精度評価実験の分析を複数被験者対応

tests/test_accuracy_report.py

@@ -7,7 +7,8 @@ from pathlib import Path
 import pytest
 
 from src.accuracy_report import (
-    Options, evaluate_trial, load_face_series, load_trials, main, percentile, run_report,
+    Options, evaluate_trial, kruskal_permutation, load_face_series, load_trials, main, percentile, run_report,
+    split_label, _signflip_p,
 )
 from src.accuracy_session import SessionRunner, build_plan
 
@@ -16,8 +17,8 @@ FPS = 30.0
 
 
 class FakeTime:
-    def __init__(self) -> None:
-        self.now = datetime(2026, 10, 4, 12, 0, 0, tzinfo=TZ)
+    def __init__(self, hour: int = 12) -> None:
+        self.now = datetime(2026, 10, 4, hour, 0, 0, tzinfo=TZ)
 
     def sleep(self, seconds: float) -> None:
         self.now += timedelta(seconds=seconds)
@@ -26,10 +27,15 @@ class FakeTime:
         return self.now.isoformat(timespec="milliseconds")
 
 
-def _simulate(tmp_path: Path, scale: float = 0.95, offset: float = -19.0, noise: float = 0.5, face_dropout: float = 0.0):
-    """Run a fake session and write a matching face_frames CSV (30 fps, yaw ramps after each cue)."""
+def _simulate(tmp_path: Path, scale: float = 0.95, offset: float = -19.0, noise: float = 0.5, face_dropout: float = 0.0,
+              hour: int = 12, seed: int = 0):
+    """Run a fake session and write a matching face_frames CSV (30 fps, yaw ramps after each cue).
+
+    ``hour`` shifts the wall clock so that several simulated subjects never overlap in time.
+    """
+    tmp_path.mkdir(parents=True, exist_ok=True)
     plan = build_plan([1, 3], [0, 5, -5, 10, -10], 3, seed=3)
-    fake, trials = FakeTime(), tmp_path / "trials.csv"
+    fake, trials = FakeTime(hour), tmp_path / "trials.csv"
     start = fake.now
 
     def operator_pause(_prompt: str) -> str:
@@ -39,10 +45,10 @@ def _simulate(tmp_path: Path, scale: float = 0.95, offset: float = -19.0, noise:
     SessionRunner(plan, trials, "S", input_fn=operator_pause, out=lambda _m: None, sleep=fake.sleep, clock=fake.clock,
                   beep=lambda _k: None).run()
     end = fake.now + timedelta(seconds=2)
-    rng = random.Random(0)
+    rng = random.Random(seed)
     marks = [(datetime.fromisoformat(r["t_turn_cue"]), float(r["true_yaw_deg"]), datetime.fromisoformat(r["t_hold_end"]))
              for r in csv.DictReader(trials.open(encoding="utf-8-sig"))]
-    path = tmp_path / "20261004_120000_exp_face_frames.csv"
+    path = tmp_path / f"20261004_{hour:02d}0000_exp_face_frames.csv"
     with path.open("w", encoding="utf-8-sig", newline="") as stream:
         writer = csv.writer(stream)
         writer.writerow(["frame_number", "timestamp", "track_id", "pose_estimated", "yaw", "raw_yaw"])
@@ -83,6 +89,9 @@ def test_full_pipeline_recovers_scale_bias_and_threshold(tmp_path: Path) -> None
     for name in ("accuracy_trials.csv", "accuracy_conditions.csv", "accuracy_thresholds.csv", "accuracy_report.md"):
         assert (tmp_path / "out" / name).is_file()
     assert "θ_th" in result["markdown"]
+    # one subject: no multi-subject sections or files
+    assert result["multi"] is None and "被験者別" not in result["markdown"]
+    assert not (tmp_path / "out" / "accuracy_loso.csv").exists()
 
