Raise an error if Kinetics400 dataset is empty

[vfdev-5]

🚀 Feature

Currently, if we construct a dataset on empty folder, it would be nice to have an error. Testing the length of the dataset is failing:

import torchvision
torchvision.__version__
> '0.8.0a0+e280f61'
dataset = torchvision.datasets.Kinetics400("/tmp/", frames_per_clip=10, step_between_clips=1, frame_rate=15)
len(dataset)
> Traceback (most recent call last):
  File "<input>", line 1, in <module>
  File "/torchvision/torchvision/datasets/kinetics.py", line 70, in __len__
    return self.video_clips.num_clips()
  File "/torchvision/torchvision/datasets/video_utils.py", line 247, in num_clips
    return self.cumulative_sizes[-1]
IndexError: list index out of range

Additional context

Without any errors related to empty dataset, reference video_classification example is failing inside the random sampler with misleading error:

  File "/vision/torchvision/datasets/samplers/clip_sampler.py", line 175, in __iter__                                                                               
    idxs_ = torch.cat(idxs)                                                                                                                                                 
RuntimeError: There were no tensor arguments to this function (e.g., you passed an empty list of Tensors), but no fallback function is registered for schema aten::_cat.  Th
is usually means that this function requires a non-empty list of Tensors.  Available functions are [CPU, CUDA, QuantizedCPU, BackendSelect, Named, AutogradOther, AutogradCP
U, AutogradCUDA, AutogradXLA, AutogradPrivateUse1, AutogradPrivateUse2, AutogradPrivateUse3, Tracer, Autocast, Batched, VmapMode].   

cc @pmeier @bjuncek

[pmeier]

Internally Kinetics400 uses datasets.folder.make_dataset() to collect the samples:

vision/torchvision/datasets/kinetics.py

Lines 47 to 49 in fc33c46

	classes = list(sorted(list_dir(root)))
	class_to_idx = {classes[i]: i for i in range(len(classes))}
	self.samples = make_dataset(self.root, class_to_idx, extensions, is_valid_file=None)

We have an error handling like you propose in datasets.DatasetFolder

vision/torchvision/datasets/folder.py

Lines 127 to 132 in fc33c46

	samples = self.make_dataset(self.root, class_to_idx, extensions, is_valid_file)
	if len(samples) == 0:
	msg = "Found 0 files in subfolders of: {}\n".format(self.root)
	if extensions is not None:
	msg += "Supported extensions are: {}".format(",".join(extensions))
	raise RuntimeError(msg)

Since this also relies on make_dataset() I'm wondering if it would make sense to move this check into make_dataset() instead implement it in every dataset that uses it. @fmassa?

[fmassa]

@pmeier sure, makes sense to move this check to be inside make_dataset