Callbacks

Callbacks are used to control to control how a runner handles the results (features, labels or model weights).

They are classes implementing a pre-defined set of functions:

• save_features
• save_label_scores
• save_bounding_boxes
• save_frames

In memory callbacks

These callbacks accumulate results in-memory. They expose their results via object attributes.

mozuma.callbacks.memory.CollectFeaturesInMemory dataclass

Callback to collect features in memory

Attributes:

Name	Type	Description
indices	list	List of dataset indices
features	numpy.ndarray	Array of features. The first dimension correspond to self.indices values.

Note

This callback works with any array-like features

save_features(self, model: Any, indices: Sequence, features: Union[torch.Tensor, numpy.ndarray]) -> None

Save features output returned by a module

Parameters:

Name	Type	Description	Default
model	Any	The MoZuMa model that produced the features	required
indices	Sequence	The list of indices as defined by the dataset	required
features	ArrayLike	The feature object as returned by the model	required

mozuma.callbacks.memory.CollectLabelsInMemory dataclass

Callback to collect labels in memory

Attributes:

Name	Type	Description
indices	list	List of dataset indices
label_scores	numpy.ndarray	Array of label scores. The first dimension correspond to self.indices values.
labels	list[str]	List of matching labels (label with maximum score)

Note

This callback works with any array-like features

save_label_scores(self, model: ModelWithLabels, indices: Sequence, labels_scores: Union[torch.Tensor, numpy.ndarray]) -> None

Save labels scores returned by a module

Parameters:

Name	Type	Description	Default
model	ModelWithLabels	The MoZuMa model that produced the label scores	required
indices	Sequence	The list of indices as defined by the dataset	required
labels_scores	ArrayLike	Contains the output score/probability for each label	required

mozuma.callbacks.memory.CollectBoundingBoxesInMemory dataclass

Callback to collect bounding boxes predictions in memory

Attributes:

Name	Type	Description
indices	list	List of dataset indices
frames	list[BatchVideoFramesPrediction[np.ndarray]]	Sequence of video frames. The first dimension correspond to self.indices values.

Note

This callback works with any array-like features

save_bounding_boxes(self, model: Any, indices: Sequence, bounding_boxes: Sequence[mozuma.predictions.BatchBoundingBoxesPrediction[Union[torch.Tensor, numpy.ndarray]]]) -> None

Save bounding boxes output of a module

Parameters:

Name	Type	Description	Default
model	Any	The model that produces the bounding boxes	required
indices	Sequence	The list of indices as defined by the dataset	required
bounding_boxes	Sequence[BatchBoundingBoxesPrediction[_ContraArrayType]]	The sequence bounding predictions	required

mozuma.callbacks.memory.CollectVideoFramesInMemory dataclass

Callback to collect video frames in memory

Attributes:

Name	Type	Description
indices	list	List of dataset indices
frames	list[BatchVideoFramesPrediction[np.ndarray]]	Sequence of video frames. The first dimension correspond to self.indices values.

Note

This callback works with any array-like features

save_frames(self, model: Any, indices: Sequence, frames: Sequence[mozuma.predictions.BatchVideoFramesPrediction[Union[torch.Tensor, numpy.ndarray]]]) -> None

Save frames extracted from a video

Parameters:

Name	Type	Description	Default
model	Any	The model that produces the video frames encoding	required
indices	Sequence	The list of indices as defined by the dataset	required
frames	Sequence[BatchVideoFramesPrediction[_ArrayType]]	The sequence of frame features and indices	required

Callbacks for training

mozuma.callbacks.states.SaveModelState dataclass

Simple callback to save model state during training.

If state are saved during training (every X epochs, see TorchTrainingOptions) the current epoch number is appended to the state_key.training_id in the following way: <state_key.training_id>-e<num_epoch>. When the training is complete, just the state_key.training_id is used.

Attributes:

Name	Type	Description
store	AbstractStateStore	Object to handle model state saving
state_key	StateKey	State identifier for the training activity.

Warning

This callback only saves the model state, thus does not create a whole training checkpoint (optimizer state, loss, etc..).

save_model_state(self, engine: Engine, model: Any) -> None

Save model state by calling the state store

Parameters:

Name	Type	Description	Default
model	Any	The model to save	required

Write your own callbacks

mozuma.callbacks.base.BaseSaveFeaturesCallback

save_features(self, model: Any, indices: Sequence, features: -_ContraArrayType) -> None

Save features output returned by a module

Parameters:

Name	Type	Description	Default
model	Any	The MoZuMa model that produced the features	required
indices	Sequence	The list of indices as defined by the dataset	required
features	ArrayLike	The feature object as returned by the model	required

mozuma.callbacks.base.BaseSaveLabelsCallback

save_label_scores(self, model: ModelWithLabels, indices: Sequence, labels_scores: -_ContraArrayType) -> None

Save labels scores returned by a module

Parameters:

Name	Type	Description	Default
model	ModelWithLabels	The MoZuMa model that produced the label scores	required
indices	Sequence	The list of indices as defined by the dataset	required
labels_scores	ArrayLike	Contains the output score/probability for each label	required

mozuma.callbacks.base.BaseSaveBoundingBoxCallback

save_bounding_boxes(self, model: Any, indices: Sequence, bounding_boxes: Sequence[mozuma.predictions.BatchBoundingBoxesPrediction[-_ContraArrayType]]) -> None

Save bounding boxes output of a module

Parameters:

Name	Type	Description	Default
model	Any	The model that produces the bounding boxes	required
indices	Sequence	The list of indices as defined by the dataset	required
bounding_boxes	Sequence[BatchBoundingBoxesPrediction[_ContraArrayType]]	The sequence bounding predictions	required

mozuma.callbacks.base.BaseSaveVideoFramesCallback

save_frames(self, model: Any, indices: Sequence, frames: Sequence[mozuma.predictions.BatchVideoFramesPrediction[-_ContraArrayType]]) -> None

Save frames extracted from a video

Parameters:

Name	Type	Description	Default
model	Any	The model that produces the video frames encoding	required
indices	Sequence	The list of indices as defined by the dataset	required
frames	Sequence[BatchVideoFramesPrediction[_ArrayType]]	The sequence of frame features and indices	required

mozuma.callbacks.base.BaseRunnerEndCallback

on_runner_end(self, model: Any) -> None

Called when the runner finishes

This can be used to do clean up. For instance if the data is being processed by a thread, this function can wait for the thread to finish.$

Parameters:

Name	Type	Description	Default
model	Any	The currently run model	required