feat: migrate parquet schema to {path, audio, label, notes}

build_parquet.py

@@ -1,112 +1,114 @@
+import json
 import tempfile
-import tarfile
 from pathlib import Path
 
 from datasets import Audio, ClassLabel, Dataset, Features, Value
 
 REPO_ROOT = Path(__file__).resolve().parent
-TAR_PATH = REPO_ROOT / "data" / "ASVspoof2019LA_eval_flac.tar.gz"
-PROTOCOL_PATH = REPO_ROOT / "protocols" / "ASVspoof2019.LA.cm.eval.trl.txt"
 PARQUET_DIR = REPO_ROOT / "data"
-EXPECTED_ROWS = 71237
+PROTOCOL_PATH = REPO_ROOT / "protocols" / "ASVspoof2019.LA.cm.eval.trl.txt"
 NUM_SHARDS = 9
+EXPECTED_ROWS = 71237
 
-FEATURES = Features(
+NEW_FEATURES = Features(
     {
-        "trial_id": Value("string"),
+        "path": Value("string"),
         "audio": Audio(sampling_rate=16000),
-        "utterance_id": Value("string"),
-        "speaker_id": Value("string"),
-        "attack_id": Value("string"),
         "label": ClassLabel(names=["bonafide", "spoof"]),
+        "notes": Value("string"),
     }
 )
 
 
-def parse_protocol(path: Path) -> list[dict]:
-    rows = []
-    with open(path) as f:
-        for line in f:
-            parts = line.strip().split()
-            if len(parts) != 5:
-                continue
-            speaker_id, utterance_id, _, attack_id, label = parts
-            rows.append(
-                {
-                    "trial_id": utterance_id,
-                    "utterance_id": utterance_id,
-                    "speaker_id": speaker_id,
-                    "attack_id": attack_id,
-                    "label": label,
-                }
-            )
-    return rows
-
-
-def verify_protocol(rows: list[dict]) -> None:
-    assert len(rows) == EXPECTED_ROWS, f"Expected {EXPECTED_ROWS} rows, got {len(rows)}"
-    labels = {r["label"] for r in rows}
-    assert labels == {"bonafide", "spoof"}, f"Unexpected labels: {labels}"
-    trial_ids = [r["trial_id"] for r in rows]
-    assert len(set(trial_ids)) == EXPECTED_ROWS, "Duplicate trial_ids found"
-    for row in rows[:10]:
-        assert row["speaker_id"].startswith("LA_"), f"Bad speaker_id: {row['speaker_id']}"
-        assert row["utterance_id"].startswith("LA_E_"), f"Bad utterance_id: {row['utterance_id']}"
-
-
-def build_parquet() -> None:
-    rows = parse_protocol(PROTOCOL_PATH)
-    verify_protocol(rows)
-    print(f"Parsed {len(rows)} protocol rows")
-
-    with tempfile.TemporaryDirectory() as tmpdir:
-        print(f"Extracting tarball to {tmpdir}...")
-        with tarfile.open(TAR_PATH, "r:gz") as tar:
-            tar.extractall(tmpdir)
-
-        flac_dir = Path(tmpdir) / "flac"
-        protocol_ids = {r["utterance_id"] for r in rows}
-        missing = [uid for uid in protocol_ids if not (flac_dir / f"{uid}.flac").exists()]
-        assert not missing, f"Missing FLAC files for {len(missing)} utterance IDs"
-        print(f"All {len(protocol_ids)} protocol FLAC files found")
+def build_notes(utterance_id: str, speaker_id: str, subset: str = "eval") -> str:
+    return json.dumps(
+        {"utterance_id": utterance_id, "speaker_id": speaker_id, "subset": subset}
+    )
 
-        for row in rows:
-            flac_path = flac_dir / f"{row['utterance_id']}.flac"
-            row["audio"] = {"bytes": flac_path.read_bytes(), "path": f"{row['utterance_id']}.flac"}
 
