Download and Kinetics 400/600/700 Datasets

test/datasets_utils.py

@@ -416,7 +416,11 @@ def _populate_private_class_attributes(cls):
                 continue
 
             defaults.append(
-                {kwarg: default for kwarg, default in zip(argspec.args[-len(argspec.defaults):], argspec.defaults)}
+                {
+                    kwarg: default
+                    for kwarg, default in zip(argspec.args[-len(argspec.defaults):], argspec.defaults)
+                    if not kwarg.startswith("_")
+                }
             )
 
             if not argspec.varkw:
@@ -637,7 +641,7 @@ def __init__(self, *args, **kwargs):
 
     def _set_default_frames_per_clip(self, inject_fake_data):
         argspec = inspect.getfullargspec(self.DATASET_CLASS.__init__)
-        args_without_default = argspec.args[1:-len(argspec.defaults)]
+        args_without_default = argspec.args[1:(-len(argspec.defaults) if argspec.defaults else None)]
         frames_per_clip_last = args_without_default[-1] == "frames_per_clip"
 
         @functools.wraps(inject_fake_data)

test/test_datasets.py

@@ -944,6 +944,27 @@ def test_not_found_or_corrupted(self):
             super().test_not_found_or_corrupted()
 
 
+class KineticsTestCase(datasets_utils.VideoDatasetTestCase):
+    DATASET_CLASS = datasets.Kinetics
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(
+        split=("train", "val"), num_classes=("400", "600", "700")
+    )
+
+    def inject_fake_data(self, tmpdir, config):
+        classes = ("Abseiling", "Zumba")
+        num_videos_per_class = 2
+        tmpdir = pathlib.Path(tmpdir) / config['split']
+        digits = string.ascii_letters + string.digits + "-_"
+        for cls in classes:
+            datasets_utils.create_video_folder(
+                tmpdir,
+                cls,
+                lambda _: f"{datasets_utils.create_random_string(11, digits)}.mp4",
+                num_videos_per_class,
+            )
+        return num_videos_per_class * len(classes)
+
+
 class Kinetics400TestCase(datasets_utils.VideoDatasetTestCase):
     DATASET_CLASS = datasets.Kinetics400
 

test/test_datasets_download.py

@@ -392,6 +392,25 @@ def widerface():
     )
 
 
+def kinetics():
+    return itertools.chain(
+        *[
+            collect_download_configs(
+                lambda: datasets.Kinetics(
+                    path.join(ROOT, f"Kinetics{num_classes}"),
+                    frames_per_clip=1,
+                    num_classes=num_classes,
+                    split=split,
+                    download=True,
+                ),
+                name=f"Kinetics, {num_classes}, {split}",
+                file="kinetics",
+            )
+            for num_classes, split in itertools.product(("400", "600", "700"), ("train", "val"))
+        ]
+    )
+
+
 def kitti():
     return itertools.chain(
         *[
@@ -440,6 +459,7 @@ def make_parametrize_kwargs(download_configs):
             usps(),
             celeba(),
             widerface(),
+            kinetics(),
             kitti(),
         )
     )

torchvision/datasets/__init__.py

@@ -20,7 +20,7 @@
 from .sbd import SBDataset
 from .vision import VisionDataset
 from .usps import USPS
-from .kinetics import Kinetics400
+from .kinetics import Kinetics400, Kinetics
 from .hmdb51 import HMDB51
 from .ucf101 import UCF101
 from .places365 import Places365
@@ -34,6 +34,6 @@
            'Omniglot', 'SBU', 'Flickr8k', 'Flickr30k',
            'VOCSegmentation', 'VOCDetection', 'Cityscapes', 'ImageNet',
            'Caltech101', 'Caltech256', 'CelebA', 'WIDERFace', 'SBDataset',
-           'VisionDataset', 'USPS', 'Kinetics400', 'HMDB51', 'UCF101',
+           'VisionDataset', 'USPS', 'Kinetics400', "Kinetics", 'HMDB51', 'UCF101',
            'Places365', 'Kitti',
            )