(file:single-animal-userguide)=

---
local:
depth: 3
---

This document covers single/standard DeepLabCut use. If you have a complicated multi-animal scenario (i.e., they look the same), then please see our maDLC user guide.

DeepLabCut offers two equivalent interfaces: a GUI for those who prefer a visual workflow (no Python knowledge required), and a Python API for users who want scripting flexibility or to integrate DeepLabCut into a larger pipeline. All workflow steps are available in both.

We assume you have DeepLabCut installed (if not, see {ref}file:how-to-install). Open a terminal and activate your conda environment:

conda activate DEEPLABCUT

On Windows, always open the terminal with administrator privileges: right-click and
select "Run as administrator".

Choose your interface below to launch DeepLabCut:

python -m deeplabcut

---
name: fig-gui-launch
alt: The DeepLabCut Project Manager GUI after launch
width: 60%
align: center
---
The DeepLabCut Project Manager GUI.

In an interactive Python session (e.g. ipython), import DeepLabCut:

import deeplabcut

As a reminder, the core functions are described in our Nature Protocols paper (published at the time of DeepLabCut version 2.0.6). Additional functions and features are continually added to the package; we recommend reading the protocol alongside this documentation.

DeepLabCut's full workflow is described in steps (A)–(N) below. Code examples throughout this page use the Python API; if you are using the GUI, the same steps are available in the corresponding panels of the Project Manager.

The function create_new_project creates a new project directory, required subdirectories, and a basic project configuration file. Each project is identified by the name of the project (e.g. Reaching), name of the experimenter (e.g. YourName), as well as the date at creation.

Thus, this function requires the user to input:

• The name of the project
• The name of the experimenter
• The full path of the videos that are (initially) used to create the training dataset.
• Optional arguments specify:
  • The working directory
  • Where the project directory will be created
  • Whether to copy the videos to the project directory

If the optional argument `working_directory` is unspecified, the
project directory is created in the current working directory.

If `copy_videos` is unspecified symbolic links
for the videos are created in the videos directory.
Each symbolic link creates a reference to a video and thus
eliminates the need to copy the entire video to the video directory (if the videos remain at the original location).
This is why administrator privileges are required for Windows users, as creating symbolic links requires them.

deeplabcut.create_new_project(
    "Name of the project",
    "Name of the experimenter",
    ["Full path of video 1", "Full path of video 2", "Full path of video 3"],
    working_directory="Full path of the working directory",
    copy_videos=True/False,
    multianimal=False
)

On Windows, input paths as:
`r'C:\Users\computername\Videos\reachingvideo1.avi'` or
`'C:\\Users\\computername\\Videos\\reachingvideo1.avi'`

You can also place `config_path` in front of `deeplabcut.create_new_project` to create a variable that holds
the path to the config.yaml file, i.e. `config_path=deeplabcut.create_new_project(...)`

This set of arguments creates a project directory with the name ++ in the working directory and creates the symbolic links to videos in the videos directory.

The project directory will have subdirectories:

<Name of the project>+<name of the experimenter>+<date of creation of the project>/
├── dlc-models/
│   ├── iteration-0/
│   ├── iteration-1/
│   └── ...
├── dlc-models-pytorch/
│   ├── iteration-0/
│   │   └── <shuffle directories>/
│   │       ├── train/
│   │       └── test/
│   ├── iteration-1/
│   └── ...
├── labeled-data/
│   └── <video subdirectories>/
├── training-datasets/
├── videos/
└── config.yaml

All the outputs generated during the course of a project will be stored in one of these subdirectories, thus allowing each project to be managed independently of other projects.

1. dlc-models and dlc-models-pytorch: These directories have the same structure but store model files for different engines:

   • dlc-models for TensorFlow
   • dlc-models-pytorch for PyTorch
   • At the top level, both contain iteration folders such as:
     • iteration-0
     • iteration-1, etc. which correspond to successive rounds of label refinement. Each iteration folder in turn contains shuffle directories, each representing a specific experiment defined by a particular train/test split and model architecture. Within each shuffle directory:
       • train/ and test/ store metadata and configuration files for the feature detectors.
       • These configuration files are written in YAML, a human-readable format that can be edited with any standard text editor.
       • The train/ folder also stores training checkpoints (snapshots), which let users reload a trained model or resume training from an intermediate checkpoint if training was interrupted.

2. labeled-data/: Contains the extracted frames used to build the training dataset. Frames from different videos are stored in separate subdirectories, and each frame filename encodes its temporal position in the source video, making it easy to trace each frame back to its origin.

3. training-datasets/: Stores the generated training datasets along with metadata describing how each dataset was created.

