| """Test whether a constant row offset materially improves cached teacher accuracy.""" |
| from pathlib import Path |
| import sys |
|
|
| ROOT = Path(r"E:\Gaze_estimation") |
| sys.path.insert(0, str(ROOT / ".codex_deps")) |
| import h5py |
| import numpy as np |
| import pandas as pd |
|
|
| OUT = ROOT / "artifacts" / "kd-teacher-trap-diagnostic" |
|
|
|
|
| def expectation(logits): |
| x = logits - logits.max(1, keepdims=True) |
| p = np.exp(x); p /= p.sum(1, keepdims=True) |
| return (p * np.arange(90)).sum(1) * 2 - 90 |
|
|
|
|
| rows = [] |
| for subject in ("p01", "p08", "p11"): |
| with h5py.File(ROOT / "data" / "processed" / f"{subject}_v16.h5", "r") as f: |
| pred = np.column_stack([expectation(f["teacher_pitch_logits"][:]), expectation(f["teacher_yaw_logits"][:])]) |
| gt = np.rad2deg(f["left_gaze"][:]) |
| for shift in range(-250, 251): |
| if shift >= 0: |
| p, g = pred[:len(pred)-shift or None], gt[shift:] |
| else: |
| p, g = pred[-shift:], gt[:len(gt)+shift] |
| mae = np.abs(p - g).mean() |
| rows.append({"subject": subject, "shift_gt_minus_teacher": shift, "overlap": len(p), "axis_mae_deg": mae}) |
| df = pd.DataFrame(rows) |
| df.to_csv(OUT / "teacher_constant_shift_audit.csv", index=False) |
| for subject in ("p01", "p08", "p11"): |
| cur = df[df.subject == subject].sort_values("axis_mae_deg") |
| zero = float(cur[cur.shift_gt_minus_teacher == 0].axis_mae_deg.iloc[0]) |
| best = cur.iloc[0] |
| print(subject, "zero", zero, "best", int(best.shift_gt_minus_teacher), float(best.axis_mae_deg), "gain", zero-float(best.axis_mae_deg)) |
|
|
