Create Python API for VideoEncoder

[Dan-Flores]

This PR creates a simple VideoEncoder class, and updates several tests to utilize the VideoEncoder.to_file pattern.

• test_bad_input_parameterized is the VideoEncoder equivalent to the AudioEncoder test that ensures general error checking occurs, while test_bad_input tests method specific errors.

• test_contiguity: is the VideoEncoder equivalent to the AudioEncoder test to ensure contiguous and non-contiguous tensors can be encoded, and are encoded equivalently.

[NicolasHug]

Just use torch.randint(0, 256, size=(num_frames, channels, height, width), dtype=torch.uint8).contiguous()

[NicolasHug]

This is good, but you should be able to check for this which is stricter:

assert  non_contiguous_frames.is_contiguous(memory_format=torch.channels_last)

[Dan-Flores]

So far, I have assumed frames will be in NCHW format. This shape is retained after the permutations, so I do not believe channels_last applies here?

[NicolasHug]

The frames are always of NCHW shape . But internally they can have any arbitrary memory layout (format). These are mostly orthogonal concepts, i.e. frames can have NCHW shape while still being represented as NHWC layout (channels last).

I realize now the non_contiguous_frames you created are indeed not channels last. But we can force them to be by changing to:

contiguous_frames.permute(0, 2, 3, 1).contiguous().permute(0, 3, 1, 2)

and you should see that assert non_contiguous_frames.is_contiguous(memory_format=torch.channels_last) is passing now. Note how we're now permuting the channels dimension from N*HW to NHW*, and back.

Your test was still correct BTW - I just think it's good to actually explicitly check for channels_last because this is how frames may arrive.