moseq_infer.py

Code adapted from the Datta Lab DataJoint Schema for Keypoint-MoSeq inference pipeline

activate(infer_schema_name, *, create_schema=True, create_tables=True, linking_module=None)

Activate this schema.

Parameters:

Name	Type	Description	Default
infer_schema_name	str	Schema name on the database server to activate the moseq_infer schema.	required
create_schema	bool	When True (default), create schema in the database if it does not yet exist.	True
create_tables	bool	When True (default), create schema tables in the database if they do not yet exist.	True
linking_module	str	A module (or name) containing the required dependencies.	None

Functions:

Name	Description
get_kpms_root_data_dir	Returns absolute path for root data director(y/ies) with all behavioral recordings, as (list of) string(s)
get_kpms_processed_data_dir	Optional. Returns absolute path for processed data.

Source code in element_moseq/moseq_infer.py

def activate(
    infer_schema_name: str,
    *,
    create_schema: bool = True,
    create_tables: bool = True,
    linking_module: str = None,
):
    """Activate this schema.

    Args:
        infer_schema_name (str): Schema name on the database server to activate the `moseq_infer` schema.
        create_schema (bool): When True (default), create schema in the database if it
                            does not yet exist.
        create_tables (bool): When True (default), create schema tables in the database
                             if they do not yet exist.
        linking_module (str): A module (or name) containing the required dependencies.

    Functions:
        get_kpms_root_data_dir(): Returns absolute path for root data director(y/ies) with all behavioral recordings, as (list of) string(s)
        get_kpms_processed_data_dir(): Optional. Returns absolute path for processed data.
    """

    if isinstance(linking_module, str):
        linking_module = importlib.import_module(linking_module)
    assert inspect.ismodule(
        linking_module
    ), "The argument 'dependency' must be a module's name or a module"
    assert hasattr(
        linking_module, "get_kpms_root_data_dir"
    ), "The linking module must specify a lookup function for a root data directory"

    global _linking_module
    _linking_module = linking_module

    # activate
    schema.activate(
        infer_schema_name,
        create_schema=create_schema,
        create_tables=create_tables,
        add_objects=_linking_module.__dict__,
    )

Model

Bases: Manual

Register a model.

Attributes:

Name	Type	Description
model_id	int)	Unique ID for each model.
model_name	varchar)	User-friendly model name.
model_dir	varchar)	Model directory relative to root data directory (e.g. kpms_project_output_dir/2024_03_21-00_51_39)
latent_dim	int)	Latent dimension of the model.
kappa	float)	Kappa value of the model.
model_desc	varchar)	Optional. User-defined description of the model

Source code in element_moseq/moseq_infer.py

@schema
class Model(dj.Manual):
    """Register a model.

    Attributes:
        model_id (int)                      : Unique ID for each model.
        model_name (varchar)                : User-friendly model name.
        model_dir (varchar)                 : Model directory relative to root data directory (e.g. `kpms_project_output_dir/2024_03_21-00_51_39`)
        latent_dim (int)                    : Latent dimension of the model.
        kappa (float)                       : Kappa value of the model.
        model_desc (varchar)                : Optional. User-defined description of the model

    """

    definition = """
    model_id                : int          # Unique ID for each model
    ---
    model_name              : varchar(64)  # User-friendly model name
    model_dir               : varchar(1000)# Model directory relative to root data directory
    model_desc=''           : varchar(1000)# Optional. User-defined description of the model
    -> [nullable] moseq_train.SelectedFullFit
    """

VideoRecording

Bases: Manual

Set of video recordings for the Keypoint-MoSeq inference.

Attributes:

Name	Type	Description
Session	foreign key)	Session key.
recording_id	int)	Unique ID for each recording.
Device	foreign key)	Device primary key.

Source code in element_moseq/moseq_infer.py

@schema
class VideoRecording(dj.Manual):
    """Set of video recordings for the Keypoint-MoSeq inference.

    Attributes:
        Session (foreign key)               : `Session` key.
        recording_id (int)                  : Unique ID for each recording.
        Device (foreign key)                : Device primary key.
    """

    definition = """
    -> Session                             # `Session` key
    recording_id: int                      # Unique ID for each recording
    ---
    -> Device                              # Device primary key
    """

    class File(dj.Part):
        """File IDs and paths associated with a given `recording_id`.

