DeepPoseKit projects can be created and managed in SimBa, and DeepPosekit tracking data can be imported into SimBA to generate machine learning behavioral classifiers.
DeepPosekit is an extremely powerful and flexible pose-estimation tracking tool that has a wide-range of features that enables user to create robust tracking models. DeepLabCut and LEAP based model architectures can also be generated within the DeepPoseKit pipeline which may allow model comparisons - please read the DeepPoseKit paper in ELife. The base DeepPoseKit package does not come with a GUI or standardized project folder structures, which users of other animal pose-estimation tools might have gotten used to. However, such folder structures are created when DeepPoseKit projects are managed through SimBA.
IMPORTANT I: Although DeepPoseKit currently does not come with a GUI, it's commandline functions offer much(!!) extended functionality over the SimBA menus accessing DeepPoseKit. We strongly urge those interested in DeepPoseKit to access its functions through the available notebooks when possible to get the full tour-de-force DeepPoseKit experience.
IMPORTANT II: For extensive details on the DeepPoseKit workflow, please see the documentation on the DeepPoseKit Github page.
Note: Depending on the specific version of the different pose estimation tools the dependencies may diverge. For example, note that DeepPoseKit require tensorflow >=1.13 and CUDA 10, while some versions of DeepLabCut run on earlier CUDA/tensorflow releases.
To start using DeepPoseKit in SimBA, begin by creating a SimBA project as decribed in-depth in Part 1 - Create a new project of the Scenario 1 Tutorial.
Next, close the Project Configuration window. In the main SimBA console window, click on Tracking. In the dropdown, chose DeepPoseKit, and click on Create DeepPoseKit project.
Once clicked, the following window should pop open:
Next to the SimBA project.in file, click on browse and select your SimBA project_config.ini file. Next to project Name, give your DeepPoseKit project a name (Note: avoid using spaces in the name. Instead, use underscore if needed). Click on Generate project to create your DeepPoseKit project folder structure.
You DeepPoseKit project will be generated within the chosen SimBA project. If you navigate to you SimBA project, and look within the project_folder\logs\measures\dpk
folder, you should see a folder named after your chosen Project Name
. The inside if this folder will contain several subfolders reminiscent of the sub-folders created by other packages:
Go ahead and close the Create DeepPoseKit project
window.
Once created, you may want to update the parent and swap columns in your generated skeleton.csv file. These columns defaults to NaN when DeepPoseKits are generated in SimBA - and they are most readily updated manually by opening the skeleton.csv file.
For more information on the parent and swap settings, please see the DeepPosekit repository.
In the main SimBA console window, click on Tracking. In the dropdown, chose DeepPoseKit, and click on Load DeepPoseKit project and the following window pops up:
When DeepPoseKit projects are created in SimBA, they each come with their own .ini file that stores your project settings chosen in the SimBA GUI. This is not the same project_config.ini that stores your main SimBA project settings, but a separate .ini file, exclusive to your DeepPoseKit project. Go ahead and click on browse File
and navigate to your DeepPoseKit .ini file located in your DeepPosekitProject (in the project_folder\logs\measures\dpk
directory). Once done, click on Load Project
.
After loading you newly created project, first navigate to the [Import videos]
tab.
- Under the
Import multiple videos
sub-heading, click onBrowse Folder
to select a folder that contains all the videos that you wish to import into your project. - Enter the file type of your videos. (e.g., mp4, avi, mov, etc. - do not enter the fullstop/period in the file-ending name) in the
Video type
entry box. - Click on
Import multiple videos
.
Note: If you have a lot of videos, it might take a few minutes before all the videos are imported into your project.
- Under the
Import single video
heading, click onBrowse File
to select your video. - Click on
Import a video
.
If you'd like to confirm that all the videos are imported into your project, then navigate to the project_folder\logs\measures\dpk\*DeepPoseKit project name*\input\videos
folder and you should see all of your imported videos.
- To create an annotation set, click on the
[Create annotation set]
tab. For more information on the different user-definable variables, refer to the DeepPosekit respository.
-
Enter an Annotation Output Name in the first entry box. *Note: Do not enter a file-ending, and avoid spaces in the annotation output name. For example, if I wish the annotation set file to be called MyAnnotationSet.h5, then I enter MyAnnotationSet in the Annotation Output Name entry box.
-
Click on
Create Annotation
. A message in the main SimBA terminal window will be printed to let you know when the annotation set creation is complete. To confirm that the annotation file has been correctly created, navigate to theproject_folder\logs\measures\dpk\*DeepPoseKit project name*\annotation_sets
folder and you should see all of your annotation file. If you named your file MyAnnotationSet in the Annotation Output Name entry box, you should see a file called MyAnnotationSet.h5.
In the DeepPoseKit window, click on the Annotator
tab and you should see the following window:
Click on Browse File
and select your annotation .h5 file located in the annotation_sets
directory. Click on Run
and the following windows should pop open, with instructions displayed on the right:
After the annotations are complete, press Esc
to save your annotations.
Click on the Train model
tab and ou should see the following window.
For documentation on the different parameters and network architectures available, see Jake Garvings' DeepPosekit documentation. In SimBA, most of the entry boxes and user-defined settings defaults to the DeepPoseKit suggested defults. To chooose the neural network archictecture, click on the NN_architecture
drop-down menu and you should see the networks available in DeepPoseKit:
To access architecture-specific settings, click on the Model Settings
button. Depending on the chosen architecture, you should see one of the three possible menus:
StackedDenseNet/Hourglass | DeepLabCut | LEAP |
---|---|---|
Again, see Jake Garvings' DeepPosekit documentation for the model-specific setting in the DeepLabCut, StackedDenseNet, StackedHourglass, and LEAP.
Once all the information is filled in, click on Train model
and the model training will start.
New models will be saved in the project_folder\logs\measures\dpk\*DeepPoseKit project name*\models
sub-directory.
- To use a DeepPoseKit-generated model to track animals in new videos, navigate to the
[Predict new video]
tab.
Next to Model path, click on Browse File
to select your DeepPoseKit-generated model. If the model was generated in SimBA, the model file should be located in the project_folder\logs\measures\dpk\*DeepPoseKit project name*\models
sub-directory (see Step 3, above).
- Next to Video folder, click on
Browse Folder
to select the folder containing your .mp4 video files that you wish to analyze. Then click onPredict
to start analyzing your videos.
Once complete, a message will be printed in in main SimBA console window. The output files, containing all your body-part locatations in each frame of the video, will be located in the project_folder\logs\measures\dpk\*DeepPoseKit project name*\predictions
sub-directory.
You can now go ahead and import these files into SimBA to generate behavioral predictive classifiers and/or anlyze behavioral repetoirs in the videos using already created predictive classifiers
To generate a visualization of analyzed videos, navigate to the [Visualiza video]
tab.
Author Simon N