Expand external dataset coverage before harder real-data training

Constraint: Open-dataset ingestion needs a way to generate multiple overlapping queries per track, otherwise training/eval coverage stays too sparse Rejected: Keep only one random external query per track | Leaves long songs underrepresented and weakens reproducibility Confidence: high Scope-risk: moderate Directive: Preserve single-query behavior as the default, but keep overlap-query generation configurable through query_stride for future corpora Tested: manifest_tools audio-dir-to-splits --help shows --query-stride; prepare-local on data/synthetic_v2/songs with query_duration=8.0 and query_stride=4.0 produced 72 queries with query_index fields Not-tested: Full end-to-end smoke-local completion on the still-running real FMA corpus with overlap-query mode enabled

Expand external dataset coverage before harder real-data training
Constraint: Open-dataset ingestion needs a way to generate multiple overlapping queries per track, otherwise training/eval coverage stays too sparse Rejected: Keep only one random external query per track | Leaves long songs underrepresented and weakens reproducibility Confidence: high Scope-risk: moderate Directive: Preserve single-query behavior as the default, but keep overlap-query generation configurable through query_stride for future corpora Tested: manifest_tools audio-dir-to-splits --help shows --query-stride; prepare-local on data/synthetic_v2/songs with query_duration=8.0 and query_stride=4.0 produced 72 queries with query_index fields Not-tested: Full end-to-end smoke-local completion on the still-running real FMA corpus with overlap-query mode enabled
cnb.bofCdSsphPA
Commit a68a7296 ... a68a7296b29419e3cac387ade21fdb03fa51121d authored 2026-06-02 15:21:48 +0800 by cnb.bofCdSsphPA
Showing 5 changed files with 85 additions and 17 deletions
acr-engine/src/data/external_adapters.py
acr-engine/src/data/manifest_tools.py
docs/CHANGELOG.md
docs/dataset-spec.md
docs/open-dataset-workflow.md
--- a/acr-engine/src/data/external_adapters.py
View file @a68a729
+++ b/acr-engine/src/data/external_adapters.py
View file @a68a729
@@ -103,6 +103,7 @@ class BaseAdapter:
        output_root: Path,
        eval_ratio: float = 0.2,
        query_duration: float = 8.0,
+        query_stride: float | None = None,
        seed: int = 42,
    ) -> Dict:
        output_root.mkdir(parents=True, exist_ok=True)
@@ -118,9 +119,16 @@ class BaseAdapter:
            str(eval_ratio),
            "--query-duration",
            str(query_duration),
+        ]
+        if query_stride is not None:
+            cmd.extend([
+                "--query-stride",
+                str(query_stride),
+            ])
+        cmd.extend([
            "--seed",
            str(seed),
-        ]
+        ])
        result = subprocess.check_output(cmd, text=True)
        summary = json.loads(result)
        summary["input_dir"] = str(input_dir)
@@ -352,6 +360,7 @@ def smoke_local_dataset(
    output_root: Path,
    eval_ratio: float,
    query_duration: float,
+    query_stride: float | None,
    seed: int,
    train_epochs: int,
    batch_size: int,
@@ -378,6 +387,7 @@ def smoke_local_dataset(
        output_root / dataset,
        eval_ratio=eval_ratio,
        query_duration=query_duration,
+        query_stride=query_stride,
        seed=seed,
    )
    manifests_dir = Path(prepare_summary["output_dir"])
@@ -433,6 +443,7 @@ def smoke_local_dataset(
        resolved_device=resolved_device,
        base_cfg=base_cfg,
    )
+    config["data"]["manifest_query_stride"] = query_stride
    report_dir.mkdir(parents=True, exist_ok=True)
    config_path.write_text(json.dumps(config, indent=2))
@@ -481,6 +492,7 @@ def main():
    p.add_argument("--output-root", default="data/external_ingested")
    p.add_argument("--eval-ratio", type=float, default=0.2)
    p.add_argument("--query-duration", type=float, default=8.0)
+    p.add_argument("--query-stride", type=float, default=None)
    p.add_argument("--seed", type=int, default=42)
    p = sub.add_parser("inspect-local")
@@ -510,6 +522,7 @@ def main():
    p.add_argument("--output-root", default="data/external_smoke")
    p.add_argument("--eval-ratio", type=float, default=0.2)
    p.add_argument("--query-duration", type=float, default=8.0)
+    p.add_argument("--query-stride", type=float, default=None)
    p.add_argument("--seed", type=int, default=42)
