Initial commit: 上分帝（Climperor）

从 dota2-draft-vision 迁出并定名，作为天梯选将识别项目起点。

Co-authored-by: Cursor <cursoragent@cursor.com>

evaluate.py

@@ -0,0 +1,77 @@
+"""Score the recogniser against every labelled frame at once.
+
+Labels live in samples/labels.json as {frame filename: 10 hero keys}, '?' for
+slots nobody has identified yet. Those slots are skipped, not counted wrong.
+
+Usage:
+    python evaluate.py                # use the full template library
+    python evaluate.py --cdn-only     # ignore templates/real, measure the fallback layer
+"""
+
+import json
+import sys
+from pathlib import Path
+
+import cv2
+
+from common import ROOT, load_config, load_template_library
+from recognize import recognize_image
+
+LABELS_PATH = ROOT / "samples" / "labels.json"
+RAW_DIR = ROOT / "samples" / "raw"
+
+
+def main() -> None:
+    cdn_only = "--cdn-only" in sys.argv
+    if not LABELS_PATH.is_file():
+        sys.exit(f"missing {LABELS_PATH}")
+    frames = json.loads(LABELS_PATH.read_text(encoding="utf-8"))["frames"]
+
+    cfg = load_config()
+    if not cfg["slots"]:
+        sys.exit("config.json has no slots - run autocalibrate.py first")
+    library = load_template_library()
+    if cdn_only:
+        library = [t for t in library if t[1] == "cdn"]
+    print(f"library: {len(library)} templates{' (cdn only)' if cdn_only else ''}")
+
+    graded = correct = skipped = 0
+    misses: list[str] = []
+    for name, truth in frames.items():
+        path = RAW_DIR / name
+        img = cv2.imread(str(path))
+        if img is None:
+            print(f"  {name}: MISSING, skipped")
+            continue
+        result = recognize_image(img, cfg, library)
+
+        hits = frame_graded = 0
+        worst = 1.0
+        for slot, expected in zip(result["slots"], truth):
+            if expected == "?":
+                skipped += 1
+                continue
+            frame_graded += 1
+            worst = min(worst, slot["score"])
+            if slot["hero"] == expected:
+                hits += 1
+            else:
+                misses.append(
+                    f"    {name} slot {slot['slot']}: expected {expected}, "
+                    f"got {slot['hero']} (raw {slot['raw_best']} "
+                    f"score {slot['score']} margin {slot['margin']})"
+                )
+        graded += frame_graded
+        correct += hits
+        flag = " ranked" if result.get("ranked_overlay") else ""
+        print(f"  {name}: {hits}/{frame_graded}  lowest score {worst:.3f}  {result['elapsed_ms']}ms{flag}")
+
+    if misses:
+        print("\nmisses:")
+        print("\n".join(misses))
+    pct = 100 * correct / graded if graded else 0
+    print(f"\ntotal: {correct}/{graded} ({pct:.1f}%){f', {skipped} unlabelled slots skipped' if skipped else ''}")
+
+
+if __name__ == "__main__":
+    main()