initial video support

docs/source/package_reference/main_classes.mdx

@@ -245,6 +245,10 @@ Dictionary with split names as keys ('train', 'test' for example), and `Iterable
 
 [[autodoc]] datasets.Image
 
+### Video
+
+[[autodoc]] datasets.Video
+
 ## Filesystems
 
 [[autodoc]] datasets.filesystems.is_remote_filesystem

src/datasets/arrow_dataset.py

@@ -77,7 +77,7 @@
 from .arrow_writer import ArrowWriter, OptimizedTypedSequence
 from .data_files import sanitize_patterns
 from .download.streaming_download_manager import xgetsize
-from .features import Audio, ClassLabel, Features, Image, Sequence, Value
+from .features import Audio, ClassLabel, Features, Image, Sequence, Value, Video
 from .features.features import (
     FeatureType,
     _align_features,
@@ -1416,9 +1416,9 @@ def save_to_disk(
         """
         Saves a dataset to a dataset directory, or in a filesystem using any implementation of `fsspec.spec.AbstractFileSystem`.
 
-        For [`Image`] and [`Audio`] data:
+        For [`Image`], [`Audio`] and [`Video`] data:
 
-        All the Image() and Audio() data are stored in the arrow files.
+        All the Image(), Audio() and Video() data are stored in the arrow files.
         If you want to store paths or urls, please use the Value("string") type.
 
         Args:
@@ -5065,7 +5065,7 @@ def _estimate_nbytes(self) -> int:
 
             def extra_nbytes_visitor(array, feature):
                 nonlocal extra_nbytes
-                if isinstance(feature, (Audio, Image)):
+                if isinstance(feature, (Audio, Image, Video)):
                     for x in array.to_pylist():
                         if x is not None and x["bytes"] is None and x["path"] is not None:
                             size = xgetsize(x["path"])
@@ -5310,7 +5310,7 @@ def push_to_hub(
         """Pushes the dataset to the hub as a Parquet dataset.
         The dataset is pushed using HTTP requests and does not need to have neither git or git-lfs installed.
 
-        The resulting Parquet files are self-contained by default. If your dataset contains [`Image`] or [`Audio`]
+        The resulting Parquet files are self-contained by default. If your dataset contains [`Image`], [`Audio`] or [`Video`]
         data, the Parquet files will store the bytes of your images or audio files.
         You can disable this by setting `embed_external_files` to `False`.
 

src/datasets/config.py

@@ -129,6 +129,7 @@
 IS_MP3_SUPPORTED = importlib.util.find_spec("soundfile") is not None and version.parse(
     importlib.import_module("soundfile").__libsndfile_version__
 ) >= version.parse("1.1.0")
+DECORD_AVAILABLE = importlib.util.find_spec("decord") is not None
 
 # Optional compression tools
 RARFILE_AVAILABLE = importlib.util.find_spec("rarfile") is not None

src/datasets/features/__init__.py

@@ -12,8 +12,10 @@
     "Image",
     "Translation",
     "TranslationVariableLanguages",
+    "Video",
 ]
 from .audio import Audio
 from .features import Array2D, Array3D, Array4D, Array5D, ClassLabel, Features, LargeList, Sequence, Value
 from .image import Image
 from .translation import Translation, TranslationVariableLanguages
+from .video import Video

src/datasets/features/features.py

@@ -43,6 +43,7 @@
 from .audio import Audio
 from .image import Image, encode_pil_image
 from .translation import Translation, TranslationVariableLanguages
+from .video import Video
 
 
 logger = logging.get_logger(__name__)
@@ -1202,6 +1203,7 @@ class LargeList:
     Array5D,
     Audio,
     Image,
+    Video,
 ]
 
 
@@ -1346,7 +1348,7 @@ def encode_nested_example(schema, obj, level=0):
             return list(obj)
     # Object with special encoding:
     # ClassLabel will convert from string to int, TranslationVariableLanguages does some checks
-    elif isinstance(schema, (Audio, Image, ClassLabel, TranslationVariableLanguages, Value, _ArrayXD)):
+    elif isinstance(schema, (Audio, Image, ClassLabel, TranslationVariableLanguages, Value, _ArrayXD, Video)):
         return schema.encode_example(obj) if obj is not None else None
     # Other object should be directly convertible to a native Arrow type (like Translation and Translation)
     return obj
@@ -1397,7 +1399,7 @@ def decode_nested_example(schema, obj, token_per_repo_id: Optional[Dict[str, Uni
         else:
             return decode_nested_example([schema.feature], obj)
     # Object with special decoding:
-    elif isinstance(schema, (Audio, Image)):
+    elif isinstance(schema, (Audio, Image, Video)):
         # we pass the token to read and decode files from private repositories in streaming mode
         if obj is not None and schema.decode:
             return schema.decode_example(obj, token_per_repo_id=token_per_repo_id)
@@ -1417,6 +1419,7 @@ def decode_nested_example(schema, obj, token_per_repo_id: Optional[Dict[str, Uni
     Array5D.__name__: Array5D,
     Audio.__name__: Audio,
     Image.__name__: Image,
+    Video.__name__: Video,
 }