deba50
 
 
 
 
 
 
 
 
 
 
 
ecaa1ff
 
 
 
 
 
 
 
 
 
11d2b0b
ecaa1ff
 
 
11d2b0b
 
 
 
 
 
 
ecaa1ff
 
 
 
 
11d2b0b
7deba50
ecaa1ff
 
 
 
7deba50
ecaa1ff
 
 
 
 
7deba50
ecaa1ff
 
 
11d2b0b
 
 
 
7deba50
11d2b0b
 
 
 
7deba50
 
 
11d2b0b
7deba50
 
ecaa1ff
 
 
 
11d2b0b
 
7deba50
ecaa1ff
 
 
 
 
7deba50
 
 
 
 
 
 
 
 
 
 
 
 
11d2b0b
 
7deba50
 
 
ecaa1ff
 
 
 
 
 
 
7deba50
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ecaa1ff
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
"""End-to-end tests for the CoReDD evaluation.

A tiny real git repository is built with pygit2 in a temp dir: one base commit and one
child (the defect) that deletes two lines across two Python files. The synthetic
labels.jsonl is built by lifting those edited lines through the very projection the
evaluation uses, so node spans are never hard-coded and the predicted lines match by
construction. A trivial noise.json drives the perturbation.
"""

from __future__ import annotations

import json

import numpy as np
import pygit2
import pytest
from sklearn.metrics import f1_score, matthews_corrcoef

from coredd import metrics
from coredd.benchmark import Benchmark, Line, Result
from coredd.history import GitRepository
from coredd.noise import NoiseModel
from coredd.perturbation import Change, interval
from coredd.projection import NodeProjection


A_BASE = b"""\
def load(path):
    data = open(path).read()
    return data
"""

A_AFTER = b"""\
def load(path):
    data = open(path).read()
"""

B_BASE = b"""\
def save(path, data):
    handle = open(path, "w")
    handle.write(data)
"""

B_AFTER = b"""\
def save(path, data):
    handle = open(path, "w")
"""


def _tree(repo, files):
    builder = repo.TreeBuilder()
    for path, data in files.items():
        builder.insert(path, repo.create_blob(data), pygit2.GIT_FILEMODE_BLOB)
    return builder.write()


@pytest.fixture
def scenario(tmp_path):
    """Build the repo and the labels, deriving node spans from the projection.

    Returns (root, labels_path). ``a.py`` line 3 (deleted) is the defective node;
    ``b.py`` line 3 (deleted) is present in N(c) but left defect-free.
    """
    root = tmp_path / "repos"
    path = root / "acme" / "widget"
    path.mkdir(parents=True)
    repo = pygit2.init_repository(str(path), bare=True)
    sig = pygit2.Signature("t", "t@t", 0, 0)

    base = repo.create_commit(
        None, sig, sig, "base", _tree(repo, {"a.py": A_BASE, "b.py": B_BASE}), []
    )
    after = repo.create_commit(
        None, sig, sig, "after",
        _tree(repo, {"a.py": A_AFTER, "b.py": B_AFTER}), [base],
    )
    base, after = str(base), str(after)

    projection = NodeProjection(GitRepository(repo))
    a_node = projection.project(after, "a.py", 3, "base")  # defective
    b_node = projection.project(after, "b.py", 3, "base")  # defect-free
    assert a_node is not None and b_node is not None

    def node(span, defective):
        entry = {
            "path": span.path, "start": list(span.start), "end": list(span.end),
            "type": span.type, "base": base, "after": after, "bin": 1,
        }
        if defective:
            entry["provenance"] = [{"side": "base", "line": 3, "corrections": [99]}]
        return entry

    record = {
        "repository": "acme/widget", "pr": 1,
        "cutoff": "2023-01-01T00:00:00+00:00",
        "nodes": [node(a_node, True), node(b_node, False)],
    }
    labels = tmp_path / "labels.jsonl"
    labels.write_text(json.dumps(record) + "\n")
    return root, labels, base, after