    p.add_argument("--train-epochs", type=int, default=1)
    p.add_argument("--batch-size", type=int, default=2)
@@ -531,6 +544,7 @@ def main():
            root,
            eval_ratio=args.eval_ratio,
            query_duration=args.query_duration,
+            query_stride=args.query_stride,
            seed=args.seed,
        )
        print(json.dumps(summary, indent=2, ensure_ascii=False))
@@ -562,6 +576,7 @@ def main():
            output_root=Path(args.output_root),
            eval_ratio=args.eval_ratio,
            query_duration=args.query_duration,
+            query_stride=args.query_stride,
            seed=args.seed,
            train_epochs=args.train_epochs,
            batch_size=args.batch_size,
--- a/acr-engine/src/data/manifest_tools.py
View file @a68a729
+++ b/acr-engine/src/data/manifest_tools.py
View file @a68a729
@@ -9,6 +9,7 @@ import random
 import shutil
 from pathlib import Path
 from typing import List, Dict
+import numpy as np
 import soundfile as sf
@@ -43,6 +44,7 @@ def build_train_eval_from_audio_dir(
    exts: tuple[str, ...] = (".wav", ".mp3", ".flac", ".ogg"),
    eval_ratio: float = 0.2,
    query_duration: float = 8.0,
+    query_stride: float | None = None,
    seed: int = 42,
 ):
    rng = random.Random(seed)
@@ -80,21 +82,32 @@ def build_train_eval_from_audio_dir(
        refs.append(ref)
        if duration >= query_duration:
-            max_offset = max(0.0, duration - query_duration)
+            if query_stride and query_stride > 0:
-            offset = rng.uniform(0.0, max_offset) if max_offset > 0 else 0.0
+                max_offset = max(0.0, duration - query_duration)
-            query = {
+                offsets = [round(x, 3) for x in np.arange(0.0, max_offset + 1e-9, query_stride).tolist()]
-                "song_id": song_id,
+                if not offsets:
-                "audio_path": str(rel),
+                    offsets = [0.0]
-                "duration": query_duration,
+                if offsets[-1] < round(max_offset, 3):
-                "type": "clean",
+                    offsets.append(round(max_offset, 3))
-                "offset": round(offset, 3),
-                "segment_type": "external_query",
-                "source_dataset": source_dataset,
-            }
-            if rng.random() < eval_ratio:
-                test.append(query)
            else:
-                train.append(query)
+                max_offset = max(0.0, duration - query_duration)
+                offsets = [round(rng.uniform(0.0, max_offset) if max_offset > 0 else 0.0, 3)]
+            for seg_idx, offset in enumerate(offsets):
+                query = {
+                    "song_id": song_id,
+                    "audio_path": str(rel),
+                    "duration": query_duration,
+                    "type": "clean",
+                    "offset": offset,
+                    "segment_type": "external_query",
+                    "source_dataset": source_dataset,
+                    "query_index": seg_idx,
+                }
+                if rng.random() < eval_ratio:
+                    test.append(query)
+                else:
+                    train.append(query)
    if len(files) >= 2 and not train and test:
        train.append(test.pop())
@@ -109,6 +122,8 @@ def build_train_eval_from_audio_dir(
        "catalog": len(refs),
        "train_queries": len(train),
        "test_queries": len(test),
+        "query_duration": query_duration,
+        "query_stride": query_stride,
        "output_dir": str(manifests_dir),
    }
@@ -209,6 +224,7 @@ def main():
    p.add_argument("--source-dataset", required=True)
    p.add_argument("--eval-ratio", type=float, default=0.2)
    p.add_argument("--query-duration", type=float, default=8.0)
+    p.add_argument("--query-stride", type=float, default=None)
    p.add_argument("--seed", type=int, default=42)
    p = sub.add_parser("inspect-audio-dir")
@@ -230,6 +246,7 @@ def main():
            source_dataset=args.source_dataset,
            eval_ratio=args.eval_ratio,
            query_duration=args.query_duration,
+            query_stride=args.query_stride,
            seed=args.seed,
        )
        print(json.dumps({"status": "ok", **summary}, ensure_ascii=False))
--- a/docs/CHANGELOG.md
View file @a68a729
+++ b/docs/CHANGELOG.md
View file @a68a729
@@ -2,6 +2,37 @@
 ## 2026-06-02
+### Stage: 为外部数据集接入增加 overlap query manifest 能力
