Datasets:
File size: 7,355 Bytes
79918b7 | 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 | """Re-align the released v1.1.0 and v1.1.1 TL biometry plans to v1.0.0's exact
image-level train/test split, WITHOUT re-running the planner.
Why: the split is an image-level partition stored only in each plan's
train_cases[]/test_cases[] arrays. The sorted()-glob fix (dbb2028) landed after
v1.0.0, so v1.1.0 reshuffled ~41% of cases vs v1.0.0 and v1.1.1 inherited that.
This makes all three releases share v1.0.0's split by relabeling (moving case
objects between train/test to match v1.0.0's membership AND order). Annotation
VALUES (slice_profiles, landmark_file, figures) are byte-identical — only the
partition changes. v1.0.0 is the read-only anchor; case sets must already match
(BraTS-MET-00232-000 was removed beforehand).
Derived sidecars carry an informational per-entry "split" label; this refreshes
only those whose source plan is v1.1.0/v1.1.1 (consumers key on
task_ID/image_file/slice_dim/slice_idx and ignore "split", but keep it honest).
removed_samples_v1.0.0_to_* and multi_cluster_samples_v1.0.0_to_v1.1.0 are
sourced from the unchanged v1.0.0 baseline and are left untouched.
Backups of every touched file go to {dataset}/dev/bak/ before any write.
"""
import argparse
import gzip
import json
import shutil
from pathlib import Path
DATASETS = ["KiPA22", "HNTSMRG24", "autoPET-III", "KiTS23", "MSD", "BraTS24"]
V0 = "benchmark_plan_biometry_v1.0.0.json.gz"
V10 = "benchmark_plan_biometry_v1.1.0.json.gz"
V11 = "benchmark_plan_biometry_v1.1.1.json.gz"
# sidecar filename -> which re-aligned plan version supplies its "split" label
# ("v1.1.0"/"v1.1.1"); files sourced from the unchanged v1.0.0 baseline are absent.
SIDECAR_SOURCE = {
"multi_cluster_samples_v1.1.0.json": "v1.1.0",
"multi_cluster_samples_v1.1.1.json": "v1.1.1",
"added_samples_v1.0.0_to_v1.1.0.json": "v1.1.0",
"added_samples_v1.0.0_to_v1.1.1.json": "v1.1.1",
"added_samples_v1.1.0_to_v1.1.1.json": "v1.1.1",
"removed_samples_v1.1.0_to_v1.1.1.json": "v1.1.0",
}
def load_gz(p):
with gzip.open(p, "rt") as f:
return json.load(f)
def dump_gz(obj, p):
with gzip.open(p, "wt") as f:
json.dump(obj, f)
def v0_partition(plan):
"""task_ID -> {'train': [case_ID in order], 'test': [case_ID in order]}."""
out = {}
for t in plan["tasks"]:
out[t["task_ID"]] = {
"train": [c["case_ID"] for c in t.get("train_cases", [])],
"test": [c["case_ID"] for c in t.get("test_cases", [])],
}
return out
def realign_plan(plan, target):
"""Repartition each task's cases to match `target` (membership + order).
Mutates `plan` in place; returns (#moved, side_map) where side_map maps
(task_ID, image_file) and (task_ID, basename) -> 'train'|'test'."""
moved = 0
side_map = {}
for t in plan["tasks"]:
tid = t["task_ID"]
by_id = {c["case_ID"]: c for c in t.get("train_cases", []) + t.get("test_cases", [])}
tgt = target[tid]
have = set(by_id)
want = set(tgt["train"]) | set(tgt["test"])
if have != want:
raise SystemExit(
f" ABORT task {tid}: case sets differ from v1.0.0 "
f"(only-here={sorted(have - want)[:5]}, only-v0={sorted(want - have)[:5]})"
)
old_train = [c["case_ID"] for c in t.get("train_cases", [])]
new_train = [by_id[cid] for cid in tgt["train"]]
new_test = [by_id[cid] for cid in tgt["test"]]
moved += sum(1 for c in new_train if c["case_ID"] not in set(old_train))
t["train_cases"] = new_train
t["test_cases"] = new_test
t["train_cases_number"] = len(new_train)
t["test_cases_number"] = len(new_test)
for side, cases in (("train", new_train), ("test", new_test)):
for c in cases:
imf = c["image_file"]
side_map[(tid, imf)] = side
side_map[(tid, Path(imf).name)] = side
return moved, side_map
def patch_sidecar(path, side_map):
"""Refresh the 'split' label of every entry from side_map. Returns
(changed, total, misses)."""
with open(path) as f:
entries = json.load(f)
changed = misses = 0
for e in entries:
key = (e["task_ID"], e["image_file"])
side = side_map.get(key) or side_map.get((e["task_ID"], Path(e["image_file"]).name))
if side is None:
misses += 1
continue
if e.get("split") != side:
e["split"] = side
changed += 1
if misses == 0:
with open(path, "w") as f:
json.dump(entries, f, indent=2)
return changed, len(entries), misses
def main():
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--data_dir", required=True,
help="Root containing per-dataset folders with the plans.")
ap.add_argument("--dry_run", action="store_true", help="Report only; write nothing.")
args = ap.parse_args()
root = Path(args.data_dir)
for ds in DATASETS:
d = root / ds
print("=" * 88)
print(ds)
print("=" * 88)
v0 = load_gz(d / V0)
target = v0_partition(v0)
side_maps = {}
# backup + realign each version's plan
for ver, fn in (("v1.1.0", V10), ("v1.1.1", V11)):
plan = load_gz(d / fn)
moved, side_map = realign_plan(plan, target)
side_maps[ver] = side_map
if not args.dry_run:
bak = d / "dev" / "bak"
bak.mkdir(parents=True, exist_ok=True)
shutil.copy2(d / fn, bak / fn)
dump_gz(plan, d / fn)
print(f" {fn}: relabeled, {moved} case(s) moved to match v1.0.0")
# patch sidecars
for fn, ver in SIDECAR_SOURCE.items():
p = d / fn
if not p.exists():
continue
if not args.dry_run:
bak = d / "dev" / "bak"
bak.mkdir(parents=True, exist_ok=True)
shutil.copy2(p, bak / fn)
changed, total, misses = patch_sidecar(p, side_maps[ver])
else:
with open(p) as f:
entries = json.load(f)
sm = side_maps[ver]
misses = sum(1 for e in entries
if (e["task_ID"], e["image_file"]) not in sm
and (e["task_ID"], Path(e["image_file"]).name) not in sm)
changed, total = -1, len(entries)
flag = f" !! {misses} unmatched -> NOT written" if misses else ""
print(f" {fn}: split refreshed on {changed}/{total} entries (src {ver}){flag}")
# ---- verification: all three plans share one split (membership + order) ----
print("=" * 88)
print("VERIFY: v1.0.0 == v1.1.0 == v1.1.1 (membership + order) per task")
print("=" * 88)
allok = True
for ds in DATASETS:
d = root / ds
p0 = v0_partition(load_gz(d / V0))
p1 = v0_partition(load_gz(d / V10))
p2 = v0_partition(load_gz(d / V11))
ok = (p0 == p1 == p2)
allok &= ok
detail = "" if ok else " <-- MISMATCH"
print(f" {ds:<13} identical={ok}{detail}")
print(f"\nALL DATASETS aligned: {allok}")
if __name__ == "__main__":
main()
|