        Attributes:
            VideoRecording (foreign key)   : `VideoRecording` key.
            file_id(int)                   : Unique ID for each file.
            file_path (varchar)            : Filepath of each video, relative to root data directory.
        """

        definition = """
        -> master
        file_id: int             # Unique ID for each file
        ---
        file_path: varchar(1000) # Filepath of each video, relative to root data directory.
        """

File

Bases: Part

File IDs and paths associated with a given recording_id.

Attributes:

Name	Type	Description
VideoRecording	foreign key)	VideoRecording key.
file_id(int)		Unique ID for each file.
file_path	varchar)	Filepath of each video, relative to root data directory.

Source code in element_moseq/moseq_infer.py

class File(dj.Part):
    """File IDs and paths associated with a given `recording_id`.

    Attributes:
        VideoRecording (foreign key)   : `VideoRecording` key.
        file_id(int)                   : Unique ID for each file.
        file_path (varchar)            : Filepath of each video, relative to root data directory.
    """

    definition = """
    -> master
    file_id: int             # Unique ID for each file
    ---
    file_path: varchar(1000) # Filepath of each video, relative to root data directory.
    """

InferenceTask

Bases: Manual

Staging table to define the Inference task and its output directory.

Attributes:

Name	Type	Description
VideoRecording	foreign key)	VideoRecording key
Model	foreign key)	Model key
PoseEstimationMethod	foreign key)	Pose estimation method used for the specified recording_id.
inference_output_dir	varchar)	Optional. Sub-directory where the results will be stored.
inference_desc	varchar)	Optional. User-defined description of the inference task.
num_iterations	int)	Optional. Number of iterations to use for the model inference. If null, the default number internally is 50.

Source code in element_moseq/moseq_infer.py

@schema
class InferenceTask(dj.Manual):
    """Staging table to define the Inference task and its output directory.

    Attributes:
        VideoRecording (foreign key)         : `VideoRecording` key
        Model (foreign key)                  : `Model` key
        PoseEstimationMethod (foreign key)   : Pose estimation method used for the specified `recording_id`.
        inference_output_dir (varchar)       : Optional. Sub-directory where the results will be stored.
        inference_desc (varchar)             : Optional. User-defined description of the inference task.
        num_iterations (int)                 : Optional. Number of iterations to use for the model inference. If null, the default number internally is 50.
    """

    definition = """
    -> VideoRecording                                       # `VideoRecording` key
    -> Model                                                # `Model` key
    ---
    -> moseq_train.PoseEstimationMethod                     # Pose estimation method used for the specified `recording_id`
    keypointset_dir               : varchar(1000)           # Keypointset directory for the specified VideoRecording
    inference_output_dir=''       : varchar(1000)           # Optional. Sub-directory where the results will be stored
    inference_desc=''             : varchar(1000)           # Optional. User-defined description of the inference task
    num_iterations=NULL           : int                     # Optional. Number of iterations to use for the model inference. If null, the default number internally is 50.
    task_mode='load'              : enum('load', 'trigger') # Task mode for the inference task
    """

Inference

Bases: Computed

Infer the model from the checkpoint file and save the results as results.h5 file.

Attributes:

Name	Type	Description
InferenceTask	foreign_key)	InferenceTask key.
inference_duration	float)	Time duration (seconds) of the inference computation.

Source code in element_moseq/moseq_infer.py

@schema
class Inference(dj.Computed):
    """Infer the model from the checkpoint file and save the results as `results.h5` file.

    Attributes:
        InferenceTask (foreign_key)         : `InferenceTask` key.
        inference_duration (float)          : Time duration (seconds) of the inference computation.
    """

    definition = """
    -> InferenceTask                        # `InferenceTask` key
    ---
    inference_duration=NULL        : float  # Time duration (seconds) of the inference computation
    """

    class MotionSequence(dj.Part):
        """Store the results of the model inference.