+完成项：
+- 修改 `acr-engine/src/data/manifest_tools.py`
+  - 为 `audio-dir-to-splits` 增加 `--query-stride`
+  - 支持按 stride 为单首歌生成多个 query
+  - 新 query 记录增加 `query_index`
+- 修改 `acr-engine/src/data/external_adapters.py`
+  - `prepare-local` / `smoke-local` 透传 `--query-stride`
+  - smoke 配置摘要里记录 `manifest_query_stride`
+- 更新 [open-dataset-workflow.md](./open-dataset-workflow.md) 与 [dataset-spec.md](./dataset-spec.md)
+验证结果：
+- CLI 验证：
+  - `manifest_tools.py audio-dir-to-splits --help` 已出现 `--query-stride`
+- 小数据验证：
+  - 执行：
+    - `prepare-local fma data/synthetic_v2/songs --query-duration 8.0 --query-stride 4.0`
+  - 返回：
+    - `train_queries = 57`
+    - `test_queries = 15`
+  - 解析 manifest 后得到：
+    - `num_queries = 72`
+    - `sample_query.offset = 0.0`
+    - `query_index = 0`
+    - `max_query_index = 2`
+结论：
+- 现在外部开源数据接入已经不再局限于“每首歌只采一个随机 query”
+- 当需要更高覆盖率时，可以直接生成多 query / overlap query manifests，用于更稳定的训练与评测
 ### Stage: 显式拆分 smoke 配置里的 8s query 与 5s training segment 语义
 完成项：
--- a/docs/dataset-spec.md
View file @a68a729
+++ b/docs/dataset-spec.md
View file @a68a729
@@ -91,13 +91,14 @@ flowchart TD
 |---|---|---:|---|
 | 训练 `SongPairDataset` | 每次采样随机取一个 5s clip | 否，**不是固定滑窗** | [acr-engine/src/data/dataset.py](../acr-engine/src/data/dataset.py) |
 | 检索 / embedding / 建索引 | `window_sec=5.0`, `stride_sec=2.5` | 是，**50% overlap** | [acr-engine/src/utils/audio.py](../acr-engine/src/utils/audio.py), [acr-engine/src/engines/ecapa_embedder.py](../acr-engine/src/engines/ecapa_embedder.py) |
-| `audio-dir-to-splits` | 每首歌只生成 1 个随机 query | 否 | [acr-engine/src/data/manifest_tools.py](../acr-engine/src/data/manifest_tools.py) |
+| `audio-dir-to-splits` 默认 | 每首歌只生成 1 个随机 query | 否 | [acr-engine/src/data/manifest_tools.py](../acr-engine/src/data/manifest_tools.py) |
+| `audio-dir-to-splits --query-stride 4.0` 例 | 对单首歌生成多个滑窗 query | 是，可配置 | [acr-engine/src/data/manifest_tools.py](../acr-engine/src/data/manifest_tools.py) |
 ### 直接回答你的问题
 - **有重叠窗口，但只在检索/索引链路里有。**
 - **当前训练主链路没有对 3 分钟 mp3 预展开成“全量重叠切片集”**，而是每次 batch 动态随机裁一个 5s 片段。
- **当前外部数据集 manifest 生成器也没有自动为每首歌生成多个重叠 query。**
+- **当前外部数据集 manifest 生成器默认仍是一首歌 1 个随机 query，但现在已经支持通过 `--query-stride` 开启多 query / overlap query 生成。**
 ---
--- a/docs/open-dataset-workflow.md
View file @a68a729
+++ b/docs/open-dataset-workflow.md
View file @a68a729
@@ -68,6 +68,7 @@ flowchart LR
 ```bash
 /usr/local/miniconda3/bin/python src/data/external_adapters.py inspect-local fma data/raw/fma_small_audio --eval-ratio 0.2 --query-duration 8.0
 /usr/local/miniconda3/bin/python src/data/external_adapters.py prepare-local fma data/raw/fma_small_audio --output-root data/external_ingested --eval-ratio 0.2 --query-duration 8.0
+/usr/local/miniconda3/bin/python src/data/external_adapters.py prepare-local fma data/raw/fma_small_audio --output-root data/external_ingested --eval-ratio 0.2 --query-duration 8.0 --query-stride 4.0
 /usr/local/miniconda3/bin/python src/data/external_adapters.py validate-local fma data/external_ingested/fma/manifests
 /usr/local/miniconda3/bin/python train.py --data data/external_ingested/fma/manifests --output data/models_fma_smoke --device cpu --epochs 1 --batch-size 2 --dry-run
 /usr/local/miniconda3/bin/python run_demo.py build-index --data data/external_ingested/fma/manifests --model data/models_fma_smoke/best_model.pt --output data/index_fma_smoke --device cpu
@@ -86,6 +87,7 @@ flowchart LR
 ```bash
 /usr/local/miniconda3/bin/python src/data/external_adapters.py smoke-local fma data/raw/fma_small_audio --output-root data/external_smoke --eval-ratio 0.2 --query-duration 8.0 --train-epochs 1 --batch-size 2
 /usr/local/miniconda3/bin/python src/data/external_adapters.py smoke-local fma data/raw/fma_small_audio --output-root data/external_smoke --eval-ratio 0.2 --query-duration 8.0 --train-epochs 1 --batch-size 2 --device auto
+/usr/local/miniconda3/bin/python src/data/external_adapters.py smoke-local fma data/raw/fma_small_audio --output-root data/external_smoke --eval-ratio 0.2 --query-duration 8.0 --query-stride 4.0 --train-epochs 1 --batch-size 2 --device auto
 ```
 真实目录放置位置可参考：
@@ -131,6 +133,8 @@ flowchart LR
  - 会一次性返回 inspect / prepare / validate / report 路径摘要
  - 现在支持 `--device cpu|cuda|auto`
  - `auto` 会在 smoke 内部解析成实际设备，避免把字符串 `auto` 直接传给 embedding/eval 侧
+  - 现在支持 `--query-stride`
+  - 当设置 `--query-stride < query-duration` 时，会为单首歌生成多个重叠 query，而不是只采 1 个随机 query
 ---