Epic-kitchens

Features

* Add `__getitem__` method to `EpicVideoDataset` to enable segments to be
retrieved by ID, rather than having to obtain the `video_segments` and filter
by ID.
* Make `idx` arg optional in `VideoDataset.load_frames`, by default we now load
all the frames for the segment, this behaviour is equivalent to
`VideoDataset.load_frames(segment, range(0, segment.num_frames)`
* Support string paths to constructor of `EpicVideoDataset` and `EpicVideoFlowDataset`
* Support test datasets in `EpicVideoDataset` and `EpicVideoFlowDataset`: Simply set
`class_type` as `None` and the video segment's `label` property will be `None`.
* Bugfix: Gulped flow frames were read in a random order (due to the results of
`glob.glob` not being sorted lexicographically), so if you have used a version
prior to 1.5.0 you should discard your gulped flow, and regulp using the fixed
adapter - our sincere apologies for the inconvenience caused.

**WARNING: People who have gulped flow**:
If you have used any previous version of the library to gulp flow,
then you should discard that flow and regulp the flow due to the previous
version of the adapter reading frames in an unsorted manner resulting in
flow segments within the gulp file being randomly shuffled.

Features

* Add `sample_transform` kwarg to `VideoDataset`, this allows you to
transform each frame or optical flow stack by providing a function that takes
in a list of PIL images and produces a list of PIL images
* Add `segment_filter` kwarg to `VideoDataset`, this allows you to selectively
filter action segments from a video by providing a function that takes a
`VideoSegment` and makes a decision on whether to include the segment in the
dataset or not, thus allowing you to filter to include or exclude specific
classes etc.

Features

* Both `epic_kitchens.preprocessing.split_segments` and
`epic_kitchens.gulp` now support reading CSV labels as well as a pickled
labels.
* Support un-narrated action segment splitting in
`epic_kitchens.preprocessing.split_segments` to enable splitting using the
test timestamp CSVs.
* Add `--unlabelled` option to `epic_kitchens.gulp` to enable gulping of test
set that doesn't have label data.

Bug fix

* Fix crash due to passing `Path` object to `os.path.lexists` in
`epic_kitchens.preprocessing.split_segments` on Python 3.5 (3.6+ supports
this)

Features

* Expose `epic_kitchens.preprocessing.split_segments` as an entrypoint
* Add docs for `epic_kitchens.preprocessing.*`

Bug fix

* `setup.py` used to import `epic_kitchens` which would fail if all dependencies
weren't already satisfied, which is the case in fresh virtual environments
causing the installation to fail, now the metadata is kept in a separate file
read in to `setup.py` to avoid this issue.