        Attributes:
            video_name (varchar)                : Name of the video.
            syllable (longblob)                 : Syllable labels (z). The syllable label assigned to each frame (i.e. the state indexes assigned by the model).
            latent_state (longblob)             : Inferred low-dim pose state (x). Low-dimensional representation of the animal's pose in each frame. These are similar to PCA scores, are modified to reflect the pose dynamics and noise estimates inferred by the model.
            centroid (longblob)                 : Inferred centroid (v). The centroid of the animal in each frame, as estimated by the model.
            heading (longblob)                  : Inferred heading (h). The heading of the animal in each frame, as estimated by the model.
        """

        definition = """
        -> master
        video_name      : varchar(150)    # Name of the video
        ---
        syllable        : longblob        # Syllable labels (z). The syllable label assigned to each frame (i.e. the state indexes assigned by the model)
        latent_state    : longblob        # Inferred low-dim pose state (x). Low-dimensional representation of the animal's pose in each frame. These are similar to PCA scores, are modified to reflect the pose dynamics and noise estimates inferred by the model
        centroid        : longblob        # Inferred centroid (v). The centroid of the animal in each frame, as estimated by the model
        heading         : longblob        # Inferred heading (h). The heading of the animal in each frame, as estimated by the model
        """

    class GridMoviesSampledInstances(dj.Part):
        """Store the sampled instances of the grid movies.

        Attributes:
            syllable (int)                  : Syllable label.
            instances (longblob)            : List of instances shown in each in grid movie (in row-major order), where each instance is specified as a tuple with the video name, start frame and end frame.
        """

        definition = """
        -> master
        syllable: int           # Syllable label
        ---
        instances: longblob     # List of instances shown in each in grid movie (in row-major order), where each instance is specified as a tuple with the video name, start frame and end frame
        """

    def make(self, key):
        """
        This function is used to infer the model results from the checkpoint file and store the results in `MotionSequence` and `GridMoviesSampledInstances` tables.

        Args:
            key (dict): `InferenceTask` primary key.

        Raises:
            FileNotFoundError: If no pca model (`pca.p`) found in the parent model directory.
            FileNotFoundError: If no model (`checkpoint.h5`) found in the model directory.
            NotImplementedError: If the format method is not `deeplabcut`.
            FileNotFoundError: If no valid `kpms_dj_config` found in the parent model directory.

        High-level Logic:
        1. Fetch the `inference_output_dir` where the results will be stored, and if it does not exist, create it.
        2. Fetch the `model_name` and the `num_iterations` from the `InferenceTask` table.
        3. Load the most recent model checkpoint and the pca model from files in the `kpms_project_output_dir`.
        4. Load the keypoint data for inference as `filepath_patterns` and format it.
        5. Initialize and apply the model with the new keypoint data.
        6. If the `num_iterations` is set, fit the model with the new keypoint data for `num_iterations` iterations; otherwise, fit the model with the default number of iterations (50).
        7. Save the results as a CSV file and store the histogram showing the frequency of each syllable.
        8. Generate and save the plots showing the median trajectory of poses associated with each given syllable.
        9. Generate and save video clips showing examples of each syllable.
        10. Generate and save the dendrogram representing distances between each syllable's median trajectory.
        11. Insert the inference duration in the `Inference` table.
        12. Insert the results in the `MotionSequence` and `GridMoviesSampledInstances` tables.
        """
        from keypoint_moseq import (
            apply_model,
            format_data,
            generate_grid_movies,
            generate_trajectory_plots,
            load_checkpoint,
            load_keypoints,
            load_pca,
            plot_similarity_dendrogram,
            plot_syllable_frequencies,
            save_results_as_csv,
        )

