Keypoint MoSeq

GitHub

Colab

Paper

Slack

License

Motion Sequencing (MoSeq) is an unsupervised machine learning method for animal behavior analysis. Given behavioral recordings, MoSeq learns a set of stereotyped movement patterns and when they occur over time. This package provides tools for fitting a MoSeq model to keypoint tracking data.

Setup

• Local installation
  • Install using conda
  • Install using pip
• Google colab

Tutorial

• Project setup
  • Edit the config file
  • Load data
  • Calibration
  • Fit PCA
• Model fitting
  • Setting hyperparameters
  • Initialization
  • Fitting an AR-HMM
  • Fitting the full model
  • Extract model results
    • Save results in csv format
  • Apply to new data
• Visualization
  • Trajectory plots
  • Crowd & grid movies
• Changepoints
  • Load data
  • Find changepoints
  • Visualize changepoints

FAQs

• Code usage
• Input data
  • How much data do I need?
  • How many/which keypoints?
  • Multiple animals
  • Keypoints are noisy
  • High proportion of NaNs
  • Should I preprocess keypoints?
  • Head-fixed animals
  • Non-rodents
  • Loading data from methods other than SLEAP or DeepLabCut
  • Size variation between animals
• Modeling
  • Validating model outputs
  • Choosing the target syllable duration
  • Number of model fitting iterations
  • Detecting existing syllables in new data
  • Continue model fitting but with new data
  • Interpreting model outputs
• Troubleshooting
  • Dead kernel
  • Out of memory
  • NaNs during fitting
  • Installation errors

Developer API

• Model fitting
• Visualization
• Input/Output
• Utilities
• Error Calibration