Spaces:
Running
Running
File size: 8,328 Bytes
d70361b | 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 | """
Ablate production detection configs against the frozen v3 Delhi test split.
Compares offline v3 operating point (DL-only @ thr=0.2) with TTA / registration
/ fusion variants so we can see what the webapp actually delivers vs the
reported Test F1 = 0.581.
Usage (from repo root):
python scripts/ablate_v3_app_pipeline.py
python scripts/ablate_v3_app_pipeline.py --configs dl_only_no_tta,full_pipeline
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from pathlib import Path
import numpy as np
from PIL import Image
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT))
try:
from dotenv import load_dotenv
load_dotenv(ROOT / ".env")
except ImportError:
pass
# Force v3 frozen weights + calibrated threshold for this benchmark.
os.environ["ADAPTFORMER_WEIGHTS"] = str(
(ROOT / "models" / "adaptformer_delhi" / "v3_frozen").resolve()
)
os.environ.setdefault("ADAPTFORMER_THRESHOLD", "0.2")
os.environ.setdefault("DETECTION_DL_THRESHOLD", "0.2")
os.environ.setdefault("DETECTION_DL_FLOOR_BASE", "0.15")
from app.detection_engine import run_detection # noqa: E402
from app.evaluation.metrics import binary_metrics # noqa: E402
from app.model_inference import ( # noqa: E402
get_calibrated_threshold,
get_loaded_weights_source,
predict_change_mask,
preload_model,
)
CONFIGS = {
"dl_only_no_tta": {
"desc": "v3 DL-only @ 0.2, TTA off (offline-style)",
"env": {"DETECTION_TTA": "off", "DETECTION_FUSION": "dl_only"},
"registration": False,
"use_engine": False,
},
"dl_only_tta": {
"desc": "v3 DL-only @ 0.2 + CPU TTA (hflip)",
"env": {"DETECTION_TTA": "hflip", "DETECTION_FUSION": "dl_only"},
"registration": False,
"use_engine": False,
},
"dl_only_reg": {
"desc": "v3 DL-only @ 0.2 + registration, TTA off",
"env": {"DETECTION_TTA": "off", "DETECTION_FUSION": "dl_only"},
"registration": True,
"use_engine": True,
},
"smart_union": {
"desc": "v3 + smart_union (floor<=0.15), TTA off, no reg",
"env": {
"DETECTION_TTA": "off",
"DETECTION_FUSION": "smart_union",
"DETECTION_DL_FLOOR_BASE": "0.15",
},
"registration": False,
"use_engine": True,
},
"hysteresis": {
"desc": "v3 + hysteresis fusion, TTA off, no reg",
"env": {"DETECTION_TTA": "off", "DETECTION_FUSION": "hysteresis"},
"registration": False,
"use_engine": True,
},
"full_pipeline": {
"desc": "Full production-like: smart_union + TTA auto + registration",
"env": {
"DETECTION_TTA": "auto",
"DETECTION_FUSION": "smart_union",
"DETECTION_DL_FLOOR_BASE": "0.15",
},
"registration": True,
"use_engine": True,
},
}
def _load_rgb(path: Path) -> np.ndarray:
if path.suffix.lower() in (".tif", ".tiff"):
from app.dda.geotiff_io import load_rgb_pil
return np.array(load_rgb_pil(path))
return np.array(Image.open(path).convert("RGB"))
def _load_test_pairs(manifest_path: Path):
data = json.loads(manifest_path.read_text(encoding="utf-8"))
pairs = []
for row in data.get("pairs", []):
before = ROOT / row["before_path"]
after = ROOT / row["after_path"]
gt_path = ROOT / row["gt_mask"]
if not (before.is_file() and after.is_file() and gt_path.is_file()):
print(f" skip missing files for {row['pair_id']}")
continue
before_arr = _load_rgb(before)
after_arr = _load_rgb(after)
gt = np.array(Image.open(gt_path).convert("L"))
pairs.append((row["pair_id"], before_arr, after_arr, gt))
return pairs
def _apply_env(overrides: dict[str, str]):
saved = {}
for key, value in overrides.items():
saved[key] = os.environ.get(key)
os.environ[key] = value
return saved
def _restore_env(saved: dict):
for key, old in saved.items():
if old is None:
os.environ.pop(key, None)
else:
os.environ[key] = old
def _eval_config(name: str, cfg: dict, pairs: list) -> dict:
saved = _apply_env(cfg["env"])