        (
            keypointset_dir,
            inference_output_dir,
            num_iterations,
            model_id,
            pose_estimation_method,
            task_mode,
        ) = (InferenceTask & key).fetch1(
            "keypointset_dir",
            "inference_output_dir",
            "num_iterations",
            "model_id",
            "pose_estimation_method",
            "task_mode",
        )

        kpms_root = moseq_train.get_kpms_root_data_dir()
        kpms_processed = moseq_train.get_kpms_processed_data_dir()

        model_dir = find_full_path(
            kpms_processed,
            (Model & f"model_id = {model_id}").fetch1("model_dir"),
        )
        keypointset_dir = find_full_path(kpms_root, keypointset_dir)

        inference_output_dir = os.path.join(model_dir, inference_output_dir)

        if not os.path.exists(inference_output_dir):
            os.makedirs(model_dir / inference_output_dir)

        pca_path = model_dir.parent / "pca.p"
        if pca_path:
            pca = load_pca(model_dir.parent.as_posix())
        else:
            raise FileNotFoundError(
                f"No pca model (`pca.p`) found in the parent model directory {model_dir.parent}"
            )

        model_path = model_dir / "checkpoint.h5"
        if model_path:
            model = load_checkpoint(
                project_dir=model_dir.parent, model_name=model_dir.parts[-1]
            )[0]
        else:
            raise FileNotFoundError(
                f"No model (`checkpoint.h5`) found in the model directory {model_dir}"
            )

        if pose_estimation_method == "deeplabcut":
            coordinates, confidences, _ = load_keypoints(
                filepath_pattern=keypointset_dir, format=pose_estimation_method
            )
        else:
            raise NotImplementedError(
                "The currently supported format method is `deeplabcut`. If you require \
        support for another format method, please reach out to us at `support@datajoint.com`."
            )

        kpms_dj_config = load_kpms_dj_config(
            model_dir.parent.as_posix(), check_if_valid=True, build_indexes=True
        )

        if kpms_dj_config:
            data, metadata = format_data(coordinates, confidences, **kpms_dj_config)
        else:
            raise FileNotFoundError(
                f"No valid `kpms_dj_config` found in the parent model directory {model_dir.parent}"
            )

        if task_mode == "trigger":
            start_time = datetime.now(timezone.utc)
            results = apply_model(
                model=model,
                data=data,
                metadata=metadata,
                pca=pca,
                project_dir=model_dir.parent.as_posix(),
                model_name=Path(model_dir).name,
                results_path=(inference_output_dir / "results.h5").as_posix(),
                return_model=False,
                num_iters=num_iterations
                or 50,  # default internal value in the keypoint-moseq function
                **kpms_dj_config,
            )
            end_time = datetime.now(timezone.utc)

            duration_seconds = (end_time - start_time).total_seconds()

            save_results_as_csv(
                results=results,
                save_dir=(inference_output_dir / "results_as_csv").as_posix(),
            )

            fig, _ = plot_syllable_frequencies(
                results=results, path=inference_output_dir.as_posix()
            )
            fig.savefig(inference_output_dir / "syllable_frequencies.png")
            plt.close(fig)

            generate_trajectory_plots(
                coordinates=coordinates,
                results=results,
                output_dir=(inference_output_dir / "trajectory_plots").as_posix(),
                **kpms_dj_config,
            )

            sampled_instances = generate_grid_movies(
                coordinates=coordinates,
                results=results,
                output_dir=(inference_output_dir / "grid_movies").as_posix(),
                **kpms_dj_config,
            )

            plot_similarity_dendrogram(
                coordinates=coordinates,
                results=results,
                save_path=(inference_output_dir / "similarity_dendrogram").as_posix(),
                **kpms_dj_config,
            )

        else:
            from keypoint_moseq import (
                filter_centroids_headings,
                get_syllable_instances,
                load_results,
                sample_instances,
            )

            # load results
            results = load_results(
                project_dir=Path(inference_output_dir).parent,
                model_name=Path(inference_output_dir).parts[-1],
            )

            # extract syllables from results
            syllables = {k: v["syllable"] for k, v in results.items()}

            # extract and smooth centroids and headings
            centroids = {k: v["centroid"] for k, v in results.items()}
            headings = {k: v["heading"] for k, v in results.items()}

            filter_size = 9  # default value
            centroids, headings = filter_centroids_headings(
                centroids, headings, filter_size=filter_size
            )

