A suite for postprocessing time-series extracted from videos of freely moving rodents using DeepLabCut

You can use this package to either extract pre-defined motifs from the time series (such as time-in-zone, climbing, basic social interactions) or to embed your data into a sequence-aware latent space to extract meaningful motifs in an unsupervised way! Both of these can be used within the package, for example, to automatically compare user-defined experimental groups.

How do I start?

Installation:

open a terminal (with python>3.6 installed) and type: pip install deepof

Before we delve in:

To start, create a folder for your project with at least two subdirectories inside, called 'Videos' and 'Tables'. The former should contain the videos you're working with (either you original data or the labeled ones obtained from DLC); the latter should have all the tracking tables you got from DeepLabCut, either in .h5 or .csv format. If you don't want to use DLC yourself, don't worry: a compatible pre-trained model for mice will be released soon!

my_project  -- Videos -> all tagged videos
            |
            |
            -- Tables -> all tracking tables (.h5 or .csv)

IMPORTANT: You should make sure that the tables and videos correspond to the same experiments. While the names should be compatible, this is handled by DLC by default.

Basic usage:

The main module with which you'll interact is called deepof.data. Let's import it and create a project:

import deepof.data
my_project = deepof.data.project(path="./my_project",
                                 arena_type="circular", # type of the filmed arena. So far, only "circular" is valid
                                 arena_dims=[380],      # diameter of the arena in milimeters
                                 smooth_alpha=0.99)

This command will create a deepof.data.project object storing all the necessary information to start. The smooth_alpha parameter will control how much smoothing will be applied to your trajectories, using an exponentially weighted average. Values close to 0 apply a stronger smoothing, and values close to 1 a very light one. In practice, we recommend values between 0.95 and 0.99 if your trajectories are not too noisy. There are other things you can do here, but let's stick to the basics for now.

One you have this, you can run you project using the .run() method, which will do quite a lot of computing under the hood (load your data, smooth your trajectories, compute distances and angles). The returned object belongs to the deepof.data.coordinates class.

my_project = my_project.run(verbose=True)

Once you have this, you can do several things! But let's first explore how the results of those computations I mentioned are stored. To extract trajectories, distances and/or angles, you can respectively type:

my_project_coords = my_project.get_coords(center=True, polar=False, speed=0, align="Nose", align_inplace=True)
my_project_dists  = my_project.get_distances(speed=0)
my_project_angles = my_project.get_angles(speed=0)

Here, the data are stored as deepof.data.table_dict instances. These are very similar to python dictionaries with experiment IDs as keys and pandas.DataFrame objects as values, with a few extra methods for convenience. Peeping into the parameters you see in the code block above, center centers your data (it can be either a boolean or one of the body parts in your model! in which case the coordinate origin will be fixed to the position of that point); polar makes the .get_coords() method return polar instead of Cartesian coordinates, and speed indicates the derivation level to apply (0 is position-based, 1 speed, 2 acceleration, 3 jerk, etc). Regarding align and align-inplace, they take care of aligning the animal position to the y Cartesian axis: if we center the data to "Center" and set align="Nose", align_inplace=True, all frames in the video will be aligned in a way that will keep the Center-Nose axis fixed. This is useful to constrain the set of movements that one can extract with out unsupervised methods.

As mentioned above, the two main analyses that you can run are supervised and unsupervised. They are executed by the .rule_based_annotation() method, and the .gmvae_embedding() methods of the deepof.data.coordinates class, respectively.

rule_based_annot = my_project.rule_based_annotation()
gmvae_embedding  = my_project.gmvae_embedding()

The former returns a deepof.data.table_dict object, with a pandas.DataFrame per experiment containing a series of annotations. The latter is a bit more complicated: it returns an array containing the encoding of the data per animal, another one with motif membership per time point (probabilities of the animal doing whatever is represented by each of the clusters at any given time), an abstract distribution (a multivariate Gaussian mixture) representing the extracted components, and a decoder you can use to generate samples from each of the extracted components (yeah, you get a generative model for free).

That's it for this (very basic) introduction. More detailed documentation, tutorials and method explanation will follow, so stay tuned!