Release dev2

[data-hound]

New release for DANNCE

New Features

• Mirror augmentation added
  • Mirror the images in half of the samples of the training set
  • If mirror augmentation is used, the right_keypoints indices and left_keypoints indices must be specified as well.
    For the skeleton, ['RHand, 'LHand', 'RFoot', 'LFoot'],
      set right_keypoints: [0, 2] and
      left_keypoints: [1, 3] in the config file
• Options for layer and instance normalization added.
• More Losses available: Huber, logcosh, gaussian cross entropy
• n-rand-views configured to choose a subset of camera views for training/testing on weights trained for a lesser number of views.
  (Note: the number of views for which the weights have been trained should match the number of views specified with n-rand-views)
• AVG+MAX training introduced where both the mean of heatmaps and the pmax of heatmaps are used for loss computation.
• Saving prediction targets when running in MAX mode enabled

Codebase Improvements

• Cluster scripts have now been remodelled to use yaml config files specific to each cluster instead of a set of shell scripts
  • Sbatch submissions for dannce and com training and prediction have been implemented.
• Logging statements added throughout the code base with options to change the log level in the config (closes Logging code type error #117, closes dannce-predict loss very small, but result same like normal but shift #138)
• DANNCE now supports mat73 format files for label3d and predictions
• Python based scripts for plotting predictions on videos added
• Newer tests using function name parsing
• Uniform nomenclature for results_dir across codebase
• Callbacks moved to a new module
• Code refactored and reformatted across the codebase
• Trimming dependencies in multiple modules
• Saving pred targets moved to end of training

Bug Fixes and Minor Enhancements

• Allow multiple read attempts when loading video
• Added sigmoid normalization for pmax in dannce-predict
• Introduced closing videos from prev experiments
• Remove GPU placement during inference
• Model summary printed during training to help debug mismatches
• Save training data into npy volumes and cleanup unused
• Remove hardcoded file names in multi-gpu

[spoonsso]

@data-hound can you please test out the README install instructions in a fresh conda environment and then run all tests? thanks!

[data-hound]

Done. The install instructions work and all tests under /tests are passing

@data-hound can you please test out the README install instructions in a fresh conda environment and then run all tests? thanks!