            # extract sample instances for each syllable
            syllable_instances = get_syllable_instances(
                syllables, min_duration=3, min_frequency=0.005
            )
            # Map each syllable to a list of its sampled events.
            sampled_instances = sample_instances(
                syllable_instances=syllable_instances,
                num_samples=4 * 6,  # minimum rows * cols
                coordinates=coordinates,
                centroids=centroids,
                headings=headings,
            )

            duration_seconds = None

        self.insert1({**key, "inference_duration": duration_seconds})

        for result_idx, result in results.items():
            self.MotionSequence.insert1(
                {
                    **key,
                    "video_name": result_idx,
                    "syllable": result["syllable"],
                    "latent_state": result["latent_state"],
                    "centroid": result["centroid"],
                    "heading": result["heading"],
                }
            )

        for syllable, sampled_instance in sampled_instances.items():
            self.GridMoviesSampledInstances.insert1(
                {**key, "syllable": syllable, "instances": sampled_instance}
            )

MotionSequence

Bases: Part

Store the results of the model inference.

Attributes:

Name	Type	Description
video_name	varchar)	Name of the video.
syllable	longblob)	Syllable labels (z). The syllable label assigned to each frame (i.e. the state indexes assigned by the model).
latent_state	longblob)	Inferred low-dim pose state (x). Low-dimensional representation of the animal's pose in each frame. These are similar to PCA scores, are modified to reflect the pose dynamics and noise estimates inferred by the model.
centroid	longblob)	Inferred centroid (v). The centroid of the animal in each frame, as estimated by the model.
heading	longblob)	Inferred heading (h). The heading of the animal in each frame, as estimated by the model.

Source code in element_moseq/moseq_infer.py

class MotionSequence(dj.Part):
    """Store the results of the model inference.

    Attributes:
        video_name (varchar)                : Name of the video.
        syllable (longblob)                 : Syllable labels (z). The syllable label assigned to each frame (i.e. the state indexes assigned by the model).
        latent_state (longblob)             : Inferred low-dim pose state (x). Low-dimensional representation of the animal's pose in each frame. These are similar to PCA scores, are modified to reflect the pose dynamics and noise estimates inferred by the model.
        centroid (longblob)                 : Inferred centroid (v). The centroid of the animal in each frame, as estimated by the model.
        heading (longblob)                  : Inferred heading (h). The heading of the animal in each frame, as estimated by the model.
    """

    definition = """
    -> master
    video_name      : varchar(150)    # Name of the video
    ---
    syllable        : longblob        # Syllable labels (z). The syllable label assigned to each frame (i.e. the state indexes assigned by the model)
    latent_state    : longblob        # Inferred low-dim pose state (x). Low-dimensional representation of the animal's pose in each frame. These are similar to PCA scores, are modified to reflect the pose dynamics and noise estimates inferred by the model
    centroid        : longblob        # Inferred centroid (v). The centroid of the animal in each frame, as estimated by the model
    heading         : longblob        # Inferred heading (h). The heading of the animal in each frame, as estimated by the model
    """

GridMoviesSampledInstances

Bases: Part

Store the sampled instances of the grid movies.

Attributes:

Name	Type	Description
syllable	int)	Syllable label.
instances	longblob)	List of instances shown in each in grid movie (in row-major order), where each instance is specified as a tuple with the video name, start frame and end frame.

Source code in element_moseq/moseq_infer.py

class GridMoviesSampledInstances(dj.Part):
    """Store the sampled instances of the grid movies.

    Attributes:
        syllable (int)                  : Syllable label.
        instances (longblob)            : List of instances shown in each in grid movie (in row-major order), where each instance is specified as a tuple with the video name, start frame and end frame.
    """

    definition = """
    -> master
    syllable: int           # Syllable label
    ---
    instances: longblob     # List of instances shown in each in grid movie (in row-major order), where each instance is specified as a tuple with the video name, start frame and end frame
    """

make(key)

This function is used to infer the model results from the checkpoint file and store the results in MotionSequence and GridMoviesSampledInstances tables.

Parameters:

Name	Type	Description	Default
key	dict	InferenceTask primary key.	required

Raises:

Type	Description
FileNotFoundError	If no pca model (pca.p) found in the parent model directory.
FileNotFoundError	If no model (checkpoint.h5) found in the model directory.
NotImplementedError	If the format method is not deeplabcut.
FileNotFoundError	If no valid kpms_dj_config found in the parent model directory.