4. videos/: Stores either the project videos themselves or symbolic links to them: If copy_videos=False (default), it contains symbolic links. If copy_videos=True, the videos are copied into the directory.

deeplabcut.add_new_videos(
    "Full path of the project configuration file",
    ["full path of video 4", "full path of video 5"],
    copy_videos=True/False
)

The *Full path of the project configuration file* will be referenced as `config_path` throughout this guide.

The project directory also contains the main configuration file called config.yaml. The config.yaml file contains many important parameters of the project. A complete list of parameters including their description can be found in {ref}Box 1 <config-box1>.

The create_new_project step writes the following parameters to the configuration file:

• Task
• Scorer
• Date
• Project_path
• A list of videos video_sets.

The first three parameters should **not** be changed.
The list of videos can be changed by adding new videos or manually removing videos.

---
name: config-box1
alt: Box 1 - Single Animal Project Configuration File Glossary
align: center
---
Single Animal project configuration file glossary

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.create_new_project.rst
```

---
width: 175px
title: colormaps
alt: DLC Utils
align: right
---

Next, open the config.yaml file, which was created during create_new_project. You can edit this file in any text editor. Familiarize yourself with the meaning of the parameters ({ref}Box 1 <config-box1>). You can edit various parameters, in particular you must add the list of bodyparts (or points of interest) that you want to track. You can also set the colormap here that is used for all downstream steps (can also be edited at anytime), like labeling GUIs, videos, etc. Here any matplotlib colormaps will do!

Please do not include spaces in the names of bodyparts.

• bodyparts: are the bodyparts of each individual (in the above list).

A good training dataset should consist of a sufficient number of frames that capture the breadth of the
behavior. This ideally implies to select the frames from different (behavioral) sessions, different lighting and
different animals, if those vary substantially (to train an invariant, robust feature detector). Thus for creating a
robust network that you can reuse in the laboratory, a good training dataset should reflect the diversity of the
behavior with respect to postures, luminance conditions, background conditions, animal identities, etc. of the data that
will be analyzed. For the simple lab behaviors comprising mouse reaching, open-field behavior and fly behavior, 100−200
frames gave good results [Mathis et al, 2018](https://www.nature.com/articles/s41593-018-0209-y). However, depending on
the required accuracy, the nature of behavior, the video quality (e.g. motion blur, bad lighting) and the context, more
or less frames might be necessary to create a good network. Ultimately, in order to scale up the analysis to large
collections of videos with perhaps unexpected conditions, one can also refine the dataset in an adaptive way (see
refinement below).

The function extract_frames extracts frames from all the videos in the project configuration file in order to create a training dataset. The extracted frames from all the videos are stored in a separate subdirectory named after the video file’s name under the ‘labeled-data’. This function also has various parameters that might be useful based on the user’s need.

deeplabcut.extract_frames(
    config_path,
    mode="automatic/manual",
    algo="uniform/kmeans",
    crop=True/False,
    userfeedback=False
)

It is advisable to keep the frame size small, as large frames increase the training and
inference time. The cropping parameters for each video can be provided in the config.yaml file (and see below).
When running the function `extract_frames`, if the parameter `crop=True`, then you will be asked to draw a box within the
GUI (and this is written to the config.yaml file).

userfeedback allows the user to specify which videos they wish to extract frames from. When set to True, a dialog will appear, where the user is asked for each video if (additional/any) frames from this video should be extracted. Use this, e.g. if you have already labeled some folders and want to extract data for new videos.

The provided function either selects frames from the videos that are randomly sampled from a uniform distribution (uniform), by clustering based on visual appearance (k-means), or by manual selection. Random uniform selection of frames works best for behaviors where the postures vary across the whole video. However, some behaviors might be sparse, as in the case of reaching where the reach and pull are very fast and the mouse is not moving much between trials. In such a case, the function that allows selecting frames based on k-means derived quantization would be useful. If the user chooses to use k-means as a method to cluster the frames, then this function downsamples the video and clusters the frames using k-means, where each frame is treated as a vector. Frames from different clusters are then selected. This procedure makes sure that the frames look different. However, on large and long videos, this code is slow due to computational complexity.

It is advisable to extract frames from a period of the video that contains interesting
behaviors, and not extract the frames across the whole video. This can be achieved by using the start and stop
parameters in the config.yaml file.
Also, the user can change the number of frames to extract from each video using the numframes2pick in the config.yaml file.

However, picking frames is highly dependent on the data and the behavior being studied. Therefore, it is hard to provide one-size-fits-all code that extracts frames to create a good training dataset for every behavior and animal. If the user feels specific frames are lacking, they can extract hand selected frames of interest using the interactive GUI provided along with the toolbox. This can be launched by using:

deeplabcut.extract_frames(config_path, "manual")

The user can use the Load Video button to load one of the videos in the project configuration file, use the scroll bar to navigate across the video and grab a frame or a range of frames to extract the frame(s) (see {numref}fig-manual-frame-selection). The user can also look at the extracted frames and e.g. delete frames (from the directory) that are too similar before reloading the set and then manually annotating them.

---
name: fig-manual-frame-selection
alt: Manual frame selection using the extract_frames GUI
width: 70%
align: center
---
Manual frame selection using the `extract_frames` GUI.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.extract_frames.rst
```