@pytest.fixture
def noise(tmp_path):
    """A noise.json with vanishing rates: the score and interval are means and
    quantiles over the perturbed draws, so near-zero alpha and gamma make them
    reproduce the metric on the released labels."""
    path = tmp_path / "noise.json"
    path.write_text(json.dumps({
        "bins": 3,
        "alpha": {str(j): [0, 10_000_000] for j in range(3)},
        "gamma": {str(j): [0, 10_000_000] for j in range(3)},
    }))
    return path


@pytest.fixture
def noisy(tmp_path):
    """A noise.json with real rates: symmetric alpha, small gamma."""
    path = tmp_path / "noisy.json"
    path.write_text(json.dumps({
        "bins": 3,
        "alpha": {"0": [1, 1], "1": [1, 1], "2": [1, 1]},
        "gamma": {"0": [1, 40], "1": [1, 40], "2": [1, 40]},
    }))
    return path


@pytest.fixture
def benchmark(scenario, noise):
    root, labels, _, _ = scenario
    return Benchmark.load(repositories=root, labels=labels, noise=noise)


@pytest.fixture
def pred_key(scenario):
    """The full prediction key ``acme/widget#1@base..after`` for the fixture change."""
    _, _, base, after = scenario
    return Benchmark.key("acme/widget", 1, base, after)


def _hit_defective():
    # a.py line 3 (base side) lifts to the defective node.
    return Line(path="a.py", line=3, side="base", score=0.9)


def test_perfect_prediction_scores_one(benchmark, pred_key):
    result = benchmark.evaluate({pred_key: [_hit_defective()]}, metric=f1_score)
    assert result.score == pytest.approx(1.0)


def test_empty_prediction_scores_zero(benchmark):
    result = benchmark.evaluate({}, metric=f1_score)
    assert result.score == pytest.approx(0.0)


def test_missing_key_is_empty_prediction(benchmark):
    result = benchmark.evaluate({"other/repo#5": [_hit_defective()]},
                                metric=f1_score)
    assert result.score == pytest.approx(0.0)


def test_prediction_against_wrong_diff_misses(benchmark):
    # Right change and lines, but a base/after that is not the released diff: the lifted
    # node carries the wrong (base, after) and matches no label node -> no hit.
    key = Benchmark.key("acme/widget", 1, "deadbeef" * 5, "cafebabe" * 5)
    result = benchmark.evaluate({key: [_hit_defective()]}, metric=f1_score)
    assert result.score == pytest.approx(0.0)


def test_monte_carlo_is_deterministic(benchmark, pred_key):
    # Same seed reproduces score and band exactly. (Under the vanishing-noise fixture
    # the draws barely differ, so seed sensitivity is not asserted here; determinism
    # is the property that matters.)
    preds = {pred_key: [_hit_defective()]}
    a = benchmark.evaluate(preds, metric=f1_score, draws=500, seed=7)
    b = benchmark.evaluate(preds, metric=f1_score, draws=500, seed=7)
    assert a.interval == b.interval
    assert a.score == b.score


def test_band_contains_score_and_reports_draws(benchmark, pred_key):
    result = benchmark.evaluate({pred_key: [_hit_defective()]},
                                metric=f1_score, draws=800, seed=1)
    lo, hi = result.interval
    assert lo <= hi
    assert result.draws == 800


def test_score_is_a_draw_statistic(scenario, noisy, pred_key):
    # Under real label noise the score is the mean over the perturbed draws, not the
    # metric on the released labels: alpha thins the marked node in many draws, so a
    # perfect prediction scores below 1.0 and the mean sits inside the band.
    root, labels, _, _ = scenario
    benchmark = Benchmark.load(repositories=root, labels=labels, noise=noisy)
    result = benchmark.evaluate({pred_key: [_hit_defective()]},
                                metric=f1_score, draws=500, seed=3)
    lo, hi = result.interval
    assert result.score < 1.0
    assert lo <= result.score <= hi


