dannce v1.2.0

This is a major update introducing new features and codebase improvements.

Contributors: Diego Aldarondo, Timothy Dunn

New Features

• Support for setups that use mirrors to provide multiple views with just a single camera.
• Multi-animal COM detection
• Support for training over very large motion capture datasets via .npy volume caching
  • Subsets of training frames can be sampled by setting num_train_per_exp
• Support for video file chunks of variable length
• Multi-gpu training, set via the multi_gpu_train config parameter
• Slurm scripts for automation, parallelization, and hyperparameter grid searches.
• Train/validation splits can now be set reproducibly using the data_split_seed config parameter
• Ability to ignore specific landmarks during training via the drop_landmark config parameter
• Support for validation over a dedicated recording via the valid_exp config parameter
• Random view sampling augmentation (on by default)

Codebase Improvements

• Additional documentation and typing hints
• Faster loading of training frames
• Improvements to metadata saving
• Refactoring interface.py, inference.py, and generator.py
• Reduced the repository download size by removing large files from the git history.

dannce v1.1.0

This release fixes bugs with metadata saving and implements a new mirror configuration mode, which instructs dannce to expect a single video with multiple views (with the aid of mirrors in the behavioral arena).

Mirror mode is selected by setting mirror to True in a config .yaml file, or with the cli. When using mirror mode, the params camera parameters variable inside *dannce.mat must indicate which views must be y-flipped, if any, using the m field.

dannce v1.0.0

This is a major update changing the way the networks are configured. This includes a major change to the expected organization of required files. Including:

• sync, calibration, and labeling folder content is now combined into a single master *dannce.mat file that lives in a particular project directory.
• COM predictions are now placed into this master *dannce.mat file after running COM prediction. The older types of COM prediction files are still generated in the com_predict_dir for legacy support, although they will likely be removed in a future release.
• config.yaml file structure has been reorganized to depend on a relatively static base config (see /configs/) containing parameters used across many recordings. This base config then points to a project-specific io.yaml file that coordinates paths to required files and can contain any other project-specific config params you want to set.
• exp.yaml files are no longer supported. Instead, use the exp block inside io.yaml
• train_DANNCE.py, predict_DANNCE.py, train_COMfinder.py, and predict_COMfinder.py have been replaced by entry points that are created when running setup.py: dannce-train, dannce-predict, com-train, com-predict.
• These entry points can now be massed command line arguments to override or add to the parameters in the config.yaml files. To see a list of command line parameters, pas --help to any of the above entry points, e.g. dannce-train --help at the command line.
• All of the underlying train/predict code can now be called as functions.
• Many parameters (e.g. extension) are now inferred automatically.
• Many parameters now have default values and thus are no longer included in the config files. But these default values can be changed by adding additional parameters to the config files or by using the command line interface. Use --help to see a full list of parameters and their default values.
• Some parameters have new names or have been combined. See --help of the DANNCE Wiki for more info. DANNCE will let you know if you are using an invalid parameter.
• Added basic image 2D (COM) and 3D (DANNCE) image augmentations, which are off by default.
• Added support for true monochrome networks.
• Other general refactoring

Thanks @diegoaldarondo for co-authoring this release and creating the CLI.

dannce v0.2.0

dannce prediction now accelerated significantly via pytorch 3D volume construction
bug fixes
refactoring

A big thanks to Dr. Kyle Severson for authoring the new torch code.

dannce v0.1.0

This is the first working version of dannce, without volume generation optimization on the GPU. Major improvements and structural changes follow in the next release.