High-level Logic: 1. Fetch the inference_output_dir where the results will be stored, and if it does not exist, create it. 2. Fetch the model_name and the num_iterations from the InferenceTask table. 3. Load the most recent model checkpoint and the pca model from files in the kpms_project_output_dir. 4. Load the keypoint data for inference as filepath_patterns and format it. 5. Initialize and apply the model with the new keypoint data. 6. If the num_iterations is set, fit the model with the new keypoint data for num_iterations iterations; otherwise, fit the model with the default number of iterations (50). 7. Save the results as a CSV file and store the histogram showing the frequency of each syllable. 8. Generate and save the plots showing the median trajectory of poses associated with each given syllable. 9. Generate and save video clips showing examples of each syllable. 10. Generate and save the dendrogram representing distances between each syllable's median trajectory. 11. Insert the inference duration in the Inference table. 12. Insert the results in the MotionSequence and GridMoviesSampledInstances tables.

Source code in element_moseq/moseq_infer.py

def make(self, key):
    """
    This function is used to infer the model results from the checkpoint file and store the results in `MotionSequence` and `GridMoviesSampledInstances` tables.

    Args:
        key (dict): `InferenceTask` primary key.

    Raises:
        FileNotFoundError: If no pca model (`pca.p`) found in the parent model directory.
        FileNotFoundError: If no model (`checkpoint.h5`) found in the model directory.
        NotImplementedError: If the format method is not `deeplabcut`.
        FileNotFoundError: If no valid `kpms_dj_config` found in the parent model directory.

    High-level Logic:
    1. Fetch the `inference_output_dir` where the results will be stored, and if it does not exist, create it.
    2. Fetch the `model_name` and the `num_iterations` from the `InferenceTask` table.
    3. Load the most recent model checkpoint and the pca model from files in the `kpms_project_output_dir`.
    4. Load the keypoint data for inference as `filepath_patterns` and format it.
    5. Initialize and apply the model with the new keypoint data.
    6. If the `num_iterations` is set, fit the model with the new keypoint data for `num_iterations` iterations; otherwise, fit the model with the default number of iterations (50).
    7. Save the results as a CSV file and store the histogram showing the frequency of each syllable.
    8. Generate and save the plots showing the median trajectory of poses associated with each given syllable.
    9. Generate and save video clips showing examples of each syllable.
    10. Generate and save the dendrogram representing distances between each syllable's median trajectory.
    11. Insert the inference duration in the `Inference` table.
    12. Insert the results in the `MotionSequence` and `GridMoviesSampledInstances` tables.
    """
    from keypoint_moseq import (
        apply_model,
        format_data,
        generate_grid_movies,
        generate_trajectory_plots,
        load_checkpoint,
        load_keypoints,
        load_pca,
        plot_similarity_dendrogram,
        plot_syllable_frequencies,
        save_results_as_csv,
    )

    (
        keypointset_dir,
        inference_output_dir,
        num_iterations,
        model_id,
        pose_estimation_method,
        task_mode,
    ) = (InferenceTask & key).fetch1(
        "keypointset_dir",
        "inference_output_dir",
        "num_iterations",
        "model_id",
        "pose_estimation_method",
        "task_mode",
    )

    kpms_root = moseq_train.get_kpms_root_data_dir()
    kpms_processed = moseq_train.get_kpms_processed_data_dir()

    model_dir = find_full_path(
        kpms_processed,
        (Model & f"model_id = {model_id}").fetch1("model_dir"),
    )
    keypointset_dir = find_full_path(kpms_root, keypointset_dir)

    inference_output_dir = os.path.join(model_dir, inference_output_dir)

    if not os.path.exists(inference_output_dir):
        os.makedirs(model_dir / inference_output_dir)

    pca_path = model_dir.parent / "pca.p"
    if pca_path:
        pca = load_pca(model_dir.parent.as_posix())
    else:
        raise FileNotFoundError(
            f"No pca model (`pca.p`) found in the parent model directory {model_dir.parent}"
        )