def test_metric_agnostic(benchmark, pred_key):
    preds = {pred_key: [_hit_defective()]}
    f1 = benchmark.evaluate(preds, metric=f1_score, draws=200)
    mcc = benchmark.evaluate(preds, metric=matthews_corrcoef, draws=200)
    rr = benchmark.evaluate(preds, metric=metrics.mrr, pooled=False, draws=200)
    accuracy = lambda yt, yp: float((np.asarray(yt) == np.asarray(yp)).mean())
    acc = benchmark.evaluate(preds, metric=accuracy, draws=200)
    for result in (f1, mcc, rr, acc):
        assert isinstance(result, Result)
        assert not np.isnan(result.score)


def _change(y_true, y_pred, scores=None):
    y_true = np.asarray(y_true, dtype=bool)
    y_pred = np.asarray(y_pred, dtype=bool)
    if scores is None:
        scores = y_pred.astype(float)
    return Change(y_true=y_true, y_pred=y_pred,
                  scores=np.asarray(scores, dtype=float),
                  bin=np.zeros(y_true.size, dtype=int))


def _noise_free():
    # Beta(1, 1e7) posteriors: alpha and gamma are effectively zero, so every draw
    # reproduces the released labels and the aggregation is observed without noise.
    return NoiseModel(1, {0: (0, 10_000_000)}, {0: (0, 10_000_000)})


def test_pooled_and_per_case_aggregate_differently():
    # Per case: F1 is 1.0 on the small change and 0.0 on the large one, mean 0.5.
    # Pooled: TP=1, FP=5, FN=1 over the concatenated nodes, F1 = 2/(2+5+1) = 0.25.
    changes = [
        _change([1], [1]),
        _change([1, 0, 0, 0, 0, 0], [0, 1, 1, 1, 1, 1]),
    ]
    pooled, _ = interval(changes, f1_score, _noise_free(), draws=20,
                         rng=np.random.default_rng(0))
    per_case, _ = interval(changes, f1_score, _noise_free(), draws=20, pooled=False,
                           rng=np.random.default_rng(0))
    assert pooled == pytest.approx(0.25)
    assert per_case == pytest.approx(0.5)


def test_pooled_rank_metric_uses_one_global_ranking(benchmark):
    # Per case AUROC is undefined for both changes (a single class each), but the
    # pooled ranking separates the positive of one change from the negatives of the
    # other: positive score 0.9 above negatives 0.5 and 0.1 gives AUC 1.0.
    changes = [
        _change([0, 0], [0, 0], scores=[0.5, 0.1]),
        _change([1], [1], scores=[0.9]),
    ]
    level, _ = interval(changes, metrics.auroc, _noise_free(), draws=20,
                        rng=np.random.default_rng(0))
    assert level == pytest.approx(1.0)


def test_per_case_drops_undefined_changes():
    # AUROC is NaN on the all-positive change and 1.0 on the separable one; the
    # per-case mean drops the undefined change rather than propagating NaN.
    changes = [
        _change([1, 1], [1, 1], scores=[0.9, 0.8]),
        _change([1, 0], [1, 0], scores=[0.8, 0.2]),
    ]
    level, _ = interval(changes, metrics.auroc, _noise_free(), draws=20, pooled=False,
                        rng=np.random.default_rng(0))
    assert level == pytest.approx(1.0)


def test_mrr_is_undefined_without_positives():
    assert np.isnan(metrics.mrr([False, False], [0.5, 0.2]))


def test_result_repr_format():
    result = Result(score=0.62, interval=(0.55, 0.68), draws=10_000)
    assert repr(result) == (
        "Score: 0.62\n"
        "95% label-uncertainty interval: [0.55, 0.68]\n"
        "Monte Carlo draws: 10,000"
    )