The toolbox provides a function label_frames which helps the user to easily label all the extracted frames using an interactive graphical user interface (GUI). The user should have already named the bodyparts to label (points of interest) in the project’s configuration file by providing a list. The following command invokes the napari-deeplabcut labelling GUI.

Check out the {ref}`napari-deeplabcut docs <file:napari-gui-landing>` for
more information about the labelling workflow.

deeplabcut.label_frames(config_path)

🎥 DEMO

It is advisable to **consistently label similar spots** (e.g., on a wrist that is very large, try
to label the same location). In general, invisible or occluded points should not be labeled by the user. They can
simply be skipped by not applying the label anywhere on the frame.

To add more labels to the existing labeled dataset, the user needs to append the new labels to the bodyparts in the config.yaml file. Thereafter, the user can call the function label_frames. A box will pop up and ask the user if they wish to display all parts, or only add in the new labels. Saving the labels after all the images are labelled will append the new labels to the existing labeled dataset.

For more information about the labelling workflow, check out the {ref}napari-deeplabcut docs <file:napari-gui-landing>.

Checking if the labels were created and stored correctly is beneficial for training, since labeling is one of the most critical parts for creating the training dataset. The DeepLabCut toolbox provides a function check_labels to do so. It is used as follows:

deeplabcut.check_labels(config_path, visualizeindividuals=True/False)

For each video directory in labeled-data this function creates a subdirectory with labeled as a suffix. Those directories contain the frames plotted with the annotated body parts. The user can double check if the body parts are labeled correctly. If they are not correct, the user can reload the frames (i.e. deeplabcut.label_frames), move them around, and click save again.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.check_labels.rst
```

(create-training-dataset)=

Only run this step **where** you are going to train the network. If you label on your laptop but
move your project folder to Google Colab or AWS, lab server, etc, then run the step below on that platform! If you
labeled on a Windows machine but train on Linux, this is handled automatically (it
saves file sets as both Linux and Windows for you).

• If you move your project folder, you must only change the project_path (which is done automatically) in the main config.yaml file - that's it - no need to change the video paths, etc! Your project is fully portable.

• Be aware you select your neural network backbone at this stage. As of DLC3+ we support PyTorch (and TensorFlow, but this will be phased out).

This function combines the labeled datasets from all the videos and splits them to create train and test datasets. The training data will be used to train the network, while the test dataset will be used for evaluating the network.

deeplabcut.create_training_dataset(config_path)

• Optional: If the user wishes to benchmark the performance of different training settings, they can create multiple training datasets by specifying an integer value for num_shuffles; see the docstring for more details.

The function creates a new shuffle(s) directory in the dlc-models-pytorch directory (dlc-models if using TensorFlow), in the current "iteration" directory. The train and test directories each have a configuration file (pytorch_config.yaml in train and pose_cfg.yaml in test for PyTorch models, pose_cfg.yaml in train and test for TensorFlow models). Specifically, the user can edit the pytorch_config.yaml (or pose_cfg.yaml) within the train subdirectory before starting the training. These configuration files contain meta information with regard to the parameters of the feature detectors. For more information about the pytorch_config.yaml file, see here (for TensorFlow-based models, see key parameters here).

A schematic view of the structure described above is:

<iteration directory>/
└── <shuffle(s) directory>/
    ├── train/
    │   └── pytorch_config.yaml or pose_cfg.yaml
    └── test/
        └── pose_cfg.yaml

At this step, for create_training_dataset you select the network you want to use, and any additional data augmentation (beyond our defaults). You can set net_type, detector_type (if using a detector) and augmenter_type when you call the function.

• Networks: ImageNet pre-trained networks OR SuperAnimal pre-trained network weights will be downloaded, as you select. You can decide to do transfer-learning (recommended) or "fine-tune" both the backbone and the decoder head. We suggest seeing our dedicated documentation on models for more information ( or the this page on selecting models for the TensorFlow engine).

🚨 If they do not download (you will see this downloading in the terminal), then you may not have permission to do
so - be sure to open your terminal "as an admin" (This is only something we have seen with some Windows users - see
the **[docs for more help!](tf-training-tips-and-tricks)**).

Data augmentation: At this stage you can also decide what type of augmentation to use. Once you've called create_training_dataset, you can edit the pytorch_config.yaml file that was created (or for the TensorFlow engine, the pose_cfg.yaml file).

• PyTorch Engine: Albumentations is used for data augmentation. Look at the pytorch_config.yaml for more information about image augmentation options.
• TensorFlow Engine: The default augmentation works well for most tasks (as shown on www.deeplabcut.org), but there are many options, more data augmentation, intermediate supervision, etc. Here are the available loaders:
  • imgaug: a lot of augmentation possibilities, efficient code for target map creation & batch sizes >1 supported. You can set the parameters such as the batch_size in the pose_cfg.yaml file for the model you are training. This is the recommended default!
  • crop_scale: our standard DLC 2.0 introduced in Nature Protocols variant (scaling, auto-crop augmentation)
  • tensorpack: a lot of augmentation possibilities, multi CPU support for fast processing, target maps are created less efficiently than in imgaug, does not allow batch size>1
  • deterministic: only useful for testing, freezes numpy seed; otherwise like default.

You can also test several models by creating the same train/test split for different networks. You can easily do this in the Project Manager GUI (by selecting the "Use an existing data split" option), which also lets you compare PyTorch and TensorFlow models.

You can now create new shuffles using the same train/test split as
existing shuffles with `create_training_dataset_from_existing_split`. This allows you to
compare model performance (between different architectures or when using different
training hyper-parameters) as the shuffles were trained on the same data, and evaluated
on the same test data!

Example usage - creating 3 new shuffles (with indices 10, 11 and 12) for a ResNet 50
pose estimation model, using the same data split as was used for shuffle 0:

```python
deeplabcut.create_training_dataset_from_existing_split(
    config_path,
    from_shuffle=0,
    shuffles=[10, 11, 12],
    net_type="resnet_50",
)
```

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.create_training_dataset.rst
```

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.create_training_model_comparison.rst
```

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.create_training_dataset_from_existing_split.rst
```

The function train_network helps the user in training the network. It is used as follows:

deeplabcut.train_network(config_path)

The set of arguments in the function starts training the network for the dataset created for one specific shuffle. Note that you can change training parameters in the pytorch_config.yaml file (or pose_cfg.yaml for TensorFlow models) of the model that you want to train (before you start training).

At user-specified iterations during training checkpoints are stored in the subdirectory train under the respective iteration & shuffle directory.

---
class: dropdown
---
Example parameters that one can call:

```python
deeplabcut.train_network(
    config_path,
    shuffle=1,
    trainingsetindex=0,
    device="cuda:0",
    max_snapshots_to_keep=5,
    displayiters=100,
    save_epochs=5,
    epochs=200,
)
```

PyTorch models in DeepLabCut 3.0 are trained for a set number of epochs, instead of a
maximum number of iterations (which is what was used for TensorFlow models). An epoch
is a single pass through the training dataset, which means your model has seen each
training image exactly once. So if you have 64 training images for your network, an
epoch is 64 iterations with batch size 1 (or 32 iterations with batch size 2, 16 with
batch size 4, etc.).

By default, the pretrained networks are not in the DeepLabCut toolbox (as they can be
more than 100MB), but they get downloaded automatically before you train.

If the user wishes to restart the training at a specific checkpoint they can specify the
full path of the checkpoint to the variable ``resume_training_from`` in the [
**pytorch_config.yaml**](
dlc3-pytorch-config) file (checkout the "Restarting Training at a Specific Checkpoint"
section of the docs) under the *train* subdirectory.

**Tip**: It is recommended to train the networks **until the loss plateaus**
(depending on the dataset, model architecture and training hyper-parameters this happens
after 100 to 250 epochs of training).

The variables ``display_iters`` and ``save_epochs`` in the [**pytorch_config.yaml**](
dlc3-pytorch-config) file allow the user to alter how often the loss is displayed
and how often the weights are stored. We suggest saving every 5 to 25 epochs.

---
class: dropdown
---
Example parameters that one can call:

```python
deeplabcut.train_network(
    config_path,
    shuffle=1,
    trainingsetindex=0,
    gputouse=None,
    max_snapshots_to_keep=5,
    autotune=False,
    displayiters=100,
    saveiters=25000,
    maxiters=300000,
    allow_growth=True,
)
```

By default, the pretrained networks are not in the DeepLabCut toolbox (as they are
around 100MB each), but they get downloaded before you train. However, if not previously
downloaded from the TensorFlow model weights, it will be downloaded and stored in a
subdirectory *pre-trained* under the subdirectory *models* in
*Pose_Estimation_Tensorflow*. At user specified iterations during training checkpoints
are stored in the subdirectory *train* under the respective iteration directory.