-        print("Building dataset...")
-        ds = Dataset.from_list(rows, features=FEATURES)
-        print(f"Dataset created: {len(ds)} rows, features: {ds.features}")
+def migrate_shards() -> None:
+    import pyarrow.parquet as pq
 
-        print(f"Sharding into {NUM_SHARDS} parquet files...")
+    parquet_files = sorted(PARQUET_DIR.glob("test-*.parquet"))
+    print(f"Found {len(parquet_files)} parquet shards")
+
+    all_rows = []
+    for pf in parquet_files:
+        table = pq.read_table(pf)
+        for i in range(table.num_rows):
+            row = {col: table.column(col)[i].as_py() for col in table.column_names}
+            all_rows.append(row)
+
+    print(f"Read {len(all_rows)} total rows")
+    assert len(all_rows) == EXPECTED_ROWS, f"Expected {EXPECTED_ROWS}, got {len(all_rows)}"
+
+    migrated = []
+    for row in all_rows:
+        utterance_id = row["utterance_id"]
+        speaker_id = row["speaker_id"]
+        label_raw = row["label"]
+        label_str = "bonafide" if label_raw == 0 else "spoof"
+        audio = row["audio"]
+
+        migrated.append(
+            {
+                "path": f"{utterance_id}.flac",
+                "audio": audio,
+                "label": label_str,
+                "notes": build_notes(utterance_id, speaker_id),
+            }
+        )
+
+    print("Building dataset with new schema...")
+    ds = Dataset.from_list(migrated, features=NEW_FEATURES)
+    print(f"Dataset: {len(ds)} rows, features: {ds.features}")
+
+    print("Verifying...")
+    sample = ds[0]
+    assert sample["audio"]["sampling_rate"] == 16000
+    assert len(sample["audio"]["array"]) > 0
+    notes_parsed = json.loads(sample["notes"])
+    assert "utterance_id" in notes_parsed
+    assert notes_parsed["utterance_id"].startswith("LA_E_")
+
+    uid_set = set()
+    path_set = set()
+    for row in migrated:
+        n = json.loads(row["notes"])
+        uid_set.add(n["utterance_id"])
+        path_set.add(row["path"])
+    assert len(uid_set) == EXPECTED_ROWS, f"Duplicate utterance_ids: {EXPECTED_ROWS - len(uid_set)}"
+    assert len(path_set) == EXPECTED_ROWS, f"Duplicate paths: {EXPECTED_ROWS - len(path_set)}"
+
+    bonafide_count = sum(1 for r in migrated if r["label"] == "bonafide")
+    spoof_count = sum(1 for r in migrated if r["label"] == "spoof")
+    print(f"  bonafide: {bonafide_count}, spoof: {spoof_count}, total: {len(migrated)}")
+
+    print(f"Writing {NUM_SHARDS} shards...")
+    with tempfile.TemporaryDirectory() as tmpdir:
         for i in range(NUM_SHARDS):
             shard = ds.shard(num_shards=NUM_SHARDS, index=i)
-            shard_path = PARQUET_DIR / f"test-{i:05d}-of-{NUM_SHARDS:05d}.parquet"
+            shard_path = Path(tmpdir) / f"test-{i:05d}-of-{NUM_SHARDS:05d}.parquet"
             shard.to_parquet(str(shard_path))
             print(f"  Wrote {shard_path.name} ({len(shard)} rows)")
 
-        print("Verifying audio decode from in-memory dataset...")
-        sample = ds[0]
-        assert "audio" in sample, "No audio column in dataset"
-        assert sample["audio"]["sampling_rate"] == 16000, f"Wrong sample rate: {sample['audio']['sampling_rate']}"
-        assert len(sample["audio"]["array"]) > 0, "Empty audio array"
-        print(f"  Audio decode OK: {sample['audio']['sampling_rate']}Hz, {len(sample['audio']['array'])} samples")
-
-    print("Verifying parquet row count...")
-    import pyarrow.parquet as pq
+        for i in range(NUM_SHARDS):
+            shard_name = f"test-{i:05d}-of-{NUM_SHARDS:05d}.parquet"
+            src = Path(tmpdir) / shard_name
+            dst = PARQUET_DIR / shard_name
+            dst.write_bytes(src.read_bytes())
 
