Spaces:
Running
Running
File size: 8,936 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 | """
Mine hard-negative tiles (parking / seasonal veg / shadows) for AdaptFormer.
For labeled Delhi pairs, run frozen v3 at 256². Where the model fires but GT is
empty (false positives), save before/after/empty-gt crops into
``data/delhi_cd/hard_negatives/`` and append them to the train manifest.
Wednesday plan: broaden beyond vegetation to parking-row texture and
illumination/shadow FPs so the retrain stops firing on them.
Usage:
python scripts/mine_hard_negatives.py
python scripts/mine_hard_negatives.py --max-pairs 40 --thr 0.2 --min-fp-frac 0.03
"""
from __future__ import annotations
import argparse
import json
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))
OUT = ROOT / "data/delhi_cd/hard_negatives"
TRAIN_MAN = ROOT / "data/delhi_cd/train/manifest.json"
def _load_model(ckpt: Path):
import torch
from transformers import AutoModel, AutoImageProcessor
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = AutoImageProcessor.from_pretrained(str(ckpt), trust_remote_code=True)
model = AutoModel.from_pretrained(
str(ckpt), trust_remote_code=True).to(device).eval()
return model, processor, device, torch
def _predict(model, processor, device, torch, before, after):
from app.model_inference import _logits_to_change_prob
b = np.array(Image.fromarray(before).resize((256, 256)))
a = np.array(Image.fromarray(after).resize((256, 256)))
inputs = processor(images=(Image.fromarray(b), Image.fromarray(a)), return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}
with torch.no_grad():
score = _logits_to_change_prob(model(**inputs).logits, torch).cpu().numpy().astype(np.float32)
if score.ndim == 3:
score = score[0]
return score, b, a
def _tag_pair(pair: dict) -> list[str]:
"""Classify why this pair is a useful hard-neg source."""
types = [str(t).lower() for t in (pair.get("change_types") or [])]
notes = (pair.get("notes") or "").lower()
tags = ["hard_negative"]
blob = " ".join(types) + " " + notes
if any(k in blob for k in ("park", "car", "vehicle", "lot")):
tags.append("parking")
if any(k in blob for k in ("vegetation", "seasonal", "crop", "field", "mixed_gsd")):
tags.append("vegetation_seasonal")
if any(k in blob for k in ("shadow", "illumination", "brightness")):
tags.append("shadow")
# Empty / near-empty GT pairs from held-out FP set are seasonal FP sources
if "false positive" in notes or "seasonal" in notes or "texture" in notes:
if "vegetation_seasonal" not in tags:
tags.append("vegetation_seasonal")
if len(tags) == 1:
tags.append("generic_fp")
return tags
def _is_fp_source(pair: dict, *, include_all: bool) -> bool:
if include_all:
return True
tags = _tag_pair(pair)
return any(t in tags for t in ("parking", "vegetation_seasonal", "shadow", "generic_fp"))
def _shadowish_fp(before256: np.ndarray, after256: np.ndarray, fp: np.ndarray) -> bool:
"""True when FP pixels look like darkening-only (shadow / illumination)."""
if int(fp.sum()) < 50:
return False
b = before256.astype(np.float32).mean(axis=2)
a = after256.astype(np.float32).mean(axis=2)
delta = a[fp] - b[fp]
return float(np.mean(delta)) < -8.0
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--ckpt", default="models/adaptformer_delhi/v3_frozen")
ap.add_argument("--thr", type=float, default=0.2)
ap.add_argument("--max-pairs", type=int, default=40)
ap.add_argument("--min-fp-frac", type=float, default=0.03,
help="Min FP pixel fraction to keep a hard-neg tile")
ap.add_argument("--max-gt-frac", type=float, default=0.02,
help="Skip tiles that already have real change")
ap.add_argument("--include-all-labeled", action="store_true",
help="Scan every labeled pair (not just veg/parking/shadow tags)")
args = ap.parse_args()
from dotenv import load_dotenv
load_dotenv(ROOT / ".env", override=True)
from app.evaluation.delhi_eval import _load_label, _load_rgb
ckpt = ROOT / args.ckpt
if not (ckpt / "config.json").is_file():
print(f"Missing checkpoint {ckpt}")
return 1
man = json.loads((ROOT / "docs/delhi_eval/manifest.json").read_text(encoding="utf-8"))
pairs = [
p for p in man.get("pairs", [])
if p.get("gt_mask") and _is_fp_source(p, include_all=args.include_all_labeled)
]
# Prefer empty/near-empty GT pairs first (pure FP sources)
def _priority(p: dict) -> tuple:
tags = _tag_pair(p)
has_park = int("parking" in tags)
has_veg = int("vegetation_seasonal" in tags)
has_sh = int("shadow" in tags)
return (-(has_park + has_veg + has_sh), p.get("pair_id") or "")
pairs = sorted(pairs, key=_priority)[: args.max_pairs]
print(f"Scanning {len(pairs)} FP-prone pairs for hard negatives...")