If the user wishes to restart the training at a specific checkpoint they can specify the
full path of the checkpoint to the variable ``init_weights`` in the **pose_cfg.yaml**
file under the *train* subdirectory (see {ref}`Box 2 <pose-cfg-box2>`).

**Tip**: It is recommended to train the networks for thousands of iterations
until the loss plateaus (typically around **500,000**) if you use batch size 1. If you
want to batch train, we recommend using Adam,
[see more here](tf-custom-image-augmentation).

The variables ``display_iters`` and ``save_iters`` in the **pose_cfg.yaml** file allows
the user to alter how often the loss is displayed and how often the weights are stored.

**maDeepLabCut recommendation:** For multi-animal projects we are using not only
different and new output layers, but also new data augmentation, optimization, learning
rates, and batch training defaults. Thus, please use a lower ``save_iters`` and
``maxiters``. I.e. we suggest saving every 10K-15K iterations, and only training until
50K-100K iterations. We recommend you look closely at the loss to not overfit on your
data. This will reduce your training time.

```{figure} images/box2-single.png
---
name: pose-cfg-box2
alt: Box 2 - Single Animal TensorFlow Configuration File Glossary
align: center
---
Single-animal TensorFlow configuration file glossary
```

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.train_network.rst
```

It is important to evaluate the performance of the trained network. This performance is measured by computing the average root mean square error (RMSE) between the manual labels and the ones predicted by DeepLabCut. The RMSE is saved as a comma-separated file and displayed for all pairs and only likely pairs (>p-cutoff). This helps to exclude, for example, occluded body parts. One of the strengths of DeepLabCut is that due to the probabilistic output of the scoremap, it can, if sufficiently trained, also reliably report if a body part is visible in a given frame. (see discussions of finger tips in reaching and the Drosophila legs during 3D behavior in Mathis et al, 2018). The evaluation results are computed by typing:

deeplabcut.evaluate_network(config_path, Shuffles=[1], plotting=True)

Setting plotting to true plots all the testing and training frames with the manual and predicted labels. The user should visually check the labeled test (and training) images that are created in the ‘evaluation-results’ directory. Ideally, DeepLabCut labeled unseen (test images) according to the user’s required accuracy, and the average train and test errors are comparable (good generalization). What (numerically) comprises an acceptable RMSE depends on many factors (including the size of the tracked body parts, the labeling variability, etc.). Note that the test error can also be larger than the training error due to human variability (in labeling, see Figure 2 in Mathis et al, Nature Neuroscience 2018).

• Shuffles: list, optional - List of integers specifying the shuffle indices of the training dataset. The default is [1]

• plotting: bool, optional - Plots the predictions on the train and test images. The default is False; if provided it must be either True or False

• show_errors: bool, optional - Display train and test errors. The default is True

• comparisonbodyparts: list of bodyparts, Default is all - The average error will be computed for those body parts only (Has to be a subset of the body parts).

• gputouse: int, optional - Natural number indicating the number of your GPU (see number in nvidia-smi). If you do not have a GPU, put None. See: https://nvidia.custhelp.com/app/answers/detail/a_id/3751/~/useful-nvidia-smi-queries

• pcutoff: float | list[float] | dict[str, float], optional (Only applicable when using the PyTorch engine. For TensorFlow, set pcutoff in the config.yaml file.) Specifies the cutoff value(s) used to compute evaluation metrics.

  • If None (default), the cutoff will be loaded from the project configuration.
  • To apply a single cutoff value to all bodyparts, provide a float.
  • To specify different cutoffs per bodypart, provide either:
    • A list[float]: one value per bodypart, with an additional value for each unique bodypart if applicable.
    • A dict[str, float]: where keys are bodypart names and values are the corresponding cutoff values. If a bodypart is not included in the provided dictionary, a default pcutoff of 0.6 will be used for that bodypart.

The plots can be customized by editing the config.yaml file (i.e., the colormap, scale, marker size (dotsize), and transparency of labels (alphavalue) can be modified). By default each body part is plotted in a different color (governed by the colormap) and the plot labels indicate their source. Note that by default the human labels are plotted as plus (‘+’), DeepLabCut’s predictions either as ‘.’ (for confident predictions with likelihood > p-cutoff) and ’x’ for (likelihood <= pcutoff).

The evaluation results for each shuffle of the training dataset are stored in a unique subdirectory in a newly created directory ‘evaluation-results-pytorch’ (‘evaluation-results’ for TensorFlow models) in the project directory. The user can visually inspect if the distance between the labeled and the predicted body parts are acceptable. In the event of benchmarking with different shuffles of same training dataset, the user can provide multiple shuffle indices to evaluate the corresponding network. Note that with multi-animal projects additional distance statistics aggregated over animals or bodyparts are also stored in that directory. This aims at providing a finer quantitative evaluation of multi-animal prediction performance before animal tracking. If the generalization is not sufficient, the user might want to:

• Check if the labels were imported correctly; i.e., invisible points are not labeled and the points of interest are labeled accurately
• Make sure that the loss has already converged
• Consider labeling additional images and make another iteration of the training dataset

Optional: You can also plot the scoremaps, locref layers, and PAFs:

deeplabcut.extract_save_all_maps(config_path, shuffle=shuffle, Indices=[0, 5])

you can drop Indices to run this on all training/testing images (this is slow!)

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.evaluate_network.rst
```