 
 def test_flip_sign_is_flagged_by_negative_slope(tmp_path: Path) -> None:
@@ -137,3 +146,97 @@ def test_cli_end_to_end_with_glob_and_plots(tmp_path: Path, capsys: pytest.Captu
     assert "しきい値" in capsys.readouterr().out
     assert (out / "plots" / "accuracy_scatter.png").is_file()  # matplotlib is in requirements.txt
     assert main(["--trials", str(trials), "--face-csv", str(tmp_path / "missing.csv")]) == 2
+
+
+# ---------------------------------------------------------------------------------------------
+# multiple subjects
+# ---------------------------------------------------------------------------------------------
+def _two_subjects(tmp_path: Path, scale_b: float = 0.80):
+    ta, fa = _simulate(tmp_path / "a", scale=0.95, offset=-19.0, hour=12)
+    tb, fb = _simulate(tmp_path / "b", scale=scale_b, offset=-5.0, hour=14, seed=1)
+    return ta, fa, tb, fb
+
+
+def test_split_label() -> None:
+    assert split_label("A=trials_a.csv") == ("A", "trials_a.csv")
+    assert split_label(r"C:\data\x=1.csv") == ("", r"C:\data\x=1.csv")
+    assert split_label("trials.csv") == ("", "trials.csv")
+
+
+def test_subject_labels_aliases_and_exclusions(tmp_path: Path) -> None:
+    header = "session_id,seq,block,condition_id,distance_m,true_yaw_deg,pitch_deg,repeat,subject,status,t_baseline_start,t_turn_cue,t_hold_end\n"
+    a, b = tmp_path / "a.csv", tmp_path / "b.csv"
+    a.write_text(header + "S,1,1,c,1,10,0,1,S01,ok,a,b,c\nS,2,1,c,1,10,0,2,S01,ok,a,b,c\n", encoding="utf-8-sig")
+    b.write_text(header + "T,1,1,c,1,10,0,1,,ok,a,b,c\nT,2,1,c,1,10,0,2,,ok,a,b,c\n", encoding="utf-8-sig")
+    assert {r["subject"] for r in load_trials([a, b])} == {"S01", "b"}      # column, then file name
+    rows = load_trials([a, b], labels={b: "S02"}, aliases={"S01": "A"})
+    assert {r["subject"] for r in rows} == {"A", "S02"} and len(rows) == 4
+    kept = load_trials([a, b], exclude_seqs=["S02:2", 1], labels={b: "S02"})
+    assert [(r["subject"], r["seq"]) for r in kept] == [("S01", "2")]
+    only = load_trials([a, b], exclude_seqs=["S02:2"], labels={b: "S02"})
+    assert [(r["subject"], r["seq"]) for r in only] == [("S01", "1"), ("S01", "2"), ("S02", "1")]
+
+
+def test_multi_subject_pools_frames_and_reports_each_subject(tmp_path: Path) -> None:
+    ta, fa, tb, fb = _two_subjects(tmp_path)
+    result = run_report([ta, tb], [fa, fb], tmp_path / "out", Options(), plots=False, labels={ta: "A", tb: "B"})
+    assert result["subjects"] == ["A", "B"] and len(result["records"]) == 60
+    assert all(r["status"] == "ok" for r in result["records"])
+    multi = result["multi"]
+    per = multi["per_subject"]
+    assert abs(per["A"]["thresholds"][1.0]["slope"] - 0.95) < 0.05
+    assert abs(per["B"]["thresholds"][1.0]["slope"] - 0.80) < 0.05
+    pooled = result["thresholds"][1.0]
+    assert pooled["n_subjects"] == 2 and 0.80 < pooled["slope"] < 0.95
+    cond = {(c["distance_m"], c["true_yaw_deg"]): c for c in result["conditions"]}
+    assert cond[(1.0, 10.0)]["n_subjects"] == 2 and cond[(1.0, 10.0)]["n_trials"] == 6
+    assert cond[(1.0, 10.0)]["sd_subjects_deg"] > 0.5        # +10 deg reads 9.5 (A) vs 8.0 (B)
