A post-deeplabcut module for kinematic analysis

This repo will continue to grow, but here are some helper functions to get you started. Note, the API is subject to change. You can run the functions on data files obtained from running inference with DeepLabCut. Currently, this code requires python 3.8 onwards. We recommend using the DEEPLABCUT conda file, and then simply run pip install dlc2kinematics within your environment.

Quick start

pip install dlc2kinematics

Useage

import dlc2kinematics

Load data:

df, bodyparts, scorer = dlc2kinematics.load_data(<path of the h5 file>)

Basic Operations

Compute velocity:

• For all bodyparts:

  df_vel = dlc2kinematics.compute_velocity(df,bodyparts=['all'])

• For only few bodyparts:

  df_vel = dlc2kinematics.compute_velocity(df,bodyparts=['nose','joystick'])

Compute acceleration:

• For all bodyparts:

  df_acc = dlc2kinematics.compute_acceleration(df,bodyparts=['all'])

• For only few bodyparts:

  df_vel = dlc2kinematics.compute_acceleration(df,bodyparts=['nose','joystick'])

Compute speed:

df_speed = dlc2kinematics.compute_speed(df,bodyparts=['nose','joystick'])

Computations in joint coordinates

To compute joint angles, we first create a dictionary where keys are the joint angles and the corresponding values are the set of bodyparts:

joints_dict= {}
joints_dict['R-Elbow']  = ['R_shoulder', 'Right_elbow', 'Right_wrist']

and compute the joint angles with

joint_angles = dlc2kinematics.compute_joint_angles(df,joints_dict)

Compute joint angular velocity with

joint_vel = dlc2kinematics.compute_joint_velocity(joint_angles)

Compute joint angular acceleration with

joint_acc = dlc2kinematics.compute_joint_acceleration(joint_angles)

Compute correlation of angular velocity

corr = dlc2kinematics.compute_correlation(joint_vel, plot=True)

Compute PCA of angular velocity with

pca = dlc2kinematics.compute_pca(joint_vel, plot=True)

PCA-based reconstruction of postures

Compute and plot PCA based on posture reconstruction with:

dlc2kinematics.plot_3d_pca_reconstruction(df_vel, n_components=10, framenumber=500,
                                     bodyparts2plot=bodyparts2plot, bp_to_connect=bp_to_connect)

UMAP Embeddings

embedding, transformed_data = dlc2kinematics.compute_umap(df, key=['LeftForelimb', 'RightForelimb'], chunk_length=30, fit_transform=True, n_neighbors=30, n_components=3,metric="euclidean")

dlc2kinematics.plot_umap(transformed_data, size=5, alpha=1, color="indigo", figsize=(10, 6))

Contributing

• If you spot an issue or have a question, please open an issue with a suitable tag.
• For code contributions:
  • please see the contributing guide.
  • Please reference all issues this PR addresses in the description text.
  • Before submitting your PR, ensure all code is formatted properly by running

    black .

    in the root directory.
  • Assign a reviewer, typically MMathisLab.
  • sign CLA.

Acknowledgements

This code is a collect of contributions from members of the Mathis Laboratory over the years. In particular (alphabetical): Michael Beauzile, Sebastien Hausmann, Jessy Lauer, Alexander Mathis, Mackenzie Mathis, Tanmay Nath, Steffen Schneider.

If you use this code, please cite:

@software{mathis_mackenzie_2020_6669074,
  author       = {Mathis, Mackenzie and
                  Lauer, Jessy and
                  Nath, Tanmay and
                  Beauzile, Michael and
                  Hausmann, Sébastien and
                  Schneider, Steffen and
                  Mathis, Alexander},
  title        = {{DLC2Kinematics: a post-deeplabcut module for 
                   kinematic analysis}},
  month        = feb,
  year         = 2020,
  publisher    = {Zenodo},
  version      = {v0.0.4},
  doi          = {10.5281/zenodo.6669074},
  url          = {https://doi.org/10.5281/zenodo.6669074}
}