The trained network can be used to analyze new videos. Novel/new videos do not need to be in the config file!. You can analyze new videos anytime by simply using the following line of code:

deeplabcut.analyze_videos(
    config_path, ["fullpath/analysis/project/videos/reachingvideo1.avi"],
    save_as_csv=True
)

There are several other optional inputs, such as:

deeplabcut.analyze_videos(
    config_path,
    videos,
    videotype="avi",
    shuffle=1,
    trainingsetindex=0,
    gputouse=None,
    save_as_csv=False,
    destfolder=None,
    dynamic=(True, .5, 10)
)

The user can choose a checkpoint for analyzing the videos. For this, the user can enter the corresponding index of the checkpoint to the variable snapshotindex in the config.yaml file. By default, the most recent checkpoint (i.e. last) is used for analyzing the video. The labels are stored in a MultiIndex Pandas Array, which contains the name of the network, body part name, (x, y) label position in pixels, and the likelihood for each frame per body part. These arrays are stored in an efficient Hierarchical Data Format (HDF) in the same directory, where the video is stored. However, if the flag save_as_csv is set to True, the data can also be exported in comma-separated values format (.csv), which in turn can be imported in many programs, such as MATLAB, R, Prism, etc.; this flag is set to False by default. You can also set a destination folder (destfolder) for the output files by passing a path of the folder you wish to write to.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.analyze_videos.rst
```

As of 2.1+ we have a dynamic cropping option. Namely, if you have large frames and the animal/object occupies a smaller fraction, you can crop around your animal/object to make processing speeds faster. For example, if you have a large open field experiment but only track the mouse, this will speed up your analysis (also helpful for real-time applications). To use this simply add dynamic=(True,.5,10) when you call analyze_videos.

"""
dynamic: tuple containing (state, detectionthreshold, margin)
"""

If state is True, then dynamic cropping will be performed. That means that if an object is detected (i.e., any body part likelihood > detectionthreshold), then object boundaries are computed according to the smallest/largest x position and smallest/largest y position of all body parts. This window is expanded by margin and from then on only the posture within this crop is analyzed (until the object is lost; i.e., < detectionthreshold). The current position is utilized for updating the crop window for the next frame (this is why the margin is important and should be set large enough given the movement of the animal).

You can also filter the predictions with a median filter (default) or with a SARIMAX model, if you wish. This creates a new .h5 file with the ending _filtered that you can use in create_labeled_video and/or plot_trajectories.

deeplabcut.filterpredictions(
    config_path,
    ["fullpath/analysis/project/videos/reachingvideo1.avi"]
)

An example call:

deeplabcut.filterpredictions(
    config_path,
    ["fullpath/analysis/project/videos"],
    videotype=".mp4",
    filtertype="arima",
    ARdegree=5,
    MAdegree=2
)

Here are parameters you can modify and pass:

deeplabcut.filterpredictions(
    config_path,
    ["fullpath/analysis/project/videos/reachingvideo1.avi"],
    shuffle=1,
    trainingsetindex=0,
    filtertype="arima",
    p_bound=0.01,
    ARdegree=3,
    MAdegree=1,
    alpha=0.01
)

Here is an example of how this can be applied to a video:

---
name: fig-filter-example
alt: Example output of filterpredictions applied to a video
width: 70%
align: center
---
Example output of `filterpredictions` applied to a video.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.filterpredictions.rst
```

The plotting components of this toolbox utilize matplotlib. Therefore, these plots can easily be customized by the end user. We also provide a function to plot the trajectory of the extracted poses across the analyzed video (see {numref}fig-trajectory-frame and {numref}fig-trajectory-plots).

deeplabcut.plot_trajectories(config_path, ['fullpath/analysis/project/videos/reachingvideo1.avi'])

It creates a folder called plot-poses (in the directory of the video). The plots display the coordinates of body parts vs. time, likelihoods vs time, the x- vs. y- coordinate of the body parts, as well as histograms of consecutive coordinate differences. These plots help the user to quickly assess the tracking performance for a video. Ideally, the likelihood stays high and the histogram of consecutive coordinate differences has values close to zero (i.e. no jumps in body part detections across frames). Example outputs are shown below.