+    print("Verifying written shards...")
     parquet_files = sorted(PARQUET_DIR.glob("test-*.parquet"))
     total_rows = sum(pq.read_metadata(f).num_rows for f in parquet_files)
-    assert total_rows == EXPECTED_ROWS, f"Expected {EXPECTED_ROWS} rows, got {total_rows}"
+    assert total_rows == EXPECTED_ROWS, f"Expected {EXPECTED_ROWS}, got {total_rows}"
 
-    bonafide_count = sum(1 for r in rows if r["label"] == "bonafide")
-    spoof_count = sum(1 for r in rows if r["label"] == "spoof")
+    t = pq.read_table(str(parquet_files[0]))
+    assert set(t.column_names) == {"path", "audio", "label", "notes"}
+    print(f"Shard 0 columns: {t.column_names}")
 
     print("All verifications passed!")
-    print(f"  Rows: {total_rows}")
-    print(f"  Bonafide: {bonafide_count}, Spoof: {spoof_count}")
-    print(f"  Shards: {NUM_SHARDS}")
 
 
 if __name__ == "__main__":
-    build_parquet()
+    migrate_shards()

data/test-00000-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7a77a95e5dc47c7aa060ad08f6754eb1f13be116064dca882efa51a0ed02159c
-size 486163192
+oid sha256:0ed869f9c1c5ef591f5abf363eac2242aab8690c8803055aad07a05abbcbb011
+size 486188677

data/test-00001-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:df51477d2b1d6ded8101a60d44b46ff73432a0776e70875f45a9a5f46593a4ea
-size 486865931
+oid sha256:f3fb70e73f4dfcc50407f3d83c180fd3e1687635d172d1ff8d51056338c03772
+size 486891443

data/test-00002-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:977623682adac9b74e6569f6025af3b9b834b958489bf67360b257db03722ceb
-size 486760857
+oid sha256:743f2f8642a190447ed76ae6e2e7d87b8a822d477a10e06eeb696b5ef2ce351b
+size 486786361

data/test-00003-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4c3adc8b93e7a0ed955f141b16d3734e76b34d62c3ca9019fed69da39fe7344b
-size 486647406
+oid sha256:ca8b3493227396651db0c3b2938fe5460b14bdb6ae9faf63ede7fbc322ae9d20
+size 486672585

data/test-00004-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:24fa803aa4784409ff63eb6da718990d2bfd4ed35a5fba17a84440bf3bd016c3
-size 485788266
+oid sha256:8368306ad7f4c8a86707832ca516151af7882abe0eb81904355a088442379d76
+size 485813748

data/test-00005-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d11611e1b3e70693996c0e2b142bf9697ed8dfc7b51ec9234f8739c110daabf
-size 483760731
+oid sha256:bebdf6bd5915a8e18f326d52d6c260550472976447a447b9d82937330aa5030f
+size 483786272

data/test-00006-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b02c1eba813f81b6f2cc160d8887381b3c38d13cf57969ef903bdee0e510362
-size 486223531
+oid sha256:04acec349a4e59e5d21cfad56b79c5590558d94537b5df849d33c12c283244ba
+size 486248843

data/test-00007-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aeb53ebfcb40ba2098afcf8fa6d376017af144c7ce8dfe9cdf75180719e720ed
-size 489676388
+oid sha256:f6431527082ef430dbe094e4faea8f2fbaff872dcdefdb972ba40f8039ca8234
+size 489701913

data/test-00008-of-00009.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d52561a7b10d4ddb344401deac00a7bb22775f2d133233298d27f5877710e76b
-size 490112838
+oid sha256:e5107356edacab203df3bda7683e378311d004815079aeccaddd8a7fbf5263e2
+size 490138442