Skip to content

Latest commit

 

History

History
153 lines (102 loc) · 4.96 KB

CHANGELOG.md

File metadata and controls

153 lines (102 loc) · 4.96 KB
Raw

Development

Version 1.7.1

Changes

  • Default path to download files is ~/.cache/epic-kitchens instead of ~/epic-kitchens

Version 1.7.0

Features

  • Add epic_kitchens.metrics and epic_kitchens.scoring containing method for computing metrics and manipulating model scores. Of particular note is epic_kitchens.metrics.compute_metrics which computes top-{1,5} accuracy, many-shot precision and recall for the verb, noun and action prediction tasks.

Dependencies

  • We now depend on scikit-learn for computing metrics.

Version 1.6.3

Dependencies

  • Depend on moviepy >= 1.0.0 due to imageio breaking ffmpeg download capabilities.

Version 1.6.2

Bug fix

  • epic_kitchens.gulp.visualisation visualisers returned ImageSequenceClip instead of IPython HTML display element, which caused the videos not to show up in Jupyter, these methods now return an IPython.display.HTML element
  • Fix bug in epic_kitchens.gulp.visualisation.combine_flow_uv_frames where hstack_frames wasn't called with width_axis correctly causing a RuntimeError

Version 1.6.1

Bug fix

  • Add sphinx 1.8.2 to requirements.txt to get RTFD to build the docs

Version 1.6.0

Features

  • Add epic_kitchens.meta package containing helpers for downloading and reading label class definitions, and training and test set annotations.
  • Add epic_kitchens.data.visualisation package containing tools that allow you to view arbitrary action segments RGB or Flow contained in a EpicVideoDataset using moviepy. This is useful when sifting through results per instance.

Changes

  • Improve documentation and bring it all in line with google doc string standards

Version 1.5.0

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.

Version 1.4.0

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.

Version 1.3.0

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.

Version 1.2.1

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)

Version 1.2.0

Features

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

Version 1.1.1

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.

Version 1.1.0

Bug fix

  • epic_kitchens.gulp in 1.0.0rc0 didn't read in the pickled dataframe before constructing the EpicDatasetAdapter causing an exception to be thrown therefore rendering the script useless.

Features

  • Change CLI interface of epic_kitchens.preprocessing.split_segments to match that of epic_kitchens.gulp in terms of argument ordering and whether arguments are mandatory or not.