File size: 4,640 Bytes
cacb325 | 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 | """verify_bundle.py — prove this bundle is byte-identical to the frozen TEST-100 benchmark.
Run it after downloading. It needs only numpy (+ the bundled JSON); no GPU, no model, no source data.
python verify_bundle.py # from inside the bundle directory
Two independent checks per slice, 800 slices total:
1. HASH IDENTITY. `clean_sha256` / `meas_sha256` recomputed from the arrays and compared byte-for-
byte against the frozen manifest. This is the binding check: it proves the pixels and the
measurements are the same ones every published number was computed on, not a re-derivation that
happens to look similar.
2. INPUT-METRIC ANCHORS. `input_psnr` / `input_ssim` are recomputed from `naive` vs `clean` with
scikit-image at its DEFAULT settings (7x7 uniform window; a hand-rolled Gaussian-window SSIM
will NOT match) and compared against the manifest at the protocol tolerances (1e-5 / 1e-5).
If your loader or metric convention differs from the protocol's, this is where it shows up.
(`input_lpips` needs the `lpips` package and a network download, so it is checked only against
the value stored in the npz, not recomputed. The stored value came from the reference run.)
Exit code 0 = every check passed. Non-zero = do not use this bundle.
"""
import hashlib, json, os, sys
import numpy as np
TOL = {"psnr": 1e-5, "ssim": 1e-5}
def sha(a):
return hashlib.sha256(np.ascontiguousarray(a).tobytes()).hexdigest()
# The protocol's own metric definitions (medical/inverse/test100.py:_psnr/_ssim): skimage with
# DEFAULT parameters -- a 7x7 UNIFORM window, not a Gaussian one -- and the estimate clipped to
# [0,1] first. Reimplementing SSIM by hand does NOT reproduce these; use skimage or skip the check.
def psnr(a, b):
from skimage.metrics import peak_signal_noise_ratio as sk_psnr
return float(sk_psnr(a, np.clip(b, 0, 1), data_range=1.0))
def ssim(a, b):
from skimage.metrics import structural_similarity as sk_ssim
return float(sk_ssim(a, np.clip(b, 0, 1), data_range=1.0))
def main():
here = os.path.dirname(os.path.abspath(__file__))
man = json.load(open(os.path.join(here, "test100_cases.json")))
blob = open(os.path.join(here, "test100_cases.json"), "rb").read()
print(f"manifest sha256 = {hashlib.sha256(blob).hexdigest()}")
have_sk = True
try:
import skimage.metrics # noqa: F401
except ImportError:
have_sk = False
print("! scikit-image not installed — skipping the metric recomputation "
"(the sha256 identity checks, which are the binding ones, still run)")
total = bad = 0
for cb in man["classes"]:
f = os.path.join(here, f"test100_{cb['class']}.npz")
if not os.path.exists(f):
print(f"MISSING {os.path.basename(f)}"); bad += 1; continue
d = np.load(f, allow_pickle=False)
clean, y, naive = d["clean"], d["y"], d["naive"]
tags = [str(t) for t in d["case_tag"]]
errs = []
for i, case in enumerate(cb["cases"]):
total += 1
if tags[i] != case["case_tag"]:
errs.append(f"{tags[i]}: tag != manifest {case['case_tag']}"); continue
if sha(clean[i].astype(np.float32)) != case["clean_sha256"]:
errs.append(f"{tags[i]}: clean_sha256 MISMATCH"); continue
# y is stored squeezed; the manifest hashed it with the solver's leading dims
if sha(y[i]) != case["meas_sha256"] and sha(y[i][None, None]) != case["meas_sha256"]:
errs.append(f"{tags[i]}: meas_sha256 MISMATCH"); continue
if have_sk:
dp = abs(psnr(clean[i], naive[i]) - case["input_psnr"])
if dp > TOL["psnr"]:
errs.append(f"{tags[i]}: input_psnr drift {dp:.2e} > {TOL['psnr']}"); continue
ds = abs(ssim(clean[i], naive[i]) - case["input_ssim"])
if ds > TOL["ssim"]:
errs.append(f"{tags[i]}: input_ssim drift {ds:.2e} > {TOL['ssim']}")
bad += len(errs)
status = "OK" if not errs else f"{len(errs)} FAILURES"
print(f" class {cb['class_id']} {cb['class']:<22} n={len(cb['cases']):<4} "
f"{clean.shape} {y.dtype} {status}")
for e in errs[:5]:
print(f" {e}")
print(f"\n{total - bad}/{total} slices verified")
if bad:
print("FAILED — do not use this bundle."); return 1
print("All checks passed: this bundle is byte-identical to the frozen TEST-100 benchmark.")
return 0
if __name__ == "__main__":
sys.exit(main())
|