fix

@@ -11,7 +11,7 @@ python results/fix_same_singer_samples.py \
     --input  results/dna_eval.csv \
     --input  results/dna_eval.csv \
     --queries downloads_db/queries.csv \
     --queries downloads_db/queries.csv \
     --output  results/dna_eval_fixed.csv \
     --output  results/dna_eval_fixed.csv \
-    --threshold 0.05
+    --threshold 0.01
 """
 """
 
 
 import argparse
 import argparse
@@ -70,20 +70,21 @@ def compute_summary(rows: list[dict], threshold: float, out_path: str) -> dict:
     """计算评测指标。"""
     """计算评测指标。"""
     tp = fp = tn = fn = 0
     tp = fp = tn = fn = 0
     upload_ms, poll_ms, total_ms = [], [], []
     upload_ms, poll_ms, total_ms = [], [], []
-    by_variant: dict[str, dict] = {}
+    # 按 (sample_class, variant) 逐行统计，避免同一 variant 跨类混淆
+    by_class_variant: dict[str, dict[str, dict]] = {}
 
 
     for r in rows:
     for r in rows:
         pred = r.get("predicted_duplicate", "").lower() == "true"
         pred = r.get("predicted_duplicate", "").lower() == "true"
         exp = r.get("expected_duplicate", "").lower() == "true"
         exp = r.get("expected_duplicate", "").lower() == "true"
         correct = r.get("correct", "").lower() == "true"
         correct = r.get("correct", "").lower() == "true"
         sample_class = r.get("sample_class", "unknown")
         sample_class = r.get("sample_class", "unknown")
-
         variant = r.get("variant", "unknown")
         variant = r.get("variant", "unknown")
-        if variant not in by_variant:
-            by_variant[variant] = {"correct": 0, "total": 0, "sample_class": sample_class}
-        by_variant[variant]["total"] += 1
+
+        cls_dict = by_class_variant.setdefault(sample_class, {})
+        entry = cls_dict.setdefault(variant, {"correct": 0, "total": 0})
+        entry["total"] += 1
         if correct:
         if correct:
-            by_variant[variant]["correct"] += 1
+            entry["correct"] += 1
 
 
         if exp and pred:
         if exp and pred:
             tp += 1
             tp += 1
@@ -127,6 +128,12 @@ def compute_summary(rows: list[dict], threshold: float, out_path: str) -> dict:
             "max": round(s[-1], 1),
             "max": round(s[-1], 1),
         }
         }
 
 
+    def _variant_stats(cls_dict):
+        return {
+            v: {"accuracy": round(d["correct"] / d["total"], 4), "total": d["total"]}
+            for v, d in sorted(cls_dict.items())
+        }
+
     return {
     return {
         "total": total,
         "total": total,
         "duplicate_threshold": threshold,
         "duplicate_threshold": threshold,
@@ -139,23 +146,9 @@ def compute_summary(rows: list[dict], threshold: float, out_path: str) -> dict:
         "upload_time_ms": _stats(upload_ms),
         "upload_time_ms": _stats(upload_ms),
         "poll_time_ms": _stats(poll_ms),
         "poll_time_ms": _stats(poll_ms),
         "by_sample_class": {
         "by_sample_class": {
-            cls: {
-                "by_variant": {
-                    v: {"accuracy": round(d["correct"] / d["total"], 4), "total": d["total"]}
-                    for v, d in sorted(by_variant.items())
-                    if d["sample_class"] == cls
-                }
-            }
+            cls: {"by_variant": _variant_stats(by_class_variant.get(cls, {}))}
             for cls in ("positive", "negative")
             for cls in ("positive", "negative")
         },
         },
-        "by_variant": {
-            v: {
-                "sample_class": d["sample_class"],
-                "accuracy": round(d["correct"] / d["total"], 4),
-                "total": d["total"],
-            }
-            for v, d in sorted(by_variant.items())
-        },
         "out": out_path,
         "out": out_path,
     }
     }
 
 
@@ -196,7 +189,7 @@ def main():
 
 
     print(f"原始行数: {len(rows)}")
     print(f"原始行数: {len(rows)}")
 
 
-    removed = []
+    relabeled = []
     kept = []
     kept = []
     for r in rows:
     for r in rows:
         if r.get("sample_class") != "negative":
         if r.get("sample_class") != "negative":
@@ -214,23 +207,30 @@ def main():
         main_dur = main_durations.get(song_id)
         main_dur = main_durations.get(song_id)
         cover_dur = cover_durations.get(r.get("audio_path", ""))
         cover_dur = cover_durations.get(r.get("audio_path", ""))
 
 
+        is_same_recording = False
         if main_dur and cover_dur:
         if main_dur and cover_dur:
-            diff_ratio = abs(cover_dur - main_dur) / main_dur
-            if diff_ratio < args.threshold:
-                removed.append(r)
-                continue
+            is_same_recording = abs(cover_dur - main_dur) / main_dur < args.threshold
         elif main_dur:
         elif main_dur:
-            # 没有 cover 时长信息但同歌手，保守移除
-            removed.append(r)
-            continue
+            is_same_recording = True  # 没有 cover 时长，保守认为同录音
+
+        if is_same_recording:
+            # 重新标注为正样本：同录音跨平台分发，应该匹配
+            r = dict(r)
+            r["expected"] = "duplicate"
+            r["sample_class"] = "positive"
+            r["expected_song_id"] = song_id
+            # 重算 correct：预测为 duplicate 才算对
+            predicted = r.get("predicted_duplicate", "").lower() == "true"
+            r["correct"] = str(predicted)
+            relabeled.append(r)
 
 
         kept.append(r)
         kept.append(r)
 
 
-    print(f"移除行数: {len(removed)}（同歌手且时长相近）")
+    print(f"重新标注行数: {len(relabeled)}（同歌手且时长相近，改为 expected=duplicate）")
     print(f"保留行数: {len(kept)}")
     print(f"保留行数: {len(kept)}")
 
 
     # 写修复后的文件（去掉冗余列）
     # 写修复后的文件（去掉冗余列）
-    DROP_COLS = {"topk_hit", "expected_duplicate", "predicted_duplicate", "sample_class"}
+    DROP_COLS = {"topk_hit", "expected", "sample_class"}
     out_fieldnames = [f for f in fieldnames if f not in DROP_COLS]
     out_fieldnames = [f for f in fieldnames if f not in DROP_COLS]
     out_path = Path(args.output)
     out_path = Path(args.output)
 
 
@@ -294,12 +294,12 @@ def main():
         for v, d in variants.items():
         for v, d in variants.items():
             print(f"  {v:20s} acc={d['accuracy']:.4f}  n={d['total']}")
             print(f"  {v:20s} acc={d['accuracy']:.4f}  n={d['total']}")
 
 
-    # 打印被移除的样本明细
-    if removed:
-        print(f"\n移除的 {len(removed)} 条负样本:")
-        for r in removed:
+    # 打印重新标注的样本明细
+    if relabeled:
+        print(f"\n重新标注的 {len(relabeled)} 条（负样本→正样本）:")
+        for r in relabeled:
             print(f"  song_id={r.get('query_song_id')} variant={r.get('variant')} "
             print(f"  song_id={r.get('query_song_id')} variant={r.get('variant')} "
-                  f"singer={r.get('singer_name')} correct_was={r.get('correct')}")
+                  f"singer={r.get('singer_name')} correct={r.get('correct')}")
 
 
 
 
 if __name__ == "__main__":
 if __name__ == "__main__":