model, processor, device, torch = _load_model(ckpt)
OUT.mkdir(parents=True, exist_ok=True)
saved = []
for p in pairs:
pid = p["pair_id"]
try:
before = _load_rgb(ROOT / p["before_path"])
after = _load_rgb(ROOT / p["after_path"])
gt = _load_label(ROOT / p["gt_mask"])
except Exception as exc:
print(f" skip {pid}: {exc}")
continue
score, b256, a256 = _predict(model, processor, device, torch, before, after)
gt256 = np.array(Image.fromarray(gt).resize((256, 256), Image.NEAREST))
pred = score >= args.thr
g = gt256 > 127
fp = pred & (~g)
fp_frac = float(fp.mean())
gt_frac = float(g.mean())
tags = _tag_pair(p)
if _shadowish_fp(b256, a256, fp) and "shadow" not in tags:
tags.append("shadow")
print(f" {pid}: gt={gt_frac:.3f} fp={fp_frac:.3f} tags={tags}")
if gt_frac > args.max_gt_frac:
continue
if fp_frac < args.min_fp_frac:
continue
hn_id = f"hn_{pid}"
before_p = OUT / f"{hn_id}_before.png"
after_p = OUT / f"{hn_id}_after.png"
gt_p = OUT / f"{hn_id}_gt.png"
Image.fromarray(b256).save(before_p)
Image.fromarray(a256).save(after_p)
Image.fromarray(np.zeros((256, 256), dtype=np.uint8)).save(gt_p)
vis = a256.copy()
vis[fp] = (255, 40, 40)
Image.fromarray(vis).save(OUT / f"{hn_id}_fp_overlay.png")
saved.append({
"pair_id": hn_id,
"before_path": str(before_p.relative_to(ROOT)).replace("\\", "/"),
"after_path": str(after_p.relative_to(ROOT)).replace("\\", "/"),
"gt_mask": str(gt_p.relative_to(ROOT)).replace("\\", "/"),
"change_types": tags,
"notes": (
f"Mined FP from {pid} (fp_frac={fp_frac:.3f}, "
f"tags={','.join(tags)})"
),
})
if not saved:
print("No hard negatives mined — try lowering --min-fp-frac or --include-all-labeled")
return 0
train = json.loads(TRAIN_MAN.read_text(encoding="utf-8"))
existing = {p.get("pair_id") for p in train.get("pairs", [])}
# Replace prior hn_* entries so the set stays fresh
train["pairs"] = [p for p in train.get("pairs", []) if not str(p.get("pair_id", "")).startswith("hn_")]
existing = {p.get("pair_id") for p in train["pairs"]}
added = 0
for row in saved:
if row["pair_id"] in existing:
continue
train.setdefault("pairs", []).append(row)
added += 1
train["n_pairs"] = len(train.get("pairs", []))
TRAIN_MAN.write_text(json.dumps(train, indent=2), encoding="utf-8")
ids_path = ROOT / "data/delhi_cd/train/pair_ids.txt"
ids = [ln.strip() for ln in ids_path.read_text(encoding="utf-8").splitlines() if ln.strip()]
ids = [i for i in ids if not i.startswith("hn_")]
for row in saved:
if row["pair_id"] not in ids:
ids.append(row["pair_id"])
ids_path.write_text("\n".join(ids) + "\n", encoding="utf-8")
summary = {
"n_saved": len(saved),
"n_added_to_train": added,
"thr": args.thr,
"ckpt": str(ckpt),
"tag_counts": {},
"pairs": saved,
"created_unix": time.time(),
}
for row in saved:
for t in row["change_types"]:
summary["tag_counts"][t] = summary["tag_counts"].get(t, 0) + 1
(OUT / "mining_summary.json").write_text(json.dumps(summary, indent=2), encoding="utf-8")
print(f"Saved {len(saved)} hard-neg tiles; added {added} to train manifest -> {OUT}")
print(f" tag_counts={summary['tag_counts']}")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|