Spaces:
Running on Zero
Running on Zero
File size: 5,648 Bytes
f234112 | 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 | """Visual debugging of the Phase 5 baseline detector.
Generates reports/figures/accident/debug/ containing ground-truth examples and
baseline-model predictions with predicted classes and confidence values, then
summarizes the baseline's failure mode (which classes it emits, confidence
distribution).
Uses the IMMUTABLE Phase 5 checkpoint (models/accident/best_detector.pth),
which has a 2-class head — loaded explicitly with num_classes=2.
"""
from __future__ import annotations
import json
import sys
from collections import Counter
from pathlib import Path
import numpy as np
import torch
if str(Path(__file__).resolve().parents[2] / "src") not in sys.path:
sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "src"))
from accident import CLASS_NAMES
from accident.dataset import AccidentParquetDataset
from accident.model import build_detector
from accident.utils import FIGURE_DIR, MODEL_DIR, get_device, seed_everything
from accident.visualization import save_ground_truth_example, save_prediction_example
NUM_EXAMPLES = 24
SCORE_THRESHOLD = 0.05
def main() -> int:
seed_everything(42)
device = get_device()
out_dir = FIGURE_DIR / "debug"
out_dir.mkdir(parents=True, exist_ok=True)
# --- Ground-truth examples -------------------------------------------------
ds = AccidentParquetDataset(split="validation", target_size=(320, 320))
# Pick a spread of images: accident-heavy, non-accident-heavy, mixed.
indices: list[int] = []
for i in range(len(ds)):
item = ds[i]
labels = item["labels"].numpy()
n_acc = int((labels == 0).sum())
n_non = int((labels == 1).sum())
if n_acc >= 2 and len(indices) % 2 == 0 and len(indices) < NUM_EXAMPLES:
indices.append(i)
elif n_non > n_acc and len(indices) < NUM_EXAMPLES:
indices.append(i)
indices = indices[:NUM_EXAMPLES]
for i in indices:
item = ds[i]
gt_labels = item["labels"].numpy()
n_acc = int((gt_labels == 0).sum())
n_non = int((gt_labels == 1).sum())
save_ground_truth_example(
item,
out_dir / f"gt_{i:04d}.png",
title=f"GT — image {i}: {n_acc} accident / {n_non} non_accident",
)
# --- Baseline predictions --------------------------------------------------
model = build_detector(num_classes=2, pretrained=False)
model.load_state_dict(
torch.load(MODEL_DIR / "best_detector.pth", map_location=device, weights_only=True)
)
model.to(device)
model.eval()
label_counts: Counter = Counter()
score_by_label: dict[int, list[float]] = {0: [], 1: []}
summary_rows: list[dict] = []
with torch.no_grad():
for i in indices:
item = ds[i]
outputs = model([item["image"].to(device)])
output = outputs[0]
keep = output["scores"] >= SCORE_THRESHOLD
pred_boxes = output["boxes"][keep].cpu().numpy()
# The baseline checkpoint was trained with the un-shifted (buggy)
# label space, where its single foreground class 1 was supervised
# ONLY by non_accident boxes. Interpret it with that identity
# mapping, not the corrected +1 mapping.
pred_labels = output["labels"][keep].cpu().numpy()
pred_scores = output["scores"][keep].cpu().numpy()
for lbl, sc in zip(pred_labels.tolist(), pred_scores.tolist()):
label_counts[lbl] += 1
score_by_label[lbl].append(sc)
save_prediction_example(
image_tensor=item["image"],
gt_boxes=item["boxes"].numpy(),
gt_labels=item["labels"].numpy(),
pred_boxes=pred_boxes,
pred_labels=pred_labels,
pred_scores=pred_scores,
save_path=out_dir / f"baseline_pred_{i:04d}.png",
title=f"Baseline prediction — image {i} (thr={SCORE_THRESHOLD})",
)
summary_rows.append(
{
"image_id": i,
"gt_accident": int((item["labels"].numpy() == 0).sum()),
"gt_non_accident": int((item["labels"].numpy() == 1).sum()),
"pred_accident": int((pred_labels == 0).sum()),
"pred_non_accident": int((pred_labels == 1).sum()),
"max_score": float(pred_scores.max()) if len(pred_scores) else 0.0,
}
)
all_scores = [s for scores in score_by_label.values() for s in scores]
report = {
"checkpoint": "models/accident/best_detector.pth (Phase 5 baseline, num_classes=2)",
"num_examples": len(indices),
"score_threshold": SCORE_THRESHOLD,
"predicted_label_counts_dataset_space": {
CLASS_NAMES.get(k, str(k)): v for k, v in sorted(label_counts.items())
},
"mean_score_by_label": {
CLASS_NAMES.get(k, str(k)): round(float(np.mean(v)), 4) if v else None
for k, v in score_by_label.items()
},
"max_score_overall": round(float(np.max(all_scores)), 4) if all_scores else None,
"images": summary_rows,
}
with open(out_dir / "baseline_debug_summary.json", "w", encoding="utf-8") as f:
json.dump(report, f, indent=2)
print("Predicted class counts (dataset space):", dict(report["predicted_label_counts_dataset_space"]))
print("Mean score by class:", report["mean_score_by_label"])
print(f"Debug visuals written to {out_dir}")
return 0
if __name__ == "__main__":
raise SystemExit(main()) |