    model_path = model_dir / "checkpoint.h5"
    if model_path:
        model = load_checkpoint(
            project_dir=model_dir.parent, model_name=model_dir.parts[-1]
        )[0]
    else:
        raise FileNotFoundError(
            f"No model (`checkpoint.h5`) found in the model directory {model_dir}"
        )

    if pose_estimation_method == "deeplabcut":
        coordinates, confidences, _ = load_keypoints(
            filepath_pattern=keypointset_dir, format=pose_estimation_method
        )
    else:
        raise NotImplementedError(
            "The currently supported format method is `deeplabcut`. If you require \
    support for another format method, please reach out to us at `support@datajoint.com`."
        )

    kpms_dj_config = load_kpms_dj_config(
        model_dir.parent.as_posix(), check_if_valid=True, build_indexes=True
    )

    if kpms_dj_config:
        data, metadata = format_data(coordinates, confidences, **kpms_dj_config)
    else:
        raise FileNotFoundError(
            f"No valid `kpms_dj_config` found in the parent model directory {model_dir.parent}"
        )

    if task_mode == "trigger":
        start_time = datetime.now(timezone.utc)
        results = apply_model(
            model=model,
            data=data,
            metadata=metadata,
            pca=pca,
            project_dir=model_dir.parent.as_posix(),
            model_name=Path(model_dir).name,
            results_path=(inference_output_dir / "results.h5").as_posix(),
            return_model=False,
            num_iters=num_iterations
            or 50,  # default internal value in the keypoint-moseq function
            **kpms_dj_config,
        )
        end_time = datetime.now(timezone.utc)

        duration_seconds = (end_time - start_time).total_seconds()

        save_results_as_csv(
            results=results,
            save_dir=(inference_output_dir / "results_as_csv").as_posix(),
        )

        fig, _ = plot_syllable_frequencies(
            results=results, path=inference_output_dir.as_posix()
        )
        fig.savefig(inference_output_dir / "syllable_frequencies.png")
        plt.close(fig)

        generate_trajectory_plots(
            coordinates=coordinates,
            results=results,
            output_dir=(inference_output_dir / "trajectory_plots").as_posix(),
            **kpms_dj_config,
        )

        sampled_instances = generate_grid_movies(
            coordinates=coordinates,
            results=results,
            output_dir=(inference_output_dir / "grid_movies").as_posix(),
            **kpms_dj_config,
        )

        plot_similarity_dendrogram(
            coordinates=coordinates,
            results=results,
            save_path=(inference_output_dir / "similarity_dendrogram").as_posix(),
            **kpms_dj_config,
        )

    else:
        from keypoint_moseq import (
            filter_centroids_headings,
            get_syllable_instances,
            load_results,
            sample_instances,
        )

        # load results
        results = load_results(
            project_dir=Path(inference_output_dir).parent,
            model_name=Path(inference_output_dir).parts[-1],
        )

        # extract syllables from results
        syllables = {k: v["syllable"] for k, v in results.items()}

        # extract and smooth centroids and headings
        centroids = {k: v["centroid"] for k, v in results.items()}
        headings = {k: v["heading"] for k, v in results.items()}

        filter_size = 9  # default value
        centroids, headings = filter_centroids_headings(
            centroids, headings, filter_size=filter_size
        )

        # extract sample instances for each syllable
        syllable_instances = get_syllable_instances(
            syllables, min_duration=3, min_frequency=0.005
        )
        # Map each syllable to a list of its sampled events.
        sampled_instances = sample_instances(
            syllable_instances=syllable_instances,
            num_samples=4 * 6,  # minimum rows * cols
            coordinates=coordinates,
            centroids=centroids,
            headings=headings,
        )

        duration_seconds = None

    self.insert1({**key, "inference_duration": duration_seconds})

    for result_idx, result in results.items():
        self.MotionSequence.insert1(
            {
                **key,
                "video_name": result_idx,
                "syllable": result["syllable"],
                "latent_state": result["latent_state"],
                "centroid": result["centroid"],
                "heading": result["heading"],
            }
        )

    for syllable, sampled_instance in sampled_instances.items():
        self.GridMoviesSampledInstances.insert1(
            {**key, "syllable": syllable, "instances": sampled_instance}
        )