dataset script

dataset.py

@@ -0,0 +1,84 @@
+import csv
+import os
+from pathlib import Path
+from typing import Dict, Iterable, Any, Tuple
+
+import datasets
+
+# -------- CONFIG -------- #
+CSV_AUDIO_COL = "audio_segment_path"
+CSV_VIDEO_NO_AUDIO_COL = "video_noaudio_segment_path"
+CSV_VIDEO_WITH_AUDIO_COL = "video_withaudio_segment_path"
+
+_DESCRIPTION = """\
+VADA-AVSR: an audio-visual dataset of non-native English ("accents") and English varieties ("dialects")
+"""
+
+_LICENSE = ""
+
+# ------------------------ #
+
+class VisualAccentDialectArchive(datasets.GeneratorBasedBuilder):
+    """DatasetBuilder for VADA (with audio + two video fields)."""
+
+    VERSION = datasets.Version("1.0.0")
+
+    def _info(self) -> datasets.DatasetInfo:
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features(
+                {
+                    "utt_id": datasets.Value("string"),
+                    "filename": datasets.Value("string"),
+                    "segment_id": datasets.Value("int64"),
+                    "start": datasets.Value("float64"),
+                    "end": datasets.Value("float64"),
+                    "confidence": datasets.Value("float64"),
+                    "text": datasets.Value("string"),
+                    "audio": datasets.Audio(),
+                    "video_no_audio": datasets.Video(),
+                    "video_with_audio": datasets.Video(),
+                }
+            ),
+            homepage="https://huggingface.co/datasets/Berkeley-NLP/visual_accent_dialect_archive",
+            license=_LICENSE,
+        )
+
+    def _split_generators(self, dl_manager: datasets.DownloadManager):
+        # Paths are relative inside the repo snapshot
+        # This loads them via dl_manager even if local or hub
+        base_dir = Path(dl_manager.download_config.dataset_path or "")
+
+        csv_file = base_dir / "test.csv"
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST,
+                gen_kwargs={"csv_file": str(csv_file), "base_path": str(base_dir)},
+            ),
+        ]
+
+    def _generate_examples(
+        self, csv_file: str, base_path: str
+    ) -> Iterable[Tuple[int, Dict[str, Any]]]:
+        base = Path(base_path)
+
+        with open(csv_file, encoding="utf-8") as f:
+            reader = csv.DictReader(f)
+            for idx, row in enumerate(reader):
+                # Make local absolute path for each media reference
+                audio_path = str(base / row[CSV_AUDIO_COL])
+                noaudio_path = str(base / row[CSV_VIDEO_NO_AUDIO_COL])
+                withaudio_path = str(base / row[CSV_VIDEO_WITH_AUDIO_COL])
+
+                yield idx, {
+                    "utt_id": str(row["utt_id"]),
+                    "filename": str(row["filename"]),
+                    "segment_id": int(float(row["segment_id"])),
+                    "start": float(row["start"]),
+                    "end": float(row["end"]),
+                    "confidence": float(row["confidence"]),
+                    "text": str(row["text"]),
+                    "audio": audio_path,
+                    "video_no_audio": noaudio_path,
+                    "video_with_audio": withaudio_path,
+                }