---
name: fig-trajectory-frame
alt: Example video frame with tracked body parts overlaid
height: 240px
align: center
---
Example video frame with tracked body parts overlaid.

---
name: fig-trajectory-plots
alt: Example plot_trajectories output
height: 250px
align: center
---
Example `plot_trajectories` output: body part coordinates, likelihoods, and consecutive displacement histograms.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.plot_trajectories.rst
```

Additionally, the toolbox provides a function to create labeled videos based on the extracted poses by plotting the labels on top of the frame and creating a video. There are two modes to create videos: 'fast' and 'slow' (but higher quality). One can use the command as follows to create multiple labeled videos:

deeplabcut.create_labeled_video(
    config_path,
    ["fullpath/analysis/project/videos/reachingvideo1.avi",
     "fullpath/analysis/project/videos/reachingvideo2.avi"],
    save_frames = True/False
)

Optionally, if you want to use the filtered data for a video or directory of filtered videos pass filtered=True, i.e.:

deeplabcut.create_labeled_video(
    config_path,
    ["fullpath/afolderofvideos"],
    videotype=".mp4",
    filtered=True
)

You can also optionally add a skeleton to connect points and/or add a history of points for visualization (see {numref}fig-skeleton-trail). To set the "trailing points" you need to pass trailpoints:

deeplabcut.create_labeled_video(
    config_path,
    ["fullpath/afolderofvideos"],
    videotype=".mp4",
    trailpoints=10
)

To draw a skeleton, you need to first define the pairs of connected nodes (in the config.yaml file) and set the skeleton color (in the config.yaml file). There is also a GUI to help you do this, used by calling deeplabcut.SkeletonBuilder(config_path), where config_path is the path to your project's config.yaml file on disk.

Here is how the config.yaml additions/edits should look (for example, on the Openfield demo data we provide):

# Plotting configuration
skeleton:
  - ["snout", "leftear"]
  - ["snout", "rightear"]
  - ["leftear", "tailbase"]
  - ["leftear", "rightear"]
  - ["rightear", "tailbase"]
skeleton_color: white
pcutoff: 0.4
dotsize: 4
alphavalue: 0.5
colormap: jet

Then pass draw_skeleton=True with the command:

deeplabcut.create_labeled_video(
    config_path,
    ["fullpath/afolderofvideos"],
    videotype=".mp4",
    draw_skeleton=True
)

You can create a video with only the "dots" plotted, i.e., in the style of Johansson, by passing keypoints_only=True:

deeplabcut.create_labeled_video(
    config_path,["fullpath/afolderofvideos"],
    videotype=".mp4",
    keypoints_only=True
)

The best quality videos are created when `fastmode=False` is passed. Therefore, when
`trailpoints` and `draw_skeleton` are used, we **highly** recommend you also pass `fastmode=False`!

---
name: fig-skeleton-trail
alt: Labeled video with skeleton overlay and trailing points
width: 40%
align: center
---
Labeled video with skeleton overlay and trailing points (`draw_skeleton=True`, `trailpoints=10`).

This function has various other parameters, in particular the user can set the colormap, the dotsize, and alphavalue of the labels in config.yaml file.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.create_labeled_video.rst
```

You can save the "skeleton" that was applied in create_labeled_videos for more computations. Namely, it extracts length and orientation of each "bone" of the skeleton as defined in the config.yaml file. You can use the function by:

deeplabcut.analyzeskeleton(
    config,
    video,
    videotype="avi",
    shuffle=1,
    trainingsetindex=0,
    save_as_csv=False,
    destfolder=None
)

