Fix dataset script

Makes the dataset script faster (no need to extract the archives) and streamable

danbooru2023.py

@@ -1,12 +1,12 @@
 import os
+import posixpath
+
 import datasets
-from huggingface_hub import HfApi
-from datasets import DownloadManager, DatasetInfo
-from datasets.data_files import DataFilesDict
+from datasets import DatasetInfo, DownloadManager
+from fsspec.core import url_to_fs
+
 
 _EXTENSION = [".png", ".jpg", ".jpeg", ".webp", ".bmp"]
-_NAME = "nyanko7/danbooru2023"
-_REVISION = "main"
 
 
 class DanbooruDataset(datasets.GeneratorBasedBuilder):
@@ -28,23 +28,17 @@ class DanbooruDataset(datasets.GeneratorBasedBuilder):
         return info
 
     def _split_generators(self, dl_manager: DownloadManager):
-        hfh_dataset_info = HfApi().dataset_info(_NAME, revision=_REVISION, timeout=100.0)
-        data_files = DataFilesDict.from_hf_repo(
-            {datasets.Split.TRAIN: ["**"]},
-            dataset_info=hfh_dataset_info,
-            allowed_extensions=["tar", ".tar"],
-        )
-        gs = []
-        for split, files in data_files.items():
-            downloaded_files = dl_manager.download_and_extract(files)
-            gs.append(datasets.SplitGenerator(name=split, gen_kwargs={"filepath": downloaded_files}))
-        return gs
-
-    def _generate_examples(self, filepath):
-        for path in filepath:
-            all_fnames = {os.path.relpath(os.path.join(root, fname), start=path) for root, _dirs, files in os.walk(path) for fname in files}
-            image_fnames = sorted([fname for fname in all_fnames if os.path.splitext(fname)[1].lower() in _EXTENSION], reverse=True)
-            for image_fname in image_fnames:
-                image_path = os.path.join(path, image_fname)
-                post_id = int(os.path.splitext(os.path.basename(image_fname))[0])
-                yield image_fname, {"post_id": post_id}
+        fs, path = url_to_fs(dl_manager._base_path)
+        urls = fs.glob(posixpath.join(path, "**/*.tar"), detail=False)
+        archives = dl_manager.download(["hf://" + url for url in urls])
+        archives = [dl_manager.iter_archive(archives) for archives in archives]
+        return [datasets.SplitGenerator(name="train", gen_kwargs={"archives": archives})]
+
+    def _generate_examples(self, archives):
+        for archive in archives:
+            for path, f in archive:
+                path_root, path_ext = os.path.splitext(path)
+                if path_ext.lower() not in _EXTENSION:
+                    continue
+                post_id = int(os.path.basename(path_root))
+                yield path, {"image": {"bytes": f.read()}, "post_id": post_id}