Won Pound by Won Pound

• Added new param --debug-side-length to music_into_network.py commands to control the vertical
  size of the debug videos.
• Switched canonical image resizing interpolation to cv2.INTER_CUBIC, resulting in better looking
  output videos.
• Note: This version that produced the production Won Pound music videos.

Production Videos

• Fixed --log argument in music_into_network.py commands.
• Fixed ability to set output side length in music_into_network.py commands.
• Added alternate frames -> video writer functionality that invokes ffmpeg directly, for better
  control of output quality. "High Quality" settings are optimized for uploading big videos to
  YouTube. Default settings still use MP4V encoder.

0.20.0 - Removed music mapped overlays

• In music_into_models.py projection-file-blend, the notion of connecting overlay appearances
  with audio activity has been removed, meaning --complexity-change-rolling-sum-window and
  --complexity-change-threshold are disabled. This functionality was underdeveloped, and didn't
  really work in a predictable way. The functionality will still be there in
  projection_file_blend_api, just not accessible via the CLI.
• In the same script, if --phash-distance, --bbox-distance and --track-length aren't provided,
  overlays won't be computed in the resulting videos.

0.19.0 - Production Synthesis Improvements

• Removed underscores from all CLI arguments.
• Refactored codebase to make sure 'network' was used over 'model' in all cases. They're networks
  not models.
  A few more changes to projection file blending:
• Added a few more arguments to music_into_models.py to be able to make selecting the production
  list of models easier. Now input args can be saved to file, and lists of models can be passed in
  as json.
• Switched phash computation to only take place on the bounding box regions of the images.
  This results to tighter looking overlays.
• Can use --wav argument multiple times now, good for working through projection files with
  multiple tracks.
• Audio files are now scaled in duration to the length of the video contained in the projection file.
  This way, the number of FFTs, and the number of final latents to blend are always the same. There's
  a check to make sure the projected frame count and the # of vectors are not too far off.

0.18.0 - Overlays

Adds overlay functionality to the music_into_models.py projection_file_blend command. Parts of the target are overlaid onto the output from the network, creating the effect that you're peering into reality.

0.17.0 - GitHub Publication

• Breaking up project and renaming it GANce for publication on GitHub.
• Added a number of examples for a blog post explaining this project here.