# TTA / fusion are read at call time; model already loaded.
f1s, ious, precs, recs = [], [], [], []
per_pair = {}
t0 = time.time()
thr = get_calibrated_threshold(0.2)
try:
for pair_id, before, after, gt in pairs:
if cfg["use_engine"]:
mask, _img, _stats, _regions = run_detection(
Image.fromarray(before),
Image.fromarray(after),
method="AI-Based Deep Learning",
enable_registration=cfg["registration"],
enable_normalization=True,
detection_sensitivity=0.5,
)
else:
mask, _score = predict_change_mask(before, after, threshold=thr)
if mask.shape != gt.shape:
import cv2
mask = cv2.resize(
mask, (gt.shape[1], gt.shape[0]), interpolation=cv2.INTER_NEAREST
)
m = binary_metrics(mask, gt)
f1s.append(m["f1"])
ious.append(m["iou"])
precs.append(m["precision"])
recs.append(m["recall"])
per_pair[pair_id] = m
finally:
_restore_env(saved)
return {
"config": name,
"desc": cfg["desc"],
"n_pairs": len(pairs),
"threshold": thr,
"mean_f1": round(float(np.mean(f1s)) if f1s else 0.0, 4),
"mean_iou": round(float(np.mean(ious)) if ious else 0.0, 4),
"mean_precision": round(float(np.mean(precs)) if precs else 0.0, 4),
"mean_recall": round(float(np.mean(recs)) if recs else 0.0, 4),
"seconds": round(time.time() - t0, 1),
"per_pair": per_pair,
"weights": get_loaded_weights_source(),
}
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--manifest",
default="data/delhi_cd/test/manifest.json",
help="Held-out v3 test split manifest",
)
parser.add_argument(
"--configs",
default=",".join(CONFIGS.keys()),
help="Comma-separated config keys",
)
parser.add_argument(
"--out",
default="runs/v3_app_ablation/results.json",
)
args = parser.parse_args()
wanted = [c.strip() for c in args.configs.split(",") if c.strip()]
unknown = [c for c in wanted if c not in CONFIGS]
if unknown:
raise SystemExit(f"Unknown configs: {unknown}. Choose from {list(CONFIGS)}")
print("Preloading v3 AdaptFormer weights...")
ok = preload_model()
print(f" loadedFrom={get_loaded_weights_source()} ok={ok} thr={get_calibrated_threshold(0.2)}")
if not ok or not get_loaded_weights_source() or "v3" not in str(get_loaded_weights_source()).lower():
raise SystemExit("Refusing to ablate: v3_frozen weights were not loaded")
pairs = _load_test_pairs(ROOT / args.manifest)
if not pairs:
raise SystemExit(f"No usable test pairs in {args.manifest}")
print(f"Evaluating {len(pairs)} test pair(s): {[p[0] for p in pairs]}\n")
rows = []
for name in wanted:
print(f"=== {name}: {CONFIGS[name]['desc']}")
row = _eval_config(name, CONFIGS[name], pairs)
rows.append(row)
print(
f" F1={row['mean_f1']:.4f} P={row['mean_precision']:.4f} "
f"R={row['mean_recall']:.4f} IoU={row['mean_iou']:.4f} "
f"({row['seconds']}s)"
)
out_path = ROOT / args.out
out_path.parent.mkdir(parents=True, exist_ok=True)
payload = {
"baseline_offline_v3": {
"test_f1": 0.5809,
"precision": 0.6782,
"recall": 0.5351,
"iou": 0.4117,
"threshold": 0.2,
},
"rows": rows,
}
out_path.write_text(json.dumps(payload, indent=2), encoding="utf-8")
print(f"\nWrote {out_path}")
best = max(rows, key=lambda r: r["mean_f1"])
print(f"Best app config: {best['config']} mean_f1={best['mean_f1']}")
if __name__ == "__main__":
main()
|