+    # leave-one-subject-out: one row per held-out subject and distance
+    loso = multi["loso"]
+    assert {(r["held_out"], r["distance_m"]) for r in loso} == {("A", 1.0), ("A", 3.0), ("B", 1.0), ("B", 3.0)}
+    assert all(r["n_train_subjects"] == 1 for r in loso)
+    # the +/- gain is symmetric in this simulation; the per-subject gains differ
+    asym = {r["subject"]: r for r in multi["asymmetry"]}
+    assert set(asym) == {"A", "B", "ALL"} and abs(asym["ALL"]["gain_diff"]) < 0.05
+    assert asym["A"]["gain_pos"] > asym["B"]["gain_pos"] + 0.05
+    assert multi["differences"] and all(r["p_permutation"] < 0.05 for r in multi["differences"])
+    for name in ("accuracy_subject_conditions.csv", "accuracy_subject_thresholds.csv", "accuracy_loso.csv",
+                 "accuracy_asymmetry.csv", "accuracy_asymmetry_cells.csv", "accuracy_subject_differences.csv"):
+        assert (tmp_path / "out" / name).is_file(), name
+    header = (tmp_path / "out" / "accuracy_conditions.csv").read_text(encoding="utf-8-sig").splitlines()[0]
+    assert header.endswith("exceed_rate,n_subjects,sd_subjects_deg")
+    for text in ("被験者別しきい値", "leave-one-subject-out", "正負方向の非対称", "A (30試行)", "B (30試行)"):
+        assert text in result["markdown"]
+
+
+def test_flip_only_the_named_subject(tmp_path: Path) -> None:
+    ta, fa, tb, fb = _two_subjects(tmp_path, scale_b=-0.95)
+    labels = {ta: "A", tb: "B"}
+    raw = run_report([ta, tb], [fa, fb], tmp_path / "o1", Options(), plots=False, labels=labels)
+    assert raw["multi"]["per_subject"]["B"]["thresholds"][1.0]["slope"] < 0
+    assert "--flip-subject B" in raw["markdown"]
+    fixed = run_report([ta, tb], [fa, fb], tmp_path / "o2", Options(flip_subjects=("B",)), plots=False, labels=labels)
+    per = fixed["multi"]["per_subject"]
+    assert per["A"]["thresholds"][1.0]["slope"] > 0.9 and per["B"]["thresholds"][1.0]["slope"] > 0.9
+    assert "--flip-subject B" not in fixed["markdown"]
+    only_a = run_report([ta, tb], [fa, fb], tmp_path / "o3", Options(), plots=False, labels=labels,
+                        exclude_subjects=["B"])
+    assert only_a["subjects"] == ["A"] and only_a["multi"] is None
+
+
+def test_cli_multi_subject_with_labels_and_subject_plots(tmp_path: Path, capsys: pytest.CaptureFixture[str]) -> None:
+    ta, fa, tb, fb = _two_subjects(tmp_path)
+    out = tmp_path / "report"
+    assert main(["--trials", f"A={ta}", f"B={tb}", "--face-csv", str(fa), str(fb), "--out", str(out)]) == 0
+    assert "被験者別しきい値" in capsys.readouterr().out
+    for name in ("accuracy_scatter.png", "accuracy_scatter_by_subject.png", "accuracy_error_by_subject.png",
+                 "accuracy_gain_by_side.png"):
+        assert (out / "plots" / name).is_file(), name
+    assert main(["--trials", f"A={ta}", "--face-csv", str(fa), "--alias", "broken"]) == 2
+
+
+def test_permutation_helpers() -> None:
+    assert _signflip_p([1.0, 1.0, 1.0, 1.0]) == pytest.approx(2 / 16)
+    assert math.isnan(_signflip_p([0.0, 0.0, 0.0]))
+    h, p = kruskal_permutation([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
+    assert h == pytest.approx(3.857, abs=0.01) and 0.05 < p < 0.2