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.analyzeskeleton.rst
```

(active-learning)=

While DeepLabCut typically generalizes well across datasets, one might want to optimize its performance in various, perhaps unexpected, situations. For generalization to large datasets, images with insufficient labeling performance can be extracted, manually corrected by adjusting the labels to increase the training set and iteratively improve the feature detectors. Such an active learning framework can be used to achieve a predefined level of confidence for all images with minimal labeling cost (discussed in Mathis et al 2018). Then, due to the large capacity of the neural network that underlies the feature detectors, one can continue training the network with these additional examples. One does not necessarily need to correct all errors as common errors could be eliminated by relabeling a few examples and then re-training. A priori, given that there is no ground truth data for analyzed videos, it is challenging to find putative “outlier frames”. However, one can use heuristics such as the continuity of body part trajectories, to identify images where the decoder might make large errors.

All this can be done for a specific video by typing (see other optional inputs below):

deeplabcut.extract_outlier_frames(config_path, ["videofile_path"])

We provide various frame-selection methods for this purpose. In particular the user can set:

outlieralgorithm: "fitting", "jump", or "uncertain"

• outlieralgorithm="uncertain": select frames if the likelihood of a particular or all body parts lies below p_bound (note this could also be due to occlusions rather than errors).

• outlieralgorithm="jump": select frames where a particular body part or all body parts jumped more than epsilon pixels from the last frame.

• outlieralgorithm="fitting": select frames if the predicted body part location deviates from a state-space model fit to the time series of individual body parts. Specifically, this method fits an Auto Regressive Integrated Moving Average (ARIMA) model to the time series for each body part. Thereby each body part detection with a likelihood smaller than p_bound is treated as missing data. Putative outlier frames are then identified as time points, where the average body part estimates are at least epsilon pixels away from the fits. The parameters of this method are epsilon, p_bound, the ARIMA parameters as well as the list of body parts to average over (can also be all).

• outlieralgorithm="manual": manually select outlier frames based on visual inspection from the user.

As an example:

deeplabcut.extract_outlier_frames(config_path, ["videofile_path"], outlieralgorithm="manual")

In general, depending on the parameters, these methods might return many more frames than the user wants to extract (numframes2pick). Thus, this list is then used to select outlier frames either by randomly sampling from this list (extractionalgorithm="uniform"), by performing extractionalgorithm="kmeans" clustering on the corresponding frames.

In the automatic configuration, before the frame selection happens, the user is informed about the amount of frames satisfying the criteria and asked if the selection should proceed. This step allows the user to perhaps change the parameters of the frame-selection heuristics first (i.e. to make sure that not too many frames are qualified). The user can run the extract_outlier_frames method iteratively, and (even) extract additional frames from the same video. Once enough outlier frames are extracted the refinement GUI can be used to adjust the labels based on user feedback (see below).

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.extract_outlier_frames.rst
```

Based on the performance of DeepLabCut, four scenarios are possible:

• (A) Visible body part with accurate DeepLabCut prediction. These labels do not need any modifications.

• (B) Visible body part but wrong DeepLabCut prediction. Move the label’s location to the actual position of the body part.

• (C) Invisible, occluded body part. Remove the predicted label by DeepLabCut with a middle click. Every predicted label is shown, even when DeepLabCut is uncertain. This is necessary, so that the user can potentially move the predicted label. However, to help the user to remove all invisible body parts the low-likelihood predictions are shown as open circles (rather than disks).

• (D) Invalid images: In the unlikely event that there are any invalid images, the user should remove such an image and their corresponding predictions, if any. Here, the GUI will prompt the user to remove an image identified as invalid.

The labels for extracted putative outlier frames can be refined by opening the GUI:

deeplabcut.refine_labels(config_path)

This will launch a GUI where the user can refine the labels.

Please refer to the {ref}napari-deeplabcut docs <file:napari-gui-landing> for more information about the labelling workflow.

After correcting the labels for all the frames in each of the subdirectories, the users should merge the dataset to create a new dataset. In this step the iteration parameter in the config.yaml file is automatically updated.

deeplabcut.merge_datasets(config_path)

Once the dataset is merged, the user can test if the merging process was successful by plotting all the labels (Step E). Next, with this expanded training set the user can now create a new training set and train the network as described in Steps F and G. The training dataset will be stored in the same place as before but under a different iteration-# subdirectory, where the # is the new value of iteration variable stored in the project’s configuration file (this is automatically done).

Now you can run create_training_dataset, then train_network, etc. If your original labels were adjusted at all, start from fresh weights (which is generally recommended), otherwise consider using your already trained network weights (see {ref}Box 2 <pose-cfg-box2>).

If after training the network generalizes well to the data, proceed to analyze new videos. Otherwise, consider labeling more data.

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.refine_labels.rst
```

---
class: dropdown
---
```{eval-rst}
.. include:: ./api/deeplabcut.merge_datasets.rst
```

We also provide two Jupyter notebooks for using DeepLabCut on both a pre-labeled dataset, and on the end user’s own dataset.

• Firstly, we prepared an interactive Jupyter notebook called Demo_yourowndata.ipynb that can serve as a template for the user to develop a project.
• Furthermore, we provide a notebook for an already started project with labeled data. The example project, named as Reaching-Mackenzie-2018-08-30 consists of a project configuration file with default parameters and 20 images, which are cropped around the region of interest as an example dataset. These images are extracted from a video, which was recorded in a study of skilled motor control in mice. Some example labels for these images are also provided. See more details here.

For stereo or multi-camera setups, see the {ref}3D overview <3D-overview>.

A collection of optional utility functions is available in the {ref}helper functions reference <helper-functions>.